content stringlengths 1 15.9M |
|---|
\section{Introduction}
\label{sec:introduction}
Knowledge of the equation of state and the phase diagram of QCD is
essential in understanding the properties of the deconfined matter
created in heavy-ion collisions as well as the properties of compact
stars and their quark cores.
In non-central heavy-ion collisions, large time-dependent magnetic
fields are generated during the experiment~\cite{harmen,tonev,bzdak}. The maximum
strength of these magnetic fields is on the order of $10^{19}$ Gauss ($qB \sim$ 6 $m_\pi^2$).
Likewise, very strong magnetic fields exist inside
magnetars~\cite{neutron}.
These may be several orders of magnitude larger than
the magnetic fields in
ordinary neutron stars.
On the surface, the magnetic field may be as strong as $10^{14}-10^{15}$ Gauss
and it could be as strong as $10^{16}-10^{19}$ Gauss in the interior
of the star. This has spurred the interest in strongly interacting matter
at finite temperature, density and magnetic field, see for example
Ref.~\cite{overview}
for a recent review.
The phase boundary in $(T,\mu_B,B$) space is therefore of great
interest;
however due to the infamous sign problem, one cannot use the standard
techniques of lattice calculations at finite $\mu_B$.
At zero $\mu_B$ and finite $B$, there is no sign problem and so
one can calculate the phase diagram in the $T,B$ plane using Monte-Carlo
methods. Recent lattice calculations~\cite{budaleik,gunnar}
suggest that for physical quark masses, the transition temperature
for the chiral transition is a decreasing function of the magnetic field
$B$, while for larger values of the quark masses corresponding to
$m_{\pi}\simeq400$ MeV the temperature is an increasing function of
$B$~\cite{sanf,negro}. The qualitative behavior of the
transition temperature for physical quark masses in is disagreement with
model calculations using either the (Polyakov-loop extended)
Nambu-Jona-Lasinio ((P)NJL)
model or the (Polyakov-loop extended) quark-meson model ((P)QM);
In these models, the critical temperature is an
increasing function of the
magnetic field, see
e.g.~\cite{fragapol,sadooghi,pnjlgat,pnjlkas,duarte,skokov,anders,grecoleik}.
Possible resolutions to the disagreement have
been suggested~\cite{bag,res1,nanko,inverse,huang,ferry,orlov}
and we will discuss these at the end of the paper.
In a previous paper~\cite{anders}, two of us used the
two-flavor three-color quark-meson model and the functional
renormalization group~\cite{wetterich}
to map out the phase diagram in the $\mu_B-T$ plane for different values
of the magnetic field (see also Refs.~\cite{ebert22,pinto}).
In the present paper, we add the Polyakov loop
to the model to
include certain aspects of confinement~\cite{fukushima,megias}.
In particular, we investigate a set of possible implementations
of the Polyakov loop
and how they effect both the chiral and deconfinement transitions.
In the context of the functional renormalization
group, this was studied in Ref.~\cite{skokov} at zero baryon chemical
potential.
The paper is organized as follows: In Sec.~\ref{sec:quarkmeson} we
briefly discuss
the functional renormalization group implementation of
the quark-meson model in a constant magnetic background.
In Sec.~\ref{sec:poly} we add the Polyakov
loop variable to the model and review the three gluonic potentials
we have used in this work. Section~\ref{sec:numerics}
explans the numerical
implementation and the effects of the various gluonic potentials.
In Sec.~\ref{sec:results} we discuss our results for the deconfinement and
chiral transitions. Finally, in Sec.~\ref{sec:conclusion}, we
summarise the main results and
comment on the disagreement between lattice and model calculations at finite
$B$ and $\mu_B=0$.
\section{Quark-meson model and the functional renormalization
group}
\label{sec:quarkmeson}
The quark meson model is the linear sigma model coupled
to two massless quark flavors via a Yukawa coupling.
The Euclidean Lagrangian for the model is
\begin{eqnarray}\nonumber
{\cal L}&=&
\bar{\psi}\bigg[
\gamma_{\mu}\partial_{\mu}-\mu\gamma_4
+g(\sigma-i\gamma_5{\boldsymbol \tau}\cdot{\boldsymbol \pi})\bigg]\psi
\label{lagra}
+{1\over2}\bigg[
(\partial_{\mu}\sigma)^2
+(\partial_{\mu}{\boldsymbol \pi})^2
\bigg]
+{1\over2}m^2\bigg[\sigma^2+{\boldsymbol \pi}^2\bigg]
\\ &&
+{\lambda\over4}\bigg[
\sigma^2+{\boldsymbol \pi}^2\bigg]^2
-h\sigma\;,
\end{eqnarray}
where $\sigma$ is the sigma field, ${\boldsymbol \pi}$ denotes the
neutral and charged pions.
${\boldsymbol \tau}$ are the Pauli matrices,
$\mu={\rm diag}(\mu_u,\mu_d)$ is the quark chemical potential,
where
$\mu_u$ and $\mu_d$ are the chemical potential for the $u$ and $d$ quarks,
respectively.
We set $\mu_u=\mu_d$ so that we are working at zero isospin
chemical potential, $\mu_I=\mbox{$1\over2$}(\mu_u-\mu_d)=0$.
The baryon chemical potential is given by $\mu_B=3\mu$.
The Euclidean $\gamma$ matrices
are given by $\gamma_j=i\gamma^j_M$, $\gamma_4=\gamma^0_M$,
and $\gamma_5=-\gamma^5_M$, where
the index $M$ denotes Minkowski space.
The fermion field is an isospin doublet,
\bq
\psi=
\left(\begin{array}{c}
u\\
d\\
\end{array}\right)\;.
\label{d0}
\end{eqnarray}
If $h=0$, Eq.~(\ref{lagra}) is invariant under $O(4)$.
If $h\neq0$, chiral symmetry is explicitly broken, otherwise
it is spontaneously broken in the vacuum. Either way, the symmetry is
reduced to $O(3)$.
Since $SU(2)_L\times SU(2)_R\sim O(4)$ and $SU(2)_V\sim O(3)$,
the quark-meson model incorporates the global symmetries of two-flavor
QCD, whether or not the $SU(2)_A$-symmetry is broken explicitly
by finite quark masses.
Chiral symmetry is broken in the vacuum by a nonzero
expectation value $\phi$ for the sigma field and we
make the replacement
\begin{eqnarray}
\sigma\rightarrow\phi+\tilde{\sigma}\;,
\end{eqnarray}
where $\tilde{\sigma}$ is a quantum fluctuating field.
The tree-level potential then becomes
\begin{eqnarray}
U_{\Lambda}&=&
{1\over2}m^2_{\Lambda}\phi^2
+{\lambda_{\Lambda}\over4}\phi^4
-h\phi\;.
\label{tree}
\end{eqnarray}
Note that we have introduced a
subscript $\Lambda$ on $U$,
$m^2$, and $\lambda$, where $\Lambda$ is the ultraviolet cutoff of the
theory. This
is a reminder that these are unrenormalized quantities\footnote{
The symmetry breaking term
is equivalent to an external field that does not flow and
therefore $h=h_{\Lambda}$.}.
We will follow Wetterich's implementation of the renormalization
group
ideas based on the effective average action
$\Gamma_k[\varphi]$~\cite{wetterich}.
This action is a functional of a set of background fields that
are denoted by $\varphi$. $\Gamma_k[\varphi]$
satisfies
an integro-differential flow equation in the variable $k$, to be specified
below.
The subscript $k$ indicates that all the modes $p$ between the ultraviolet
cutoff $\Lambda$ of the theory and $k$ have been integrated out.
When $k=\Lambda$ no modes have been integrated out and $\Gamma_{\Lambda}$
equals the classical action $S$. On the other hand, when $k=0$, all the
momentum modes have been integrated out and $\Gamma_0$ equals the full
quantum effective action. The flow equation then describes the flow in the
space of effective actions as a function of $k$.
In order to implement the renormalization
group ideas, one introduces a regulator function $R_k(p)$.
The function $R_k(p)$
is
large for $p < k$ and small for $p > k$ whenever $0 < k < \Lambda$,
and $R_{\Lambda} (p) =\infty$. These properties ensure
that the modes below $k$ are heavy and decouple, and only the modes between $k$
and the UV cutoff $\Lambda$ are light and integrated out.
The choice of regulator function
has been discussed in detail in the literature and some choices are
better than others due both to their analytical and stability
properties, see for example~\cite{litim}.
The flow equation for the effective action cannot be solved exactly
so one must make tractable and yet physically sound approximations.
The first approximation in a derivative expansion is the local-potential
approximation
and in this case the flow equation for $\Gamma_k$ reduces to a
flow equation for an effective potential $U_k(\phi)$.
In the case of a constant magnetic field,
the differential equation for $U_k$ first appeared in~\cite{skokov}
and a derivation can be found in the preceeding work~\cite{anders}. It reads
\begin{eqnarray}\nonumber
\partial_k U_k
&=&
{k^4\over12\pi^2}
\left\{
{1\over\omega_{1,k}}\left[1+2n_B(\omega_{1,k})\right]
+{1\over\omega_{k,2}}\left[1+2n_B(\omega_{2,k})\right]
\right\}
\\ &&\nonumber
+k{|qB|\over2\pi^2}\sum_{m=0}^{\infty}{1\over\omega_{1,k}}
\sqrt{k^2-p^2_{\perp}(q,m,0)}\,\theta\left(k^2-p^2_{\perp}(q,m,0)\right)
\left[1+2n_B(\omega_{1,k})\right]
\\ && \nonumber
-{N_c\over2\pi^2}k\sum_{s,f,m=0}^{\infty}
{|q_fB|\over\omega_{q,k}}
\sqrt{k^2-p^2_{\perp}(q_f,m,s)}\,\theta\left(k^2-p^2_{\perp}(q_f,m,s)\right)
\left[1-n^+_F(\omega_{q,k})-n^-_F(\omega_{q,k})
\right]\;,
\\ &&
\label{flowu}
\end{eqnarray}
where we have defined
$\omega_{1,k}=\sqrt{k^2+U_k^{\prime}}\,$,
$\omega_{2,k}=\sqrt{k^2+U^{\prime}+2U_k^{\prime\prime}\rho}\,$,
$\omega_{q,k}=\sqrt{k^2+2g^2\rho}\,$,
$p^2_{\perp}(q,m,s)=(2m+1-s)|qB|\,$,
$n_B(x)=1/(e^{\beta x}-1)\,$, $\rho={1\over2}\phi^2$ and
$n_F^{\pm}(x)=1/(e^{\beta(x\pm\mu)}+1)$, however the fermionic
distribution functions will be transformed to Eqs.~(\ref{gen1}) and (\ref{gen2})
when we add the Polyakov loop.
At zero temperature, the Bose distribution function vanishes and the
Fermi distribution function becomes a step function.
Furthermore, if we set $\mu=0$, this step function
vanishes and we obtain the flow equation in the vacuum.
\section{Adding the Polyakov loop }
\label{sec:poly}
The Polyakov loop $\Phi$ is given by the thermal expectation value of the
trace of the Wilson line, i.e.\
\begin{eqnarray}
\Phi={1\over N_c}\langle
{\rm Tr}_c\,L\rangle\;,
\end{eqnarray}
where the trace is in color space and
\begin{equation}
L = {\cal P} \exp\left[i\int_0^{\beta}
d\tau \,A_4
\right]\;,
\end{equation}
where $A_4=iA_0$ and $A_0=\delta_{\mu0}{\cal A}_a^{\mu}t^a$.
Here ${\cal A}_a^{\mu}$ are the $SU(3)_c$ gauge fields and
the generators are $t^a={1\over2}\lambda^a$,
where $\lambda^a$ are the Gell-Mann matricies.
The Wilson line is a complex variable and so $\Phi$
is not equal to
${\bar \Phi}={1\over N_c}\langle{\rm Tr}_c\,L^\dagger\rangle$ in general.
It is known that
$\Phi=\bar{\Phi}$ at mean field level, but in the present work this is
only true at zero baryon chemical potential.
The Polyakov loop is an order parameter for deconfinement in pure-glue
QCD. Under the center symmetry $Z_N$, it transforms as
$\Phi\rightarrow e^{2\pi in/N_c}$, where $n=0,1,2...,N_c-1$.
At low temperatures, i.e.\ in the confined phase we have $\Phi\approx0$,
while in the deconfined phase we have $\Phi\approx1$.
Coupling the Polyakov loop to the QM model gives a more physically accurate
model of the quark sector and allows us to explore both the chiral
and deconfinement transitions of low energy QCD.
This is done by introducing a constant
background temporal
gauge field ${\cal A}_a^0$
via the covariant derivative and adding a phenomenological
potential for the gluonic sector, as discussed below.
The Polyakov gauge is particularly convenient for calculations as
the Wilson line is then a diagonal matrix, $L=e^{i(\lambda^3A_3+\lambda^8A_8)/2T}$.
Utilizing this and the mean field solution for the effective potential the
quark distribution
functions are found to be transformed from the standard Fermi-Dirac distribution functions to
\begin{align}
\label{gen1}
n_F^+(\Phi, {\bar \Phi}; T, \mu) &= \frac{1+2{\bar \Phi}e^{\beta(E_q-\mu)}
+ \Phi e^{2\beta(E_q-\mu)}}{1+3{\bar\Phi} e^{\beta(E_q-\mu)}+3\Phi e^{2\beta(E_q-\mu)}
+e^{3\beta(E_q-\mu)}}\;,\\
n_F^-(\Phi, {\bar \Phi}; T, \mu) &=n_F^+({\bar \Phi}, \Phi; T, -\mu) \; .
\label{gen2}
\end{align}
\noindent These are then substituted back into the
renormalization group flow equation (\ref{flowu}).
This form is a particularly promising result, as in the confining
limit ($\Phi$ and ${\bar\Phi} \rightarrow 0$) we obtain a Fermi-Dirac-like
distribution function for states of three quarks, however as $\Phi$ and
${\bar\Phi} \rightarrow 1$ the functions $n_F^{\pm}$ are equal to the
standard Fermi-Dirac distribution functions, as they should be.
A number of forms for the gluonic potentials have been proposed
and investigated at mean field level for the PNJL model~\cite{lourenco}
and the PQM model with $\mu=0$~\cite{schaefer2010}. In
this work we will investigate three different gluon potentials.
Since the Polyakov loop variable is the order parameter for the $Z(3)$
center symmetry of pure-glue QCD, a Ginzburg-Landau type potential
should incorporate this. A polynomial expansion then leads
to~\cite{polyakovpot1}
\begin{align}
\frac{U_{\rm poly}}{T^4} &= -\frac{b_2(T)}{2} \Phi{\bar\Phi} -\frac{b_3}{6}
\big( \Phi^3+{\bar\Phi}^{3} \big) + \frac{b_4}{4} \big( \Phi{\bar\Phi} \big)^2
\; , \label{poly}
\end{align}
where the coefficients are
\begin{eqnarray}
b_2(T) &=& 6.75 - 1.95\left(\frac{T_0}{T}\right) + 2.624\left(\frac{T_0}
{T}\right)^2 - 7.44\left(\frac{T_0}{T}\right)^3 \; , \\
b_3 &=& 0.75\;,\\
b_4 &=& 7.5\;.
\end{eqnarray}
The coefficients $b_2(T)$, $b_3$, and $b_4$ are chosen such that the
Polyakov loop potential reproduces the equation of state and
temperature dependence of $\Phi$ around the transition at $\mu=0$.
The parameter $T_0$ is the transition temperature
for pure-glue QCD
lattice calculations~\cite{tc270}.
In Refs.~\cite{polyakovpot2,polyakovpot3}, another form for the
Polyakov loop potential based on the $SU(3)$ Haar measure was proposed:
\begin{align}
\frac{U_{\rm log}}{T^4} &= -\frac{a(T)}{2} \Phi{\bar\Phi} +b(T) \ln
\left[ 1- 6\,{\bar\Phi}\Phi + 4 \big( \Phi^3+{\bar\Phi}^{3} \big) -
3\big( {\bar\Phi}\Phi \big)^2 \right] \; , \label{log}
\end{align}
where the coefficients are
\begin{align}
a(T) &= 3.51 - 2.47\left(\frac{T_0}{T}\right) + 15.2\left(\frac{T_0}{T}
\right)^2 \; , \\
b(T) &= -1.75\left(\frac{T_0}{T}\right)^3\;.
\end{align}
We note that the logarithmic term ensures that the magnitude of
$\Phi$ and $\bar{\Phi}$ are constrained to be in the region between
$-1$ and $1$, i.e.\ the possible attainable
values for the
normalized trace of an element of the $SU(3)$.
Finally, Fukushima proposed a Polyakov loop potential in~\cite{polyakovpot4}
\begin{align}
\frac{U_{\rm Fuku}}{T^4} &= -\frac{b}{T^3} \left( 54e^{-a\,T_0/T}\Phi{\bar\Phi}
+\ln \left[ 1- 6\,\Phi{\bar\Phi} + 4 \big( \Phi^3+{\bar\Phi}^{3} \big)
- 3\big( \Phi{\bar\Phi} \big)^2 \right] \right)\; , \label{Fuku}
\end{align}
where the constants are $a=664/270$ and $b=(196.2\textrm{ MeV})^3$ and we
have added dependence upon the transition temperature, $T_0$.
A problem with all the Polyakov loop potentials proposed is that
they are independent of the number of flavors and of the baryon chemical
potential. However, we know that, for example, the transition temperature
for the deconfinement transition is a function of $N_f$.
In other words, one ought to incorporate the back-reaction from the
fermions to the gluonic sector. In Ref.~\cite{bj}, the authors
use perturbative arguments to estimate the effects of the number
of flavors and the baryon chemical potential on the transition temperature
$T_0$. The functional form of $T_0$ is~\cite{herbst}
\begin{equation}
T_0 = T_\tau e^{-1/(\alpha_0 \, b(N_f,\mu))}\;,
\label{nfmub}
\end{equation}
where
\begin{equation}
b(N_f,\mu) = \frac{1}{6\pi}(11N_c - 2N_f) - \frac{16}{\pi}N_f \frac{\mu^2}
{(\hat{\gamma} \, T_\tau)^2}\;,
\label{bpar}
\end{equation}
and $T_\tau = 1.77 \textrm{ GeV}$, $\alpha_0 = 0.304$. ${\hat\gamma}$ controls
the curvature of $T_0$ as a function of $\mu$, and again
following~\cite{herbst} we experiment with a range of values to study the
effects. This is further discussed in the following section.
Let us finally make a few remarks about the sign problem.
At finite baryon chemical potential, QCD has a sign problem due to
a complex fermion determinant. This implies that the action is complex
and one cannot use standard Monte-Carlo techniques based on importance
sampling. Also effective models
such as the PNJL and PQM models have a sign problem at finite
baryon chemical potential as discussed in
Refs.~\cite{sign1,sign2,sign3,pbarp} for example. The sign problem in these
models
shows up as an imaginary part of the effective potential and one must
therefore consider it as a complex function of complex
variables $\Phi$ and $\bar{\Phi}$.
There are two ways out. One way is to restrict the
Polyakov loop variables to be real as in Ref.~\cite{fukushima}.
This is the approach we will follow in the present paper.
The other option is to split the effective potential into a real part
and an imaginary part~\cite{sign3,pbarp} and treat the imaginary part
as a perturbation. While this is no longer the case when including
perturbative corrections, at the mean-field level this implies $\Phi=\bar{\Phi}$.
\section{Numerical implementation and the glue potential}
\label{sec:numerics}
To find the equilibrium state values of the order parameters
$\phi$, $\Phi$ and ${\bar\Phi}$ we numerically solve the flow equation
(\ref{flowu}) with the boundary condition specified by the tree level
potential, Eq.~(\ref{tree}), on a grid in $\phi$-$\Phi$-${\bar\Phi}$-space
with $\phi \in [0,126]$ MeV and $\Phi , {\bar\Phi} \in [0,1]$ ($\Phi$
and ${\bar\Phi}$ are real, as discussed Sec.~\ref{sec:poly}).
Doing this at various values of $T$, $B$ and $\mu$ gives
us $U_{k=0}(\phi,\Phi,{\bar\Phi};T,B,\mu)$, which we construct as a dimensionless quantity.
In the derivation of the flow
equation we have used $O(4)$ symmetry, thus for the boundary condition of
the flow we set $h=0$, then when minimising with respect to $\phi$ we
minimise $U_{k=0} - h\phi$. The resulting surface, $U_{k=0}(\Phi , {\bar\Phi})$
is very smooth thus we use interpolation to save computation time. Additional
runs at intermediate values show that errors due to the interpolation are
on the order of 0.1\%. Before we minimise with respect to the deconfinement
order parameters we must add the gluonic potential. Thus $\Phi$
and ${\bar\Phi}$ are obtained from the minimisation of
$U_{k=0}(\Phi , {\bar\Phi}) + U_{\textrm{glue}}(\Phi , {\bar\Phi})/\Lambda^4$,
where `glue' stands for one of `poly', `log' or `Fuku' as given in
Sec.~\ref{sec:poly}.
We use the following (dimensionless) bare parameters: $m_\Lambda^2 = 0.075$,
$\lambda_\Lambda = 9.2$, $g=3.2258$ and $h=0.0146$ and we have
$\Lambda=500$~MeV which
give constituent quark masses of 300~MeV, a sigma mass of $\sim$478~MeV and pion
masses of $\sim$140~MeV, that is, our results are calculated at the physical
point. Changing the energy of the ultraviolet cutoff from 500 to 800 MeV, gives an
increase of approximately 3\% to the chiral phase transition at low $\mu$, and
approximately 10\% at low $T$.
Additional details about the implementation at $\Phi={\bar\Phi}=1$ can
be found in~\cite{anders}.
As the results presented here are calculated at the physical point all of the
phase transitions are
crossover `transitions' and thus all critical temperatures are
pseudo-critical temperatures.
We must therefore define how we can calculate these transitions.
Since we have discretized the variables in the computation
of the effective potential, calculating the inflection point directly from
the output data is very inaccurate. Thus
one way to define the transition temperature is to fit the data points
for the order parameter in question with a function and then define the
transition
temperature, $T_{\textrm{x}}$, as the inflection point of the fitted curve.
For the chiral transition we use this method,
with the fit based on $\arctan(x)$.
However, using this method for the deconfinement transition we run into
problems as the functional form of the underlying curve changes with
changing $\mu$ (see the left panel of Fig.~\ref{dcop}). An alternative way of
defining this transition is
when the order parameter, $\Phi(T)$, is equal to ${1\over2}$, this we define
as $T_{\Phi/2}$.
To find this point we interpolate with third-order polynomial interpolation.
Fig~\ref{tcdef} illustrates this for $\mu=0$.
The left panel shows the data points (crosses) for $\phi$
as a function of $T$. The open circle indicates the inflection point
of the fitted curve, i.e. $T_{\phi}$, while the cross indicates the
temperature when the normalized chiral order parameter satisfies
$\phi/\phi(T=0)={1\over2}$, we denote this $T_{\phi/2}$.
The right panel shows the same, but now for the deconfinement
order parameter $\Phi$ and the green curve is now the interpolation used to
determine $T_{\Phi/2}$.
\begin{figure}[htb]
\begin{center}
\setlength{\unitlength}{1mm}
\includegraphics[width=16.0cm]{TransitionDef01.eps}
\caption{Methods used to determine the transition temperatures for
the chiral transition (left) and deconfinement transition (right).
Both plots are for $\mu=0$. See main text for details.}
\label{tcdef}
\end{center}
\end{figure}
Following Ref.~\cite{bj}, we introduced an $N_f$ and $\mu_B$-dependent
transition temperature $T_0$ via Eq.~(\ref{nfmub}).
In Fig.~\ref{dcgamma}, we show the effects of varying the parameter
$\hat{\gamma}$ in Eq.~(\ref{bpar}) on the deconfinement transition
in the $\mu-T$ plane for zero magnetic field and utilizing the polynomial gluonic potential, Eq.~(\ref{poly}). The solid lines show
$T_{\Phi/2}$
while the dashed lines show $T_{{\bar\Phi}/2}$
for the same values of $\hat{\gamma}$. We note that both $\Phi$
and $\bar{\Phi}$ are real and coincide for $\mu=0$
but differ at non-zero $\mu$.
Furthermore, for a $\mu_B$-independent $T_0$
($=208$ MeV)
the transition temperature
is almost independent of the baryon chemical potential $\mu$
(magenta lines).
The red, green, and blue lines show the results for $\hat{\gamma}=0.8$,
$0.9$, and $1.0$, respectively. The bending of the curves decreases
as a function of $\hat{\gamma}$ which is reasonable since this parameter enters
in the denominator the parametrization (\ref{bpar}) of $b(N_f,\mu)$.
Finally, we remark that the qualitative behavior is the same for finite
magnetic field $B$. We will present more results for various $B$-fields
in the next section.
\begin{figure}[htb]
\begin{center}
\setlength{\unitlength}{1mm}
\includegraphics[width=12.0cm]{DeconGamma03.eps}
\caption{Phase diagram for the deconfinement transition with
$B=0$ and various values of the parameter
$\hat{\gamma}$. See main text for details.}
\label{dcgamma}
\end{center}
\end{figure}
In Fig.~\ref{dcop}, we show
the order parameter $\Phi(T)$ as a function of
$T$ for $\mu=0$ (blue), $\mu=210$ (green),
$\mu=260$ (red), and $\mu=290$ (magenta) with and without
a $\mu$ dependent gluonic potential.
In the left panel, the results are for $T_0=T_0(N_f,\mu)$,
while in the right panel $T_0=T_0(N_f,0)$ i.e.\ independence
from $\mu$.
Comparing the two panels we see the result shown in Fig.~\ref{dcgamma},
that only with a $\mu$ dependent transition temperature $T_0$ do we obtain
significant change in the deconfinement order parameter when varying $\mu$.
Additionally we see in the right panel that at high $\mu$ (magenta in
particular)
there is an initial increase in $\Phi$ around $T=50$ MeV, which comes from the
mesonic and fermionic
potential, $U_{k=0}$, and then around 208 MeV there is the typical increase,
driven
largely by the gluonic potential, $U_{\textrm{glue}}$. We then see in the left
panel, with
a $\mu$ dependent $T_0$, that the effect of $U_{\textrm{glue}}$ mirrors that of
$U_{k=0}$
and the deconfinement transition thus decreases with increasing $\mu$.
Figure~\ref{dcop} also illustrates the aforementioned difficulties
in defining the deconfinement transition at large $\mu$. It is seen
that $T_{\Phi/2} \sim T_\Phi$
at low $\mu$, but for $\mu \gtrsim 230$ MeV this is no longer true.
In addition to this,
the numerics become more time consuming at low $T$,
thus for values of $T \gtrsim 30$ MeV our results
only approximate the behavior of the model.
For these reasons we have only calculated the
phase diagram up to $\mu=290$ MeV.
\footnote{We
have also observed the splitting of the chiral transition reported
in~\cite{herbst13} without the Polyakov loop, but have not
resolved that region in detail with the Polyakov loop.}
\begin{figure}[htb]
\begin{center}
\setlength{\unitlength}{1mm}
\includegraphics[width=16.0cm]{DeconOrderParam03.eps}
\caption{Order parameter $\Phi$ as a function of
$T$ for various values of the chemical potential $\mu$ with (left) and
without (right) a $\mu$ dependent gluonic transition temperature, $T_0$.
$+\,$s are the data points, lines are the interpolations thereof,
$\square\,$s give $T_{\Phi/2}$ and $\circ\,$s approximate $T_{\Phi}$}
\label{dcop}
\end{center}
\end{figure}
In Fig.~\ref{gluepot}, we show the phase diagram for the deconfinement
transition with the three different glue potentials introduced in Sec.~\ref{sec:poly} at $B=0$. The blue lines are the
polynomial potential~(\ref{poly}), the red lines are the
logarithmic potential~(\ref{log}), and the green lines are the
Fukushima potential~(\ref{Fuku}).
The black line shows
the transition temperature $T_0=T_0(N_f,\mu,\hat{\gamma}=0.9)$
for pure glue for comparison.
We note that the black curve is almost the same as the
curve for the Fukushima potential (red), implying that the coupling to
the quarks has almost no influence on the deconfinement transition.
As was observed in~\cite{polyakovpot2} we find with the
logarithmic potential that $\Phi=\bar{\Phi}$ for all values of $\mu$,
we also find this to be true with the Fukushima potential. We also find
with the Fukushima potential, and to a lesser degree with the logarithmic
potential,
that the deconfinement transition temperature is dominated by the gluonic
potential.
This was also backed up by direct investigation of the $\Phi$ and ${\bar\Phi}$
as functions of $T$.
\begin{figure}[htb]
\begin{center}
\setlength{\unitlength}{1mm}
\includegraphics[width=12.0cm]{GluePotentials05.eps}
\caption{Phase diagram for the deconfinement transition for different
glue potentials and $B=0$.
Also shown is the transition temperature
$T_0=T_0(N_f,\mu,\hat{\gamma}=0.9)$ for pure glue for comparison.}
\label{gluepot}
\end{center}
\end{figure}
In Fig.~\ref{gluepot2}, we show the phase diagram for the chiral transition
using the different gluonic potentials. We also show the phase diagram
for the quark-meson model without the Polyakov loop, i.e. for
$\Phi=1$.
The lines show that the particular form of the gluonic potential
is
not as influential as we saw in the case of the deconfinement transition.
At zero $\mu$ and $B$, $T_\phi$ decreases by 2\% and 3\%
for the logarithmic and Fukushima potentials respectively. Only with
$\mu \gtrsim 260$ MeV do we see a significantly larger deviation than this.
\begin{figure}[htb]
\begin{center}
\setlength{\unitlength}{1mm}
\includegraphics[width=12.0cm]{ChiralDiffGlueGamma902.eps}
\caption{Phase diagram for the chiral transition for different
glue potentials and $B=0$.
Also shown is the transition temperature for $\Phi=1$, i.e.\ for
the quark-meson model without the Polyakov loop.}
\label{gluepot2}
\end{center}
\end{figure}
\newpage
\section{Results at finite magnetic field}
\label{sec:results}
In this section, we will present our main results and discuss them in some
detail. In Fig.~\ref{phase}, we show the phase diagram for the
chiral and the deconfinement transitions for $B=0$ (blue lines)
and for $|qB|=5.3m_{\pi}^2$. The results are obtained using the
polynomial glue potential~(\ref{poly}). We will discuss the
results in detail in connection with Fig.~\ref{big}, where we
show the chiral and deconfinement transition temperatures as a function of $B$
for different values of $\mu$.
\begin{figure}[htb]
\begin{center}
\setlength{\unitlength}{1mm}
\includegraphics[width=12.0cm]{PhaseDiag02.eps}
\caption{Phase diagram for
the deconfinement and chiral transitions for
$B=0$ and the largest magnetic
field, $|qB|=5.3$ $m_{\pi}^2$ with the Polynomial potential.}
\label{phase}
\end{center}
\end{figure}
In Fig.~\ref{big}, we show the
transition temperatures for the chiral and deconfinement transitions
as functions of $B$ for different values of $\mu$.
The solid blue lines indicate the chiral transition, $T_\phi$,
while the dashed green lines are $T_{\Phi/2}$
and the dashed red lines are $T_{{\bar\Phi}/2}$.
In the left upper panel, $\mu=0$ and $\Phi={1\over2}$ and $\bar{\Phi}={1\over2}$ coincide
for all B,
in agreement with our earlier remarks about the sign problem.
We note that the transition temperature for the chiral transition
is increasing for values of $\mu$ up to approximately $\mu=210$ MeV where
it is flat (lower middle panel). For larger chemical potentials,
the transition temperature for chiral transition is a decreasing function.
This shows the magnetic catalysis for small $\mu$ and inverse catalysis
for large $\mu$ which we discuss below.
For nonzero $\mu$ we see that the splitting between $\Phi$ and $\bar{\Phi}$
increases with $\mu$ and also with the strength of the magnetic field $B$.
For small values of $\mu$, $T_{\Phi/2}$ and
$T_{{\bar\Phi}/2}$ are almost independent of
$B$, while for large values, $T_{\Phi/2}$ increases with
increasing $B$ while $T_{{\bar\Phi}/2}$ decreases with $B$.
This behavior indicates that the relative importance
of the fermionic and mesonic fields also increases with larger $B$ and $\mu$
although we have not identified a mechanism behind this behavior.
\begin{figure}[htb]
\begin{center}
\setlength{\unitlength}{1mm}
\includegraphics[width=15.0cm]{SixPlots02.eps}
\caption{Transition temperatures for the chiral and deconfinement transitions
as functions of $B$ for different values of $\mu$. Solid blue lines denote $T_\phi$ while dashed lines correspond to the deconfinement transition with green giving $T_{\Phi/2}$, red giving $T_{{\bar\Phi}/2}$.
\label{big}
\end{center}
\end{figure}
\begin{figure}[htb]
\begin{center}
\setlength{\unitlength}{1mm}
\includegraphics[width=12.0cm]{ChiralWithWithoutL02.eps}
\caption{Phase diagram for the chiral transition for different values of
the magnetic field $B$ with (solid lines) and without (dashed lines)
the Polyakov loop using the polynomial potential.
Inset shows the critical temperature as a function of $B$
for $\mu=0$ with (solid lines) and without (dashed lines) the Polyakov loop.}
\label{cata}
\end{center}
\end{figure}
In Fig.~\ref{cata}, we show the phase diagram for the chiral phase
transition for different values of the magnetic field $B$ with
coupling to the Polyakov loop variable (solid lines) and without
(dashed lines). Inset shows the transition temperature as a function
of $B$ for vanishing $\mu$ in the two cases.
We first notice that the critical temperature
increases with the magnetic field for small values of the chemical
potential $\mu$.
The basic mechanism is that of magnetic catalysis~\cite{dcsm1,dcsm2,catarev},
namely that the chiral condensate increases as a function of
the magnetic field. It is interesting to note that the
increase of the transition temperature
as a function of $B$ is smaller when we couple the chiral sector
to the gluonic sector.
For large values of the chemical potential
$\mu$, the critical temperature is a decreasing function of
the magnetic field. This is inverse catalysis~\cite{inverse0,inverse2}.
We also find that the transition temperature is increased signficantly for
all values of $\mu$ with the addition of the Polyakov loop. Below $\mu\sim200$
MeV $T_\phi$ increases by approximately
25\% and above this density we find greater
increases in $T_\phi$. The Polyakov loop acts to suppress the finite temperature,
fermionic
contribution to the effective potential at all temperatures, although
particularly at
low temperatures. Thus we expect some increase in $T_\phi$ but its magnitude is
of
interest as it shows that the confining dynamics does play an important role in
the chiral transition within this model.
In this region we find $T_{\phi\textrm{, Fuku}} - T_{\phi\textrm{, log/poly}} \approx 20$
MeV.
The relative increase in magnetic field is more greatly affected by the choice
of potential,
with the relative increase in $T_\phi$ being approximately 20\% less with the
logarithmic and
Fukushima potentials as opposed to the polynomial potential shown in
Fig.~\ref{cata}.
Very recently, the existence of a new critical point associated with the
deconfinement transition of strongly interacting matter at finite $T$
and $B$, but vanishing $\mu$ has been suggested~\cite{cohen}.
The basic idea is that quarks effectively decouple in
the presence of very large magnetic
fields due to their increasing mass as a function of $B$.
In this case, one should be able to describe the system
with an effective theory of pure gluondynamics.
Although this effective theory is anisotropic, it is likely that it
has a first-order transition just like isotropic pure-glue QCD.
Since QCD with physical quark masses exhibit a crossover and not
a first-order transition, there ought to be a critical point
in the $T-B$ plane,
where the line of first-order transition ends.
However we find no evidence
within the range of magnetic fields we examine of a transformation from
the observed cross-over transition to a first order transition for the
deconfinement order parameter.
\section{Summary and outlook}
\label{sec:conclusion}
In this work we have used the functional renormalization group to calculate
the phase diagram with respect to the chiral and deconfinement transitions
for the Polyakov loop extended quark-meson model. We first investigated the
effects of the gluonic potential, showing that the
deconfinement transision is quantitavely dependent upon the exact
implementation,
and in some cases even qualitatively dependent. Most noticibly
$T_{\Phi/2} - T_{{\bar\Phi}/2}$
is only non-zero when using the polynomial potential~(\ref{poly}). This
potential was also
the least dominating in that the fermionic and mesonic degrees of freedom had a
much
larger effect upon the deconfinement order parameters, $\Phi$ and ${\bar\Phi}$.
However for
all three potentials the gluonic potential dominated the dynamics. At
high $\mu$ we see a double humped structure in these order parameters.
This made the evaluation of $T_{\Phi/2}$ and $T_{{\bar\Phi}/2}$ difficult and
we can not find a first order transition around $\mu\sim300$ MeV (given by
Herbst et al.~\cite{herbst}) although we saw indications of this.
We find magnetic catalysis at
low $\mu$ in agreement with other model calculations,
however we see a weakening
of its effects with the addition of the Polyakov loop. At large $\mu$ the
inverse magnetic catalysis found in the quark-meson model~\cite{anders}
is also found here.
When using the polynomial potential we a find splitting of
$T_{\Phi/2}$ and $T_{{\bar\Phi}/2}$ at non-zero $\mu$. This splitting increases with
increasing
magnetic field strength and quark chemical potential (other than for the very
highest $\mu$ value).
In addition
$T_\phi$ increases significantly
for all values of $\mu$ shows that the Polyakov loop plays an important role
in the chiral transition. In contrast to the confinement transition,
we found that the chiral transition is not sensitive to the choice of
the gluon potential.
In the recent papers~\cite{budaleik,inverse}, the authors
suggest a resolution
of the discrepancy between the
model calculations and the lattice simulations.
The chiral condensate can be written as
\begin{eqnarray}
\langle\bar{\psi}\psi\rangle
&=&{1\over{\cal Z}(B)}\int d{\cal U}
e^{-S_g}\det(D\!\!\!\!/(B)+m){\rm Tr}
(D\!\!\!\!/(B)+m)^{-1}\;,
\label{det}
\end{eqnarray}
where the partition function is
\begin{eqnarray}
{\cal Z}(B)&=&\int d{\cal U}
e^{-S_g}\det(D\!\!\!\!/(B)+m)\;,
\end{eqnarray}
and $S_g$ is the pure-glue action. Thus there are two contributions to the
chiral condensate, namely the operator itself
(called valence contribution)
and the change of typical gauge configurations sampled, coming
from the determinant in Eq.~(\ref{det})
(called sea contribution). At least for
small magnetic fields one can disentangle these contributions by
defining
\begin{eqnarray}
\langle\bar{\psi}\psi\rangle^{\rm val}
&=&{1\over{\cal Z}(0)}\int d{\cal U}
e^{-S_g}\det(D\!\!\!\!/(0)+m){\rm Tr}
(D\!\!\!\!/(B)+m)^{-1}\;,
\\
\langle\bar{\psi}\psi\rangle^{\rm sea}
&=&{1\over{\cal Z}(B)}\int d{\cal U}
e^{-S_g}\det(D\!\!\!\!/(B)+m){\rm Tr}
(D\!\!\!\!/(0)+m)^{-1}\;.
\end{eqnarray}
At zero temperature, both contributions are positive leading
to magnetic catalysis. At temperatures around the transition temperature,
the valence condensate is still positive while the sea condensate
is negative. Hence there is a competition between the two leading to a
net inverse catalysis. The sea contribution can be viewed as a back reaction
of the fermions on the gauge fields and this effect is not present in the
model calculations as there are no dynamical gauge fields.
If such a back reaction can be incorporated in the model calculations,
one may be able to obtain agreement with the lattice simulations.
One way of doing this is by using a $B$-dependent
parametrization
of the
transition temperature~\cite{ferry}
in analogy with the flavor and $\mu_B$ dependence that we see
here to be critical to realistic mapping of the phase diagram.
We will report on this in a future publication~\cite{future}.
\section*{Acknowledgments}
The authors would like to thank Jonas R. Glesaaen for valuable
discussions.
|
\section{Introduction and results
In 1923, C. Loewner derived a differential equation for a certain family of conformal mappings to
attack the Bieberbach conjecture,
see \cite{Loewner:1923}. Loewner's method has been extended and turned out to be a useful tool
within complex analysis. In particular, the Loewner differential equations provide a powerful
tool for the description of the growth of slits in a given planar domain.
After O. Schramm discovered Stochastic Loewner Evolution (or Schramm Loewner Evolution, SLE)
in \cite{MR1776084}, it became clear that those models have many applications in different
mathematical and physical disciplines, especially in statistical physics.\\
In the classical setting, Loewner theory describes the evolution of a family of simply connected,
proper subsets of the complex plane $\mathbb{C},$ which are all conformally equivalent to the
unit disc $\D:=\{z\in\mathbb{C} \;|\; |z|<1\}$ according to the Riemann Mapping Theorem.
In 1943, Y. Komatu showed that Loewner's ideas are not confined to the simply connected case:
He derived a Loewner equation for the growth of a slit within a doubly connected domain,
see \cite{KomatuZweifach}. In \cite{Komatu}, he considered a generalization of Loewner's
differential equation to a more general finitely connected domain. Komatu's ideas have been applied and
extended by several authors. Recently, R. Bauer and R. Friedrich derived a radial and a
chordal Komatu-Loewner equation to deal with the growth of a (stochastic) slit in a multiply
connected domain, see \cite{BauerFriedrichCSD} and \cite{BauerFriedrichCBC}. In the radial
setting, the slit grows within a \textit{circular slit disk} $D$, i.e.
$D= \D \setminus (C_1\cup...\cup C_N)$, $N\in \N_0,$ where each
$C_j\subset\D$ is a circular arc centered at $0$ such that
$C_j\cap C_k=\emptyset$ whenever $j\not=k.$ Note that every $N$-connected domain $\Omega$ can
be mapped onto such a circular slit disk $D$ by a conformal map $f: \Omega \rightarrow D.$
This mapping is unique if we require the normalization $f(z_0)=0,$ $f'(z_0)>0$ for some
$z_0\in \Omega,$ see \cite{ConwayII}, Chapter 15.6.\\
In \cite{BoehmLauf}, W. Lauf and the first author generalized the radial Komatu-Loewner
equation for the growth of several slits:\\
Let $\Omega$ be an arbitrary circular slit disk and let
$\gamma_1,...,\gamma_m:[0,T]\to\overline{\Omega}$ be parametrizations of pairwise disjoint simple
curves such that $\gamma_k(0)\in\partial \D$ and $\gamma_k(0,T]\subset
\Omega\setminus\{0\}$ for all $k=1,...,m.$ \\
Furthermore, if $D$ is a circular slit disk and $u\in\partial \D$, we denote by
$w\mapsto \Phi(u,w;D)$ the unique conformal mapping from $D$ onto the right half-plane minus
slits parallel to the imaginary axis with $\Phi(u,u;D)=\infty$ and $\Phi(u,0;D)=1.$\\
We summarize one of the main results of \cite{BoehmLauf} in the following theorem:
\begin{satz}[Corollary 5 in \cite{BoehmLauf}] \label{The:KomLowEqu}
Let $\Omega_t=\Omega \setminus \bigcup_{k=1}^m \gamma_j[0,t]$ and denote by $g_t$ the unique
conformal mapping $g_t: \Omega_t \to D_t$ where $D_t$ is a circular slit disk
and $g_t(0)=0,$ $g'_t(0)>0.$\\
Then there exists a Lebesgue measure zero set $\mathcal{N}$ such that for every
$z\in \Omega_T$ the function $t\mapsto g_t(z)$ is differentiable on
$[0,T]\setminus \mathcal{N}$ with
\begin{equation}\label{KLE}
\dot{g}_t(z) = g_t(z) \sum_{k=1}^m \lambda_k(t)\cdot \Phi(\xi_k(t),g_t(z);D_t),
\end{equation}
where the continuous function $t\mapsto\xi_k(t)\in \partial\D$ is the image of
$\gamma_k(t)$ under the map $g_t$ and the coefficient functions $t\mapsto \lambda_k(t)$ are measurable with $\lambda_k(t)\geq 0$ for every
$t\in[0,T]$.
\end{satz}
\begin{remark} The continuous functions
$\xi_k:[0,T]\to\partial\D$ are usually called \emph{driving functions}. \\
Informally, the coefficient function $\lambda_k(t)$ corresponds to the speed
of growth of the slit parametrized by $\gamma_k.$ From the normalization $\Phi(u,0;D)=1$ it follows that $g'_t(0)=e^{\int_0^t \sum_{k=1}^m \lambda_k(\tau)\; d\tau}.$ \\
\end{remark}
\begin{figure}[h]
\centering \includegraphics[width=135mm]{Input/Pic.pdf}
\caption{The mappings $z\mapsto g_t(z)$ and $w\mapsto \Phi(u,w,D_t)$ from Theorem \ref{The:KomLowEqu}.}
\end{figure}
Note that there are no further assumptions on the parametrizations of the slits
$\Gamma_k:=\gamma_k[0,T]$ in Theorem \ref{The:KomLowEqu}. \\
Now suppose we are given only the circular slit disk $\Omega$ and the slits
$\Gamma_1,...,\Gamma_m$ without parametrization. Roughly speaking, we address the question if it is possible to find a simple form of equation (\ref{KLE}) such that it has still enough parameters to generate the slits $\Gamma_1,...,\Gamma_m$, but, on the
other hand, the choice of those parameters is unique. We will see that this is possible and,
moreover, it will turn out that equation (\ref{KLE}) is satisfied for \textit{all} $t$
in this case.
The latter can be interpreted as a generalization of Loewner's original idea of finding a parametrization of an arbitrary curve such that the family of certain associated conformal mappings is differentiable. In some sense, this problem is related to Hilbert's fifth problem of finding differentiable structures for continuous groups, see \cite{Goodman}.
\medskip
In a first step, we can choose parametrizations such that the mappings $g_t$ satisfy $g_t'(0)=e^t,$ i.e. $\sum_{k=1}^m \lambda_k(t)\equiv 1.$ However, there are many of such parametrizations when $m\geq 2$.
In this work we will show that there exist unique parametrizations of the slits
$\Gamma_1,...,\Gamma_m$, such that the corresponding Komatu-Loewner equation is satisfied
for all $t$ where all coefficient functions $\lambda_k(t)$ are constant and sum up to 1.
\medskip
We need one further notation:
Let $\Omega$ be a circular slit disc and let $f:\Omega\setminus(\Gamma_1\cup\ldots\cup\Gamma_m)
\to D$ be the unique conformal mapping onto a
circular slit disk $D$ with $f(0)=0$ and $f'(0)>0$, then $f'(0)> 1$ and the
\emph{logarithmic mapping radius} of $\Omega\setminus(\Gamma_1\cup\ldots\cup\Gamma_m)$ is
defined to be the real number $\log f'(0)> 0$. The inequality $f'(0)>1$ is
an immediate consequence of Lemma \ref{Lem:PropLMR} b).
\begin{theorem} \label{The:ConCoeffi}
Let $L$ be the logarithmic mapping radius of $\Omega\setminus(\Gamma_1\cup\ldots\cup\Gamma_m)$.
There exist unique continuous parametrizations
$$
\delta_1:[0,L]\to\Gamma_1,..., \delta_m:[0,L]\to\Gamma_m
$$
and unique $\lambda_1,\ldots,\lambda_m\in(0,1)$ with $\sum_{k=1}^m\lambda_k=1$ such that
the following holds: \\
Let $\Omega_t=\Omega \setminus \bigcup_{k=1}^m \delta_j[0,t]$ and denote by $h_t$ the unique
conformal mapping $h_t: \Omega_t \to D_t$ where $D_t$ is a circular slit disk and
$h_t(0)=0,$ $h'_t(0)>0.$\\
Then $h_t'(0)=e^t$ and for every $z\in \Omega_L,$ the function $t\mapsto h_t(z)$ is
differentiable on $[0,L]$ with
\begin{equation}\label{KLE2}
\dot{h}_t(z) = h_t(z)\sum_{k=1}^m \lambda_k \cdot \Phi(\xi_k(t),h_t(z);D_t),
\end{equation}
where $\xi_k(t)\in \partial\D$ is the image of $\delta_k(t)$ under the map $h_t$.
Moreover, the driving functions $t\mapsto \xi_k(t)$ are continuous.
\end{theorem}
In the simply connected case, we have $D_t=\D$ for all $t$ and equation (\ref{KLE2}) can be
written down explicitly as
$$
\Phi(u,w,\D)=\frac{u+w}{u-w}.
$$
Furthermore, $m$ coefficients $\lambda_1,...,\lambda_m\in(0,1)$ and $m$ continuous driving
functions $\xi_1,...,\xi_m:[0,L]\to\partial\D$ determine the unique solution to
equation (\ref{KLE2}) in the simply connected case.
If the solution generates slit mappings, then the parametrization of these slits are
uniquely determined by the driving functions and the coefficients.\\
Thus we can formulate the simply connected case of Theorem \ref{The:ConCoeffi} as follows.
\begin{corollary}
Let $\Gamma_1,\ldots,\Gamma_m$ be disjoint slits in $\D$ and let $L$ be the logarithmic
mapping radius of $\D\setminus(\Gamma_1\cup\ldots\cup\Gamma_m)$.
Then there exist unique $\lambda_1,...,\lambda_m\in(0,1)$ and unique continuous
driving functions $\xi_1,\ldots,\xi_m:[0,L]\to\partial\D$ such that the solution of the
Loewner equation
\begin{equation*}
\dot{h}_t(z) = h_t(z)\sum_{k=1}^m \lambda_k \cdot \frac{\xi_k(t)+h_t(z)}{\xi_k(t)-h_t(z)},
\qquad h_0(z)=z,
\end{equation*}
generates the slits $\Gamma_1,\ldots,\Gamma_m,$ i.e. $h_L$ maps
$\D\setminus(\Gamma_1\cup\ldots\cup\Gamma_m)$ conformally onto $\D$.
\end{corollary}
\begin{remark}
In \cite{Prokhorov:1993}, D. Prokhorov has proven the existence and uniqueness of constant
coefficients for several slits in the simply connected case under the assumption that all
slits are \emph{piecewise analytic}. This theorem forms the basis for Prokhorov's study of
extremal problems for univalent functions in \cite{Prokhorov:1993} by using control-theoretic
methods.
Our proof shows that one can drop any assumption on the regularity of the slits in order
to generate them with constant coefficients.
\end{remark}
We shall give the details of the proof only in the case $m=2$, i.e.~for two slits,
in order to allow a simple notation. The general case of $m \ge 2$ slits can be proved
inductively in exactly the same way.\\
Our proof combines some technical tools from \cite{BoehmLauf} that were used to prove
Theorem \ref{The:KomLowEqu} and an idea from \cite{RothSchl}, where a similar result was
proven for the so called chordal Loewner equation in the simply connected case.\\
The rest of this paper is organized as follows.
In Section 2 we describe the setting for the proof and cite some technical results
from \cite{BoehmLauf}.
The proof of Theorem \ref{The:ConCoeffi} is divided into two parts: In Section 3 we prove
the existence statement and in Section 4 we give the proof of the uniqueness statement
of Theorem \ref{The:ConCoeffi}.
\section{The setting for the proof}
Suppose $\Omega$ is a circular slit disk and $\Gamma_1$, $\Gamma_2$ are disjoint slits,
i.e. there are continuous, one-to-one functions $\gamma_1, \gamma_2:[0,T]\to\overline{\Omega}$
with $\Gamma_k=\gamma_k[0,T]$ such that $\Gamma_1\cap\Gamma_2=\emptyset$,
$\gamma_k(0)\in\partial \D$ and $\gamma_k(0,T]\subset \Omega\setminus\{0\}$ for $k= 1,2.$ \\
Let the function $g_t$ be defined as in Theorem \ref{The:KomLowEqu}.
As $g'_t(0)$ is monotonically increasing (see Lemma \ref{Lem:PropLMR}) and $g'_0(0)=1$,
we can assume without loss of generality that $g'_t(0)=e^t.$
Otherwise, we can simultaneously reparameterize the two slits.
Consequently we have $T=L$, where $L$ denotes the logarithmic mapping radius of
$\Omega\setminus(\Gamma_1\cup\Gamma_2)$.\\
Furthermore, we assume that $L=1$ in order to simplify some of the notations.\\
For every $t,\tau\in[0,1]$ we let $f_{t,\tau}$ be the unique conformal mapping from
$\Omega\setminus(\gamma_1[0,t]\cup \gamma_2[0,\tau])$ onto a circular slit disk with
$f_{t,\tau}(0)=0$ and $f_{t,\tau}'(0)>0,$ and we define
$$
\lmr(t,\tau):=\log(f'_{t,\tau}(0)).
$$
Now, in order to prove Theorem \ref{The:ConCoeffi}, we have to show that there exist
\begin{itemize}
\item two uniquely determined increasing homeomorphisms $u,v:[0,1]\to[0,1]$ and
\item a uniquely determined $\lambda_0\in(0,1),$
\end{itemize}
such that the Komatu-Loewner equation for the slits $\delta_1:=\gamma_1\circ u$ and
$\delta_2:=\gamma_2\circ v$ is satisfied for all $t\in[0,1]$ with $\lambda_1(t)= \lambda_0$
and $\lambda_2(t)=1-\lambda_0$ and $\lmr(u(t),v(t))=t$ for all $t\in[0,1].$\\
First, we summarize some basic properties of $\lmr$ in the following lemma.
\begin{lemma}${}$\label{Lem:PropLMR}
\begin{enumerate}
\item[(a)] \label{Pro:ComConf} The function $\lmr(t,\tau)$ is continuous in $[0,1]^2$.
\item[(b)] \label{Pro:Monfunf} The function $lmr(t,\tau)$ is strictly increasing with respect to $t$ and $\tau$ respectively.
\item[(c)] \label{Pro:DifQuoAbs2} For every $\epsilon>0$ there exists a $\delta>0$ so that for all
$0\leq \lt < \gt \leq 1$ and $0\leq \ltau <\gtau \leq 1$ with $|\lt-\gt|<\delta$ and $|\ltau-\gtau|<\delta$
the following holds:
\[
1-\epsilon <\frac{\lmr(\lt,\ltau)-\lmr(\gt,\ltau)}{\lmr(\lt,\gtau)-
\lmr(\gt,\gtau)} < 1+\epsilon.
\]
\end{enumerate}
\end{lemma}
\begin{proof}
See Proposition 6, 8 and 15 in \cite{BoehmLauf}.
\end{proof}
Furthermore, we will use a dynamic interpretation of the coefficient functions from
Theorem \ref{The:KomLowEqu}. Let $a,b:[0,1]\to[0,1]$ be arbitrary strictly increasing
homeomorphisms and let $Z=\{t_0,\ldots,t_s\}$ be a partition of the interval $[0,t]$
for a fixed $t\in(0,1]$, i.e. $0=t_0<t_1<...<t_s=t.$ We will denote by
$|Z|:= \max_{j\in\{0,\ldots,s-1\}}(t_{j+1}-t_j)$ the norm of $Z.$ Now we define the two sums
\begin{align*}
S_1(a,b,t,Z)&:= \sum_{l=0}^{s-1} \big[\lmr(a(t_{l+1}),b(t_{l})) - \lmr(a(t_{l}),b(t_l))\big],\\
S_2(a,b,t,Z)&:= \sum_{l=0}^{s-1} \big[\lmr(a(t_l),b(t_{l+1})) - \lmr(a(t_{l}),b(t_l))\big].
\end{align*}
The following proposition relates the limit of $S_j$ for $|Z|\to0$ to the coefficient
functions $\lambda_j$ of Theorem \ref{The:KomLowEqu}.
\begin{proposition} \label{Pro:ProCalLam}
Let $a,b:[0,1]\to[0,1]$ be two increasing self-homeomorphisms and let $g_t(z)$
denote the conformal mapping for the slits $\gamma_1\circ a$ and
$\gamma_2\circ b$ from Theorem \ref{The:KomLowEqu}. Assume that
$g'_t(0)=e^t$, i.e. $\lmr(a(t),b(t))=t,$ for all $t\in[0,1]$. Then the limits
\[
c_1(t):=\lim_{|Z|\rightarrow 0} S_1(a,b,Z,t)\quad \text{and}
\quad c_2(t):=\lim_{|Z|\rightarrow 0} S_2(a,b,Z,t)
\]
exist and form two increasing and Lipschitz continuous functions
$c_1,c_2:[0,1]\rightarrow[0,\infty)$ with $c_1(0)=c_2(0)=0$ and $c_1(t)+c_2(t)=t$.
Furthermore, if $c_j$ is differentiable in $t_0$ for $j=1$ and $j=2,$ then
the differential equation (\ref{KLE}) holds for $g_{t_0}$ with
$$
\lambda_j(t_0) = \dot c_j(t_0).
$$
In this case, $\lambda_1(t_0)$ is equal to the first derivative of the
function $t\mapsto\lmr(a(t),b(t_0))$ in $t_0$ and
$\lambda_2(t_0)$ is equal to the first derivative of the function
$t\mapsto\lmr(a(t_0),b(t))$ in $t_0$.
\end{proposition}
\begin{proof}
This follows immediately from Proposition 16 in \cite{BoehmLauf}, Theorem 2
in \cite{BoehmLauf} and the definition of $\lambda_j$ in \cite{BoehmLauf}.
\end{proof}
Beside $S_1$ and $S_2$ we define for a partition $Z=\{t_0,\ldots,t_s\}$ of the interval $[0,t]$
\begin{align*}
\tilde S_1(a,b,t,Z)&:= \sum_{l=0}^{s-1} \big[\lmr(a(t_{l+1}),b(t_{l+1})) -
\lmr(a(t_{l}),b(t_{l+1}))\big],\\
\tilde S_2(a,b,t,Z)&:= \sum_{l=0}^{s-1} \big[\lmr(a(t_{l+1}),b(t_{l+1})) -
\lmr(a(t_{l+1}),b(t_l))\big].
\end{align*}
If $g'_t(0)=e^t$ holds for every $t\in[0,1]$, then it is easy to see that $S_1(a,b,t,Z)+
\tilde S_2(a,b,t,Z)=t$ and $S_2(a,b,t,Z) + \tilde S_1(a,b,t,Z)=t$.
By Proposition \ref{Pro:ProCalLam} we see
\[
\lim_{|Z|\rightarrow 0} \tilde S_1(a,b,Z,t) = t-c_2(t)=c_1(t),\quad
\lim_{|Z|\rightarrow 0} \tilde S_2(a,b,Z,t) = t-c_1(t)=c_2(t).
\]
\section{Existence}
Now we are able to prove the existence part of Theorem \ref{The:ConCoeffi}.
Recall that we have to show the existence of two strictly increasing homeomorphisms
$u,v:[0,1]\to[0,1]$ and a $\lambda_0\in(0,1),$
such that the Komatu-Loewner equation for the slits $\gamma_1\circ u$ and
$\gamma_2\circ v$ is satisfied for all $t\in[0,1]$ with $\lambda_1(t)= \lambda_0$
and $\lambda_2(t)=1-\lambda_0$ and $\lmr(u(t),v(t))=t$ for all $t\in[0,1].$
The proceeding of this proof is as follows.
\begin{enumerate}
\item First of all we will use a Bang-Bang method introduced in \cite{RothSchl} to construct
two sequences $(u_n)_{n\in\IN}$ and $(v_n)_{n\in\IN}$ of increasing self-homeomorphisms
of $[0,1]$.
\item By using a diagonal argument on $u_n$ and $v_n$ we will find two subsequences
$(u_n^*)_{n\in\IN}$ and $(v_n^*)_{n\in\IN}$ which converge pointwise on a dense set
$S\subset [0,1]$ to increasing functions $u$ and $v$ respectively.
The functions $u$ and $v$ can be extended to continuous functions
defined on $[0,1]$, with $u(1)=1=v(1)$.
Furthermore, we will get $\lambda_0\in [0,1]$ by the construction of $u$ and $v$.
\item Next we will derive a connection between the sum $S_1(u_n^*,v_n^*,t,Z)$
and the sum $S_1(u,v,t,Z)$ for a given partition $Z$ of the interval $[0,t]$.
\item Moreover, we will find a connection between $S_1(u_n^*,v_n^*,t,Z)$ and $\lambda_0$.
\item By combining these results we will find $S_1(u,v,t,Z)\rightarrow \lambda_0 t$ if
$|Z|\rightarrow 0$. Furthermore, as a consequence of this, we will find $\lambda_0\in(0,1)$.
\item Next will show that $u$ and $v$ are strictly increasing, i.e. both functions are
increasing self-homeomorphisms of $[0,1].$
\item Finally we will obtain the Komatu-Loewner-Equation with constant coefficients
$\lambda_0$ and $1-\lambda_0$ for the parametrizations $u$ and $v$.
\end{enumerate}
\begin{proof}[Proof of Theorem \ref{The:ConCoeffi} (Existence)]
${}$
\begin{selflist}
\item To construct $u_n$ and $v_n$, we first extend both $\gamma_1$ and $\gamma_2$ to an interval
$[0,T^*],$ $T^*>1,$ such that $\gamma_1[0,T^*]$ and $\gamma_2[0,T^*]$ are still
disjoint slits and $\lmr(T^*,0)\geq1,$ $\lmr(0,T^*)\geq 1.$
Let $n\in \IN$ and $\lambda\in[0,1]$.
We let $t_{0,n}=\tau_{0,n}=0$ and for $k\in\{1,\ldots, n\}$ we define $t_{k,n}>0$
and $\tau_{k,n}>0$ recursively as the unique values with
$$
\lmr(t_{k,n}, \tau_{k-1,n})-\lmr(t_{k-1,n},\tau_{k-1,n})=
\frac{\lambda}{n}, \quad \lmr(t_{k,n}, \tau_{k,n})-\lmr(t_{k,n}, \tau_{k-1,n})=
\frac{1-\lambda}{n}.
$$
Since $(t,\tau)\mapsto\lmr(t,\tau)$ is strictly increasing in both variables,
see Lemma \ref{Lem:PropLMR} b), we get
\begin{align*}
\lmr(t_{n,n},\tau_{n,n})&=1 \le \lmr(T^*,0) < \lmr(T^*,\tau_{n,n})\\
\lmr(t_{n,n},\tau_{n,n})&=1 \le \lmr(0,T^*) < \lmr(t_{n,n},T^*).
\end{align*}
Consequently $t_{n,n},\tau_{n,n}\le T^*$.\\
Furthermore, note that the values $t_{k,n}=t_{k,n}(\lambda)$ and
$\tau_{k,n}=\tau_{k,n}(\lambda)$ depend continuously on $\lambda$:
This follows easily by induction and the continuity
and strict monotonicity of the function
$(t,\tau)\mapsto \lmr(t,\tau)$, see Lemma \ref{Pro:Monfunf}.
Consequently, for every $n\in\IN$, we can find a value $\lambda_n\in(0,1)$ with
$t_{n,n}(\lambda_n)=1$.
Now we define a sequence of functions $u_n:[0,1]\rightarrow [0,t_{n,n}]$ and
$v_n:[0,1]\rightarrow [0,\tau_{n,n}]$. Define
\[
u_n\Big(\frac{k}{2^n}\Big) := t_{k,2^n}(\lambda_{2^n}), \quad
v_n\Big(\frac{k}{2^n}\Big) := \tau_{k,2^n}(\lambda_{2^n})
\]
for all $k=0,\ldots,2^n$. The values of $u_n$ and $v_n$ between the supporting
points are defined by linear interpolation. An immediate consequence of this
construction is
\begin{align} \label{Equ:1}
\lmr\bigg(u_n\Big(\frac{k}{2^n}\Big),v_n\Big(\frac{k}{2^n}\Big) \bigg)
=\lmr\big( t_{k,2^n}(\lambda_{2^n}),\tau_{k,2^n}(\lambda_{2^n}) \big)=\frac{k}{2^n}.
\end{align}
\item
Since $\lambda _{2^n}$ is bounded, we find a subsequence
$(m_{k,0})_{k\in\N}$
such that $(\lambda_{2^{m_{k,0}}})_{k\in\N}$ is convergent with
the limit $\lambda_0\in[0,1]$.
Next we set
\[
S:=\bigcup_{n=1}^\infty S_n, \quad
S_n:=\Big\{ \frac{k}{2^n}\, \big|\, k=0,\ldots, 2^n\Big\}.
\]
$S$ is a dense and countable subset of $[0,1]$.
Denote by $a:\IN\rightarrow S$ a bijective mapping.\\
Since the sequences $(u_{m_{k,0}}(a_1))_{k\in\IN}$ and $(v_{m_{k,0}}(a_1))_{k\in\IN}$
are bounded (by $T^*$), we find a subsequence $(m_{k,1})_{k\in\IN}$
of $(m_{k,0})_{k\in\IN}$, so that $(u_{m_{k,1}}(a_1))_{k\in\IN}$ and
$(v_{m_{k,1}}(a_1))_{k\in\IN}$ are convergent.
Inductively, we define $(m_{k,l})_{k\in\N}$, $l\in\N$, to be a subsequence of
$(m_{k,l-1})_{k\in\N}$ such that $(u_{m_{k,l}}(a_l))$ and $(v_{n_{k,l}}(a_l))$
are convergent.
Consequently we can define sequences
$u_n^*:= u_{m_{n,n}}$ and $v_n^*:= v_{m_{n,n}}$ which are (pointwise)
convergent in $S$.
We denote by $u$ and $v$ the limit function, i.e.
\[
u(t):= \lim_{n\rightarrow \infty} u^*_n(t),\quad
v(t):= \lim_{n\rightarrow \infty} v^*_n(t)
\]
for all $t\in S$. Moreover we set $\lambda_n^*:=\lambda_{2^{m_{n,n}}}$
and $S^*_n:=S_{m_{n,n}}$.
By using equation (\ref{Equ:1}) we get
$\lmr\big(u_n^*(t),v_n^*(t)\big)=t$ for $t\in S$ if $n$ is big
enough. Consequently we find by using Lemma \ref{Pro:ComConf} a)
\begin{align} \label{Equ:2}
\lmr\big(u(t),v(t)\big)=
\lim_{n\rightarrow\infty} \lmr\big(u_n^*(t),v_n^*(t)\big) = t
\end{align}
for all $t\in S$.
Furthermore, since $t\mapsto u_n^*(t)$ and $t\mapsto v_n^*(t)$ are strictly
increasing, the functions $t\mapsto u(t)$ and $t\mapsto v(t)$ are increasing too.
Moreover $u$ and $v$ can be extended in a continuous and unique way to [0,1].
To see this, let $t_0\in(0,1)$ and define
\[
t_1:=\lim_{t\nearrow t_0 \atop t\in S} u(t),\quad
t_2:=\lim_{t\searrow t_0 \atop t\in S} u(t),\quad
\tau_1:=\lim_{t\nearrow t_0 \atop t\in S} v(t),\quad
\tau_2:=\lim_{t\searrow t_0 \atop t\in S} v(t).
\]
Thus we find by Lemma \ref{Pro:ComConf} a) and equation (\ref{Equ:2})
\[
\lmr(t_1,\tau_1)=\lim_{t\nearrow t_0\atop t\in S} \lmr(u(t),v(t)) = t_0
=\lim_{t\searrow t_0\atop t\in S} \lmr(u(t),v(t)) = \lmr(t_2,\tau_2).
\]
Since $(t,\tau)\mapsto\lmr(t,\tau)$ is strictly increasing in both variables
and $t_1\le t_2$ and $\tau_1\le\tau_2$, we find $t_1=t_2$ and $\tau_1=\tau_2$.
If $t_0\in\{0,1\}$ we can argue in the same way, so $t\mapsto u(t)$ and
$t\mapsto v(t)$ are continuous in $[0,1]$.
Summarizing, $u$ and $v$ are continuous and increasing in [0,1] with $u(1)=1$ and
$v(1)=1$.
For later use we define $h^{[n]}_{t,\tau}:=f_{u^*_n(t),v^*_n(\tau)}$ and
$h_{t,\tau}:=f_{u(t),v(\tau)}$.
\item Next we show that for every fixed $\epsilon >0$, fixed $t\in S$ and a fixed
partition $Z\subset S$ of the interval $[0,t]$, there exists an $n_0\in\IN$ so that
\[
|S_1(u^*_n,v^*_n,t,Z) - S_1(u,v,t,Z)| < \epsilon
\]
holds for all $n\ge n_0$, where $Z=\{t_0,t_1,\ldots, t_s\}$.
Fix $\epsilon>0$.
As the function $(t,\tau)\mapsto \lmr(t,\tau)$ is (uniformly) continuous in
$[0,T^*]^2$ by Lemma \ref{Pro:ComConf} a), there exists $\delta >0$ such that
\begin{align*}
|\lmr(\lt,\ltau)-\lmr(\gt,\gtau)|<\frac{\epsilon}{2s}
\quad\text{whenever} \quad |\lt-\gt|,\, |\ltau-\gtau|<\delta.
\end{align*}
Since $Z\subset S$, we find an $n_0\in\IN$ so that $|u_n^*(t_l)-u(t_l)|$,
$|v_n^*(t_l)-v(t_l)|<\delta$ holds for all $l=0,\ldots, s$ and $n\ge n_0$.
Consequently we find
\begin{align*}
|&S_1(u^*_n,v^*_n,t,Z) - S_1(u,v,t,Z)|\\
&= \Big|\sum_{l=0}^{s-1} \lmr(h^{[n]}_{t_{l+1},t_l})- \lmr(h^{[n]}_{t_l,t_l})
+ \sum_{l=0}^{s-1} \lmr(h_{t_{l+1},t_l})-\lmr(h_{t_l,t_l})\Big|\\
&\le \sum_{l=0}^{s-1} \big|\lmr(h^{[n]}_{t_{l+1},t_l})-\lmr(h_{t_{l+1},t_l})\big|
+ \sum_{l=0}^{s-1} \big|\lmr(h^{[n]}_{t_l,t_l})-\lmr(h_{t_l,t_l})\big|
\le 2 s \frac{\epsilon}{2s} = \epsilon.
\end{align*}
\item For now we fix $t\in S$. We show that for all $\epsilon >0$ we find
a $\mu>0$ so that for all partitions $Z\subset S$ of $[0,t]$ with $|Z|<\mu$
there exists an $m_0 \in \IN$ so that for all $n\ge m_0$ we have
\[
|S_1(u^*_n,v^*_n,t,Z)-\lambda_0 t|<\epsilon.
\]
Let $\epsilon>0$. Then there exists $\delta>0$ such that the inequality from
Lemma \ref{Pro:DifQuoAbs2} c) holds. Since the functions $t\mapsto u(t)$
and $t\mapsto v(t)$ are (uniformly) continuous we get
\begin{align*}
\exists \mu>0:\; |\lt-\gt|<\mu \Rightarrow
|u(\lt)-u(\gt)|, |v(\lt)-v(\gt)|<\frac{\delta}{2}.
\end{align*}
Denote by $Z=\{t_0,\ldots,t_s\}$ a partition of $[0,t]$ with $|Z|<\mu$ and
$Z\subset S$.
Then we find an $m_0\in\IN$ with $Z\subset S^*_n$, $t\in S^*_n$ and
\[
|u_n^*(t_l)-u(t_l)|,\, |v_n^*(t_l)-v(t_l)|< \frac{\delta}{4}
\]
for all $n\ge m_0$ and all $l=0,\ldots, s$.
As a consequence we get
\begin{multline*}
|u^*_n(t_{l+1})-u^*_n(t_l)|\\
\le |u^*_n(t_{l+1})-u(t_{l+1})| + |u(t_{l+1})-u(t_l)| + |u(t_l)-u^*_n(t_l)|<
\frac{\delta}{4} + \frac{\delta}{2} + \frac{\delta}{4} = \delta
\end{multline*}
for all $n\ge m_0$ and all $l=0,\ldots, s$. In an analog way we get
$|v^*_n(t_{l+1})-v^*_n(t_l)|<\delta$.
Next, we set $S^*_n(t):=S^*_n\cap[0,t]$.
$S^*_n(t)$ is a partition of the interval
$[0,t]$ and we write $S^*_n(t)=\{t^*_0,\ldots, t^*_{s^*}\}$.
\begin{align*}
|\lambda^*_n t-&S_1(u^*_n,v^*_n,t,Z)| = |S_1(u^*_n,v^*_n,t,S^*_n(t)) -
S_1(u^*_n,v^*_n,t,Z)|\\
&=\sum_{j=0}^{s^*-1}\big|[\lmr(h^{[n]}_{t^*_{j+1},t^*_j})-
\lmr(h^{[n]}_{t^*_j,t^*_j})]-
[\lmr(h^{[n]}_{t^*_{j+1},\phi(t^*_j)})-\lmr(h^{[n]}_{t^*_j,\phi(t^*_j)})]\big|
\end{align*}
where $\phi(t^*_j):=t_l$ if $t^*_j\in[t_l,t_{l+1})$ with $j=0,\ldots,s^*$ and
$l=0,\ldots,s$. Since $|v^*_n(t_{l+1})-v^*_n(t_l)|<\delta$, we have
$|v^*_n(\phi(t_l))-v^*_n(t_l)|<\delta$ for all $n\ge m_0$ and all
$l=0,\ldots, s$. Thus we get
\begin{align*}
|\lambda^*_n t-&S_1(u^*_n,v^*_n,t,Z)|\\
&= \sum_{j=0}^{s^*-1} |\lmr(h^{[n]}_{t^*_{j+1},t^*_j})-\lmr(h^{[n]}_{t^*_j,t^*_j})|
\cdot \left|1-\frac{\lmr(h^{[n]}_{t^*_{j+1},\phi(t^*_j)})-
\lmr(h^{[n]}_{t^*_j,\phi(t^*_j)})}{\lmr(h^{[n]}_{t^*_{j+1},t^*_j})-
\lmr(h^{[n]}_{t^*_j,t^*_j})} \right|
\end{align*}
Since $|u^*_n(t^*_{j+1})-u^*_n(t^*_j)|<\delta$ and
$|v^*_n(\phi(t^*_j))-v^*_n(t^*_j)|<\delta$ for all $n\ge m_0$ and all
$l=0,\ldots, s$, we have by Lemma \ref{Pro:DifQuoAbs2} c)
\[
|\lambda^*_n t-S_1(u^*_n,v^*_n,t,Z)| \le \epsilon
\sum_{j=0}^{s^*-1} \big(\lmr(h^{[n]}_{t^*_{j+1},t^*_j})-
\lmr(h^{[n]}_{t^*_j,t^*_j})\big)\le\epsilon
\]
for all $n\ge m_0$. The last inequality can be proven by using the monotonicity of
$(t,\tau)\mapsto\lmr(t,\tau)$ as follows
\begin{align*}
\sum_{j=0}^{s^*-1} &\big(\lmr(h^{[n]}_{t^*_{j+1},t^*_j})-
\lmr(h^{[n]}_{t^*_j,t^*_j})\big)\\[-0.5\baselineskip]
&\le \sum_{j=0}^{s^*-1} \big(\lmr(h^{[n]}_{t^*_{j+1},t^*_j})-
\lmr(h^{[n]}_{t^*_j,t^*_j})\big)
+\sum_{j=0}^{s^*-1} \big(\lmr(h^{[n]}_{t^*_{j+1},t^*_{j+1}})-
\lmr(h^{[n]}_{t^*_{j+1},t^*_j})\big)\\
&= \lmr(h^{[n]}_{t^*_{s^*},t^*_{s^*}}) - \lmr(h^{[n]}_{t^*_0,t^*_0}) =
\lmr(h^{[n]}_{t,t}) = t \le 1
\end{align*}
The assertion follows now, since $\lambda^*_n$ converges to $\lambda_0$.
\item If we put 3) and 4) together we find for every $\epsilon>0$ a $\mu>0$ so that for all
partitions $Z\subset S$ of the interval $[0,t]$ with $|Z|<\mu$ the
inequality
\begin{align} \label{Equ:4}
|S_1(u,v,t,Z)-\lambda_0 t|<\epsilon
\end{align}
holds.
As a consequence of this, we will show next that $\lambda_0\ne 0$. \\
So let us assume the opposite, namely $S_1(u,v,1,Z)\rightarrow 0$ for
$|Z|\rightarrow 0$, where $Z\subset S$ denotes a partition of the interval $[0,1]$.
Let $\delta>0$ such that the inequality from Proposition \ref{Pro:DifQuoAbs2} c) holds for
$\epsilon=\frac{1}{2}$. Since $t\mapsto u(t)$ and $t\mapsto v(t)$ are
(uniformly) continuous in $[0,1]$, we get
\[
\exists \mu>0:\, |\lt-\gt|<\mu \Rightarrow |u(\lt)-u(\gt)|,\,
|v(\lt)-v(\gt)|<\delta
\]
Denote by $Z:=\{t_0,\ldots,t_s\}\subset S$ a partition of the interval
$[0,1]$ with $|Z|<\mu$.
Consequently, we have for all $n\in\IN$ with $Z\subset S_n:=\{t^*_0,\ldots,t^*_{2^n}\}$
as before
\begin{align*}
0<\sigma:=S_1(u,v,1,Z)=\sum_{j=0}^{2^n-1} \lmr(h_{t^*_{j+1},\phi(t^*_j)})-
\lmr(h_{t^*_j,{\phi(t^*_j)}}),
\end{align*}
where $\phi(t^*_j):=t_l$ if $t^*_j\in[t_l,t_{l+1})$ with $j=0,\ldots,2^n$ and
$l=0,\ldots,s$. Hence $|v(\phi(t^*_l))-v(t^*_l)|<\delta$, so we get by
Lemma \ref{Pro:DifQuoAbs2} c)
\begin{align*}
S_1(u,v,1,Z) \le (1+\epsilon)\sum_{j=0}^{2^n-1} \lmr(h_{t^*_{j+1},t^*_j})-
\lmr(h_{t^*_j,t^*_j}) = (1+\epsilon)S_1(u,v,1,S_n)
\end{align*}
Thus we get $\frac{2}{3}\sigma\le S_1(u,v,1,S_n)$, so $S_1(u,v,1,S_n)$ does not tend to
zero as $n$ tends to infinity. This is a contradiction, so $\lambda_0\ne 0$.
\item Next we will show, as another consequence of equation (\ref{Equ:4}), that the
function $t\mapsto u(t)$ is strictly increasing.
For this purpose, we assume the opposite. Let $t_1<t_2$ with
$u(t_1)=u(t_2)$. Without loss of generality we can assume $t_1,t_2\in S$.
So we find $n_0\in\IN$ with $t_1,t_2\subset S_n$ for all $n\ge n_0$.
Let $\epsilon:=\frac{1}{2}(t_2-t_1)\lambda_0$. Then there exist $\mu_1,\mu_2>0$ such
that
\[
|S_1(u,v,t_1,Z_1)-\lambda_0 t_1|<\epsilon\quad \text{and} \quad
|S_1(u,v,t_2,Z_2)-\lambda_0 t_2|<\epsilon
\]
for all partitions $Z_1$ of $[0,t_1]$ and $Z_2$ of $[0,t_2]$ with $|Z_1|<\mu_1$
and $|Z_2|<\mu_2$. Consequently we find an $m_0\ge n_0$ with
$|S_{m_0}|<\min(\mu_1,\mu_2)$, so we get
\[
\lambda_0(t_2-t_1) < 2\epsilon,
\]
since $S_1(u,v,t_1,S_{m_0}(t_1)) = S_1(u,v,t_2,S_{m_0}(t_2))$.
This is a contradiction to $\lambda_0\ne 0$, so
$t\mapsto u(t)$ needs to be strictly increasing.
By applying the steps 3) - 6) to the second slit with $\tilde S_2$ instead of $S_1$, we
get $\lambda_0\ne 1$ and the strict monotonicity of $t\mapsto v(t)$.
\item Since $u$ and $v$ are strictly increasing self-homeomorphisms of $[0,1]$,
we can apply Proposition \ref{Pro:ProCalLam} (for $a=u$ and $b=v$)
to get Lipschitz continuous and increasing functions $c_1$ and $c_2$.
By equation (\ref{Equ:4}) we see
\[
c_1(t)=\lambda_0 t,\quad c_2(t)=(1-\lambda_0)t
\]
for all $t\in S$. Since $S$ is dense in $[0,1]$ this relation holds for
all $t\in[0,1]$. \\
Finally, from Proposition \ref{Pro:ProCalLam} it follows that for every $z\in \Omega_1$ the differential equation
\[\label{eq:twoslits}
\dot h_t(z) = h_t(z)\big( \lambda_0 \Phi(\xi_1(t),h_t(z),D_t) + (1-\lambda_0)
\Phi(\xi_2(t),h_t(z),D_t)\big)
\]
holds for all $t\in[0,1]$, where $h_t:=h_{t,t}$ and $\xi_j$ is the driving function for $\Gamma_j.$
As $t\mapsto \gamma_1(u(t))$ and $t\mapsto \gamma_2(v(t))$ are continuous,
the continuity of the driving functions $t\mapsto \xi_1(t)$ and
$t\mapsto \xi_2(t)$ follows directly from Proposition 7 in \cite{BoehmLauf}.
\end{selflist}
\end{proof}
\section{Uniqueness}
\begin{proof}[Proof of Theorem \ref{The:ConCoeffi} (Uniqueness)]
Let us denote by $u_1, u_2, v_1,v_2$ increasing self-homeomorphisms of $[0,1]$ with
$\lmr(u_1(t),v_1(t))=\lmr(u_2(t),v_2(t))=t$ for all $t\in[0,1]$ such that the functions
$h_t:= f_{u_1(t),v_1(t)}$ and $g_t:=f_{u_2(t),v_2(t)}$ satisfy the differential equations
\begin{eqnarray*}
\dot h_t(z) &=& h_t(z)\big( \lambda_1 \Phi(\xi_1(t),h_t(z),D_t) + (1-\lambda_1)
\Phi(\xi_2(t),h_t(z),D_t)\big),\\
\dot g_t(z) &=& g_t(z)\big( \lambda_2 \Phi(\zeta_1(t),g_t(z),E_t) + (1-\lambda_2)
\Phi(\zeta_2(t),g_t(z),E_t)\big)
\end{eqnarray*}
for all $t\in[0,1]$ with coefficients $0<\lambda_2\le\lambda_1<1$ and
continuous driving functions $\xi_1,\xi_2,\zeta_1,\zeta_2,$
where $\xi_j$ and $\zeta_j$ correspond to the slit $\Gamma_j.$
The continuity of the driving functions is an immediate consequence of Proposition 7
in \cite{BoehmLauf}, since $u_1, u_2, v_1,v_2$ are increasing self-homeomorphisms of $[0,1]$.
By $D_t$ and $E_t$ we denote
the circular slit disks that are uniquely determined by $u_1, v_1$ and $u_2, v_2$ respectively.
\begin{selflist}
\item First of all we will show $\lambda_1=\lambda_2$, so let us assume
$\lambda_1>\lambda_2$.
The differential equations immediately imply $h_t'(0)=g_t'(0)=e^t$ for all $t\in[0,1]$
and by Proposition \ref{Pro:ProCalLam} we get Lipschitz-continuous functions
$c_1$ and $c_2$ for the case $a=u_1$ and $b=v_1$.
These functions are differentiable a.e. and it holds
\[
\dot c_1(t) = \lambda_1, \quad
\dot c_2(t) = 1-\lambda_1, \quad \text{a.e. in} \; [0,1].
\]
This is based on the fact, that the functions
\[
z\mapsto \Phi(\xi_1(t),h_t;D_t),\quad
z\mapsto \Phi(\xi_2(t),h_t;D_t)
\]
are for fixed $t$ linear independent and equation (\ref{KLE2}).\\
Analogously, Proposition \ref{Pro:ProCalLam} gives us two Lipschitz-continuous
functions $d_1$ and $d_2$ for the case $a=u_2$ and $b=v_2$ with
\[
\dot d_1(t) = \lambda_2, \quad \dot d_2(t)= 1-\lambda_2 \quad \text{a.e. in}
\;[0,1].
\]
An immediate consequence of the Lipschitz continuity is
$$
c_1(t)=\lambda_1 t,\; d_1(t)=\lambda_2 t,\;
c_2(t)=(1-\lambda_1)t,\; d_2(t)=(1-\lambda_2)t \quad
\text{for all} \; t\in[0,1].
$$
Furthermore we set
\[
x_j(t):=\lmr\big(u_j(t),v_j(0)\big) = \lmr\big(u_j(t),0\big)
\]
for $j=1,2$. Denote by $0<t_0\le 1$ the first positive time when $u_1(t_0)=u_2(t_0)$.
Consequently $v_1(t_0)=v_2(t_0)$ and $x_1(t_0)=x_2(t_0)$ by normalization and the
monotonicity of $(t,\tau)\mapsto \lmr(t,\tau)$ in each variable.
Since
$$
\dot x_1(0) = \lambda_1 > \lambda_2 = \dot x_2(0)
$$
by Proposition \ref{Pro:ProCalLam}, we have $x_1(t)>x_2(t)$
and as a consequence $u_1(t)>u_2(t)$ for all $t\in (0,t_0)$. Consequently we have also
$\lmr\big(u_1(t),v_1(t_0)\big)>\lmr\big(u_2(t),v_2(t_0)\big)$ for all
$t\in(0,t_0)$. Thus we get
\begin{multline*}
\lmr\big(u_2(t),v_2(t_0)\big) < \lmr\big(u_1(t),v_1(t_0)\big) <
\lmr\big(u_1(t_0),v_1(t_0)\big) = \lmr\big(u_2(t_0),v_2(t_0)\big) = t_0
\end{multline*}
if $t<t_0$. This implies
\[
\frac{\lmr\big(u_1(t_0),v_1(t_0)\big)-\lmr\big(u_1(t),v_1(t_0)\big)}{t_0-t} <
\frac{\lmr\big(u_2(t_0),v_2(t_0)\big)-\lmr\big(u_2(t),v_2(t_0)\big)}{t_0-t}
\]
for all $t<t_0$. If $t$ tends to $t_0$ we get $\lambda_1\le \lambda_2$
by Proposition \ref{Pro:ProCalLam}. This is a contradiction, so $\lambda_1=\lambda_2=:\lambda$.
\item Next we prove the uniqueness of the parametrizations $u$ and $v$, i.e. we show
$u_1\equiv u_2$. By using the result from the first part, we have
\[
c_1(t)=d_1(t) = \lambda t,\quad
c_2(t)=d_2(t) = (1-\lambda)t,
\]
for all $t\in[0,1]$.\\
Again, we extend both $\gamma_1$ and $\gamma_2$ to an interval
$[0,T^*],$ $T^*>1,$ such that $\gamma_1[0,T^*]$ and $\gamma_2[0,T^*]$ are still
disjoint slits and $\lmr(T^*,0)\geq1,$ $\lmr(0,T^*)\geq 1.$
Let $t\in(0,1)$ be fixed. Next we denote by $t_0=0<t_1<\ldots<t_n$
and $\tilde t_0=0<\tilde t_1<\ldots< \tilde t_n$ the unique values such that
\begin{eqnarray*}
&&\lmr\big(u_1(t_{l+1}),v_1(t_l)\big) - \lmr\big(u_1(t_l),v_1(t_l)\big)\\ &=&
\lmr\big(u_2(\tilde t_{l+1}),v_2(\tilde t_l)\big) -
\lmr\big(u_2(\tilde t_l),v_2(\tilde t_l)\big) =
\frac{\lambda}{n} t
\end{eqnarray*}
holds.
By induction, it is easy to see that
$$
u_1(t_l)= u_2(\tilde t_l) \quad \text{and} \quad
v_1(t_l) = v_2(\tilde t_l) \quad \text{for all}\;\, l=1,\ldots,n.
$$
Furthermore, the values $|t_{l+1}-t_l|$ and $|\tilde t_{l+1}-\tilde t_l|$
$(l=0,\ldots, n-1)$ become arbitrary small, if $n$ is big enough.
Since $t_n$ and $\tilde t_n$ are bounded by $T^*$, we find convergent subsequences
$(t_{n_j})_{j\in\IN}$ and $(\tilde t_{n_j})_{j\in\IN}$.
The limit of both sequences needs to be $t$ since the sum
\[
\sum_{l=0}^{n-1} \lmr\big(u_j(t_{l+1}),v_j(t_l)\big) -
\lmr\big(u_j(t_l),v_j(t_l)\big),
\]
converges to $c_1(t)=d_1(t)=\lambda t$ for each $j=1,2$.
Finally, as a consequence of the continuity of $t\mapsto u_j(t)$, we get
$u_1(t)=u_2(t)$, so the proof is complete.
\end{selflist}
\end{proof}
|
\section{Analysis}
\label{sec:analysis}
\begin{definition}
\label{def:state-space}
Let $G$ and $Q$ be two graphs, and $O$ an ordering of the vertices of $Q$. A state space is the set of all partial solutions that algorithm $findSolutions$ explores to find the solutions of $Q$ over $G$ using $O$. We denote this state space as $\Psi(G, Q, O)$.
\end{definition}
\begin{definition}
\label{def:inner-ordering}
Let $Q$ be a graph, $O$ an ordering of the vertices of $Q$, and $O'$ an ordering of a subset of the vertices of $Q$. $O'$ is an inner ordering of $O$ if $O' \subseteq O$, and the vertices in $O'$ appear in the same order as in $O$.
\end{definition}
\begin{theorem}
\label{theo:partial-solutions}
Let $G$ and $Q$ be two graphs, $O$ an ordering of the vertices of $Q$, and $O'$ an inner ordering of the vertices of $M \subseteq MHC(Q)$ such that $M \subseteq Q$. The partial solutions computed by algorithm $findSolutions$ using $O$ are included in the partial solutions computed by algorithm $findSolutions$ using $O'$.
\end{theorem}
\begin{proof}
Let $\Psi(G, Q, O)$ and $\Psi(G, Q, O')$ be the state spaces of algorithm $findSolutions$ using $O$ and $O'$, respectively. The proof follows from reductio ad absurdum: assume that there exists a partial solution $s \in \Psi(G, Q, O')$ such that $s \notin \Psi(G, Q, O)$. If this occurs, it is because there exists a pair $(k = v)$, in which $k \in vertices(Q)$ and $v \in vertices(G)$, such that it does not belong to any partial solution in $\Psi(G, Q, O)$.
To compute $(k = v)$, algorithm $findSolutions$ unifies $G$ and $q$, which is the graphlet that results from substituting it using the previous partial solution $current$ (see lines~\ref{line:find-solutions:get-graphlet}--~\ref{line:find-solutions:unify} in Figure~\ref{fig:find-solutions}). Since $current$ belongs to both $\Psi(G, Q, O)$ and $\Psi(G, Q, O')$, $q$ is also computed when calling algorithm $findSolutions$ using $O$. Therefore, the same set of partial solutions are computed when unifying $G$ and $q$, in which $(k = v)$ belongs to one or more partial solutions in $\Psi(G, Q, O)$, and $s \in \Psi(G, Q, O)$, which is a contradiction. As a conclusion, the partial solutions computed by algorithm $findSolutions$ using $O$ are included in the partial solutions computed by algorithm $findSolutions$ using $O'$.
\end{proof}
\section{Conclusions}
\label{sec:conclusions}
In this paper, we present a new vertex-at-a-time proposal to match a query graph over a data graph. Our proposal is based on the concept of graphlets, each of which comprises a vertex of a graph, all of the immediate neighbors of that vertex, and all of the edges that relate those neighbors. Furthermore, our proposal use minimum hub covers, each of which comprises a subset of vertices in a query graph that account for all of the edges in that query graph. The main challenge of using minimum hub covers is to select the best one among all possible minimum hub covers for a given query graph and the best ordering to be processed. We have discussed three different aspects that must be taken into account, i.e., the connection of the vertices in the minimum hub cover, the selectivity of the vertices and the use of the data graph. As a product of this discussion, we have devised a number of heuristics to select the best minimum hub cover and the best ordering.
We have presented the algorithms to implement our proposal, an implementation of these algorithms that is based on XQuery and RDF, and some evaluation results of this implementation. Our main conclusion is that we have experimentally validated that using our heuristics we are able to select the best minimum hub cover and the best ordering. Furthermore, our results experimentally validate that it is better to use a minimum hub cover to process a query graph instead of using any random ordering of these vertices.
\section{Evaluation}
\label{sec:evaluation}
In the evaluation of our proposal, our main goal is to experimentally validate that the heuristics that we define in Section~\ref{sec:mhc} actually select the best minimum hub cover and the best ordering, i.e., there exists an optimal ordering of a minimum hub cover that performs better than any other random ordering of the minimum hub covers or the complete set of vertices. To validate this, we ran our implementation using our motivating example.
\begin{table}
\centering
\begin{rctabular}{|l|l||l|l|}{Ordering & Explored & Ordering & Explored}
$(3, 5, 8)$ & 189 & $(6, 7, 1)$ & 297 \\
$(3, 8, 5)$ & 189 & $(6, 7, 2)$ & 297 \\
$(7, 6, 1)$ & 207 & $(6, 7, 3)$ & 297 \\
$(7, 6, 2)$ & 207 & $(6, 1, 7)$ & 425 \\
$(7, 6, 3)$ & 207 & $(6, 2, 7)$ & 425 \\
$(8, 5, 3)$ & 211 & $(6, 3, 7)$ & 425 \\
$(5, 8, 3)$ & 211 & $(3, 7, 6)$ & 1081 \\
$(1, 6, 7)$ & 233 & $(7, 3, 6)$ & 1171 \\
$(2, 6, 7)$ & 233 & $(1, 7, 6)$ & 2029 \\
$(3, 6, 7)$ & 245 & $(2, 7, 6)$ & 2029 \\
$(8, 3, 5)$ & 279 & $(7, 1, 6)$ & 2131 \\
$(5, 3, 8)$ & 279 & $(7, 2, 6)$ & 2131
\end{rctabular}
\caption{Orderings of the minimum hub covers and the solutions explored.}
\label{tab:ordering-explored}
\end{table}
In our motivating example, our proposal outputs 24 solutions and Table~\ref{tab:ordering-explored} shows the number of solutions explored when using different minimum hub covers and orderings. The best orderings are $(3, 5, 8)$ and $(3, 8, 5)$, each of which explores 189 solutions, which is what we predicted using our heuristics.
In addition to the minimum hub covers, we computed the number of solutions explored when processing all of the query graphlets using all possible orderings. The best ordering among these is $(3, 5, 8, 6, 1, 2, 4, 7)$, in which our proposal explores 309 solutions. Note that the initial vertices correspond to the best ordering of the minimum hub cover, which entails that we have already computed all of the solutions after processing these three vertices. Therefore, we are performing an extra computation that it is not mandatory. On the contrary, the worst ordering is $(5, 2, 7, 1, 8, 4, 6, 3)$, which explores 8815 solutions.
These results experimentally validate that it is better to use a minimum hub cover to process a query graph instead of using any random ordering of these vertices.
\section{Graph matching}
\label{sec:graph-matching}
\begin{figure}
\begin{listing}[99]
$findSolutions$
input\tab[2]$graph{:} Graph$; $query{:} Graph$; $V{:}$ List of
\tab[6]$Vertex$; $i{:} Integer$; $current{:} Solution$
output\tab $O{:}$ Set of $Solution$
variables $q{:} Graphlet$; $P, R{:}$ Set of $Solution$
// If $i$ is at the end of the order
if $|V| = i$ \label{line:find-solutions:initial-if}
$O$ := $\{current\}$ \label{line:find-solutions:add-current-solution}
else
// Substitute the current graphlet using the
// current solution
$q$ := $getGraphlet(query, V[i])$ \label{line:find-solutions:get-graphlet}
$q$ := $substitute(g, current)$ \label{line:find-solutions:substitute}
// Compute partial solutions by unification
$P$ := $unify(graph, q)$ \label{line:find-solutions:unify}
// If the graphlet does not unify
if $P = undefined$ \label{line:find-solutions:not-unify-1}
$O$ := $undefined$ \label{line:find-solutions:not-unify-2}
// If the graphlet unifies but it is ground
else if $P = \emptyset$ \label{line:find-solutions:empty-unify-1}
$O$ := $findSolutions(graph, query, V, i + 1, current)$ \label{line:find-solutions:empty-unify-2}
else
$O$ := $\emptyset$
// Iterate over the valid partial solutions
for each $p{:} Solution \mid p \in P \cdot isValid(p, current)$ do \label{line:find-solutions:valid-iteration}
$R$ := $findSolutions(graph, query, V, i + 1, p \cup current)$ \label{line:find-solutions:recursive-call}
// If the recursive solutions unify
if $R \neq undefined$
$O$ := $O \cup R$ \label{line:find-solutions:add-partial}
end if
end for
// If the graphlet does not unify
if $O = \emptyset$ \label{line:find-solutions:no-partial-1}
$O$ := $undefined$ \label{line:find-solutions:no-partial-2}
end if
end if
end if
\end{listing}
\caption{Algorithm to find the solutions of a query over a data graph.}
\label{fig:find-solutions}
\end{figure}
Our proposal takes a data graph, a query graph, and an ordering of query vertices to be processed as input. It computes all of the subgraphs by matching the query over the data and, finally, it outputs a number of solutions, each of which comprises the data vertices that match the query vertices. Figure~\ref{fig:find-solutions} shows the algorithm to compute these solutions. Note that, for the sake of brevity, we present the recursive algorithm, whose initial call is the following: $findSolutions(graph, query, V, 0, \emptyset)$.
In our proposal, the recursion ends when we reach the end of the ordering, i.e., $|V| = i$, in which we add the current solution to the final set (see lines~\ref{line:find-solutions:initial-if}--~\ref{line:find-solutions:add-current-solution}). For the rest of the cases, we first substitute the current graphlet of the query graph using the current solution (see lines~\ref{line:find-solutions:get-graphlet}--~\ref{line:find-solutions:substitute}), the intuition behind this is that the current solution may comprise some ground query vertices, which can be replaced in the graphlet by the actual data to narrow the search. With the substituted graphlet, we try to find all of the possible combinations of data vertices for the query vertices (see line~\ref{line:find-solutions:unify}), which result in a number of partial solutions. We perform this by unifying the substituted graphlet with the graphlets in the data graph. Note that a data graphlet $g_d = {<}v_d, N_d, B_d{>}$ unifies with a query graphlet $g_q = {<}v_q, N_q, B_q{>}$ if $|N_q| \leq |N_d|$, $|B_q| \leq |B_d|$, and there exists a solution $s$ such that $s(v_q) = v_d$, $s(N_q) \subseteq N_d$, and $s(B_q) \subseteq B_d$. This unification process depends on the technology used to implement it; additional details are presented in Section~\ref{sec:implementation}.
After the computation of these partial solutions, we have to check if the query has unified with the data. In the case that it does not unify, we output $undefined$ as the final solution, since this mean that the query has no solution (see lines~\ref{line:find-solutions:not-unify-1}--~\ref{line:find-solutions:not-unify-2}). If it unifies and the set of partial solutions is empty, this entails that the query is ground (see lines~\ref{line:find-solutions:empty-unify-1}--~\ref{line:find-solutions:empty-unify-2}), i.e., all of the vertices of the substituted query graphlet belong to the data graph. So we continue with the following vertex in the ordering by performing a recursive call.
Another possibility is that the partial solutions are not empty, in which we iterate over the set of partial solutions that are valid (see line~\ref{line:find-solutions:valid-iteration}). A partial solution $p$ is valid with respect to a current solution $current$ if the query vertices are not repeated ($keys(p) \cap keys(current) = \emptyset$), and the data vertices are not repeated ($values(p) \cap values(current) = \emptyset$). The intuition behind this is that a query and/or data vertex must appear only once in the solution that results from combining $p$ and $current$. Therefore, for each valid partial solution, we continue with the following vertex in the ordering, and the combination of $p$ and $current$ as the next current solution (see line~\ref{line:find-solutions:recursive-call}). If the resulting solutions are not $undefined$, we add them to the output solutions (see line~\ref{line:find-solutions:add-partial}). Finally, if we do not add any solution to the output set, this means that we have found no solution for this query, so we return $undefined$ (see lines~\ref{line:find-solutions:no-partial-1}--~\ref{line:find-solutions:no-partial-2}).
\begin{example}
\label{ex:graph-matching}
To illustrate how our proposal works, we use our motivating example (see Section~\ref{sec:preliminaries}). We also use the following minimum hub cover and ordering: $(3, 6, 7)$; so $findSolutions(graph, query, (3, 6, 7), 0, \emptyset)$ is the first call to our algorithm, in which $graph$ and $query$ comprise all of the data and query graphlets that we presented above. Since $i = 0 \neq |V| = 3$, we retrieve the graphlet of vertex $3$ and substitute it using the current solution, which is empty, so the substituted graphlet is exactly the same.
In the next step, we unify graphlet $3$ with the data graphlets; note that graphlet $3$ comprises 4 neighbors and 2 boundaries, so it is necessary to unify with data graphlets with, at least, 4 neighbors and 2 boundaries. In this case, it can be unified with $b$, $j$, $k$, and $o$ graphlets; however, graphlet $3$ cannot be unified with graphlet $b$ since both boundaries of $b$ comprises vertex $j$, which is not supported by the boundaries of graphlet $3$. For the rest of the graphlets, we compute a number of partial solutions and we iterate over them. All of these partial solutions are valid since the current solution is empty. One of these partial solutions is the following: $p_{01} = \{3 = k, 1 = a, 2 = j, 4 = l, 6 = m\}$; then, we perform a recursive call: $findSolutions(graph, query, (3, 6, 7), 1, p_{01})$.
In this recursive call $i = 1 \neq |V| = 3$, so we retrieve the graphlet of vertex $6$ and substitute it using the current solution, which results in ${<}m, \{k, l, 5, 8\}, \{\{k, l\}\}{>}$ graphlet. We unify it with the data graph and retrieve two valid partial solutions: $p_{11} = \{5 = l, 8 = n\}$ and $p_{12} = \{5 = n, 8 = l\}$. We iterate over them and perform recursive calls; unfortunately, both partial solutions do not unify with the data graph, i.e., it does not exist a data graphlet with ${l, n}$ neighbors and zero boundaries. Therefore, we return $undefined$ as the solution of this recursive call.
Another valid partial solution is $p_{02} = \{3 = k, 1 = l, 2 = m, 4 = a, 6 = j\}$. We perform a recursive call in which $i = 1$, and we retrieve and substitute graphlet $6$ using $p$, which results in ${<}j, \{k, a, 5, 8\}, \{\{k, a\}\}{>}$ graphlet. We unify this graphlet with the data graph and retrieve several valid partial solutions, e.g., $p_{13} = \{5 = c, 8 = e\}$. In the next step, we call $findSolutions(graph, query, (3, 6, 7), 2, p_{02} \cup p_{13})$ and substitute graphlet $7$ with the current solution. We obtain the following graphlet: ${<}7, \{c, e\}, \{\}{>}$, which unifies with the data graph and produces only one partial solution: $p_{21} = \{7 = d\}$. Finally, we call $findSolutions(graph, query, (3, 6, 7), \\ 3, p_{02} \cup p_{13} \cup p_{21})$, in which $i = 3 \neq |V| = 3$, so we add the following solution to the final set: $\{1 = l, 2 = m, 3 = k, 4 = a, 5 = c, 6 = j, 7 = d, 8 = e\}$.
\end{example}
\section{Implementation}
\label{sec:implementation}
\begin{figure}
\centering
\begin{minipage}{0.45\hsize}
\begin{listing}
<graph>
<graphlet vertex=``a''>
<neighbor>b</neighbor>
<neighbor>j</neighbor>
<neighbor>k</neighbor>
<boundary>
<vertex>b</vertex>
<vertex>j</vertex>
</boundary>
<boundary>
<vertex>j</vertex>
<vertex>k</vertex>
</boundary>
</graphlet>
\dots
</graph>
\end{listing}
\end{minipage}
\quad
\begin{minipage}{0.45\hsize}
\begin{listing}
<graph>
<graphlet vertex=``1''>
<neighbor>2</neighbor>
<neighbor>3</neighbor>
<boundary>
<vertex>2</vertex>
<vertex>3</vertex>
</boundary>
</graphlet>
<graphlet vertex=``2''>
<neighbor>1</neighbor>
<neighbor>3</neighbor>
<boundary>
<vertex>1</vertex>
<vertex>3</vertex>
</boundary>
</graphlet>
\dots
</graph>
\end{listing}
\end{minipage}
\caption{Graph examples in XML.}
\label{fig:xml-graph-examples}
\end{figure}
We implemented a research prototype of our proposal using Java~1.6, eXist-DB~2.1, and Jena~2.10.1. Furthermore, we used Guava~15.0 to implement auxiliary operations, and JGraph~2.2 to draw graphs in a GUI utility. Every graph is transformed into its graphlets, which are stored as an XML object in eXist-DB. Figure~\ref{fig:xml-graph-examples} presents a part of the XML object of the data and query graphlets in our motivating example. The unification process in our implementation is divided into two steps, namely: XQuery extraction and SPARQL unification.
In the first step, our goal is to select those graphlets in the data graph that can be unified with the single query graphlet. Recall that this query graphlet may comprise also ground data, so we take the query graphlet as input and output a query in XQuery to retrieve the data graphlets. In Example~\ref{ex:graph-matching}, we retrieve all of the data graphlets that can unify with graphlet $3$, which comprises 4 neighbors and 2 boundaries, so we issue the following query:
\begin{listing}
for \$x in doc(``DataGraph.xml''/graph/graphlet)
where count(\$x/neighbor) $\geq$ 4 and
count(\$x/boundary) $\geq$ 2
return \{\$x\}
\end{listing}
Furthermore, we also retrieve the data graphlets that can unify with graphlet ${<}m, \{k, l, 5, 8\}, \{\{k, l\}\}{>}$, which also comprises 4 neighbors and 2 boundaries, so we issue the following query:
\begin{listing}
for \$x in doc(``DataGraph.xml''/graph/graphlet)
let \$b := \$x/boundary
where count(\$x/neighbor) $\geq$ 4 and
count(\$x/boundary) $\geq$ 2 and
\$x[@vertex = ``m''] and
\$b/vertex = ``k'' and
\$b/vertex = ``l''
return \{\$x\}
\end{listing}
In the second step step, we wish to retrieve all of the possible combinations of data vertices. To perform this, we iterate over the whole set of data graphlets that we have retrieved in the previous set. We transform each of them into a set of RDF triples, e.g., the following triples correspond to the RDF transformation of data graphlet $a$ in our motivating example:
\begin{listing}
a neighbor b, j, k ;
boundary {\_:}B1, {\_:}B2 .
{\_:}B1 vertex b, j .
{\_:}B2 vertex j, k .
\end{listing}
Then, we transform the data graphlet into a SPARQL query that is posed over the RDF data. In Example~\ref{ex:graph-matching}, for query graphlet $3$, we issue the following query:
\begin{listing}
SELECT DISTINCT ?1 ?2 ?3 ?4 ?6
WHERE \{
?3 neighbor ?1, ?2, ?4, ?6 ;
boundary {\_:}X1, {\_:}X2 .
{\_:}X1 vertex ?1, ?2 .
{\_:}X2 vertex ?4, ?6 .
\}
\end{listing}
Note that this query may retrieve the same values for different variables that must be filtered out. Furthermore, for query graphlet ${<}m, \{k, l, 5, 8\}, \{\{k, l\}\}{>}$, we issue the following query:
\begin{listing}
SELECT DISTINCT ?5 ?8
WHERE \{
m neighbor k, l, ?5, ?8 ;
boundary {\_:}X1 .
{\_:}X1 vertex k, l .
\}
\end{listing}
\section{Introduction}
Nowadays, there is an increasing interest in using graphs to represent data~\cite{journals/prl/WeberLD12}. This is due to the fact that graphs allow to model complex relationships that cannot be easily captured using traditional data models like the relational model~\cite{conf/sigmod/HeS08}. One of the crucial tasks of graph management is graph matching, which aims to find all of the subgraphs in a data graph that match a query graph~\cite{journals/jiis/BhattacharjeeJ12}.
In the literature, proposals to perform graph matching can be broadly classified into two different categories: graph-at-a-time and vertex-at-a-time. In the former, the whole query graph is processed at the same time, in which GraphQL~\cite{conf/sigmod/HeS08}, QuickSI~\cite{journals/pvldb/ShangZLY08}, Ullman~\cite{journals/jacm/Ullmann76} and VFLib~\cite{conf/iciap/CordellaFSV99} are the most representative proposals. In the latter, a single vertex of the query graph is processed at the same time, in which SAPPER~\cite{journals/pvldb/ZhangYJ10} and TALE~\cite{conf/icde/TianP08} are the most representative proposals. The most appealing advantage of the vertex-at-a-time proposals is that they are able to prune the search space using matching conditions, such as vertex indices~\cite{journals/pvldb/ZhangYJ10}, path indices~\cite{conf/icpr/GiugnoS02}, or frequent structures~\cite{conf/edbt/ZhangLY09}.
Our goal is to devise a new vertex-at-a-time proposal that is based on graphlets, each of which is a three tuple that comprises a vertex of a graph, all of the immediate neighbors of that vertex, and all of the edges that relate those neighbors. In our proposal, both the data and query graphs are represented as sets of graphlets. To process the query graph over the data graph, we first reduce the number of graphlets in the query graph based on a new notion of edge covering in graph theory, called the minimum hub cover, which comprises a subset of vertices in a graph that account for all of the edges in that graph. Note that there may exist more than one minimum hub covers for the same query graph. Then, our proposal sort the vertices of the minimum hub cover to reduce the search space as much as possible. Finally, we use a number of algorithms to compute the subgraphs of the data graph that match the query graph.
In this paper, we assume that the minimum hub covers for a given query graph are already computed. In Section~\ref{sec:preliminaries}, we introduce the main concepts that we use throughout the paper. We present a discussion on how to select the best minimum hub cover and the best ordering of this minimum hub cover in Section~\ref{sec:mhc}. Section~\ref{sec:graph-matching} describes the algorithms to match a query graph over a data graph. In Section~\ref{sec:implementation}, we describe an implementation of the previous algorithms using XQuery and RDF. Section~\ref{sec:evaluation} presents some initial results that we have obtained using this implementation. Finally, Section~\ref{sec:conclusions} recaps on our main conclusions.
\section{On selecting and ordering minimum hub covers}
\label{sec:mhc}
The main challenge of using minimum hub covers is to select the best one among all possible minimum hub covers for a given query graph, and the best ordering to be processed, so we explore the minimum number of solutions. To perform this, we focus on three different aspects that must be taken into account: 1) The connection of the vertices in the minimum hub cover; 2) The selectivity of the vertices; and 3) The use of the data graph. In the rest of this section we discuss about these three aspects.
Regarding the connection of vertices, it is mandatory to take the connection of the vertices into account when selecting and ordering the minimum hub covers. This entails that, if vertices $u$ and $v$ appear consecutively in the ordering of a minimum hub cover, both graphlets have to be connected by, at least, one vertex. This avoids performing cartesian products when exploring the solutions to the query graph. For instance, in our motivating example, for the minimum hub cover $\{1, 6, 7\}$, we must avoid orderings $(1, 7, 6)$ and $(7, 1, 6)$ since graphlets $1$ and $7$ do not have any vertices in common. The rest of the orderings are suitable candidates since graphlets $1$ and $6$ have vertex $3$ in common, and graphlets $6$ and $7$ have vertices $5$ and $8$ in common.
Regarding the selectivity of the vertices, it is mandatory to ensure the maximum selectivity of the vertices, i.e., we must select first those graphlets that filter the maximum number of graphlets in the data graph. To compute this, we rely on the concept of free neighbors, which are those neighbors that do not appear in the boundaries of a data graphlet. One heuristic is that the most selective graphlet is the one that has more boundaries and less free neighbors. Furthermore, if the number of boundaries is equal to zero, another heuristic is that the most selective graphlet is the one that has more free neighbors.
\begin{table}
\centering
\begin{rctabular}{|r|r|r|}{Graphlet & Boundaries & Free neighbors}
1 & 1 & 0 \\
2 & 1 & 0 \\
3 & 2 & 0 \\
4 & 1 & 0 \\
5 & 1 & 0 \\
6 & 1 & 2 \\
7 & 0 & 2 \\
8 & 0 & 2
\end{rctabular}
\caption{Boundaries and free neighbors of our motivating example.}
\label{tab:graphlet-boundaries-free}
\end{table}
In our motivating example, the number of boundaries and free neighbors of the graphlets for the query graph are shown in Table~\ref{tab:graphlet-boundaries-free}. Using the previous heuristics, we have to select the most selective graphlet among the following: 1, 2, 3, 5, 6, 7, and 8. In this case, 3 is the most selective graphlet since the number of boundaries is two, which is the maximum number of boundaries. We have already discarded minimum hub covers $\{1, 6, 7\}$ and $\{2, 6, 7\}$, since they do not contain graphlet 3. Then, in the following step, we have to select the most selective graphlet among the following: 5, 6, 7, and 8; the most selective one is graphlet 5, which has more boundaries and less free neighbors than the others, so we discard minimum hub cover $\{3, 6, 7\}$. Therefore, using our heuristics, the best minimum hub cover and ordering is: $(3, 5, 8)$.
Regarding the use of the data graph, the previous heuristics take only the query graph into account; however, there may be some scenarios in which the data graph helps us to prune our search. For instance, in our motivating example, graphlet 6 has only two possible solutions: graphlets $b$ and $j$ in the data graph. Therefore, to avoid the exploration of a large number of solutions, it is mandatory to discover as soon as possible that graphlet 6 has only $b$ and $j$ as solutions. In these cases, it is possible to store a number of metadata statistics about the data graph to help us prune the search space.
\section{Preliminaries}
\label{sec:preliminaries}
\begin{figure}
\subfigure[Data graph.]{
\includegraphics[scale=0.75]{fig/Preliminaries/data-example}
}
\quad
\subfigure[Query graph.]{
\includegraphics[scale=0.75]{fig/Preliminaries/query-example}
}
\caption{Motivating example.}
\label{fig:motivating-example}
\end{figure}
To illustrate our proposal, we use the data and query graphs that are shown in Figure~\ref{fig:motivating-example}. The data graph comprises a total number of 16 vertices and 24 edges connecting those vertices. The graphlets that result of these data graph are the following:
\begin{formulae}[l l l]
{<}a, & \{b, j, k\}, & \{\{b, j\}, \{k, j\}\}{>} \\
{<}b, & \{a, c, j, m\}, & \{\{a, j\}, \{c, j\}\}{>} \\
{<}c, & \{b, d, j\}, & \{\{b, j\}\}{>} \\
{<}d, & \{c, e\}, & \{\}{>} \\
{<}e, & \{d, j, o\}, & \{\{j, o\}\}{>} \\
{<}h, & \{i, o\}, & \{\}{>} \\
{<}i, & \{h, j\}, & \{\}{>} \\
{<}j, & \{a, b, c, e, k, i, o\}, & \{\{a, b\}, \{a, k\}, \{b, c\}, \{e, o\}\}{>} \\
{<}k, & \{a, j, l, m\}, & \{\{a, j\}, \{l, m\}\}{>} \\
{<}l, & \{k, m\}, & \{\{k, m\}\}{>} \\
{<}m, & \{b, k, l, n\}, & \{\{k, l\}\}{>} \\
{<}n, & \{m\}, & \{\}{>} \\
{<}o, & \{e, h, j, p, q, r\}, & \{\{e, j\}, \{p, q\}\}{>} \\
{<}p, & \{o, q\}, & \{\{o, q\}\}{>} \\
{<}q, & \{o, p\}, & \{\{o, p\}\}{>} \\
{<}r, & \{o\}, & \{\}{>}
\end{formulae}
Each graphlet is a three tuple in which the first element is the main vertex $v$, the second element is the set of neighbors of $v$, i.e., those vertices that are related to $v$ by an edge, and the third element is the set of boundaries, each of which is a pair of vertices that belong to the neighbors of $v$ and there exists an edge between them. Therefore, we represent a graphlet as the following tuple: ${<}v, N, B{>}$, in which $N$ is the set of neighbors and $B$ is the set of boundaries. A graph is represented as a set of graphlets, and we use the following auxiliary functions to have access to all of the vertices and edges of a given graph $g$: $vertices(g)$ and $edges(g)$, respectively.
The query graph in Figure~\ref{fig:motivating-example} comprises a total number of 8 vertices and 10 edges, whose graphlets are the following:
\begin{formulae}[l l l]
{<}1, & \{2, 3\}, & \{\{2, 3\}\}{>} \\
{<}2, & \{1, 3\}, & \{\{1, 3\}\}{>} \\
{<}3, & \{1, 2, 4, 6\}, & \{\{1, 2\}, \{4, 6\}\}{>} \\
{<}4, & \{3, 6\}, & \{\{3, 6\}\}{>} \\
{<}5, & \{6, 7\}, & \{\}{>} \\
{<}6, & \{3, 4, 5, 8\}, & \{\{3, 4\}\}{>} \\
{<}7, & \{5, 8\}, & \{\}{>} \\
{<}8, & \{6, 7\}, & \{\}{>}
\end{formulae}
Our goal is to find all of the subgraphs in the data graph that match the query graph. The final result is a set of solutions, each of which comprises a set of pairs in which the first element, the key, is a vertex in the query graph, and the second element, the value, is a vertex in the data graph. Therefore, we represent a solution $s$ of a query graph $q$ over a data graph $g$ as a set of pairs $(u, v)$, in which $u \in vertices(q)$ and $v \in vertices(g)$. We also define two auxiliary functions: $keys$ and $values$; the former retrieves all of the vertices of the query graph in the solution, and the latter retrieves all of the vertices of the data graph in the solution. Furthermore, every solution has to comprise all of the vertices in the query graph and a subset of vertices of the data graph, i.e., $keys(s) = vertices(q)$ and $values(g) \subseteq vertices(g)$. A sample solution in our motivating example is the following: $\{(1, m), (2, l), (3, k), (4, a), (5, c), (6, j), (7, d), (8, e)\}$.
Our proposal relies on using minimum hub covers, each of which contains a subset of vertices in the query graph whose graphlets cover all of the edges of the original query graph. This entails that processing a minimum hub cover is exactly the same as processing all of the graphlets of the query graph. Therefore, $M \subseteq vertices(q)$ is a minimum hub cover of $q$ if $M$ is the smallest set, and these two conditions hold:
\begin{itemize}
\item There exists an edge ${<}u, v{>} \in edges(g)$, such that $u \in M$ or $v \in M$.
\item There exist two edges ${<}u, v{>} \in edges(g)$ and ${<}v, x{>} \in edges(g)$, such that $x \in M$.
\end{itemize}
Note that there may exist more than one minimum hub cover for the same query graph. For instance, in our motivating example, there exist four minimum hub covers: $\{1, 6, 7\}$, $\{2, 6, 7\}$, $\{3, 6, 7\}$, and $\{3, 5, 8\}$. Now, the challenge is to select the minimum hub cover and the ordering that will produce the minimum number of solutions to explore and, therefore, the minimum processing time. |
\section{Introduction}
In the paper \cite{ETe} Emparan and Teo considered an
electrostatic solution of the Einstein-Maxwell equations for a
non-extreme dihole -- a configuration consisting of two non-
extremal Reissner-Nordstr\"om black holes \cite{Rei,Nor} endowed
with equal masses and opposite charges of the same magnitude.
Extension of their results to the case of rotation was referred to
by those authors as an interesting work to be done in the future.
The problem of obtaining such an extension seems to have been
attacked in a recent paper of Cabrera-Munguia et al \cite{CLL} who
constructed a specific 4-parameter exact solution for two
oppositely charged counter-rotating Kerr-Newman (KN) black holes
\cite{NCC}. Though technically the paper \cite{CLL} is correct,
the solution itself is, however, frankly unphysical because of the
presence in it of non-vanishing magnetic charges induced by
rotation of electric charges. As a consequence, the usual Smarr
mass formula \cite{Sma} does not hold for the black-hole
constituents comprising that binary configuration and, moreover,
the authors of \cite{CLL} failed to find a concise expression for
the quantity $\sigma$ which would be similar for instance to that of
the physically parametrized Bret\'on-Manko (BM) solution
\cite{BMa,MRR}.
The present paper aims at working out a correct 4-parameter model
for stationary diholes with antiparallel rotation of its
constituents. To accomplish this task, we shall make use of a
5-parameter asymptotically flat specialization of the
Ernst-Manko-Ruiz (EMR) equatorially antisymmetric solution
\cite{EMR1,SRo} possessing arbitrary parameters of electric and
magnetic dipole moments, which will enable us to eliminate the
individual magnetic charges of the constituents by choosing
appropriately the values of the latter dipole parameters. This
will secure the validity of Smarr's formula for each black-hole
constituent and, in turn, will enable us to find a simple
expression for $\sigma$ in terms of the Komar quantities \cite{Kom}
which is a key point for elaborating the physical parametrization
of the whole model. After having reached our main objective, we
will prove that, similar to the BM model of equally charged
counter-rotating black-hole constituents, the configuration
obtained for KN black holes with opposite charges verifies (and
actually saturates) the inequality for interacting black holes
with struts recently derived by Gabach Clement \cite{Gab}.
\section{The 5-parameter asymptotically flat EMR solution\\
in $\sigma$-representation}
A key result of the paper \cite{CLL} is its expressions (14) for
the axis data $e(z)$ and $f(z)$ allowing one to construct in a
straightforward manner the corresponding entire metric with the
aid of the general formulas of the extended $N$-soliton solution
\cite{RMM} (we also refer the interested reader to Appendix of
Ref.~\cite{SRo} for a complete set of algebraic relations involved
in the construction procedure of the $N=2$ case). However, an
explanation the authors of \cite{CLL} give to the origin of those
expressions -- the solution (11) of a complicated system of
algebraic equations for certain metric functions -- is in fact
misleading, and below we will give a veritable simple derivation
of their data~(14) with one additional arbitrary real parameter
representing a magnetic dipole moment.
As a starting point of the derivation procedure we take the
following axis data obtained in the paper~\cite{EMR1} for an
equatorially antisymmetric spacetime \cite{EMR2} with both
electric and magnetic dipole moments:
\begin{equation} e(z)=\frac{z^2-b_1z+b_2}{z^2+b_1z+b_2}, \quad
f(z)=\frac{c_2}{z^2+b_1z+b_2}, \label{ad6} \end{equation}
where $b_1$, $b_2$ and $c_2$ are arbitrary complex constants.
Mention that this data rewritten in an equivalent representation
was used in the paper \cite{EMR1} for constructing the
corresponding Ernst potentials \cite{Ern} ${\mathcal{E}}$ and $\Phi$ of a
6-parameter EMR solution. If nevertheless one opts to work
directly with the above (\ref{ad6}), then the asymptotic flatness
of the solution implies immediately that $b_1$ is a real constant
related to the total mass $2M$ of the binary configuration as
$b_1=2M$. Choosing then the constant $c_2$ in the form
$c_2=2(q+ib)$, the real parameters $q$ and $b$ being associated,
respectively, with the electric and magnetic dipole moments, and
also formally setting $b_2=c-i\delta$, we arrive at the
5-parameter axis data
\begin{equation} e(z)=\frac{e_-}{e_+}, \quad f(z)=\frac{2(q+ib)}{e_+}, \quad
e_\mp=z^2\mp 2Mz+c-i\delta, \label{ad5} \end{equation}
in which the real constants $c$ and $\delta$ should yet be related
to some physical or geometrical characteristics.
Recall now that the extended multi-soliton solutions involve the
constants $\a_n$ which satisfy the algebraic equation \cite{Sib}
\begin{equation} e(z)+\bar e(z)+2f(z)\bar f(z)=0 \label{sib} \end{equation}
(the bar over a symbol means complex conjugation), and in the
equatorially antisymmetric case these can be chosen in the form
\begin{equation} \a_1=-\a_4=\frac{1}{2}R+\sigma, \quad \a_2=-\a_3=\frac{1}{2}R-\sigma,
\label{alf} \end{equation}
where $R$ is a real constant representing the coordinate
separation of the sources, and the parameter $\sigma$ can take
non-negative real or pure imaginary values (see Fig.~1). Then
instead of the constants $c$ and $\delta$ from the axis data
(\ref{ad5}) one is able to introduce the new parameters $R$ and
$\sigma$ by equating coefficients at the same powers of $z$ on the two
sides of the equation
\begin{equation} \frac{e_-}{e_+}+\frac{\bar e_-}{\bar
e_+}+\frac{4(q^2+b^2)}{e_+\bar e_+}=
\frac{2\prod_{n=1}^4(z-\a_n)}{e_+\bar e_+}, \label{eq} \end{equation}
thus yielding
\begin{equation} c=2M^2-\frac{1}{4}R^2-\sigma^2, \quad
\delta=\sqrt{(R^2-4M^2)(M^2-\sigma^2)-4(q^2+b^2)}. \label{cd} \end{equation}
Accounting for (\ref{cd}), the 5-parameter axis data (\ref{ad5})
finally assumes the form
\begin{equation} e(z)=\frac{e_-}{e_+}, \quad f(z)=\frac{2(q+ib)}{e_+}, \quad
e_\mp=z^2\mp 2Mz+2M^2-\frac{1}{4}R^2-\sigma^2-i\delta, \label{ad5n}
\end{equation}
and, by setting $b=0$ in (\ref{cd}) and (\ref{ad5n}), one recovers
the axis data (14) of \cite{CLL} (and consequently the quantities
$\beta_{1,2}$ and $f_{1,2}$ in (11) of \cite{CLL} via the simple
fraction decomposition of $e(z)$ and $f(z)$). It is worth noting
that this procedure of changing parameters in the axis data was
already described in application to the case of identical
counter-rotating uncharged black holes \cite{MRRS} and, moreover,
has been recently used for obtaining a physical parametrization of
the BM solution \cite{MRR}. Furthermore, by virtue of the
equatorial antisymmetry, the axis condition for the solution
defined by the axis data (\ref{ad5n}) is satisfied automatically,
and therefore there is no need to solve any additional algebraic
equations for the metric functions.
As it is straightforward to elaborate by purely algebraic
computing the explicit form of the Ernst potentials defined by the
axis data (\ref{ad5n}), as well as the form of the corresponding
metric functions $f$, $\gamma$ and $\omega$ entering the
stationary axisymmetric line element
\begin{equation} d s^2=f^{-1}[e^{2\gamma}(d\rho^2+d z^2)+\rho^2d\varphi^2]-f(d
t-\omega d\varphi)^2, \label{pap} \end{equation}
below we will restrict ourselves to only writing out the final
expressions which reproduce and generalize the analogous formulas
of the paper \cite{CLL}. Then for ${\mathcal{E}}$ and $\Phi$ we have
\begin{eqnarray} {\mathcal{E}}&=&\frac{A-B}{A+B}, \quad \Phi=\frac{C}{A+B}, \nonumber\\
A&=&R^2[M^2(R^2-4\sigma^2)-4(q^2+b^2)](R_+-R_-)(r_+-r_-)+
4\sigma^2[M^2(R^2-4\sigma^2) \nonumber\\
&&+4(q^2+b^2)](R_+-r_+)(R_--r_-)+2R\sigma(R^2-4\sigma^2)[R\sigma(R_+r_-+R_-r_+)
\nonumber\\ &&+i\delta(R_+r_--R_-r_+)], \nonumber\\
B&=&2MR\sigma(R^2-4\sigma^2)[R\sigma(R_++R_-+r_++r_-)-(2M^2-i\delta)(R_+-R_--r_++r_-)],
\nonumber\\
C&=&4(q+ib)R\sigma[(R+2\sigma)(R\sigma-2M^2-i\delta)(r_+-R_-)
+(R-2\sigma)(R\sigma+2M^2+i\delta) \nonumber\\ &&\times(r_--R_+)],
\label{EF} \end{eqnarray}
where
\begin{equation} R_\pm=\sqrt{\rho^2+(z+{\textstyle\frac{1}{2}}R\pm\sigma)^2},
\quad r_\pm=\sqrt{\rho^2+(z-{\textstyle\frac{1}{2}}R\pm\sigma)^2},
\label{Rpm} \end{equation}
while the metric functions are given by the expressions
\begin{eqnarray} f&=&\frac{A\bar A-B\bar B+C\bar C}{(A+B)(\bar A+\bar B)},
\quad e^{2\gamma}=\frac{A\bar A-B\bar B+C\bar C}
{16R^4\sigma^4(R^2-4\sigma^2)^2R_+R_-r_+r_-}, \nonumber\\
\omega&=&-\frac{{\rm Im}[2G(\bar A+\bar B)+C\bar I]}
{A\bar A-B\bar B+C\bar C}, \nonumber \\
G&=&-zB+R\sigma\{2R[M^2(R^2-4\sigma^2)-2(q^2+b^2)](R_-r_--R_+r_+) \nonumber\\
&&+4\sigma[M^2(R^2-4\sigma^2)+2(q^2+b^2)](r_+r_--R_+R_-)
\nonumber\\
&&+M(R+2\sigma)[(R-2\sigma)^2(R\sigma+2M^2-i\delta)-8(q^2+b^2)](R_--r_+)
\nonumber\\
&&+M(R-2\sigma)[(R+2\sigma)^2(R\sigma-2M^2+i\delta)+8(q^2+b^2)](R_+-r_-)\},
\nonumber\\
I&=&-zC+4M(q+ib)[R^2(2M^2-2\sigma^2+i\delta)(R_+r_++R_-r_-)
\nonumber\\
&&+2\sigma^2(R^2-4M^2-2i\delta)(R_+R_-+r_+r_-)]-2(q+ib)(R^2-4\sigma^2)
\nonumber\\ &&\times\{2M[(R\sigma+2M^2+i\delta)R_+r_-
-(R\sigma-2M^2-i\delta)R_-r_+] +R\sigma \nonumber\\
&&\times[(R\sigma+6M^2+i\delta)(R_++r_-)+ (R\sigma-6M^2-i\delta)(R_-+r_+)
+8MR\sigma]\}. \label{mf} \end{eqnarray}
The $t$ and $\varphi$ components of the electromagnetic
4-potential are defined by the formulas
\begin{equation} A_t=-{\rm Re}\left(\frac{C}{A+B}\right), \quad A_\varphi={\rm
Im}\left(\frac{I}{A+B}\right), \label{tph} \end{equation}
and these complete the general mathematical description of the
5-parameter EMR solution in $\sigma$-parametrization.
At this point, several remarks on the formulas
(\ref{EF})-(\ref{tph}) might be appropriate. First, the above
representation of the 5-parameter EMR solution is fully equivalent
to the known description of that solution worked out in the papers
\cite{EMR1,SRo} (with the NUT parameter $\nu$ set equal to zero).
Second, it is highly important to underline that the arbitrary
parameter $\sigma$ of the solution is not restricted exclusively to
real values (contrary to what was assumed in \cite{CLL}) but can
also take pure imaginary values determining the hyperextreme part
of the solution. The significance of this point will be fully
understandable later on when we express $\sigma$ in terms of the Komar
quantities. Third, the $\sigma$- representation of the EMR solution
should be only considered as an intermediate parametrization that
could be suitable for elaborating the final physical
representation in which $\sigma$ must be replaced by a rotation
parameter.
To gain a better insight into the structure of the EMR solution,
let us consider its first four Beig-Simon multipole moments
\cite{BSi,SBe,Sim} which can be found with the aid of the
Hoenselaers-Perj\'es procedure \cite{HPe,SAp}:
\begin{eqnarray} M_0&=&2M, \quad M_1=0, \quad
M_2=\frac{1}{2}M(R^2-8M^2+4\sigma^2), \quad M_3=0, \nonumber \\
J_0&=&J_1=0, \quad J_2=2M\delta, \quad J_3=0, \nonumber \\
Q_0&=&0, \quad Q_1=2q, \quad Q_2=0, \quad
Q_3=\frac{1}{2}q(R^2-8M^2+4\sigma^2) -2b\delta, \nonumber \\ B_0&=&0,
\quad B_1=2b, \quad B_2=0, \quad B_3=\frac{1}{2}b(R^2-8M^2+4\sigma^2)
+2q\delta \label{mm} \end{eqnarray}
($M_i$, $J_i$, $Q_i$ and $B_i$ define, respectively, the mass,
angular momentum, electric and magnetic multipole moments), whence
it follows the asymptotic flatness of the solution ($J_0=0$), the
total mass $M_0$ of the configuration being equal to $2M$, and
total angular momentum $J_1$ being zero due to counterrotation. In
the absence of net charges the parameters $q$ and $b$ define the
electric and magnetic dipole moments, respectively, which means
that the two sources in the EMR solution are endowed with opposite
electric and magnetic charges.
It is clear from the above form of the multipole moments that the
special $b=0$ case of the EMR solution considered in \cite{CLL} is
characterized by zero total magnetic dipole moment, and this fact
explains the intrinsically non-physical character of that
particular solution. Indeed, the magnetic dipole moment $2b$ of
the 5-parameter EMR solution is a result of the following two
non-zero contributions -- one coming from the rotating electric
charges and the other originated by the opposite magnetic charges.
The electric contribution is twice the magnetic dipole moment
created by one rotating electric charge, so by demanding $b=0$,
Cabrera-Munguia et al introduced in \cite{CLL} a specific
non-vanishing magnetic dipole moment due to magnetic charges,
antiparallel to that created by electric charges. Those authors
apparently confused the case of counter-rotating opposite charges
and the BM configuration in which the counter-rotating charges
have the same signs and hence the total magnetic and electric
dipole moments are both equal to zero intrinsically. Therefore, a
physically meaningful dihole solution arising from the 5-parameter
EMR configuration must have zero individual magnetic charges and,
at the same time, a non-zero magnetic dipole moment generated by
counterrotation of opposite electric charges.
The individual magnetic charges in the 5-parameter EMR solution
can be eliminated by means of the condition \cite{Tom,MMR}
\begin{equation} A_t(\rho=0,z=\a_1)-A_t(\rho=0,z=\a_2)=0, \label{cond} \end{equation}
which can be easily solved for $b$. Then from (\ref{EF}),
(\ref{Rpm}) and (\ref{tph}) we get
\begin{equation}
b^2=\frac{4q^2[(R^2-4M^2)(M^2-\sigma^2)-4q^2]}{(R^2-4M^2)^2+16q^2},
\label{b} \end{equation}
and this condition, together with the formulas
(\ref{EF})-(\ref{tph}) with real $\sigma$, provide one with a
$\sigma$-representation of the physically meaningful 4-parameter model
for a stationary black dihole whose constituents are
counter-rotating. It can be verified by a direct calculation that
each black-hole constituent of such a model verifies the
well-known Smarr mass formula identically.
\section{The 4-parameter dihole solution in physical\\ parametrization}
As was already mentioned, the $\sigma$-representation of the solution
is only an intermediate step on the way of obtaining the physical
parametrization in terms of the Komar quantities. Once the
$\sigma$-representation is known, our further actions are the
following: we must first try to express the parameter $q$ in terms
of the individual Komar charge $Q$ of any of the dihole
constituents, thus rewriting the solution in the parameters $M$,
$R$, $Q$ and $\sigma$, and then find the form of $\sigma$ in terms of $M$,
$R$, $Q$ and $J$, $J$ being the individual Komar angular momentum,
from Smarr's mass formula, by considering the latter an algebraic
equation for $\sigma$.
The mass formula for black holes discovered by Smarr \cite{Sma}
relates the mass $M$, angular momentum $J$ and charge $Q$ of a
black hole to several quantities evaluated on the horizon: the
surface gravity $\kappa$, horizon's area $S$ and angular velocity
$\Omega^H$, and the electric potential $\Phi^H$. The formula reads
\begin{equation} M=\frac{1}{4\pi}\kappa S+2J\Omega^H+Q\Phi^H
=\sigma+2J\Omega^H+Q\Phi^H, \label{Sma} \end{equation}
the Komar quantities $M$, $J$ and $Q$ being defined by the
integrals \cite{Tom}
\begin{eqnarray} M&=&-\frac{1}{8\pi}\int_{H}
\omega\Omega_{,z} d\varphi d z, \label{kq1}\\
J&=&\frac{1}{8\pi}\int_{H}\omega
\left[-1-{\textstyle\frac12}\omega\Omega_{,z}+\tilde
A_\varphi A'_{\varphi,z}+(A_\varphi A'_\varphi)_{,z}\right] d\varphi d z, \label{kq2}\\
Q&=&\frac{1}{4\pi}\int_{H} \omega A'_{\varphi,z} d\varphi d z,
\label{kq3} \end{eqnarray}
with $\Omega={\rm Im}({\mathcal{E}})$, $A'_\varphi={\rm Im}(\Phi)$, $\tilde
A_\varphi=A_\varphi+\omega A_t$ (note that the metric functions
$\omega$ and $\gamma$, as well as the potential $\tilde
A_\varphi$, take constant values on the horizon), while the form
of the constants $\kappa$, $S$, $\Omega^H$ and $\Phi^H$ is given
by the formulas \cite{Tom,Car}
\begin{equation} \kappa=\sqrt{-\omega^{-2}e^{-2\gamma}}, \quad
S=4\pi\sigma\sqrt{-\omega^{2}e^{2\gamma}}, \quad
\Omega^H=\omega^{-1}, \quad \Phi^H=-A_t-\Omega^H A_\varphi.
\label{kap} \end{equation}
For our dihole solution, the calculation of the individual charge
$Q$ of the upper black-hole constituent, whose horizon is
represented by the null hypersurface $\rho=0$,
${\textstyle\frac12}R-\sigma\le z\le {\textstyle\frac12}R+\sigma$, leads
to a cubic equation for $q$ which has to be solved in order to
pass from the latter $q$ to the Komar $Q$ in the formulas
determining the solution. It is remarkable, however, that the need
to solve a cubic equation can be circumvented by an appropriate
change of the parameter $q$. Thus, by introducing a new parameter
${\sf{q}}$ via the relation
\begin{equation} {\sf{q}}^2=q^2+b^2, \label{qnew} \end{equation}
which has some analogy with a duality rotation of the
electromagnetic Ernst potential, we find from (\ref{b}) and
(\ref{qnew}) the form of $q$ and $b$ in terms of ${\sf{q}}$:
\begin{eqnarray} q&=&{\sf{q}}(R^2-4M^2)/\tau, \quad b=2{\sf{q}}
\delta'/\tau, \nonumber\\
\delta'&=&\sqrt{(R^2-4M^2)(R^2-\sigma^2)-4{\sf{q}}^2}, \quad
\tau=\sqrt{(R^2-4M^2)(R^2-4\sigma^2)-16{\sf{q}}^2}, \label{qb} \end{eqnarray}
and this redefinition of the parameter $q$ permits us to obtain
from (\ref{kq3}) a simple expression for the Komar charge $Q$ in
terms of ${\sf{q}}$:
\begin{equation} Q=-2{\sf{q}}(R+2M)/\tau, \label{QK} \end{equation}
whence we readily get the inverse dependence of ${\sf{q}}$ on $Q$:
\begin{equation}
{\sf{q}}=-\frac{Q\sqrt{(R^2-4M^2)(R^2-4\sigma^2)}}{2\sqrt{(R+2M)^2+4Q^2}}.
\label{qQ} \end{equation}
The above formula for ${\sf{q}}$ permits us, by rewriting the dihole
solution in terms of the parameters $M$, $R$, $Q$ and $\sigma$, to
obtain the quantities $\Omega^H$ and $\Phi^H$ that we need for
finding $\sigma$:
\begin{eqnarray} &&\Omega^H=\frac{\sqrt{(R-2M)[(R+2M)^2+4Q^2]
[(R+2M)(M^2-\sigma^2)-Q^2(R-2M)]}}{(R+2M)[2M(R+2M)(M+\sigma)-Q^2(R-4M-2\sigma)]},
\nonumber\\ &&\Phi^H=\frac{Q(R-2M)[(R+2M)(M+\sigma)+2Q^2]
}{(R+2M)[2M(R+2M)(M+\sigma)-Q^2(R-4M-2\sigma)]}. \label{OF} \end{eqnarray}
Finally, after the substitution of (\ref{OF}) into the Smarr
formula (\ref{Sma}) in which we can put $J=Ma$, $a$ being the
angular momentum per unit mass of the upper black hole, we obtain
by a simple calculation the desired expression for $\sigma$ in terms
of the physical quantities $M$, $a$, $Q$ and $R$:
\begin{equation} \sigma=\sqrt{M^2-\left(\frac{M^2a^2[(R+2M)^2+4Q^2]}
{[M(R+2M)+Q^2]^2}+Q^2\right)\frac{R-2M}{R+2M}}. \label{sig} \end{equation}
This formula for $\sigma$ is the central result of our paper. Now the
dihole solution can be rewritten in the physical parameters, its
Ernst potentials ${\mathcal{E}}$ and $\Phi$ assuming the form
\begin{eqnarray} {\mathcal{E}}&=&\frac{A-B}{A+B}, \quad \Phi=\frac{C}{A+B}, \nonumber\\
A&=&R^2(M^2-Q^2\nu)(R_+-R_-)(r_+-r_-)+
4\sigma^2(M^2+Q^2\nu)(R_+-r_+)(R_--r_-) \nonumber\\
&&+2R\sigma[R\sigma(R_+r_-+R_-r_+)+iMa\mu(R_+r_--R_-r_+)], \nonumber\\
B&=&2MR\sigma[R\sigma(R_++R_-+r_++r_-)-(2M^2-iMa\mu)(R_+-R_--r_++r_-)],
\nonumber\\
C&=&2C_0R\sigma[(R+2\sigma)(R\sigma-2M^2-iMa\mu)(r_+-R_-) +(R-2\sigma) \nonumber\\
&&\times(R\sigma+2M^2+iMa\mu)(r_--R_+)], \label{EFn} \end{eqnarray}
where the dimensionless quantities $\mu$, $\nu$ and $C_0$ are
defined as
\begin{equation} \mu=\frac{R^2-4M^2}{M(R+2M)+Q^2}, \quad
\nu=\frac{R^2-4M^2}{(R+2M)^2+4Q^2}, \quad
C_0=-\frac{Q(R^2-4M^2+2iMa\mu)}{(R+2M)(R^2-4\sigma^2)}, \label{mnc}
\end{equation}
and the final form of the metric coefficients $f$, $\gamma$ and
$\omega$ is the following:
\begin{eqnarray} f&=&\frac{A\bar A-B\bar B+C\bar C}{(A+B)(\bar A+\bar B)},
\quad e^{2\gamma}=\frac{A\bar A-B\bar B+C\bar C}
{16R^4\sigma^4R_+R_-r_+r_-}, \quad
\omega=-\frac{{\rm Im}[2G(\bar A+\bar B)+C\bar I]}
{A\bar A-B\bar B+C\bar C}, \nonumber \\
G&=&-zB+R\sigma\{R(2M^2-Q^2\nu)(R_-r_--R_+r_+)
+2\sigma(2M^2+Q^2\nu)(r_+r_--R_+R_-)
\nonumber\\
&&+M[(R+2\sigma)(R\sigma-2M^2+iMa\mu)+2(R-2\sigma)Q^2\nu](R_+-r_-)
\nonumber\\
&&+M[(R-2\sigma)(R\sigma+2M^2-iMa\mu)-2(R+2\sigma)Q^2\nu](R_--r_+)\},
\nonumber\\
I&=&-zC+2C_0M[R^2(2M^2-2\sigma^2+iMa\mu)(R_+r_++R_-r_-)
\nonumber\\
&&+2\sigma^2(R^2-4M^2-2iMa\mu)(R_+R_-+r_+r_-)]-C_0(R^2-4\sigma^2) \nonumber\\
&&\times\{2M[R\sigma(R_+r_--R_-r_+)+(2M^2+iMa\mu)(R_+r_-+R_-r_+)] +R\sigma[R\sigma \nonumber\\
&&\times(R_++R_-+r_++r_-)+(6M^2+iMa\mu)(R_+-R_--r_++r_-)
+8MR\sigma]\}. \label{mfn} \end{eqnarray}
It should be noted that $\sigma$ in the above (\ref{EFn})-(\ref{mfn})
is no longer an independent parameter, having conceded that role
to the constant $a$. From (\ref{sig}) it follows that $\sigma$,
depending on interrelations between the parameters $M$, $a$, $Q$
and $R$, can automatically take on (non-negative) real or pure
imaginary values, thus describing not only the binary
configurations of black holes but also of hyperextreme objects.
That is why $\sigma$'s taking pure imaginary values (along with the
real ones) in the EMR solution (\ref{EF}) is highly important for
the mathematical equivalence of the parameter sets ($M$, $Q$, $R$,
$\sigma$) and ($M$, $Q$, $R$, $a$), and consequently for the
correctness of the entire reparametrization procedure.
\section{The limits and physical properties of dihole solution}
The main limits of the dihole solution can be well seen from the
formula (\ref{sig}) for $\sigma$. Thus, in the absence of rotation
($a=0$) the solution reduces to the Emparan-Teo electrostatic
non-extreme dihole spacetime \cite{ETe} whose physical form was
found in the paper \cite{CGM}. In the pure vacuum limit ($Q=0$)
the solution represents a vacuum specialization of the BM
equatorially antisymmetric binary configuration whose physical
parametrization was elaborated in the paper \cite{MRRS} on the
basis of Varzugin's expression for the quantity $\sigma$ \cite{Var}.
When $R\to\infty$ (no interaction between the dihole
constituents), one gets from (\ref{sig}) $\sigma=(M^2-a^2-Q^2)^{1/2}$,
which is characteristic of a single KN black hole.
By construction, the upper KN constituent has mass $M$, angular
momentum $Ma$ and charge $Q$, whereas the analogous
characteristics of the lower constituent are $M$, $-Ma$ and $-Q$,
respectively. The strut separating the two constituents provides
us with the information about the interaction force \cite{Isr},
the latter being defined by the expression
\begin{eqnarray} {\mathcal
F}=\frac14(e^{-\gamma_0}-1)&=&\frac{M^2(R+2M)^2+Q^2R^2}
{(R+2M)^2(R^2-4M^2)} \nonumber\\
&=&\frac{1}{R^2-4M^2}\left(M^2+Q^2-
\frac{4MQ^2(R+M)}{(R+2M)^2}\right), \label{F} \end{eqnarray}
where $\gamma_0$ is the value of the metric function $\gamma$ on
the strut, and one can see that ${\mathcal F}$ cannot take zero
value at any finite separation $R$ of the constituents, so that
the strut is irremovable generically. It is worth mentioning that
the formula (\ref{F}) differs from the analogous expression
obtained in \cite{CLL}, as our ${\mathcal F}$ does not contain a
non-vanishing unphysical contribution due to magnetic charges.
Turning now to the thermodynamical quantities $\kappa$, $S$,
$\Omega^H$ and $\Phi^H$ entering the Smarr mass formula
(\ref{Sma}), it may be observed that these must be calculated only
for the upper black-hole constituent because the analogous set for
the lower constituent is just $\kappa$, $S$, $-\Omega^H$ and
$-\Phi^H$. Then for the upper constituent we get
\begin{eqnarray} \kappa&=&\frac{R\sigma[(R+2M)^2+4Q^2]}
{(R+2M)^2[2(M+\sigma)(MR+2M^2+Q^2)-Q^2(R-2M)]}, \label{kS1}\\
S&=&4\pi\left(1+\frac{2M}{R}\right) \left(2M(M+\sigma)-
\frac{Q^2(R-2M)(R-2\sigma)}{(R+2M)^2+4Q^2}\right) \nonumber \\
&=&\frac{4\pi}{R(R+2\sigma)}\left((R+2M)^2(M+\sigma)^2
+\frac{M^2a^2(R^2-4M^2)^2}{(MR+2M^2+Q^2)^2}\right), \label{kS2}\\
\Omega^H&=&\frac{Ma[2(M-\sigma)(MR+2M^2+Q^2)-Q^2(R-2M)]}
{(4M^2a^2+Q^4)(MR+2M^2+Q^2)}, \label{kS3}\\
\Phi^H&=&\frac{Q[Q^2(M-\sigma)(MR+2M^2+Q^2)+2M^2a^2(R-2M)]}
{(4M^2a^2+Q^4)(MR+2M^2+Q^2)}, \label{kS4} \end{eqnarray}
where $S$ is given in two different forms suitable for recovering
the known limiting cases. The substitution of
(\ref{kS1})-(\ref{kS4}) into (\ref{Sma}) shows that the above
expressions satisfy identically Smarr's formula for black holes.
It has been recently shown \cite{MRR} that the equally charged
black-hole constituents of the BM configuration saturate the
Gabach-Clement inequality for black holes with struts \cite{Gab}
which reads
\begin{equation} \sqrt{1+4{\mathcal F}}\ge\frac{\sqrt{(8\pi J)^2+(4\pi
Q^2)^2}}{S}. \label{GC} \end{equation}
In this respect it would be interesting to clarify whether the
oppositely charged constituents of our dihole model saturate the
inequality (\ref{GC}) too. The saturation means that the extremal
dihole constituents must satisfy (\ref{GC}) with the equality
sign. The extremality condition $\sigma=0$ yields from (\ref{sig}) the
value of $a$ at which the black-hole degeneration occurs:
\begin{equation} a^2=\frac{(MR+2M^2+Q^2)^2[M^2(R+2M)-Q^2(R-2M)]}
{M^2(R-2M)[(R+2M)^2+4Q^2]}, \label{ext} \end{equation}
and by substituting this $a$ into (\ref{kS2}) and (\ref{GC}) we
get ($J=Ma$)
\begin{eqnarray} &&S=\frac{4\pi(R+2M)^2[2M^2(R+2M)-Q^2(R-4M)]}
{R[(R+2M)^2+4Q^2]}, \nonumber\\ &&(8\pi J)^2+(4\pi Q^2)^2
=\frac{16\pi^2(R+2M)[2M^2(R+2M)-Q^2(R-4M)]^2}
{(R-2M)[(R+2M)^2+4Q^2]}. \label{SI} \end{eqnarray}
Taking into account that ${\mathcal F}$ does not depend explicitly
on $a$, it is easy to check that (\ref{F}) and (\ref{SI}) verify
the equality in (\ref{GC}) identically. Therefore, independently
of whether the KN black holes in a binary system have equal or
opposite charges, the interaction force between them is governed
by the Gabach-Clement inequality. Mention here one more common
feature shared by the BM and dihole configurations -- the extreme
limit is achieved in both of them at a larger absolute value of
$a$ (for some given $M$ and $Q$) than in the case of a single KN
black hole whose extremality condition is simply $a^2=M^2-Q^2$.
Mention that in the paper \cite{CLL} the Ernst potentials defining
the extreme limit are given with errors. Therefore, we find it
useful to give below the expressions for these potentials and
corresponding metric functions of the entire 5-parameter EMR
solution in the extreme limit $\sigma\to 0$:
\begin{eqnarray} {\mathcal{E}}&=&\frac{A-B}{A+B}, \quad \Phi=\frac{C}{A+B}, \nonumber \\
f&=&\frac{N}{D}, \quad e^{2\gamma}=\frac{N} {\a^8(x^2-y^2)^4},
\quad \omega=-\frac{4\a^2\delta y(x^2-1)(1-y^2)W} {N}, \nonumber\\
A&=&M^2\a^2(x^4-1)+\a^2(\a^2-M^2)(x^2-y^2)^2+(q^2+b^2)(1-y^4)
\nonumber\\ &&+2i\a^2\delta(x^2-2x^2y^2+y^2), \nonumber\\ B&=&2M\a
x[\a^2(x^2-y^2)-(M^2-i\delta)(1-y^2)], \nonumber\\ C&=&-2(q+ib)y
[\a^2(x^2-y^2)-(M^2+i\delta)(1-y^2)], \nonumber\\
I&=&-\a xyC-2(q+ib)(M+\a x)(1-y^2) [(M+\a x)^2+(M^2-\a^2)y^2
+i\delta(1+y^2)], \nonumber\\ N&=&[M^2\a^2(x^2-1)^2+\a^2
(\a^2-M^2)(x^2-y^2)^2-(q^2+b^2)(1-y^2)^2]^2 \nonumber\\
&&-16\a^4\delta^2x^2y^2(x^2-1)(1-y^2), \nonumber\\
D&=&\{M^2\a^2(x^4-1)+\a^2(\a^2-M^2)(x^2-y^2)^2
+(q^2+b^2)(1-y^4)+2M\a x \nonumber\\
&&\times[\a^2(x^2-y^2)-M^2(1-y^2)]\}^2
+4\a^2\delta^2[\a(x^2-2x^2y^2+y^2)+Mx(1-y^2)]^2, \nonumber\\
W&=&M\a^2[(\a^2-M^2)(x^2-y^2)(3x^2+y^2) +M^2(3x^4+6x^2-1)+8M\a
x^3] \nonumber\\ &&-(q^2+b^2)[4\a xy^2-M(1-y^2)^2], \nonumber\\
\delta&=&\sqrt{M^2(\a^2-M^2)-q^2-b^2}, \quad \a=\frac{1}{2}R,
\label{extreme} \end{eqnarray}
where the prolate spheroidal coordinates ($x$, $y$) are related to
the cylindrical coordinates ($\rho$,~$z$) by the formulas
\begin{equation} x=\frac{1}{2\a}(r_++r_-), \quad y=\frac{1}{2\a}(r_+-r_-),
\quad r_\pm=\sqrt{\rho^2+(z\pm\a)^2}, \label{xy} \end{equation}
and where we have also given the explicit extremal form of the
function $I$ defining the magnetic potential $A_\varphi$ via
formula (\ref{tph}).
As it follows from (\ref{mm}), the magnetic field in our dihole
solution differs considerably from that of the particular $b=0$
specialization of the EMR solution considered in \cite{CLL}: in
the former case it has a dipole character, while in the latter
case it behaves itself like a magnetic octupole ($B_3=2q\delta$).
In Figs.~2 and 3 this difference is illustrated by the plots of
magnetic lines of force for two characteristic particular cases.
\section{Conclusions}
In our paper we succeeded in elaborating a physically consistent
4-parameter model for stationary diholes which generalizes the
known dihole electrostatic solution earlier obtained by Emparan
and Teo. The model is comprised of two identical (up to the sign
of charges) counter-rotating KN black holes supported from falling
onto each other by a massless strut, and its finding and correct
mathematical description has turned out more sophisticated than in
the case of counter-rotating equally charged KN black holes
represented by the BM solution because the knowledge of a more
general 5-parameter EMR solution was needed for getting rid of the
specific individual magnetic charges initially present in the
dihole components. The solution's physical representation was
advantageous for a direct check that the binary configuration it
describes really saturates the Gabach-Clement inequality for
interacting black holes.
Since the aforementioned inequality also takes into account the
possibility for the black holes to carry magnetic charges, we
would like to mention that our dihole solution can be very easily
generalized to the case when the two KN constituents, besides the
opposite electric charges, would have arbitrary opposite magnetic
charges too, thus representing a pair of dyons \cite{Sch}. To
introduce an arbitrary magnetic charge ${\mathcal B}$ into our
dihole model, one only needs to make the following substitutions
in the formulas (\ref{sig})-(\ref{mfn}): change $Q$ to ${\mathcal
Q}=Q-i{\mathcal B}$, and $Q^2$ to $|{\mathcal Q}|^2=Q^2+{\mathcal
B}^2$ in all the occurrences. For instance, our expression
(\ref{sig}) for $\sigma$ will then assume the form
\begin{equation} \sigma=\sqrt{M^2-\left(\frac{M^2a^2[(R+2M)^2+4|{\mathcal Q}|^2]}
{[M(R+2M)+|{\mathcal Q}|^2]^2}+|{\mathcal
Q}|^2\right)\frac{R-2M}{R+2M}}. \label{sigB} \end{equation}
We underline that the parameter ${\mathcal B}$ thus introduced
will be a genuine individual magnetic charge of the upper black
hole, and this can be readily verified by means of the formula
\begin{equation} {\mathcal B}=\frac{1}{4\pi}\int_{H} \omega A_{t,z} d\varphi d
z. \label{mag} \end{equation}
It is easy to see that the dyonic dihole model, which is of course
equivalent to the 5-parameter EMR solution, will also saturate the
Gabach-Clement inequality because the electric and magnetic
charges $Q_i$ and ${\mathcal B}_i$ enter that inequality only in
the combination $Q_i^2+{\mathcal B}_i^2$ \cite{Gab}. Mention also
that the introduction of the magnetic charge does not actually
modify the Smarr mass formula (\ref{Sma}), provided the
substitutions described above are carried out properly in the term
$Q\Phi^H$.
\section*{Acknowledgments}
V.S.M. is grateful to the Relativity group of School of
Mathematical Sciences, Queen Mary, University of London, with
special thanks to Juan Valiente-Kroon, for their kind hospitality
and for providing excellent working conditions during his visit,
when part of this work was done. The research was partially
supported by CONACyT of Mexico.
|
\section{Introduction}
For an integer $n$ put $P(n)$ for the maximum prime factor of $n$ with the convention that $P(0)=P(\pm 1)=1$. A lot of work has been done understanding
the distribution of $P(p-1)$ for prime numbers $p$. The extreme cases $P(p-1)=2$ and $P(p-1)=(p-1)/2$ correspond to Fermat primes and Sophie-Germain primes, respectively.
Not only do we not know if there are infinitely many primes of these kinds, but we do not know whether for each $\varepsilon>0$ arbitrarily small there exist infinitely many primes
$p$ with $P(p-1)<p^{\varepsilon}$ or $P(p-1)>p^{1-\varepsilon}$.
For a set ${\mathcal C}$ of positive integers and a positive real number $x$ we put ${\mathcal C}(x)={\mathcal C}\cap [1,x]$. Let
$$
{\mathcal P}_\varepsilon := \{p \textrm{ prime : } P(p-1) \geq p^{1-\varepsilon}\}, \quad \kappa(\varepsilon) = \liminf_{x \to \infty} \frac{\#{\mathcal P}_\varepsilon(x)}{\pi(x)}.
$$
Goldfeld proved in \cite{Gold} that $\kappa(1/2) \geq 1/2 $. It is not known whether ${\mathcal P}_{{1}/{2} }$ has a relative density, nor what this density could be in case it exists. Fouvry \cite{Fou85}, showed that there exists $\varepsilon_0\in (0,1/3)$ such that $\kappa({\varepsilon_0})>0$. Baker and Harman \cite{HaBa96}, found $0<\varepsilon_{1}<\varepsilon_0$ such that ${\mathcal P}_{\varepsilon_{1}}$ is infinite.
In this article, we generalize Goldfeld's result in two different directions. First, we estimate from below the lower density of ${\mathcal P}_{\varepsilon}$ for all $\varepsilon \in [1/2, 3/4]$. Secondly, we estimate the counting function of the set of square free positive integers having prime divisors that, when shifted, share a large common prime factor. Both questions are motivated by a technique used in \cite{BM} to bound from below the degree of the field of coefficients of newforms in terms of the level. A feature of the method in loc. cit. is that what is needed are values of $\varepsilon$ such that $\kappa(\varepsilon)$ is as large as possible. Since $\kappa(\varepsilon)$ is clearly an increasing function of $\varepsilon$, in contrast with the aforementioned works, which are focused in dealing with smaller and smaller values of $\varepsilon$, here we concentrate on the case where this parameter is bigger than $1/2$.
We obtain the following results.
\begin{theorem}
\label{primo}
Let $0 \leq \alpha \le {1}/{4}$. Let
$$N_\alpha= \{ p \textrm{ prime such that } P(p-1) \geq p^{{1}/{2} -\alpha} \}.$$
Then
$$\# N_\alpha(x) \geq \left(\frac{1}{2} +\alpha\right) \frac{x}{\log x} + E(x);\quad E(x)=\left\{\begin{matrix} {\displaystyle{O\left(\frac{x\log\log x}{(\log x)^2}\right)}} & (\alpha<1/4)\\
{\displaystyle{O\left(\frac{x}{(\log x)^{5/3}}\right)}} & (\alpha=1/4).\\
\end{matrix}\right.
$$
The implied constant depends on $\alpha$. In particular,
$$
\kappa(1/2+\alpha)\ge 1/2+\alpha\quad {\text{for~all}}\quad \alpha\in [0,1/4].
$$
\end{theorem}
The case $\alpha=0$ is Goldfeld's result mentioned above. Our proof of Theorem \ref{primo} follows closely his method.
For any $k\ge 1$ and $a\in (0,1/k)$, let
$$
{\mathcal A}_{k,a}=\{n=p_1\cdots p_k, P\big(\gcd(p_1-1,\ldots,p_k-1)\big)>n^{a}\}.
$$
By Goldfeld's result, $\#{\mathcal A}_{1,1/2}(x)\asymp x/\log x$. Here, we prove the following result.
\begin{theorem}\label{compuesto}
If $k\ge 2$ and $a\in \big[1/(2k),17/(32k)\big)$ are fixed, then
\begin{equation}
\label{eq:main}
\frac{x^{1-a(k-1)}}{(\log x)^{k+1}}\ll \#{\mathcal A}_{k,a}(x)\ll \frac{x^{1-a(k-1)}(\log\log x)^{k-1}}{(\log x)^2}.
\end{equation}
\end{theorem}
The case $a=1/(2k)$ is important for the results from \cite{BM}. We have the estimate
\begin{equation}\label{crudo}
\#{\mathcal A}_{k,1/(2k)}(x)=x^{1/2+1/2k+o(1)}, \quad x\to\infty.
\end{equation}
Goldfeld's method does not seem to extend to the situation in Theorem \ref{compuesto} (see the last section). Instead, we follow a more direct method. For the lower bound, we rely on a refined version of the Brun-Titchmarsh inquality due to Banks and Shparlinsky \cite{BS}.
We remark that both theorems presented here remain valid if, instead of considering large factors of $p-1$, we look at large factors $p+n$ for an arbitrary nonzero fixed integer $n$.
We leave as a problem for the reader to determine the exact order of magnitude of $\#{\mathcal A}_{k,a}(x)$, or an asymptotic for it.
Throughout this paper, we use $p,~q,~r$ with or without subscripts for primes. We use the Landau symbols $O,~o$ and the Vinogradov symbols $\ll$ and $\gg $ with their regular meaning.
The constants implied by them might depend on some other parameters such as $\alpha,~k,~\varepsilon$ which we will not indicate.
\section{Proof of Theorem \ref{primo}}
We follow Goldfeld's general strategy. Put $c={1}/{2}-\alpha$. So, ${1}/{4} \le c \leq {1}/{2}$. Let $$N_c'(x)=\#\{ p \leq x : p \textrm{ is prime and } P(p-1) \geq x^c \}.$$
Since $\#N_\alpha(x) \geq N_c'(x)$, it is enough to give a lower bound for $N_c'(x)$.
Put
$$
M_c(x)= \sum_{p \leq x} \sum_{\substack{\ell\mid p-1\\ \ell \geq x^c}} \log \ell,$$ where $p$ and $\ell$ denote primes. Since
$$
\sum_{\substack{\ell \mid p-1 \\ \ell \geq x^c}} \log \ell \left\{ \begin{array}{ll}
= 0, & \textrm{ if } P(p-1) < x^c; \\
\leq \log x, & \textrm{ otherwise }
\end{array} \right.$$
we have that
$$
M_c(x) \leq \log x \sum_{\substack{p \leq x \\ P(p-1) \geq x^c}} 1 = N_c'(x) \log x.
$$
Hence, $N_c'(x) \geq M_c(x)/\log x.$ Then, in order to prove Theorem \ref{primo}, it is enough to show that
\begin{equation}\label{primera reduccion}
M_c(x) = (1-c)x + F(x),\quad F(x)=\left\{\begin{matrix} {\displaystyle{O_c\left(\frac{x\log \log x}{\log x} \right)}}, & (c>1/4);\\
{\displaystyle{O\left(\frac{x}{(\log x)^{2/3}}\right)}}, & (c=1/4).\\
\end{matrix}\right.
\end{equation}
We denote by $\Lambda(\cdot)$ the von Mangoldt's function. As usual, $\pi(x;b,a)$ is the number of primes $q\le x$ in the arithmetic progression $q\equiv a\pmod b$.
We define
$$
L(x;u,v)=\sum_{u<m\le v} \Lambda(m)\pi(x;m,1).
$$
\begin{lemma}
\label{log}
Assume $1/4\le c\le 1/2$. Then
$$
L(x;x^c,x) = M_c(x) + O\left(\frac{x^{{7}/{6} -{2c}/{3}}}{(\log x)^r}\right),
$$
where $r=0$ when $c>1/4$ and $r=2/3$ when $c=1/4$.
\end{lemma}
\begin{proof}
Let $0<d<1-c$ be a real number and $r \in (0,1)$. We assume that $x$ is large enough so that the inequality $x^{1-d}(\log x)^{r}<x$ holds. We put
\begin{eqnarray*}
M_1^d(x) & = & \sum_{\substack{x^c < \ell^k \leq x^{1-d}(\log x)^{r} \\ \ell \textrm{ prime, } k\geq 2 }} \pi (x;\ell^k,1)\log \ell\\
M_2^d(x) & = & \sum_{\substack{x^{1-d}(\log x)^{r} < \ell^k \leq x \\ \ell \textrm{ prime, } k\geq 2 }} \pi (x;\ell^k,1)\log \ell.
\end{eqnarray*}
Hence,
\begin{equation}\label{sumas}
L(x;x^c,x)-M_c(x) = M_1^d(x) + M_2^d(x).
\end{equation}
Using the Brun-Titchmarsh inequality, we have that
\begin{eqnarray*}
M_1^d(x) & \ll & \frac{x}{\log x } \sum_{\substack{x^c < \ell^k \leq x^{1-d}(\log x)^{r} \\ \ell \textrm{ prime, } k\geq 2 }} \frac{ \log \ell}{\ell^{k-1}(\ell-1)} \\
& \le & \frac{x}{\log x} \sum_{\substack{\ell \leq x^{(1-d)/2}(\log x)^{r/2}\\ \ell \textrm{ prime}}} 2\log\ell \sum_{k\ge c\log x/\log \ell} \frac{1}{\ell^k}\\
& \le & \frac{x}{\log x} \sum_{\ell \leq x^{(1-d)/2}(\log x)^{r/2}} \frac{4\log x}{x^c}\\
& = & 4x^{1-c}\pi\left(x^{(1-d)/2}(\log x)^{r/2}\right)\\
& \ll & \frac{x^{1-c+(1-d)/2}}{(\log x)^{1-r/2}}.
\end{eqnarray*}
On the other hand, for an integer $m>x^{1-d}(\log x)^{\alpha}$, we have that
$$
\pi (x;m,1) < \sum_{\substack{n \leq x\\ n \equiv 1 \pmod m }} 1 \leq \frac{x}{m} < \frac{x^d}{(\log x)^{\alpha}}.
$$
Hence,
\begin{eqnarray*}
M_2^d(x) & < & \frac{x^d}{(\log x)^{\alpha}} \sum_{\substack{x^{1-d}(\log x)^{\alpha} < \ell^k \le x \\ \ell \textrm{ prime, } k\geq 2 }} \log \ell \\
& \ll & \frac{x^d}{(\log x)^{\alpha}} (\log x) \pi(\sqrt{x}) \ll \frac{x^{d + \frac{1}{2} }}{(\log x)^{\alpha}}.
\end{eqnarray*}
Using \eqref{sumas}, we obtain
$$
L_c(x)-M_c(x)= O\left( \frac{x^{1-c+(1-d)/2}}{(\log x)^{1-r/2}} + \frac{x^{d + \frac{1}{2}}}{(\log x)^{r}}\right).$$
We take $d=2/3(1-c)$ and then both exponents of $x$ above are equal and evaluate to $7/6-2/3 c$. Taking $r=0$ when $c<1/4$ and $r=2/3$ when $c=1/4$, we obtain the desired estimate.
\end{proof}
\begin{lemma}
\label{BT}
Assume that $c\in (0,1/2]$. Then, for $B>0$, we have
$$
L\left(x;{x^c}/{(\log x)^B}, x^c\right) =O\left(\frac{x \log \log x}{\log x} \right), \quad (x \rightarrow \infty).
$$
\end{lemma}
\begin{proof}
This follows immediately from the Brun-Titchmarsh inequality (see, for example, equation (3) in \cite{Gold}).
\end{proof}
\begin{lemma}\label{BV}
Assume that $c\in (0,1/2]$. Then, there exists $B>0$ such that
$$
L\left(x;1,{x^c}/{(\log x)^B}\right) = cx+O\left(\frac{x \log \log x}{\log x} \right), \quad (x \rightarrow \infty).
$$
\end{lemma}
\begin{proof} This follows easily from the Bombieri-Vinogradov theorem (see, for example, equation (2) in \cite{Gold}).
\end{proof}
\medskip
\noindent {\it Proof of Theorem \ref{primo}}: We have (see p. 23 in \cite{Gold}),
\begin{equation}\label{basico}
L(x;1,x) = x + O\left(\frac{x}{\log x}\right), \quad (x \rightarrow \infty).
\end{equation}
Take $B>0$ as in Lemma \ref{BV}.
Since
$$
L(x;1,x) =L\left(1,\frac{x^c}{(\log x)^B}\right) + L\left(\frac{x^c}{(\log x)^B}, x^c\right) + L(x;x^c,x),
$$
the result follows by combining \eqref{primera reduccion} and Lemmas \ref{log}, \ref{BT} and \ref{BV}.
\qed
\section{Proof of Theorem \ref{compuesto}}
\subsection{The upper bound}
Let $x$ be large. It is sufficient to prove the upper bound indicated at \eqref{eq:main} for the number of integers $n\in {\mathcal A}_{k,a}\cap [x/2,x]$, since then the upper bound will follow by changing $x$ to $x/2$, then to $x/4$ and so on, and summing up the resulting estimates.
So, we assume that $n\ge x/2$ is in ${\mathcal A}_{k,a}(x)$.
Then $n=p_1\cdots p_k\le x$, where $p_1\le p_2\le \cdots \le p_k$, and $p_i=p\lambda_i+1$ for $i=1,\ldots,k$, where
$$
p>n^{a}>(x/2)^{a}.
$$
Note that
$$
p^k \lambda_1\cdots \lambda_k\le \phi(n)<n<x.
$$
Thus, $p<x^{1/k}$. Let ${\mathcal B}_1(x)$ be the set of such $n\le x$ such that $\lambda_k\le x^{\delta}$, where $\delta=\delta_{k}=15(k-1)/(32 k^2)$. Since $\lambda_1\le \cdots \le \lambda_k$, we get that $\lambda_i\le x^{\delta}$ for all $i=1,\ldots,k$. This shows that
\begin{equation}
\label{eq:B2}
\#{\mathcal B}_1(x)\le \pi(x^{1/k}) (x^{\delta})^k<x^{1/k+15(k-1)/(32k)}=o(x^{1-a(k-1)})\quad (x\to\infty),
\end{equation}
where we used the fact that $1/k+15(k-1)/(32k)<1-a(k-1)$, which holds for all $k\ge 2$ and $a\in (0,17/(32k))$.
From now on, we assume that $n\in {\mathcal B}_2(x)=\left({\mathcal A}_k\cap [x/2,x]\right)\backslash {\mathcal B}_1(x)$. Fix the primes $p_1\le \cdots\le p_{k-1}$. Then $p$ is fixed, $p_k\le x/(p_1\ldots p_{k-1})$ and $p_k\equiv 1\pmod p$. The number of such primes is, by the Brun-Titchmarsch theorem (see \cite{MV}), at most
$$
\pi(x/(p_1\ldots p_{k-1}); p,1)\le \frac{2x}{(p-1) p_1\ldots p_{k-1} \log(x/(pp_1\ldots p_{k-1}))}.
$$
Since $x/(p p_1\ldots p_{k-1})>\lambda_k>x^{\delta}$, we get that the last bound is at most
$$
\ll \frac{x}{(\log x)pp_1\ldots p_{k-1}}.
$$
Keeping $p$ fixed and summing up the above bound over all ordered $k-1$-tuples of primes $(x/2)^{a}<p_1\le \cdots\le p_{k-1}\le x$ such that $p_i\equiv 1\pmod p$ for $i=1,\ldots,k-1$, we get a bound of
\begin{equation}
\label{eq:bound}
\frac{x}{(\log x)p}\left(\sum_{\substack{q\le x\\ q\equiv 1\pmod p}} \frac{1}{q}\right)^{k-1}\ll
\frac{x(\log\log x)^{k-1}}{(\log x)p^k},
\end{equation}
where we used the fact that
$$
\sum_{\substack{q\le x\\ q\equiv 1\pmod p}} \frac{1}{q}\ll \frac{\log\log x}{p}
$$
uniformly in $(x/2)^{a}\le p\le x^{1/k}$, which follows from the Brun-Titchmarsch theorem by partial summation. Summing up the above bound \eqref{eq:bound} over all $p>(x/2)^{a}$ gives
\begin{eqnarray}
\label{eq:B3}
\# {\mathcal B}_2(x) & \ll & \frac{x(\log\log x)^{k-1}}{\log x} \sum_{(x/2)^{a}<p\le x^{1/k}} \frac{1}{p^k}\nonumber\\
& \ll & \frac{x(\log\log x)^{k-1}}{\log x} \int_{(x/2)^{a}}^{x^{1/k}} \frac{d \pi(t)}{t^k} \nonumber\\
& \ll & \frac{x(\log\log x)^{k-1}}{\log x} \left( \frac{1}{t^{k-1} \log t} \Big|_{t=(x/2)^{a}}^{t=x^{1/k}}+\int_{(x/2)^{a}}^{x^{1/k}} \frac{dt}{t^k \log t}\right)\nonumber\\
& \ll & \frac{x(\log\log x)^{k-1}}{\log x}\left(\frac{1}{x^{a(k-1)}\log x}\right)\nonumber\\
& \ll & \frac{x^{1-a(k-1)}(\log\log x)^{k-1}}{(\log x)^2}.
\end{eqnarray}
The upper bound follows from \eqref{eq:B2} and \eqref{eq:B3}.
\subsection{The lower bound}
The following result is Lemma 2.1 in \cite{BS}.
\begin{lemma}
\label{lem:BS}
There exist functions $C_2(\nu)>C_1(\nu)>0$ defined for all real numbers $\nu\in (0,17/32)$ such that for every integer $u\ne 0$ and positive real number $K$, the inequalities
$$
\frac{C_1(\nu)y}{p\log y}<\pi(y;p,u)<\frac{C_2(\nu) y}{p\log y}
$$
hold for all primes $p\le y^{\nu}$ with $O(y^{\nu}/(\log y)^K)$ exceptions, where the implied constant depends on $u,~\nu,~K$. Moreover, for any fixed $\varepsilon>0$, these functions can be chosen to satisfy the following properties:
\begin{itemize}
\item $C_1(\nu)$ is monotonic decreasing, and $C_2(\nu)$ is monotonic increasing;
\item $C_1(1/2)=1-\varepsilon$ and $C_2(1/2)=1+\varepsilon$.
\end{itemize}
\end{lemma}
So, we take $y=x^{1/k}$ and consider primes $p\in {\mathcal I}=[y^{ak},2y^{ak}]$. Then $2y^{ak}=y^{\nu}$, where $\nu=ak+(\log 2)/(ak\log y)<17/32$ for all $x$ sufficiently large. So, let $\varepsilon>0$ be such that
$a<17/32-\varepsilon$ and assume that $x$ is sufficiently large such that $\log 2/(\log y)<\varepsilon/2$.
Then, by Lemma \ref{lem:BS} with $u=1$ and $K=2$, the set ${\mathcal P}$ of primes $p\le 2y$ such that
$$
\pi(y;p,1)>\frac{C_1(17/32-\varepsilon/2) y}{p\log y }
$$
contains all primes $p\le 2y^{ak}$ with $O(y^{ak}/(\log y)^2)$ exceptions. Thus, the number of primes $p\in {\mathcal P}\cap {\mathcal I}$ satisfies
$$
\#\left({\mathcal P}\cap {\mathcal I}\right)\ge \pi(2y^{ak})-\pi(y^{ak})-O\left(\frac{y^{a}}{(\log y)^2}\right)>\frac{y^{ak}}{\log y}
$$
for all $x$ sufficiently large independently in $k$ and $a$. Consider numbers of the form $n=p_1\cdots p_k$, where $p_1<\cdots<p_k\le y$ are all primes congruent to $1$ modulo $p$. Furthermore, it is clear that
$p=P(p_i-1)$ for $i=1,\ldots,k$. Note that $n\le x$. The number of such $n$ is,
for $p$ fixed,
$$
\binom{\pi(y;p,1)}{k}\gg \left(\frac{y}{p\log y}\right)^k\gg \frac{x}{p^k (\log x)^k}.
$$
Summing up the above bound over $p\in {\mathcal P}\cap {\mathcal I}$, we get that
\begin{eqnarray*}
\#{\mathcal A}_{k,a}(x) & \gg & \frac{x}{(\log x)^k} \sum_{p\in {\mathcal P}\cap {\mathcal I}} \frac{1}{p^k}\gg
\frac{x}{(\log x)^k} \left(\frac{\#\left({\mathcal P}\cap {\mathcal I}\right)}{y^{ak^2}}\right)\\
& \gg & \frac{xy^{ak}}{y^{ak^2} (\log x)^{k}\log y}\gg \frac{x^{1-a(k-1)}}{(\log x)^{k+1}},
\end{eqnarray*}
which is what we wanted.
\section{Comments and Remarks}
It is not likely that Goldfeld's method extends to the situation considered in Theorem \ref{compuesto}. As we have seen, the proof of Theorem \ref{primo} is based on the identity \eqref{basico}. Then, Mertens's theorem, the Brun-Titchmarsh inequality and the Bombieri-Vinogradov theorem are used to extract the desired estimate out of it. If we try to follow the same strategy to prove Theorem \ref{compuesto}, for example with $a=1/(2k)$, we are then led to replace the left hand side of \eqref{basico} by
$$
L_k(x):= \sum_{m \leq x^{1/k} }^{ } \Lambda(m) \pi_k(x;m,1),
$$
where $\pi_k(x;m,1) = \# \{ n \in \mathcal{A}_k(x) : p | n \Rightarrow p \equiv 1 \mod m\}.$ Let $\pi_k(x)$ denote the number of squarefree integers up to $x$ having exactly $k$ prime factors. Then, letting $p_1, p_2, \ldots, p_k$ denote primes,
\begin{eqnarray}
L_k(x) &=& \sum_{\substack{p_1< p_2<\cdots< p_k \\ p_1p_2 \cdots p_k \leq x}}^{ } \sum_{\substack{m\mid {\text{\rm{gcd}}} (p_i-1) \\ 1\le i\le k}}\Lambda(m) \nonumber \\
&=& \sum_{\substack{p_1 <p_2 <\cdots < p_k \\ p_1p_2 \cdots p_k \leq x}}^{ } \log \left( {\text{\rm{gcd }}}\left(p_i-1:1\le i\le k \right)\right) \nonumber \\
&\geq & (\log 2) \pi_k(x) \gg_k \frac{x(\log \log x)^{k+1}}{\log x}, \quad x \rightarrow \infty. \nonumber
\end{eqnarray}
In view of \eqref{crudo}, we see that $L_k(x)$ grows much faster, when $k \geq 2$, than the counting function we are interested in. Hence, it is unlikely that $L_k(x)$ can be used to obtain information on the growth of $\mathcal{A}_k(x)$.
\medskip
{\bf Acknowledgement.} We thank N. Billerey for stimulating questions. Part of this work was done during a visit of F.~L.
at the Mathematics Department of the Universidad de Valparaiso during 2013 with a MEC project from CONICYT. He thanks the people of this Department for their
hospitality. F.~L. was also partially supported by a Marcos Moshinsky Fellowship and
Projects PAPIIT IN104512,
CONACyT 163787 and CONACyT 193539. R.M. is partially supported by FONDECYT grant 11110225.
|
\section{Introduction}
The investigation of thermodynamics of computation and information processing predates the era of computer technology~\cite{Demon}.
From the viewpoint of fundamental physics, this research area is closely related to the foundation of the second law of thermodynamics, which dates back to Maxwell's thought experiment on ``Maxwell's demon'' in the ninteenth century~\cite{Maxwell}. Several decades later, Szilard made the first crucial step toward the quantitative understanding of the relationship between information and thermodynamics~\cite{Szilard}. In his thought experiment with a single-particle heat engine, which is called the Szilard engine, the demon can extract $k_{\rm B}T \ln 2$ of work from a single heat bath through feedback control by using one bit of information, where $k_{\rm B}$ is the Boltzmann constant and $T$ is the temperature of the bath.
By developing Szilard's observation, Brillouin investigated the relationship between thermodynamic and informational entropies~\cite{Brillouin1,Brillouin2}.
Recently, a kind of the Szilard engine have been experimentally demonstrated for the first time~\cite{Toyabe} with a small thermodynamic engine that works at the level of thermal fluctuations.
Another prominent observation was made by Landauer~\cite{Landauer,Landauer2}. He pointed out that, to erase one bit of information from a memory, at least $k_{\rm B}T \ln 2$ of heat should be emitted into a heat bath and the same amount of work is needed. This has been referred to as the Landauer principle.
Bennett also discussed thermodynamics of computation by considering the concept of logical reversibility and its relationship to thermodynamics~\cite{Bennett0,Bennett1,Bennett2}.
Moreover, Zurek discussed thermodynamics of computation in terms of algorithmic complexity~\cite{Zurek1,Zurek2}.
Later, the Landauer principle has been studied in various aspects~\cite{Plenio,Maruyama}; it has been derived in terms of statistical mechanics in some setups~\cite{Shizume,Piechocinska,Esposito2}, has been demonstrated in concrete systems theoretically~\cite{Lutz1,Lutz2,Lambson,Aurell,Diana1} and experimentally~\cite{Berut,Berut2}, has sparked up intense discussions on its validity~\cite{Goto,Allahverdyan3,Horhammer,Vaccaro,Maroney0,Norton,Barkeshli,Maroney,Turgut,Sagawa-Ueda2}.
In a rather different context, the recent advancements in nonequilibrium statistical mechanics have revealed a fundamental aspect of the second law of thermodynamics~\cite{Jarzynski1,Crooks1,Crooks2,Jarzynski2,Seifert,Kawai,Jarzynski3,Crooks3,Kurchan,Tasaki0,Campisi,Sagawa3}.
This has enabled us to more quantitatively and comprehensively understand thermodynamics of computation and information processing.
In terms of theory, the second law of thermodynamics and the fundamental nonequilibrium relations (e.g, the fluctuation theorem and the Jarzynski equality) have been generalized to information processing processes such as measurement, information erasure, and feedback control~\cite{Nielsen,Touchette,Touchette2,Sagawa-Ueda1,Cao2,Jacobs,Sagawa-Ueda3,Suzuki,Horowitz1,Horowitz2,Morikuni,Abreu,Horowitz3,Sagawa-Ueda4,Abreu2,Still,Vedral,Sagawa-Ueda2012,Granger,Ito2,Deffner,Tasaki,HSP,Sagawa-Ueda-NJP}, where information contents and thermodynamic quantities are treated on an equal footing.
In particular, the concept of thermodynamic reversibility in the presence of feedback control has been established~\cite{Horowitz2,Sagawa-Ueda4}, and several concrete models of thermodynamically reversible information processing have been proposed~\cite{Jacobs,Horowitz2,Abreu,Horowitz3,Sagawa-Ueda4,HSP}.
More recently, a variety of autonomous Maxwell's demons has attracted much attention~\cite{
Mandal,Barato,Strasberg,Mandal2,Barato2}.
In terms of experiment, a generalized Jarzynski equality with feedback control has been verified~\cite{Toyabe}.
The Jarzynski equality for information erasure has also been investigated experimentally~\cite{Berut2}.
In light of these advancements, we are now in the position to revisit and to further clarify the fundamental concepts in thermodynamics of computation and information processing.
In this paper, we review and investigate the long-standing fundamental problems in thermodynamics of computation, shedding new light on the celebrated Landauer principle from the viewpoint of modern statistical physics.
In particular, we clarify the fundamental relationship between the thermodynamic and logical reversibilities.
Moreover, on the basis of the second law of thermodynamics, we derive universal thermodynamic inequalities that set the lower bounds of the work requirement and the heat emission during a computation.
Our observation does not contradict the conventional Landauer principle, but extends it to much broader class of memories that perform computation.
While we assume that a memory is a classical system, the generalization of our arguments to the quantum cases is straightforward~\cite{Sagawa3}.
This paper is organized as follows. In Sec.~2, we discuss the concept of thermodynamic reversibility in terms of the total entropy production in the whole system including a heat bath. In Sec.~3, we discuss the concept of logical reversibility, and illustrate several typical examples of reversible and irreversible computations. In Sec.~4, we discuss the conventional Landauer principle, and clarify the relationship between thermodynamic and logical reversibilities with the standard setup of the Landauer principle. In Sec.~5, we formulate the general setup of thermodynamic computation, and derive the second law of thermodynamics for computational processes.
In Sec.~6, we discuss the situation that there are two memories; this setup enables us to analyze measurement and feedback, which constitute the typical situation of Maxwell's demon. In Sec.~7, we conclude this paper.
\section{Thermodynamic Reversibility}
In this section, we clarify the concept of thermodynamic reversibility.
We consider a time evolution of a thermodynamic system in the presence of heat bath(s).
We assume that the system may be driven from equilibrium by changing external parameters such as the volume of the gas.
First of all, we roughly characterize the thermodynamic reversibility according to the standard definition in thermodynamics~\cite{Callen,Prigogine}:\\
\\
{\it A physical process is thermodynamically reversible if and only if its time-reversal is not prohibited by the second law of thermodynamics. Otherwise, the process is thermodynamically irreversible.}\\
\\
In more precise, we need to consider the ensemble of the system, because, from the microscopic point of view, the dynamics of the phase-space point of the system is stochastic due to thermal fluctuations. The thermodynamic property of the system is described by the probability distribution on the phase space. The thermodynamic reversibility can then be characterized as follows~\cite{Jarzynski1,Crooks1,Crooks2,Jarzynski2,Seifert,Kawai,Crooks3}:\\
\\
{\it A physical process is thermodynamically reversible if and only if the time evolution of the probability distribution in the process can be time-reversed, where the change of the external parameters is also time-reversed, and the signs of the amounts of the work and the heat are changed.}\\
\\
A process can become thermodynamically reversible in the quasi-static limit, where the change of the external parameters is much slower than the relaxation time of the system, and the state of the system can always be regarded in thermal equilibrium during the process.
For example, the quasi-static and isothermal compression and expansion of the gas in a box are both thermodynamically reversible, as they are the time-reversal with each other.
We note that all quasi-static processes are not necessarily reversible; there may be quasi-static but irreversible processes such as a weak constant of two baths with different temperatures~\cite{Callen}.
To characterize the thermodynamic reversibility in a more quantitative way, we discuss the concept of entropy production.
For simplicity, we assume that there is a single heat bath at inverse temperature $\beta := (k_{\rm B}T)^{-1}$. The generalization of the following arguments to the cases with multiple baths is straightforward.
Let $\mathcal Y$ be the phase space of the system, $y, y' \in \mathcal Y$ be the initial and final phase-space points, and $P[y]$ and $P'[y']$ be their probabilities.
We consider the Shannon entropy of the probability distribution on the phase space~\cite{Shannon,Cover-Thomas}. The initial and final entropies are respectively given by
\begin{equation}
S := - \int_{y \in \mathcal Y} dy P[y] \ln P[y], \ \ S' := - \int_{y' \in \mathcal Y} dy' P'[y'] \ln P'[y'].
\end{equation}
We denote as $Q$ the average heat absorbed by the system from the heat bath during the dynamics.
The total entropy production in the relevant total system (i.e., the whole ``universe'') including the bath is then given by~\cite{Crooks1,Crooks2,Jarzynski2,Seifert}
\begin{equation}
\Delta S_{\rm tot} := \Delta S -\beta Q,
\label{total_entropy_production}
\end{equation}
where $\Delta S := S' - S$ is the change in the Shannon entropy of the system. We note that $-\beta Q$ is regarded as the change in the entropy of the heat bath.
The second law of thermodynamics can then be expressed as~\cite{Crooks1,Crooks2,Jarzynski2,Seifert,Callen,Prigogine}
\begin{equation}
\Delta S_{\rm tot} \geq 0,
\label{second1}
\end{equation}
or equivalently,
\begin{equation}
\Delta S \geq \beta Q.
\label{second1_1}
\end{equation}
We stress that inequality~(\ref{second1}) holds for any initial and final nonequilibrium distributions of the system (i.e., for any $P[y]$ and $P'[y']$).
Inequality~(\ref{second1}) can be derived on the basis of nonequilibrium statistical mechanics~\cite{Sagawa3}. In fact, inequality (\ref{second1}) is a straightforward consequence of the fluctuation theorem~\cite{Crooks1,Crooks2,Jarzynski2,Seifert}.
On the basis of inequality~(\ref{second1}), the thermodynamic reversibility can be characterized in terms of the entropy production~\cite{Callen,Prigogine}: \\
\\
{\it A physical process is thermodynamically reversible if the equality in (\ref{second1}) is achieved (i.e., the total entropy production is zero). }\\
\\
In fact, if $\Delta S_{\rm tot} > 0$ holds in a process, its time-reversal is impossible because of $\Delta S_{\rm tot} < 0$.
The thermodynamically reversible condition $\Delta S_{\rm tot} = 0$ can be achieved in the quasi-static limit, where the system is always in thermal equilibrium during the process.
We stress that a process can be thermodynamically reversible even if there is an entropy transfer from the system to the bath or vice versa. If the amount of the increase/decrease in entropy in the system is the same as the amount of the decrease/increase in the bath, the total amount of the entropy increase in the whole universe is zero, where the equality in (\ref{second1}) is achieved and the process is thermodynamically reversible.
As a simple example, we consider a quasi-static isothermal expansion of an ideal gas with $N$ particles. Starting from a thermal equilibrium state, we expand the gas isothermally and quasi-statically, doubling its volume. The entropy of the system is then increased by $N \ln 2$, and the heat emission to the bath is given by $- Q = \beta^{-1} N \ln 2$ (i.e., the entropy of the bath is decreased by $N \ln 2$). Therefore, the total entropy production is given by
\begin{equation}
\Delta S_{\rm tot} = ( -N \ln 2) + N \ln 2 = 0,
\end{equation}
which implies that the quasi-static isothermal expansion is thermodynamically reversible. Similarly, the quasi-static isothermal compression is also thermodynamically reversible.
We now discuss the relationship between the foregoing thermodynamic perspective and the microscopic perspective based on the reversible Hamiltonian dynamics.
One of the crucial progresses in modern statistical physics is that it has succeeded to reconcile these two perspectives~\cite{Jarzynski1,Jarzynski2,Kawai}.
In fact, the second law (\ref{second1}) can be derived on the basis of the fluctuation theorem, where one needs essentially only two assumptions: the microscopic dynamics satisfies the Liouville theorem, and the initial distribution of the bath is the canonical distribution~\cite{Jarzynski2}.
A crucial observation here is that the Shannon entropy on the whole phase space of the system and the bath is different from the thermodynamic entropy in the whole universe; the former is conserved in any Hamiltonian dynamics, while the latter is increased in thermodynamically irreversible processes even if the underlying microscopic Hamiltonian dynamics is reversible.
The change in the thermodynamic entropy is equal to the relative entropy (i.e., the Kullback-Leibler divergence) between the probability distribution in the final state and a reference probability distribution such as the canonical distribution~\cite{Jarzynski3}.
We again stress that the thermodynamic entropy is not equivalent to the Shannon entropy of the total system including the system and the bath.
Correspondingly, we have not considered the Shannon entropy in the bath in the foregoing argument, but regarded the heat transfer $-\beta Q$ as the entropy change in the bath, which is consistent with the picture that the relative entropy corresponds to the entropy production in the whole universe~\cite{Jarzynski2,Jarzynski3,Sagawa3}.
The thermodynamic irreversibility (i.e., the positive entropy production in the whole universe) can also be characterized by the gap between the probability distributions of microscopic trajectories in the forward process and the backward one~\cite{Kawai,Crooks3,Jarzynski3}.
We note that the same argument is also valid for quantum systems that obey unitary dynamics~\cite{Kurchan,Tasaki0,Campisi,Sagawa3}, by replacing the Shannon entropy by the von Neumann one.
The concept of thermodynamic reversibility discussed in this manuscript is consistent with the above-mentioned observation based on the fluctuation theorem, and therefore has a rigid theoretical foundation that is consistent with the microscopic reversible physics.
In terms of the fluctuation theorem, $\Delta S_{\rm tot} = 0$ holds if and only if the probability distribution of the trajectories in the phase space is the same as that of the time-reversed trajectories, which has been explicitly discussed in, for example, Ref.~\cite{Crooks3,Jarzynski3}.
We now assume that the initial and final distributions are the canonical distributions:
\begin{equation}
P[y] = e^{\beta (F - E[y])}, \ P'[y'] = e^{\beta (F'- E'[y'])},
\end{equation}
where $E[y]$ ($E'[y']$) and $F$ ($F'$) are the initial (final) energy and the initial (final) free energy of the system, respectively. In this case, we obtain
\begin{equation}
S = \beta (E - F ), \ \ S' = \beta (E' - F' ),
\end{equation}
where $E$ ($E'$) is the ensemble average of the initial (final) energy. Therefore,
\begin{equation}
\Delta S = \beta (\Delta E - \Delta F),
\end{equation}
where $\Delta E := E' -E$ and $\Delta F := F' -F$.
Let $W$ be the average work performed on the system during the process. The total entropy production is then given by
\begin{equation}
\Delta S_{\rm tot} = \beta (W - \Delta F),
\end{equation}
where we used the first law of thermodynamics:
\begin{equation}
W = \Delta E - Q.
\end{equation}
Therefore, the second law~(\ref{second1}) reduces to
\begin{equation}
W \geq \Delta F,
\label{second2}
\end{equation}
where the equality can be achieved in the quasi-static limit.
If the final distribution is different from the canonical distribution, we can show that
\begin{equation}
\Delta S_{\rm tot} \leq \beta ( W - \Delta F),
\label{S_F_noneq}
\end{equation}
where the final free energy is given by the equilibrium one corresponding to the final Hamiltonian.
In this case, we again obtain inequality~(\ref{second2}).
We note that the equality in (\ref{S_F_noneq}) is achieved if the final distribution is given by the canonical distribution.
\section{Logical Reversibility}
In this section, we discuss the concept of logical reversibility, and introduce the logical entropy that is decreased by logically irreversible computation.
\subsection{Definition and examples}
We consider a memory that has several logical states, in which information is stored.
A computational process is performed by changing an input logical state into output one with a certain algorithm.
Let $\mathcal M$ and $\mathcal M'$ be the sets of input and output logical states, respectively.
We assume that these are finite sets.
For example, if the input consists of $n$ bits, we set $\mathcal M = \{ 0,1 \}^n$.
We note that the logical states do not have one-to-one correspondence to the physical phase space, as discussed in Sec.~4 in detail.
We next formulate (deterministic) computational processes.
In this paper, a computational process $\hat C$ is defined as a map
\begin{equation}
\hat C: \mathcal M \to \mathcal M'.
\end{equation}
We also call $\hat C$ as a gate.
We note that $\hat C$ is not necessarily a surjection.
In order to discuss the logical reversibility and its relationship to thermodynamics, we do not need the detailed characterization of computable functions in terms of computability theory~\cite{Moore}; the following argument is applicable to any map $\hat C : \mathcal M \to \mathcal M'$.
We now define the logical reversibility~\cite{Landauer2,Bennett0,Bennett1,Bennett2}: \\
\\
{\it A computational process $\hat C$ is logically reversible if and only if it is an injection.
In other words, $\hat C$ is logically reversible if and only if, for any output logical state, there is a unique input logical state.
Otherwise, $\hat C$ is logically irreversible.}\\
\\
We note that, if $\mathcal M' = \mathcal M$, a computational process is reversible if and only if it is a bijection (i.e., an injection and a surjection).
In general, a computational process is reversible if and only if we can precisely estimate the input state from the output state.
In fact, if a computational process is reversible and $\hat C$ is an injection, we can define the reversed computational process
\begin{equation}
\hat C^{-1}: \mathcal M' \to \mathcal M,
\end{equation}
where the domain of $\hat C^{-1}$ is given by the image of $\hat C$ denoted as $\hat C (\mathcal M) \subset \mathcal M'$.
We next discuss several simple examples of computation.
We first consider the case that both of the input and output are one bit so that $\mathcal M = \mathcal M' = \{ 0,1 \}$.
A simple example of reversible gate is NOT, which is defined as
\begin{equation}
0 \mapsto 1, \ \ 1 \mapsto 0.
\end{equation}
This is clearly a bijection, and the reversal of NOT is also NOT.
A simple example of irreversible gate is the information erasure that is referred to as ERASE:
\begin{equation}
0 \mapsto 0, \ \ 1 \mapsto 0.
\end{equation}
This is not a bijection, as the logical state is always $0$ after the computation; we cannot estimate the input state from the output state.
We next consider the case that both of the input and output are two bits so that $\mathcal M = \mathcal M' = \{ 0,1 \}^2 = \{ 00, 01, 10, 11 \}$.
A simple example of reversible gate is CNOT (controlled-NOT), which is defined as
\begin{equation}
00 \mapsto 00, \ \ 01 \mapsto 01, \ \ 10 \mapsto 11, \ \ 11 \mapsto 10,
\label{CNOT}
\end{equation}
where the first (left) bit is the control bit and the second (right) bit is the target bit. If the control gate is $1$, CNOT behaves as NOT on the target bit. Otherwise, CNOT is just identity. CNOT is a bijection and its reversal is also CNOT.
CNOT can be used for a measurement (i.e., the copy of information); if the input of the target bit is $0$, the input of the control bit is copied to the output of the target bit by CNOT:
\begin{equation}
00 \mapsto 00, \ \ \ 10 \mapsto 11.
\label{CNOT_measurement}
\end{equation}
CNOT can also be used for feedback control, where the control bit is the feedback controller and the target bit is to be controlled.
In the case of feedback control, we exchange the roles of the bits from the case of measurement~(\ref{CNOT_measurement}); we regard the first (left) bit as the target bit and the second (right) bit as the control bit for feedback control.
Before the feedback, the input states of the two bits are assumed to be the same, which is realized after the measurement. The target bit is then flipped if the control bit is $1$:
\begin{equation}
00 \mapsto 00, \ \ 11 \mapsto 01,
\label{CNOT_feedback}
\end{equation}
where the output of the target bit is $0$ irrespective of its input.
We next consider the case that the input is two bit and the output is one bit, where $\mathcal M = \{ 0,1 \}^2$ and $\mathcal M' = \{ 0, 1 \}$. In this case, any computational process is irreversible. In fact, the numbers of the elements in $\mathcal M$ and $\mathcal M'$ are $4$ and $2$, respectively, and therefore any map from $\mathcal M$ to $\mathcal M'$ cannot be an injection. An example of such an irreversible gate is XOR, which is defined as
\begin{equation}
00 \mapsto 0, \ \ 01 \mapsto 1, \ \ 10 \mapsto 1, \ \ 11 \mapsto 0.
\end{equation}
\subsection{Reversible extension}
We next show that any irreversible computation can be extended to a reversible computation, which has been discussed by Bennett in detail~\cite{Bennett0}. In fact, we can show the following theorem: \\
\\
{\it For any $\hat C : \mathcal M \to \mathcal M'$, there exist a finite set $\mathcal M''$ and a map $\hat C_{\rm ex}: \mathcal M \to \mathcal M' \times \mathcal M''$, such that $\hat C_{\rm ex}$ is logically reversible and the restriction of $\hat C_{\rm ex}$ on $\mathcal M \to \mathcal M'$ is equivalent to $\hat C$.} \\
\\
Here, $\mathcal M''$ can be regarded as an ancilla or an environment.
We call $\hat C_{\rm ex}$ the reversible extension of $\hat C$.
We note that the reversible extension is not unique.
Before the proof of the theorem, we illustrate the reversible extension of ERASE. Let $\mathcal M'' = \{ 0,1 \}$. We then define $\hat C_{\rm ex} : \{ 0,1 \} \to \{ 0,1 \}^2$ by
\begin{equation}
0 \mapsto 00, \ \ \ 1 \mapsto 01,
\label{ERASE_ex}
\end{equation}
where the first (left) bit of the output is in $\mathcal M$. $\hat C_{\rm ex}$ is clearly an injection and therefore reversible. Moreover, its restriction on $\mathcal M \to \mathcal M'$ is equivalent to ERASE. Intuitively, this extension describes that the erased information can be kept in ancilla $\mathcal M''$ (i.e., can remain in the environment).
We now show a simple constructive proof of the theorem: we set $\mathcal M'' = \mathcal M$ and define $\hat C_{\rm ex}$ as
\begin{equation}
\hat C_{\rm ex} (m) = (\hat C(m), m ) \in \mathcal M' \times \mathcal M'',
\label{reversible_ex}
\end{equation}
which satisfies the condition of the theorem. (Q.E.D.)
The extension of ERASE~(\ref{ERASE_ex}) is a special case of extension~(\ref{reversible_ex}). We note that extension~(\ref{reversible_ex}) may be redundant in general. For example, in the case of XOR, the extension~(\ref{reversible_ex}) is given by
\begin{equation}
00 \mapsto 000, \ \ 01 \mapsto 101, \ \ 10 \mapsto 110, \ \ 11 \mapsto 011,
\end{equation}
where $\mathcal M'' = \{ 0, 1 \}^2$. However, there is a simpler reversible extension of XOR:
\begin{equation}
00 \mapsto 00, \ \ 01 \mapsto 10, \ \ 10 \mapsto 11, \ \ 11 \mapsto 01,
\end{equation}
where $\mathcal M'' = \{ 0 , 1 \}$.
\subsection{Entropy change in computation}
We next consider the concept of logical entropy, which is defined by the Shannon entropy of the logical states.
We consider probability distribution on $\mathcal M$ (i.e., $P[m]$ for $m \in \mathcal M$) with $\sum_{m \in \mathcal M} P[m] = 1$. After computation $\hat C$, the probability distribution on $\mathcal M'$ (i.e., $P'[m']$ for $m' \in \mathcal M'$) is given by
\begin{equation}
P' [m'] = \sum_{m: \ \hat C (m) = m'} P[m],
\label{prob_com1}
\end{equation}
where the sum in the right-hand side is taken over $m$ satisfying $\hat C (m ) = m'$.
If the computation is reversible, Eq.~(\ref{prob_com1}) reduces to
\begin{equation}
P' [m'] = P[\hat C^{-1} (m') ] \ \ ({\rm if} \ m' \in \hat C (\mathcal M)), \ \ \ P' [m'] = 0 \ \ {\rm (otherwise)},
\label{prob_com2}
\end{equation}
which describes the conservation of probability.
We now define the initial and final logical entropies by~\cite{Shannon,Cover-Thomas}
\begin{equation}
H (\mathcal M) := - \sum_{m \in \mathcal M} P[m] \ln P[m],
\end{equation}
\begin{equation}
H' (\mathcal M') := - \sum_{m' \in \mathcal M'} P'[m'] \ln P'[m'].
\end{equation}
If the computation is logically reversible, the logical entropy does not change. In fact, by using Eq.~(\ref{prob_com2}), we obtain
\begin{equation}
\eqalign{
H' (\mathcal M') &= - \sum_{m' \in \hat C (\mathcal M)} P[\hat C^{-1} (m') ] \ln P[\hat C^{-1} (m') ] \\
&= - \sum_{m \in \mathcal M} P[m] \ln P[m] = H (\mathcal M).
}
\end{equation}
In contrast, if the computation is logically irreversible, the logical entropy states is decreased:
\begin{equation}
H(\mathcal M) \geq H'(\mathcal M').
\label{Shannon_decrease}
\end{equation}
In fact, by noting that $P'[m'] = \sum_{m : \ \hat C (m) = m'} P[m]$ (i.e., $P[m] / P'[m']$ is a probability distribution over $m$'s that satisfy $\hat C (m) = m'$), we obtain
\begin{equation}
H(\mathcal M) - H'(\mathcal M') = - \sum_{m'} P' [m'] \sum_{m : \ \hat C (m ) = m'} \frac{P[m]}{P'[m']} \ln \frac{P[m]}{P'[m']} \geq 0.
\end{equation}
We note that inequality~(\ref{Shannon_decrease}) is a special case of the data processing inequality in information theory~\cite{Cover-Thomas}.
In the case of the reversible extension~(\ref{reversible_ex}), the logical entropy of the extended logical states does not change: $H(\mathcal M ) = H' (\mathcal M' \times \mathcal M'')$. From the subadditivity of the Shannon entropy, we have
\begin{equation}
H'(\mathcal M' \times \mathcal M'') \leq H'(\mathcal M') + H'( \mathcal M''),
\label{subadditivity}
\end{equation}
and therefore, together with inequality (\ref{Shannon_decrease}),
\begin{equation}
H'(\mathcal M') \leq H(\mathcal M ) \leq H'(\mathcal M') + H'( \mathcal M'').
\end{equation}
We discuss simple examples with $\mathcal M = \mathcal M' = \{ 0,1 \}$. Let $p := P[0]$ and $p' := P' [0]$ be the probabilities of the input $0$ and output $0$, respectively. In the case of NOT, $p' = 1-p$ and therefore $H(\mathcal M) = - p \ln p - (1-p) \ln (1-p) = H'(\mathcal M')$. In the case of ERASE, $p' = 1$ for any $p$. Therefore, $H(\mathcal M) = - p \ln p - (1-p) \ln (1-p)$ and $H' (\mathcal M') = 0$, which implies that the logical entropy is decreased by $H(\mathcal M)$ by information erasure.
In the case of the reversible extension of erasure~(\ref{ERASE_ex}), the equality in~(\ref{subadditivity}) is achieved with $ H'(\mathcal M') = 0$ and $H'( \mathcal M'') = - p \ln p - (1-p) \ln (1-p)$. Therefore, the decrease in the entropy of the memory is compensated for by the increase in the entropy of the ancilla or environment, such that the total entropy is conserved.
\section{Conventional Landauer Principle}
In this section, we discuss the conventional Landauer principle~\cite{Landauer}, and clarify the relationship between the thermodynamic and logical reversibilities in the standard setup of information erasure.
We consider the information erasure of one bit of information from a memory in the presence of a single heat bath at inverse temperature $\beta$.
As a simple and conventional setup, we consider a binary symmetric potential as a physical model of the memory (Fig.~1 (a))~\cite{Landauer,Landauer2,Bennett1}, where the height of the barrier is assumed to be much larger than the thermal fluctuation.
If the particle is in the left (right) well, the logical state is ``$0$'' (``$1$'').
An idealized model of the double-well memory is shown in Fig.~1 (b), where the left (right) box corresponds to the left (right) well, and the wall at the center of the box corresponds to the barrier of the binary potential.
We note that these two models are not completely equivalent.
In fact, the width of the barrier is finite in the double-well memory but infinitely small in the two-box memory.
We also note that the height of the barrier in the double-well memory can be very large but is still finite theoretically, while the barrier in the two-box memory can be regarded as perfectly impenetrable.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=75mm]{Fig1.png}
\end{center}
\caption{(a) Schematic of a memory with a binary symmetric potential. If the particle is in the left (right) well, the logical state is $0$ ($1$). (b) A model of the memory with two boxes, which is an idealization of the binary-potential memory (a). If the particle is in the left (right) box, the logical state is $0$ ($1$).}
\end{figure}
Before the erasure, the probability of ``$0$'' and ``$1$'' are assumed to be equally $1/2$. After the erasure, the particle is in ``$0$'' with unit probability.
The logical entropy is $\ln 2$ before the erasure, while it is $0$ after the erasure. Therefore, the logical entropy changes by $\Delta H = -\ln 2$ during the erasure.
We assume that the initial probability distribution in the memory is thermal equilibrium, and that the final probability distribution is in conditional thermal equilibrium only in the well of ``$0$.''
The crucial observation in the Landauer principle~\cite{Landauer,Landauer2,Bennett0,Bennett1,Bennett2} is that the logical entropy must be included as a part of the total entropy, and be treated on an equal footing with the thermodynamic entropy.
The change in the total entropy of the memory is then given by
\begin{equation}
\Delta S = \Delta H = - \ln 2,
\end{equation}
and therefore, the second law~(\ref{second1}) reduces to
\begin{equation}
- Q \geq \beta^{-1} \ln 2,
\label{Landauer1}
\end{equation}
where $- Q$ is the heat that is emitted to the heat bath during the erasure.
Inequality~(\ref{Landauer1}) is the conventional Landauer principle and the right-hand side is called the Landauer bound. Inequality~(\ref{Landauer1}) implies that the decrease in the entropy by $\ln 2$ in the memory must be compensated for by the increase in the entropy of the bath by at least $\ln 2$, which is accompanied by the inevitable heat emission of $- Q = \beta^{-1} \ln 2$.
We note that the increase in the entropy of the bath corresponds to the increase in the entropy of ancilla $\mathcal M''$ in terms of the reversible erasure~(\ref{ERASE_ex}) in Sec.~3.2.
The Landauer principle can then be summarized as follows: \\
\\
{\it A positive amount of the heat emission is inevitable during the logically irreversible information erasure.}\\
\
We next consider the work needed for the erasure.
Since the internal energy of the memory does not change during the erasure, we have $- Q = W$ from the first law of thermodynamics. Therefore, we obtain the minimal work needed for the information erasure:
\begin{equation}
W \geq \beta^{-1} \ln 2.
\label{Landauer2}
\end{equation}
Inequality~ (\ref{Landauer2}) is also referred to as the Landauer principle.
If the information erasure is quasi-static, the equalities in (\ref{Landauer1}) and (\ref{Landauer2}) can be achieved as
\begin{equation}
- Q = \beta^{-1} \ln 2, \ \ W = \beta^{-1} \ln 2.
\end{equation}
A concrete protocol of such a quasi-static erasure will be discussed below.
Therefore, the total entropy production is given by
\begin{equation}
\Delta S_{\rm tot} = (-\ln 2) - (- \ln 2) = 0,
\end{equation}
which implies that the quasi-static erasure is thermodynamically reversible, while it is logically irreversible.
We now conclude that:\\
\\
{\it The logically irreversible erasure can be performed in a thermodynamically reversible manner in the quasi-static limit.}\\
\\
This does not contradict the conventional Landauer principle.
In fact, the logical reversibility is defined only by the reversibility of the logical states, which is related only to the logical entropy. In contrast, the thermodynamic reversibility is related to the reversibility of the relevant total system (i.e., the whole universe) including the heat bath, and to the total entropy production as discussed in Sec.~2.
Therefore, these two reversibilities are not equivalent in general.
We note that, if the erasure is not quasi-static but is performed with a finite velocity, the erasure becomes thermodynamically irreversible. We summarize the thermodynamic and logical reversibilities for the information erasure in Table 1.
\begin{table}
\caption{Summary for the conventional setup of information erasure.}
\begin{center}
\begin{tabular}{| l || l | l | }\hline
{} & Quasi-static & Finite-velocity \\ \hline \hline
Thermodynamically & reversible & irreversible \\ \hline
Logically & irreversible & irreversible \\ \hline
Heat emission & $ = \beta^{-1} \ln 2 $ & $> \beta^{-1} \ln 2 $ \\ \hline
\end{tabular}
\end{center}
\end{table}
Let us discuss the above point in more detail. During the information erasure, the entropy is transferred from the logical (i.e., accessible) degrees of freedom in the memory into the environmental (i.e., inaccessible or microscopic) degrees of freedom in the bath.
Even if the erased information may still be kept in the degrees of freedom of the environment in principle, one cannot access or recognize in practice such dissipated information in the bath in a real computation.
This implies that the erasure is logically irreversible (i.e., the accessibility to the stored information is lost).
We note that the logical reversibility depends on the boundary between the logical and environmental degrees of freedom.
For example, in terms of the reversible extension of the erasure~(\ref{ERASE_ex}), the erased information is kept in ancilla $\mathcal M''$.
If all of the microscopic physical states in the whole universe were regarded as ``logical states,'' the information erasure could be logically reversible.
In contrast, the above-mentioned entropy transfer from the accessible to inaccessible degrees of freedom does not imply the thermodynamic irreversibility.
The second law of thermodynamics is relevant only to the reversibility and irreversibility of the whole universe from the macroscopic point of view, as discussed in Sec.~II.
We stress that the thermodynamic reversibility does not depend on the boundary between the logical and environmental degrees of freedom.
This shows the fundamental difference between the logical and thermodynamic reversibilities.
\
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=100mm]{Fig2.png}
\end{center}
\caption{
(a) Information erasure with the two-box memory (Fig.~1 (b)). The logical state of the memory is initially $0$ or $1$ with equal probability $1/2$, which corresponds to one bit of information. To erase information, we remove the wall, and compress the box from right quasi-statically and isothermally. The logical state of the memory then becomes $0$ with unit probability. While the information erasure is logically irreversible, this quasi-static erasure protocol achieves the Landauer bound~(\ref{Landauer1}), and is thermodynamically reversible.
(b) Information erasure in the memory with the binary symmetric potential. The erasure in (a) with the two-box memory is an idealization of this erasure.
(c) The time-reversal of the information erasure. The logical state is initially $0$ with unit probability. It finally becomes $0$ or $1$ with equal probability $1/2$, where the probability distribution is the same as that before the erasure. This process is the time-reversal of (a) in terms of the time evolution of the probability distribution.}
\end{figure}
A simple model of the information erasure that achieves the Landauer bound is illustrated as follows.
We consider the information erasure with the two-box memory shown in Fig.~2 (a), which is an idealization of the information erasure with the binary-potential model as shown in Fig.~2 (b).
Let $V/2$ be the volume of each box.
Before the erasure, the particle is in the left or right box with equal probability $1/2$. We then remove the wall without performing any work on the memory, and next compress the box from right to left quasi-statically and isothermally. The particle is then in the left box with unit probability. The work that is performed on the memory during the compression process is given by
\begin{equation}
W = \int_{V/2}^V \frac{\beta^{-1}}{V'} dV' = \beta^{-1} \ln 2,
\end{equation}
where we used the equation of states of the ideal gas with a single particle.
Therefore, the Landauer bound~(\ref{Landauer2}) is achieved in this erasure protocol.
By noting that $Q = -W$ holds in this case, the equality in (\ref{Landauer1}) is also achieved.
We show in Fig.~2 (c) the time-reversal of the foregoing protocol of the information erasure with the two-box memory. The logical state of the memory is initially $0$ with unit probability, which is the same as that after the information erasure. We then expand the left box quasi-statically and isothermally, so that the volume of the box becomes twice. During this process, we can extract $\beta^{-1} \ln 2$ of work, and $\beta^{-1}\ln 2$ of heat is absorbed by the memory from the bath. We next insert a wall at the center of the box, and the final logical state of the memory becomes $0$ or $1$ with equal probability $1/2$. While the final logical state may be different from the pre-erasure logical state for the individual processes, the probability distribution of the final logical states are the same as that of the pre-erasure logical states. In fact, the protocol shown in Fig.~2 (c) is the time-reversal of that in Fig.~2 (a) in terms of the ensemble; the time evolutions of their probability distributions are the time-reversal with each other. We stress that the thermodynamic reversibility is defined in terms of the ensemble.
We discuss how to achieve the quasi-static limit in the information erasure, in particular for the removal process of the barrier. In the case of the two-box memory shown in Fig.~2 (a), the velocity of the removal of the wall does not affect the probability distribution of the position of the particle, which is the same before and after the removal. Therefore, we can rapidly remove the wall even in the quasi-static limit. We note that, if the particle is quantum, its wave function can be affected by the velocity of the removal of the wall~\cite{SWKim}.
On the other hand, in the case of the binary-potential memory shown in Fig.~2 (b), the velocity of the removal of the barrier affects the probability distribution of the particle, because the barrier is not infinitely thin but has a finite width.
In this case, one needs infinitely slow change in the height of the barrier to achieve the quasi-static limit.
The higher the barrier is, the more time one needs to achieve the quasi-static limit, because the relaxation time over the two wells becomes exponentially larger as the barrier becomes higher.
This makes it hard to achieve the quasi-static limit with the binary-potential memory in practice.
\section{Thermodynamics of Computation}
In this section, we discuss the general formulation of thermodynamics of computation, and derive the general formulas that set the fundamental lower bounds of the work requirement and the heat emission during a computation. Moreover, we clarify the relationship between the thermodynamic and logical reversibilities in the general setup.
\subsection{General Setup}
We first consider the physical structure of the memory. In general, the logical states do not have one-to-one correspondence to the physical states of the memory; there may be a lot of microscopic physical states that correspond to a single logical state.
Let $\mathcal Y$ be the phase space of the memory, where each phase-space point $y \in \mathcal Y$ describes a microscopic physical state of the memory, and let $\mathcal M$ be the set of the possible input logical states.
To relate the physical states to logical ones, we decompose $\mathcal Y$ into subspaces $\mathcal Y_m$'s ($m \in \mathcal M$), where $\mathcal Y_m$ and $\mathcal Y_n $ do not overlap with each other for $m \neq n$, i.e., $\cup_{m \in \mathcal M} \mathcal Y_m = \mathcal Y$ and $\mathcal Y_m \cap \mathcal Y_n = \phi$ ($m \neq n$) with $\phi$ the empty set. If the phase-space point of the memory is in $\mathcal Y_m$ before the computation, we define that the input logical sate is $m$. We call $\mathcal Y_m$ an input logical subspace associated with $m$.
Figure~3 shows simple examples of phase-space separations with binary potentials of the memory. Figure~3 (a) shows a symmetric potential, which is the same as the memory illustrated in Fig.~1 (a). In this case, $0$ and $1$ correspond to the left and right well, respectively, and the phase-space volume of $\mathcal Y_0$ is the same as that of $\mathcal Y_1$.
Figure~3 (b) shows an asymmetric potential, where $0$ and $1$ correspond to the left and right well, respectively. In this case, the phase-space volume of $\mathcal Y_0$ is different from that of $\mathcal Y_1$.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=75mm]{Fig3.png}
\end{center}
\caption{
(a) Separation of phase-space $\mathcal Y$ into subspaces $\mathcal Y_0$ and $\mathcal Y_1$ with equal volumes, which respectively correspond to the left and right well of the binary symmetric potential.
(b) Separation of phase-space $\mathcal Y$ into subspaces $\mathcal Y_0$ and $\mathcal Y_1$ with different volumes, which respectively correspond to the left and right well of the asymmetric potential.}
\end{figure}
We consider probability distributions of the physical states in $\mathcal Y$.
Let $y \in \mathcal Y$ be the initial phase-space point before the computation, $P [y]$ be its probability, $m \in \mathcal M$ be the initial logical state such that $y \in \mathcal Y_m$, and $P [m]$ be its probability that satisfies
\begin{equation}
P[m] = \int_{y \in \mathcal Y_m} dy P[y].
\end{equation}
The probability distribution of $y$ under the condition of $m$ is written as $P [y|m]$, which takes nonzero value only if $y \in \mathcal Y_m$. We then have
\begin{equation}
P [y] = \sum_{m \in \mathcal M} P [y|m] P [m].
\end{equation}
We also consider output logical states and phase-space points after the computation. Let $\mathcal M'$ be the set of the output logical states, and $\mathcal Y'_{m'}$ be the logical subspace associated with $m' \in \mathcal M'$, where $\cup_{m' \in \mathcal M'} \mathcal Y'_{m'} = \mathcal Y$ and $\mathcal Y'_{m'} \cap \mathcal Y'_{n'} = \phi$ ($m' \neq n'$).
Let $y' \in \mathcal Y$ be the final phase-space point after the computation, $P' [y']$ be its probability, $m' \in \mathcal M'$ be the final logical state such that $y' \in \mathcal Y'_{m'}$, and $P' [m']$ be its probability.
The probability of $y'$ under the condition of $m'$ is written as $P' [y'|m']$, which satisfies
\begin{equation}
P'[y'] = \sum_{m' \in \mathcal M'} P' [y'|m'] P'[m'].
\end{equation}
\subsection{Entropy Balance}
We next consider the changes in entropies during the computation.
Before the computation, the Shannon entropy of the physical states is given by
\begin{equation}
S (\mathcal Y ) = -\int_{y \in \mathcal Y} dy P [y ] \ln P [ y ],
\end{equation}
and the logical entropy of the input is given by
\begin{equation}
H (\mathcal M) = -\sum_{m \in \mathcal M} P [m] \ln P [m].
\end{equation}
The conditional Shannon entropy inside $\mathcal Y_m$ is given by
\begin{equation}
S (\mathcal Y | m) = -\int_{y \in \mathcal Y_m} dy P [y | m] \ln P [ y| m],
\end{equation}
whose ensemble average over $m \in \mathcal M$ is
\begin{equation}
S (\mathcal Y | \mathcal M) := \sum_m P [m] S (\mathcal Y | m).
\end{equation}
Due to a general formula in probability theory~\cite{Cover-Thomas}, these entropies satisfy
\begin{equation}
S (\mathcal Y) = H (\mathcal M) + S (\mathcal Y | \mathcal M),
\label{entropy_balance1}
\end{equation}
which implies that the total entropy is given by the sum of the logical entropy of $\mathcal M$ and the average of the conditional entropies of $\mathcal Y_m$'s.
Intuitively, the fluctuation over the whole phase space can be decomposed into that over the logical states and that over the internal physical states in the individual logical subspaces.
We also consider the entropies after the computation in the same manner:
\begin{equation}
H' (\mathcal M') = -\sum_{m' \in \mathcal M'} P' [m'] \ln P' [m'],
\end{equation}
\begin{equation}
S' (\mathcal Y ) = -\int_{y' \in \mathcal Y} dy' P'[y' ] \ln P' [ y' ].
\end{equation}
\begin{equation}
S' (\mathcal Y | m') = -\int_{y' \in \mathcal Y'_{m'}} dy' P' [y' | m'] \ln P' [ y'| m'],
\end{equation}
and
\begin{equation}
S' (\mathcal Y | \mathcal M') := \sum_{m'} P' [m'] S' (\mathcal Y | m').
\end{equation}
They also satisfy the decomposition formula:
\begin{equation}
S' (\mathcal Y) = H' (\mathcal M') + S' (\mathcal Y | \mathcal M').
\label{entropy_balance2}
\end{equation}
Therefore, the total entropy change during the computation is decomposed as
\begin{equation}
\Delta S = \Delta H + \Delta S_{\rm in},
\label{total_entropy_change}
\end{equation}
where
\begin{equation}
\Delta S := S' (\mathcal Y) - S(\mathcal Y),
\label{entropy_change1}
\end{equation}
\begin{equation}
\Delta H := H'(\mathcal M') - H(\mathcal M),
\end{equation}
\begin{equation}
\Delta S_{\rm in} := S' (\mathcal Y | \mathcal M') - S(\mathcal Y | \mathcal M).
\end{equation}
Here, $\Delta S$ is the change in the total entropy, $\Delta H$ is the change in the logical entropy, and $\Delta S_{\rm in}$ is the average of the change in the physical entropy in the individual logical subspaces.
We note that $\Delta H \leq 0$ holds for logically irreversible computations, while $\Delta H = 0$ holds for logically reversible computations.
As a special case, we consider the information erasure.
We refer to one of the logical states in $\mathcal M'$ as the ``standard state,'' which we denote by $0 \in \mathcal M'$.
The information erasure is defined as the process in which the output logical state is in the standard state with unit probability (i.e., $P'[0] = 1$ and $P'[m'] = 0$ if $m' \neq 0$) for any probability distribution $P[m]$ of the input logical states.
In this case, $H' (\mathcal M') = 0$ holds by definition, and therefore $\Delta H = - H(\mathcal M')$. The change in the physical entropy is given by
\begin{equation}
\Delta S_{\rm in} = S' (\mathcal Y' | 0) - S(\mathcal Y | \mathcal M).
\end{equation}
We note that $\Delta S \neq - H(\mathcal M')$ if $\Delta S_{\rm in} \neq 0$.
To clarify the role of $\Delta S_{\rm in}$, we consider a simple case with $\mathcal M = \mathcal M'$ and $\mathcal Y_m = \mathcal Y'_m$. We assume that the initial and final distributions inside the individual logical subspaces are the same (i.e., $P[y|m] = P'[y|m]$ for any $y$ and $m$), and therefore $S(\mathcal Y| m) = S'(\mathcal Y|m)$ holds for any $m$. On the other hand, the probability distribution over $\mathcal M$ changes from $P[m]$ to $P'[m]$ during the computation. In this case, we have
\begin{equation}
\Delta S_{\rm in} = \sum_{m \in \mathcal M} (P'[m] - P[m]) S(\mathcal Y | m).
\end{equation}
If $S(\mathcal Y | m) $ does not depend on $m$, we have $\Delta S_{\rm in} = 0$ for any $P[m]$ and $P'[m]$.
For example, in the case of the symmetric memory in Fig.~3 (a), $S(\mathcal Y | 0) = S(\mathcal Y | 1) $ holds if the conditional probability distributions in the individual wells are in thermal equilibrium. If $\Delta S_{\rm in} = 0$, the internal fluctuations inside the individual logical subspaces do not contribute to the change in the total entropy (i.e., $\Delta S = \Delta H$).
In contrast, if $S(\mathcal Y | m) $ depends on $m$, $\Delta S_{\rm in} \neq 0$ in general.
For example, in the case of the asymmetric memory in Fig.~3 (b), $S(\mathcal Y | 0) $ and $S(\mathcal Y | 1) $ are different with each other.
If $\Delta S_{\rm in} \neq 0$, the internal fluctuations inside the individual wells contribute to the change in the total entropy, and therefore $\Delta S \neq \Delta H$.
The role of the asymmetry of the potential has been discussed in Refs.~\cite{Barkeshli,Norton,Maroney,Turgut,Sagawa-Ueda2}.
\subsection{Generalized Landauer Principle}
We now consider the second law of thermodynamics for computation.
We assume that the memory is attached to a single heat bath at inverse temperature $\beta$ during the computation.
The total entropy production is given by Eq.~(\ref{total_entropy_production}), which leads to
\begin{equation}
\Delta S_{\rm tot} = \Delta H + \Delta S_{\rm in} - \beta Q,
\label{total_entropy1}
\end{equation}
where $Q$ is the average heat that is absorbed by the memory from the bath. Therefore, the second law~(\ref{second1_1}) is given by
\begin{equation}
\Delta H \geq \beta Q - \Delta S_{\rm in},
\label{g_Landauer1}
\end{equation}
or equivalently,
\begin{equation}
-\beta Q \geq - \Delta H - \Delta S_{\rm in},
\label{g_Landauer2}
\end{equation}
which gives the fundamental lower bound of the heat emission into the bath during the computation.
We refer to inequalities~(\ref{g_Landauer1}) and (\ref{g_Landauer2}) as the generalized Landauer principle, and the right-hand side of (\ref{g_Landauer2}) as the generalized Landauer bound.
Several inequalities that are similar to or essentially equivalent to (\ref{g_Landauer2}) have been obtained in Refs.~\cite{Maroney,Turgut,Sagawa-Ueda2}.
In the special case of $\Delta S_{\rm in} = 0$, inequality~(\ref{g_Landauer2}) reduces to
\begin{equation}
-\beta Q \geq - \Delta H.
\end{equation}
In this case, if the computation is logically irreversible, the heat emission $-Q$ is nonnegative because of $\Delta H \geq 0$.
In the case of the information erasure, the heat emission is bounded as
\begin{equation}
-\beta Q \geq H (\mathcal M) - \Delta S_{\rm in},
\label{g_Landauer3}
\end{equation}
where $\Delta S_{\rm in}$ is the modification term to the original Landauer principle due to the change in the fluctuations inside the individual logical subspaces.
If $\Delta S_{\rm in} = 0$, we obtain
\begin{equation}
-\beta Q \geq H (\mathcal M),
\end{equation}
which is the conventional Landauer principle~\cite{Landauer,Shizume,Piechocinska,Esposito2}.
In the conventional setup of the Landauer principle with a symmetric memory such as Fig.~1 (a) and Fig.~3 (a), the decrease in the entropy of the logical states should be compensated for by the increase in the entropy of the bath, which is accompanied by at least $\beta^{-1} H (\mathcal M)$ of heat emission into the bath.
In contrast, in the case of the asymmetric memories such as Fig.~3 (b), the decrease in the entropy of the logical states can be compensated for not only by the increase in the entropy of the bath, but also by that inside the individual logical subspaces.
Here, the heat emission is determined only by the change in the entropy of the bath, but not by that inside the logical subspaces.
Therefore, the lower bound of the heat emission can be different from $\beta^{-1} H (\mathcal M)$ as shown in inequality~(\ref{g_Landauer3}); in particular, the heat emission can be smaller than $\beta^{-1} H (\mathcal M)$ if $\Delta S_{\rm in} > 0$.
To illustrate the above situation, we consider the situation with $\mathcal M = \mathcal M' = \{ 0, 1 \}$ and $\mathcal Y_m = \mathcal Y'_m$ for $m=0,1$. We assume that $S (\mathcal Y | 0) = S' (\mathcal Y | 0 )$ holds, which implies that the probability distribution inside the standard state is the same before and after the erasure. We then have
\begin{equation}
\eqalign{
\Delta S_{\rm in} &= S (\mathcal Y | 0) - \left( P[0] S (\mathcal Y | 0) + P[1] S (\mathcal Y | 1) \right) \\
&= P[1] \left( S (\mathcal Y | 0) - S (\mathcal Y | 1) \right).
}
\end{equation}
Therefore, if $P [1] \neq 0$ and $S (\mathcal Y | 0) > S (\mathcal Y | 1)$, we have $\Delta S_{\rm in} > 0$.
It is natural to assume that $S (\mathcal Y | 0) \neq S (\mathcal Y | 1)$ in asymmetric memories.
\subsection{Thermodynamic and logical reversibilities}
We now summarize the relationship between the thermodynamic and logical reversibilities in the general setup.
On the basis of the argument in Sec.~5.2, the conventional Landauer principle discussed in Sec.~4 needs to be modified in general, and the generalized Landauer principle is stated as follows: \\
\\
{\it The decrease in the entropy of the logical states during a logically irreversible computation can be compensated for not only by the increase in the entropy of the heat bath, but also by that of the physical states inside the individual logical subspaces, where only the former determines the amount of the heat emission.}\\
\
Since the generalized Landauer principle (\ref{g_Landauer1}) is equivalent to the second law of thermodynamics~(\ref{second1}), the equality in (\ref{g_Landauer1}) can be achieved in the quasi-static limit (i.e., in the case of the quasi-static computation) where the total entropy production is zero. In this case, the computational process is thermodynamically reversible. Therefore, we conclude that:\\
\\
{\it Any logically irreversible computation can be performed in a thermodynamically reversible manner in the quasi-static limit, and therefore, the thermodynamic and logical reversibilities are not equivalent with each other.}\\
\\
These are generalizations of the arguments in Sec.~4. We again stress that our observations here do not contradict the conventional Landauer principle; the logical reversibility is related to the change in the entropy of the logical states, while the thermodynamic reversibility is related to the change in the entropy of the whole universe that consists of the logical and physical states of the memory and the physical states of the heat bath.
\subsection{Work requirement for computation}
We next consider the thermodynamic work needed for computation.
Let $E [y]$ with $y \in \mathcal Y$ be the initial Hamiltonian of the memory. We define the conditional free energy in logical subspace $\mathcal Y_m$ as
\begin{equation}
F_m := -\beta^{-1} \int_{y \in \mathcal Y_m} dy e^{-\beta E[y]}.
\end{equation}
We assume that the memory is initially in the conditional canonical distribution in the individual logical subspaces, which is given by
\begin{equation}
P [y | m ] = e^{\beta (F_m - E [y])}
\end{equation}
for $y \in \mathcal Y_m$, and otherwise $P [y | m ] = 0$. The total probability distribution is then given by
\begin{equation}
P [y] = \sum_{m \in \mathcal M} P[m] \chi (y, m) e^{\beta (F_m - E [y])},
\end{equation}
where $\chi (y, m)$ is the characteristic function which takes one if $y \in \mathcal Y_m$ and zero otherwise.
The conditional entropy is then given by
\begin{equation}
S(\mathcal Y | m) = \beta (E_m - F_m),
\end{equation}
where
\begin{equation}
E_m := \int_{y \in \mathcal Y_m} dy P[y|m] E[y].
\end{equation}
Therefore, we obtain
\begin{equation}
S(\mathcal Y | \mathcal M) = \beta ( E - F ),
\end{equation}
where
\begin{equation}
E := \sum_{m \in \mathcal M} P[m] E_m = \int_{y \in \mathcal Y} dy P[y] E[y]
\end{equation}
is the average energy, and
\begin{equation}
F := \sum_{m \in \mathcal M} P[m] F_m
\end{equation}
is the average free energy.
We also consider the final Hamiltonian $E'[y']$ after the computation. Correspondingly, every argument about the final state is parallel to that about the initial one in the previous paragraph. To show what a quantity is about the final states after the computation, we use notation of prime.
We then have
\begin{equation}
\Delta S_{\rm in} = \beta ( \Delta E - \Delta F ),
\label{Sint_F}
\end{equation}
where $\Delta E := E' - E $ is the change in the average energy, and $\Delta F := F' - F$ is the change in the average free energy.
We note that, if the final distribution is different from the conditional canonical distribution, we can show an inequality:
\begin{equation}
\Delta S_{\rm in} \leq \beta ( \Delta E - \Delta F ),
\label{entropy}
\end{equation}
where the equality in (\ref{entropy}) is achieved if and only if the output state is in the conditional canonical distribution.
Therefore, the total entropy production~(\ref{total_entropy1}) satisfies
\begin{equation}
\Delta S_{\rm tot} \leq \beta ( W - \Delta F) + \Delta H,
\label{total_entropy2}
\end{equation}
where $W$ is the work performed on the memory, and we used the first law of thermodynamics
\begin{equation}
\Delta E = Q + W.
\end{equation}
Therefore, by applying the second law~(\ref{second1}), we obtain the generalized Landauer principle in terms of the work:
\begin{equation}
\beta W \geq - \Delta H + \beta \Delta F,
\label{g_LandauerW1}
\end{equation}
which gives the minimal work needed for the computation.
In the case of information erasure, inequality~(\ref{g_LandauerW1}) reduces to
\begin{equation}
\beta W_{\rm eras} \geq H (\mathcal M) + \beta \Delta F,
\label{g_Landauer_work}
\end{equation}
where $\Delta F$ is the modification term to the conventional Landauer principle~(\ref{Landauer2}). As shown in Eq.~(\ref{Sint_F}), there are two contributions to $\Delta F$: the changes in the entropy and the energy inside the individual logical subspaces, where the former is $\Delta S_{\rm in}$ and the latter is $\Delta E$.
We can also regard $W - \Delta F$ as the energy cost needed for the information erasure, whose lower bound is given by the conventional Landauer bound $\beta^{-1} H(\mathcal M)$.
Several inequalities that are similar to or essentially equivalent to the generalized Landauer principle (\ref{g_Landauer_work}) have been obtained in Refs.~\cite{Maroney,Turgut,Sagawa-Ueda2}.
In general, we call the memory as symmetric if $F_m$ does not depend on $m$. The memory shown in Fig.~3 (a) is symmetric in this sense because of $F_0 =F_1$. In contrast, the memory shown in Fig.~3 (b) is asymmetric because of $F_0 \neq F_1$.
When $\mathcal M = \mathcal M'$, $\mathcal Y_m = \mathcal Y'_{m'}$ with $m=m'$, $F_m = F'_{m'}$ with $m=m'$, and $F_m$ does not depend on $m$, then we have $\Delta F=0$ for any $P[m]$ and $P'[m]$.
As a simple example of information erasure with an asymmetric memory, we consider a model of the memory shown in Fig.~4 (a) that is in contact with a heat bath at inverse temperature $\beta$. If the barrier is much higher than the thermal fluctuation, this model is idealized by a memory with two boxes with different volumes as shown in Fig.~4 (b). Let $t:1-t$ ($0 < t < 1$) be the ratio of the volumes of the boxes, which characterizes the ratio of the phase-space volumes inside the individual logical states.
If the memory is symmetric, $t = 1/2$.
We assume that $\mathcal M = \mathcal M' = \{ 0,1 \}$ and $\mathcal Y_m = \mathcal Y'_m$ with $m=0,1$, and write $p := P[0]$. We note that $t$ is in general different from $p$, because the initial state is not necessarily in global thermal equilibrium over the whole phase space.
The protocol of the information erasure is as follows (see also Fig.~4 (c)). We first move the wall quasi-statically to the position where the ratio of the two volumes is given by $p:1-p$.
During this process, we perform the work of $\beta^{-1} [ p \ln (t/p) + (1-p) \ln (1-t/1-p)]$ on average. We then remove the wall without any work. We next compress the box from the right quasi-statically, and the ratio of the two volumes is given by $t:1-t$ in the final stage, where the logical state is ``$0$'' with unit probability. During this compression, we perform the work of $-\beta^{-1}\ln t$. The total work performed on the memory is then given by
\begin{equation}
\beta W_{\rm eras} = - p \ln p - (1-p) \ln (1-p) + (1-p \ln \frac{1-t}{t}).
\label{erasure_asymmetric}
\end{equation}
We note that the initial logical entropy is given by $H(\mathcal M) = - p \ln p - (1-p ) \ln (1-p)$. The free-energy difference between two logical states are given by $F_0 - F_1 = \beta^{-1} \ln [(1-t)/t]$, and therefore $\Delta F = \beta^{-1} (1-p) \ln [(1-t)/t]$. Therefore, Eq.~(\ref{erasure_asymmetric}) achieves the generalized Landauer bound in~(\ref{g_Landauer_work}), which implies that this protocol of the information erasure is thermodynamically reversible.
\
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=100mm]{Fig4.png}
\end{center}
\caption{
(a) An asymmetric memory with phase-space separation $\mathcal Y_0$ and $\mathcal Y_1$ with different volumes.
(b) An asymmetric memory with two boxes with different volumes, which is an idealization of the asymmetric memory in (a).
(c) Information erasure with the asymmetric memory (b). The logical state of the memory is initially $0$ or $1$ with equal probability $1/2$, which corresponds to one bit of information. We move the wall quasi-statically and isothermally to the center of the box, remove the wall, and compress the box from the right quasi-statically and isothermally so that the final logical state is $0$ with unit probability. This erasure protocol achieves the generalized Landauer bound in~(\ref{g_Landauer_work}), and therefore is thermodynamically reversible.}
\end{figure}
We next consider a simple example of the information erasure with $\mathcal M \neq \mathcal M'$. We assume that $\mathcal M = \{ 0,1 \}$ and that the memory is initially binary symmetric as shown in Fig.~1 (a), which is idealized by the two-boxes memory with equal volumes as shown in Fig.~1 (b). We then assume that $\mathcal M' = \{ 0 \}$ in the final stage, i.e., the output logical state is only the standard state; the corresponding potential model is shown in Fig.~5 (a). This is idealized by a memory with a single box as shown in Fig.~5 (b). For simplicity, we set $P[0] = 1/2$. In this case, the information erasure is just the removal of the wall as shown in Fig.~5 (c), and therefore $W_{\rm eras} = 0$. This erasure protocol achieves the generalized Landauer bound in~(\ref{g_Landauer_work}), since $H(\mathcal M) = \ln 2$, $F_0 - F'_0 = F_1 - F'_0 = \beta^{-1} \ln 2$, and $\Delta F = - \beta^{-1} \ln 2$. Therefore, this erasure protocol is thermodynamically reversible. In fact, if we insert a wall to the center of the single box after the erasure, the probability distribution of the logical states returns to the initial one, where a particle is in one of the two boxes with equal probability $1/2$.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=100mm]{Fig5.png}
\end{center}
\caption{
(a) A memory with a single-well potential, whose logical state is only the standard state $m'=0$.
(b) A memory with a single box, which is an idealization of the single-well memory in (a).
(c) Information erasure from the symmetric two-box memory [Fig.~1 (b)] to the single-box memory (b). The initial logical state of the memory is $m=0$ or $1$ with equal probability $1/2$, which corresponds to one bit of information. We then only remove the wall so that the final logical state becomes $m'=0$, which can be regarded as the information erasure that achieves the generalized Landauer bound in~(\ref{g_Landauer_work}).
(d) Information erasure where the initial logical state is $0$ ($1$) with probability $p$ ($1-p$). We first move the wall quasi-statically and isothermally so that the ratio of the volumes of the boxes becomes $p:1-p$. We then remove the wall, and the final logical state becomes $m'=0$ with unit probability. This erasure also achieves the generalized Landauer bound in~(\ref{g_Landauer_work}).}
\end{figure}
If $p := P[0] \neq 1/2$ in the above model, we quasi-statically move the wall to the position where the ratio of the volumes $p:1-p$, and then remove the wall (see all Fig.~5 (d)). In this case, the work performed on the memory is given by $\beta W_{\rm eras} = - \ln 2 - p \ln p - (1-p) \ln (1-p)$. Since $H(\mathcal M) = - p \ln p - (1-p) \ln (1-p)$ and $\beta \Delta F = - \ln 2$, this erasure protocol also achieves the generalized Landauer bound in~(\ref{g_Landauer_work}).
\section{Two Memories}
In this section, we consider the cases that the memory consists of two sub-memories. In particular, we discuss measurement and feedback control between the two sub-memories, which constitute the typical setup of Maxwell's demon~\cite{Demon}.
\subsection{General argument}
We assume that any logical state in $\mathcal M$ is a pair of two logical states of sub-memories that we refer to as S and D; any input logical state $m \in \mathcal M$ is written as $m = (s,d)$ where $s$ and $d$ are the logical states of memories S and D, respectively.
Let $\mathcal M^{\rm S}$ ($\mathcal M^{\rm D}$) be the set of logical states of S (D) with $\mathcal M = \mathcal M^{\rm S} \times \mathcal M^{\rm D}$, where $s \in \mathcal M^{\rm S}$ and $d \in \mathcal M^{\rm D}$.
We also consider the physical states of the memories. Let $\mathcal Y^{\rm S}$ ($\mathcal Y^{\rm D}$) be the set of physical states of S (D) with $\mathcal Y = \mathcal Y^{\rm S} \times \mathcal Y^{\rm D}$.
Let $y := (u,v)$ for any initial physical state $y \in \mathcal Y$ with $u \in \mathcal Y^{\rm S}$ and $v \in \mathcal Y^{\rm D}$.
Let $\mathcal Y_m := \mathcal Y_s^{\rm S} \times \mathcal Y_d^{\rm D}$.
The conditional probability of $(u,v)$ under the condition of $(s,d)$ is given by $P[u,v | s,d]$ that takes nonzero value only if $u \in \mathcal Y_s^{\rm S}$ and $v \in \mathcal Y_d^{\rm D}$. By applying the argument in Sec.~5.1 to the present situation, we have
\begin{equation}
P[u,v] = \sum_{(s,d) \in \mathcal M} P[u,v | s,d] P[s,d],
\end{equation}
whose Shannon entropy is given by Eq.~(\ref{entropy_balance1}).
The mutual information plays a crucial role in the presence of two memories; it characterizes the correlation between them~\cite{Shannon,Cover-Thomas}. The mutual information between the physical states and that between the logical states are respectively given by
\begin{equation}
I (\mathcal Y^{\rm S} : \mathcal Y^{\rm D}) = S (\mathcal Y^{\rm S}) + S(\mathcal Y^{\rm D}) - S(\mathcal Y),
\end{equation}
\begin{equation}
I (\mathcal M^{\rm S} : \mathcal M^{\rm D}) = H (\mathcal M^{\rm S}) + H(\mathcal M^{\rm D}) - H(\mathcal M).
\end{equation}
The mutual information between the internal states is
\begin{equation}
I (\mathcal Y^{\rm S} : \mathcal Y^{\rm D} | s, d) = S (\mathcal Y^{\rm S} | s) + S(\mathcal Y^{\rm D} | d) - S(\mathcal Y | s,d),
\end{equation}
whose ensemble average over $(s,d)$ is given by
\begin{equation}
\eqalign{
I (\mathcal Y^{\rm S} : \mathcal Y^{\rm D} | \mathcal M) &:= \sum_{(s,d) \in \mathcal M } I (\mathcal Y^{\rm S} : \mathcal Y^{\rm D} | s, d)
P[s,d]\\
&= S (\mathcal Y^{\rm S} | \mathcal M^{\rm S}) + S(\mathcal Y^{\rm D} | \mathcal M^{\rm D}) -S(\mathcal Y | \mathcal M).
}
\end{equation}
Therefore, we obtain
\begin{equation}
I (\mathcal Y^{\rm S} : \mathcal Y^{\rm D}) = I (\mathcal M^{\rm S} : \mathcal M^{\rm D}) + I (\mathcal Y^{\rm S} : \mathcal Y^{\rm D} | \mathcal M),
\label{two_decom1}
\end{equation}
which implies that the total correlation between the two memories can be decomposed into the correlation between their logical states and that between their internal physical states in the individual logical subspaces. We note that $I (\mathcal Y^{\rm S} : \mathcal Y^{\rm D} | \mathcal M) = 0$ if $P[u,v | s,d] = P [u |s,d] P[v|s,d]$.
We also note that
\begin{equation}
S (\mathcal Y^{\rm S}) = H(\mathcal M^{\rm S}) + S(\mathcal Y^{\rm S} | \mathcal M^{\rm S}),
\label{two_decom2}
\end{equation}
\begin{equation}
S (\mathcal Y^{\rm D}) = H(\mathcal M^{\rm D}) + S(\mathcal Y^{\rm D} | \mathcal M^{\rm D}).
\label{two_decom3}
\end{equation}
The sum of Eqs.~(\ref{two_decom1}), (\ref{two_decom2}), and (\ref{two_decom3}) leads to Eq.~(\ref{entropy_balance1}).
We also consider the probability distributions, the Shannon entropy, and the mutual information after the computation in the parallel manner to those before the computation; for example, we write $\mathcal M' := \mathcal M'_{\rm S} \times \mathcal M'_{\rm D}$, $m' := (s',d')$ with $s' \in \mathcal M'^{\rm S}$ and $d' \in \mathcal M'^{\rm D}$. To show what a quantity is about the states after computation, we use notation of prime.
The change in the total Shannon entropy is then given by
\begin{equation}
\Delta S = \Delta H^{\rm S} + \Delta H^{\rm D} - \Delta I + \Delta S_{\rm in}^{\rm S} + \Delta S_{\rm in}^{\rm D} - \Delta I_{\rm in},
\end{equation}
where
\begin{equation}
\Delta S := S'(\mathcal Y') - S(\mathcal Y),
\end{equation}
\begin{equation}
\Delta H^{\rm S} := H'(\mathcal M'^{\rm S}) - H(\mathcal M^{\rm S}), \ \ \Delta H^{\rm D} := H'(\mathcal M'^{\rm D}) - H(\mathcal M^{\rm D}),
\end{equation}
\begin{equation}
\Delta I := I' (\mathcal M'^{\rm S} : \mathcal M'^{\rm D}) - I (\mathcal M^{\rm S} : \mathcal M^{\rm D}),
\end{equation}
\begin{equation}
\Delta S_{\rm in}^{\rm S} := S'(\mathcal Y'^{\rm S} | \mathcal M'^{\rm S} ) - S(\mathcal Y^{\rm S} | \mathcal M^{\rm S}), \ \ \Delta S_{\rm in}^{\rm D} := S'(\mathcal Y'^{\rm D} | \mathcal M'^{\rm D}) - S(\mathcal Y^{\rm D} | \mathcal M^{\rm D}),
\end{equation}
\begin{equation}
\Delta I_{\rm in} := I' (\mathcal Y'^{\rm S} : \mathcal Y'^{\rm D} | \mathcal M') - I (\mathcal Y^{\rm S} : \mathcal Y^{\rm D} | \mathcal M).
\end{equation}
Here, $\Delta H^{\rm S}$ and $\Delta H^{\rm D}$ describe the changes in the logical entropies, $\Delta I$ describes the change in the logical correlation, $\Delta S_{\rm in}^{\rm S}$ and $\Delta S_{\rm in}^{\rm D}$ describe the change in the internal physical entropies in the individual logical subspaces, and $\Delta I_{\rm in}$ describes the change in the internal correlation between the individual logical subspaces.
Therefore, the total entropy production is given by
\begin{equation}
\Delta S_{\rm tot} = \Delta H^{\rm S} + \Delta H^{\rm D} - \Delta I + \Delta S_{\rm in}^{\rm S} + \Delta S_{\rm in}^{\rm D} - \Delta I_{\rm in} - \beta Q.
\label{two_total}
\end{equation}
We next assume that the energies of two memories are additive before the computation, $E[y] = E^{\rm S}[u] + E^{\rm D} [v]$, where $E^{\rm S}[u] $ ($E^{\rm D} [v]$) is the Hamiltonian of S (D).
This assumption implies that the interaction Hamiltonian between the memories is negligible before the computation.
The corresponding free energies are given by
\begin{equation}
F_s^{\rm S} := - \beta^{-1} \ln \int_{u \in \mathcal Y_s} e^{- \beta E^{\rm S}[u]}, \ \ F_d^{\rm D} := - \beta^{-1} \ln \int_{v \in \mathcal Y_d} e^{- \beta E^{\rm D}[v]}.
\end{equation}
In this case, the conditional distribution is given by
\begin{equation}
P[u,v |s,d] = P[u|s]P[v|d],
\end{equation}
where, if $u \in \mathcal Y_s^{\rm S}$ and $v \in \mathcal Y_d^{\rm S}$,
\begin{equation}
P[u|s] = e^{\beta (F_s^{\rm S} - E^{\rm S} [u])}, \ \ P[v|d] = e^{\beta (F_d^{\rm D} - E^{\rm D} [v])}.
\end{equation}
We also consider the energies in the memories after the computation, where the energies are also assumed to be additive. The argument about energy and the free energy after the computation is parallel to that before the computation. To show what a quantity is about after the computation, we use notation of prime. We note that the energies are not necessarily additive during the computation, since the interaction Hamiltonian between the memories may become nonzero during the computation.
If the initial probability distribution is the conditional probability distribution, the total entropy production satisfies Eq.~(\ref{total_entropy2}). Therefore, we obtain the lower bound of the work that is needed for the computation with two memories:
\begin{equation}
W \geq - \Delta H + \Delta F,
\end{equation}
where
\begin{equation}
\Delta H = \Delta H^{\rm S} + \Delta H^{\rm D} - \Delta I,
\end{equation}
\begin{equation}
\Delta F := \Delta F^{\rm S} + \Delta F^{\rm D},
\end{equation}
\begin{equation}
\Delta F^{\rm S} := \sum_{s \in \mathcal M^{\rm S}} P[s] F_s^{\rm S}, \ \ \Delta F^{\rm D} := \sum_{d \in \mathcal M^{\rm D}} P[d] F_d^{\rm D}.
\end{equation}
In the special case of $\Delta H^{\rm S} = \Delta H^{\rm D} = 0$ and $\Delta F = 0$, we have
\begin{equation}
- W \leq - \Delta I,
\end{equation}
which is the work extraction by using the correlation, which can be demonstrated by Maxwell's demon~\cite{Demon}.
\subsection{Measurement and feedback}
In this subsection, we consider measurement and feedback as a special case of the argument in Sec.~6.1 with the assumption of the initial and final conditional canonical distributions.
We first consider a measurement process, where memory D performs a measurement on memory S. In other words, the information in memory S is copied to memory D.
The logical state of D is initially the standard state $0^{\rm D}$ with unit probability, and the logical entropy of S is initially given by $H^{\rm S}$.
We assume that the initial logical entropy in D and the initial mutual information between the logical states are both zero.
We also assume that the two memories are in the conditional canonical distributions before and after the measurement.
The two memories interact with each other, and make a correlation between the logical states.
We assume that the logical state of S does not change in time during the measurement.
When both of the two memories are one bit, a typical example of such measurement is given by CNOT~(\ref{CNOT_measurement}), where the measurement is error-free (i.e., the copy of information is perfect).
We denote by $I$ the correlation between the logical states after the measurement, and denote by $H^{\rm D}$ the logical entropy of D after the measurement.
We note that the mutual information satisfies the following inequalities~\cite{Shannon,Cover-Thomas}:
\begin{equation}
0 \leq I \leq H^{\rm S}, \ \ 0 \leq I \leq H^{\rm D}.
\end{equation}
Applying Eq.~(\ref{total_entropy2}) to the measurement, the total entropy production is given by
\begin{equation}
\Delta S_{\rm tot} = \beta (W_{\rm meas} - \Delta F_{\rm meas}^{\rm D}) + H^{\rm D} - I,
\label{tot_meas}
\end{equation}
where $W_{\rm meas}$ is the work performed on the memories, and $\Delta F_{\rm meas}^{\rm D}$ is the change in the average free energy of D.
Therefore, we obtain
\begin{equation}
W_{\rm meas} \geq I - H^{\rm D} + \Delta F_{\rm meas}^{\rm D},
\label{meas_cost1}
\end{equation}
which gives the minimal work needed for the measurement.
Several inequalities that are essentially equivalent to (\ref{meas_cost1}) have been obtained in Refs.~\cite{Sagawa-Ueda2,Sagawa-Ueda2012,Sagawa-Ueda-NJP}.
If $ I = H^{\rm D}$ holds and memory D is symmetric (i.e., $\Delta F_{\rm meas}^{\rm D} = 0$), inequality~(\ref{meas_cost1}) reduces to
\begin{equation}
W_{\rm meas} \geq 0,
\label{bound_Bennett}
\end{equation}
which is the bound discussed by Bennett~\cite{Bennett1}. Since the energies of the memories do not change in time during the measurement if the memory is symmetric, inequality~(\ref{bound_Bennett}) is equivalent to
\begin{equation}
- Q_{\rm meas} \geq 0,
\label{bound_Bennett1}
\end{equation}
where $Q_{\rm meas}$ is the heat absorbed by the memories during the measurement. Inequalities~(\ref{bound_Bennett}) and (\ref{bound_Bennett1}) imply that the lower bounds of the work requirement and the heat emission during the measurement are both zero.
We consider a simple example where the logical states of S and D before and after the measurement are all one bit, which is a conventional setup of the measurement~\cite{Bennett1}.
We assume that S is symmetric (i.e., $F^{\rm S}_0 = F^{\rm S}_1$) before and after the measurement; in this case, memory S can be modeled by two boxes (see also Fig.~6 (a)).
Before the measurement, the logical state $(s,d)$ is $(0,0)$ or $(1,0)$ with equal probability $1/2$.
The measurement protocol is given by CNOT where S is the control bit and D is the target bit, and then the final logical state is $(0,0)$ or $(1,1)$ with equal probability $1/2$. In this case, $I = H^{\rm D} = \ln 2$ and $\Delta F_{\rm meas}^{\rm D} = 0$. Therefore, $W_{\rm meas} = 0$ and $- Q_{\rm meas} = 0$ are achieved in the quasi-static limit, where the measurement is thermodynamically reversible as well as logically reversible. We summarize the thermodynamic and logical reversibilities for the conventional setup of measurement in Table 2, which is contrastive to Table 1 for the information erasure.
\begin{table}
\caption{Summary for the conventional setup of measurement.}
\begin{center}
\begin{tabular}{| l || l | l | }\hline
{} & Quasi-static & Finite-velocity \\ \hline \hline
Thermodynamically & reversible & irreversible \\ \hline
Logically & reversible & reversible \\ \hline
Heat emission & $ = 0 $ & $> 0 $ \\ \hline
\end{tabular}
\end{center}
\end{table}
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=100mm]{Fig6.png}
\end{center}
\caption{
(a) Model of measurement with two symmetric memories. The two memories are initially not correlated. The logical state of D quasi-statically changes only if the logical state of S is initially in $1$, and the memories finally have one bit of correlation. This measurement achieves the equality in~(\ref{bound_Bennett}), and is thermodynamically reversible.
(b) Model of feedback control, where the two symmetric memories initially share one bit of correlation. The logical state of S quasi-statically changes only if the logical state of D is initially in $1$. Due to the feedback control, the initial correlation vanishes and the logical entropy of S is decreased by one $\ln 2$. This feedback protocol achieves the equality in~(\ref{feedback_cost1}), and is thermodynamically reversible.
(c) Model of feedback control, where the two symmetric memories initially share one bit of correlation. The left (right) box of S is expanded quasi-statically and isothermally if the initial logical state of D is $0$ ($1$), so that the final logical subspace of S is the whole phase space. Due to the feedback control, the initial correlation vanishes and $\beta^{-1}\ln 2$ of work is extracted from S, which is equivalent to the case of the Szilard engine. This feedback protocol also achieves the equality in~(\ref{feedback_cost1}), and is thermodynamically reversible.}
\end{figure}
We next consider a feedback control process followed by the measurement, where memory D performs feedback control on memory S.
This situation can be regarded as a typical setup of Maxwell's demon (see also Sec.~6.3), where D is the demon and S is the engine to be controlled.
We assume that the two memories are in the conditional canonical distributions before and after the feedback control.
The logical state of S changes depending on the logical state of D after the measurement.
We assume that memory D does not evolve in time during the feedback control.
For simplicity, we assume that the final correlation between the logical states between the memories is zero after the feedback control, such that the change in the logical mutual information is given by $-I$. We also assume that the output logical state of S is the standard state $0^{\rm S}$ with unit probability after feedback control such that the change in the logical entropy of S is given by $-H^{\rm S}$.
When both of the two memories are one bit, a typical example of such feedback control is also given by CNOT~(\ref{CNOT_feedback}).
Applying Eq.~(\ref{total_entropy2}) to the feedback control, the total entropy production is given by
\begin{equation}
\Delta S_{\rm tot} = \beta (W_{\rm fb} - \Delta F_{\rm fb}^{\rm S}) - H^{\rm S} + I,
\label{tot_feedback}
\end{equation}
where $W_{\rm fb}$ is the work performed on the memories, and $\Delta F_{\rm fb}^{\rm S}$ is the change in the average free energy of S.
Therefore, we obtain
\begin{equation}
W_{\rm fb} \geq - I + H^{\rm S} + \Delta F_{\rm fb}^{\rm S},
\label{feedback_cost1}
\end{equation}
which gives the minimal work needed for the feedback control.
Several inequalities that are essentially equivalent to (\ref{feedback_cost1}) has been obtained in Refs.~\cite{Sagawa-Ueda1,Sagawa-Ueda3,Sagawa-Ueda4,Sagawa-Ueda2012,Sagawa-Ueda-NJP}; in these previous works, $H^{\rm S} + \Delta F_{\rm fb}^{\rm S}$ is denoted by, for example, just $\Delta F_{\rm fb}^{\rm S}$.
We consider a simple situation where the logical states of S and D before and after feedback are all one bit.
We assume that S is symmetric (i.e., $F^{\rm S}_0 = F^{\rm S}_1$) before and after the feedback (see also Fig.~6 (b)).
Before feedback, the logical state $(s,d)$ is $(0,0)$ or $(1,1)$ with equal probability $1/2$. The feedback protocol is given by CNOT that is logically reversible, where D is the control bit and S is the target bit, and then the final logical state is $(0,0)$ or $(0,1)$ with equal probability $1/2$. In this case, $I = H^{\rm S} = \ln 2$ and $\Delta F_{\rm fb}^{\rm S} = 0$ hold. Therefore, inequality~(\ref{feedback_cost1}) reduces to $W_{\rm fb} \geq 0$. In the quasi-static limit, $W_{\rm fb} = 0$ is achieved, which implies that we can reduce the entropy of S (i.e., $H^{\rm S} = \ln 2$) without performing any positive amount of work on S. We note that, in the conventional thermodynamics, we need a positive amount of work to isothermally decrease the entropy. In contrast, in the present situation, the mutual information plays the role of the resource to decrease the entropy of S.
We next consider another situation that there is a single logical state $0^{\rm S}$ in S after the feedback, which corresponds to the whole phase-space of S (see also Fig~6 (c)). In this case, $I = H^{\rm S} = \ln 2$ and $\Delta F_{\rm fb}^{\rm S} = - \beta^{-1} \ln 2$ hold. Therefore, inequality~(\ref{feedback_cost1}) reduces to $W_{\rm fb} \geq - \beta^{-1} \ln 2$. In the quasi-static limit, $W_{\rm fb} = - \beta^{-1} \ln 2$ is achieved, where we extract $\beta^{-1} \ln 2$ of work by the feedback control; the mutual information is the resource of the work extraction. This situation is equivalent to the case of the conventional Szilard engine~\cite{Szilard}.
\subsection{On Maxwell's demon}
We consider Maxwell's demon as a special example of our general argument.
A typical situation of Maxwell's demon consists of measurement and feedback control.
In the setup of Sec.~6.2, memory D plays the role of the demon and memory S is the system to be measured and controlled.
By summing up inequalities~(\ref{meas_cost1}) and (\ref{feedback_cost1}), we obtain
\begin{equation}
W_{\rm meas} + W_{\rm fb} \geq H^{\rm S} - H^{\rm D} + \Delta F_{\rm meas}^{\rm S} + \Delta F_{\rm fb}^{\rm D},
\label{total_work1}
\end{equation}
where the contribution of the mutual information vanishes in the right-hand side.
If we perform the information erasure from D after the feedback control, the total work is given by
\begin{equation}
W_{\rm meas} + W_{\rm fb} + W_{\rm eras} \geq H^{\rm S} + \Delta F_{\rm meas}^{\rm S},
\label{total_work2}
\end{equation}
where we used the generalized Landauer principle~(\ref{g_Landauer_work}).
Here, $H^{\rm S} + \Delta F_{\rm meas}^{\rm S}$ on the right-hand side can be regarded as an effective free energy of S, which vanishes in, for example, the case of the Szilard engine discussed in Sec.~6.2. If $H^{\rm S} + \Delta F_{\rm meas}^{\rm S} = 0$ holds, inequality~(\ref{total_work2}) reduces to
\begin{equation}
W_{\rm meas} + W_{\rm fb} + W_{\rm eras} \geq 0,
\end{equation}
which implies that we cannot extract any work from the entire process.
We stress that, before the information erasure, the total entropy productions~(\ref{tot_meas}) and (\ref{tot_feedback}) are always nonnegative (i.e., $\Delta S_{\rm tot}\geq 0$) for the individual processes of measurement and feedback.
This confirms that measurement and feedback control are individually consistent with the second law of thermodynamics, without considering the information erasure.
As shown in Sec.~6.2, both of measurement and feedback are logically reversible in typical situations, and can be thermodynamically reversible in the quasi-static limit.
On the other hand, the information erasure is logically irreversible, while it can be thermodynamically reversible in the quasi-static limit.
Therefore, the entire process of Maxwell's demon including the erasure is logically irreversible, but can be thermodynamically reversible in the quasi-static limit, where $\Delta S_{\rm tot} = 0$ holds for the individual processes of the measurement, feedback control, and information erasure.
\section{Conclusions}
In this paper, we have discussed the relationship between computation and the second law of thermodynamics. In particular, we have clarified the fundamental relationship between the thermodynamic and logical reversibilities.
In Sec.~2, we have discussed the concept of thermodynamic reversibility. A physical process is reversible if and only if its time-reversal is not prohibited by the second law of thermodynamics. In more precise, a physical process is thermodynamically reversible if the total entropy production in the whole universe (\ref{total_entropy_production}) is zero during the process (i.e., $\Delta S_{\rm tot} = 0$). The total entropy production consists of the increases in the entropy of the system and that of the heat bath.
In Sec.~3, we have discussed the concept of logical reversibility. A computational process is logically reversible if and only if it is an injection. The Shannon entropy of the logical states decreases if a computation is irreversible, while it does not change if a computation is reversible. A typical example of irreversible computation is the information erasure.
In Sec.~4, we have discussed the conventional setup of the information erasure with a binary symmetric memory (Fig.~1). On the basis of the second law of thermodynamics, we have confirmed the conventional Landauer principle, which states that at least $\beta^{-1} \ln 2$ of heat should be emitted into the heat bath during the information erasure of one bit of information. We have clarified that the information erasure can be thermodynamically reversible in the quasi-static limit, where the heat emission equals $\beta^{-1} \ln 2$. The crucial observation here is that the thermodynamic reversibility is related to the entropy production in the whole universe, while the logical reversibility is related only to the change in the entropy of the logical states. Such a thermodynamically reversible erasure is illustrated in Fig. 2.
In Sec.~5, we have discussed the general theory of thermodynamics of computation. In particular, we have derived a generalized Landauer principle (\ref{g_Landauer2}). In the case of an asymmetric memory such as Fig.~3 (b) and Fig.~4 (a), the lower bound of the heat emission should be modified from the original Landauer bound; the second term on the right-hand side of (\ref{g_Landauer2}) originates from the asymmetry of the memory. In fact, the amount of the heat emission is determined by the increase in the entropy of the heat bath, while the change in the entropy of the logical states of the memory can be compensated for not only by the change in the entropy of the heat bath, but also by that of the internal physical states of the memory if it is asymmetric. We have clarified the general relationship between the thermodynamic and logical reversibilities, which is summarized in Sec.~5.4; any logically irreversible computation can be thermodynamically reversible in the quasi-static limit.
In Sec.~6, we have discussed the case that there are two memories, and derived a generalized Landauer principle for two memories, which includes the mutual information. In particular, we have discussed measurement and feedback control, which play the crucial roles in the case of Maxwell's demon.
Our results quantitatively extend the conventional Landauer principle to general situations including asymmetric memories, and clarify the subtle and fundamental relationship between the thermodynamic and logical reversibilities. The derived inequalities are based on the second law of thermodynamics, where the information contents are included as well as thermodynamic quantities. Thermodynamics of computation would serve as the foundation to analyze computation at the level of thermal fluctuations with, for example, artificial nanodevices and biological nanomachines.
\ack
The author is grateful to Masahito Ueda, Jordan M. Horowitz, and Juan M. R. Parrondo for valuable discussions, and to Naoto Shiraishi for carefully reading the manuscript.
This work was supported by JSPS KAKENHI Grant Nos. 25800217 and 22340114, and by Platform for Dynamic Approaches to Living System from MEXT, Japan.
\section*{References}
|
\section{Introduction}
Gravitational waves (GW) generated during the coalescence of compact binary systems (CBC) are the only mean to directly probe the space-time in its genuine dynamical regime. Despite the fact that General Relativity (GR) has so far passed all experimental tests with great success \cite{Will2006}, those tests were performed in situations where the field is weak or stationary and the full non-linear dynamics of GR were not explored.
On the other hand, in compact objects like neutron stars or black holes coalesce, they approach orbital velocities as high as $50\%$ the speed of light, when very close to merging with the companion.
The LIGO \cite{ILigo,ALigo} and Virgo \cite{IVirgo,IVirgo2,IVirgo3,AVirgo} ground based gravitational waves observatories are currently undergoing major upgrades, and are scheduled to go back online in 2015 and 2016 respectively~\cite{ObserveScenario}, collecting data with a sensitivity that should allow for a few up to a few tens of CBC detections per year. The exact number will depend on the actual sensitivity reached by the instrument, as well as on the formation rate of compact binary systems, which is still rather uncertain \cite{rates}.
The worldwide network of GW detectors will continue expand during the decade, with LIGO India \cite{Indigo} and Kagra \cite{LCGT} joining operations by 2020. Additional instruments will dramatically improve the sky localization accuracy of GW sources, as well as increase the number of detectable sources \cite{Schutz:2011}.
The prospect of frequent detections calls for the use of a Bayesian framework, which allows the information from each signal to be used to either infer some underlying general property of the observations (see for example \cite{delpozzo12,tayloretal12} for applications in the context of cosmology), or accumulate evidence for a specific model to explain the observations (see \cite{delpozzoetal13} in the context of measuring the neutron star equation of state and \cite{lietal12a} in the context of tests of GR).
Regarding the strong field deviations from GR, the following questions seem interesting:
\begin{enumerate}
\item Will LIGO and Virgo be able to confidently recognize a deviation from GR in a detected signal?\label{Item.Tiger}
\item Should a deviation from GR be visible, wil it be possible to associate the GW signal with a given alternative theory of gravity?\label{Item.OnProgress}
\item Should a deviation be present but not taken into account in the analysis, how would this affect the estimation of physical (e.g. source masses) and extrinsic (e.g. distance) parameters of the source? \label{Item.This}
\end{enumerate}
The answers will obviously depend on the actual nature and magnitude of the deviation from GR, which may be one of the existing proposed alternative theories (see \cite{cornishetal11} and references therein for a list of alternative theories of gravity) or something unanticipated.
Some literature exists which answers the point \ref{Item.Tiger} above: the authors and collaborators have shown \cite{lietal12a,lietal12b} how advanced LIGO and Virgo will be sensitive to quite generic (heuristic) deviations from GR.
They have built a pipeline (TIGER) which works with \emph{any kind} of deviation from GR, thus not requiring the data analyst to know the deviation's form.
Its efficiency was tested by simulating several kinds of deviations from GR, of comparable magnitude, and it has been found that deviations from GR will eventually be evident by combining evidence (in a Bayesian framework) from several signals.
The exact number of detections required to confidently claim a deviation from GR, depends on what deviation was added, but it usually is $\mathcal{O}(10)$.
Even though this proves that non-GR effects beyond solar system and pulsar tests can be measured, it implies that the biases of the point \ref{Item.This} above will be present not only for unmeasurable deviations (\emph{stealth bias})\footnote{One may wonder whether a deviation so small that it can be hardly measured using model selection can have large effects on the estimation of the GR parameters. The answer is usually yes, as model selection will only work if the extra likelihood gained by taking the non GR parameters into account is higher than the penalty paid for having extra parameters (Occam Razor). Because GW detections are noisy, the parameter estimation code will usually be able to shift the GR parameters, hence the bias, to accommodate the deviation.}, but also for deviations whose measurability requires building up evidence with several signals.
That implies that, should a deviation from GR be present, it may not be discovered immediately, and GR waveforms might be used for the first few detections, introducing a bias in the parameter estimation process.
The idea of ``fundamental bias'', i.e. bias in estimated gravitational wave parameters induced by the assumption that GR is correct, was first introduced in \cite{yunespretorius09}. Subsequently, \cite{cornishetal11} coined the term ``stealth bias'' to describe the class of fundamental biases that cannot be corrected a posteriori, since the data do not provide enough evidence to favor an alternative theory of gravity.
With the use of analytical approximations, \cite{vallisneriyunes13} follows up by exploring the conditions, expressed in terms of signal-to-noise (SNR) ratio and magnitude of the deviation from GR, in which single events will be affected by stealth bias. They conclude that significant systematic bias might occur, even for deviations that are not yet excluded by observational constraints.
However, \cite{vallisneriyunes13} takes an approach which is valid for loud signals. Because we expect the sources to be distributed uniformly in co-moving volume, the majority of the gravitational wave events will be weak, with SNR close to the threshold necessary for a confident detection.
For this reason, a full numerical study as close to the real data analysis process as possible is necessary to get more general answers which are valid for signals in a broad range of SNR and parameters.
In this paper, we investigate the transition regime from stealth bias to fundamental bias, considering heuristic deviations from GR that are too small to be confidently detected from any single source observation, but that can be detected after multiple detections. In particular, we focus on the inferences that can be drawn about the class of observed systems from the measurement of the masses. We find that, before enough evidence is accumulated to detect a deviation from GR, the mass measurements can be heavily biased when measured with GR templates. We thus recommend that any astrophysical conclusion drawn by gravitational waves observations should be explicitly conditional on the validity of GR.
The rest of this paper is structured as follows. In section \ref{Sec.Method} we present the experimental set up and the example deviations from GR that were considered; in section \ref{Sec.Results} we present results for each of the deviations; finally, in section \ref{Sec.Conclusions} we draw conclusions and discuss our findings.
\section{Method}\label{Sec.Method}
The bias induced by an unaccounted for deviation from GR in the detected gravitational wave will strongly depend on the exact shape and magnitude of the deviation. Thus, to perform our analysis, we had to choose how to modify the GW signals. While the most natural choice would have been to select some proposed alternative theories of gravity, we have decided not to do so. The reason is twofold: (i) at the moment of writing (with the exception of the investigation of a Massive Gravity theory in \cite{delpozzoetal11,cornishetal11}) no full Bayesian analysis has been performed to check to what extent those theories can be confirmed or ruled out with GW observations; (ii) the class of alternative theories for which usable waveforms are available is very limited, thus limiting the scope of our investigation.
Particularly in view of (ii), we have picked some of the heuristic GR deviations investigated in \cite{lietal12a,lietal12b}.
We chose deviations that are detectable, but only when evidence from $\mathcal{O}(10)$ detections is accumulated. Until then, the analysis with standard GR templates may not be unambiguous and, when a detection is made, no final statements may be made about, e.g., the nature of the source (e.g. a binary black hole (BBH) vs a black hole - neutron star (NSBH) or a binary neutron star (BNS)).
Specifically, in addition to standard GR waveforms, we have considered three possible deviations from GR:
\begin{enumerate}
\item $10\%$ deviation in the 1.5 post-Newtonian (PN) phase coefficient (\cite{lietal12a},~IV~A.1);
\item an extra term in the phase of the GW, corresponding to a ``1.25'' PN order (\cite{lietal12a},~IV~C);
\item an extra term in the phase, whose frequency content depends on the total mass of the binary system (\cite{lietal12b},~3.1);
\end{enumerate}
The first step was to generate a catalog of $150$ events, with component masses uniformly distributed in the range $[1.2-2.8] M_\odot$. The position and orientation parameters were uniformly distributed on the unit sphere, while the distances were distributed uniformly in co-moving volume, in the distance range $[50-400]$ Mpc, keeping only events with network SNRs in a realistic range [10-25].
In addition to the standard GR catalog, three catalogs of non-GR signals were generated by assuming the same events as in the GR catalog but adding the deviations described in the aforementioned list. Henceforth we will refer to those modified catalogs as \emph{d$\chi_3$ } (1.5PN deviation), \emph{NonPN } (``1.25PN'' deviation) and \emph{powerM } (deviation with a mass dependent power of the frequency).
All the signals in the four catalogs (three non-GR plus one GR), were thus identical except for the eventual non-GR contribution to the phase.
Each signal in the four catalogs was analyzed using \verb+lalinference_nest+, \cite{S6PE,lalinfPaper}, a Bayesian parameter estimation code based on the Nested Sampling \cite{Skilling} algorithm. The analysis for the three non-GR catalogs were performed using GR templates, thus simulating the situation in which a non-GR deviation is present in the waveform, but it is not accounted for in the analysis.
We also analyzed the GR catalog (using GR templates) to have an idea of the typical uncertainties and (eventual) biases due to poor sampling, noise, etc., in the ``optimal'' case when the template perfectly matches the injected\footnote{We use here the LIGO/Virgo jargon whereby injection means the process of adding a simulated signal into the noise data stream.} signal.
Each of the simulated waveforms was added to zero mean stationary Gaussian noise with a power spectral density corresponding to the design sensitivity of advanced LIGO and Virgo \cite{commissioningdocument}, as coded in the \verb+lalsimulation+ library \cite{LAL}.
The signals were generated using the so-called TaylorF2 approximant (\cite{BuonannoOverlap}), as produced by the \verb+lalsimulation+ package which is part of the LIGO Algorithm Library \cite{LAL}, considering phase contributions up to the 3.5 PN order (O PN in amplitude). No spins were considered in the waveform due to computing limitations. Generic spins would force us to use time domain waveforms (e.g. SpinTaylorT4, \cite{BuonannoOverlap}) which are much slower to calculate, making it impractical for our large scale research program.
The TaylorF2 approximant is written in the frequency domain, as:
\beq
h(f) = \frac{1}{D} \frac{\mathcal{A}(\theta,\phi,\iota,\psi,\mathcal{M},\eta)}{\sqrt{\dot{F}(\mathcal{M},\eta;f)}} f^{2/3}\,e^{i\Psi(t_c, \phi_c, \mathcal{M},\eta;f)},
\label{Eq.TaylorF2}
\end{equation}
where $D$ is the luminosity distance to the source, $(\alpha,\delta)$ are right ascension and declination, $(\iota,\psi)$ give the orientation of the orbital plane with respect to the line of sight, $\mathcal{M}$ is the chirp mass, and $\eta$ is the symmetric mass ratio. They are defined in terms of the component masses $(m_1,m_2)$, as: $\eta = m_1 m_2/(m_1 + m_2)^2$ and $\mathcal{M} = (m_1 + m_2)\,\eta^{3/5}$. $t_c$ and $\phi_c$ are the time and phase at coalescence, respectively.
$\dot{F}(\mathcal{M},\eta;f)$ is an expansion in powers of the frequency $f$ with coefficients that depend on mass (and eventual spins) and
\beqa
\label{psiidef}
\Psi^{\mathrm{GR}}(t_c, \phi_c, \mathcal{M},\eta;f)&=& 2\pi f t_c - \phi_c - \pi/4 \\
&+& \sum_{i=0}^7 \left[\psi_i + \psi_{i}^{(l)} \ln f \right]\,f^{(i-5)/3}.\nonumber
\label{phase}
\eeqa
The explicit forms of the coefficients $\psi_i$ and $\psi_{i}^{(l)}$ in $(\mathcal{M},\eta)$ are given, for example, in \cite{mais10}.
Each of the signals in the non-GR catalogs had an extra term added to the phase. These were:
\begin{enumerate}
\item d$\chi_3$ : For the d$\chi_3$ deviation, the 1.5 PN phase terms were shifted by 10\%, $\psi_3 \rightarrow \psi_3 (1+0.1)$.
That is: $\Psi = \Psi^{\mbox{GR}} + 0.1 \frac{3}{128 \eta} (-16\pi) (\pi \mathcal{M} f)^{-\frac{2}{3}}$
\item NonPN : In this case the power of the $(\pi \mathcal{M} f)$ term is not normally present in the PN series (that would be $[i-5]/3$ with i an integer): $\Psi = \Psi^{\mbox{GR}} - 2.2 \frac{3}{128 \eta} (\pi \mathcal{M} f)^{-\frac{5}{6}}$. The prefactor $-2.2$ was chosen to make the magnitude of this deviation comparable to the previous one at a reference frequency of 150Hz for a system of $1.5-1.5 M_\odot$.
\item powerM : Finally, for the powerM catalog, the power of $(\pi \mathcal{M} f)$ was a function of the total mass of the system $M$: $\Psi = \Psi^{\mbox{GR}} +\frac{3}{128 \eta} (\pi \mathcal{M} f)^{-2+\frac{M}{3 M_\odot}}$. Here again the prefactor was such that the magnitude of this deviation is comparable to the other two, at the same reference frequency and mass.
\end{enumerate}
The reader is referred to~\cite{lietal12a,lietal12b} for more details about these deformation, their magnitude and their measurability with Advanced LIGO and Virgo.
\section{Results}\label{Sec.Results}
\subsection{GR injection with GR recovery }
In this section we describe the performance of the parameter estimation (PE) process in the case where both injection and template obey GR. This will serve as a reference for the analysis of non-GR signals. The results will be presented in detail because, to the best of our knowledge, a systematic and statistically large set of events analyzed with the ``full'' PE pipeline (as opposed to Fisher Matrix results, e.g. \cite{Vitale2011}) is not present in the literature.
It is a common assumption that the chirp mass $\mathcal{M}$ is well estimated, with sub-percent relative errors (e.g. \cite{Vitale2011,zanolinetal10,arunetal05,arunetal05erratum} with Fisher Matrix, \cite{S6PE} with the MCMC code), while the component masses are estimated with $\mathcal{O}(10)$ percent error, which will make it hard to infer the nature of the source (i.e. whether it was a BNS, NHBS or low mass BBH) \cite{hannametal13}.
Our findings confirm this assumption; the chirp mass is estimated with relative errors (i.e. standard deviation divided by the injected value: $\Gamma_\alpha\equiv \sigma_\alpha/\alpha_{\mbox{true}}$) that are never larger than $\pm0.1\%$ across the whole SNR range, while typical values are $\sim 0.03-0.04\%$ . This is shown in Fig. \ref{Fig.MchirpBPGRGR}: The boxes are logarithmically spaced to take into account the fact that there were more events at low SNR than at high SNR, so that at least 10 events (unless otherwise indicated) are contained in each box.
\begin{figure}[h!]
\includegraphics[width=\columnwidth]{1.png}
\caption{Chirp mass percentage relative error, $\Gamma_\mathcal{M}$, for the GR catalog set of events as a function of the optimal SNR. The boxes indicate the lower to upper quartile values of the data, with a line at the median. The whiskers show 1.5 times the interquantile range, while the symbols are the remaining data points. In every SNR bin under consideration the relative error for $\mathcal{M}$ is always smaller than $0.1\%$, and usually $\lesssim 0.03\%$ for medium-high SNR signals}\label{Fig.MchirpBPGRGR}
\end{figure}
As for the symmetric mass ratio $\eta = m_1 m_2/(m_1 + m_2)^2$, we find that the error can be as large as $\sim 5\%$, but is generally $\lesssim 2\%$ for medium-low SNR events, and $\lesssim 1\%$ for louder events, as shown in Fig. \ref{Fig.etaBPGRGR}.
\begin{figure}[h!]
\includegraphics[width=\columnwidth]{2.png}
\caption{Symmetric mass ratio percentage relative error, $\Gamma_\eta$, for the GR catalog set of events as a function of the optimal SNR. The bias can be as large as $\sim 5\%$, but, in general, are $\sim 2\%$ events with medium SNR ($\sim 15$), and $\lesssim 1\%$ for louder event.}\label{Fig.etaBPGRGR}
\end{figure}
In particular, we find that for $50\%$ ($90\%$) of the signals the relative error for the symmetric mass ratio is smaller than $1\%$ ($2\%$).
We would like to draw some conclusions about how often the PE code will make the correct inference on the nature of the source, i.e. how precisely the component masses can be estimated.
Because the maximum mass of a neutron star is a function of the yet unknown equation of state, e.g. \cite{lattimer12}, we recognize that, for the time being, there can \emph{not} be an uncontroversial choice for this upper bound.
We thus choose the reasonable value of $2.0 M_\odot$ \cite{AntoniadisAl}. Henceforth, we will label as ``neutron star'' (``black hole'') an object lighter (heavier) than $2.0M_\odot$.
Having made clear our choice and its limitations, we can now calculate how often a system that was injected into a mass bin, e.g. BNS, is correctly assigned to the corresponding mass bin at a two sigma confidence level.
We must also consider the possibility that we will not be able to make a decision (i.e. that the error bars on the component masses are such that we cannot decide where to put the signal) or that a wrong inference will be made, assigning the injection to a different mass bin than where it was injected. This point will become more important when non-GR injections will be performed.
Our findings are reported in Table. \ref{Tab.GR_GR}: nearly half of the signals are assigned to the correct mass bin, at 95\% confidence level.
Even more important, we notice that signals are either assigned to the correct mass bin or not assigned at all, and that none of the signals are assigned to the wrong mass bin. We will see that the situation is very different whit non-GR injections.
\begin{table}[h]
\begin{tabular}{|c|c|c|c|c|c|}
\hline
&&\multicolumn{3}{c|}{Rec. as}&\multicolumn{1}{c|}{Unassigned}\\
&& BNS & NSBH & BBH & at $2\sigma$\\
\hline
\multirow{3}{3mm}{\begin{sideways}\parbox{15mm}{Inj. as}\end{sideways}}& BNS & 45\%& 0 & 0&55\%\\ [1.1ex] \cline{2-6}
&NSBH & 0 & 44\% & 0&56\% \\ [1.1ex]\cline{2-6}
&BBH & 0 & 0 & 48\%&52\% \\ [1.1ex]\cline{2-6}
\hline
\end{tabular}
\caption{Fractions of signals recovered as a BNS, NSBH or BBH at two sigma confidence level (see the text for details). Nearly half of the times the code is able to infer correctly what had been injected. \label{Tab.GR_GR}}
\end{table}
We also have checked how the efficiency in assigning injections to the correct mass bin depends on the injected chirp mass. As expected, the efficiency is quite high ($\sim60\%$ for low mass BNS and $\sim 90\%$ for high mass BBH) for systems which are either very light or very heavy, and decreases significantly for systems with chirp mass in the range $\sim 1.4-2.0$ where the error bars on the component masses can easily cross the BNS-NSBH or NSBH-BBH boundary, and no decision can be made.
A summary of the GR analysis is reported in Table~\ref{Tab.GR_stat}. The first five rows reports statistics for
for $\mathcal{M},\eta,m_{1,2},D$ and $\iota$.
We quote the median and $90^{\mbox{th}}$ percentile of the relative errors $\Gamma$ and the absolute value of the \emph{effect size} $\Sigma$, defined as:
\beq\label{Eq.Esize}
\Sigma_\alpha\equiv\frac{\bar{\alpha} - \alpha_{\mbox{true}}}{\sigma_{\alpha}};
\end{equation}
where $\bar{\alpha}$ and $\sigma_\alpha$ are the median and standard deviation of the posterior distribution of $\alpha$, and $\alpha_{\mbox{true}}$ is the injected ``true'' value. The effect size represents the offset in units of standard deviation.
For all parameter, 50\% (90\%) of signals have a median which is found within $\sim 0.6$ ($\sim 1.5$) standard deviations. These values are consistent with the expectations for random Gaussian variables, thus showing the robustness of the parameter estimation algorithm.
The bottom row of Table~\ref{Tab.GR_stat} focuses on sky localization performances: the first two columns report the median and $90^{\mbox{th}}$ percentile for the 90\% confidence level sky error area (in square degrees), the numbers in brackets correspond to a selection of events having SNR above 8 in \emph{all} three instruments. Finally, the last two columns report the angle (in degrees) between the injected sky position and the maximum likelihood point; 90\% of the signals that had been detected in the three interferometers were found at less than 2 degrees from the true position, and with a sky area smaller than 33 $\mathrm{deg}^2$.
\begin{table}[h]
\begin{tabular}{|c|c|c|c|c|}
\hline
&\multicolumn{2}{c|}{$\Gamma$}&\multicolumn{2}{c|}{$|\Sigma|$}\\
& 50\% & 90\% & 50\% &90\%\\
\hline
$\mathcal{M}$ & 0.03\%& 0.05\% & 0.6&1.4\\ [1.1ex] \cline{1-5}
$\eta$ & 1.1\% & 2.2\% & 0.5&1.5 \\ [1.1ex]\cline{1-5}
$m_{1,2}$ &$4.5$\% & $6.8$\% &$0.7$&$1.4$ \\ [1.1ex]\cline{1-5}
$D$ & 20.7\% & 31.3\% & 0.6&1.5 \\ [1.1ex]\cline{1-5}
$\iota$ & 32.5\% & 104.7\% & 0.6&1.6 \\ [1.1ex]\cline{1-5}
\hline\hline
&\multicolumn{2}{c|}{Sky Error [deg$^2$]}&\multicolumn{2}{c|}{Sky offset [deg]}\\
& 50\% & 90\% & 50\% &90\%\\
$\delta\Omega_{90}$ & 15.3 (4.5) & 95.3 (33.3) & 1.7 (0.8) &25 (2.2) \\ [1.1ex]\cline{1-5}
\hline
\end{tabular}
\caption{Summary of errors for the GR analysis. The first two columns, $\Gamma$, quote the $50^{\mbox{th}}$ and $90^{\mbox{th}}$ percentile for the relative errors. The last two columns, $|\Sigma|$, quote the $50^{\mbox{th}}$ and $90^{\mbox{th}}$ percentile for the effect size, eq.\ref{Eq.Esize}. The numbers for each component mass are similar, thus we quote their means in the $m_{1,2}$ row. For the sky localization accuracy (last row) the last two columns report the percentiles on the angle offset between the injected and median recovered sky position. The numbers in brackets refer to events which have an SNR above 8 in all detectors. \label{Tab.GR_stat}}
\end{table}
The following subsections will be devoted to analyses of the non-GR catalogs. As a general statement, because all the variations from GR we considered affect the phase of the waveform, we would expect that intrinsic parameters, i.e. the mass parameters, are the most affected by the unaccounted deviation. We will see that this is generally the case.
\subsection{d$\chi_3$ injection with GR recovery }
In this section we report the analysis of the d$\chi_3$ catalog, i.e. the signals in which a 10\% deviation in the 1.5PN phase terms is present. Here and in what follows we will use two figures of merit for the stealth bias: the relative offset with respect to the injected value:
$$
\Delta_\alpha \equiv \frac{\bar{\alpha} - \alpha_{\mbox{true}}}{\alpha_{\mbox{true}}}
$$
and the effect size, $\Sigma$, defined in eq. \ref{Eq.Esize}
We found that the chirp mass estimation, Fig. \ref{Fig.MchirpBPdchiGR}, is only mildly affected, with relative offsets that, even if larger (up to $\sim 10$ times) than the typical uncertainties for this parameter in the GR case, are still well below the percent level. As a consequence, the selection of BNS events for a test of GR based on the measured chirp mass \cite{AgathosAl}, will not be affected by stealth bias.
\begin{figure}[h!]
\includegraphics[width=\columnwidth]{3.png}
\caption{$\Delta_\mathcal{M}$ for the d$\chi_3$ catalog set of events as a function of the optimal SNR. Line, box, whiskers and symbols are the same as in Fig.~\ref{Fig.etaBPGRGR}. In each SNR bin, the median $\mathcal{M}$ and the injected one differ by $\sim 0.1$\%. While the offset is a factor of three larger than the typical errors for the GR catalog, it is still a fraction of percent.}\label{Fig.MchirpBPdchiGR}
\end{figure}
On the other hand, the mass ratio is heavily affected by the presence of the d$\chi_3$ deviation. This finding is not totally unexpected, cfr. Fig. 3 in \cite{lietal12a}.
The relative offset is $~ -15 \%$ for all the signals, while the measurement becomes more precise for loud signals. $\Sigma_\eta$ thus gets larger and larger, with the loudest events being measured $\sim 20 $ standard deviations away from the injected values, Fig.~\ref{Fig.etaBPdchiGR}.
\begin{figure}[h!]
\includegraphics[width=\columnwidth]{4.png}
\caption{$\Sigma_\eta$ for the d$\chi_3$ catalog set of events as a function of the optimal SNR. Line, box, whiskers and symbols are the same as in Fig.~\ref{Fig.etaBPGRGR}. The median recovered value is more than 5 standard deviations away from the injected value for low SNR events, and gets further and further as the SNR increases, with the loudest events having $\Sigma_\eta \sim -20$. This is not due to the bias becoming larger for louder signals but to the standard deviation becoming smaller.}\label{Fig.etaBPdchiGR}
\end{figure}
Since the bias is always negative, the parameter estimation algorithm systematically \emph{underestimates} the value of the mass ratio.
In other words, each system is seen as having component masses that are more different than reality. This is also shown in Table \ref{Tab.dchi_GR} in which the overwhelming majority of the signals are identified as NSBH systems.
Note the difference compared to the GR catalog, Table~\ref{Tab.GR_GR}. In the GR case systems were either not assigned to any mass bin or correctly identified. In the case in which a d$\chi_3$ departure from GR is present, systems are misidentified 95\% of the times for the BNS case and 98\% of the times for BBH. The NSBH bin is now 100\% (it was 44\% for GR), since all the signals for which a decision could not be made in the GR catalog are being pushed toward very low mass ratios.
Not a single BNS or BBH gets assigned to the correct mass bin.
\begin{table}[h]
\begin{tabular}{|c|c|c|c|c|c|}
\hline
&&\multicolumn{3}{c|}{Rec. as}&\multicolumn{1}{c|}{Unassigned}\\
&& BNS & NSBH & BBH & at $2\sigma$\\
\hline
\multirow{3}{3mm}{\begin{sideways}\parbox{15mm}{Inj. as}\end{sideways}}& BNS & 0& 95\% & 0&5\%\\ [1.1ex] \cline{2-6}
&NSBH & 0 & 100\% & 0&0 \\ [1.1ex]\cline{2-6}
&BBH & 0 & 98\% & 0&2\% \\ [1.1ex]\cline{2-6}
\hline
\end{tabular}
\caption{Fraction of signals recovered as a BNS, NSBH or BBH at two sigma confidence level for d$\chi_3$ injections. }\label{Tab.dchi_GR}
\end{table}
The seriousness of the bias is evident from Fig. \ref{Fig.etaComparison} in which we compare the distribution of the injected mass ratios (in red) the posterior medians for the GR (in blue) and d$\chi_3$ catalog (in green). As expected, in the GR catalog the median estimated values match closely the injected ones. On the contrary, in the d$\chi_3$ catalog the distribution of the posterior medians is shifted towards smaller values of $\eta$, and has barely any overlap with the distribution of the injected values. The distribution of the recovered symmetric mass ratios for d$\chi_3$ peaks at $\eta \sim 0.21$ , which means that the most massive star is seen as twice as massive as than the lighter object.
\begin{figure}[h!]
\includegraphics[width=\columnwidth]{5.png}
\caption{Distribution for the injected values of $\eta$ (in red) compared with the distributions of posterior medians as measured in the GR catalog (in blue) and the d$\chi_3$ catalog (in green). The d$\chi_3$ distribution is clearly offset from both the injections and the GR.}\label{Fig.etaComparison}
\end{figure}
We further verified whether the bias on the symmetric mass ratio depends on the injected values of the masses of the system, and we found that it does not. Indeed, the non-GR phase terms at a given frequency do not vary significantly over the mass parameters range we are probing here. On the other hand, the powerM deviation shows a much stronger dependence on the masses, with the magnitude of the shift inversely proportional to the masses. We will indeed see in Sec.~\ref{SubS.PowerM} that the bias is larger for BNS.
Finally, we remark that, as expected with phase-only deviations, none of the extrinsic parameters shows significant biases with respect to the GR catalog.
\subsection{NonPN injection with GR recovery }
In this subsection we investigate the NonPN catalog, for which the deviation from GR predictions can be, with an abuse of notation, dubbed ``1.25PN''.
For this catalog the relative offset introduced in the chirp mass estimation is moderate, even though generally larger than for the d$\chi_3$ catalog, being usually around $0.25-0.3$\% for all the event.
We find that the offset in $\eta$ is comparable, in both magnitude and sign, to what was seen for the d$\chi_3$ runs, i.e. the posterior medians are systematically \emph{underestimated}. This is shown in Fig.~\ref{Fig.etaBPOne25GR}.
\begin{figure}[h!]
\includegraphics[width=\columnwidth]{6.png}
\caption{$\Sigma_\eta$ for the NonPN set of events as a function of the optimal SNR. Line, box, whiskers and symbols are the same as in Fig.~\ref{Fig.etaBPGRGR}. The median recovered value is more than 5 standard deviations away from the injected value for low SNR events, and gets further and further as the SNR increases, with the loudest events having $\Sigma_\eta \sim -20$ }\label{Fig.etaBPOne25GR}
\end{figure}
It is thus not surprising that our findings resemble the $d\chi_3$ catalog. Nearly the totality of events are seen as NSBH, Table~\ref{Tab.nonPN_GR}.
\begin{table}[h]
\begin{tabular}{|c|c|c|c|c|c|}
\hline
&&\multicolumn{3}{c|}{Rec. as}&\multicolumn{1}{c|}{Unassigned}\\
&& BNS & NSBH & BBH & at $2\sigma$\\
\hline
\multirow{3}{3mm}{\begin{sideways}\parbox{15mm}{Inj. as}\end{sideways}}& BNS & 0& 100\% & 0&0\\ [1.1ex] \cline{2-6}
&NSBH & 0 & 100\% & 0&0 \\ [1.1ex]\cline{2-6}
&BBH & 0 & 98\% & 0&2\% \\ [1.1ex]\cline{2-6}
\hline
\end{tabular}
\caption{Fraction of signals recovered as a BNS, NSBH or BBH at two sigma confidence level for NonPN signals.}\label{Tab.nonPN_GR}
\end{table}
\subsection{powerM injection with GR recovery}\label{SubS.PowerM}
In this subsection we describe our findings for the powerM catalog. As described in Sec.~\ref{Sec.Method}, these signals are characterized by the presence of an extra term in the GW phase whose power of the frequency is a function of the total mass of the system. This is a rather different, and richer, situation than d$\chi_3$ or NonPN , where all signals had non-GR shifts with the same frequency dependence, and we may thus expect the resulting bias to manifest itself differently.
We find that the bias in the chirp mass, Fig. \ref{Fig.MchirpBPMGR}, is usually smaller than in the other catalogs, even though several outliers are present for which the offset is several tenths of percent. We also notice that, unlike the other two deviations, the sign of the bias is not the same for all events, but tends to be negative for low-mass events and slightly positive for the most massive sources. Such an effect should not come as a total surprise, given that, by its very nature, this deviation strongly depends on the mass of the system.
\begin{figure}[h!]
\includegraphics[width=\columnwidth]{7.png}
\caption{$\Delta_\mathcal{M}$ for the powerM catalog set of events as a function of the optimal SNR. Line, box, whiskers and symbols are the same as in Fig.~\ref{Fig.etaBPGRGR}.}\label{Fig.MchirpBPMGR}
\end{figure}
As for $\eta$, Fig.~\ref{Fig.etaBPMGR}, we observe that the sign of the bias is now positive. Component masses will thus be seen as more equal than they actually are.
\begin{figure}[h!]
\includegraphics[width=\columnwidth]{8.png}
\caption{Effect size for $\eta$ in the powerM catalog set of events as a function of the optimal SNR. Line, box, whiskers and symbols are the same as in Fig.~\ref{Fig.etaBPGRGR}. The effect is generally smaller than what seen in the d$\chi_3$ or NonPN catalogs, even if several outliers exist, which are found at $\gtrsim 5$ sigmas from the injected value. We also notice the presence of both positive and negative biases, the sign depending on the injected chirp mass (see text).}\label{Fig.etaBPMGR}
\end{figure}
That is shown in Table~\ref{Tab.powerM_GR}: 64\% of BNS injections are now being recovered as BNS. That is more than in the GR injections case, where the number was 45\%. The explanation is that some of the events for which a decision could not be made for GR injections have been pushed up to higher $\eta$, to the BNS cell.
The same line of thought applies to the NSBH injections: 44\% of them were correctly recognized in the GR catalog while only 29\% are still seen as NSBH, and 9\% are mislabeled as BNS.
Finally, we do not see much difference for the BBH injections, which is due to the fact that the magnitude of the powerM non-GR shift gets very small for chirp masses $\gtrsim 2 M_\odot$ (see Fig.~\ref{Fig.MchirpEtaScatMGR}, top panel).
\begin{table}[h]
\begin{tabular}{|c|c|c|c|c|c|}
\hline
&&\multicolumn{3}{c|}{Rec. as}&\multicolumn{1}{c|}{Unassigned}\\
&& BNS & NSBH & BBH & at $2\sigma$\\
\hline
\multirow{3}{3mm}{\begin{sideways}\parbox{15mm}{Inj. as}\end{sideways}}& BNS & 64\%& 0 & 0&36\%\\ [1.1ex] \cline{2-6}
&NSBH & 9\% & 29\% & 0&62\% \\ [1.1ex]\cline{2-6}
&BBH & 0 & 0& 50\%&50\% \\ [1.1ex]\cline{2-6}
\hline
\end{tabular}
\caption{Fraction of signals recovered as a BNS, NSBH or BBH at two sigma confidence level for powerM injections. }\label{Tab.powerM_GR}
\end{table}
We would thus expect, just from back of the envelop calculations, low-mass events to be more heavily biased than higher mass events; that is indeed what we have found.
The top panel in Fig. \ref{Fig.MchirpEtaScatMGR} shows the bias for the chirp mass (colorbar) for the various events, labeled by their injected $\mathcal{M}$ and $\eta$. The circles are proportional to the loudness of the event. It is clear how the bias for the chirp mass strongly depends on the injected $\mathcal{M}$, and is more important for \emph{low} chirp mass systems.
The equivalent plot for the mass ratio $\eta$, Fig.~\ref{Fig.MchirpEtaScatMGR} bottom panel, shows dependence on both the injected $\mathcal{M}$ and $\eta$, with large-$\eta$ events getting a smaller bias on $\eta$.
That is easily understandable. Because the effect of the powerM deviations from GR is to \emph{increase} the recovered value of $\eta$, events which were already very close to the upper bound ($\eta=0.25$) will be less biased for the simple reason that their posterior distribution cannot move any higher. For those events the effects of the non-GR shift will mostly be to narrow down the $\eta$ posterior distribution.
\begin{figure}[h!]
\includegraphics[width=\columnwidth]{9a.png} \\
\includegraphics[width=\columnwidth]{9b.png}
\caption{Top: bias for the estimation of $\mathcal{M}$ as function of the injected values of $\mathcal{M}$ and $\eta$.
Bottom: bias for the estimation of $\eta$ as function of the injected values of $\mathcal{M}$ and $\eta$. The bias becomes larger for smaller injected values of $\mathcal{M}$.}\label{Fig.MchirpEtaScatMGR}
\end{figure}
\section{Conclusions}\label{Sec.Conclusions}
In this work we have shown how parameter estimation for gravitational wave signals can be strongly affected by deviations from general relativity, should they not be taken into account in the waveform used as templates.
For the purpose of illustration, we have considered three heuristic deviations from GR, whose non-GR nature would \emph{eventually} be recognizable using evidence from several, $\mathcal{O}(10)$, detections.
We created a catalog of $150$ signals, that were analyzed a total of 4 times: injecting the ``correct'' GR signal, or injecting one of the 3 deviations from GR.
We have seen that the effect of those deviations on the mass ratio can be very different. It can lead to a heavy systematic underestimation of $\eta$ (d$\chi_3$ and NonPN catalog), where the mass ratio is biased toward 2:1, i.e. $\eta\sim 0.22$, or larger, or to an overestimation of $\eta$, powerM catalog, where the distribution of recovered $\eta$'s is pushed toward the upper boundary at 0.25 and the systems seen as equal-mass.
For the sake of argument, and without claim of astrophysical validity, we have labeled as neutron star (black hole) compact objects lighter (heavier) than $2M_\odot$. We have shown that when GR templates are used for injection and for the analysis, our current parameter estimation algorithms are able to recognize the nature of the injected systems, at the $2\sigma$ confidence level, $\sim$ 50\% of the time. Moreover, none of the GR signals were assigned to the wrong source class.
The situation was reversed when injections were allowed to depart from GR and analyzed using GR templates. For the d$\chi_3$ (NonPN ) catalog, for example, 95\% (100\%) of the injected BNS were mistaken for NSBH. For the powerM catalog, only $29\%$ of the injected NSBH was recognized as such, while $9\%$ of them were mistaken for BNS.
Even though the numerical details of our findings would change if different mass thresholds were to be chosen, our conclusions can be summarized in two points:
\begin{enumerate}
\item[i)] if the templates used are a good representation of the detected signals, $\sim 50\%$ of the times we can infer the nature of the detected signal at a $2-\sigma$ confidence level;
\item[ii)] if the templates do not match the signal waveforms well, the measured component masses of a system are an unreliable (and potentially disastrous) indicator of the class of the system.
\end{enumerate}
Therefore, any future inference that will be drawn from an in-depth analysis of GW signals with state-of-the-art parameter estimation algorithms, \emph{is critically and explicitly dependent on the underlying theory of gravity assumed}.
The study herein reported focused on the bias introduced by the presence of non-GR phase terms in the signal waveform when those are not present in the template. However, it is easy to appreciate that similar effects will be introduced by other possible mismatches (e.g. unknown large Post Newtonian orders, tidal effects, spins, eccentricity). It is therefore imperative for the GW community to concentrate on the development of as accurate waveforms as possible or of methods to be robust against the potential systematics that our approximate waveforms might introduce.
The uniqueness of the non-GR stealth bias, however, is that it cannot be eliminated with more precise numerical simulations or analytic models, as it represents the very uncertainty on our understanding of gravity in its strong-field regime.
Furthermore, the conclusions drawn from studies that rely on the correct identification of the component objects of a compact binary source (\emph{e.g.} the measurement of the differential rate of coalescence in each class of systems or the measurement of the mass function) will have to be conditional on the assumption that GR is the correct description of the physics of the system. However, it is not farfetched to assume that if a deviation from GR is eventually detected, all inferences will be corrected \emph{a posteriori}.
In conclusion, the somewhat exotic, but with very real effects, ``stealth bias'' is nothing more than a consequence of our assumptions about the theory of gravity describing the process of gravitational radiation. This is a common phenomenon in every inference process, since the conclusions \emph{always} depend on the assumptions, but it is particularly worrying in the gravitational wave physics context, since it deals with the very foundations of our understanding of gravity.
\section{Acknowledgments}
S.V. acknowledges the support of the National Science Foundation and the LIGO Laboratory. LIGO was constructed by the California Institute of Technology and Massachusetts Institute of Technology with funding from the National Science Foundation and operates under cooperative agreement PHY-0757058.
W.D.P. is supported by the research program of the Foundation for Fundamental Research on Matter (FOM), which is partially supported by the Netherlands Organisation for Scientific Research (NWO). The work was funded in part by a Leverhulme Trust research project grant.
The authors would like to acknowledge the LIGO Data Grid clusters, without which the simulations could not have been performed. Specifically, these include the computing resources supported by National Science Foundation awards PHY-0923409 and PHY-0600953 to UW-Milwaukee. Also, we thank the Albert Einstein Institute in Hannover, supported by the Max-Planck-Gesellschaft, for use of the Atlas high-performance computing cluster.
We would also like to thank Trevor Sidery, Christopher Berry, Neil Cornish, Andy Lundgren, Ilya Mandel, Chris Van Den Broeck, Ruslan Vaulin, Nico Yunes and Rai Weiss for useful comments and suggestions.
We are particularly grateful to Reed Essick for his comments and for carefully reading the manuscript.
This is LIGO document number P1300188.
|
\section{Introduction}
\label{sec:intro}
\input{introduction-radu}
\section{Related Work}
\label{sec:sota}
While vision and hearing have been mainly addressed separately, several behavioral, electrophysiological and imaging
studies \cite{Calvert04}, \cite{Ghazanfar06}, \cite{Senkowski08} postulate that the fusion of different sensorial
modalities is an essential component of perception. Nevertheless, computational models of audio-visual fusion and their
implementation on robots remain largely unexplored.
The problem of integrating data gathered with physically different sensors, e.g., cameras and microphones, is extremely
challenging. Auditory and visual sensor-data correspond to different physical phenomena which must be interpreted in a
different way. Relevant visual information must be inferred from the way light is reflected by scene objects and valid
auditory information must be inferred from the perceived signals such that it contains the properties of one or several
emitters. The spatial and temporal distributions of auditory and visual data are also very different. Visual data are
spatially dense and continuous in time. Auditory data are spatially sparse and intermittent since in a natural
environment there are only a few acoustic sources. These two modalities are perturbed by different phenomena such as
occlusions and self-occlusions (visual data) or ambient noise and echoic environments (auditory data).
Despite all these challenges, numerous researchers investigated the fusion of auditory and visual cues in
a variety of domains such as event classification \cite{Natarajan12}, speech recognition \cite{Barker09}, sound source
separation \cite{Naqvi10}, speaker tracking \cite{Hospedales08}, \cite{Gatica07} and speaker diarization
\cite{Noulas12}. However, these approaches are not suitable for robots either because the algorithmic
complexity is too high, or because methods use a distributed sensor network or because the amount of
training data needed is too high, drastically reducing the robots' adatableness. Unfortunately, much less effort has been
devoted to design audio-visual fusion methods for humanoid robots. Nevertheless, there are some interesting works
introducing methods specifically conceived for humanoid robots on speech recognition \cite{Nakadai04}, beat tracking
\cite{Itohara11}, \cite{Itohara12}, active audition \cite{Kim07} and sound recognition \cite{Nakamura11a}. All
these methods deal with the detection and localisation problem by using a combination of off-the-shelf algorithms,
suitable for humanoid robots. Albeit, all these approaches lack from a framework versatile enough to be used in other
situations than the ones they are specifically designed for.
Finding the potential speakers and assessing their speaking status is a pillar task, on which all applications mentioned
above rely. In other words, providing a robust framework to count how many speakers are in the scene, to localize them
and to ascertain their speaking state, will definitely increase the performance of many audio-visual perception
methods. This problem is particularly interesting in the case of humanoid robots, because the framework must be
designed for \textit{untethered} interaction using a set of \textit{robocentric sensors}. That is to say that the
cameras and microphones are mounted onto a robotic platform that freely interacts with the unconstrained AV events
(i.e., people). As a consequence, the use of any kind of distributed sensor network, e.g. close-range microphones and
speaker-dedicated cameras, is not appropriated. Likewise, the algorithms should be light enough to satisfy the
constraints associated to real-time processing with a humanoid robot.
The existing literature on speaker detection and localisation can be grouped into two main research lines. On one side,
many statistical non-parametric approaches have been developed. Indeed, \cite{Gurban06}, \cite{Besson08a} and
\cite{Besson08b} investigate the use of information theory-based methods to associate auditory and visual data in order
to detect the active speaker. Similarly, \cite{Barzelay07} proposes an algorithm matching auditory and visual onsets.
Even though these approaches show very good performance results, they use speaker/object dedicated cameras, thus
limiting the interaction. Moreover, the cited non-parametric approaches need a lot of training data. The outcome of
such training steps is also environment-dependent. Consequently, implementing such methods on mobile platforms results
in systems with almost no practical adaptability.
On the other side, several probabilistic approaches have been published. In \cite{Khalidov08}, \cite{Khalidov11a}, the
authors introduce the notion of conjugate GMM for audio-visual fusion. Two GMMs are estimated, one for each modality
(vision and auditory) while the two mixture parameter sets are constrained through a common set of \textit{tying
parameters}, namely the 3D locations of the AV events being sought. Recently in \cite{Noulas12}, a factorial HMM is
proposed to associate auditory, visual and audio-visual features. All these methods simultaneously detect and localize
the speakers but they are not suitable for real-time processing, because of their algorithmic complexity. \cite{Kim07}
proposed a Bayesian framework inferring the position of the active speaker and combining a sound source localisation
technique with a face tracking algorithm on a robot. The reported results are good in the case of one active speaker,
but show bad performance for multiple/far speakers. This is due to the fact that the proposed probabilistic framework is
not able to correctly handle outliers. In \cite{Alameda11}, the authors use a 1D GMM to fuse the auditory and visual
data, building AV clusters. The probabilistic framework is able to handle the outliers thanks to one of the
mixture components. However, the algorithm presented in the paper is not light enough for real-time processing.
Unlike these recent approaches, we propose a novel hybrid deterministic/probabilistic model for audio-visual detection
and localisation of speaking people. Up to the authors' knowledge, we introduce the very first model with the
following remarkable attributes all together: (i) theoretically sound and solid, (ii) designed to process robocentric
data, (iii) accommodating different visual and auditory features, (iv) robust to noise and outliers, (v) requiring a
once-and-forever tiny calibration step guaranteeing the adaptability of the system, (vi) working on unrestricted indoor
environments, (vii) handling a variable number of people and (viii) implemented on a humanoid platform.
\section{A Hybrid Deterministic/Probabilistic Model}
\label{sec:problem}
We introduce a multimodal deterministic/probabilistic fusion model for audio-visual detection and localisation of
speaking people that is suitable for real-time applications. The algorithms derived from that hybrid model aim to count
how many speakers are there, find them in the scene and ascertain when they speak. In other words, we seek for the
number of potential speakers, $N\in\mathbb{N}$, their positions $\Svect_n\in\mathbb{S}$ ($\mathbb{S}\subset\mathbb{R}^3$
is the scene space) and their speaking state $e_n\in\{0,1\}$ (0 -- \textit{not speaking} and 1 -- \textit{speaking}).
In order to accomplish the detection and localization of speakers, auditory and visual features are extracted from the
raw signals (sound track and image flow), during a time interval $\Delta t$. We assume $\Delta t$ to be short enough
such that the speakers remain approximately in the same 3D location and long enough to capture small
displacements and oscillatory movements of the head, hands, torso and legs. The auditory and visual features extracted
during $\Delta t$ are denoted by $\amat=\left\{\avect_1,\ldots,\avect_k,\ldots,\avect_K\right\}\subset\mathbb{A}$ and by
$\vmat=\left\{\vvect_1,\ldots,\vvect_m,\ldots,\vvect_M\right\}\subset\mathbb{V}$ respectively, where $\mathbb{A}$
($\mathbb{V}$) is the auditory (visual) feature space.
We aim to solve the task from the auditory and visual observations. That is, we want to compute the
values of $N$, $\{\Svect_n\}_{n=1}^N$ and $\{e_n\}_{n=1}^N$, that best explain the extracted features $\amat$ and
$\vmat$. Therefore, we need a framework that encompasses all (hidden and observed) variables and that accounts for the
following challenges: (i)~the visual and auditory observations lie in physically different spaces with different
dimensionality, (ii) the object-to-observation assignments are not known in advance, (iii) both visual and auditory
observations are contaminated with noise and outliers, (iv) the relative importance of the two types of data is
unassessed, (v) the position and speaking state of the speakers has to be gauged and (vi) since we want to be able to
deal with a variable number of AV objects over a long period of time, the number of AV object that are effectively
present in the scene must be estimated.
We propose a hybrid deterministic/probabilistic framework performing audio-visual fusion, seeking for the desired
variables and accounting for the outlined challenges. On one hand, the deterministic components allow us to model those
characteristics of the scene that are known with precision in advance. They may be the outcome of a very accurate
calibration step, or the direct consequence of some geometrical or physical properties of the sensors. On the other
hand, the probabilistic components model random effects. For example, the feature noise and outliers, which is a
consequence of the contents of the scene as well as the feature extraction procedure.
\subsection{The Deterministic Model}
In this section we delineate the deterministic components of our hybrid model: namely the visual and auditory mappings.
Because the scene space, the visual space and the auditory space are different we need two mappings: the first one,
${\cal A}:\mathbb{S}\rightarrow\mathbb{A}$, links the scene space to the auditory space and the second one, ${\cal
V}:\mathbb{S}\rightarrow\mathbb{V}$, links the scene space to the visual space. Both mappings are represented in
\reffig{fig:mapping}. An AV object placed at $\Svect$ in the scene space, is virtually placed at ${\cal A}(\Svect)$ in
the auditory space and at ${\cal V}(\Svect)$ in the visual space.
\begin{figure}
\centering
\includegraphics[width=0.7\linewidth]{IntroImage/MappingsSimple.pdf}
\caption{Perceptual auditory ($\cal A$) and visual ($\cal V$) mappings of NAO. The extracted auditory
$\avect_k$ and lay around ${\cal
A}(\Svect)$ and ${\cal V}(\Svect)$ respectively. An audio-visual mapping needs to be built to link the two observations
spaces.}
\label{fig:mapping}
\end{figure}
The definition of ${\cal A}$ and ${\cal V}$ provide a link between the two observations spaces, which corresponds
either to ${\cal A}\circ{\cal V}^{-1}$ or to ${\cal V}\circ{\cal A}^{-1}$. Depending on the extracted features and on
the sensors, the mappings ${\cal A}$ and ${\cal V}$ may be invertible. If that is not the case, ${\cal A}\circ{\cal
V}^{-1}$ or ${\cal V}\circ{\cal A}^{-1}$ should be estimated through a learning procedure. There are several works
already published dealing with this problem in different ways. In \cite{Alameda11,Sanchez12}, ${\cal V}$ is invertible
and ${\cal A}$ is known, so building ${\cal A}\circ{\cal V}^{-1}$ is straightforward. In sound source localization
approaches (inter alia \cite{Nakadai04}) ${\cal A}$ is invertible and ${\cal V}$ is known so ${\cal V}\circ{\cal
A}^{-1}$ is easily constructed. In \cite{Khalidov08,Khalidov11a}, none of the mappings are inverted, but used to tie the
parameters of the probabilistic model. So the link between $\mathbb{A}$ and $\mathbb{V}$ is not used explicitly, but
implicitly. In \cite{Butz05,Kidron05,Kidron07,Liu08}, the scene space is undetermined and the authors learn a common
representation space (the scene space) at the same time they learn both mappings.
In our case, we chose to extract 3D visual feature points, and represent them in the scene coordinate system (see
\refsec{sec:visual_features}). Thus, the mapping ${\cal V}$ is the identity, which is invertible. The auditory features
correspond to the Interaural Time Differences (see \refsec{sec:auditory_features}), and a direct path propagation model
defines ${\cal A}$. The mapping ${\cal A}\circ{\cal V}^{-1}$ is accurately built from the geometric and physical models
estimated through a calibration step (see \refsec{sec:calibration}). Consequently, we are able to map the visual
features $\vmat=\{\vvect_1,\ldots,\vvect_M\}$ onto the auditory space $\mathbb{A}$. We will denote the projection of
$\vvect_m$ by $\tilde{\vvect}_m$:
\[\tilde{\vvect}_m=({\cal A}\circ{\cal V}^{-1})(\vvect_m).\]
Summarizing, we use the mapping from $\mathbb{V}$ to $\mathbb{A}$ to map all visual features onto the auditory space.
Hence, all extracted features lie, now, in the same space, and we can perform the multimodal fusion in there.
\subsection{The Probabilistic Model}
Thanks to the link built in the previous section, we obtain a set of projected visual features
$\tilde{\vmat}=\{\tilde{\vvect}_1,\ldots,\tilde{\vvect}_M\}$, laying
in the same space as the auditory features $\amat$. These features need to be grouped to construct audio-visual
objects. However, we do not know which observation is generated by which object. Therefore, we introduce two sets of
hidden variables $\Zvect$ and $\Wvect$:
\begin{eqnarray*}
\Zvect &=& \{Z_1,\ldots,Z_m,\ldots,Z_M\} \\
\Wvect &=& \{W_1,\ldots,W_k,\ldots,W_K\},
\end{eqnarray*}
accounting for the observation-to-object assignment. The notation $Z_m=n$ ($m\in\{1,\ldots,M\}$,
$n\in\{1,\ldots,N+1\}$) means that the projected visual observation $\tilde{\vvect}_m$ was either generated by the
$n\th$ 3D object ($n\in\{1,\ldots,N\}$) or it is an outlier ($n=N+1$). Similarly, the variable $W_k$ is associated to
the auditory observation $\avect_k$.
We formulate the multimodal probabilistic fusion model under the assumption that all observations $\tilde{\vvect}_m$ and
$\avect_k$ are independent and identically distributed. The $n\th$ AV object generates both visual and auditory
features normally distributed around ${\cal A}(\Svect_n)$ and both the visual and auditory outliers are uniformly
distributed in $\mathbb{A}$. Therefore, we write:
\begin{equation*}
\textrm{P}(\tilde{\vvect}_m|Z_m=n,\Theta) = \left\{\begin{array}{ll}
{\cal N}(\tilde{\vvect}_m;\mu_n,\sigma_n) & n=1,\ldots,N \\
{\cal U}(\tilde{\vvect}_m;\mathbb{A}) & n=N+1 \\
\end{array}\right..
\end{equation*}
where $\Theta$ contains the Gaussian parameters, that is $\mu_n = {\cal A}(\Svect_n)$ and $\sigma_n$ (the mean and the
standard deviation of the $n\th$ Gaussian). The exact same rule holds for $\textrm{P}(\avect_m|W_m=n,\Theta)$. Thus we
can define a generative model for the observations $x\in\mathbb{A}$:
\begin{equation}
\p\left(x;\Theta\right) = \sum_{n=1}^N \pi_n\,{\cal N}(x;\mu_n,\sigma_n)
+ \pi_{N+1}\,{\cal U}(x;\mathbb{A}),
\label{eq:prob_model}
\end{equation}
where $\pi_n$ is the prior probabilities of the $n\th$ mixture component. That is,
$\pi_n=\textrm{P}(Z_m=n)=\textrm{P}(W_k=n)$, $\forall n,m,k$. The prior probabilities satisfy $\sum_{n=1}^{N+1}
\pi_n= 1$. Summarizing, the model parameters are:
\begin{equation}
\Theta = \{\pi_1,\ldots,\pi_{N+1},\mu_1,\ldots,\mu_N,\sigma_1,\ldots,\sigma_N\}.
\end{equation}
Under the probabilistic framework described, the set of parameters is estimated within a maximum likelihood
formulation:
\begin{equation}
{\cal L}\left(\tilde{\vmat},\amat;\Theta\right) =
\sum_{m=1}^M \log\textrm{p}\left(\tilde{\vvect}_m;\Theta\right) +
\sum_{k=1}^K\log\textrm{p}\left(\avect_k;\Theta\right).
\label{eq:log-like}
\end{equation}
In other words, the optimal set of parameters is the one maximizing the log-likelihood function \refeq{eq:log-like},
where $\p$ is the generative probabilistic model in \refeq{eq:prob_model}. Unfortunately, direct maximization of
\refeq{eq:log-like} is an intractable problem. Equivalently, the expected complete-data log-likelihood will be
maximized~\cite{Dempster77} (see \refsec{sec:inference}).
We recall that the ultimate goal is to determine the number $N$ of AV events, their 3D
locations $\Svect_1,\ldots,\Svect_n,\ldots,\Svect_N$ as well as their auditory activity $e_1,\ldots,e_n,\ldots,e_N$.
However, the 3D location parameters can be computed only indirectly, once the multimodal mixture's parameters $\Theta$
have been estimated. Indeed, once the auditory and visual observations are grouped in $\mathbb{A}$, the
$\tilde{\vvect}_m \leftrightarrow \vvect_m$ correspondences are used to infer the locations $\Svect_n$ of the AV
objects and the grouping of the auditory observations $\amat$ is used to infer the speaking state $e_n$ of the AV
objects. The choice of $N$ as well as the formulas for $\Svect_n$ and $e_n$ are given in Sections~\ref{sec:bic}
and~\ref{sec:output_estimators} respectively. Before these details are given and in order to fix ideas, we devote next
section to describe the auditory and visual features, justify the existence of ${\cal V}^{-1}$ and detail the
calibration procedure leading to a highly accurate mapping ${\cal A}\circ{\cal V}^{-1}$.
\section{Finding Auditory and Visual Features}
\label{sec:features}
In this section we describe the auditory (\refsec{sec:auditory_features}) and the visual (\refsec{sec:visual_features})
features we extract from the raw data. Given this features, the definition of ${\cal A}$ and ${\cal V}$ is
straightforward. However, the computation of the mapping's parameters is done through a calibration
procedure detailed in \refsec{sec:calibration}.
\subsection{Auditory Features}
\label{sec:auditory_features}
An auditory observation $\avect_k$ corresponds to an Interaural Time Difference (ITD) between the left and right
microphones. Because the ITDs are real-valued, the auditory feature space is $\mathbb{A}=\mathbb{R}$. One ITD value
corresponds to the different of time of arrival of the sound signal between the left and right microphones. For
instance, the sound wave of a speaker located in the left-half of the scene will obviously arrive earlier to the left
microphone than to the right microphone. We found that the method proposed in \cite{Christensen07} yields
very good results that are stable over time. The relationship between an auditory source located at $\Svect
\in\mathbb{R}^3$ and an ITD observation $\avect$ depends on the relative position of the acoustic source with respect to
the locations of the left and right microphones, $\Mvect_L$ and $\Mvect_R$. If we assume direct sound propagation and
constant sound velocity $\nu$, this relationship is given by the mapping ${\cal A}:\mathbb{S}\rightarrow\mathbb{A}$
defined as:
\begin{equation}
{\cal A}(\Svect) = \frac{\|\Svect-\Mvect_L\|-\|\Svect-\Mvect_R\|}{\nu} .
\label{eq:itd}
\end{equation}
\subsection{Visual Features}
\label{sec:visual_features}
The visual observations are 3D points extracted using binocular vision. We used two types of features: the
Harris-Motion 3D (HM3D) points and the faces 3D (F3D).
\begin{description}
\item [HM3D] The first kind of features we extracted are called Harris-motion points. We first detect Harris interest
points \cite{Harris88} in the left and right image pairs of the time interval $\Delta t$. Second, we only
consider a subset of theses points, namely those points where motion occurs. For each interest-point image location
$(u,v)$ we consider the image intensities at the same location $(u,v)$ in the subsequent images and we compute a
temporal intensity standard deviation $\tau_{(u,v)}$ for each interest point. Assuming stable lighting condition
over $\Delta t$, we simply classify the interest points into static ($\tau_{(u,v)} \leq \tau_M$) and dynamic
($\tau_{(u,v)} > \tau_M$) where $\tau_M$ is a user-defined threshold. Third, we apply a standard stereo matching
algorithm and a stereo reconstruction algorithm \cite{Hartley04} to yield a set of 3D features $\vmat$ associated with
$\Delta t$.
\item [F3D] The second kind of features are the 3D coordinates of the speakers' faces. They are obtained
using the face detector in~\cite{Waldboost-CVPR-2005}. More precisely, the center of the bounding box retrieved by the
face detector is matched to the right image and the same stereo reconstruction algorithm as in HM3D is used to
obtain $\vmat$.
\end{description}
Both 3D features are expressed in cyclopean coordinates~\cite{Hansard08}, which are also the scene coordinates.
Consequently, the visual mapping ${\cal V}$ is the identity mapping. In conclusion, because we are able to
accurately model the geometry of the visual sensors, we can assume that ${\cal V}$ is invertible and explicitly
build the linking mapping ${\cal A}\circ{\cal V}^{-1}$.
\subsection{Calibration}
\label{sec:calibration}
In the two previous sections we described the auditory and the visual features respectively. As a consequence, the
mappings ${\cal A}$ and ${\cal V}$ are defined. However, we made implicit use of two, a priory unknown, objects. On one
hand the stereo-matching and the 3D reconstruction algorithms need the so-called stereo-calibration. That is, the
projection matrices corresponding to the left and right cameras which are estimated using \cite{Bouguet08}. It
is worth to remark that the calibration procedure allows us to accurately represent any point in the field-of-view of
both cameras as a 3D point. On the other hand, and in order to use ${\cal A}$, we need to know the positions of the
microphones $\Mvect_L$ and $\Mvect_R$ in the scene coordinate frame, which is slightly more complex. Since the scene
coordinates are the same as the visual coordinates, we refer to this as ``audio-visual calibration''. We manually
measure the values of $\Mvect_L$ and $\Mvect_R$ with respect to the stereo rig. However, because these measurements are
imprecise, an affine correction model needs to be applied:
\begin{equation}
\overline{\cal A}(\Svect)=c_1\,{\cal A}(\Svect)+c_0 = c_1\,\frac{\|\Svect-\Mvect_L\|-\|\Svect-\Mvect_R\|}{\nu} + c_0,
\label{eq:itd_up}
\end{equation}
where $c_1$ and $c_0$ are the adjustment coefficients. In order to estimate $c_1$ and $c_0$, a person with a
speaker held just below the face moves in a zig-zag trajectory in the entire visual field of view of the
two cameras. The 3D position of the person's face and the ITD values were extracted. We used white
noise because it correlates very well resulting in a single sharp peak in ITD space. In many experiments, we
did not observe any effect of reverberations, because the reverberant components are suppressed by the direct component
of the long lasting white noise signal. The optimal values for $c_1$ and $c_0$, in the least square sense, were computed
from these data. \reffig{fig:calibration} shows the extracted ITDs (red-circle), the projected faces before (blue) and
after (green) the affine correction. We can clearly see how the affine transformation enhanced the audio-visual
linking mapping. Hence the projected visual features have the following expression:
\begin{equation}
\tilde{\vvect}_m = (\overline{\cal A}\circ{\cal V}^{-1})(\vvect_m) =
c_1\,\frac{\|\vvect_m-\Mvect_L\|-\|\vvect_m-\Mvect_R\|}{\nu} + c_0.
\label{eq:final_av_mapping}
\end{equation}
The outlined calibration procedure has three main advantages: (i) it requires very few training data, (ii) it lasts a
long period of time and (iii) it is environment-independent, thus guaranteeing the system's adaptability. Indeed, in our
case, the calibration ran on a one-minute audio-visual sequence and has been successfully used for the last 18
months in several rooms, including project demonstrations and conference exhibits. Consequently, the robustness of the
once-for-all tiny audio-visual calibration step is proved up to a large extent.
\begin{figure}[t]
\centering
\includegraphics[width=0.46\textwidth]{images/AV_calibration.pdf}
\caption{Affine correction of the audio-visual calibration. Extracted ITD values are plot in red-circled. F3D features
projected into the ITD space using Equation \refeq{eq:itd} are plot in blue. F3D features projected using Equation
\refeq{eq:final_av_mapping}, that is after the audio-visual calibration step, are plot in green.}
\label{fig:calibration}
\end{figure}
\section{Multimodal Inference}
\label{sec:inference}
In \refsec{sec:problem} we set up the maximum-likelihood framework to perform AV fusion. The 3D visual features are
mapped into the auditory space $\mathbb{A}$ through the audio-visual mapping $(\overline{\cal A}\circ{\cal V}^{-1})$.
This mapping takes the form in \refeq{eq:final_av_mapping} when using the auditory and visual features described in
\refsec{sec:features}. However, three of the initial issues remain unsolved: (i) the relative importance of each
modality, (ii) the estimates for $\Svect_n$ and $e_n$ and (iii) the variable number of AV objects, $N$. In this
Section we described EM-based method solving the ML problem with hidden variables and accounting for these unsolved
issues.
\subsection{Visual guidance}
\label{sec:visem}
Previous papers do not agree on how to balance the relative importance of each modality. After a deep analysis of the
features' statistics, we choose to use the visual information to guide the clustering process of the sparse auditory
observations. Indeed, because the HM3D visual features are more dense and have better temporal continuity than the ITD
values, we start by fitting a 1D GMM to the projected visual features $\{\tilde{\vvect}_m\}_{m=1}^M$. This is done with
the standard EM algorithm \cite{Bishop06}. In the E step of the algorithm the posterior probabilities $\alpha_{mn} =
\textrm{P}(Z_m=n|\tilde{\vmat},\Theta)$ are updated via the following formula:
\begin{equation}
\label{eq:alpha_m}
\alpha_{mn} = \frac{\pi_n\, \textrm{P}(\tilde{\vvect}_m|Z_m=n,\Theta)}{\sum_{i=1}^{N+1} \pi_i\,
\textrm{P}(\tilde{\vvect}_m|Z_m=i,\Theta)}.
\end{equation}
The M step is devoted to maximize the expected complete data log-likelihood with respect to the parameters, leading to
the standard formulas (with $\bar{\alpha}_n = \sum_{m=1}^M \alpha_{mn}$):
\begin{eqnarray*}
\pi_n &=& \frac{\bar{\alpha}_n}{M}, \\
\mu_n &=& \frac{1}{\bar{\alpha}_n}\sum_{m=1}^M \alpha_{mn}\tilde{\vvect}_m, \\
\sigma_n^2 &=& \frac{1}{\bar{\alpha}_n}\sum_{m=1}^M \alpha_{mn}(\tilde{\vvect}_m-\mu_n)^2.
\end{eqnarray*}
Once the model is fitted to the projected visual data, i.e., the visual information has already been probabilistically
assigned to the $N$ objects, the clustering process proceeds by including the auditory information. Hence, we are faced
with a constrained maximum-likelihood estimation problem: maximize \refeq{eq:log-like} subject to the constraint that
the posterior probabilities $\alpha_{mn}$ were previously computed. This leads to \textit{vision-guided EM fusion
algorithm} in which the E-step only updates the posterior probabilities associated with the auditory observations
while those associated with the visual observations remain unchanged. This semi-supervision strategy was introduced in
the context of text classification \cite{Nigam00,Miller03}. Here it is applied to enforce the quality and reliability of
one of the sensing modalities within a multimodal clustering algorithm. To summarize, the E-step of the algorithm
updates only the posterior probabilities of the auditory observations $\beta_{kn}=\textrm{P}(W_k =
n|\amat,\Theta)$:
\begin{equation}
\label{eq:beta_kn}
\beta_{kn} = \frac{\pi_n\,\textrm{P}(\avect_k|W_k = n,\Theta)}{\sum_{i=1}^{N+1}\pi_i\,\textrm{P}(\avect_k|W_k =
i,\Theta)},
\end{equation}
while keeping the visual posterior probabilities, $\alpha_{mn}$, constant.
The M-step has a closed-form solution and the prior probabilities are updated with:
\[ \pi_n = \frac{\gamma_n}{M+K},\quad n=1,\ldots,N+1, \]
with $\gamma_n = \sum_{m=1}^M \alpha_{mn} + \sum_{k=1}^K \beta_{kn} = \bar{\alpha}_n + \bar{\beta}_n$.
The means and variances of the current model are estimated by combining the two modalities:
\begin{equation}
\mu_n = \frac{1}{\gamma_n} \left(\sum_{m=1}^M \alpha_{mn}\,\tilde{\vvect}_m + \sum_{k=1}^K
\beta_{kn}\,\avect_k\right),
\label{eq:update_mu}
\end{equation}
\begin{equation}
\sigma^2_n = \frac{\sum_{m=1}^M \alpha_{mn}\,(\tilde{\vvect}_m-\mu_n)^2 +
\sum_{k=1}^K \beta_{kn}\,(\avect_k-\mu_n)^2}{\gamma_n}.
\label{eq:update_sigma}
\end{equation}
\subsection{Counting the number of speakers}
\label{sec:bic}
Since we do not know the value of $N$, a reasonable way to proceed is to estimate the parameters $\Theta_N$ for
different values of $N$ using the method delineated in the previous section. Once we estimated the maximum likelihood
parameters for models with different number of AV objects, we need a criterion to choose which is the best one. This is
estimating the number of AV objects (clusters) in the scene. BIC~\cite{Schwarz78} is a well known criterion to choose
among several maximum likelihood statistical models. BIC is often chosen for this type of tasks due to its attractive
consistency properties~\cite{Keribin00}. It is appropriate to use this criterion in our framework, due to the fact that
the statistical models after the \textit{vision-guided EM algorithm}, fit the AV data in an ML sense. In our case,
choosing among these models is equivalent to estimate the number of AV events $\hat{N}$. The formula to compute
the BIC score is:
\begin{equation}
\textrm{BIC}(\tilde{\vvect},\amat,\Theta_N) = {\cal L}\left(\tilde{\vvect},\amat;\Theta_N\right) -
\frac{D_N\,\log(M+K)}{2},
\label{eq:criteria}
\end{equation}
where $D_N = 3N$ is the number of free parameters of the model.
The number of AV events is estimated by selecting the statistical model corresponding to the maximum score:
\begin{equation}
\hat{N} = \arg\max\limits_N \textrm{BIC}(\tilde{\vvect},\amat,\Theta_N).
\label{eq:estimate-N}
\end{equation}
\subsection{Detection and localisation}
\label{sec:output_estimators}
The selection on $N$ leads to the best maximum-likelihood model in the BIC sense. That is, the set of parameters that
best explain the auditory and visual observations $\amat$ and $\tilde{\vmat}$. In the following, $\vmat$ are used
to estimate the 3D positions in the scene and $\amat$ to estimate the speaking state of each AV object.
The locations of the AV objects are estimated thanks to the one-to-one correspondence between 3D visual features and
the 1D projected features, $\tilde{\vvect}_m\leftrightarrow\vvect_m$. Indeed, the probabilistic assignments of the
projected visual data onto the 1D clusters, $\alpha_{mn}$, allow us to estimate $\Svect_n$ through:
\begin{equation}
\hat{\Svect}_n = \frac{1}{\bar{\alpha}_n} \sum_{m=1}^M \alpha_{mn} \vvect_m.
\label{eq:hat_svect}
\end{equation}
The auditory activity associated to the $n\th$ speaker is estimated as follows ($\tau_A$ is a user-defined threshold):
\begin{equation}
\hat{e}_n =
\left\{
\begin{array}{ll}
1& \mbox{if } \bar{\beta}_n > \tau_A\\
0 & \mbox{otherwise}
\end{array}
\right.
\label{eq:audio_act}
\end{equation}
This two formulas account for the last remaining issue: the 3D localization and speaking state estimation of the AV
objects. Next section describes some practical considerations to take into account when using this EM-based AV fusion
method. Afterward, in \refsec{sec:mgrh}, we summarize the method by providing an algorithmic scheme of the
multimodal inference procedure.
\subsection{Practical Concerns}
\label{sec:practical_concerns}
Even though the EM algorithm has proved to be the proper (and extremely powerful) methodology to solve ML problems with
hidden variables, in practice we need to overcome two main hurdles. First, since the log-likelihood function has many
local maxima and EM is a local optimization technique, a very good initialization is required. Second, because real
data is finite and may not strictly follow the generative law of probability \refeq{eq:prob_model}, the consistency
properties of the EM algorithm do not guarantee that the model chosen by BIC is meaningful regarding the application.
Thus, a post-processing step is needed in order to include the application-dependant knowledge. In all, we must
account for three practical concerns: (i) EM initialization, (ii) eventual shortage of observations and (iii) the
probabilistic model does not fully correspond to the observations.
It is reasonable to assume that the dynamics of the AV objects are somehow constrained. In other words, the positions
of the objects at a time interval are close to the positions at the previous time interval. Hence, we use the model
computed in the previous time interval to initialize the EM based procedure. More precisely, if we denote by $N^{(p)}$
the number of AV objects found in the previous time interval, we initialize a new 1D GMM with $N$ clusters,
for $N\in\{0,\ldots,N_{\textrm{max}}\}$. In the case $N\leq N^{(p)}$, we take the $N$ clusters with the
highest weight. For $N>N^{(p)}$, we incrementally split a cluster at its mean into two clusters. The cluster to be
split is selected on the basis of a high Davies-Bouldin index \cite{Davies79}:
\[ DW_i = \max_{j\neq i} \frac{\sigma_i + \sigma_j}{\|\mu_i - \mu_j\|}. \]
We chose to split the cluster into two clusters in order to detect AV objects that have recently appeared in
the scene, either because they were outside the field of view, or because they were occluded by another AV object. This
provides us with a good initialization. In our case the maximum number of AV objects is $N_{\textrm{max}} = 10$.
A shortage of observations usually leads to clusters whose interactions may describe an overall pattern, instead of
different components. We solve this problem by merging some of the mixture's components. There are several techniques to
merge clusters within a mixture model (see \cite{Henning10}). Since the components to be merged lie around the same
position and have similar spread, the \textit{ridgeline} method \cite{Surajit05} best solves our problem.
Finally, we need to face the fact that the probabilistic model does not fully represent the observations. Indeed, we
observed the existence of spurious clusters. Although the 3D visual observations associated with these clusters may be
uniformly distributed, their projections onto the auditory space $\tilde{\vvect}_m$ may form a spurious cluster. Hence
these clusters are characterized by having their points distributed near some hyperboloid in the 3D space (hyperboloids
are the level surfaces of the linking mapping defined in \refeq{eq:final_av_mapping}). As a
consequence, the volume of the back-projected 3D cluster is small. We discard those clusters whose covariance matrix
has a small determinant. Similarly as in \refeq{eq:hat_svect}, the clusters' covariance matrix is estimated
via:
\begin{equation}
\hat{\Sigmamat}_n = \frac{1}{\bar{\alpha}_n} \sum_{m=1}^M \alpha_{mn}
\left(\vvect_m-\hat{\Svect}_n\right)\left(\vvect_m-\hat{\Svect}_n\right)^\top.
\label{eq:hat_svect}
\end{equation}
\subsection{Motion-Guided Robot Hearing}
\label{sec:mgrh}
\refalg{alg:procedure} below summarizes the proposed method. It takes as input the visual (MH3D) and
auditory (ITD) observations gathered during a time interval $\Delta t$. The algorithm's output is the estimated
number of clusters $\hat{N}$, the estimated 3D positions of the AV events $\{\hat{\Svect}_n\}_{n=1}^{\hat{N}}$ as well
as their estimated auditory activity $\{\hat{e}_n\}_{n=1}^{\hat{N}}$. Because the grouping process is supervised by the
HM3D features, we name the procedure \textit{Motion-Guided Robot Hearing}. The algorithm starts by mapping the visual
observations onto the auditory space by means of the linking mapping defined in \refeq{eq:final_av_mapping}. Then, for
$N\in\{1,\ldots,N_\textrm{max}\}$ it iterates through the following steps: (a)~Initialize a model with $N$ components
using the output of the previous time interval (\refsec{sec:practical_concerns}), (b) apply EM using the selected $N$ to
model the 1D projections of the visual data (\refsec{sec:visem}), (c) apply the \textit{vision-guided EM fusion}
algorithm to both the auditory and projected visual data (\refsec{sec:visem}) in order to perform audio-visual
clustering, and (d) compute the BIC score associated with the current model, i.e., \refeq{eq:criteria}. This allows the
algorithm to select the model with the highest BIC score, i.e., \refeq{eq:estimate-N}. The post-processing step is then
applied to the selected model (\refsec{sec:practical_concerns}) prior to computing the final output
(\refsec{sec:output_estimators}).
\begin{algorithm}
\caption{Motion-Guided Robot Hearing}
\label{alg:procedure}
\begin{algorithmic}[1]
\STATE \textbf{Input:} HM3D, $\left\{\vvect_m\right\}_{m=1}^M$, and ITD, $\left\{\avect_k\right\}_{k=1}^K$, features.
\STATE \textbf{Output:} Number of AV events $\hat{N}$, 3D localization $\left\{\hat{\Svect}_n\right\}_{n=1}^{\hat{N}}$
and auditory status $\left\{\hat{e}_n\right\}_{n=1}^{\hat{N}}$
\STATE Map the visual features onto the auditory space, $\tilde{\vvect}_m = (\overline{\cal A}\circ{\cal
V}^{-1})(\vvect_m)$ \refeq{eq:final_av_mapping}.
\FOR{$N = 1 \to N_\textrm{max}$}
\STATE \textbf{(a) } { Initialize the model with $N$ clusters (\refsec{sec:practical_concerns}).}
\STATE \textbf{(b) } { Apply EM clustering to $\{\tilde{\vvect}_m\}_{m=1}^M$ (\refsec{sec:visem}).}
\STATE \textbf{(c) } { Apply the \textit{Vision-guided EM fusion} algorithm to cluster the audio-visual data
(\refsec{sec:visem}).}
\STATE \textbf{(d) } { Compute the BIC score \refeq{eq:criteria}}.
\ENDFOR
\STATE Estimate the number of clusters based on the BIC score \refeq{eq:estimate-N}.
\STATE Post-processing (\refsec{sec:practical_concerns}).
\STATE Compute the final outputs $\{\hat{\Svect}_n\}_{n=1}^{\hat{N}}$ and $\{\hat{e}_n\}_{n=1}^{\hat{N}}$
(\refsec{sec:output_estimators}).
\end{algorithmic}
\end{algorithm}
\section{Implementation on NAO}
\label{sec:robot_implem}
The previous multimodal inference algorithm has desirable statistical properties and good performance (see
\refsec{sec:results}). Since our final aim is to have a stable component working on a humanoid robot (i.e., able to
interact with other components), we reduced the computational load of the AV fusion algorithm. Indeed, we adapted the
method described in \refsec{sec:inference} to achieve a light on-line algorithm working on mobile robotic platforms.
In order to reduce the complexity, we substituted the Harris-Motion 3D point detector (HM3D) with the face 3D detector
(F3D), described in \refsec{sec:visual_features}. F3D replaces hundreds of HM3D points with a few face locations in 3D,
$\{\vvect_m\}_{m=1}^M$. We then consider that the potential speakers correspond to the detected faces. Hence we set
$N=M$ and $\Svect_n=\vvect_n$, $n=1,\ldots,N$. This has several crucial consequences. First, the number of AV objects
corresponds to the number of detected faces; the model selection step is not needed and the EM algorithm does
not have to run $N_{\textrm{max}}$ times, but just once. Second, because the visual features provide a good
initialization for the EM (by setting $\mu_n=(\overline{\cal A}\circ{\cal V}^{-1})(\Svect_n)$), the visual EM is not
required and the hidden variables $\Zvect$ do not make sense anymore. Third, since the visual features are not used as
observations in the EM, but to initialize it, the complexity of the \textit{vision-guide EM fusion} algorithm is ${\cal
O}(NK)$ instead of ${\cal O}\left(N(K+M)\right)$. This important because the number of HM3D points is much bigger than
the number of ITD values, i.e., $M\gg K$. Last, because the visual features provide the $\Svect_n$'s, there is not need
to estimate them through \refeq{eq:hat_svect}.
\subsection{Face-Guided Robot Hearing}
\label{sec:fgrh}
The resulting procedure is called \textit{Face-Guided Robot Hearing} and it is summarized
in \refalg{alg:procedure_simple} below. It takes as input the detected heads ($\Svect_1,\ldots,\Svect_N$) and the
auditory ($\amat$) observations gathered during a time interval $\Delta t$. The algorithm's output is the estimated
auditory activity $\{\hat{e}_n\}_{n=1}^{N}$.
\begin{algorithm}
\caption{Face-Guided Robot Hearing}
\label{alg:procedure_simple}
\begin{algorithmic}[1]
\STATE \textbf{Input:} Faces' position $\{\Svect_n\}_{n=1}^N$ and auditory $\left\{\avect_k\right\}_{k=1}^K$
features.
\STATE \textbf{Output:} AV objects' auditory status $\left\{\hat{e}_n\right\}_{n=1}^{\hat{N}}$.
\STATE Map the detected heads onto the auditory space, $\mu_n = (\overline{\cal A}\circ{\cal V}^{-1})(\Svect_n)$
\refeq{eq:final_av_mapping}.
\STATE Apply EM clustering to $\{\avect_k\}_{k=1}^K$ (\refsec{sec:visem}).
\STATE Compute the final outputs $\{\hat{e}_n\}_{n=1}^{\hat{N}}$ (\refsec{sec:output_estimators}).
\end{algorithmic}
\end{algorithm}
\subsection{System Architecture}
\label{sec:architecture}
We implemented our method using several components which are connected by a middleware called
Robotics Services Bus (RSB) \cite{Wienke2011}. RSB is a platform-independent event-driven middleware specifically
designed for the needs of distributed robotic applications. It is based on a logically unified bus which can span over
several transport mechanisms like network or in-process communication. The bus is hierarchically structured using scopes
on which events can be published with a common root scope. Through the unified bus, full introspection of the event flow
between all components is easily possible. Consequently, several tools exist which can record the event flow and replay
it later, so that application development can largely be done without a running robot. RSB events are automatically
equipped with several timestamps, which provide for introspection and synchronization abilities. Because of these
reasons RSB was chosen instead of NAO's native framework NAOqi and we could implement and test our algorithm remotely
without performance and deployment restrictions imposed by the robot platform. Moreover, the resulting implementation
can be reused for other robots.
One tool available in the RSB ecosystem is an event synchronizer, which synchronizes
events based on the attached timestamps with the aim to free application developers from such a generic task. However,
several possibilities of how to synchronize events exist and need to be chosen based on the intended application
scenario. For this reason, the synchronizer implements several strategies, each of them synchronizing events from
several scopes into a resulting compound event containing a set of events from the original scopes. We used two
strategies for the implementation.
The \emph{ApproximateTime} strategy is based on the algorithm available in \cite{ApproximateTime} and outputs sets of
events containing exactly one event from each scope. The algorithm tries to minimize the time between the earliest and
the latest event in each set and hence well-suited to synchronize events which originate from the same source (in the
world) but suffered from perception or processing delays in a way that they have non-equal timestamps. The
second algorithm, \emph{TimeFrame}, declares one scope as the primary event source and for each event received here, all
events received on other scopes are attached that lie in a specific time frame around the timestamp of the source event.
\emph{ApproximateTime} is used in our case to synchronize the results from the left and right camera as frames in
general form matching entities but due to independent grabbing of both cameras have slightly different timestamps.
Results from the stereo matching process are synchronized with ITD values using the \emph{TimeFrame} strategy because
the integration time for generating ITD values is much smaller than for a vision frame and hence multiple ITD values
belong to a single vision result.
\subsection{Modular Structure}
\label{sec:modular_structure}
The implementation is divided into components shown in the pipeline of \reffig{fig:pipeline}. Components are
color-coded: modules provided by the RSB middleware (white), auditory (red) and visual (green) processing, audio-visual
fusion (purple) and the visualization tool (blue) described at the end of this Section.
The visual processing is composed by five modules. \emph{Left video} and \emph{Right video} stream the images
received at left and right cameras. The \emph{Left face detection} module extracts the faces from the left image.
These are then synchronized with the right image in \emph{Face-image synchronization}, using the \emph{ApproximateTime}
strategy. The \emph{F3D Extraction} module computes the F3D features. A new audio-visual head for NAO was used for this
implementation. The new head (see \reffig{fig:nao_head}) is equipped with a pair of cameras and four microphones, thus
providing a synchronized VGA stereoscopic image flow as well as four audio channels.
\begin{figure}
\centering
\includegraphics[width=0.4\linewidth]{images/nao-new}
\caption{Within this work we used a new audio-visual head that is composed of a synchronized camera pair and two
microphones. This ``orange'' head replaces the former ``blue'' head and is fully interfaced by the RSB middleware
previously described in this section.}
\label{fig:nao_head}
\end{figure}
The auditory component consists of three modules. Interleaved audio samples coming from the four microphones of NAO are
streamed by the \emph{Interleaved audio} module. The four channels are deinterleaved by the \emph{Sound
deinterleaving} module, which outputs the auditory flows corresponding to the left and right microphones. These flows
are stored into two circular buffers in order to extract the ITD values (\emph{ITD extraction} module).
Both visual and auditory features flow until the \emph{Audio-visual synchronization} module; the \emph{TimeFrame}
strategy is used here to find the ITD values coming from the audio pipeline associated to the 3D positions of the
faces coming from the visual processing. These synchronized events feed the \emph{Face-guided robot hearing} module,
which is in charge of estimating the speaking state of each face, $e_n$.
Finally, we developed the module \emph{Visualization}, in order to get a better insight of the proposed algorithm. A
snapshot of this visualization tool can be seen in \reffig{fig:vtool}. The image consists of three parts. The top-left
part with a blue frame is the original left image plus one rectangle per detected face. In addition to the face's
bounding box, a solid circle is plot on the face of the actor codding the emitting sound probability, the higher it is,
the darker the circle. The top-right part, framed in green, is a bird-view of the scene, in which the detected heads
appear as circles. The bottom-left part, with a red frame, represents the ITD space. There, both the mapped heads
(ellipses) and the histogram of ITD values are plot.
\begin{figure}
\centering
\includegraphics[width=0.6\textwidth]{images/visualization3.jpg}
\caption {Snapshot of the visualization tool. The top-left (blue-framed) image is the original left image plus one
bounding box per detected face. In addition, an intensity-coded circle appears when the speaker is active. The darker
the color is, the higher the speaking probability is. The top-right (green-framed) image corresponds to the bird-view of
the scene, in which each circle corresponds to a detected head. The bottom-left (red-framed) image represents the ITD
space. The projected faces are represented by an ellipse and the histogram of extracted ITD values is plot.}
\label{fig:vtool}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=0.6\linewidth]{images/functionalDiagram}
\caption {Modular structure of the \textit{Face-Guided Robot Hearing} procedure implemented on NAO. There are five types
of modules: streaming \& synchronization (white), visual processing (green), auditory processing (red), audio-visual
fusion (purple) and visualization (blue).}
\label{fig:pipeline}
\end{figure}
\subsection{Implementation Details}
Some details need to be specified regarding the implementation of the face-guided robot hearing method. First, the
integration window $F$ and the frame shift $f$ of the ITD extraction procedure. The bigger the integration window is
the more reliable the ITD values are and the more expensive its computation becomes. Similarly, the smaller $f$ is the
more ITD observations are extracted and the more computational load we have. A good compromise between low computational
load, high rate, and reliability of ITD values was found for $W=150$~ms and $f=20$~ms. We also used an activity
threshold: when the energy of the sound signals is lower than $E_A=0.001$, the window is not processed. Thus saving
computational time for other components in the system when there are no emitted sounds. Notice that this parameter could
be controlled by a higher level module which would learn the characteristics of the scene and infer the level of
background noise. We initialize $\sigma_n^2=10^{-9}$, since we found this value big enough to take into account the
noise in the ITD values and small enough to discriminate speakers that are close to each other. The threshold $\tau_A$
has to take into account how many audio observations ($K$) are gathered during the current time interval $\Delta t$ as
well as the number of potential audible AV objects ($N$). For instance, if there is just one potential AV object, most
of the audio observations should be assigned to it, whereas if there are three of them the audio observations may be
distributed among them (in case all of them emit sounds). The threshold $\tau_A$ was experimentally set to $\tau_A = K /
(N+2)$. The entire pipeline was running on a laptop with an i7 processor at $2.5$~GHz.
\section{Results}
\label{sec:results}
In order to evaluate the proposed approach, we ran three sets of experiments. First, we evaluated the Multimodal
Inference method described in \refsec{sec:inference} on synthetic data. This allowed us to assess the quality of the
model on a controlled scenario, where the feature extraction did not play any role. Second, we evaluated the
\textit{Motion-Guided Robot Hearing} method on a publicly available dataset, thus assessing the quality of the entire
approach. Finally, we evaluated the \textit{Face-Guided Robot Hearing} implemented on NAO, which proves that the
proposed hybrid deterministic/probabilistic framework is suitable for robot applications.
In all our experiments we used a time interval of 6 visual frames, $\Delta t=0.4s$; time in which approximately 2,000
HM3D observations and 20 auditory observations are extracted. A typical set of visual and auditory observations are
shown in Figures~\ref{fig:example-Vobservations} and~\ref{fig:example-AVobservations}. Indeed,
\reffig{fig:example-Vobservations} focuses on the extraction of the HM3D features: the Harris interest point
detection, filtered by motion, matched between images and reconstructed in 3D. \reffig{fig:example-AVobservations}
shows the very same 3D features projected in to the ITD space. Also, the ITD values extracted during the same time
interval are shown. These are the input features of the \textit{Motion-Guided Robot Hearing} procedure. Notice that both
auditory and visual data are corrupted by noise and by outliers. Visual data suffer from reconstruction errors either
from wrong matches or from noisy detection. Auditory data suffer from reverberations, which enlarge the pics' variances,
or from sensor noise which is sparse along the ITD space.
\begin{figure}
\centering
\begin{tabular}{cc}
\subfloat[]{\includegraphics[width=0.3\columnwidth]{images/leftHarrisPoints1Frame_173.pdf}} &
\subfloat[]{\includegraphics[width=0.3\columnwidth]{images/rightHarrisPoints1Frame_173.pdf}} \\
\subfloat[]{\includegraphics[width=0.3\columnwidth]{images/leftHarrisMotionPoints1Frame_173.pdf}} &
\subfloat[]{\includegraphics[width=0.3\columnwidth]{images/rightHarrisMotionPoints1Frame_173.pdf}}\\
\multicolumn{2}{c}{\subfloat[]{\includegraphics[width=0.3\columnwidth]{images/Points3D_173.pdf}}}\\
\end{tabular}
\caption{Interest points as detected in the left (a) and right (b) images. Dynamic interest points detected in the left
(c) and the right (d) images. (e) HM3D visual observations, $\{\vvect_m\}_{m=1}^M$. Most of the background (hence
static) points are filtered out from (a) to (c) and from (b) to (d). It is worth noticing that the reconstructed HM3D
features suffer from reconstruction errors.}
\label{fig:example-Vobservations}
\end{figure}
\begin{figure}
\centering
\begin{tabular}{cc}
\subfloat[]{\includegraphics[width=0.3\columnwidth]{images/ProjectedVisual_173.pdf}} &
\subfloat[]{\includegraphics[width=0.3\columnwidth]{images/NewAuditoryHistogram.pdf}}
\end{tabular}
\caption{Observation densities in the auditory space $\mathbb{A}$: (a) of the projected HM3D features,
$\{\tilde{\vvect}_m\}_{m=1}^M$, and (b) of the ITD features, $\{\avect_k\}_{k=1}^K$. In this particular example, we
observe three moving objects (corresponding to the three people in the images). In addition, two of them are emitting
sound (left and middle) and one is silent (right). We remark that auditory as well as visual observations are
contaminated by noise (enlarging the Gaussian variances) and by outliers (uniformly distributed in the auditory
feature space).}
\label{fig:example-AVobservations}
\end{figure}
To quantitatively evaluate the localization results, we compute a distance matrix between the detected clusters and
the ground-truth clusters. The cluster-to-cluster distance corresponds to the Euclidean distance between cluster means.
Let $\mathbf{D}$ be the distance matrix, then entry $D_{ij}=\|\mu_i-\hat{\mu}_j\|$ is the distance from the $i\th$
ground-truth cluster to the $j\th$ detected cluster. Next, we associate at most one ground-truth cluster to each
detected cluster. The assignment procedure is as follows. For each detected cluster we compute its ground-truth
nearest cluster. If it is not closer than a threshold $\tau_{\textrm{loc}}$ we mark it as a \textit{false positive},
otherwise we assign the detected cluster to the ground-truth cluster. Then, for each ground-truth cluster we determine
how many detected clusters are assigned to it. If there is none, we mark the ground-truth cluster as \textit{false
negative}. Finally, for each of the remaining ground-truth clusters, we select the closest (\textit{true positive})
detected cluster among the ones assigned to the ground-truth cluster and we mark the remaining ones as \textit{false
positives}. We can evaluate the localization error and the auditory state for those clusters that have been
correctly detected . The localization error corresponds to the Euclidean distance between the means. Notice that by
choosing $\tau_{\textrm{loc}}$, we fix the maximum localization error allowed. The auditory state is counted as
\textit{false positive} if detected audible when silent, \textit{false positive} if detected silent when audible and
\textit{true positive} otherwise. $\tau_{\textrm{loc}}$ was set to $0.35\,\textrm{m}$ in all the experiments.
\subsection{Results on Synthetic Data}
\label{sec:synthetic_data}
Four synthetic sequences containing one to three AV objects were generated. These objects can move and they are not
necessarily visible/audible along the entire sequence. \reftab{tab:video_results_synth} shows the visual evaluation of
the method when tested with synthetic sequences. The sequence code name describes the dynamic character of the sequence
(\textit{Sta} means static and \textit{Dyn} means dynamic) and the varying number of AV objects in the scene
(\textit{Con} means constant number of AV objects and \textit{Var} means varying number of AV objects). The columns show
different evaluation quantities: FP (\textit{false positives}), i.e., AV objects found that do not really exist, FN
(\textit{false negatives}), i.e., present AV objects that were not found, TP (\textit{true positives}) and ALE (average
localization error). Recall that we can compute the localization error just for the true positives. First, we observe
that the right detection rate is always above 65\%, increasing to 96\% in the case where there are 3 visible static
clusters. We also observe that the fact that the number of AV objects in the scene varies does not impact the
localization error. The effect on the localization error is due, hence, to the dynamic character of the scene; if the AV
objects move or not. The third observation is that both the dynamic character of the scene and the varying number of
clusters have a lot of impact on the detection rate.
\begin{table}
\caption{Visual evaluation of results obtained with synthetic sequences. \textit{Sta}/\textit{Dyn} states for static or
dynamic scene; the AV objects move or do not move. \textit{Var}/\textit{Con} states for varying or constant number of AV
objects. FP stands for false positives, FN for false negatives, TP for true positives and ALE for average
localization
error (expressed in meters).\vspace{0.2cm}}
\label{tab:video_results_synth}
\centering
{\small
\begin{tabular}{lcccc}
\toprule
Seq. & FP & FN & TP & ALE [m] \\
\midrule
\textit{StaCon} & 12 & 16 (3.9\%) & 392 (96.1\%) & 0.03 \\
\textit{DynCon} & 43 & 139 (34.1\%) & 269 (65.9\%) & 0.10 \\
\textit{StaVar} & 46 & 69 (30.1\%) & 160 (69.9\%) & 0.03 \\
\textit{DynVar} & 40 & 82 (35.9\%) & 147 (64.1\%) & 0.11 \\
\bottomrule
\end{tabular}}
\end{table}
\begin{table}
\caption{Audio evaluation of the results obtained with synthetic sequences. \textit{Sta}/\textit{Dyn} states for static
or dynamic scene; the AV objects move or do not move. \textit{Var}/\textit{Con} states for varying or constant number of
AV objects.\vspace{0.2cm}}
\label{tab:audio_results_synth}
\centering
{\small
\begin{tabular}{lccc}
\toprule
Seq. & FP & FN & TP \\
\midrule
\textit{StaCon} & 161 & 33 (13.4\%) & 214 (86.6\%) \\
\textit{DynCon} & 144 & 56 (21.2\%) & 208 (78.8\%) \\
\textit{StaVar} & 53 & 33 (18.8\%) & 143 (81.2\%) \\
\textit{DynVar} & 56 & 34 (19.7\%) & 139 (80.3\%) \\
\bottomrule
\end{tabular}}
\end{table}
\reftab{tab:audio_results_synth} shows the auditory evaluation of the method when tested with synthetic sequences. The
remarkable achievement is the high number of right detections, around 80\%, in all cases. This means that neither the
dynamic character of the scene nor the fact that the number of AV objects varies have an impact on sound detection. It
is also true that the number of false positives is large in all the cases.
\subsection{Results on Real Data}
\label{sec:results_real}
The \textit{Motion-Guided Robot Hearing} method was tested on the CTMS3 sequence of the CAVA data set \cite{Arnaud08}.
The CAVA (\textit{computational audio-visual analysis}) data set was specifically recorded to test various real-world
audio-visual scenarios. The CTMS3
sequence\footnote{\url{http://perception.inrialpes.fr/CAVA_Dataset/Site/data.html\#CTMS3}} consists on three
people freely moving in a room and taking speaking turns. Two of them count in English (one, two, three, ...) while the
third one counts in Chinese. The recorded signals, both auditory and visual, enclose the difficulties found in natural
situations. Hence, this is a very challenging sequence: People come in and out the visual field of the two cameras, hide
each other, etc. Aside from the speech sounds, there are acoustic reverberations and non-speech sounds such as those
emitted by foot steps and clothe chafing. Occasionally, two people speak simultaneously.
\begin{figure}
\centering
\begin{tabular}{ccc}
\subfloat[]{\label{fig:ctms3_179}\includegraphics[width=0.3\linewidth]{images/full_179}} \hspace{-0.3cm} &
\subfloat[]{\label{fig:ctms3_222}\includegraphics[width=0.3\linewidth]{images/full_222}} \hspace{-0.3cm} &
\subfloat[]{\label{fig:ctms3_254}\includegraphics[width=0.3\linewidth]{images/full_254}} \\
\subfloat[]{\label{fig:ctms3_307}\includegraphics[width=0.3\linewidth]{images/full_307}} \hspace{-0.3cm} &
\subfloat[]{\label{fig:ctms3_259}\includegraphics[width=0.3\linewidth]{images/full_259}} \hspace{-0.3cm} &
\subfloat[]{\label{fig:ctms3_297}\includegraphics[width=0.3\linewidth]{images/full_297}} \\
\subfloat[]{\label{fig:ctms3_349}\includegraphics[width=0.3\linewidth]{images/full_349}} \hspace{-0.3cm} &
\subfloat[]{\label{fig:ctms3_191}\includegraphics[width=0.3\linewidth]{images/full_191}} \hspace{-0.3cm} &
\subfloat[]{\label{fig:ctms3_204}\includegraphics[width=0.3\linewidth]{images/full_204}} \\
\end{tabular}
\caption{Results obtained with the CTMS3 sequence from the CAVA data set. The ellipses correspond to the 3D covariance
matrices projected onto the image. The circle at each ellipse center illustrates the auditory activity: speaker emitting
a sound (white) or being silent (black) during each time interval. The plot associated with each image shows the
auditory observations as well as the fitted 1D mixture model.}
\label{fig:results_ctms3}
\end{figure}
\reffig{fig:results_ctms3} shows the results obtained with nine time intervals chosen to show both successes and
failures of our method and to allow to qualitatively evaluate it. \reffig{fig:ctms3_179} shows one extreme case, in
which the distribution of the HM3D observations associated to the person with the white T-shirt is clearly not Gaussian.
\reffig{fig:ctms3_222} shows a failure of the \textit{ridgeline} method, used to merge Gaussian components, where
two different clusters are associated into one. \reffig{fig:ctms3_254} is an example with too few observations.
Indeed, the BIC points as optimal the model with no AV objects, thus considering all the observations to be outliers.
\reffig{fig:ctms3_307} clearly shows that our approach cannot deal with occluded objects, because of the
instantaneous processing of robocentric data, the person occluded will never be detected.
Figures~\ref{fig:ctms3_259},~\ref{fig:ctms3_297} and~\ref{fig:ctms3_349} are examples of success. The three speakers are
localised and their auditory status correctly guesses. However, the localisation accuracy is not good in these cases,
because one or more covariance matrices are not correctly estimated. The grouping of AV observations is, then, not well
conducted. Finally, Figures~\ref{fig:ctms3_191} and~\ref{fig:ctms3_204} show two case in which the \textit{Motion-Guided
Robot Hearing} algorithms works perfectly, three people are detected and their speaking activity is correctly assessed
from the ITD observations. In average, the method correctly detected 187 out of 213 objects (87.8\%) and correctly
detected the speaking state in 88 cases out of 147 (59.9\%).
\subsection{Results on NAO}
\label{sec:results_2}
To validate the \textit{Face-Guided Robot Hearing} method using NAO, we performed a set of experiments with five
different scenarios. The scenarios were recorded in a room around $5\times5$ meters with just a sofa and 3 chairs where
NAO and the other persons sat respectively. We designed five scenarios to test the algorithm in different conditions in
order to identify its limitations. Each scenario is repeated several times and consists on people counting from one up
to sixteen.
In scenario \textbf{S1}, only one person is in the room sitting in front of the robot and
counting. In the rest of the scenarios (\textbf{S2}-\textbf{S5}) three persons are in the room. People are not always in
the field of view (FoV) of the cameras and sometimes they move. In scenario \textbf{S2} three persons are sitting and
counting alternatively one after the other. The configuration of scenario \textbf{S3} is similar to the one of
\textbf{S2}, but one person is standing instead of sitting. These two scenarios are useful to determine the precision of
the ITDs and experimentally see if the difference of height (elevation) affects the quality of the extracted ITDs. The
scenario \textbf{S4} is different from \textbf{S2} and \textbf{S3} because one of the actors is outside the FoV. This
scenario is used to test if people speaking outside the FoV affect the performance of the algorithm. In the last
scenario (\textbf{S5}) the three people are in the FoV, but they count and speak independently of the other actors.
Furthermore, one of them is moving while speaking. With \textbf{S5}, we aim to test the robustness of the method to
dynamic scenes.
In \reffig{fig:results} we show several snapshots of our visualization tool. These frames are selected from the
different scenarios aiming to show both the successes and the failures of the implemented system. \reffig{fig:s1} shows
an example of perfect alignment between the ITDs and the mapped face, leading to a high speaking probability. A similar
situation is presented in \reffig{fig:s21}, in which among the three people, only one speaks. A failure of the ITD
extractor is shown in \reffig{fig:s4}, where the actor in the left is speaking, but no ITDs are extracted. In
\reffig{fig:s51} we can see how the face detector does not work correctly: two faces are missing, one because of the
great distance between the robot and the speaker, and the other because it is partially out of the field of
view. \reffig{fig:s52} shows a snapshot of an AV-fusion failure, in which the extracted ITDs are not significant enough
to set a high speaking probability. The \reffig{fig:s22}, \reffig{fig:s31} and \reffig{fig:s32} show the effect of
reverberations. While in \reffig{fig:s32} we see that the reverberations lead to the wrong conclusion that the actor
on the right is speaking, we also see that the statistical framework is able to handle
reverberations (\reffig{fig:s22} and \reffig{fig:s31}), hence demonstrating the robustness of the proposed
approach.
\reftab{tab:scenarios} shows the results obtained on scenarios (that were manually annotated). First of all we notice
the small amount of false negatives: the system misses very few speakers. A part from the first scenario (easy
conditions), we observe some false positives. These false positives are due to reverberations. Indeed, we notice how the
percentage of FP is severe in \textbf{S5}. This is due to the fact that high reverberant sounds (like hand claps) are
also present in the audio stream of this scenario. We believe that an ITD extraction method more robust to
reverberations will lead to more reliable ITD values, which in turn will lead to a better active speaker detector. It is
also worth to notice that actors in different elevations and non-visible actors do not affect the performance of the
proposed system, since the results obtained in scenarios \textbf{S2} to \textbf{S4} are comparable.
\begin{table}
\centering
{\small
\begin{tabular}{cccc}
\toprule
& FP & FN & TP\\
\midrule
\textbf{S1} & 13 & 23 (13.4\%) & 149 (86.6\%)\\
\textbf{S2} & 22 & 31 (14.9\%) & 176 (85.1\%)\\
\textbf{S3} & 19 & 20 (11.3\%) & 157 (88.7\%)\\
\textbf{S4} & 37 & 12 (6.7\%) & 166 (93.3\%)\\
\textbf{S5} & 53 & 32 (19.0\%) & 136 (81.0\%)\\
\bottomrule
\end{tabular}}
\caption{Quantitative evaluation of the proposed approach for the five scenarios. The columns represent, in order: the
amount of correct detections (CD), the amount of false positives (FP), the amount of false negatives (FN) and the
total number of counts (Total).}
\label{tab:scenarios}
\end{table}
\begin{figure}
\centering
\begin{tabular}{cc}
\subfloat[\textbf{S1}]{\label{fig:s1}\includegraphics[width=0.45\linewidth]{images/frame_0000000255.jpg}}
&
\subfloat[\textbf{S2}]{\label{fig:s21}\includegraphics[width=0.45\linewidth]{images/frame_0000000315.jpg}}
\\
\subfloat[\textbf{S4}]{\label{fig:s4}\includegraphics[width=0.45\linewidth]{images/frame_0000000264.jpg}}
&
\subfloat[\textbf{S5}]{\label{fig:s51}\includegraphics[width=0.45\linewidth]{images/frame_0000000691.jpg}}
\\
\subfloat[\textbf{S5}]{\label{fig:s52}\includegraphics[width=0.45\linewidth]{images/frame_0000000142.jpg}}
&
\subfloat[\textbf{S2}]{\label{fig:s22}\includegraphics[width=0.45\linewidth]{images/frame_0000000326.jpg}}
\\
\subfloat[\textbf{S3}]{\label{fig:s31}\includegraphics[width=0.45\linewidth]{images/frame_0000000390.jpg}}
&
\subfloat[\textbf{S3}]{\label{fig:s32}\includegraphics[width=0.45\linewidth]{images/frame_0000000285.jpg}}
\\
\end {tabular}
\caption{Snapshots of the visualization tool. Frames selected among the five scenarios to show the method's strengths
and weaknesses. The faces' bounding box are shown superposed to the original image (top-left). The bird-view of the
scene is shown in the top-right part of each subimage. The histogram of ITD values as well as the projected faces are
shown in the bottom-left. See \refsec{sec:modular_structure} for how to interpret the images above.}
\label{fig:results}
\end{figure}
\section{Conclusions and Future Work}
\label{sec:conc}
This paper introduces a multimodal hybrid probabilistic/deterministic framework for simultaneous detection and
localization of speakers. On one hand, the deterministic component takes advantage of the geometric and physical
properties associated with the visual and auditory sensors: the audio-visual mapping $(\overline{\cal A}\circ{\cal V})$
allows us to transform the visual features from the 3D space to an 1D auditory space. On the other hand,
the probabilistic model deals with the observation-to-speaker assignments, the noise and the outliers. We propose a new
multimodal clustering algorithm based on a 1D Gaussian mixture model, an initialization procedure, and a model selection
procedure based on the BIC score. The method is validated on a humanoid robot and interfaced through the RSB middleware
leading to a platform-independent implementation.
The main novelty of the approach is the visual guidance. Indeed, we derived to EM-based procedures for
\textit{Motion-Guided} and \textit{Face-Guided} robot hearing. Both algorithms provide the number of speakers,
localize them and ascertain their speaking status. In other words, we show how one of the two modalities can be
used to supervise the clustering process. This is possible thanks to the audio-visual calibration procedure that
provides an accurate projection mapping $(\overline{\cal A}\circ{\cal V})$. The calibration is specifically designed
for robotic usage since it requires very few data, it is long-lasting and environment-independent.
The presented method solves several open methodological issues: (i) it fuses and clusters visual and auditory
observations that lie in physically different spaces with different dimensionality, (ii) it models and estimates the
object-to-observation assignments that are not known, (iii) it handles noise and outliers mixed with both visual and
auditory observations whose statistical properties change across modalities, (iv) it weights the relative importance of
the two types of data, (v) it estimates the number of AV objects that are effectively present in the scene during a
short time interval and (vi) it gauges the position and speaking state of the potential speakers.
One prominent feature of our algorithm is its robustness. It can deal with various kinds of perturbations, such as the
noise and outlier encountered in unrestricted physical spaces. We illustrated the effectiveness and robustness of our
algorithm using challenging audio-visual sequences from a publicly available data set as well as using the humanoid
robot NAO in regular indoor environments. We demonstrated good performance on different scenarios involving
several actors, moving actors and non-visible actors. Interfaced by means of the RSB middleware, the
\textit{Face-Guided Robot Hearing} method processes the audio-visual data flow from two microphones mounted inside the
head of a companion robot with noisy fans and two cameras at a rate of 17~Hz.
There are several possible ways to improve and to extend our method. Our current implementation relies more on the
visual data than on the auditory data, although there are many situations where the auditory data are more reliable. The
problem of how to weight the relative importance of the two modalities is under investigation. Our algorithm can also
accommodate other types of visual cues, such as 2D or 3D optical flow, body detectors, etc., or auditory cues, such as
Interaural Level Differences. In this paper we used one pair of microphones, but the method can be easily extended to
several microphone pairs. Each microphone pair yields one ITD space and combining these 1D spaces would provide a much
more robust algorithm. Finally, another interesting direction of research is to design a dynamic model
that would allow to initialize the parameters in one time interval based on the information extracted in several
previous time intervals. Such a model would necessarily involve dynamic model selection, and would certainly help to
guess the right number of AV objects, particularly in situations where a cluster is occluded but still in the visual
scene, or a speaker is highly interfered by another speaker/sound source. Moreover, this future dynamic model selection
should be extended to provide for audio-visual tracking capabilities, since they enhance the temporal coherence of the
perceived audio-visual scene.
\section*{Acknowledgments}
This work was partially funded by the HUMAVIPS FP7 European Project FP7-ICT-247525.
\bibliographystyle{these}
|
\section{Introduction}
\label{sec:intro}
Virtual machines provide a uniform and portable environment for high energy physics applications~\cite{hepvirt11}.
The same virtual machine can be used for both development and for deployment on cloud infrastructures so that very same operating system libraries that are used to develop physics algorithms are also used to execute them.
As such, virtual machines provide a way to escape the so called ``dependency hell'', i.\,e.\ the problem that applications might not work or behave differently when using slightly different (versions of) system libraries.
Furthermore, virtual machines remove the need to port applications to different operating systems.
This is one of the key benefits when it comes to working with volunteers that provide free CPU cycles on a variety of platforms (Windows, Linux, Mac)~\cite{lhcathome12}.
Properly archived, virtual machines are also an important ingredient for the long-term preservation of experiment data because they contain the complex historic software stack necessary to interpret data files~\cite{dphep12}.
The hard disk resembling a virtual machine is distributed in the form of a hard disk image file.
There are two contradicting approaches to creating virtual machine images.
The first approach creates a hard disk image from a standard installation of an operating system.
Such images fit a large base of different applications and use cases, such as web servers, machines for interactive login, or graphical workstations.
The image size is rather large, typically a few gigabytes, and due to the many packages contained in an image, images quickly become outdated.
The second approach creates a \emph{virtual appliance}.
The virtual appliance comprises a ``just enough operating system'', that is a stripped down version of an operating system tailored to only one or few applications and use cases.
An image size of a virtual appliance can be as small as a few hundred megabytes.
As we have seen maintaining the CernVM virtual appliance\footnote{\url{http://cernvm.cern.ch}}, even for a minimalistic image it can take hours to prepare package repositories, build and compress virtual machine images in various formats, and to test them.
This delay between making a change and verifying the result of the change restricts the speed at which the virtual appliance can be developed.
On the user's end, despite the fact that sophisticated tools have been developed to manage virtual machine images~\cite{vmfs07,qcow208,sheepdog10,imagedist11,swift12}, the image distribution problem is best mitigated by small images.
With all its libraries, tools, and configuration data, an operating system is a complex software stack.
The problem of distributing complex and frequently changing software stacks to virtual machines were previously faced for experiment applications.
Experiment applications are now widely distributed by the CernVM File System (CernVM-FS)~\cite{cvmfs11}, a caching, read-only file system that downloads software bits and pieces on demand from an HTTP content delivery network.
Here, we bring the use of CernVM-FS to the next level and use it to distribute as well the operating system to a virtual machine.
The result is a tiny virtual machine image only consisting of a Linux kernel and the CernVM-FS client.
The actual operating system is then booted from CernVM-FS.
Instead of a ``just enough operating system'', this approach leads to an \emph{operating system on demand}.
The rest of the paper is structured as follows.
Section~\ref{sec:components} describes the components of a $\mu$CernVM system.
Section~\ref{sec:stack} describes the virtual machine's root file system stack and the connection between $\mu$CernVM and the operating system on CernVM-FS.
Section~\ref{sec:os} describes the maintenance of an operating system on CernVM-FS; in particular we focus on exploiting the internal versioning of CernVM-FS that allows the very same virtual machine image to instantiate any operating system ever released on CernVM-FS.
Section~\ref{sec:updates} describes how $\mu$CernVM based systems are updated.
Section~\ref{sec:related} compares $\mu$CernVM to similar approaches.
\section{Building blocks}
\label{sec:components}
\begin{figure}
\caption{
A $\mu$CernVM based virtual machine is twofold.
The $\mu$CernVM image contains a Linux kernel, the AUFS union file system, and a CernVM-FS client.
The CernVM-FS client connects to a special repository containing the actual operating system.
The two CernVM-FS repositories contain the operating system and the experiment software.
}
\label{fig:components}
\begin{center}
\input{img/uconcept-size}
\end{center}
\end{figure}
Figure~\ref{fig:components} shows the key components of the $\mu$CernVM image.
The image consists of a Linux kernel and an init ramdisk that includes the CernVM-FS client.
The Linux kernel is ``virtualization friendly''.
It is slim as it only requires device drivers for the handful of hypervisors around.
Table~\ref{tab:kernel} compares the size of the $\mu$CernVM kernel with an Scientific Linux 6 kernel.
On the other hand, the $\mu$CernVM kernel has the latest para-virtualized drivers which are not necessarily found in Scientific Linux or other distribution kernels.
\begin{table}
\caption{\label{tab:kernel}Comparison of the size of the $\mu$CernVM kernel and the Scientific Linux 6 kernel}
\begin{center}
\input{img/tabkernel}
\end{center}
\end{table}
Furthermore, the $\mu$CernVM kernel contains the \emph{union file system}~\cite{unionfs04} AUFS\footnote{\url{http://aufs.sourceforge.net}}.
The union file system is needed because CernVM-FS is a read-only file system.
Local changes, such as of files in /etc or /var, cannot be written to CernVM-FS.
Instead, the union file system transparently redirects such changes to a locally attached scratch hard disk.
The local hard disk is also used to store the CernVM-FS cache.
Note that the scratch hard disk does \emph{not} need to be distributed.
It can be created instantaneously when instantiating the virtual machine as an empty, sparse file.
The init ramdisk contains the CernVM-FS client and a steering script.
The purpose of the steering script is to create the virtual machine's root file system stack that is constructed by unifying the CernVM-FS mount point with the writable scratch space.
To do so, the steering script can process contextualization information (sometimes called ``user data'') from various sources, such as OpenStack, OpenNebula, or Amazon~EC2.
Based on the contextualization information, the CernVM-FS repository and the repository version is selected.
The amount of data that needs to be loaded in order to boot the virtual machine is very little.
The image itself sums up to some \SI{12}{\mega\byte}.
In order to boot Scientific Linux 6 from CernVM-FS, the CernVM-FS client downloads additional \SI{100}{\mega\byte}.
The CernVM-FS infrastructure used to distribute experiment software can be reused.
In comparison, the (already small) CernVM~2.6 virtual appliance sums up to \SIrange{300}{400}{\mega\byte} that needs to be fully loaded and decompressed upfront before the boot process can start.
As a result, booting a $\mu$CernVM virtual machine starts practically instantaneously so that it can be, for instance, integrated with a web site that starts a virtual machine on the click of a button.\footnote{An example of such a web site is a volunteer computing project by the CERN theory group: \url{http://crowdcrafting.org/app/cernvm}}
\section{The $\mu$CernVM root file system stack}
\label{sec:stack}
At the beginning of the Linux boot process, in the so called \emph{early user space}, the Linux kernel uses a root file system in memory provided by the init ramdisk.
The purpose of the early user space is to load the necessary storage device drivers to access the actual root file system.
Once the actual root file system is available, the system switches its root file system to the new root mount point after which the previous root file system becomes useless and is removed from memory.
Figure~\ref{fig:rootfs} shows the transformation of the file system tree in the early user space in $\mu$CernVM.
First, the scratch hard disk is mounted on /root.rw.
$\mu$CernVM grabs the first empty hard disk or partition attached to the virtual machine, or remaining free space on the boot hard disk.
It automatically partitions, formats, and labels the scratch space.
Due to the file system label, $\mu$CernVM finds an already prepared scratch space on next boot.
The scratch space is used as a persistent writable overlay for local changes to the root file system and as a cache for the CernVM-FS client that loads the operating system.
Secondly, a CernVM-FS repository containing a template operating system installation is mounted on /root.ro.
The file system unification of the CernVM-FS mount point and the directory containing the persistent overlay is mounted on /root.
Finally, the /root directory is installed as new root file system.
The /root.ro and /root.rw directories are projected into the final root file system via \emph{bind mounts}.
The rest of the init ramdisk is removed from memory.
\begin{figure}
\caption{\label{fig:rootfs}
Transformation of the root file system stack.
Left hand side: the root file system of the early user space.
The /root directory is an AUFS mount point.
Right hand side: the final root file system stack.
}
\begin{center}
\scriptsize\input{img/rootfs}
\end{center}
\end{figure}
Although $\mu$CernVM and the operating system loaded by $\mu$CernVM are independent from each other, there are a few possible points of interaction between the two components.
These interaction points are implemented through special files in the root directory.
The following files are created or recognized by $\mu$CernVM:
\begin{description}
\item[/.cvmfs\_pids] In this file, $\mu$CernVM stores the process IDs (PIDs) of the CernVM-FS process used to mount the operating system repository.
The PIDs help operating system scripts to avoid accidental killing of these vital processes.
\item[/.ucernvm\_pinfiles]
This file is provided by the operating system repository and contains a list of files that should be pinned in the cache.
The idea is to always keep a minimal set of files available in the cache that allows for recovery from a broken network connection.
\item[/.ucernvm\_bootstrap]
This is a shell script provided by the operating system repository.
It is sourced just before the root file system is switched and allows for custom actions.
\end{description}
Special care has to be taken in the shutdown script of the operating system.
Typically, the shutdown script does not expect an AUFS file system stack as root file system.
Therefore, the script needs to be modified to first remount the underlying read-write layer (/mnt/.rw) in read-only mode before remounting the root file system itself in read-only mode.
Only then the machine can be halted without risking any file system corruption.
\section{Versioned operating system on CernVM-FS}
\label{sec:os}
An easy way to provide an operating system template installation on CernVM-FS is to use the operating system's package manager to install the desired packages in the CernVM-FS repository area.\footnote{Like \texttt{yum --installroot /cvmfs/sl6.cern.ch install glibc emacs \dots}}
New and updated packages can then be installed incrementally by the package manager on top of the existing installation.
While being very fast, this results, however, in an ever-changing operating system directory tree.
It can quickly become difficult to trace back which update introduced or fixed a certain problem and which state of the operating system directory tree is mounted by virtual machines.
We have tackled this problem for \emph{CernVM 3}, a Scientific Linux 6 based operating system on CernVM-FS.
CernVM 3 benefits from versioning on three levels that ensure traceability and a well-defined, predictable virtual machine (see Figure~\ref{fig:version}).
\begin{enumerate}
\item The \emph{cernvm-system} meta package provides the notion of a well defined operating system.
The meta package has no payload but contains only package dependencies.
In case of the cernvm-system package, the dependencies are fully versioned and no dependencies are missing, i.\,e.\ there is no degree of freedom for the dependency resolution of this package.
\item Three different CernVM-FS repositories are maintained, \emph{development}, \emph{testing}, and \emph{production}, that reflect different levels of maturity of the cernvm-system meta package.
Furthermore, separation of repositories allows for injection of a security hotfix in the production repository while continue to develop in the other repositories.
\item To avoid silent and unwanted updates of the virtual machine operating system, $\mu$CernVM exploits the fact that CernVM-FS is a versioning file system.
CernVM-FS creates a snapshot whenever changes are published.
These snapshots remain available and they can be named.
The concept and the implementation is similar to a \emph{tag} in the git versioning system.
CernVM-FS clients can, through mount options, select a particular snapshot on mount.\footnote{This feature is so far untapped by experiments' usage of CernVM-FS. Instead experiments use different directories for different software releases, reflecting the fact that many versions of the experiment software need to be available on the same worker node at the same time.}
On first boot, $\mu$CernVM mounts the newest available snapshot of the given repository and it will stay on this snapshot during reboots unless instructed otherwise.
Snapshots are named after the version number of the cernvm-system meta package.
Contextualization can be used to select another snapshot to mount, letting $\mu$CernVM go back in time and instantiate a historic data processing environment.
\end{enumerate}
\begin{figure}
\caption{\label{fig:version}
CernVM 3 versioning scheme.
Every released cernvm-system meta package is resembled as a versioned snapshot in CernVM-FS.
The named branches \emph{development}, \emph{testing}, and \emph{production} provide entry points to versioned snapshots at different levels of maturity.
}
\begin{center}
\scriptsize\input{img/versioning2}
\end{center}
\end{figure}
The three levels of versioning allow the $\mu$CernVM image to always instantiate the newest available stable operating system version.
At the same time, the very same image provides instant access to every other operating system version ever released in the given repository.
\section{System updates}
\label{sec:updates}
While some virtual machines are short-lived and only instantiated for a particular computing job, others can run for many weeks or months.
A typical example of a long-lived virtual machine is an interactive virtual machine on an end user's laptop.
It is desirable to provide a long-lived virtual machine with bug fixes and security updates without the need to reconfigure the virtual machine or to replace the image file.
This section describes how to update both kernel and init ramdisk on $\mu$CernVM and the operating system on CernVM-FS.
\subsection{Updates of $\mu$CernVM}
The $\mu$CernVM image can be used as a read-only CD-ROM image.
Hence writing updated data to the image is not possible.
However, updated versions of the Linux kernel and the init ramdisk can be dropped into a predefined location on the scratch hard disk.
Upon boot, after mounting the scratch hard disk, $\mu$CernVM uses the \emph{kexec} facility~\cite{kexec03} in order to reboot on the fly into the updated kernel and init ramdisk.
\subsection{Updates of an operating system on CernVM-FS}
Once booted, a $\mu$CernVM based virtual machine can be kept up to date by standard means of the package manager provided by the operating system.
This approach, however, accumulates more and more local changes on the writable overlay stored on the scratch hard disk.
Over time, more and more disk space is spent on the writable overlay and the local operating system diverges from the one provided on CernVM-FS.
Instead it is preferable to pick up updates from CernVM-FS.
To this end, $\mu$CernVM can be instructed to ``unpin'' the currently mounted CernVM-FS snapshot and to mount the newest available snapshot on next boot.
Changing the snapshot can result in conflicts with the local changes on the writable overlay.
For instance, a user might have replaced /usr/bin/gcc while at the same time the updated snapshot provides a newer version of /usr/bin/gcc as well.
The conflict resolution in $\mu$CernVM is guided by the widely used file system hierarchy standard~\cite{fhs04}: $\mu$CernVM resolves conflicts in /etc and /var by keeping the locally modified versions, whereas in all other directories the version from CernVM-FS has precedence.
Furthermore, packages installed by the user need to be reinserted into the package database of the updated snapshot.
Finally, the system account databases (/etc/passwd, /etc/group, \dots) are merged record by record as follows
\begin{itemize}
\item Local passwords have precedence
\item Group membership is merged
\item Conflicting user ids or group ids from the new snapshot are mapped to the locally defined ids.
If necessary, user ids or group ids from the new snapshot are remapped to previously unused ids.
Remapping of user ids and group ids is supported by CernVM-FS, in the same way NFS clients support id mappings.~\cite{rfc1813}
\end{itemize}
This approach to updates assumes that operating system components comply with common standards (for instance, that components do not store configuration data in /usr/bin).
While first observations with CernVM 3 are positive, it requires more experience to verify the approach in practice.
\section{Related Work}
\label{sec:related}
With live CDs and network booted Linux systems, $\mu$CernVM shares the fact that the root file system is read-only and local changes are written to an overlay area by means of a union file system.
Live CDs need to be loaded as a whole before the boot process can start and the writable overlay is in memory.
Typical network boot setups rely on NFS, which cannot be used for wide-area setups.
Also, booting from the network relies on a network boot protocol (e.\,g.\ PXE) instead of using a minimal virtual machine image.
An idea very similar to $\mu$CernVM was implemented as \emph{HTTP-FUSE Xenoppix}~\cite{httpfuse06}.
This system consists of a kernel and an init ramdisk that downloads the actual root file system on demand via HTTP.
Important features of CernVM-FS are not implemented though, such as versioning and digitally signed network traffic.
Unlike $\mu$CernVM, HTTP-FUSE Xenoppix runs only on the Xen hypervisor.
There is no discussion of system updates in HTTP-FUSE Xenoppix.
The \emph{vagrant} tool set greatly facilitates building images for virtual appliances, especially ``virtual development environments''~\cite{vagrant12}.
Despite the convenient user interface, hard disk images containing an full operating system must still be distributed.
Similarly, \emph{CoreOS}\footnote{\url{http://coreos.com}} simplifies massive server deployment.
It consists of a tiny Linux system just enough to run lightweight virtual machines (``containers'') but it is not a virtual machine by itself.
Instead of loading files on demand, CoreOS manages entire containers.
\section{Conclusion}
We have presented $\mu$CernVM, a \SI{12}{\mega\byte} virtual machine image that loads the actual operating system files on demand through CernVM-FS.
Updates to the virtual machine do not require changing the image but they are simply distributed through changes in the CernVM-FS repository.
Moreover, the very same tiny image can boot any version of the operating system ever released on CernVM-FS, which facilitates the long-term preservation of data processing environments.
We have verified our approach by creating CernVM~3, the SL6 based successor of CernVM~2, on top of $\mu$CernVM.\footnote{The CernVM~3 build system is available under \url{https://github.com/cernvm}.}
CernVM~3 runs on VMware, VirtualBox, Xen, KVM, and Hyper-V hypervisors as well as in OpenStack, OpenNebula, and Amazon EC2 clouds.
CernVM~3 can build itself and it runs experiment software from the four LHC experiments.
A systematic validation of experiment software remains to be done.
\section*{References}
|
\section{Introduction}
Cavity linewidth narrowing has been suggested to have great potential in many different areas such as laser stabilization \cite{Shevy2010a,Lukin1998}, high-resolution spectroscopy \cite{Lukin1998}, enhanced light matter interaction and compressed optical energy \cite{Baba2008}. We show more than four orders of magnitude cavity linewidth narrowing, which, to the best of our knowledge, is more than two orders of magnitude larger than demonstrated with other techniques. Previously 10 to 20 times linewidth narrowing has been shown using either EIT \cite{Wu2008,Zhang2008,Wang2000a}, and recently two orders of magnitude was shown using coherent population oscillation (CPO) in combination with a cavity dispersive effect \cite{Grinberg2012a}. We also demonstrate several cavity modes within the slow light transmission window, something which we are not aware of having been demonstrated using EIT or CPO. The present results are obtained using spectral hole-burning in rare-earth-ion doped crystals \cite{Shakhmuratov2005,Lauro2009,Walther2009a} and we discuss the properties and potential of slow light structures created with this method in these materials.
\par
In this paper, a cavity formed by depositing mirrors directly onto a praseodymium doped $Y_2SiO_5$ crystal and (near) persistent spectral hole burning (PSHB) is employed to create a very strong dispersion. A sharp dispersion slope reduces the photon group velocity, and therefore increases the effective photon lifetime in the cavity compared to a non-dispersive cavity (some times referred to as cold cavity \cite{Siegman1985}).
\par
Generally the mode spacing in a Fabry-P\'{e}rot cavity, $\Delta\nu$, is given by \cite{Siegman1985}
\begin{equation}\label{eq:Mode_Spacing}
\Delta\nu=\frac{c}{2L}\frac{1}{n_g(\nu)}=\frac{c}{2L}\frac{1}{n+\nu\frac{dn}{d\nu}}=\frac{v_g(\nu)}{2L}
\end{equation}
where c is the speed of light in vacuum, $\nu$ is the light frequency, $n$ is the real part of the index of refraction (for the phase velocity), $v_g(\nu)$ is the group velocity and $n_g(\nu)$ is the index of refraction for the group velocity. For the present work it is useful to briefly analyse the mode spacing relation.
\par
The resonance condition for a Fabry-P\'{e}rot cavity of length L may be expressed as $m(\lambda/2)=L$, where $m$ is an integer and the wavelength $\lambda=c/(n \nu)$. Thus
\begin{equation} {\label{eq:cmodes}}
m \frac{c}{2L} = n \nu
\end{equation}
Differentiating Eq. \ref{eq:cmodes} gives
\begin{equation} {\label{eq:dcmodes}}
\frac{c}{2L} \delta m = n \delta \nu + \nu \delta n
\end{equation}
Dividing the left (right) hand side of Eq. \ref{eq:dcmodes} with the left (right) hand side of Eq. \ref{eq:cmodes} yields
\begin{equation} {\label{eq:dcmodes2}}
\frac{\delta m}{m} = \frac{\delta \nu}{\nu} + \frac{\delta n}{n}
\end{equation}
\begin{figure}[ht]
\includegraphics[width=8cm]{Readout.pdf}
\caption{(Color online) (a) Non-dispersive cavity transmission spectrum. (b) Cavity transmission spectrum (red trace), best guess for the absorption profile around the $\approx18\;$MHz spectral hole (black trace) inside the cavity. The corresponding real refractive index, $n$, calculated from the black trace absorption profile is shown as a blue trace. (c) Cavity transmission spectrum for the absorption profile about a MHz spectral hole.}
\label{Readout}
\end{figure}
where $n = n(\nu)$ is a function of frequency. Normally when the frequency is changed $\delta \nu/\nu \gg \delta n/n$, but in the case of significant slow light effects $ n \ll \nu (d n / d\nu)$ and the second term on the right hand side in Eq. \ref{eq:dcmodes2} is much larger than the first. Thus the cavity mode spacing is basically completely determined by the dispersion while the impact of the relative change in the frequency is negligible. Below it is shown how we modify the cavity absorption to enter this regime where the Fabry-P\'{e}rot cavity mode spacing is completely determined by the frequency dispersion.
\par
First, we examine the cavity transmission far away from the absorbing inhomogeneous Pr ion transition (non-dispersive cavity). The crystal cavity is $\approx6\;$mm long, the reflectivity was specified to $R_1\approx R_2\approx 95\%$ and $Y_2SiO_5$ has a real index of refraction of $n\approx1.8$. The cavity mode spacing for this crystal is $\Delta \nu_{non-dispersive}\approx 13\;$GHz and the transmission peak linewidth is $\delta \nu_{non-dispersive}\approx 1\;$GHz.
A frequency scan across two cavity modes is shown in Fig. \ref{Readout}a. The reason for the comparatively large cavity linewidth (low finesse) could be due to improper matching to the spatial cavity mode \cite{Sabooni2013,Sabooni2012c}. We expect that the high frequency wing of the cavity modes is due to higher order transverse modes in the cavity. The small extra peak to the left of the modes is the cavity mode for the orthogonal polarization. The $Y_2SiO_5$ crystal is birefringent and cavity modes for the two polarizations along the two principal axes will be displaced relative to each other \cite{Sabooni2013a}. Both modes are seen because the input polarization has a small angle relative to the $D_2$ (principal) axes (see in Fig.\ref{setup}b).
Second, persistent spectral hole burning is employed to manipulate the ion absorption distribution in the inhomogeneously broadened ($\sim$ 9 GHz) $^3H_4-\!^1D_2$ transition of the $0.05\%$ doped $Pr^{3+}\,:Y_2SiO_5$ crystal. A frequency stabilized dye laser \cite{Drever1983,Leibrandt2011} at $\lambda_{vac}=605.976\;$nm is used to remove the absorbing ions within an $18\;$MHz wide spectral region (black trace, Fig. \ref{Readout}b) by a series of laser pulses. The laser pulses optically pump the ions to the excited state, from where they decay back to one of the hyperfine ground levels (see Fig. \ref{setup}a). A typical value for the ground state hyperfine population decay is $\sim$40 seconds, while it could be 40 minutes in a weak magnetic field ($\approx 0.01\;$T) \cite{Ohlsson2003}, and several weeks for other rare-earth-ion-doped crystals \cite{Konz2003}. A detailed description of the procedure for creating a spectral transmission window can be found in Ref. \cite{Amari2010}. An arbitrary-waveform-generator controlled double-pass acousto-optic modulator (AOM), which can tailor the light amplitude, phase and frequency, see Fig. \ref{setup}b, is used to form the necessary laser pulses. As is shown in Fig. \ref{Readout}b the modified ion absorption profile (black trace) has close to zero absorption within about 18 MHz. The absorption profile is taken in a part of the 6 mm crystal which is not reflection coated, i.e. outside the cavity. In this way the frequency resolved absorption measurement can be recorded without being affected by the cavity mode structure. Admittedly, this can only be seen as an estimate of the absorption structure inside the cavity. Previously we have measured the absorption, $\alpha L$, in a 1 mm crystal of the same dopant concentration to be approximately equal to 2. The left axis in Fig. \ref{Readout}b has been scaled to be consistent with this value. In order to verify that absorption and scattering losses are indeed small in the 18 MHz transmission window, we compared the transmission through the crystal (without the cavity) in the black-trace spectral transmission region in Fig. \ref{Readout}b, with the transmission through the crystal when the laser is detuned by about 4 nm from the 9 GHz inhomogeneous line center. It was not possible to detect a difference between the transmission in the 18 MHz transmission region and when the laser was detuned about 4 nm ($\sim$4 THz) away from the absorption line. From this we estimate that any absorption in the slow light transmission window is in any case below $<$ 0.1 dB/cm. The bulk loss in $Y_2SiO_5$ has previously been estimated to $<$ 0.003 dB/cm \cite{Goto2010}.
\par
Experimental line narrowing data is shown in Figs \ref{Readout}b and \ref{Readout}c. In Fig. \ref{Readout}b there are, five cavity transmission peaks within the $\approx18\;$MHz transparent spectral region. The spectral width of the cavity transmission is reduced from $\delta \nu_{non-dispersive}\approx 1\;$GHz in the non-dispersive cavity case (Fig. \ref{Readout}a) to $\delta \nu_{dispersive}\approx 600\;$kHz in the dispersive cavity case. In Fig. \ref{Readout}c the transmission window is about a MHz, yielding a steeper dispersion curve and the cavity line width is about 30 kHz, i.e. a reduction of about 30000 relative to the non-dispersive cavity line width. At the same time the cavity mode spacing has decreased from 13 GHz to about 220 kHz (factor of 60000).
\begin{figure}[ht]
\includegraphics[width=8cm]{setup.pdf}
\caption{(Color online) Experimental set up. (a) The hyperfine splitting of the $^3H_4-\!^1D_2$ transition of site I $Pr^{3+}\!:\!Y_2SiO_5$ is shown \cite{Equall1995,Rippe2005}. (b) A double pass acousto-optic modulator (AOM) is employed to tailor the optical pulses out of the continuous output from a narrow ($<\,$kHz) linewidth laser. A photo diode, PD, monitors the light transmitted through the cavity. The specified reflectivity of the crystal input and output facets is $R_1=R_2=95\%$. $D_1$, $D_2$ and $b$, show the crystal principal axes orientations.}
\label{setup}
\end{figure}
\par
Eq. \ref{eq:Mode_Spacing} shows that it is possible to control the cavity mode spacing by controlling the group refractive index, $n_g(\nu)$.
The (real part of the) refractive index can be calculated from the ion absorption frequency distribution $\alpha(\nu)$ via the Kramers-Kronig relations. This will give the dispersion and group refractive index. The refractive index as calculated from the absorption (black line) is shown by the blue line in Fig. \ref{Readout}b. As shown in Fig. \ref{Readout} the cavity mode spacing and cavity line width can be controllably varied over several orders of magnitude by engineering the ion absorption frequency distribution.
\par
Although spectral hole-burning based slow light structures have been discussed previously \cite{Shakhmuratov2005,Lauro2009} and slow light structures in general have been analysed extensively, e.g. Ref. \cite{Boyd2011} and references therein, we think it could be relevant to discuss the properties and potential of the absorption structuring techniques shown here. Due to the cavity-linewidth narrowing, the $6\;$mm long cavity can have a longitudinal mode spacing of $\approx 220\;$kHz, which in vacuum would correspond to that of a $\approx 700\;$m long cavity! The transversal part of the mode however still behaves as that of a $6\;$mm long cavity, which means that it is possible to combine a narrow beam with fairly even beam size with a small cavity mode spacing. The longitudinal and transversal cavity mode behaviours are thus decoupled by more than four orders of magnitude.
\begin{figure}[ht]
\includegraphics[width=8cm]{Gaussian.pdf}
\caption{(Color online) Transmitted light to photo diode PD (a) When a 20 ns long pulse (black trace, centred at time t=0) is sent into the cavity in presence of the $\approx$18 MHz spectral hole. (b) When a 1 $\mu$s long pulse (black trace, centred at time t=0) is sent into the cavity in presence of the $\approx$3 MHz spectral hole. Both signals are normalized.}
\label{Gaussian}
\end{figure}
\par
In order to further examine the cavity properties, a short pulse was sent into the cavity. The red solid trace in Fig. \ref{Gaussian}a shows the light transmitted to the photo diode, PD, for a 20 ns long input pulse (black dashed trace, centred at time t=0). With a 18 MHz slow light transmission window, the 6 m long pulse is compressed to about 2.5 mm and bounces back and forth in the cavity with some light leaking out after each round trip. Fig. \ref{Gaussian}b shows the corresponding data with a 3 MHz transmission window where the cavity round trip time is increased to well over a microsecond. The group velocity, $v_g$, is approximately given by \cite{Walther2009a,Shakhmuratov2005}
\begin{equation} {\label{eq:group_velocity}}
v_g=\frac{2\pi\Gamma}{\alpha}
\end{equation}
where $\Gamma$ is the spectral width of the transmission window and $\alpha$ is the absorption coefficient immediately outside the spectral transmission window (i.e. the effective absorption depth of the transmission window). In general, \textit{e.g.} Ref. \cite{Lauro2009}, the time-bandwidth product, $TB$, for any slow light structure of this type of length $l$ approximately is
\begin{equation} {\label{eq:TB}}
TB=\frac{l}{v_g}\Gamma=\frac{\alpha l}{2\pi}
\end{equation}
Thus, the bandwidth is determined by the transmission window $\Gamma$, which readily can be changed by a new optical pumping sequence. The group velocity is then determined by the absorption coefficient of the surrounding structure which, for a given crystal, can most easily be changed by moving the transmission window to different positions within the inhomogeneous line. The further out from line center, higher the group velocity. Finally, the time delay is then set by the crystal length. It is noteworthy that all these parameters can be varied independently, which offers a good opportunity to test, for example, non-linear enhancement effects due to slow light in a well controlled environment. As an example a $TB\,\approx\,10$ for $\Gamma=1$ MHz, $\alpha\approx50$ cm$^{-1}$, and $l=$12 mm is reported in Ref. \cite{Zhang2012}.
\par
Potential applications of controlled spectral engineering of slow light structures in cavities will now be discussed. The fact that the lifetime of a light pulse in the cavity and the cavity Q-value increases by several orders of magnitude is interesting because whispering gallery mode rare earth crystal resonators with Q-values in the $10^6$ range have been demonstrated \cite{McAuslan2011b} and Q-values up to $10^{10}$ are predicted \cite{McAuslan2011b}. These numbers could potentially be enhanced by four orders of magnitude by modifying the absorption profiles by optical pumping as is done here. Also the loss rate due to scattering in the cavity, could be strongly reduced. For example, if the fractional scattering loss, $I_f$, for a pulse spending time, $t$, in the cavity is proportional to $1-e^{-\alpha_0l_0}$, where $\alpha_0$ is the scattering coefficient and $l_0$ is the length traversed by the pulse, then we obtain $I_f =1-e^{-\alpha_0l_0}= 1-e^{-\alpha_0v_gt}$, where $v_g$ is the group velocity. Thus if we would like to delay light by a given time, \textit{t}, a material with a low group velocity can strongly reduce scattering losses. In general these type of slow light effects can be interesting for solid state materials which as, compared to vacuum or gas, normally will experience larger scattering.
\par
It is also possible to dynamically tune the speed of light propagating through the material using an electric field. In non-centrosymmetric sites the rare earth ions have a permanent electric dipole moment which is different in the ground and excited state. This means that the ion transition frequency is shifted by an external electric field \cite{Macfarlane1994}, an effect which \textit{e.g.} has been used when creating high efficiency quantum memories in these materials \cite{Hedges2010} or to move the transmission window in frequency \cite{Beavan2013}. Electric fields can equally well be used to change the width of the spectral transmission window \cite{Persson2001} which changes the index of refraction. The group velocity can in this way readily be modulated at tens of MHz rates. As can be seen from Fig. \ref{Gaussian} this rate can be much faster than the time it takes for the pulse to propagate through the cavity. Thus it may be interesting to investigate the possibility to adiabatically tune the wavelength inside the cavity by modifying the refractive index as is done in photonic crystal cavities \textit{e.g.} Ref. \cite{Preble2007}, but now on a time scale which differs by six orders of magnitude.
\par
Low phase noise and very narrow line width laser systems have wide ranging applications as stable frequency references and sources. A passive laser line narrowing technique based on spectral hole burning (SHB) was demonstrated recently. Selective filtering of weak spectral components outside the laser centre frequency by transmitting the laser beam through a highly absorbing SHB material resulted in phase noise suppression by several tens of dB outside the central laser frequency \cite{Thiel2011}. Using the cavity line narrowing techniques presented in this paper, transmission windows narrower than the sub kHz homogeneous lines of the material can be obtained. Further, the region outside the transmission window would be non-transparent but light outside the center frequency is now reflected instead of being absorbed, which has several advantages.
\par
Spontaneous parametric down-conversion is widely used in quantum optics and quantum information science as a source for entangled photon pairs and as single photon sources. It is often desirable to generate entangled photons with a very narrow bandwidth, for example to match the bandwidth of an optical quantum memory or the bandwidth of an optical transition \cite{Clausen2011,Saglamyurek2011}. Photons with narrow bandwidth can be generated by cavity-enhanced spontaneous down conversion, where a periodically-poled non-linear crystal is enclosed inside a cavity \cite{Bao2008}. A rare-earth-ion-doped slow light cavity could be combined with a periodically-poled non-linear crystal. A single narrow cavity transmission peak could be created within the spectral transmission window of the rare-earth-ion-doped cavity crystal. Since the rare-earth-ion inhomogeneous absorption peak can be hundreds of gigahertz wide, this constitutes the only transparent region.
\par
The present cavity has the unusual property of supporting several modes with equal wavelength but with different frequencies. From Eq. \ref{eq:cmodes} it can be seen that for any mode number, \textit{m}, mathematically there is an infinite number of combinations of $\nu$ and $n$ which will fulfil the equation. The present cavity indeed has more than one mode supported for certain mode numbers. For instance, in Fig. \ref{Readout}b, the two first(last) modes, which are about 4 MHz apart, have the same mode number as the two cavity modes just below(above) the rare earth absorption line, which are separated by about 13 GHz.
\par
In conclusion we have shown more than 4 orders of magnitude cavity linewidth narrowing caused by off-resonant interaction with praseodymium ions doped in an inorganic crystal. Several cavity modes are shown within the 18 MHz slow light transmission window. The crystal can readily be reprogrammed by optical pumping or dynamically tuned by external electric fields to yield other cavity line widths or light group velocities. It is suggested that the combination of slow light structures and whispering galley modes in rare earth crystal resonators might give exceptionally high cavity Q values.
\par
This work was supported by the Swedish Research Council, the Knut \& Alice Wallenberg Foundation, the Crafoord Foundation, the EC FP7 Contract No. 247743 (QuRep). The research leading to these results also received funding from the People Programme (Marie Curie Actions) of the European Union's Seventh Framework Programme FP7 (2007-2013) under REA grant agreement no. 287252(Marie Curie Action). Finally, we are grateful to Dr. Mikael Afzelius for several valuable discussions.
|
\section{Introduction}Lie Rinehart pairs generalize the algebraic structure of
vector fields and smooth functions to commutative algebras and Lie algebras,
which are some kind of modules with
respect to each other. In particular any Lie algebra together with its
underlying field defines a Lie Rinehart pair.
Given such a pair, we look at its exterior algebra, that is the
exterior power of the Lie partner seen as a module with respect to the commutative
algebra. In case of vector fields and smooth functions, this is precisely
the algebra of \textit{multivector fields}.
The exterior power of an ordinary Lie algebra has many structures.
Scientist with a more algebraic background eventually look on it as a
particular codifferential graded \textit{coalgebra}, where the
codifferential encodes the Lie algebra structure \cite{MM},
while scientist coming more from differential
geometry likely see it as another graded Lie algebra with respect
to the \textit{Schouten-Nijenhuis} bracket \cite{PM}.
We will show that the codifferential approach is not necessarily well
defined with respect to the additional module structure, but that the
Schouten-Nijenhuis bracket is natural.
Then we provide
a non negatively graded Lie $\infty$-algebra
on the exterior power, which comes with a natural injection of
the original Lie Rinehart pair. In contrast to ordinary Lie theory this injection
is not a single map, but a whole sequence of maps.
Those functions are usually called
\textit{weak} Lie $\infty$-morphisms and we explain them in more detail in appendix A.
On the level of objects, this will merely be a shift in perspective,
but we get a considerably richer theory on the level of morphisms,
since we gain access to functions, which are much more flexible then
ordinary map of (graded) Lie algebras.
\section{The Schouten-Nijenhuis Algebra of a Lie Rinehart pair} We start our work
with a short introduction to Lie Rinehart pairs.
We look at their exterior powers and show that in
contrast to ordinary Lie algebras, in general there is no well
defined (co)differential in this setting anymore. Then we introduce the
Schouten-Nijenhuis bracket, well known from differential geometry.
\subsection{Lie Rinehart pairs}
In what follows $\mathfrak{g}$ will always be a real Lie algebra, that is a $\mathbb{R}$-vector space together with an antisymmetric, bilinear map,
\begin{equation}
[\cdot,\cdot]: \mathfrak{g}\times\mathfrak{g}\to\mathfrak{g}
\end{equation}
called \textbf{Lie bracket}, such that for any three vector $x_1$, $x_2$ and $x_3\in\mathfrak{g}$ the \textbf{Jacobi identity}
$[x_1,[x_2,x_3]]+[x_2,[x_3,x_1]]+[x_3,[x_1,x_2]]=0$ is satisfied.
In addition $A$ will always be a real
associative and commutative algebra with unit,
that is a $\mathbb{R}$-vector space together with an associative and commutative, bilinear map
\begin{equation}
\cdot : A \times A \to A
\end{equation}
called multiplication and a unit $1_A\in A$.
According to a better readable text,
we frequently suppress the symbol of the multiplication in $A$ and just
write $ab$ instead of $a\cdot b$.
Moreover $Der(A)$ will be the Lie algebra of derivations of $A$,
that is the vector space of linear endomorphisms of $A$,
with $D(ab)=D(a)b+aD(b)$ and Lie bracket
$[D,D'](a):=D(D'(a))-D'(D(a))$ for any $a,b\in A$ and $D,D'\in Der(A)$.
Before we get to Lie Rinehart pairs, it is handy to define Lie algebra modules first:
\begin{definition}[Lie algebra module]
Let $\mathfrak{g}$ be a real Lie algebra, $A$ an $\mathbb{R}$-algebra and
$D:\mathfrak{g}\to Der(A)$ a Lie algebra morphism. Then $A$ is called
a \textbf{Lie algebra module} (or just $\mathfrak{g}$-module) and $D$
is called the $\mathfrak{g}$\textbf{-scalar multiplication}.
\end{definition}
Now a \textit{Lie Rinehart pair} is nothing but a Lie algebra
and an associative algebra,
each of them being a module with respect to the other, such that
a particular compatibility equation of their multiplications is satisfied:
\begin{definition}[Lie Rinehart Pair] Let $A$ be an associative and
commutative algebra with unit, $\mathfrak{g}$ a Lie algebra and
$\cdot_A: A\times \mathfrak{g} \to \mathfrak{g}$ as well as
$D: \mathfrak{g}\to Der(A)\;;\; x\mapsto D_x$ maps, such that
$A$ is a $\mathfrak{g}$-module with $\mathfrak{g}$-scalar multiplication
$D$, the vector space $\mathfrak{g}$ is
an $A$-module with $A$-scalar multiplication $\cdot_A$ and
the \textbf{Leibniz rule}
\begin{equation}
[x,a\cdot_A y] = D_x(a)\cdot_A y + a \cdot_A[x,y]
\end{equation}
is satisfied for any $x,y\in\mathfrak{g}$ and $a\in A$. Then
$\left(A,\mathfrak{g}\right)$ is called a \textbf{Lie Rinehart pair}.
\end{definition}
This can be defined more general over arbitrary ground rings with unit and also
with respect to non commutative algebras $A$. We stick to the commutative
situation, since we need that property to define exterior powers later on.
A more general introduction can be found in \cite{JH} and in the references therein.
The two most extreme examples coming from commutative algebras on one side and
Lie algebras on the other:
\begin{example}
For any commutative and associative algebra with unit $A$, a
Lie pair is given by $(A, Der(A))$, together with the standard
action of $Der(A)$ on $A$ and the standard $A$-module structure of $Der(A)$.
\end{example}
\begin{example}
Any real Lie algebra $\mathfrak{g}$ is a $\mathbb{R}$-module
with respect to its ordinary scalar multiplication
and togeter with the \textit{trivial action} of $\mathfrak{g}$ on $\mathbb{R}$, given by
$$
D:\mathfrak{g}\times\mathbb{R} \to\mathbb{R}\;;\; (x,\lambda)\mapsto D_x(\lambda):=0,
$$
the pair $\left(\mathbb{R},\mathfrak{g}\right)$ becomes a Lie Rinehart pair.
\end{example}
As mentioned before the archetypical example is
provided by smooth functions and vector fields on a differentiable manifold:
\begin{example}
Let $M$ be a differentiable manifold, $C^\infty(M)$ the algebra of smooth,
real valued functions and $\mathfrak{X}(M)$ the Lie algebra of vector
fields on $M$. $\mathfrak{X}(M)$ is a $C^\infty(M)$-module and
vector fields acts as derivations on smooth functions, that is the
map
$$D:\mathfrak{X}(M)\times C^\infty(M) \to C^\infty(M)\;;\;
(X,f) \mapsto D_X(f):= X(f)
$$
satisfies the equation
$D_X(fg)=D_X(f)g + fD_X(g)$. Moreover the Leibniz rule
$[X,fY]=D_X(f)Y+f[X,Y]$ holds and it follows that
$(C^\infty(M),\mathfrak{X}(M))$ is a Lie Rinehart pair.
\end{example}
The Lie structure can be extended into a graded Lie algebra on the
direct sum of the partners, concentrated in degrees zero and one.
This appears in \cite{GR}:
\begin{definition}[Associated Lie algebra]
Let $(A,\mathfrak{g})$ be a Lie Rinehart pair. Its
\textbf{associated} (graded) Lie algebra is the direct
sum $A\oplus\mathfrak{g}$ seen as a graded vector space,
with $A$ concentrated in degree zero, $\mathfrak{g}$ concentrated
in degree one and Lie bracket defined by
\begin{equation}
\begin{array}{crcl}
[\cdot,\cdot]: & A\oplus\mathfrak{g} \times A\oplus\mathfrak{g} &\to&
A\oplus\mathfrak{g}\\
& \left((a,x),(b,y)\right) &\mapsto& \left(D_x(a)+D_y(b),[x,y]\right)\;.
\end{array}
\end{equation}
\end{definition}
In particular this means, that we can see any Lie Rinehart pair as a
graded Lie algebra and that it makes sense to talk about (graded)
Lie algebra morphisms in their context. The following proposition
justifies the definition:
\begin{prop}$(A\oplus\mathfrak{g},[\cdot,\cdot])$ is a graded Lie
algebra.
\end{prop}
\begin{proof}
$A\oplus\mathfrak{g}$ is a graded vector space by definition.
To see graded symmetry of the bracket, we only need to
consider mixed expressions, where we compute
$[(a,0),(0,x)]=(D_x(a),0)=[(0,x),(a,0)]=
(-1)^{|a||x|}[(x,0),(a,0)]$ on scalars $a\in A$ and vectors
$x\in\mathfrak{g}$.
To see the graded symmetric Jacobi identity, observe that it has to vanish,
whenever at least two arguments are scalars, since the left side of the identity
is an expression, homogeneous of degree $-2$. If all arguments are vectors, it becomes the usual Jacobi identity of $\mathfrak{g}$ and the
remaining cases are seen from
$D_y\circ D_x-D_x\circ D_y +D_{[x,y]}=0$.
\begin{comment}
\begin{align*}
&[[(a,0),(0,x)],(0,y)]+(-1)^{|x||y|}[[(a,0),(0,y)],(0,x)]\\
&+(-1)^{|a||x|+|a||y|}[[(0,x),(0,y)],(a,0)]\\
&=[(D_x(a),0),(0,y)]-[(D_y(a),0),(0,x)]+[(0,[x,y]),(a,0)]\\
&=(D_y(D_x(a))-D_x(D_y(a))+D_{[x,y]}(a),0)
=(D_{[y,x]}(a)+D_{[x,y]}(a),0)\\
&=0
\end{align*}
\end{comment}
\end{proof}
Morphisms of Lie Rinehart pairs are pairs
of appropriate algebra maps, which interact properly with
respect to the additional module structures \cite{GR}:
\begin{definition}[Lie Rinehart Morphism]
Let $(A,\mathfrak{g})$ and $(B,\mathfrak{h})$ be two Lie Rinehart pairs.
A \textbf{morphism of Lie Rinehart pairs} is a pair of maps $(f,g)$, such
that $f:A\to B$ is a morphism of associative and commutative, real algebras
with unit, $g:\mathfrak{g}\to\mathfrak{h}$ is a morphism of Lie
algebras and the equations
\begin{equation}\label{LR-morph}
\begin{array}{ccc}
g(a\cdot_A x)=f(a)\cdot_B g(x)
& and &
f(D_x(a))=D_{g(x)}(f(a))
\end{array}
\end{equation}
are satisfied for any $a\in A$ and $x\in \mathfrak{g}$.
\end{definition}
This is the correct definition of a
morphism in the setting of Lie Rinehart pairs, since all structure is respected
properly:
\begin{corollary}
Let $(f,g):(A,\mathfrak{g})\to(B,\mathfrak{h})$ be a morphism of
Lie Rinehart pairs. The image $(f(A),g(\mathfrak{g}))$ is a
Lie Rinehart pair and
$(f,g):A\oplus\mathfrak{g}\to B\oplus\mathfrak{h}$ is a morphism of
graed Lie algebras.
\end{corollary}
\begin{proof} The first structure equation of (\ref{LR-morph}) implies that
the vector space $g(\mathfrak{g})$ is a $f(A)$-module and the second
that $f(A)$ is a $g(\mathfrak{g})$-module. To see the Leibniz equation, compute
\begin{align*}
&[g(x),f(a)\cdot_B g(y)] = [g(x),g(a\cdot_A y)] = g([x,a\cdot_A y])
=g(D_x(a)\cdot_A y + a \cdot_A[x,y])=\\
&f(D_x(a))\cdot_B g(y) + f(a) \cdot_Bg([x,y])
=D_{g(x)}(f(a))\cdot_B g(y) + f(a) \cdot_B[g(x),g(y)]\;.
\end{align*}
The second part is a consequence of (\ref{LR-morph}).
\begin{comment}
\begin{align*}
&(f,g)([(a,x),(b,y)])=(f,g)(D_x(b)+D_y(a),[x,y])
=(f(D_x(b)+D_y(a)),g([x,y]))\\
&=(D_{g(y)}(f(a))+D_{g(x)}(f(b)),[g(x)),g(y)]
=[(f(a),g(x)),(f(b),g(y))]\\
&=[(f,g)(a,x),(f,g)(b,y)]\;.
\end{align*}
\end{comment}
\end{proof}
\subsection{The Exterior Algebra}
For any $n\in\mathbb{N}$, let $\otimes^n_A\mathfrak{g}$ be the $n$-fold
tensor product of the $A$-module $\mathfrak{g}$
with $\otimes^0_A\mathfrak{g}:=A$. Since $A$ is commutative,
$\otimes^n_A\mathfrak{g}$ is an $A$-module and we write
$a\cdot_Ax$ for the $A$-scalar multiplication of any $a\in A$ and
$x\in\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A$. Note that in general any tensor can be
expressed in terms of vectors:
\begin{prop}\label{simple_tensor}
Let $A$ be an associative and commutative algebra with unit,
$M$ an $A$-module and $\otimes^n_AM$ the appropriate $n$-fold tensor product.
Then any $x\in \otimes^n_AM$ is an $A$-linear combination of
\textbf{simple tensors}, i.e. there is a finite index set $I$, tensors
$x_{i,1}\otimes_A\cdots\otimes_A x_{i,n}\in \otimes^n_AM$ and
scalars $a_i\in A$, such that
\begin{equation}
x=\textstyle\sum_{i\in I}a_i \cdot_A x_{i,1}\otimes_A\cdots\otimes_A x_{i,n}\;.
\end{equation}
\end{prop}
\begin{proof}See for example (\cite{MH}).
\end{proof}
We call such a sum an $A$-\textbf{linear combination}. However in general these
$A$-linear combinations are not unique.
Back on Lie Rinehart pairs $(A,\mathfrak{g})$,
the \textbf{tensor algebra} of the $A$-module $\mathfrak{g}$ is the
direct sum of all $n$-fold $A$-tensor products
$$
T_A\mathfrak{g}:=\textstyle\bigoplus_{n=0}^\infty\bigotimes^n_A\mathfrak{g}\;,
$$
together with an associative but not commutative multiplication given by
concatenation of tensors
$\otimes_A: T_A\mathfrak{g} \times T_A\mathfrak{g} \to
T_A\mathfrak{g}\;;\; (x,y)\mapsto x\otimes_A y$.
This product has a unit $1_A\in \otimes^0_A\mathfrak{g}\simeq A$.
As usual we get the exterior power as the quotient of the tensor power
and the submodule generated by all simples tensors with 'repeated vector products':
\begin{definition}[Exterior Algebra] For any Lie Rinehart pair $(A,\mathfrak{g})$
and $n\in\mathbb{N}_0$, let $\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}^n_A\mathfrak{g}:=\otimes^n_A\mathfrak{g}/J^n$ be
the quotient module of the $n$-th tensor product and the submodule $J^n$,
spanned by all $x_1\otimes\cdots\otimes x_n$ with
$x_i = x_j$ for some $i = j$. Then the direct sum
\begin{equation}
\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A:=\textstyle\bigoplus_{n=0}^\infty\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}^n_A\mathfrak{g}
\end{equation}
together with the quotient
${\scriptstyle \;\wedge\;}: \raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A \times \raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A \mapsto
\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A\;;\; (x,y)\mapsto x{\scriptstyle \;\wedge\;} y$ of the
$A$-tensor multiplication, is called the \textbf{exterior algebra} of
$(A,\mathfrak{g})$ and the product is called the \textbf{exterior product}.
\end{definition}
We write $x_1{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_n\in\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}^n_A\mathfrak{g}$ for the
coset of any tensor $x_1\otimes\cdots\otimes x_n\in \otimes^n_A\mathfrak{g}$ and
in particular $\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}^0_A\mathfrak{g}\simeq A$ and
$\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}^1_A\mathfrak{g}\simeq \mathfrak{g}$, since $J^0=\{0\}$ and $J^1=\{0\}$.
If at least one factor in an exterior product is of
tensor degree zero, we sometimes write $a\cdot_A x$ instead of $a{\scriptstyle \;\wedge\;} x$,
to stress that the exterior product is just $A$-scalar multiplication in
that case.
\begin{example}
If $(C^\infty(M),\mathfrak{X}(M))$ is the Lie Rinehart pair of smooth
functions and vector field, the exterior algebra
$\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{X}(M)_{C^\infty(M)}$ is the algebra of
\textbf{multivector fields}.
\end{example}
The exterior algebra is an $A$-module and
any exterior tensor can be written as a sum (not just a linear combination) of
simple exterior tensors:
\begin{prop}Let $(A,\mathfrak{g})$ be a Lie Rinehart pair and
$\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A$ its exterior algebra. Then
$\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A$ is an $A$-module and any tensor
$x\in \raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A$ is an $A$-linear combination of simple exterior products.
In particular there is a finite index set $I$, scalars $a_i\in A$ and
simple exterior tensors
$x_{i,1}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{i,n_i}\in\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A$, such that
\begin{equation}
x= \textstyle\sum_{i\in I}a_i\cdot x_{i,1}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{i,n_i}\;.
\end{equation}
Moreover any tensor $x\in\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A$ is a finite sum
(not just a linear combination) of simple tensors, that is there is a finite index set $I$ and simple exterior tensors
$x_{i,1}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{i,n_i}\in\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A$, such that
\begin{equation}
x= \textstyle\sum_{i\in I}x_{i,1}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{i,n_i}\;.
\end{equation}
\end{prop}
\begin{proof}Since $T_A\mathfrak{g}$ is an $A$-module, so is $\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A$.
For the first equation, observe that
by prop (\ref{simple_tensor})
the $A$-tensor algebra $T_A\mathfrak{g}$ is spanned by simple
tensors $x_1\otimes_A\cdots\otimes_A x_n$. It follows that the quotient
$\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A$ is spanned by the appropriate cosets
$x_1{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_n$.
To see the second equation apply
$a\cdot x_1{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_n =
(a\cdot_A x_1){\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_n$ to the first one.
\end{proof}
Morphisms of Lie Rinehart pairs prolong naturally to morphisms of
exterior algebras defined as the direct sum of the scalar part and
the exterior tensor power of the Lie algebra part. The compatibility
conditions (\ref{LR-morph}) then guarantee that this map is well defined
as a morphism of exterior algebras over different scalars.
\begin{definition}(Associated Morphism)
Let $(f,g):(A,\mathfrak{g})\to (B,\mathfrak{h})$ be a morphism of Lie Rinehart pairs.
The map
\begin{equation}
\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f:\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A\to\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{h}_B
\end{equation} defined
on scalars $a\in\bbwedge{0}\mathfrak{g}_A$ by
$f(a)$ and on simple tensors
$x_1{\scriptstyle \;\wedge\;}_A\cdots{\scriptstyle \;\wedge\;}_A x_n\in\bbwedge{n}\mathfrak{g}_A$ by
$g(x_1){\scriptstyle \;\wedge\;}_B\cdots{\scriptstyle \;\wedge\;}_B g(x_n)$ and then extended to all of
$\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A$ by $A$-additivity, is called
the \textbf{associated morphism} of $(f,g)$.
\end{definition}
The following proposition shows that associated morphisms are well defined
as morphisms of exterior algebras over different scalars and that the construction
is natural:
\begin{prop}
Let $(f,g):(A,\mathfrak{g})\to (B,\mathfrak{h})$ be a morphism of Lie Rinehart pairs.
Its associated morphism $\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f$ is a well defined morphism of
exterior algebras over modules of different rings and in particular the equations
$$
\begin{array}{ccc}
\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f((a\cdot_A x){\scriptstyle \;\wedge}_A y)= \raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f(x{\scriptstyle \;\wedge}_A (a\cdot_A y)) & and &
\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f(x {\scriptstyle \;\wedge}_A y)= \raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f(x){\scriptstyle \;\wedge}_B \raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f(y)
\end{array}
$$ are satisfied for all $a\in A$ and $x,y\in\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A$.
If $(h,i):(B,\mathfrak{h})\to (C,\mathfrak{i})$ is another morphism of
Lie Rinehart pairs then
$$
\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} i_h \circ \raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f = \raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} (i\circ g)_{h\circ f}\;.
$$
\end{prop}
\begin{proof}Follows from (\ref{LR-morph}), since the exterior product
$\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g$ is natural.
\begin{comment}
$\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f$ has to commute with all structure on the exterior
algebras, but it only remains to show that it behaves
well with respect to the module scalar multiplications:
On scalars and vectors this follows directly from the properties of the Lie Rinehart
morphism and on simple tensors of higher order we have
\begin{align*}
&\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f((a\cdot x_1){\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_n)=
g(a\cdot_A x_1){\scriptstyle \;\wedge\;} \cdots {\scriptstyle \;\wedge\;} g(x_n)=
f(a)\cdot_B g(x_1){\scriptstyle \;\wedge\;} \cdots {\scriptstyle \;\wedge\;} g(x_n)=\\
&g(x_1){\scriptstyle \;\wedge\;} \cdots {\scriptstyle \;\wedge\;} g(a\cdot_A x_n)=
\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f(x_1{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} (a\cdot_A x_n))
\end{align*}
and similar for all other permutations. The second part follows from the
structure equations (\ref{LR-morph}), since
$\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} i_h \circ \raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f(a)=h\circ f (a) = \raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}(i\circ g)_{h\circ f}(a)$
and
\begin{multline*}
\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} i_h \circ \raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f(x_1{\scriptstyle \;\wedge}_A \cdots {\scriptstyle \;\wedge}_A x_n)
=\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} i_h (g(x_1){\scriptstyle \;\wedge}_B\cdots{\scriptstyle \;\wedge}_B g(x_n))=\\
(i\circ g)(x_1){\scriptstyle \;\wedge}_B\cdots{\scriptstyle \;\wedge}_B (i\circ g)(x_n)
=\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}(i\circ g)_{h\circ f}(x_1{\scriptstyle \;\wedge}_A \cdots {\scriptstyle \;\wedge}_A x_n)\;.
\end{multline*}
\end{comment}
\end{proof}
Exterior algebras are equipped with a
$\mathbb{Z}$-grading coming from the tensor degree, which we need
in the definition of a \textit{graded symmetric} Lie $\infty$-algebra later on.
In addition we need a 'reduced' grading to understand the symmetry
of the traditional Schouten-Nijenhuis bracket:
\begin{definition}(Gradings)
Let $(A,\mathfrak{g})$ be a Lie Rinehart pair and $\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A$ its
exterior algebra. An element $x\in\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}^n_A\mathfrak{g}$ is called
\textbf{homogeneous} and the integer $n$ is called the
\textbf{tensor} degree of $x$, written as
\begin{equation}
|x|:=n\;.
\end{equation}
In addition the \textbf{antisymmetric} degree of
a homogeneous element $x\in\bbwedge{n}\mathfrak{g}_A$ is the tensor degree, but reduced by one and written as
\begin{equation}\label{antisymm_grading}
deg(x):=n-1\;.
\end{equation}
\end{definition}
Calling the reduced grading 'antisymmetric', will become clear in the next
section, since the traditional Schouten-Nijenhuis bracket behaves
antisymmetric, with respect to this grading. The reader familiar with
$\mathbb{Z}$-graded abelian groups will further notice, that the term is the
correct one, related to the transition between \textit{symmetric} and
\textit{antisymmetric} \cite{FM}.
With respect to the tensor grading the exterior algebra is concentrated
in non negative degrees and tensors of degree zero are
precisely the scalars $a\in A$.
Note that the exterior product is mostly seen as some kind of antisymmetric operation,
but since
\begin{equation}
x{\scriptstyle \;\wedge\;} y = (-1)^{|x||y|}y{\scriptstyle \;\wedge\;} x
\end{equation}
holds on homogeneous tensors,
it is in fact a \textbf{graded symmetric} product with respect to the tensor grading,
while it has no symmetry at all with respect to the antisymmetric grading.
Moreover it is a graded bilinear map, homogeneous of degree zero with
respect to the tensor grading and homogeneous of degree one with respect to
the antisymmetric grading.
\subsection{The nonexistence of a natural codifferential}
Every graded Lie algebra comes with a (co)differential
on its reduced symmetric tensor coalgebra \cite{MM} and
considering an ordinary Lie algebra as $\mathbb{Z}$-graded
but concentrated in degree one only, we get a (co)differential on its
exterior power.
Taking this into account, one would guess that such a
(co)differential is defined for any Lie Rinehart pair, but as we will
see the technique does not apply here anymore. In fact
on the archetypical example of smooth functions and
vector fields, we can not naturally define a (co)differential
on the exterior algebra other than the zero operator.
This is not a contradiction to the previous mentioned coalgebraic
approach, since we have to deal with the additional $A$-module structure,
which happens to be trivial on the Lie Rinehart pair $(\mathbb{R},\mathfrak{g})$ of a Lie algebra.
The following theorem should be seen as a counterexample, giving a Lie Rinehart
pair, without a non trivial (co)differential on its exterior algebra:
\newpage
\begin{theorem}
Let $M$ be a Hausdorff, separable, finite dimensional and
differentiable manifold, $(C^\infty(M),\mathfrak{X}(M))$ the Lie Rinehart pair of smooth
functions and vector fields on $M$ and $\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{X}(M)_{C^\infty(M)}$
its exterior algebra of \textit{multivector fields}. Then the only
naturally defined (co)differential
$$
d:\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}{}\mathfrak{X}(M)_{C^\infty(M)}\to \raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}{}\mathfrak{X}(M)_{C^\infty(M)},
$$
homogeneous of degree $-1$ with respect to the tensor grading, is the zero operator.
\end{theorem}
\begin{proof} We look at a slightly more general situation and proof that
for any $n\in\mathbb{N}$ there is no naturally defined map
$d:\bbwedge{n}\mathfrak{X}(M)_{C^\infty(M)}\to \bbwedge{n-1}\mathfrak{X}(M)_{C^\infty(M)}$
other than the zero map at all.
Since the codifferential is homogeneous of degree $-1$, this
will include the situation of the theorem.
As we require this map to be a \textit{natural operator}, we
can use the technique of \cite{KMS}. According to theorem 14.18 in \cite{KMS},
those operators of order $r$ are in one to one correspondence with
maps $f:T^{r+1}_m(\bbwedge{n}\mathbb{R}^m)\to \bbwedge{n-1}\mathbb{R}^m$, equivariant with
respect to the associated action of the jet group $G_m^{r+1}$ on domain and
codomain for any $m\in\mathbb{N}$.
From \cite{KMS} prop 14.20 we know further
$$
T^r_m(\bbwedge{n}\mathbb{R}^m)\simeq
(\bbwedge{n}\mathbb{R}^m)\oplus(\bbwedge{n}\mathbb{R}^m\otimes\mathbb{R}^{m*})\oplus\cdots\oplus
(\bbwedge{n}\mathbb{R}^m\otimes S^{r}\mathbb{R}^{m*})
$$
and write
$x^{j_1,\ldots,j_n}$, $x^{j_1\ldots,j_n,}{}_{j_{n+1}}$,
... , $x^{j_1,\ldots,j_n,}{}_{j_{n+1},\ldots,j_{n+r}}$ for the
local coordinates of $T^r_m(\bbwedge{n}\mathbb{R}^m)$, that are the prolongations of the
canonical coordinates $x^j$ in $\mathbb{R}^m$. In particular any such expression is
antisymmetric in all upper and symmetric in all lower indices.
The required actions $l^{r+1}$ of $G^{r+1}_m$ on $T^r_m(\bbwedge{n}\mathbb{R}^m)$ are pretty
involved, but fortunately we only need the actions of the general linear
group $Gl(m)$, seen as a subset of the $(r+1)$-th order jet group
$G^{r+1}_m$. According to \cite{KMS} proposition 14.20 this restricted action is
purely tensorial and given by
$\tilde{x}^{j_1,\ldots,j_n}= x^{i_1,\ldots,i_n}
a^{j_1}{}_{i_1}\cdots a^{j_n}{}_{i_n}$,
$\tilde{x}^{j_1,\ldots,j_n,}{}_{j_{n+1}}=
x^{i_1,\ldots,i_n,}{}_{i_{n+1}}a^{j_1}{}_{i_1}\cdots
a^{j_n}{}_{i_n}a^{i_{n+1}}{}_{j_{n+1}}
$ $,\ldots,$\\
$
\tilde{x}^{j_1,\ldots,j_n,}{}_{j_{n+1},\ldots, j_{n+r}}=
x^{i_1,\ldots,i_n,}{}_{i_{n+1},\ldots,i_{n+r}}
a^{j_n}{}_{i_n}\cdots a^{j_n}{}_{i_n}a^{i_{n+1}}{}_{j_{n+1}}\cdots
a^{i_{n+r}}{}_{j_{n+r}}
$
and the restricted action on $\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}^{n-1}\mathbb{R}^m$ is tensorial, too.
Now choose some real $\lambda>0$ and consider the general linear
transformations $a^j{}_i\in Gl(m)$, defined by
$a^i{}_i=\lambda$ and $a^j{}_i=0$ for $i\neq j$. The action on
$\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}^{n-1}\mathbb{R}^m$ and $T^r_m(\bbwedge{n}\mathbb{R}^m)$ is particularly easy to compute and given by
$l^{r+1}(a^j{}_i;x^{j_1,\ldots,j_n})=\lambda^{n-1}x^{j_1,\ldots,j_n}$ and
\begin{multline*}
l^{r+1}(a^j{}_i;x^{j_1,\ldots,j_n},x^{j_1,\ldots,j_n,}{}_{j_{n+1}},\ldots,
x^{j_1,\ldots,j_n,}{}_{j_{n+1},\ldots,j_{n+r}})=\\
\left(\lambda^nx^{j_1,\ldots,j_n},\lambda^{n+1} x^{j_1,\ldots,j_n,}{}_{j_{n+1}},\ldots,
\lambda^{n+r} x^{j_1,\ldots,j_n,}{}_{j_{n+1},\ldots,j_{n+r}}\right)\;.
\end{multline*}
It follows that any map $f:T^r_m(\bbwedge{n}\mathbb{R}^m)\to \bbwedge{n-1}\mathbb{R}^m$,
equivariant with respect to the action $l^{r+1}$
has to satisfy the homogenity condition
\begin{multline*}
\lambda^{n-1}f\left(x^{j_1,\ldots,j_n},x^{j_1,\ldots,j_n,}{}_{j_{n+1}},\ldots,
x^{j_1,\ldots,j_n,}{}_{j_{n+1},\ldots,j_{n+r}}\right)=\\
f\left(\lambda^n x^{j_1,\ldots,j_n},\lambda^{n+1} x^{j_1,\ldots,j_n,}{}_{j_{n+1}},
\ldots,\lambda^{n+r} x^{j_1,\ldots,j_n,}{}_{j_{n+1},\ldots,j_{n+r}}\right)
\end{multline*}
for all real $\lambda>0$, but by the homogeneous function theorem \cite{KMS} (24.1)
such a map (other than the zero morphism) only exists if the equation
$$n d_1 + (n+1)d_2 \cdots + (n+r)d_r = n-1$$ has solutions
$d_1,\ldots,d_r\in \mathbb{N}$, which it hasn't. It follows that the zero map is the
only equivariant function and consequently the only natural operator is the zero operator.
\end{proof}
In \cite{GR} Rinehart defined a (co)differential for any Lie Rinehart pair,
but on the tensor product of the exterior algebra and the universal enveloping
algebra of the Lie algebra instead. In case of smooth functions and
vector fields this gives a structure dual to the usual De Rham complex of
differential forms.
\begin{remark} On the Lie Rinehart pair $(\mathbb{R},\mathfrak{g})$ of
a Lie algebra with its trivial $\mathfrak{g}$-module structure
on $\mathbb{R}$, a (co)differential
$d:\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g} \to \raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}$
is defined by $d(\lambda)=0$ on scalars $\lambda\in \mathbb{R}$ as well as $d(x)=0$
on vectors $x\in\mathfrak{g}$ and by
$$d(x_1{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_n):=
\textstyle\sum_{s\in Sh(2,n-2)}e(s)\,
[x_{s(1)},x_{s(2)}]{\scriptstyle \;\wedge\;} x_{s(3)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(n)}
$$
on simple tensors
$x_1{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_n\in \raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} \mathfrak{g}$
and is then extend to
all of $\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}$ by linearity. Except for the degree
zero part (which is trivial)
this is the 'coalgebraic' (co)differential as it appears for example in \cite{MM}.
\end{remark}
If we try to define a similar map on the exterior power of an
arbitrary Lie Rinehart pair, the operator is not necessarily well defined
with respect to the additional module structure and we could face situations
like
$$
d((a\cdot x_1){\scriptstyle \;\wedge\;} x_2)\neq
d(x_1{\scriptstyle \;\wedge\;}(a\cdot x_2))\;.
$$
\subsection{The Schouten-Nijenhuis bracket}
The Lie bracket on vector fields can be extended to a
graded Lie bracket on \textit{multivector fields}, usually called
Schouten-Nijenhuis bracket \cite{CM},\cite{PM}. We show that this can be generalized
verbatim to arbitrary Lie Rinehart pairs.
\begin{definition}[Schouten-Nijenhuis bracket]
Let $(A,\mathfrak{g})$ be a Lie Rinehart pair and $\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A$ its
exterior algebra. The map
\begin{equation}
\left[\cdot\;,\cdot\right]: \raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A \times \raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A \to
\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A\;,
\end{equation}
defined by $[a,b]=0$ as well as $[x,a]=[a,x]=D_x(a)$ on scalars $a,b\in A$ and
vectors $x\in\mathfrak{g}$ and by
\begin{multline}\label{SN_1}
[x_1{\scriptstyle \;\wedge\;}\cdots {\scriptstyle \;\wedge\;} x_n,y_1{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} y_m]=\\
\textstyle\sum_{i,j}(-1)^{i+j}[x_i,y_j]{\scriptstyle \;\wedge\;} x_1{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;}
\widehat{x_i}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_n{\scriptstyle \;\wedge\;}
y_1{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} \widehat{y_j}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} y_m
\end{multline}
on simple tensors
$x_{1}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{n}$,
$y_{1}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} y_{m}\in\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A$
and then extend to all of $\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A$ by $A$-additivity, is called the
\textbf{(antisymmetric) Schouten Nijenhuis bracket} of $(A,\mathfrak{g})$.
\end{definition}
This is the traditional definition as it appears for example in \cite{CM}.
In case of simple tensors an equivalent but more symmetric expression
for (\ref{SN_1}) is given by
$$\label{SN_2}
\textstyle\sum_{s\in Sh(1,n-1)}\sum_{t\in Sh(1,m-1)}
e(s)e(t)[x_{s_1},y_{t_1}]{\scriptstyle \;\wedge\;} x_{s_2}
{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s_n}{\scriptstyle \;\wedge\;} y_{t_2}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} y_{t_m}\;.
$$
Care has to be taken, to get the symmetry of the Schouten-Nijenhuis bracket right.
In fact we have to consider the antisymmetric grading (\ref{antisymm_grading}) of
$\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A$ to understand it properly. With this grading
the common commutation equation
$$
[x,y]=-(-1)^{(|x|-1)(|y|-1)}[y,x]
$$
suddenly becomes more conceptual and just says that the Schouten-Nijenhuis bracket is
\textbf{graded antisymmetric} with respect to the \textit{antisymmetric grading}.
Later we have to deal with a graded symmetric incarnation of the bracket and
that's why we call this one the \textit{antisymmetric} bracket.
Proofing its properties has been done in the situation of multivector fields
at many places before \cite{CM}, \cite{PM} and we only recapitulate the basic facts
for completeness:
\begin{theorem}Let $(A,\mathfrak{g})$ be a Lie Rinehart pair with
exterior algebra $\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A$. The
Schouten-Nijenhuis bracket $[.,.]$ is a $\mathbb{R}$-bilinear, graded antisymmetric operator,
homogeneous of degree zero with respect to the
antisymmetric grading and in particular the equation
$$
\begin{array}{cc}
[x,y]=-(-1)^{deg(x)deg(y)}[y,x],&
[x,y{\scriptstyle \;\wedge\;} z] = [x,y]{\scriptstyle \;\wedge\;} z + (-1)^{deg(x)(deg(y)-1)}y{\scriptstyle \;\wedge\;} [x,z]
\end{array}
$$
as well as the graded Jacobi equation in its antisymmetric incarnation
$$
(-1)^{deg(x)deg(z)}[x,[y,z]]+(-1)^{deg(x)deg(y)}[y,[z,x]]
+(-1)^{deg(y)deg(z)}[z,[x,y]] = 0
$$
are satisfied for any homogeneous tensors $x,y,z\in\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A$.
\end{theorem}
\begin{proof}On homogeneous tensors $x$ and $y$ we get
$|[x,y]|=|x|+|y|-1$ and the bracket is homogeneous of tensor
degree $-1$. This in turns gives $deg([x,y])=|x|+|y|-2=deg(x)+deg(y)$ and consequently
the bracket is homogeneous of degree zero with respect to the antisymmetric
grading.
All other properties are computed verbatim as for the Schouten-Nijenhuis
bracket of multivector fields.
\begin{comment}
$\mathbb{R}$-bilinearity follows from $\mathbb{R}$-bilinearity of all involved operators in
the definition (\ref{SN_1}) of the bracket. Moreover on homogeneous tensors
$x$ and $y$ we get
$|[x,y]|=|x|+|y|-1$ and the bracket is homogeneous of tensor
degree $-1$. This in turns gives $deg([x,y])=|x|+|y|-2=deg(x)+deg(y)$ and consequently
the bracket is homogeneous of degree zero with respect to the antisymmetric
grading.
On scalars and vectors, graded antisymmetry follows directly from the defining
equations and on simple tensors it is best seen from the 'symmetric expression'
(\ref{SN_2}). Since any homogeneous tensor is a sum of simple tensors, this in
turn shows that the bracket is graded antisymmetric for arbitrary homogeneous
tensors.
To see the graded Jacobi identity let $x,y,z\in\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A$ be
homogeneous tensors. Then
\begin{align*}
(-1)^{deg(x)deg(z)}[x,[y,z]]+(-1)^{deg(x)deg(y)}[y,[z,x]]
+(-1)^{deg(y)deg(z)}[z,[x,y]]
\end{align*}
\end{comment}
\end{proof}
Now lets look at exactly the same situation, but from the tensor grading of
$\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A$. There is a 'natural transformation' between the tensor
and the antisymmetric grading
called \textit{decalag\'e morphism} \cite{FM}, which can be used to
transform graded antisymmetric operators into graded symmetric ones and
vis versa. Applied to the antisymmetric Schouten-Nijenhuis
bracket, this gives a graded \textit{symmetric} operator:
\begin{definition}Let $(A,\mathfrak{g})$ be a Lie Rinehart pair and
$[\cdot,\cdot]$ the antisymmetric Schouten-Nijenuis bracket on $\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A$.
Then the operator
\begin{equation}
\{\cdot,\cdot\}: \raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}\times\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}\to
\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}
\end{equation}
defined by $\{x,y\}=e(x)[y,x]$
on homogeneous tensors $x,y\in\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}$ and extended to all of
$\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}$ by $A$-additivity is called the
\textbf{symmetric Schouten-Nijenhuis bracket}.
\end{definition}
All properties of the antisymmetric bracket transform properly under the
decalgn\'e morphism and both brackets coincides with the original Lie bracket
on vectors:
\begin{corollary}The symmetric Schouten-Nijenhuis bracket $\{.,.\}$
is a $\mathbb{R}$-bilinear, graded symmetric operator,
homogeneous of degree $-1$ with respect to the
tensor grading. In particular the symmetry equation as well as the Jacobi equation
$$\label{symmetric_jacobi}
\begin{array}{cc}
\{x,y\}=e(x,y)\{y,x\}, &
\textstyle\sum_{s\in Sh(2,1)}e\left(s;x_1,x_2,x_3\right)
\{\{x_{s(1)},x_{s(2)}\},x_{s(3)}\}=0
\end{array}$$
is satisfied for all homogeneous tensors $x,y,z\in\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A$
and on vectors $x,y\in\bbwedge{1}\mathfrak{g}$, the bracket equals the original
Lie bracket of $\mathfrak{g}$ that is
$\{x,y\}=[x,y]$.
\end{corollary}
\begin{proof}This follows from the properties of the decalgn\'e morphism or
can else be verified by simple computations.
\begin{comment}
\begin{align*}
\{x,y\}
&=e\left(x\right)\left[y,x\right]
=e\left(x\right)\left(e\left(x,y\right)e\left(x\right)e\left(y\right)\right)
\left[x,y\right]=e\left(x,y\right)\{y,x\}\;.
\end{align*}
Considering the definition of shuffle permutations as in (),
equation () becomes
$$
\{\{x_1,x_2\},x_3\}+e(x_2,x_3)\{\{x_1,x_3\},x_2\}
+e(x_1,x_2)e(x_1,x_3)\{\{x_2,x_3\},x_1\}=0
$$
and we can proof this by a direct computation using the graded antisymmetric
Jacobi identity of the antisymmetric Schouten-Nijenhuis bracket as in ():
\begin{align*}
&\{\{x_1,x_2\},x_3\}
+e(x_2,x_3)\{\{x_{1},x_{3}\},x_{2}\}
+e(x_1,x_2)e(x_1,x_3)\{\{x_{2},x_{3}\},x_{1}\}\\
&=-e(x_1)e(x_2)[x_{3},\{x_{1},x_{2}\}]
-e(x_2,x_3)e(x_{1})e(x_{3})[x_{2},\{x_{1},x_{3}\}]\\
&\phantom{=.}
-e(x_1,x_2)e(x_1,x_3)e(x_{2})e(x_{3})[x_{1},\{x_{2},x_{3}\}]\\
&=-e(x_2)[x_{3},[x_{2},x_{1}]]
-e(x_2,x_3)e(x_{3})[x_{2},[x_{3},x_{1}]]\\
&\phantom{=.}
-e(x_1,x_2)e(x_1,x_3)e(x_{3})[x_{1},[x_{3},x_{2}]]\\
&=-e(x_1,x_2)e(x_1,x_3)e(x_2,x_3)e(x_{2})[x_{1},[x_{2},x_{3}]]
-e(x_2,x_3)e(x_{3})[x_{2},[x_{3},x_{1}]]\\
&\phantom{=.}
-e(x_1,x_2)e(x_1)[x_{3},[x_{1},x_{2}]]\\
&=-e(x_1,x_2)e(x_2,x_3)e(x_1)e(x_2)e(x_3)(e(x_1,x_3)e(x_1)e(x_3)[x_{1},[x_{2},x_{3}]]\\
&\phantom{=.}
+e(x_1,x_2)e(x_1)e(x_2)[x_{2},[x_{3},x_{1}]]
+e(x_2,x_3)e(x_2)e(x_3)[x_{3},[x_{1},x_{2}]])\\
&=e(x_1,x_2)e(x_2,x_3)e(x_1)e(x_2)e(x_3)\left(0\right)\\
&=0
\end{align*}
\end{comment}
\end{proof}
According to a better readable text we will always use the Koszuls sign
conventions (\ref{Koszul_convention}), when it comes to expressions, which
are graded symmetric with respect to the tensor grading.
Any morphism of Lie Rinehart pairs prolongs
to a morphism of exterior algebras and the following corollary
shows that this is in fact a morphism of graded Lie algebras with respect
to the Schouten-Nijenuis bracket:
\begin{prop}\label{schouten-morphism}
Let $(f,g):(A,\mathfrak{g})\to (B,\mathfrak{h})$ be a morphism of Lie Rinehart pairs.
The associated morphism $\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f:\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A\to\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{h}_B$ of
exterior algebras is a morphism of graded Lie algebras,
with respect to the Schouten-Nijenuis bracket.
\end{prop}
\begin{proof}$\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f$ is homogeneous of degree zero with respect to the
tensor grading. The rest follows, since $\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g$ is natural and $g$ a
Lie algebra morphism.
\begin{comment}
Moreover we compute
$\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f([a,b])=\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f(0)=f(0)=[f(a),f(b)]=
[\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f(a),\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f(b)]$ on scalars $a,b\in\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}^0_A\mathfrak{g}$ and
$\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f([a,x])= \raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f(D_x(a))=f(D_x(a))=
D_{g(x)}(f(a))=[f(a),g(x)]=[\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f(a),\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f(x)]$ on scalars and
vectors $x\in\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}^0_A\mathfrak{g}$. On simple tensors
$x_1{\scriptstyle \;\wedge\;} \cdots{\scriptstyle \;\wedge\;} x_n$,
$y_1{\scriptstyle \;\wedge\;} \cdots{\scriptstyle \;\wedge\;} y_m\in\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}$ we use
$\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f(x_1{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_n)=
g(x_1){\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} g(x_n)$ and compute
$\left[\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f\left(x_1{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_n\right),
\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f\left(y_1{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} y_m\right)\right]=$
$\textstyle\sum_{i,j}(-1)^{i+j}[g(x_i),g(y_j)]{\scriptstyle \;\wedge\;}$
${\scriptstyle \;\wedge\;} g(x_1){\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;}
\widehat{g(x_i)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} g(x_n)
{\scriptstyle \;\wedge\;}
g(y_1){\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} \widehat{g(y_j)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} g(y_m)=
$
\begin{align*}
&\textstyle g(\sum_{i,j}(-1)^{i+j}[x_i,y_j]{\scriptstyle \;\wedge\;} x_1{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;}
\widehat{x_i}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_n{\scriptstyle \;\wedge\;}
y_1{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} \widehat{y_j}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} y_m)\\
&=\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f([x_1{\scriptstyle \;\wedge\;} \cdots{\scriptstyle \;\wedge\;} x_n,y_1{\scriptstyle \;\wedge\;} \cdots{\scriptstyle \;\wedge\;} y_m])\;.
\end{align*}
The theorem then follows since any tensor is a sum of simple tensors.
\end{comment}
\end{proof}
\begin{remark}
Since $\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}^0\mathfrak{g}\simeq A$ and $\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}^1\mathfrak{g}_A\simeq \mathfrak{g}$,
there is a natural injection
$A\oplus\mathfrak{g}\hookrightarrow\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A$ of graded vector spaces
and since the Schouten-Nijenhuis bracket coincides with the bracket of
$A\oplus\mathfrak{g}$
on scalars and vectors, this is in fact a morphism of graded Lie algebras.
\end{remark}
\section{The Lie $\infty$-Algebra of a Lie Rinehart pair}We expand the
Schouten-Nijenhuis algebra into a
Lie $\infty$-algebra with non trivial higher brackets. Since the zero
morphism is the only general (co)differential in this setting,
the 'unary' bracket has to vanish.
The structure we obtain is particularly simple and merely a change
in perspective. Its real advantage lies in the fact,
that we gain access to a lot more morphisms.
Morphisms which are just not there, when we restrict to the Schouten-Nijenhuis
picture.
Finally we show that there is a weak injection of any Lie Rinehart pair into its
associated Lie $\infty$-algebra.
\begin{definition}[Higher Lie Brackets]Let $(A,\mathfrak{g})$ be a Lie
Rinehart pair with exterior algebra $\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A$ and
$[\cdot,\cdot]$ the antisymmetric Schouten-Nijenhuis bracket. Then the
\textbf{Lie n-bracket}
\begin{equation}
\{\cdot,\cdots,\cdot\}_n : \raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A \times \cdots \times
\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A \to \raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A
\end{equation}
is defined for any integer $n \geq 2$ and homogeneous tensors
$x_1,\ldots,x_n\in \raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A$ by
\begin{multline*}
\{x_1,\ldots,x_n\}_n:=\\ \textstyle\sum_{s\in Sh(2,n-2)}e(s;x_1,\ldots,x_n)e(x_{s(1)})\;
x_{s(n)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(3)}{\scriptstyle \;\wedge\;}[x_{s(2)},x_{s(1)}]
\end{multline*}
and then extended to all of $\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A$ by $A$-additivity.
\end{definition}
In particular the Lie $2$-bracket is just the symmetric Schouten-Nijenhuis
bracket. If we referee to the Lie $n$-bracket for any $n \in \mathbb{N}$, we
consider the zero operator as the 'Lie 1-bracket',
that is we write $\{\;\cdot\;\}_1$ for the zero operator
in this context.
The following proposition provides the technical details to
show that the sequence of Lie $n$-brackets defines a
non negatively graded Lie $\infty$-algebra on the exterior power of any Lie
Rinehart pair.
\begin{prop}The Lie $n$-bracket $\{\cdot,\cdots,\cdot\}_n$
is a graded symmetric, $n$-linear operator, homogeneous of
tensor degree $-1$ for any $n\in\mathbb{N}$ and for $p$, $q\in\mathbb{N}$ with $p+q=n+1$
and $p>1$ as well as $q>1$ the equation
\begin{equation}\label{jacobi_like}
\textstyle\sum_{s\in Sh(q,p-1)}e(s;x_1,\ldots,x_n)
\{\{x_{s(1)},\ldots,x_{s(q)}\}_q,x_{s(q+1)},\ldots,x_{s(n)}\}_p=0
\end{equation}
is satisfied.
\end{prop}
\begin{proof}
The exterior product is homogeneous of degree zero and the
antisymmetric Schouten-Nijenhuis bracket is
homogeneous of degree $-1$,
with respect to the tensor grading. It follows that any $n$-ary bracket
is homogeneous of degree $-1$.
To proof graded symmetry observe that the exterior product is graded
symmetric with respect to the tensor grading and that the expression
$e(x_1)[x_2,x_1]$ is precisely the graded symmetric Schouten-Nijenhuis bracket.
The symmetry of the Lie $n$-bracket then follows since it is a graded symmetric
composition of both operators.
\begin{comment}
To proof graded symmetry, it is enough to show it on transpositions
of consecutive neighbors. Suppose $j\in\mathbb{N}$ with $1\leq j\leq (n-1)$.
Then regarding a transposition $j\leftrightarrow j+1$,
there are three cases in the definition of the $n$-ary bracket to
consider:
First there are the shuffles such that both $x_j$ and $x_{j+1}$ are
'left to the bracket'. In this case we can apply the graded symmetry of
the wedge product (with respect to the tensor grading) to get:
$$
x_\bullet{\scriptstyle \;\wedge\;} \cdots {\scriptstyle \;\wedge\;} x_{j+1}{\scriptstyle \;\wedge\;} x_j
{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} [x_\bullet,x_\bullet]=
e(x_j,x_{j+1})x_\bullet{\scriptstyle \;\wedge\;} \cdots {\scriptstyle \;\wedge\;} x_{j}{\scriptstyle \;\wedge\;} x_{j+1}
{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} [x_\bullet,x_\bullet]
$$
Second there is a single shuffle, such that both $x_j$ and $x_{j+1}$
are 'inside the Schouten-Nijenhuis' bracket. In this case we use the
graded antisymmetry of $[\cdot,\cdot]$ to get
$$
e(x_j)x_\bullet{\scriptstyle \;\wedge\;} \cdots {\scriptstyle \;\wedge\;} x_\bullet{\scriptstyle \;\wedge\;} [x_{j+1},x_j]=
e(x_j,x_{j+1})e(x_{j+1})x_\bullet{\scriptstyle \;\wedge\;} \cdots {\scriptstyle \;\wedge\;} x_\bullet{\scriptstyle \;\wedge\;}
[x_{j},x_{j+1}]
$$
Last there are shuffles $s\in Sh(2,n-2)$ with
$e(s;x_1,\ldots,x_n)x_\bullet{\scriptstyle \;\wedge\;} \cdots{\scriptstyle \;\wedge\;} x_{j+1}{\scriptstyle \;\wedge\;} [x_j,x_\bullet]$ and shuffles $s'\in Sh(2,n-2)$ with
$e(s';x_1,\ldots,x_n)
x_\bullet{\scriptstyle \;\wedge\;} \cdots{\scriptstyle \;\wedge\;} x_{j}{\scriptstyle \;\wedge\;} [x_{j+1},x_\bullet]$.
and we have
\begin{multline*}
e(s;x_1,\ldots,x_n)x_\bullet{\scriptstyle \;\wedge\;} \cdots{\scriptstyle \;\wedge\;} x_{j+1}{\scriptstyle \;\wedge\;} [x_j,x_\bullet]=\\
e(x_j,x_{j+1})e(s';x_1,\ldots,x_n)
x_\bullet{\scriptstyle \;\wedge\;} \cdots{\scriptstyle \;\wedge\;} x_{j}{\scriptstyle \;\wedge\;} [x_{j+1},x_\bullet]
\end{multline*}
for each such pair.
Taking all these cases together we arrive at the
graded symmetry of the $n$-ary bracket
$$
\{x_1,\ldots,x_j,x_{j+1},\ldots,x_n\}=
e(x_j,x_{j+1})\{x_1,\ldots,x_{j+1},x_j,\ldots,x_n\}
$$
for permutations of consecutive neighbors and hence on any permutation.
\end{comment}
Now to see that any of the 'Jacobi-like' shuffle sums
$$
\textstyle\sum_{s\in Sh\left(q,p-1\right)}e\left(s;x_1,\ldots,x_n\right)
\{\{x_{s(1)},\ldots,x_{s(q)}\}_q,x_{s(q+1)},\ldots,x_{s(n)}\}_p
$$
vanishes, observe that for $n=3$ and $p=q=2$ this is nothing but the
ordinary (graded symmetric) Jacobi expression (\ref{symmetric_jacobi}).
To see it for $n\geq 4$,
let $x_1,\ldots,x_n\in\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A$ be homogeneous and
$p,q\in\mathbb{N}$ with $p+q=n+1$ and $p,q\geq 2$.
First use the graded symmetry of the brackets, to rewrite the shuffle sum
into a sum over arbitrary permutations. This gives
$$\textstyle\frac{1}{q!\left(p-1\right)!}\sum_{s\in S_n}
e\left(s;x_1,\ldots,x_n\right)
\{\{x_{s(1)},\ldots,x_{s(q)}\}_q,x_{s(q+1)},\ldots,x_{s(n)}\}_p\;.
$$
Then apply the definition of $\{\cdot,\ldots,\cdot\}_q$.
Writing $Sh_{\{s(1),\ldots,s(q)\}}\left(i,j\right)$ for
the set of all $(i,j)$-shuffles but explicit as permutations of the set $\{s(1),\ldots,s(q)\}$
the previous expression becomes
\begin{multline*}
\textstyle\frac{1}{q!\left(p-1\right)!}\sum_{s\in S_n}
\sum_{t\in Sh_{\{s(1),\ldots,s(q)\}}\left(2,q-2\right)}
e\left(s;x_1,\ldots,x_n\right)\,e\left(t;x_{s(1)},\ldots,x_{s(q)}\right)\cdot \\
\phantom{=.}\cdot e\left(x_{ts(1)}\right)\{x_{ts(q)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{ts(3)}{\scriptstyle \;\wedge\;}[\,x_{ts(2)},x_{ts(1)}\,],
x_{s(q+1)},\ldots,x_{s(n)}\}_p\;.
\end{multline*}
Now observe, that for any $s\in S_n$
and shuffle $t\in Sh_{\{s(1),\ldots,s(q)\}}(2,q-2)$,
the permutation $(ts(1),\ldots,ts(q),s(q+1),\ldots,s(n))$ is again an element of
$S_n$ and since there are precisely $\frac{q!}{(q-2)!2!}$ many
shuffles in $Sh_{\{s(1),\ldots,s(q)\}}(2,q-2)$ we can just 'absorb' the second sum
over shuffles in the previous expression into the sum over general permutation. After reindexing we get:
\begin{multline*}
\textstyle\frac{1}{q!(p-1)!}\frac{q!}{(q-2)!2!}\sum_{s\in S_n}
e\left(s;x_1,\ldots,x_n\right)e\left(x_{s(1)}\right)\cdot\\
\phantom{=.}\cdot
\{x_{s(q)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(3)}{\scriptstyle \;\wedge\;}[\,x_{s(2)},x_{s(1)}\,],
x_{s(q+1)},\ldots,x_{s(n)}\}_p
\end{multline*}
Then apply the definition of the bracket a second time.
Care has to be taken regarding the first element.
In fact there are two possible positions for it:
At the most right position, i.e as the second argument inside of the Schouten-Nijenhuis bracket or at the most right position outside of the bracket. Taking
this into account split the expression into two parts according to the position
of the first element. If $p=2$ the second shuffle sum is omitted:
\begin{align*}
&\textstyle\frac{1}{2(p-1)!(q-2)!}\sum_{s\in S_n}
\sum_{t\in Sh_{\{s(q+1),\ldots,s(n)\}}(1,p-2)} e\left(s;x_1,\ldots,x_n\right)\cdot\\
&\phantom{=.}\cdot e\left(t;x_{s(q+1)},\ldots,x_{s(n)}\right)e\left(x_{s(1)}\right)
e\left(x_{s(q)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;}
x_{s(3)}{\scriptstyle \;\wedge\;}[\,x_{s(2)},x_{s(1)}\,]\,\right)\cdot\\
&\phantom{=.}\cdot x_{ts(n)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{ts(q+2)}{\scriptstyle \;\wedge\;}
[x_{ts(q+1)},x_{s(q)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(3)}{\scriptstyle \;\wedge\;}[x_{s(2)},x_{s(1)}]]\\
&+\textstyle\frac{1}{2\left(p-1\right)!\left(q-2\right)!}\sum_{s\in S_n}
\sum_{t\in Sh_{\{s(q+1),\ldots,s(n)\}}\left(2,p-3\right)}
e\left(s;x_1,\ldots,x_n\right)\,\cdot\\
&\phantom{=.}\cdot e\left(t;x_{s(q+1)},\ldots,x_{s(n)}\right)
e\left(x_{s(q)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(3)}{\scriptstyle \;\wedge\;}
[x_{s(2)},x_{s(1)}],x_{ts(q+1)}\right)\,\cdot\\
&\phantom{=.}\cdot
e\left(x_{s(q)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(3)}{\scriptstyle \;\wedge\;}
[x_{s(2)},x_{s(1)}],x_{ts(q+2)}\right)e\left(x_{s(1)}\right)
e\left(x_{ts(q+1)}\right)\,\cdot\\
&\phantom{=.}\cdot
x_{ts(n)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{ts(q+3)}{\scriptstyle \;\wedge\;}
x_{s(q)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(3)}{\scriptstyle \;\wedge\;}[x_{s(2)},x_{s(1)}]{\scriptstyle \;\wedge\;}
[x_{ts(q+2)},x_{ts(q+1)}]
\end{align*}
Again we 'absorb' the additional shuffle sums in both cases into the
appropriate sum over general permutations. After reindexing and simplification
this becomes
\begin{comment}
\begin{align*}
&-\textstyle\frac{1}{2(p-1)!(q-2)!}\frac{(p-1)!}{1!(p-2)!}\sum_{s\in S_n}
e\left(s;x_1,\ldots,x_n\right)
e\left(x_{s(2)}\right)\cdots e\left(x_{s(q)}\right)\,\cdot\\
&\phantom{=.}\cdot
x_{s(n)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s_(q+2)}{\scriptstyle \;\wedge\;}
[\,x_{s(q+1)},x_{s(q)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(3)}{\scriptstyle \;\wedge\;}
[\,x_{s(2)},x_{s(1)}\,]\,]\\
&+\textstyle\frac{1}{2(p-1)!(q-2)!}\frac{(p-1)!}{2!(p-3)!}\sum_{s\in S_n}
e\left(s;x_1,\ldots,x_n\right)
e\left(x_{s(q)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(3)}{\scriptstyle \;\wedge\;}
[x_{s(2)},x_{s(1)}],x_{s(q+1)}\right)\,\cdot\\
&\phantom{=.}\cdot
e\left(x_{s(q)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(3)}{\scriptstyle \;\wedge\;}
[x_{s(2)},x_{s(1)}],x_{s(q+2)}\right)e\left(x_{s(1)}\right)
e\left(x_{s(q+1)}\right)\,\cdot\\
&\phantom{=.}\cdot
x_{s(n)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(q+3)}{\scriptstyle \;\wedge\;}
x_{s(q)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(3)}{\scriptstyle \;\wedge\;}[x_{s(2)},x_{s(1)}]{\scriptstyle \;\wedge\;}
[x_{s(q+2)},x_{s(q+1)}]\\
&=-\textstyle\frac{1}{2(q-2)!(q-2)!}\sum_{s\in S_n}
e\left(s;x_1,\ldots,x_n\right)
e\left(x_{s(2)}\right)\cdots e\left(x_{s(q)}\right)\,\cdot\\
&\phantom{=.}\cdot
x_{s(n)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s_(q+2)}{\scriptstyle \;\wedge\;}
[\,x_{s(q+1)},x_{s(q)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(3)}{\scriptstyle \;\wedge\;}
[\,x_{s(2)},x_{s(1)}\,]\,]\\
&+\textstyle\frac{1}{4(q-2)!(p-3)!}\sum_{s\in S_n}
e\left(s;x_1,\ldots,x_n\right)
e\left(x_{s(1)},x_{s(q+1)}\right)\cdots e\left(x_{s(q)},x_{s(q+1)}\right)\,\cdot\\
&\phantom{=.}\cdot
e\left(x_{s(1)},x_{s(q+2)}\right)\cdots e\left(x_{s(q)},x_{s(q+2)}\right)
e\left(x_{s(1)}\right)e\left(x_{s(q+2)}\right)\,\cdot\\
&\phantom{=.}\cdot
x_{s(n)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(q+3)}{\scriptstyle \;\wedge\;}
x_{s(q)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(3)}{\scriptstyle \;\wedge\;}[\,x_{s(2)},x_{s(1)}\,]
{\scriptstyle \;\wedge\;} [x_{s(q+2)},x_{s(q+1)}]
\end{align*}
\end{comment}
\begin{align}
&-\textstyle\frac{1}{2(q-2)!(p-2)!}\sum_{s\in S_n}
e(s;x_1,\ldots,x_n)e(x_{s(2)})\cdots e(x_{s(q)})\,
x_{s(n)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(q+2)}{\scriptstyle \;\wedge\;}\nonumber\\
&\phantom{=.}
{\scriptstyle \;\wedge\;} [\,x_{s(q+1)},x_{s(q)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(3)}{\scriptstyle \;\wedge\;}
[x_{s(2)},x_{s(1)}\,]\,]\label{eq_19}\\
&+\textstyle\frac{1}{4(q-2)!(p-3)!}\sum_{s\in S_n}
e\left(s;x_1,\ldots,x_n\right)e\left(x_{s(2)}\right)e\left(x_{s(3)}\right)\,
x_{s(n)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(5)}{\scriptstyle \;\wedge\;}\nonumber\\
&\phantom{=.}
{\scriptstyle \;\wedge\;} [x_{s(4)},x_{s(3)}]{\scriptstyle \;\wedge\;} [x_{s(2)},x_{s(1)}]\nonumber
\end{align}
and again the second sum is omitted for $p=2$. Now lets look on both sums separately. From
\begin{align*}
&e(x_2)e(x_3)[x_4,x_3]{\scriptstyle \;\wedge\;} [x_2,x_1]
= e(x_2)e(x_3)e([x_2,x_1],[x_4,x_3])[x_2,x_1]{\scriptstyle \;\wedge\;} [x_4,x_3]\\
&=-e(x_1,x_3)e(x_1,x_4)e(x_2,x_3)e(x_2,x_4)e(x_1)e(x_4)[x_2,x_1]{\scriptstyle \;\wedge\;} [x_4,x_3]
\end{align*}
we see that the expression
\begin{multline}\label{eq_aa}
\textstyle\sum_{s\in S_n}e(s;x_1,\ldots,x_n)e(x_{s(2)})e(x_{s(3)})\;\cdot\\
\cdot x_{s(n)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(5)}{\scriptstyle \;\wedge\;}[x_{s(4)},x_{s(3)}]{\scriptstyle \;\wedge\;}
[x_{s(2)},x_{s(1)}]
\end{multline}
vanishes, since for any permutation $s\in S_n$ there is precisely one permutation
$t\in S_n$ with $t=(s(3),s(4),s(1),s(2),s(5),\ldots,s(n))$
and then the term for $t$ cancel against the term for $s$.
Now only the first sum in (\ref{eq_19}) remains, but for $q=2$ it
vanishes too due to the graded Jacobi equation of the symmetric
Schouten-Nijenhuis bracket.
\begin{comment}
\begin{align*}
0&=\textstyle\sum_{s\in Sh(2,1)}e(s;x_{1},x_{2},x_{3})
\{\{x_{s(1)},x_{s(2)}\},x_{s(3)}\}\\
&=\textstyle\frac{1}{2}\sum_{s\in S_{3}}e(s;x_{1},x_{2},x_{3})
\{\{x_{s(1)},x_{s(2)}\},x_{s(3)}\}\\
&=\textstyle\frac{1}{2}\sum_{s\in S_{3}}e(s;x_{1},x_{2},x_{3})e(x_{s(1)})
\{[x_{s(2)},x_{s(1)}],x_{s(3)}\}\\
&=\textstyle\frac{1}{2}\sum_{s\in S_{3}}e(s;x_{1},x_{2},x_{3})e(x_{s(1)})
e([x_{s(2)},x_{s(1)}])[x_{s(3)},[x_{s(2)},x_{s(1)}]]\\
&=-\textstyle\frac{1}{2}\sum_{s\in S_{3}}e(s;x_{1},x_{2},x_{3})
e(x_{s(2)})[x_{s(3)},[x_{s(2)},x_{s(1)}]]
\end{align*}
\end{comment}
This proofs the equation for $q=2$.
For $q\geq 3$ use the Poisson identity
$[x,y{\scriptstyle \;\wedge\;} z]=[x,y]{\scriptstyle \;\wedge\;} z+e(x,y)e(y)y{\scriptstyle \;\wedge\;} [x,z]$
of the antisymmetric Schouten-Nijenhuis bracket and omit the constant
factor in (\ref{eq_19}). After simplification we get
\begin{comment}
\begin{align*}
&\textstyle\sum_{s\in S_n}e\left(s;x_1,\ldots,x_n\right)
e\left(x_{s(2)}\right)\cdots e\left(x_{s(q)}\right)\,\cdot\\
&\phantom{=.}\cdot
x_{s(n)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(q+2)}{\scriptstyle \;\wedge\;}
[x_{s(q+1)},x_{s(q)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(3)}]
{\scriptstyle \;\wedge\;}[x_{s(2)},x_{s(1)}]\\
&+\textstyle\sum_{s\in S_n}e\left(s;x_1,\ldots,x_n\right)
e\left(x_{s(2)}\right)\cdots e\left(x_{s(q)}\right)\,\cdot\\
&\phantom{=.}\cdot
e\left(x_{s(q+1)},x_{s(q)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(3)}\right)\,
e\left(x_{s(q)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(3)}]\right)\,\cdot\\
&\phantom{=.}\cdot
x_{s_n}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s_{q+2}}{\scriptstyle \;\wedge\;}
x_{s_q}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s_3}{\scriptstyle \;\wedge\;}[x_{s_{q+1}},[x_{s_2},x_{s_1}]]
\end{align*}
\end{comment}
\begin{align}\label{eq_21}
&\textstyle\sum_{s\in S_n}e\left(s;x_1,\ldots,x_n\right)
e\left(x_{s(2)}\right)\cdots e\left(x_{s(q)}\right)\,\cdot\nonumber\\
&\phantom{=.}\cdot
x_{s(n)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(q+2)}{\scriptstyle \;\wedge\;}
[x_{s(q+1)},x_{s(q)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(3)}]
{\scriptstyle \;\wedge\;}[x_{s(2)},x_{s(1)}]\\
&+\textstyle\sum_{s\in S_n}e\left(s;x_1,\ldots,x_n\right)
e\left(x_{s(2)}\right)
x_{s(n)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(4)}{\scriptstyle \;\wedge\;}[x_{s(3)},[x_{s(2)},x_{s(1)}]]
\nonumber\,.
\end{align}
The second sum vanishes due to the graded Jacobi equation of the
symmetric Schouten-Nijenhuis bracket. For $q=3$ the first sum vanishes too,
since we arrive at situation (\ref{eq_aa}). This proofs the equation for
$q=3$.
For $q\geq 4$ we apply the Poisson identity to the remaining part again and
after simplification we get
\begin{comment}
\begin{align*}
&\textstyle\sum_{s\in S_n}e\left(s;x_1,\ldots,x_n\right)
e\left(x_{s(2)}\right)\cdots e\left(x_{s(q)}\right)\,\cdot\\
&\phantom{=.}\cdot
x_{s(n)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(q+2)}{\scriptstyle \;\wedge\;}
[x_{s(q+1)},x_{s(q)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(3)}]
{\scriptstyle \;\wedge\;}[x_{s(2)},x_{s(1)}]\\
&=\textstyle\sum_{s\in S_n}e\left(s;x_1,\ldots,x_n\right)
e\left(x_{s(2)}\right)\cdots e\left(x_{s(q)}\right)\,\cdot\\
&\phantom{=.}\cdot
x_{s(n)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(q+2)}{\scriptstyle \;\wedge\;}
[x_{s(q+1)},x_{s(q)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(4)}]{\scriptstyle \;\wedge\;} x_{s(3)}
{\scriptstyle \;\wedge\;}[x_{s(2)},x_{s(1)}]\\
&+\textstyle\sum_{s\in S_n}e\left(s;x_1,\ldots,x_n\right)
e\left(x_{s(2)}\right)\cdots e\left(x_{s(q)}\right)\,\cdot\\
&\phantom{=.}\cdot
e(x_{s(q+1)},x_{s(q)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(4)})
e(x_{s(q)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(4)})\\
&\phantom{=.}\cdot
x_{s(n)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(q+2)}{\scriptstyle \;\wedge\;}
x_{s(q)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(4)}{\scriptstyle \;\wedge\;}
[x_{s(q+1)},x_{s(3)}]{\scriptstyle \;\wedge\;}[x_{s(2)},x_{s(1)}]\\
&=\textstyle\sum_{s\in S_n}e\left(s;x_1,\ldots,x_n\right)
e\left(x_{s(2)}\right)\cdots e\left(x_{s(q)}\right)\,\cdot\\
&\phantom{=.}\cdot
e(x_{s(3)},[x_{s(q+1)},x_{s(q)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(4)}])\\
&\phantom{=.}\cdot
x_{s(n)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(q+2)}{\scriptstyle \;\wedge\;}
x_{s(3)}{\scriptstyle \;\wedge\;}[x_{s(q+1)},x_{s(q)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(4)}]
{\scriptstyle \;\wedge\;}[x_{s(2)},x_{s(1)}]\\
&+\textstyle\sum_{s\in S_n}e\left(s;x_1,\ldots,x_n\right)
e\left(x_{s(2)}\right)e\left(x_{s(3)}\right)
x_{s(n)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(5)}{\scriptstyle \;\wedge\;}
[x_{s(4)},x_{s(3)}]{\scriptstyle \;\wedge\;}[x_{s(2)},x_{s(1)}]\\
&=\textstyle\sum_{s\in S_n}e\left(s;x_1,\ldots,x_n\right)
e\left(x_{s(2)}\right)\cdots e\left(x_{s(q)}\right)\,\cdot\\
&\phantom{=.}\cdot e(x_{s(3)})
e(x_{s(3)},x_{s(4)})\cdots e(x_{s(3)},x_{s(q+1)})\\
&\phantom{=.}\cdot
x_{s(n)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(q+2)}{\scriptstyle \;\wedge\;}
x_{s(3)}{\scriptstyle \;\wedge\;}[x_{s(q+1)},x_{s(q)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(4)}]
{\scriptstyle \;\wedge\;}[x_{s(2)},x_{s(1)}]\\
&+\textstyle\sum_{s\in S_n}e\left(s;x_1,\ldots,x_n\right)
e\left(x_{s(2)}\right)e\left(x_{s(3)}\right)
x_{s(n)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(5)}{\scriptstyle \;\wedge\;}
[x_{s(4)},x_{s(3)}]{\scriptstyle \;\wedge\;}[x_{s(2)},x_{s(1)}]\\
&=\textstyle\sum_{s\in S_n}e\left(s;x_1,\ldots,x_n\right)
e\left(x_{s(2)}\right)\widehat{e\left(x_{s(4)}\right)}
\cdots e\left(x_{s(q)}\right)\,\cdot\\
&\phantom{=.}\cdot
e(x_{s(4)},x_{s(5)})\cdots e(x_{s(4)},x_{s(q+1)})\\
&\phantom{=.}\cdot
x_{s(n)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(q+2)}{\scriptstyle \;\wedge\;}
x_{s(4)}{\scriptstyle \;\wedge\;}[x_{s(q+1)},x_{s(q)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(3)}]
{\scriptstyle \;\wedge\;}[x_{s(2)},x_{s(1)}]\\
&+\textstyle\sum_{s\in S_n}e\left(s;x_1,\ldots,x_n\right)
e\left(x_{s(2)}\right)e\left(x_{s(3)}\right)
x_{s(n)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(5)}{\scriptstyle \;\wedge\;}
[x_{s(4)},x_{s(3)}]{\scriptstyle \;\wedge\;}[x_{s(2)},x_{s(1)}]\\
&=repeat-transpositions
\end{align*}
\end{comment}
\begin{align*}
&\textstyle\sum_{s\in S_n}e\left(s;x_1,\ldots,x_n\right)
e\left(x_{s(2)}\right)e\left(x_{s(3)}\right)
x_{s(n)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(5)}{\scriptstyle \;\wedge\;}
[x_{s(4)},x_{s(3)}]{\scriptstyle \;\wedge\;}[x_{s(2)},x_{s(1)}]\\
&+\textstyle\sum_{s\in S_n}e\left(s;x_1,\ldots,x_n\right)
e\left(x_{s(2)}\right)\cdots e\left(x_{s(q-1)}\right)\,\cdot\\
&\phantom{=.}\cdot
x_{s(n)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;}
x_{s(q+1)}{\scriptstyle \;\wedge\;}[x_{s(q)},x_{s(q-1)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(3)}]
{\scriptstyle \;\wedge\;}[x_{s(2)},x_{s(1)}]
\end{align*}
Again the first sum vanishes since it is the same situation as in (\ref{eq_aa})
and the second sum vanishes for $q=4$ due to the same reason.
For $q>4$ the second sum equals the first in (\ref{eq_21}) but for $q-1$ instead.
Consequently we have to repeat the last computation
$(q-4)$-times, to arrive at an expression that is equal to (\ref{eq_aa}) and hence
vanishes. This proofs the equation.
\end{proof}
Taking into account, that the unary bracket $\{\cdot\}_1$
has to be the zero operator, we can combine
the brackets into a Lie $\infty$-algebra on the exterior power of any Lie Rinehart
pair:
\begin{theorem}[The Lie $\infty$-algebra]
Let $(A,\mathfrak{g})$ be a Lie Rinehart pair with exterior
power $\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A$. Then
$\left(\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A,(\{\cdot,\ldots,\cdot\}_k)_{k\in\mathbb{N}}\right)$
is a Lie $\infty$-algebra, concentrated in non negative degrees.
\end{theorem}
\begin{proof}
All operators are graded symmetric and homogeneous of degree $-1$ with respect
to the tensor grading and the weak Jacobi identities (\ref{sh_Jacobi})
follow from (\ref{jacobi_like}), since $\{\cdot\}_1$ is the zero operator.
Moreover the exterior algebra is concentrated in non-negative degrees with
respect to the tensor grading.
\end{proof}
\begin{remark}
Note that the Lie $\infty$-structure is particularly simple in this case:
The (co)differential is the zero operator and from (\ref{jacobi_like}) we see, that
each particular shuffle sum already vanishes for fixed $p$ and $q$ in the
weak Jacobi identities (\ref{sh_Jacobi}).
\end{remark}
The following theorem shows, that the construction is natural with respect to
morphisms of Lie Rinehart pairs. In fact any morphism of Lie Rinehart pairs
gives rise to a \textit{strict} morphism of Lie $\infty$-algebras:
\begin{theorem}
Let $(f,g):(A,\mathfrak{g})\to (B,\mathfrak{h})$ be a morphism of Lie Rinehart pairs.
The associated exterior algebra morphism
$\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f:\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A\to\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{h}_B$ is a strict morphism of
Lie $\infty$-algebras.
\end{theorem}
\begin{proof}We need to show that $\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f$ commutes
with the Lie $n$-bracket for any $n\in\mathbb{N}$.
For $n=1$ this is trivial and for $n=2$ this is proposition (\ref{schouten-morphism}).
For $n\geq 3$ it follows, since $\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g$ commutes with the exterior
product and the Schouten-Nijenhuis bracket.
\begin{comment}
and homogeneous tensors
$x_1,\ldots,x_n\in\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A$ compute
\begin{align*}
&\{\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f(x_1),\ldots,\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f(x_1)\}_n=
\textstyle\sum_{s\in Sh(2,n-2)}e(s;x_1,\ldots,x_n)e(\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f(x_{s(1)}))\;\cdot\\
&\phantom{=.}\cdot\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f(x_{s(n)}){\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;}
\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f(x_{s(3)}){\scriptstyle \;\wedge\;}[\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f(x_{s(2)}),\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f(x_{s(1)})]\\
&=\textstyle\sum_{s\in Sh(2,n-2)}e(s;x_1,\ldots,x_n)e(\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f(x_{s(1)}))\;\cdot\\
&\phantom{=.}\cdot\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f(x_{s(n)}){\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;}
\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f(x_{s(3)}){\scriptstyle \;\wedge\;} \raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f([x_{s(2)},x_{s(1)}])\\
&=\textstyle\sum_{s\in Sh(2,n-2)}e(s;x_1,\ldots,x_n)e(x_{s(1)})
\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f(x_{s(n)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;}
x_{s(3)}){\scriptstyle \;\wedge\;} [x_{s(2)},x_{s(1)}])\\
&= \raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$} g_f (\{x_1,\ldots x_n\}_n)\;.
\end{align*}
Since any tensor is a finite sum of homogeneous tensors, the theorem follows.
\end{comment}
\end{proof}
The natural injection
$A\oplus\mathfrak{g}\hookrightarrow \raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A\;;\;
(a,x)\mapsto(a,x)$
can't be a morphism of Lie $\infty$-algebras,
since it has to commute with all higher brackets,
but these brackets are zero on $A\oplus\mathfrak{g}$.
However as the following theorem shows, a
natural injection now comes as a weak morphism of Lie $\infty$-algebras:
\begin{definition}Let $(A,\mathfrak{g})$ be a Lie Rinehart pair with
associated graded Lie algebra $A\oplus\mathfrak{g}$,
exterior algebra $\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A$ and
\begin{equation}
\begin{array}{crcl}
i_n: & A\oplus\mathfrak{g}\times \cdots \times A\oplus\mathfrak{g} &\to &
\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A\\
& (x_1,\ldots,x_n) & \mapsto &
(-1)^{n-1}(n-1)\,!\cdot x_n{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_1
\end{array}
\end{equation}
for any $n\in \mathbb{N}$. Then the sequence $i_\infty:=(i_n)_{n\in\mathbb{N}}$ is called the
\textbf{natural injection} of the Lie Rinehart pair into its exterior
Lie $\infty$-algebra.
\end{definition}
The following theorem shown, that this sequence of multilinear maps is in fact
a morphism of Lie $\infty$-algebras:
\begin{theorem}The natural injection
$i_\infty: A\oplus\mathfrak{g}\to \raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A$ is a morphism of Lie $\infty$-algebras.
\end{theorem}
\begin{proof}Any map $i_{n}$ is graded symmetric and homogeneous of
tensor degree zero, since the same holds for the exterior product.
In (the Lie $\infty$-algebra) $A\oplus\mathfrak{g}$, only the binary bracket
does not vanish and since $\{\cdot\}_1$ is the zero operator, the general structure equation (\ref{lie-infty-morph})
of a Lie $\infty$-algebra morphism simplifies for any $n\geq 2$ into
\begin{multline*}
\textstyle\sum_{s\in Sh(2,n-2)}e(s)
i_{n-1}([x_{s(1)},x_{s(2)}],x_{s(3)},\ldots,x_{s(n)})=\\
\textstyle\sum_{p=2}^n
\frac{1}{p!}\sum_{s\in Sh(j_{1},\ldots,j_{p})}^{j_{1}+\ldots+j_{p}=n}
e(s)\{i_{j_{1}}(x_{s(1)},\ldots,x_{s(k_{1})}),...,
i_{j_{p}}(x_{s(n-j_{p}+1)},\ldots,x_{s(n)})\}_p\;.
\end{multline*}
Now assume $n\geq 2$ and $2\leq p \leq n$ as well as $j_1+\ldots+j_p=n$ for positive
integers $j_k$. We use the graded symmetry of the Lie $n$-brackets and the
maps $i_n$,
to rewrite the shuffle sum at the right side of the structure equation into a sum over arbitrary permutations:
\begin{multline*}
\textstyle\sum_{s\in Sh(j_{1},\ldots,j_{p})}e(s)
\{i_{j_{1}}(x_{s(1)},\ldots,x_{s(j_{1})}),
\ldots,i_{j_{p}}(x_{s(n-j_{p}+1)},\ldots,x_{s(n)})\}_{p}=\\
\textstyle\frac{1}{j_{1}!\cdots j_{p}!}\sum_{s\in S_{n}}e(s)
\{i_{j_{1}}(x_{s(1)},\ldots,x_{s(j_{1})}),
\ldots,i_{j_{p}}(x_{s(n-j_{p}+1)},\ldots,x_{s(n)})\}_{p}
\end{multline*}
To reorganize this, define $j_0:=0$ and write
$X_{k}^{s}:=i_{j_{k}}(x_{s(j_{k-1}+1)},...,x_{s(j_{k-1}+j_{k})})$
for any given vectors $x_1,\ldots,x_n\in\mathfrak{g}$, permutation $s\in S_n$
and $1\leq k \leq p$.
Then $|X_{k}^s|=j_{k}$ and we can abbreviate the previous expression into
$$\textstyle\frac{1}{j_{1}!\cdots j_{p}!}
\sum_{s\in S_{n}}e(s)\{X_{1}^s,\ldots,X_{p}^s\}_{p}\;.$$
Applying the definition of the Lie $p$-bracket
is now straight forward and leads to the expression
\begin{multline*}
\textstyle\frac{1}{j_{1}!\cdots j_{p}!}\sum_{s\in S_{n}}e(s)
\sum_{t\in Sh(2,p-2)}e(t,X_{1}^{s},\ldots,X_{p}^{s})\cdot\\
\cdot e(X_{t(1)}^{s})X_{t(p)}^{s}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;}
X_{t(3)}^{s}{\scriptstyle \;\wedge\;}[X_{t(2)}^{s},X_{t(1)}^{s}]\;.
\end{multline*}
Substituting the definition of each $i_{j_k}$ back and using
$e(X_{t(1)}^{s})=(-1)^{j_{t(1)}}$ this rewrites into
\begin{multline*}
\textstyle\frac{1}{j_{1}!\cdots j_{p}!}\sum_{s\in S_{n}}e(s)
\sum_{t\in Sh(2,p-2)}e(t,X_{1}^{s},\ldots,X_{p}^{s})
(-1)^{(j_{t(p)}-1)}\cdots(-1)^{(j_{t(1)}-1)}\cdot\\
\cdot(-1)^{j_{t(1)}}(j_{t(p)}-1)!\cdots(j_{t(1)}-1)!
(x_{s(j_{t(p)-1}+j_{t(p)})}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(j_{t(p)-1}+1)})
{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;}\\
{\scriptstyle \;\wedge\;}(x_{s(j_{t(3)-1}+j_{t(3)})}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(j_{t(3)-1}+1)})
{\scriptstyle \;\wedge\;}\\
{\scriptstyle \;\wedge\;}[x_{s(j_{t(2)-1}+j_{t(2)})}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(j_{t(2)-1}+1)},
x_{s(j_{t(1)-1}+j_{t(1)})}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(j_{t(1)-1}+1)}]
\end{multline*}
and since $e(s)$ as well as $e(t,X_1^s,\ldots,X^s_p)$ keeps properly track of the signs
we can reindex this. After simplification using
$j_1+\cdots +j_p=n$ we get
\begin{multline*}
\textstyle
\frac{1}{j_{1}\cdots j_{p}}\sum_{s\in S_{n}}e(s)
\sum_{t\in Sh(2,p-2)}
(-1)^{j_{t(1)}+n-p}\;
x_{s(n)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(j_{t(1)}+j_{t(2)}+1)}{\scriptstyle \;\wedge\;}\\
{\scriptstyle \;\wedge\;} [x_{s(j_{t(1)}+j_{t(2)})}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(j_{t(1)}+1)},
x_{s(j_{t(1)})}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(1)}]\;.
\end{multline*}
Since all arguments are actually vectors, we can apply the
symmetric defining expression (\ref{SN_1}) of the Schouten-Nijenuis bracket
to simplify this further into
\begin{comment}
\begin{align*}
&\textstyle[y_{m}\wedge...\wedge y_{1},x_{n}\wedge...\wedge x_{1}]
=\textstyle(-1)^{\frac{m(m-1)}{2}}(-1)^{\frac{n(n-1)}{2}}
[y_{1}\wedge...\wedge y_{m},x_{1}\wedge...\wedge x_{n}]\\
&=\textstyle\sum_{s\in Sh(1,m-1)}\sum_{t\in Sh(1,n-1)}
(-1)^{\frac{m(m-1)}{2}}(-1)^{\frac{n(n-1)}{2}}e(s)e(t)
[y_{s(1)},x_{t(1)}]\wedge y_{s(2)}\wedge...\wedge y_{s(m)}
\wedge x_{t(2)}\wedge...\wedge x_{t(n)}\\
&=\textstyle\sum_{s\in Sh(1,m-1)}\sum_{t\in Sh(1,n-1)}
(-1)^{\frac{m(m-1)}{2}}(-1)^{\frac{n(n-1)}{2}}
(-1)^{m+n-2}e(s)e(t)
y_{s(2)}\wedge...\wedge y_{s(m)}\wedge x_{t(2)}\wedge...\wedge x_{t(n)}
\wedge[y_{s(1)},x_{t(1)}]\\
&=\textstyle\sum_{s\in Sh(1,m-1)}\sum_{t\in Sh(1,n-1)}
(-1)^{\frac{m(m-1)}{2}+\frac{(m-1)(m-2)}{2}}
(-1)^{\frac{n(n-1)}{2}+\frac{(n-1)(n-2)}{2}}
(-1)^{m+n}e(s)e(t)
y_{s(m)}\wedge...\wedge y_{s(2)}\wedge x_{t(n)}\wedge...\wedge x_{t(2)}\wedge
[y_{s(1)},x_{t(1)}]\\
&=\textstyle\sum_{s\in Sh(1,m-1)}\sum_{t\in Sh(1,n-1)}
(-1)^{\frac{m^{2}-m+m^{2}-m-2m+2}{2}}
(-1)^{\frac{n^{2}-n+n^{2}-n-2n+2}{2}}
(-1)^{m+n}e(s)e(t)
y_{s(m)}\wedge...\wedge y_{s(2)}\wedge x_{t(n)}\wedge...\wedge x_{t(2)}\wedge
[y_{s(1)},x_{t(1)}]\\
&=\textstyle\sum_{s\in Sh(1,m-1)}\sum_{t\in Sh(1,n-1)}
(-1)^{\frac{2m^{2}-4m+2}{2}}(-1)^{\frac{2n^{2}-4n+2}{2}}(-1)^{m+n}e(s)e(t)
y_{s(m)}\wedge...\wedge y_{s(2)}\wedge x_{t(n)}\wedge...\wedge x_{t(2)}\wedge
[y_{s(1)},x_{t(1)}]\\
&=\textstyle\sum_{s\in Sh(1,m-1)}\sum_{t\in Sh(1,n-1)}
(-1)^{m^{2}+1}(-1)^{n^{2}+1}(-1)^{m+n}
e(s)e(t)y_{s(m)}\wedge...\wedge y_{s(2)}\wedge x_{t(n)}\wedge...\wedge x_{t(2)}\wedge
[y_{s(1)},x_{t(1)}]\\
&=\textstyle\sum_{s\in Sh(1,m-1)}\sum_{t\in Sh(1,n-1)}(-1)^{m^{2}+n^{2}+m+n}
e(s)e(t)y_{s(m)}\wedge...\wedge y_{s(2)}\wedge x_{t(n)}\wedge...\wedge x_{t(2)}\wedge
[y_{s(1)},x_{t(1)}]\\
&=\textstyle\sum_{s\in Sh(1,m-1)}\sum_{t\in Sh(1,n-1)}
(-1)^{m(m+1)+n(n+1)}e(s)e(t)
y_{s(m)}\wedge...\wedge y_{s(2)}\wedge x_{t(n)}\wedge...\wedge x_{t(2)}\wedge
[y_{s(1)},x_{t(1)}]\\
&=\textstyle\sum_{s\in Sh(1,m-1)}\sum_{t\in Sh(1,n-1)}e(s)e(t)
y_{s(m)}\wedge...\wedge y_{s(2)}\wedge x_{t(n)}\wedge...\wedge x_{t(2)}\wedge
[y_{s(1)},x_{t(1)}]
\end{align*}
\end{comment}
\begin{multline*}
\textstyle(-1)^{n+p}\frac{1}{j_{1}\cdots j_{p}}\sum_{s\in S_{n}}e(s)
\sum_{t\in Sh(2,p-2)}(-1)^{j_{t(1)}}
\sum_{q\in Sh(1,j_{t(1)}-1)}\sum_{r\in Sh(1,j_{t(2)}-1)}\\
e(q)e(r)\;x_{s(n)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(j_{t(1)}+j_{t(2)}+1)}
{\scriptstyle \;\wedge\;} x_{rs(j_{t(1)}+j_{t(2)})}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{rs(j_{t(1)}+2)}
{\scriptstyle \;\wedge\;}\\
{\scriptstyle \;\wedge\;} x_{qs(j_{t(1)})}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{qs(2)}{\scriptstyle \;\wedge\;}
[x_{rs(j_{t(1)}+1)},x_{qs(1)}]\;.
\end{multline*}
Now observe, that for any $s\in S_n$ and shuffle
$q \in Sh(1,j_{t(1)}-1)$, the permutation
$(qs(1),\ldots,qs(j_{t(1)}), s(j_{t(1)}+1),\ldots,s(n))$ is again an element of
$S_n$ and since there are precisely $j_{t(1)}$ many shuffles in
$Sh(1,j_{t(1)}-1)$ we can just
'absorb' the appropriate sum over shuffles in the previous expression into the sum over
general permutation. The same is true for the shuffles $r \in Sh(1,j_{t(2)}-1)$.
After reindexing we get:
\begin{multline*}
\textstyle(-1)^{n+p}\frac{1}{j_{1}\cdots j_{p}}\sum_{s\in S_{n}}e(s)
\sum_{t\in Sh(2,p-2)}(-1)^{j_{t(1)}}j_{t(1)}j_{t(2)}\cdot\\
\cdot x_{s(n)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(j_{t(1)}+2)}{\scriptstyle \;\wedge\;}
x_{s(j_{t(1)})}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(2)}{\scriptstyle \;\wedge\;}[x_{s(j_{t(1)}+1)},x_{s(1)}]=
\end{multline*}
$$
\textstyle(-1)^{n+p-1}\frac{1}{j_{1}\cdots j_{p}}
\sum_{1\leq l<m\leq p}j_{l}\;j_{m}\;\sum_{s\in S_{n}}e(s)
x_{s(n)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(3)}{\scriptstyle \;\wedge\;} [x_{s(2)},x_{s(1)}]
$$
Using this we are able to rewrite the right side of the defining structure equation
into
\begin{multline*}
\textstyle(-1)^{n+1}\sum_{p=2}^n\frac{(-1)^p}{p!}\sum_{j_1+\ldots+j_p=n}
\frac{1}{j_{1}\cdots j_{p}}
\sum_{1\leq l<m\leq p}\;j_{l}\;j_{m}\;\sum_{s\in S_{n}}e(s)\cdot\\
\cdot x_{s(n)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(3)}{\scriptstyle \;\wedge\;} [x_{s(2)},x_{s(1)}]
\end{multline*}
and in addition, the left side of the defining structure equation can be rewritten
as
\begin{multline*}
\textstyle(-1)^{n-2}\frac{(n-2)!}{2(n-2)!}\sum_{s\in S_n}e(s)
x_{s(n)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(3)}{\scriptstyle \;\wedge\;} [x_{s(1)},x_{s(2)}]=\\
\textstyle(-1)^{n+1}\frac{1}{2}\sum_{s\in S_n}e(s)
x_{s(n)}{\scriptstyle \;\wedge\;}\cdots{\scriptstyle \;\wedge\;} x_{s(3)}{\scriptstyle \;\wedge\;} [x_{s(2)},x_{s(1)}]\;.
\end{multline*}
Consequently the theorem follows since
\begin{equation}\label{combi_equation}
\textstyle\sum_{p=2}^n\frac{(-1)^p}{p!}\sum_{j_1+\ldots+j_p=n}
\frac{1}{j_{1}\cdots j_{p}}\sum_{1\leq l<m\leq p}j_{l}\cdot j_{m}=\frac{1}{2}
\end{equation}
for any $n\geq 2$.
(This identity was communicated by Gjergji Zaimi at mathoverflow) To see it
consider the generating function
$$
\textstyle\sum_{p\geq 2} \frac{(-1)^p}{p!}\binom{p}{2}\left(x+\frac{x^2}{2}+\frac{x^3}{3}+\cdots\right)^{p-2}\left(x^2+2x^3+3x^4+\cdots\right)\;.$$
The coefficient of $x^n$ is precisely the left side of (\ref{combi_equation})
and to show that it actually equals $\frac{1}{2}$ use
$$
\begin{array}{ccc}
\textstyle x^2+2x^3+3x^4+\cdots=\frac{x^2}{(1-x)^2}& and &
\frac{e^{-t}}{2}=\sum_{p\geq 2}\frac{(-1)^p}{p!}\binom{p}{2} t^{p-2}
\end{array}
$$
to simplify the generating function to
$$
\textstyle\frac{e^{ln(1-x)}}{2}\frac{x^2}{(1-x)^2}=
\frac{x^2}{2}+\frac{x^3}{2}+\frac{x^4}{2}+\frac{x^5}{2}+\cdots
$$
\end{proof}
\section{Conclusion and Outlook}
We defined a Lie $\infty$-structure on the exterior power $\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A$
of any Lie Rineard pair, but there is more structure on $\raisebox{0.3ex}{${\scriptstyle\bigwedge\,}$}\mathfrak{g}_A$.
In fact one should look at the interaction of the higher brackets with the
exterior product, to eventually come to some kind of $\infty$-Gerstenhaber structure.
Moreover one should look for generalizations of the Leibniz rule
to the higher brackets.
\begin{appendix}
\section{Lie $\infty$-algebras} We recall the most basic stuff
about Lie $\infty$-algebras. There are many incarnations of them
\cite{LV}, \cite{MM}, but we will only look at their
graded symmetric, 'many brackets' version, since that picture fits nicely into
the Schouten calculus and is moreover useful when it comes to actual computations.
Lie $\infty$-algebras are defined on $\mathbb{Z}$-graded vector spaces and
consequently we recall them first:
\subsection{Graded Vector Spaces}
In what follows $\mathbb{K}$ will always be a field and $\mathbb{Z}$ the Abelian
group of integers with respect to addition.
A $\mathbb{Z}$-graded $\mathbb{K}$-vector space $V$ is the direct sum
$\oplus_{n\in \mathbb{Z}} V_n$ of $\mathbb{K}$-vector spaces $V_n$.
Since this is a coprodut, there are natural injections
$i_n:V_n \to V$ and a vector is called \textbf{homogeneous of degree} $n$
if it is in the image of the injection $i_n$.
In that case we write $deg(v)$ or $|v|$ for its degree.
According to a better readable text we just write graded vector space
as a shortcut for $\mathbb{Z}$-graded $\mathbb{K}$-vector space.
A morphism $f : V \to W$ of graded vector spaces, homogeneous of degree $r$,
is a sequence of linear maps $f_n : V_n \to W_{n+r}$ for any $n\in \mathbb{Z}$ and
the integer $r\in\mathbb{Z}$ is called the degree of $f$, denoted by $deg(f)$ (or $|f|$).
For any $n\in\mathbb{N}$, an $n$-multilinear map $f: V_1 \times \cdots \times V_n \to W$,
homogeneous of degree $r$ is a sequence of $n$-multilinear maps
$f_{k} : (V_1)_{n_1} \times \ldots \times (V_k)_{n_k} \to W_{\sum n_i+r}$ for all $j_i\in \mathbb{Z}$ with $\sum j_i=k$.
The $\mathbb{Z}$-graded tensor product $V \otimes W$ of two graded vector spaces
$V$ and $W$ is given by
$$
\textstyle\left(V \otimes W \right)_n :=
\oplus_{i+j=n}\left( V_i \otimes W_j\right)
$$
and the Koszul commutativity constraint
$\tau: V \otimes W \to W \otimes V$ is on homogeneous elements
$v\otimes w \in V \otimes W$ defined
by
$$\tau(v \otimes w):=(-1)^{deg(v)deg(w)} w \otimes v$$ and then extended to
$V\otimes W$ by linearity.
\begin{remark}
We define the symbols $e(v):=(-1)^{deg(v)}$,
$e(v,w):=(-1)^{deg(v)deg(w)}$. The \textbf{Koszul sign}
$e(s;v_1,\ldots,v_k) \in \{-1,+1\}$ is defined for any permutation $s\in S_k$ and
any homogeneous vectors $v_1,\ldots,v_k\in V$ by
\begin{equation}\label{Koszul_convention}
v_1\otimes \ldots \otimes v_k= e(s;v_1,\ldots,v_k)
v_{s(1)}\otimes \ldots \otimes v_{s(k)}.
\end{equation}
In an actual computation it can be determined by the following rules:
When a permutation $s\in S_k$ is a transposition $j\leftrightarrow j+1$
of consecutive neighbors,
then $e(s;v_1,\ldots,v_k)= (-1)^{deg(v_j)\cdot deg(v_{+1})}$ and
if $t\in S_k$ is another permutation, then
$e(ts;v_1,\ldots,v_k)=e(t;v_{s(1)},\ldots,v_{s(k)})e(s;v_1,\ldots,v_k)$.
\end{remark}
A graded $k$-linear morphism $ f: \BIGOP{\times}^k V \to W$ is called
\textbf{graded symmetric} if
$$f(v_1,\ldots,v_k) = e(s;v_1,\ldots,v_k)f(v_{s(1)},\ldots,v_{s(k)})$$
for all $s\in S_k$.
\subsection{Shuffle Permutation}
Let $S_k$ be the symmetric group, i.e the group of all bijective maps
of the ordinal $\ordinal{k}$.
\begin{definition}[Shuffle Permutation] For any $p,q\in \mathbb{N}$
a $(p,q)$-shuffle is a permutation
$s\in S_{p+q}$ with $s(1)<\ldots<s(p)$ and $s(p+1)<\ldots<s(p+q)$.
We write $Sh(p,q)$ for the set of all $(p,q)$-shuffles.
More generally for any $p_1,\ldots,p_n\in\mathbb{N}$ a $(p_1,\ldots,p_n)$-shuffle
is a permutation $s\in S_{p_1+\cdots+p_n}$ with
$s(p_{j-1}+1)<\ldots<s(p_{j-1}+p_{j})$.
We write $Sh(p_1,\ldots,p_n)$ for the set of all $(p_1,\ldots,p_n)$-shuffles.
\end{definition}
For more on shuffles, see for example at \cite{RS}.
\subsection{Lie $\infty$-algebas} On the structure level Lie $\infty$-algebras generalize
(differential graded) Lie-algebras to a setting where the Jacobi identity isn't
satisfied any more, but holds up to particular higher brackets. This can be defined
in many different ways \cite{LV},
but the one that works best for us is its 'graded symmetric, many bracket' version.
\begin{definition} A \textbf{Lie $\infty$-algebra} $\left(V,(D_k)_{k\in \mathbb{N}}\right)$
is a $\mathbb{Z}$-graded $\mathbb{R}$-vector space $V$, together with a sequence $(D_k)_{k\in \mathbb{N}}$ of
graded symmetric, $k$-multilinear maps $D_k : \BIGOP{\times}^k V \to V$,
homogeneous of of degree $-1$, such that the \textbf{weak Jacobi equations}
$$\label{sh_Jacobi}
\textstyle\sum_{i+j=n+1} \left(\sum_{s\in Sh(j,n-j)}
e\left(s;v_1,\ldots,v_n\right)D_i\left(D_j \left(
v_{s_1}, \ldots, v_{s_j} \right), v_{s_{j+1}}, \ldots, v_{s_n}\right)\right) = 0
$$
are satisfied for any integer $n \in \mathbb{N}$ and any vectors $v_1,\ldots,v_n \in V$.
\end{definition}
In particular Lie $\infty$-algebras generalizes ordinary Lie algebras, if the grading is chosen
right:
\begin{example}[Lie Algebra] Every Lie algebra $\left(V,[\cdot,\cdot]\right)$ is a
Lie $\infty$-algebra if we consider $V$ as concentrated in degree one and define $D_k=0$ for any $k \neq 2$ as well as $D_2(\cdot,\cdot):=[\cdot,\cdot]$.
\end{example}
Very different from common Lie theory is, that a morphism
of Lie $\infty$-algebras is not necessarily just a single map. In fact such a
morphism is a \textit{sequence} of maps, satisfying a particular structure
equation. To understand how these morphisms emerge, look for example at \cite{MM}.
\begin{definition}For any two Lie $\infty$-algebras $(V,(D_k)_{k\in\mathbb{N}})$ and
$(W,(l_k)_{k\in\mathbb{N}})$ a \textbf{morphism of Lie $\infty$-algebras} is a sequence
$(f_k)_{k\in\mathbb{N}}$ of graded symmetric, $k$-multilinear maps
$$
f_k: V \times \cdots \times V \to W
$$
homogeneous of degree $0$, such that the structure equation
\begin{multline}\label{lie-infty-morph}
\textstyle\sum_{p+q=n+1}\left({\sum_{s\in Sh(q,p-1)}e\left(s\right)}
f_p\left(D_{q}\left(v_{s(1)},\ldots,v_{s(q)}\right),
v_{s(q+1)},\ldots, v_{s(n)}\right)\right)=\\
\textstyle\sum_p\frac{1}{p!}\sum^{k_1+\ldots +k_p =n}_{s\in Sh(k_1,\ldots,k_p)}
e\left(s\right)
l_p\left(f_{k_1}\left(v_{s(1)},\ldots,v_{s(k_1)}\right),
\ldots,
f_{k_p}\left((v_{s(n-k_p+1)},\ldots,v_{s(n)}\right)\right)
\end{multline}
is satisfied for any $n\in\mathbb{N}$ and any vectors $v_1,\ldots,v_n\in V$.
The morphism is called \textbf{strict},
if in addition $f_k=0$ for all $k \geq 2$,that is, if the morphism is a
single map, that commutes with all brackets.
\end{definition}
\end{appendix}
|
\section{Introduction}
Graphene is a well known allotrope of carbon in which the carbons bond in a planar $sp^2$ configuration\cite{wallace, novoselov2004}. As a result, the single graphene sheet is effectively two-dimensional\cite{neto2009,saito1998}. Of the four $n=2$ electrons which occupy the outer shell of a carbon atom, three are in $sp^2$ orbitals and form in-plane $\pi$ bonds between the nearest-neighbor carbon atoms, while the fourth occupies a $2p_z$ orbital. These $2p _z$ orbitals form a band of states which is responsible for many of the characteristic electronic properties of graphene\cite{neto2009}. Among these properties are a zero band gap at the so-called Dirac point, an electronic dispersion relation that, near the Dirac point, is equivalent to that of massless Dirac fermions, and spin-orbit coupling which is believed to be small because of the low atomic number of carbon\cite{neto2009,saito1998}. Graphene has a vast number of potential applications, including photo-voltaic cells\cite{wang2008}, ultracapacitors\cite{liu2011, stoller2008}, and spin-transport electronics\cite{yazyev2008,tombros,swartz}.
Recently, a number of researchers have carried out experimental and theoretical investigations into the effects of adatoms and impurities on both the band structure and localized magnetic moments in graphene. Among these are theoretical studies of carbon vacancies in graphene\cite{Lehtinen2004, Skrypnyk, Yazyev2007}, hydrogen atoms on the surface of graphene\cite{sofo2012}, and several other types of disorder in graphene\cite{pereira2008,wehling}. In several of these cases and in other work\cite{wehling,Yuan2010,Rakhmanov12}, impurity effects have been treated using a tight-binding model for the electronic structure of graphene and impurities, vacancies, or adatoms. These calculations have, however, either been carried out numerically, or in the limit of energies close to the Dirac point, where the graphene density of states can be approximated as linear\cite{wehling1}.
Density-functional calculations for adatoms on graphene have also been carried out. They have shown that the introduction of an adatom bonded to the surface of graphene can lead to a quasi-localized state with an energy near the Fermi energy and that the wave function of this quasi-localized state includes contributions from the orbitals of neighboring carbon atoms\cite{sofo2012, Yazyev2007}. In some cases it has been found that, even if the introduced defect or adatom is non-magnetic, a localized magnetic moment can form at the defect site\cite{Yazyev2007}.
In this paper, we extend the previous work on graphene with adatoms in two ways. First, we show that the tight-binding model for adatoms on graphene can, in the limit of low concentrations, be solved analytically in the absence of electron-electron interactions. Specifically, we obtain analytical expressions for the local density of states (LDOS) on the adatom, the total density of states (TDOS) of the adatom-graphene system, and the spectral function $A({\bf k}, E)$ for an electron with Bloch vector ${\bf k}$ and energy $E$ in graphene in the presence of the adatom\cite{Bena2009}. All these results are expressed as a function of the graphene density of states, which itself is known analytically for a nearest-neighbor tight binding model\cite{Yuan2010,Hobson1953}.
Secondly, we calculate the magnetic properties of the system using the Hubbard model for the electron-electron interaction, which we treat using a standard mean-field approximation. This treatment leads to a transition between a non-magnetic and magnetic state above a critical value of $U$ which depends on the parameters of the tight-binding model. In the presence of a finite $U$ our model is basically a special case of the well-known Anderson model\cite{anderson1961}, but with a linear rather than a constant density of states near the Fermi energy. Our results include not only the magnetic moment on the adatom, but also that on the graphene sheet and the charge transfer from the adatom to the sheet, all as functions of the model parameters.
The remainder of the article is arranged as follows: In Sec.\ II, we describe the model tight-binding Hamiltonian for the graphene-adatom system. We also describe its generalization to include electron-electron interaction on the adatom via a Hubbard $U$ term and the mean-field treatment of this term. In Sec.\ III, we describe the Green's function method used to analytically calculate the local density of states on the adatom, the total density of states on the graphene lattice in the presence of the adatom, and the spectral function. In Sec.\ IV, we present numerical results for the local densities of states, total density of states, and the spectral function for the graphene-adatom system. We also give the local densities of states in the presence of a finite Hubbard $U$ within mean-field theory, and give the magnetic moment induced by the adatom, as well as the charge transfer from the adatom to the graphene, as a function of the adatom parameters. In Sec.\ V, we give a concluding discussion.
\section{The Model Hamiltonian}\label{model}
Graphene is composed of two inter-penetrating triangular lattices, which we will label $\alpha$ and $\beta$, and thus two carbon atoms per primitive cell. In the present work, we are interested in a system consisting of a perfect lattice of graphene plus a single adatom, which we will assume has one atomic orbital. We also assume that the adatom lies at the so-called $T$ site, above one of the carbon atoms. It has been found, using {\it ab initio} electronic structure calculations, that a several species of adatoms, including hydrogen, fluorine, and gold, do occupy a location above one of the carbon atoms\cite{nakada,chan,ishii}.
\begin{figure}[ht]
\centering
\includegraphics[width=0.45\textwidth]{first_zone.eps}
\caption{(Color Online) Left: graphene crystal structure showing the two interpenetrating lattices labeled as $\alpha$ and $\beta$ and the set of nearest neighbor vectors ${\bf d}_1$, ${\bf d}_2$, and ${\bf d}_3$.
Right: first Brillouin zone for graphene showing the high symmetry points. This figure is a modified version of one shown in Ref.\ \cite{neto2009}. }
\label{fig:symmetry_points}
\end{figure}
We treat the graphene-adatom system using a tight-binding Hamiltonian that can, in principle, include hopping between any two carbon atoms. For pure graphene, the Hamiltonian can be written in terms of the creation and annihilation operators for electrons of spin $\sigma$ on a site in the $n^{th}$ primitive cell of the $\alpha$ and $\beta$ sublattices. We denote the creation (annihilation) operators for the $\alpha$ and $\beta$ sublattices by $a_{n\sigma}^\dag$ ($a_{n\sigma}$) and $b_{n\sigma}^\dag$ ($b_{n\sigma}$). The corresponding tight-binding Hamiltonian $H_0$ for graphene may be written in real space as
\begin{widetext}
\begin{equation}
H_0 = -\sum_{ n,\delta,\sigma}(t_{\alpha \beta,\delta}a_{n,\sigma}^\dag b_{n+\delta,\sigma} + h.\ c.\ ) -
\sum_{n,\delta\neq 0,\sigma}t_{\alpha\alpha,\delta}(a^\dag_{n,\sigma}a_{n+\delta,\sigma}+b^\dag_{n,\sigma} b_{n+\delta,\sigma}).
\label{eq:h0real}
\end{equation}
Here $t_{\alpha\beta,\delta}$ and $t_{\alpha\alpha,\delta}$ ($t_{\alpha\beta},t_{\alpha\alpha}>0$) are hopping integrals which take an electron of spin $\sigma$ ($\sigma = \pm 1/2$) from a lattice site to a neighboring lattice site, and $\delta$ denotes a Bravais lattice vector of the triangular lattice. The first sum represents hopping between sublattices and therefore all possible Bravais lattice vectors are summed over, whereas the second sum only includes hopping between sites on the same sublattice and thus, $\delta =0$ is not allowed. We have assumed that the hopping integrals for hopping on the same sublattice are identical for the $\alpha$ and $\beta$ sublattices and therefore set $t_{\alpha\alpha}=t_{\beta\beta}$.
Eq. (\ref{eq:h0real}) can be Fourier transformed as
\begin{eqnarray}\label{eq:h0kspace}
H_0 & = & \sum_{{\bf k},\sigma}H_{0,{\bf k},\sigma}; \nonumber \\
H_{0,{\bf k},\sigma} &=& -t_{\alpha\beta}({\bf k})a^\dag_{{\bf k},\sigma}b_{{\bf k},\sigma}-t^*_{\alpha\beta}({\bf k}) a_{{\bf k},\sigma}b^\dag_{{\bf k},\sigma}-t_{\alpha\alpha}({\bf k})[a^\dag_{{\bf k},\sigma}a_{{\bf k},\sigma} + b^\dag_{{\bf k},\sigma}b_{{\bf k},\sigma}],
\end{eqnarray}
\end{widetext}
where $t_{\alpha\beta}({\bf k})= \sum_{\bf \delta}e^{i{\bf k}\cdot{\bf \delta}}t_{\alpha\beta,{\bf \delta}}$ and
$t_{\alpha\alpha}({\bf k})=\sum_{{\bf \delta}\neq 0}e^{i{\bf k}\cdot{\bf \delta}}t_{\alpha\alpha,{\bf \delta}}$.
In the limit of only nearest neighbor hopping, $t_{\alpha\alpha,\delta} = 0$, and $t_{\alpha\beta,\delta} = 0$ except for the three nearest neighbors. In this limit, we write $t_{\alpha\beta}({\bf k}) = t({\bf k})$. The three nearest neighbor vectors are shown in Fig.\ \ref{fig:symmetry_points}, where they are denoted by ${\bf d}_1$, ${\bf d}_2$, and ${\bf d}_3$.
In this limit,
\begin{equation}\label{eq:CCham}
H_{0,{\bf k},\sigma} = -t({\bf k}) a^\dagger_{{\bf k},\sigma} b_{{\bf k,\sigma}}-t^*({\bf k})a_{{\bf k},\sigma}b^\dagger_{{\bf k},\sigma}.
\end{equation}
Here the operator $a_{{\bf k},\sigma}^\dag = \frac{1}{\sqrt{N}}\sum_n a_{n,\sigma}\exp(i{\bf k}\cdot{\bf \delta}_n)$, where $N$ is the number of primitive cells in the graphene lattice and ${\bf \delta}_n$ is the n$^{th}$ Bravais lattice vector of the triangular lattice; an analogous definition holds for $b_{{\bf k},\sigma}^\dag$. The sum over ${\bf k}$ is confined to the first Brillouin zone and $t({\bf k})$ is given by \cite{Rakhmanov12}
\begin{equation}\label{eq: tk_energy}
t({\bf k})=t\left[ 1+ 2\exp\left(\frac{3 i k_x a_0}{2}\right) cos\left(\frac{\sqrt{3} k_y a_0}{2}\right)\right].
\end{equation}
In Eq.\ (\ref{eq: tk_energy}), $t$ is the hopping energy between nearest neighbor carbon atoms ($t = 2.8 \ eV$ for graphene\cite{Rakhmanov12}), and $a_0= 1.42 \AA$ is the nearest-neighbor bond length\cite{Yazyev2007}.
\begin{figure}[ht]
\centering
\includegraphics[width=0.5\textwidth]{analytic_dos.eps}
\caption{(Color Online) Density of states per spin and per primitive cell $\rho_{0}(E)$ for the tight-binding model defined by Eqs.\ (\ref{eq:CCham}) and (\ref{eq: tk_energy}) for the $p_z$ orbitals of graphene, as calculated from (\ref{eq:analytic_dos})\cite{Hobson1953,Yuan2010}. Following Ref.\ \cite{Rakhmanov12}, we assume $t = 2.8 \ eV$ for graphene.}
\label{fig:analytic}
\end{figure}
We wish to investigate what happens to the density of states when when an isolated adatom is adsorbed onto the host graphene at a $T$ site\cite{nakada,chan}. The extra piece of the tight-binding Hamiltonian, $H_I$, due to the adatom may be written in real space as
\begin{equation}
H_I = \epsilon_0 \sum_\sigma h_{0,\sigma}^\dag h_{0,\sigma} - t^\prime\sum_\sigma \left(h_{0,\sigma}^\dag a_{0,\sigma} + h_{0,\sigma}a_{0,\sigma}^\dag\right).
\label{eq:hireal}
\end{equation}
Here $h_{0,\sigma}^\dag$ and $h_{0,\sigma}$ are creation and annihilation operators for an electron of spin $\sigma$ ($\sigma = \pm 1/2$) at the site of the adatom, which we assume is located at the site $0$ of the $\alpha$ sublattice, $\epsilon_0$ is the on-site energy of an electron on that site (relative to the Dirac point of the pure graphene band structure), and $t^\prime$, ($t^\prime>0$), is the energy for an electron to hop between the adatom and the carbon atom at the site $0$ of the $\alpha$ sub-lattice.
We now wish to express $H_I$ in terms of Bloch eigenstates of $H_0$. These eigenstates may be written as two-component column vectors with components $\psi_1({\bf k})$ and $\psi_2({\bf k})$ satisfying the eigenvalue equation (hereafter we suppress the spin subscript until needed)
\begin{equation}
\begin{pmatrix}
\epsilon_{\bf k}+t_{\alpha\alpha}({\bf k}) & t_{\alpha\beta}({\bf k}) \\ t^*_{\alpha\beta}({\bf k}) & \epsilon_{\bf k}+t_{\alpha\alpha}({\bf k})
\end{pmatrix}
\begin{pmatrix}
\psi_{1,{\bf k}} \\ \psi_{2,{\bf k}}
\end{pmatrix}
=0.
\vspace{0.0 in}
\end{equation}
The solution to this eigenvalue problem is
\begin{equation}\label{eq:energy_graphene}
\epsilon_{\bf k} = -t_{\alpha\alpha}({\bf k})\pm|t_{\alpha\beta}({\bf k})|,
\end{equation}
which gives the tight-binding band structure of pure graphene\cite{Rakhmanov12}, and the corresponding eigenvectors satisfy
\begin{equation}
\psi_{1,{\bf k}} = \mp e^{-i\phi_{\bf k}}\psi_{2,{\bf k}},
\end{equation}
where the phase factor $e^{-i\phi_{\bf k}}$ is given by
\begin{equation}\label{eqn:phase}
e^{-i\phi_{\bf k}} = \frac{t_{\alpha\beta}({\bf k})}{|t_{\alpha\beta}({\bf k})|}.
\end{equation}
We can then write the destruction operator for a Bloch electron in the upper band as
\begin{equation}
\gamma_{{\bf k},1} = \frac{1}{\sqrt{2}}\left(e^{i\phi_{\bf k}}a_{\bf k} + b_{\bf k}\right),
\end{equation}
and in the lower band as
\begin{equation}
\gamma_{{\bf k},2} = \frac{1}{\sqrt{2}}\left(e^{i\phi_{\bf k}}a_{\bf k} - b_{\bf k}\right),
\end{equation}
where we have defined $\gamma_{{\bf k},1}$ and $\gamma_{{\bf k},2}$ to be properly normalized, so that, for example, the anticommutator
$\{\gamma_{{\bf k},1}^{\hspace{1pt }} , \gamma^\dag_{{\bf k},1}\} = 1$.
With these definitions, we can now use the inverse Fourier transform of the $\gamma_{{\bf k},1}$ and $\gamma_{{\bf k},2}$ to obtain
\begin{equation}
a_n = \frac{1}{\sqrt{2N}}\sum_{\bf k}e^{-i{\bf k}\cdot{\bf \delta}_n}e^{-i\phi_{\bf k}}(\gamma_{{\bf k},1} + \gamma_{{\bf k},2}).
\end{equation}
Thus, we can rewrite the impurity Hamiltonian (\ref{eq:hireal}) in momentum space as
\begin{widetext}
\begin{equation}
H_I = \epsilon_0 h_0^\dag h_0 - \frac{t^\prime}{\sqrt{2N}} \left[ h_0^\dag \sum_{\bf k}e^{-i\phi_{\bf k}}(\gamma_{{\bf k},1} +\gamma_{{\bf k},2}) + h.c.\right].
\label{himp}
\end{equation}
Thus, in $H_I$, the creation and annihilation operators of the adatom are connected to every eigenstate of the graphene band structure by matrix elements of equal magnitude (though different phase). For a hydrogen adatom, we take $\epsilon_0 = 0.4 \ eV$ and $t' = 5.8
\ eV$, as found in \cite{Rakhmanov12}. The one-electron Hamiltonian, $H_0 + H_I$, is a special case of the Anderson impurity model\cite{anderson1961}, where the impurity state is coupled to all the band electron states by matrix elements of equal magnitude.
We also include in our calculation the effects of an on-site electron-electron interaction of the Hubbard form,
\begin{equation}\label{hubbard_u}
H_U= U n_{0\uparrow} n_{0,\downarrow},
\end{equation}
where $n_{0,\sigma} = h_{0,\sigma} ^\dag h_{0,\sigma}$ is the number of electrons with spin $\sigma$ on the hydrogen site. For a hydrogen adatom we take $U$ to be the difference between the ionization potential and the electron infinity providing us with a numberical value of $U \sim 12.85 eV = 4.59t$\cite{pariser,lykk}.
The Hubbard term given in Eq. (\ref{hubbard_u}) is quartic in the creation and annihilation operators. Therefore, in order to calculate the properties of the Hamiltonian including this term, we use a standard mean field theory to rewrite this term (see, e.\ g., Ref.\ \cite{Fazekas1999}) in the form
\begin{equation}\label{MFT_U}
H_U \sim U \left[h_{0\uparrow}^\dagger h_{0\uparrow} \langle n_{0\downarrow}\rangle+h_{0\downarrow}^\dagger h_{0\downarrow} \langle n_{0\uparrow}\rangle- \langle n_{0\uparrow}\rangle \langle n_{0\downarrow}\rangle\right].
\end{equation}
\end{widetext}
With this approximation, the total Hamiltonian, consisting of the sum of Eqs.\ (\ref{eq:CCham}), (\ref{himp}), and (\ref{MFT_U}), becomes quadratic in electron creation and annihilation operators, and can be diagonalized. The Fermi energy, total energy, and magnetic properties of the system can then be obtained by an iterative process as described below. The electronic density of states corresponding to the one-electron Hamiltonian, $H_0+H_I$, can be obtained analytically, as we describe below, which makes the calculation of the total energy and the magnetic properties quite simple.
\section{Green's function, Density of States, and Spectral Function of Graphene-Adatom System}
\subsection{Green's Function}
We use the single particle Green's function approach to calculate the local and total density of states of the graphene-adatom system, initially omitting the Hubbard-U term. We continue to suppress the spin degree of freedom since, in the absence of the Hubbard term, spin just gives an extra factor of 2. To that end, we first introduce the resolvent operator
\begin{equation}
G(z) = \frac{1}{z - H},
\label{eq:resolvent}
\end{equation}
where $z = E + i\eta,$ ($\eta \rightarrow 0^+$), and $H = H_0 + H_I$. If there are $2N$ carbon atoms and 1 adatom, $G(z)$ can be expressed as an $(2N+1)\times (2N+1)$-dimensional matrix. It is convenient to use the 2N Bloch states (corresponding to $N {\bf k}$ values) created by the operators $\gamma_{{\bf k},1}^\dag$ and $\gamma_{{\bf k},2}^\dag$ as the basis for this matrix, plus the adatom orbital corresponding to $h_0^\dag$. If we let the adatom orbital correspond to the first of the $(2N+1)$ states, then one can easily write out the matrix $z-H$ of which $G(z)$ is the inverse.
\subsection{Density of States}
We denote the local electronic density of states per spin on the adatom site by $\rho_{00}(E)$. $\rho_{00}(E)$ is related to $G(z)$ by
\begin{equation}
\rho_{00}(E) = -\frac{1}{\pi}\mathrm{Im} G_{00}(z) = -\frac{1}{\pi}\mathrm{Im}\langle 0|\frac{1}{z - H}|0\rangle.
\label{eq:ldos}
\end{equation}
Here $z = E +i\eta$, ($\eta \rightarrow 0^+$), and $\langle 0|1/(z-H)|0\rangle$ denotes the matrix element of $1/(z-H)$ evaluated at the location of the adatom, which we take to be above the atom $0$ on the $\alpha$ sub-lattice. This corresponds to an element in the first row and first column of the matrix $(z-H)^{-1}$. We can obtain this matrix element as
\begin{equation}
G_{00}(z) = \frac{\mathrm{cof}_{00}(z-H)}{\mathrm{det}(z-H)},
\label{eq:cofdet}
\end{equation}
where cof$_{00}(z-H)$ denotes the cofactor of the element in the first row and first column of the matrix $z-H$, and the denominator is the determinant of $z-H$. Both quantities are readily evaluated, and the result for $\rho_{00}(E)$ is
\begin{equation}\label{eq:local_dos}
\rho_{00}(E) = -\frac{1}{\pi}\mathrm{Im}\left(\frac{1}{z - \epsilon_0 - \frac{t^{\prime 2}}{2N}{\cal G}_0(z)}\right) _{z=E+i0+},
\end{equation}
where
\begin{equation}\label{eq:script_G}
{\cal G}_0(z) =\sum_{{\bf k},\lambda=1}^2\left(\frac{1}{z-\epsilon_{{\bf k},\lambda}}\right) \equiv \mathrm{Tr}\left(\frac{1}{z - H_0}\right)
\end{equation}
and $\epsilon_{{\bf k}, 2} = -\epsilon_{{\bf k},1}$ is given by Eq.\ (\ref{eq:energy_graphene}).
$\mathrm{Im} \ {\cal G}_0(z)$ is related to the (unperturbed) graphene density of states per graphene unit cell (per spin), which we denote $\rho_{0}(E)$, by
\begin{equation}\label{eq:dos_analytic}
-\frac{1}{\pi}\mathrm{Im}\ {\cal G}_0(E+i\eta)=N \rho_{0}(E),
\end{equation}
with $\eta \rightarrow 0^+$ \cite{anderson1961,wehling}. For the form of $H_0$ which includes only nearest neighbor hopping, $\rho_0(E)$ is\cite{Hobson1953,Yuan2010}
\begin{numcases}{\rho_{0}(E)=}\label{eq:analytic_dos}
\frac{2 E}{t^2 \pi^2} \frac{1}{\sqrt{f(x)}}K\left( \frac{4x }{f(x)}\right) & $0<x<1$ \\
\frac{2 E}{t^2 \pi^2}\frac{1}{\sqrt{4x}}K\left(\frac{f(x)}{4x}\right) &$1<x<3$, \nonumber
\end{numcases}
where $x= E/t$, $f(x) = (1+x)^2 - \frac{(x^2-1)^2}{4}$ and $K(m)$ is the elliptic integral of the first kind. We plot Eq. (\ref{eq:analytic_dos}) in Fig.\ (\ref{fig:analytic}) normalized such that $\int_{-3t}^0\rho_0(E)dE = 1$.
$\mathrm{Re} \ {\cal G}_0(E)$ is related to Eq. (\ref{eq:analytic_dos}) via the principal value integral\cite{anderson1961,wehling}
\begin{equation}\label{eq:principle_real}
\mathrm{Re} \ {\cal G}_0(E) = N \ \mathrm{P}\left(\int_{-3t}^{3t}\frac{\rho_0(E')}{E-E'}dE'\right),
\end{equation}
where the integral runs over the range where $\rho_0(E^\prime) \neq 0$. The density of states on the carbon sites (per spin) in the presence of an adatom may be written as $\rho_g(E) = -\frac{1}{\pi}\mathrm{Im}\sum_{{\bf k},\lambda}\langle {\bf k},\lambda|\frac{1}{z - H}|{\bf k},\lambda\rangle$, where $z = E +i\eta$, ($\eta\rightarrow 0^+$). Each of the elements of this sum can be computed using the analog of Eq.\ (\ref{eq:cofdet}), with the result
\begin{equation}\label{eq:change_carbon}
\rho_{g}(E) = N\rho_{0}(E) +\frac{1}{\pi}\mathrm{Im}\left(\frac{t^{\prime 2}/2N}{z-\epsilon_0-\frac{t^{\prime 2}}{2N}{\cal G}_0(z)}\frac{d{\cal G}_0}{dz}\right)_{z=E+i0^+}.
\end{equation}
\begin{widetext}
The {\it total} density of states per spin is is the sum of the expressions in Eqs.\ (\ref{eq:local_dos}) and (\ref{eq:change_carbon}), and can be rearranged to have the form
\begin{equation}\label{eq:total_dos}
\rho_{tot}(E) = \rho_g(E) + \rho_{00}(E) = N\rho_{0}(E) -\frac{1}{\pi}\mathrm{Im}\left( \frac{d}{dz}\mathrm{ln}[z-\epsilon_0-\frac{t^{\prime 2}}{2N}{\cal G}_0(z)]\right)_{z=E+i0^+}.
\end{equation}
\subsection{Spectral Function}
We can use an analogous approach to calculate the spectral function $A({\bf k}, E)$. $A({\bf k}, E)$ represents the probability density that an electron with Bloch wave-vector ${\bf k}$ has energy $E$, and is given by
\begin{equation}
A({\bf k}, E) = -\frac{1}{\pi}\mathrm{Im}\sum_\lambda\langle {\bf k}, \lambda|\left(\frac{1}{z-H}\right)|{\bf k}, \lambda \rangle.
\end{equation}
These matrix elements can be evaluated using the methods of the preceding section, with the result
\begin{equation}
A({\bf k}, E) = -\frac{1}{\pi}\mathrm{Im}\sum_{\lambda=1}^2\left[\frac{1}{z-\epsilon_{{\bf k},\lambda}} +
\frac{t^{\prime 2}}{2N}\left(\frac{1}{z-\epsilon_0 - \frac{t^{\prime 2}}{2N}{\cal G}_0(z)}\right)\frac{1}{(z-\epsilon_{{\bf k},\lambda})^2}\right]_{z=E+i0^+}.
\label{eq:spectral}
\end{equation}
It is convenient to write this spectral function in terms of a self-energy $\Sigma_\lambda({\bf k}, E)$ as
\begin{equation}
A({\bf k}, E) = -\frac{1}{\pi}\mathrm{Im}\left(\sum_{\lambda=1}^2\frac{1}{z - \epsilon_{{\bf k},\lambda} - \Sigma_\lambda({\bf k}, z)}\right)_{z=E+i0^+},
\label{eq:selfe}
\end{equation}
where $\Sigma_\lambda({\bf k}, E)$ is readily computed by equating Eqs. (\ref{eq:spectral}) and (\ref{eq:selfe}). To first order in $1/N$, $\Sigma_\lambda({\bf k}, z)$ is found to be independent of both $\lambda$ and ${\bf k}$ and to take the form
\begin{equation}
\Sigma_\lambda({\bf k}, z) = \frac{t^{\prime 2}}{2N}\left(\frac{1}{z-\epsilon_0-\frac{t^{\prime 2}}{2N}{\cal G}_0(z)}\right)_{z=E+i0^+}.
\label{eq:selfe1}
\end{equation}
All these equations [(\ref{eq:local_dos}), (\ref{eq:script_G}), (\ref{eq:dos_analytic}), (\ref{eq:principle_real})-(\ref{eq:total_dos}), (\ref{eq:selfe}), and (\ref{eq:selfe1})] remain valid for non-nearest-neighbor hopping; only the form of the graphene density of states has to be changed.
\end{widetext}
\subsection{Effects of Electron-Electron Interaction; Spin-Polarized Density of States and Magnetic Moment}\label{sectionD}
Finally, we discuss the effects of including a non-zero Hubbard term $H_U$ [Eq.\ (\ref{hubbard_u})] in the Hamiltonian. If we treat $H_U$ by mean-field theory [Eq.\ (\ref{MFT_U})], then the densities of states for spin-up and spin-down electrons may be different. We can calculate these partial densities of states self-consistently as follows. First, we make an initial assumption for the value of $\langle n_{0\uparrow}\rangle$ and $\langle n_{0\downarrow}\rangle$. Then the effective on- site energy for an up-spin electron on the hydrogen adatom is obtained by making the replacement
\begin{equation}\label{eq:includespin}
\epsilon_{0,\uparrow} \rightarrow \epsilon_0 + U\langle n_\downarrow\rangle ,
\end{equation}
with a corresponding expression for $\epsilon_{0,\downarrow}$. Given $\epsilon_{0,\uparrow}$ and $\epsilon_{0,\downarrow}$, we can compute the local densities of states $\rho_{00,\uparrow}$ and $\rho_{00,\downarrow}$ using the appropriate generalizations of Eq.\ (\ref{eq:local_dos}); we can also obtain the total densities of states $\rho_{tot,\uparrow}$ and $\rho_{tot,\downarrow}$ using the corresponding generalizations of Eq. (\ref{eq:total_dos}). The Fermi energy, $E_F$, is then obtained from the condition
\begin{equation}\label{number_dos}
2N+1= \int_{-3t}^{E_F}\rho_{tot}(E) dE,
\end{equation}
where we assume one adatom, $2N$ carbon sites, and $\rho_{tot}(E) =\rho_{tot,\uparrow}(E) + \rho_{tot,\downarrow}(E)$. Given $E_F$, we then recalculate $\langle n_{0,\uparrow}\rangle$ and $\langle n_{0,\downarrow}\rangle$. The procedure is repeated until successive iterations do not lead to a significant change in $\langle n_{0,\uparrow}\rangle$ and $\langle n_{0,\downarrow}\rangle$. In practice, we require that these quantities change by no more than $\pm 0.001 n_a$
on successive iterations (here $n_a$ is the number of adatoms in the calculation). Typically, about twenty iterations of the self-consistent equations are needed to attain this degree of convergence, as discussed further below.
Once $E_F$ has been found, the total magnetic moment $\mu_T$ of the system is obtained from
\begin{equation}
\label{eq:magmom}
\frac{\mu_T}{\mu_B} = \int_{-3t}^{E_F}\left[\rho_{tot,\uparrow}(E)-\rho_{tot,\downarrow}(E)\right]dE,
\end{equation}
where $\mu_B$ is the Bohr magneton.
In the limit $U \rightarrow \infty$, the mean-field version of the Hubbard model can be done without iteration. In this limit, only one of the quantities $\langle n_\uparrow\rangle$ or $\langle n_\downarrow \rangle$ is non-zero. The reason is that if, say, $\langle n_\uparrow\rangle$ is non-zero, then the energy to put a spin-down electron on the adatom becomes infinite, and hence the number of spin-down electrons must be zero. To be definite, we assume that $\langle n_\downarrow \rangle = 0$. In that case, we just have $\epsilon_{0,\uparrow} = \epsilon_0$, and $\epsilon_{0,\downarrow} \rightarrow \infty$. The total density of states for the up spins will then be given by Eq. (\ref{eq:total_dos}), while that for the down spins is just that of unperturbed graphene: $\rho_{tot,\downarrow} = N\rho_0(E)$.
In the limit $U\rightarrow \infty$ the Fermi energy, $E_F$, is obtained from Eq. (\ref{number_dos}) and may be simplified to
\begin{widetext}
\begin{equation}\label{eq:extra_adatom}
\int_0^{E_F} 2N\rho_0(E)dE -
\frac{1}{\pi}\mathrm{Im}\ln\left[\frac{E_F -\epsilon_0-\frac{t^{\prime2}}{2N}{\cal G}_0(E_F)}{-3t-\epsilon_0-\frac{t^{\prime2}}{2N}{\cal G}_0(-3t)}\right] = 1.
\end{equation}
\end{widetext}
Once $E_F$ has been obtained, the magnetic moment $\mu_T$ can be again found using Eq.\ (\ref{eq:magmom}). Since both $\rho_\uparrow(E)$ and $\rho_\downarrow(E)$ are available analytically, using Eqs. (\ref{eq:magmom}) and (\ref{eq:extra_adatom}), $\mu_T$ is easily computed in closed form.
\section{Numerical Results}
In Fig.\ \ref{fig:ldos_hydrogen}(a), we plot the local density of states $\rho_{00}(E)$ for parameters appropriate to a hydrogen adatom on graphene with $U = 0$, as calculated from Eq.\ (\ref{eq:local_dos}). We use the parameters $t = 2.8 \ eV$, $t^\prime = 5.8 \ eV$, and $\epsilon_0 = 0.4\ eV$, as given by Ref.\ \cite{Rakhmanov12} for a hydrogen adatom. In Fig.\ \ref{fig:ldos_hydrogen}(b), we plot the {\it change} in the total density of states produced by a single hydrogen atom, i.\ e., the quantity $\rho_{tot}(E) - N\rho_0(E)$ for the three cases of Fig.\ \ref{fig:ldos_hydrogen}(a), calculated using Eqs. (\ref{eq:ldos}) and (\ref{eq:total_dos}).
Next, we calculate both the spectral function $A({\bf k}, E)$ for $U = 0$ and the spin polarized spectral function $A_\sigma({\bf k}, E)$ for $U = \infty$ as functions of $E$ for several values of ${\bf k}$, using the parameters of Fig.\ \ref{fig:analytic}. $A_\sigma({\bf k}, E)$ is obtained using a generalization of Eq.\ (\ref{eq:selfe}) in the limit $U \rightarrow \infty$ as discussed in subsection \ref{sectionD}. The resulting spectral functions are shown in Fig.\ \ref{fig:spectral} through first order in $1/N$. The contribution from the adatom appears as the sharp spike near $E_F=0.173 \ eV$, while the contribution from the graphene sheet is shown as broadened peaks near the values of $\epsilon_{{\bf k},i}$ $(i = 1, 2)$ for the three choices of ${\bf k}$. The self energy term [Eq.\ (\ref{eq:selfe1})] controls the width of the graphene resonances. The integral of the graphene sheet's contribution to the spectral function will be of order $N$ times larger then that of the adatom. Furthermore, the width of the graphene peaks in the spectral function is proportional to the density of adatoms.
\begin{widetext}
\begin{figure}[h t]
\centering
\includegraphics[width =0.9 \textwidth]{Uzeroldostdos.eps}
\caption{(Color Online) (a). Local density of states per spin on the adatom site, $\rho_{00}$(E), with $U=0$ for a graphene sample with N = 500 graphene unit cells (1000 C atoms) and one adatom. We assume the model described in the text [Eqs.\ (\ref{eq:CCham}), (\ref{eq: tk_energy}), and (\ref{himp})]. Black curve: $t^\prime = 5.8 \ eV$, $\epsilon_0 = 0.4 \ eV$;
blue curve: t$^\prime = 1.0 \ eV$, $\epsilon_0 = 0.4 \ eV$;
and red curve: t$^\prime = 1.0 \ eV$, $\epsilon_0 = -0.4 \ eV$. In all three cases, $t = 2.8 \ eV$. The Fermi energy is calculated using Eq.\ (\ref{number_dos}) and gives $E_F = 0.236 \ eV$.
(b). The change in the density of states per spin due to the adatom for the three cases shown in (a). In both (a) and (b), the insets are enlargements of the region between -1 eV and +1 eV.}
\label{fig:ldos_hydrogen}
\end{figure}
\end{widetext}
\begin{figure}
\centering
\includegraphics[width = 0.5 \textwidth]{spec_function_line.eps}
\caption{(Color Online) (a)-(c). Spectral function $A({\bf k}, E)$ at $U=0$ for the graphene-adatom system, as calculated using Eqs.\ (\ref{eq:selfe}) and (\ref{eq:selfe1}) for three values of ${\bf k}$ corresponding to $M$, $K$, and $\Gamma$ respectively, assuming a nearest-neighbor tight-binding band. We use $t=2.8 \ eV$ , $t'=5.8 \ eV$, $\epsilon_0 = 0.4 \ eV$, and $N = 500$.
(d). The spin polarized spectral function for the point $M$ and the case $U \rightarrow \infty$ in the mean-field approximation, using the same parameters as in (a)-(c). $A_{\sigma}({\bf k }, E)$ for the majority spin component is shown in black and that of the minority component in red. For the minority spin component, $A_{\sigma}({\bf k}, E)$ is just that of the unperturbed graphene sheet, i.\ e., delta functions at the unperturbed pure graphene energy eigenvalues for all values of ${\bf k}$. For the case $U = 0$, and for the majority spin at $U = \infty$, the adatom contribution occurs near $E_F$, while the graphene sheet contribution corresponds to broadened peaks near the unperturbed energy eigenvalues $\epsilon_{\bf k}$ given in Eq. (\ref{eq:energy_graphene}). For points $\Gamma$ and $M$, the integral of the peak near $E \sim 0$ is of order 1/(2N) times those of the main peaks in the spectral function, and thus vanishes as $N \rightarrow \infty$. }
\label{fig:spectral}
\end{figure}
\begin{widetext}
\begin{figure}[h t]
\centering
\includegraphics[width=1.0\textwidth]{parameter_vary.eps}
\caption{(Color Online) Spin-polarized local density of states (sLDOS) per adatom for both the majority spin (black solid line) and minority spin (red solid line), as obtained by substituting Eq.\ (\ref{eq:includespin}) into Eq.\ (\ref{eq:local_dos}). We use $t = 2.8 \ eV$ and the following
values of the on-site energy $\epsilon_0$, hopping energy $t^\prime$, and Hubbard energy $U$:
(a). $(t',\epsilon_0,U) = (5.8,0.4,4.59 t)$; (b). $(5.8,0.4,10t)$; (c)$(5.8,0.4,\infty)$ (d). $(5.8,0.0,\infty)$; (e). $(5.8,-0.4,\infty)$; (f). $(5.8,-1.0,\infty)$; (g). $(1.8,0.4,\infty)$; (h). $(7.8,0.4,\infty)$.}
\label{fig:changeparameter}
\end{figure}
\end{widetext}
\begin{figure}[hbt]
\centering
\includegraphics[width=0.5\textwidth]{magmom.eps}
\caption{Total magnetic moment on the adatom and the graphene ($\mu_T$), in units of $\mu_B$, versus $U$ for two values of $t^\prime/t$ (Red curve: $t^\prime = 1.0t$, Black curve $t^\prime = 2.0t$). In this plot, we hold $\epsilon_0 = 0.4 eV $ and $t= 2.8 eV$ as we vary $U$ for two different choices of $t^\prime$. In both cases, the moment seems to become nonzero at a characteristic value of $U$, which depends on $t^\prime$, $t$, and $\epsilon_0$. Lines connect calculated points.}
\label{fig:magmom4}
\end{figure}
Using the mean-field methods described in section \ref{model} we can calculate a variety of other spin-independent and spin-dependent properties of the adatom-graphene system. These include $\rho_{00,\uparrow}(E)$ and $\rho_{00,\downarrow}(E)$, the local density of states of up and down spin on the adatom; the induced magnetic moment on the adatom ($\mu_a$) and in the entire system of graphene sheet plus adatom ($\mu_T$); and the net charge transfer from the adatom to the sheet, all as functions of the parameters $U$, $\epsilon_0$, and $t^\prime$. The magnetic moment on the adatom site is $\mu_a = (\langle n_\uparrow\rangle - \langle n_\downarrow\rangle)\mu_B$. $\langle n_\sigma\rangle$ is obtained from
\begin{equation}
\langle n_\sigma \rangle = \int_{-\infty}^{E_F} \rho_{00,\sigma}(E) dE,
\end{equation}
where $\rho_{00,\sigma}$ ($\sigma = \uparrow or \downarrow$) is defined using the appropriate generalization of Eq.\ (\ref{eq:local_dos}). The total magnetic moment is given in Eq. (\ref{eq:magmom}). The net charge transfer from adatom to the graphene lattice is obtained by first integrating $\rho_{00,\uparrow}+\rho_{00,\downarrow}$ up to the Fermi energy, to obtain the net number of electrons on the adatom, then subtracting this quantity from the adatom valence $Z$ (i.\ e., for hydrogen, Z = 1) to obtain the net charge transfer.
We have carried out these calculations for various values of the adatom on-site energy $\epsilon_0$, Hubbard parameter $U$, and hopping energy $t^\prime$. In Table I we summarize the results above for parameters appropriate to a hydrogen adatom and summarize the trends when the various adatom parameters are varied. Additional results are shown in Figs.\ \ref{fig:ldos_hydrogen} and \ref{fig:changeparameter}. As can be seen in Table I and Fig.\ \ref{fig:changeparameter}, the parameter values thought to be appropriate to an $H$ adatom ($U = 4.59t$, $t^\prime = 5.8 \ eV$, and $\epsilon_0= 0.4 \ eV$), lead to a very small magnetic moment on the adatom though there is an increase in both the LDOS and TDOS close to the Fermi energy ($E_F = 0.173 \ eV$; see Fig.\ \ref{fig:changeparameter}(b)).
\begin{table}[h t]
\begin{center}
\begin{tabular}{| c | c | c| c|c|c| c |}
\hline
\multicolumn{7}{ | c| } {Summary of Numerical Results} \\
\hline
U & $\epsilon_0$ & $t^\prime$& $\mu_a$ ($\mu_B$)& $\mu_T $ & Charge & $E_F$\\
& (eV)&(eV) & per & ($\mu_B$) & Transfer& (eV) \\
& & & adatom & & ($|e|$) & \\
\hline\hline
$0.0t ^a $ & $ 0.4 $ & 5.8 & $0.0$ & 0.0 & $ 0.372$ &0.372\\
\hline
$ 4.59 t^b $ & $0.4 $& 5.8 & 2.67E-4 & 1.13E-3 & 0.695& 0.173\\
\hline
$ 10.0 t$& 0.4& 5.8& 3.11E-3&7.67E-3 & 0.758 &0.236 \\
\hline
$ \infty $ & $0.4 $ & $5.8$ & 0.300 &0.871 & 0.699 &0.111\\
\hline
$\infty $ & $ 0.0 $ & 5.8 & 0.260 & 0.927 & 0.738 & 0.050 \\
\hline
$\infty$ & $-0.4$ & 5.8& 0.338 &0.958 &0.662 &0.050 \\
\hline
$\infty$ & $-1.0 $& 5.8 &0.360 &0.990 & 0.639 &-0.01 \\
\hline
$\infty$ & 0.4 &1.8 &0.358 & 0.506 &0.641 & 0.236 \\
\hline
$\infty$& 0.4& 7.8& 0.219& -0.11 &0.780 &0.236 \\
\hline
\end{tabular}
\caption{Magnetic moment on the adatom ($\mu_a$), total magnetic moment on the graphene-adatom system ($\mu_T$) (both in units of $\mu_B$), and charge transferred from the adatom to the graphene lattice (in units of an electron charge), for various choices of $U$ , on-site energy $\epsilon_0$, and hopping energy $t^\prime $. Note that $U=4.59 t$, $\epsilon_0=0.4 \ eV$, and $t^\prime=5.8 \ eV$ corresponds to the expected parameters of a hydrogen adatom. When $U \rightarrow \infty$, we find a spin polarized state near the Fermi energy. The magnetic moment calculated on the adatom is done using a combination of Eqs. (\ref{MFT_U}) and (\ref{eq:local_dos}) and the magnetic moment on the sheet is calculated using Eq. (\ref{eq:magmom}).
\\ (a)- Spin polarized LDOS plotted in Fig. (\ref{fig:ldos_hydrogen}). \\(b)- Spin polarized LDOS plotted in Fig. (\ref{fig:changeparameter}b).}
\end{center}
\end{table}
In general, as seen in Table I, for sufficiently large $U$, a nonzero magnetic moment develops on both the adatom and the graphene sheet. The moment on the adatom is of order $0.3 \ \mu_B$ in this limit, for the given parameters, while the sum of the moments on the adatom and the sheet approaches $\mu_B$ in this limit. We also find in all of our calculations that a large fraction (typically $0.6-0.7$ of the electron) is transferred from the adatom to the graphene sheet for the parameters we consider.
For the parameters appropriate to hydrogen adatoms, the model predicts no, or only a very small, induced magnetic moment. A possible explanation is that our model assumes no lattice distortion due to the adatom. But DFT calculations have shown that the surface of graphene is warped due to the addition of an adatom. This warping could change the distance between the adatom and the neighboring carbon atoms, and hence possibly the value of the Coulomb integral.
In Fig.\ \ref{fig:magmom4}, we show the total magnetic moment of the system as given by Eq.\ (\ref{eq:magmom}) plotted as a function of $U$, for various values of $t^\prime$. In each case shown, $\epsilon_0 = 0.4 \ eV$ and $t = 2.8 \ eV$. In all the plots, there is an apparent threshold behavior: the moment becomes nonzero only if $U$ exceeds a threshold value which depends on $t^\prime$ as well as on $U$. While these calculations are done using a simple mean-field approximation, they seem to be consistent with other work on related models\cite{anderson1961,fradkin}.
\section{Discussion}
Using a tight-binding model we have calculated the local and total density of states and the spectral function for a system consisting of a single adatom in a $T$ site on graphene. Because the hopping integral from the adatom to a graphene Bloch state has the same magnitude for any ${\bf k}$, we have shown that these quantities can be calculated analytically. This simplification holds even if we do not make the oft-used linear approximation\cite{wehling} for the graphene density of states near the Dirac point. It is also valid even if we include non-nearest-neighbor hopping in the tight-binding graphene Hamiltonian. Since our numerical results give both the local and total density of states, we can compute the charge transfer from the adatom to the graphene. Our numerical results show that, for most parameters we consider, this charge transfer is a substantial fraction of an electron (approximately $70 \%$ for parameters appropriate to hydrogen).
Because the calculations are at low adatom concentration, the adatom-induced density of states is linear in concentration. Other work has treated the same system at higher adatom concentration, but only numerically\cite{Rakhmanov12}. In future work, it might be possible to treat the present model analytically at higher concentrations, at least approximately. It would also be of interest to include effects of lattice distortions, which are known to exist when adatoms bind to graphene \cite{boukhvalov}, and which can lead to a large increase in spin-orbit interactions \cite{neto,Gmitra}. Such spin-orbit interactions would likely have a large effect on the magnetic properties arising from the adatom.
We have also calculated the magnetic properties induced by the adatom, using a Hubbard model treated within mean-field theory. For all choices of the Hamiltonian parameters, we find that there is a critical value of the Hubbard $U$ above which the density of states near the Fermi energy is spin-polarized and a net induced magnetic moment is formed. The appearance of this magnetic moment was predicted long ago to occur within mean-field theory for models with a slowly
varying density of states near the Fermi energy\cite{anderson1961}. Here, it is also found to occur in a system with a roughly linear density of states near $E_F$. While the mean-field approximation is probably unreliable for this model, we note that a similar threshold for moment formation was also found, within the Kondo Hamiltonian, for a system with a linear density of states\cite{fradkin}. Since, at large $U$, it is known that the Hubbard model can be approximately transformed into the Kondo model\cite{schrieffer}, it seems plausible that there could really be a threshold behavior in the Hubbard model with a linear density of state such as is found here, even though we use a mean field theory to obtain it.
A somewhat counterintuitive result of our calculations is that, as $t^\prime$ increases, the value of $U$ needed to induce a magnetic moment becomes smaller. Since a larger $t^\prime$ suggests that it is easier for the electron to hop from the impurity to the graphene, one might expect that a moment on the impurity atom would be less likely to form. A possible explanation is that the larger $t^\prime$ also causes the peak in the impurity density of states to shift closer to the Dirac point, where the graphene density of states is smaller. Thus, there are fewer final states available for an electron to hop into, and hence, the electron is less likely to hop, thus increasing the likelihood of moment formation on the impurity.
In our approximation we also calculate the spectral function of our graphene-adatom system to first order in $1/N$. The main effect of the adatom is, as expected, simply to broaden the delta-function peaks that the spectral function would exhibit in an ideal graphene lattice. In our approach, this broadening, and the shape of the spectral line, are computed analytically. To the same order, we find that both the adatom contribution to the spectral function at $E\sim0$ and the broadening of the graphene spectral lines will vanish as $N\rightarrow \infty$.
In summary, we have, using a single-particle Green's function approach together with a tight binding Hamiltonian in the limit of no electron-electron correlations, obtained analytical equations for the LDOS, TDOS, and spectral function for adatoms on the surface of graphene. Using the same model with a finite Hubbard energy $U$, we find that a magnetic moment is induced both on the adatom and nearby on the graphene sheet above a critical value of $U$ which depends on the other model parameters. These results are not only of intrinsic interest but also may be useful in understanding the behavior of a variety of adatoms on graphene.
\section{Acknowledgments}
This work was supported by the Center for Emerging Materials at The Ohio State University, an NSF MRSEC (Grant No.\ DMR0820414) . We thank Prof.\ Roland Kawakami for valuable discussions.
|
\section{Introduction}
We consider the following surface growth equation for
the height $u(t,x)\in\mathbb{R}$ at time $t>0$
over a point $x\in[0,2\pi]$
\begin{align}
u_t = - u_{xxxx} - ({u_x}^2)_{xx}
\qquad x \in [0,2\pi], \; t \in [0,T] \label{eqn:sfg}
\end{align}
with periodic boundary conditions and subject to a moving frame,
which yields the zero-average condition
$ \int_0^{2\pi}u(x,t)\;\mathrm{d}x = 0$.
This equation, usually with additional noise terms,
was introduced as a phenomenological model for
the growth of amorphous surfaces \cite{Si-Pl:94b,Ra-Li-Ha:00a},
and was also used
to describe sputtering processes \cite{Cu-Va-Ga:05};
see \cite{blomkerromito12} for a detailed list of references.
Based on the papers
\cite{robinsonetal07, dashtirobinson08, robinsonrubiosadowski13}
which develop the theory of `numerical verification of regularity'
for the 3D Navier--Stokes equations,
our aim here is to establish and implement numerical algorithms
to prove rigorously global existence and uniqueness
of solutions of (\ref{eqn:sfg}).
Despite being scalar the
equation has surprising similarities to 3D Navier--Stokes equations
\cite{blomkerflandoliromito09, blomkerromito09, blomkerromito12}.
It allows for a global energy estimate in $L^2$ and uniqueness
of smooth local solutions for initial conditions in a critical
Besov-type space that contains $C^0$ and $H^{1/2}$,
see \cite{blomkerromito12}
(similar results for the 3D Navier--Stokes equations can
be found in \cite{KocTat01}).
Here we focus on the one-dimensional model, since in this case
more efficient numerical methods are available, and the calculations
would be significantly slower in higher dimension.
Moreover, for the two-dimensional case the situation of
energy estimates seems even worse,
as global existence could only be established in $H^{-1}$ using the
non-standard energy $\int_0^{2\pi} {\rm e}^{u(x)}\,{\rm d} x$,
see \cite{MW:11} for details. Nevertheless, we believe that it should
be possible to treat the 2D case using similar methods,
but the analysis becomes more delicate since in two dimensions $H^1$
is the critical space (see \cite{blomkerromito09, blomkerromito12}).
Rigorous methods for proving numerically the existence of
solutions for PDEs are a recent and active research field.
In addition to the approach taken here there are methods based
on topological arguments
like the Conley index, see
\cite{paapeetal08, daylessardmischaikov07, zgliczynski10}, for example.
For solutions of elliptic PDEs there are methods using Brouwer's fixed-point
theorem, as discussed in the review article \cite{Pl:08} and
the references therein.
Our approach is based on \cite{robinsonetal07}
and similar to the method proposed in \cite{morosipizzocchero12}.
The key point is the derivation of a scalar ODE for
the $H^1$-norm of the difference of an arbitrary approximation,
that satisfies the boundary conditions, to the
solution. The coefficients of this ODE
depend only on the numerical data (or any other approximation used).
As long as the solution of the ODE stays finite, one can
rely on the continuation property of unique local solutions,
and thus have a smooth unique solution up to a blowup time of the ODE.
A similar approach using an integral equation
based on the mild formulation was proposed in
\cite{morosipizzocchero08, morosipizzocchero11}.
In order to establish a bound on the blow-up time for the ODE,
one can either proceed analytically or numerically. We propose
two analytical methods: one, based on the standard Gronwall Lemma,
enforces a `small data' hypothesis and adds little to standard
analytical existence proofs. The second is based on an explicit
analytical upper bound to the ODE solution. A variant of this, a
hybrid method in which one applies an analytical upper bound on a
succession of small intervals of length $h>0$ to the numerical
solution and then restarts the argument, appears the most promising,
and a formal calculation indicates that the upper bound from the
third method in the limit of step-size to zero
converges to the solution of the ODE.
In order to derive the ODE for the $H^1$-error,
we use standard a-priori estimates. While the stability of
the linear term $-u_{xxxx}$ means that these `worst case'
estimates are still sufficient, an interesting alternative
approach in a slightly different context is proposed in
\cite{nakaohashimoto09, nakaokinoshitakinura12}, where
the spectrum of the linearized operator
(here $Lv=-v_{xxxx} + (v_x\varphi_x)_{xx}$, where $\varphi$
is some given numerical data) is analysed
with a rigorous numerical method, which in the case of an
unstable linear operator yields substantially better results,
at the price of a significantly higher computational time.
This will be the subject of future research.
The paper is organized as follows.
In Section \ref{sec:apa} we establish the a-priori
estimates for the $H^1$-error between solutions
and the numerical data, which in the end gives an ODE depending on the
numerical data only.
Section \ref{sec:ode} provides the ODE estimates necessary for our
three methods, while Section \ref{sec:res} states the main results.
In the final Section \ref{sec:num}, we compare our methods using
numerical experiments.
\section{A-priori analysis}
\label{sec:apa}
In this section we establish upper bounds for the $H^1$-norm of the error
\[ d(x,t) := u(x,t) - \varphi(x,t),\]
where $u$ is a solution to our surface growth equation (\ref{eqn:sfg})
and $\varphi$ is any arbitrary, but sufficiently smooth approximation,
that satisfies the boundary conditions.
Since we know $\varphi$, if we can control the $H^1$ norm of
$d$ then we control the $H^1$ norm of $u$.
For the following estimates and results, we define the
$H^p$-norm, $p \geq 1$,
of a function $u$ by
\[
\| u \|_{H^p} := \left\| \partial_x^p u \right\|_{L^2},
\]
which is equivalent to the standard $H^p$-norm as we only consider functions
with vanishing mean, i.e.
$\int_0^{2\pi}u(x,t)\;\mathrm{d}x = 0$.
Note that in this setting Agmon's inequality
\begin{equation}\label{Agmon}
\|u\|_{L^\infty}\le\|u\|_{L^2}^{1/2}\|u_x\|_{L^2}^{1/2}
\end{equation}
holds with optimal constant $1$, which follows using the Fourier expansion.
A very important property of the
surface growth equation (\ref{eqn:sfg}) is the existence of
local solutions, which are smooth in space and time.
This result is given by the following theorem from \cite{blomkerromito09}
(Theorem 3.1).
\begin{theorem}\label{thm:regularity}
Let $u_0 \in H^1$, then there exists a time $\tau(u_0) > 0$ such that
there is a unique solution $u \in C^0([0,\tau(u_0)),H^1)$
satisfying
\begin{enumerate}
\item[1)] if $\tau(u_0) < \infty$, then
$\limsup\limits_{t\rightarrow\infty} \|u(t)\|_{H^1} = \infty$.
\item[2)] $u$ is $C^\infty$ in both, space and time, for all
$(t,x) \in (0,\tau(u_0)) \times [0,2\pi]$.
\end{enumerate}
\end{theorem}
Note that the theorem implies that lack of blowup in $H^1$ is sufficient
to ensure that the solution exists for all time and is smooth.
In particular, all of the manipulations we make in what follows are
valid until the blowup time.
Throughout the rest of the paper we consider the solutions with
initial data in $H^1$ whose existence is guaranteed by
Theorem \ref{thm:regularity}, and approximations
$\varphi \in H^4_{\text{per}}$ in space and $H^1$ in time.
\subsection{Energy estimate}
\label{sec:Energyestimate}
In this section we prove the key estimate
(\ref{eqn:upperBound}) on which the theorems of the following sections are
based.
If we use the surface growth equation (\ref{eqn:sfg}) to find
the evolution of $d(x,t)$ and defining the residual of
the approximation $\varphi$ by
\[ \RES := \varphi_t + \varphi_{xxxx}
+ ({\varphi_x}^2)_{xx} ,
\]
then we have
\[
d_t = - d_{xxxx} - ({u_x}^2)_{xx}
+ ({\varphi_x}^2)_{xx} - \RES.
\]
By replacing $u$ with $d+\varphi$ we obtain
\[
d_t
= - d_{xxxx} - ({d_x}^2)_{xx}
- 2 (d_x \varphi_x)_{xx}
- \RES.
\]
For the $H^1$-norm we have
\begin{align*}
\frac{1}{2} \partial_t \|d_x\|^2 = &
\underbrace{\left\langle d_{xx} , d_{xxxx}
\right\rangle}_{\mathrm{A}}
+ \underbrace{2\<d_{xx},\left(d_x\varphi_x\right)_{xx}\>
}_{\mathrm{B}}
+ \underbrace{\<d_{xx},({d_x}^2)_{xx}\>}_{\mathrm{C}}
+ \underbrace{\<d_{xx},\RES\> }_{\mathrm{D}},
\end{align*}
where $\langle \cdot , \cdot \rangle$ is the $L^2$ scalar product.
Now consider these terms separately. Integrating by parts
we obtain $\mathrm{A}=- \left\|d_{xxx} \right\|_{L^2}^2$.
Secondly,
\begin{align*}
\mathrm{B} = -2 \int_0^{2\pi} d_{xxx}(d_x\varphi_x)_x\,\mathrm{d}x
= \;\int_0^{2\pi} \left(d_{xx}\right)^2
\varphi_{xx} \; \mathrm{d}x
- 2 \int_0^{2\pi} d_{xxx}d_x\varphi_{xx}\; \mathrm{d}x \;
\end{align*}
and so
\begin{align*}
|\mathrm{B}| \leq & \; \left\|d_{xx}\right\|_{L^2}^2
\left\|\varphi_{xx} \right\|_{L^{\infty}}
+ 2 \left\|d_{xxx} \right\|_{L^2} \left\|d_x\right\|_{L^2}
\left\|\varphi_{xx}\right\|_{L^\infty} \\
\leq & \; 3 \left\|d_{xxx} \right\|_{L^2} \left\|d_x\right\|_{L^2}
\left\|\varphi_{xx}\right\|_{L^\infty} \\
\leq & \; \frac{1}{4} \left\| d_{xxx}\right\|_{L^2}^2
+ 9 \left\| d_x \right\|_{L^2}^2 \left\|
\varphi_{xx} \right\|_{L^\infty}^2,
\end{align*}
using interpolation and Young's inequality.
For $C$ we have
\begin{align*}
\mathrm{C} = - \int_0^{2\pi} ({d_x}^2)_xd_{xxx}
\; \mathrm{d}x
= -2 \int_0^{2\pi} d_xd_{xx}d_{xxx}
\; \mathrm{d}x,
\end{align*}
hence using Agmon's inequality (\ref{Agmon}), interpolation,
and Young's inequality,
\begin{align*}
|\mathrm{C}| \leq & \; 2 \left\|d_x\right\|_{L^2}
\left\|d_{xx} \right\|_{L^\infty}
\left\|d_{xxx} \right\|_{L^2}\\
\leq &\; 2 \left\|d_x \right\|_{L^2}
\left\|d_{xx} \right\|_{L^2}^{\frac{1}{2}}
\left\|d_{xxx} \right\|_{L^2}^{\frac{3}{2}}\\
\leq & \; 2 \left\|d_x \right\|_{L^2}^{\frac{5}{4}}
\left\|d_{xxx} \right\|_{L^2}^{\frac{7}{4}}\\
\leq & \; \frac{1}{4} \left\|d_{xxx} \right\|_{L^2}^2
+ K \left\|d_x \right\|_{L^2}^{10},
\end{align*}
where $K=7^7/4$; and for the remaining term
\begin{align*}
|\mathrm{D}| \leq \; \left\| \RES \right\|_{H^{-1}}
\left\| d_{xxx} \right\|_{L^2}
\le \frac{1}{4} \left\| d_{xxx}
\right\|_{L^2}^2 + \left\| \RES
\right\|_{H^{-1}}^2.
\end{align*}
Combining these estimates
and applying Poincaré inequality with the optimal constant
$\omega = 1$, we obtain
\begin{align} \label{eqn:upperBound}
\frac{1}{2} \partial_t \|d\|_{H^1}^2 &\leq - \frac{1}{4}
\left\| d\right\|_{H^3}^2 + K
\left\|d_x \right\|_{L^2}^{10} +
\left\| \RES \right\|_{H^{-1}}^2 + 9 \left\| d \right\|_{H^1}^2
\left\| \varphi_{xx} \right\|_{L^\infty}^2 \nonumber \\
& \leq K \left\|d \right\|_{H^1}^{10} +
\Big( 9 \left\| \varphi_{xx} \right\|_{L^\infty}^2 - \frac{1}{4}
\Big) \left\|d \right\|_{H^1}^2 + \left\| \RES \right\|_{H^{-1}}^2
\end{align}
which is a scalar differential inequality of type
\begin{equation}
\dot\xi \leq b\xi^5 +\left(a(t)-c\right)\xi + f(t),
\label{eqn:estType}
\end{equation}
and by standard ODE comparison principles
a solution of the equality in (\ref{eqn:estType})
provides an upper bound for $\|d \|_{H^1}^2$.
\subsection{Time and smallness conditions}
We need two important properties of the surface growth model,
which we will prove now. These are for equations like
Navier–Stokes well known facts, namely:
that smallness of the solution implies global
uniqueness and that solutions are actually small after some time by
energy-type estimates. These results go back to Leray (\cite{leray34}),
more modern discussions can be found in \cite{constantinfoias88} (Theorem 9.3)
and in a setting that parallels the treatment here in
\cite{robinsonsadowski08}.
For our model similar results for the critical $H^{1/2}$-norm
can be found in \cite{blomkerromito09}. But for our results,
we need to derive the precise values of constants
in the $H^{1}$-norm, which were not determined before.
First, if the $H^1$-norm of a solution $u$ is smaller than some
constant $\varepsilon_0$,
we have global regularity of $u$.
\begin{theorem}[Smallness Condition]\label{thm:smallness}
If for some $t \in [0,T]$ one has that $\| u(t) \|_{H^1}$
is finite on $[0,t]$ and
\[
\| u(t) \|_{H^1} < \frac{1}{2} =: \varepsilon_0,
\]
then we have global regularity (and thus uniqueness) of
the solution $u$ on $[0,\infty)$.
\end{theorem}
\begin{proof}
This is established by almost
the same estimates derived for the parts (A) and (C) in Section
\ref{sec:Energyestimate} and Young's inequality with constant $\delta >0$.
To be more precise:
\begin{align*}
\frac{1}{2} \partial_t \|u\|_{H^1}^2
&= - \|u_{xxx} \|_{L^2}^2 + \int_0^{2\pi}
u_{xx}({u_x}^2)_{xx}
\; \mathrm{d}x\\
&\leq - \| u \|_{H^3}^2 + 2 \| u \|_{H^3}^\frac{7}{4}
\| u \|_{H^1}^\frac{5}{4} \\
&\leq - \| u \|_{H^3}^2 + 2 \cdot
\Big( \delta \| u \|_{H^3}^2 +
\left( \frac{8}{7} \delta \right)^{-7} \frac{1}{8}
\| u \|_{H^1}^{10} \Big)\\
&\leq - \| u \|_{H^3}^2 \Big( 1 - 2\delta -
\left( \frac{8}{7} \delta \right)^{-7} \cdot \frac{1}{4}
\| u \|_{H^1}^{8} \Big).
\end{align*}
If $1 - 2\delta - \left( \frac{8}{7} \delta \right)^{-7}
\cdot \frac{1}{4} \| u\|_{H^1}^{8} > 0$, then we obtain a global
bound on $\| u \|_{H^1}^2$. The optimal choice for the constant
from Young inequality is $\delta = \frac{7}{16}$ and with this
value it follows, that if $\| u(t) \|_{H^1} < \frac{1}{2}$ we have a negative
derivative and the norm decays over time and is therefore bounded.
\end{proof}
The second property is that, based on the smallness condition, we can
determine a time $T^*$, only depending on the initial value $u(0)$,
such that $\| u(T^*) \|_{H^1} < \varepsilon_0.$
\begin{theorem}[Time Condition]\label{thmtc}
If a solution u is regular up to time
\[
T^*(u(0)) := \frac{1}{\varepsilon_0^2} \| u(0) \|_{L^2}^2
= 4 \left\| u(0) \right\|_{L^2}^2,
\]
then we have global regularity of the solution $u$.
\end{theorem}
At the risk of labouring the point, we only need to verify regularity of
a solution starting at $u(0)$ up to time $T^*(u(0))$, and
from that point on regularity is automatic.
\begin{proof}
As an a-priori estimate we have
\[
\partial_t \| u \|_{L^2}^2
= - \left\|u_{xx} \right\|_{L^2}^2
\]
and thus
\[
\int_0^T \left\|u_x(s) \right\|_{L^2}^2 \; \mathrm{d}s
\leq \int_0^T \left\| u_{xx}(s) \right\|_{L^2}^2 \; \mathrm{d}s
\leq \left\| u(0) \right\|_{L^2}^2
\]
where we used the Poincaré inequality with constant $\omega=1$.
If we now assume that
$\left\| u_x(s) \right\|_{L^2} > \varepsilon_0$ for all
$s \in [0,T]$, then
\[
T \varepsilon_0^2 < \left\| u(0) \right\|_{L^2}^2
\quad \text{or}\quad
T < \frac{1}{\varepsilon_0^2} \left\| u(0) \right\|_{L^2}^2
\]
This means, that if we wait until time
$T^* := \frac{1}{\varepsilon_0^2} \left\| u(0) \right\|_{L^2}^2$,
we know that $\left\| u(t) \right\|_{H^1} \leq \varepsilon_0$
for at least one $t \in [0,T^*]$
and we have global regularity by the smallness condition,
if there was no blowup before time $T^*$.
\end{proof}
\section{ODE estimates}
\label{sec:ode}
We present several methods to bound solutions of
ODEs of the type (\ref{eqn:upperBound}).
In this section we give the results for the scalar ODE,
and present applications in the next section.
Let us first state a lemma of Gronwall type, based on comparison
principles for ODEs, for which we will only give the idea of a proof.
\begin{lemma}[Gronwall]
\label{lem:Gronwall}
Let $a,b \in L^1([0,T],\mathbb{R})$ and
$x \in W^{1,1}([0,T],\mathbb{R})\cap C^0([0,T],\mathbb{R})$ such that
\[
\dot x \leq a(t)x + b(t) \qquad \forall t \in [0,T] .
\]
Then for all $t \in [0,T]$
\[
x(t) \leq \exp \Big( \int_0^t a(s) \; \mathrm{d}s \Big) x(0)
+ \int_0^t \exp \Big( \int_s^t a(r) \; \mathrm{d}r \Big)
b(s) \; \mathrm{d}s
\;.
\]
\end{lemma}
\begin{proof}[Idea of Proof]
Consider the function
\[
u(t)=x(t)\exp\{- \int_0^t a( s)ds\}
\quad \text{with} \quad
u'( t) \leq b( t) \exp\{- \int_0^t a(s)ds\}.
\]
Integrating and solving for $x$ yields the result.
\end{proof}
\begin{lemma}\label{lem:cp}
Consider two functions
$x,u \in W^{1,1}([0,T],\mathbb{R}^+_0) \cap C^0([0,T],\mathbb{R}^+_0)$
such that
\[ \dot x \leq c(t)x^p + e(t) \qquad x(0) = x_0
\]
with $p > 1$, $c \in L^1([0,T],\mathbb{R}^+_0)$ and
$e \in L^1([0,T],\mathbb{R}^+_0)$, and let $u$ be the solution of
\[
\dot u = c(t)u^p \qquad u(0) = x_0 + \int_0^T e(s)
\; \mathrm{d}s .
\]
Then $x(t) \leq u(t)$ for all $t \in [0,T]$.
\end{lemma}
\begin{proof}
First note that if $e \equiv 0 $ on $[0,T]$ then by using the
standard comparison principle it follows that
$u(t) \geq x(t)$ for all $t \in [0,T]$.
So now we assume that $\int_0^T e(s) \; \mathrm{d}s > 0$. For a
contradiction, suppose that there exists a time $t^* \in [0,T]$ such
that $t^*:= \inf \left\{t > 0:x(t) = u(t) \right\}$.
Because of the continuity of $u(t)$ and $x(t)$, and $u(0) > x(0)$ due
to our initial assumption $\int_0^T e(s) \; \mathrm{d}s > 0$,
it follows that $t^* > 0$. From the definition $u(t) > x(t)$ for
all $t \in [0,t^*)$, and thus
\begin{align*}
0 = u(t^*) - x(t^*) & \geq u(0) - x(0)
- \int_0^{t^*} e(s) \; \mathrm{d}s +
\int_0^{t^*} c(s)\left( u(s)^p - x(s)^p \right) \; \mathrm{d}s\\
& = \int_{t^*}^T e(s) \; \mathrm{d}s
+ \int_0^{t^*} c(s)( u(s)^p - x(s)^p)
\; \mathrm{d}s,
\end{align*}
which is strictly positive provided that
$\int_0^{t^*} c(s) \; \mathrm{d}s >0$.
If $\int_0^{t^*} c(s) \; \mathrm{d}s = 0$, then as $c\ge0$ we obtain
\[
x(t) \leq x(0) + \int_0^{t} e(s) \; \mathrm{d}s
\leq x(0) + \int_0^{T} e(s) \; \mathrm{d}s = u(t)
\quad \forall t \in [0,t^*],
\]
and we can repeat the above argument on the interval
$[t^*,T]$ to obtain a contradiction.\end{proof}
\begin{theorem}[CP-Type I] \label{thm:cpType}
Assume $x \in W^{1,1}([0,T],\mathbb{R}^+_0)
\cap C^0([0,T],\mathbb{R}^+_0)$ such that
\[ \dot x \leq c(t)x^p + e(t), \qquad x(0) = x_0 \]
with $p > 1$, $c \in L^1([0,T],\mathbb{R}^+_0)$ and
$e \in L^1([0,T],\mathbb{R}^+_0)$. Then for all $t\in [0,T]$,
as long as the right-hand side is finite,
\[
x(t) \leq \Big( x_0 + \int_0^t e(s) \; \mathrm{d}s \Big)
\Big\{ 1 -(p-1)\Big[ x_0 + \int_0^t e(s) \;
\mathrm{d}s \Big]^{p-1} \!\!\!\int_0^t c(s) \;
\mathrm{d}s \Big\}^{-\frac{1}{p-1}}\;.
\]
\end{theorem}
\begin{proof}
Given the setting of Lemma \ref{lem:cp}, we can solve for $u(t)$.
As $\mathrm{d}u = c(t) u^p \; \mathrm{d}t$,
a straightforward calculation shows that
\[
u(t) = u(0)\Big( 1 -(p-1) u(0)^{p-1}
\int_0^t c(s) \; \mathrm{d}s \Big)^{-\frac{1}{p-1}}
\]
as long as the right-hand side is finite.
Thus for all $t \in [0,T]$, as long as the right-hand side is finite,
\begin{align*}
x(t) &\leq \Big( x_0 + \int_0^T e(s) \; \mathrm{d}s \Big) \\
& \; \times \Big\{ 1 -(p-1)\Big[ x_0 + \int_0^T e(s) \; \mathrm{d}s
\Big]^{p-1} \!\!\!\int_0^t c(s) \; \mathrm{d}s \Big\}^{-\frac{1}{p-1}}
\end{align*}
This holds particularly when $T=t$.
\end{proof}
We now extend this result to differential inequalities of the form
\[ \dot x \leq b(t)x^p + a(t)x + f(t), \]
where $p>1$, $f,b \in L^1([0,T],\mathbb{R}^+_0)$ and
$a \in L^1([0,T],\mathbb{R})$, as our inequality
(\ref{eqn:upperBound}) is of this type.
\begin{corollary}[CP-Type II] \label{cor:cpType}
Assume $x \in W^{1,1}([0,T],\mathbb{R}^+_0) \cap C^0([0,T],\mathbb{R}^+_0)$
such that
\[ \dot x \leq b(t)x^p + a(t)x + f(t), \]
with $p>1$, $b,f \in L^1([0,T],\mathbb{R}^+_0)$ and
$a \in L^1([0,T],\mathbb{R})$. Then for all $t\in [0,T]$, as
long as the right-hand side is finite,
\begin{align*}
x(t) \leq \; {\rm e}^{A(t)} & \Big( x_0 + \int_0^t
\tilde{f}(s) \; \mathrm{d}s \Big) \\
&\times\Big\{ 1 - (p-1)\cdot \Big[ x_0 + \int_0^t \tilde{f}(s)
\; \mathrm{d}s \Big]^{p-1} \int_0^t \tilde{b}(s) \;
\mathrm{d}s \Big\}^{-\frac{1}{p-1}}
\end{align*}
where
\[
\tilde{b}(t) = \; b(t) {\rm e}^{(p-1) A(t)}, \quad
\tilde{f}(t) = \; {\rm e}^{-A(t)} f(t) ,
\quad\text{and}\quad A(t) = \int_0^t a(s) \; \mathrm{d}s \;.
\]
\end{corollary}
\begin{proof}
Consider the substitution $y(t) = {\rm e}^{- A(t)} x(t)$ with
$ A(t) = \int_0^t a(s) \; \mathrm{d}s $. It follows that
\begin{align*}
\dot y &= - a(t)y + {\rm e}^{- A(t)}\dot x \\
& \leq - a(t)y + {\rm e}^{- A(t)} \left( b(t)x^p + a(t)x + f(t) \right) \\
&= \underbrace{b(t) {\rm e}^{(p-1) A(t)}}_{\tilde{b}(t)}\,
y^p + \underbrace{{\rm e}^{-A(t)} f(t)}_{\tilde{f}(t)}
\end{align*}
with $\tilde{b}(t) \geq 0$ and $\tilde{f}(t) \geq 0$ for all $t \in [0,T]$.
Here we can apply Theorem \ref{thm:cpType} and obtain
\[
y(t) \leq \Big( y_0 + \int_0^t \tilde{f} \;\mathrm{d}s \Big)
\Big\{ 1 -(p-1)\Big[ y_0 + \int_0^t \tilde{f} \;
\mathrm{d}s \Big]^{p-1} \!\!\! \int_0^t \tilde{b} \;
\mathrm{d}s \Big\}^{-\frac{1}{p-1}}.
\]
Now substitute back with $x(t) = {\rm e}^{A(t)} y(t)$.
\end{proof}
\section{Verification methods}
\label{sec:res}
We now outline three techniques for numerical verification.
All of them are based on the key estimate (\ref{eqn:upperBound})
for the difference $d$ between a smooth approximation
$\varphi$ and a smooth local solution. The first is additionally
based on the simple Gronwall Lemma \ref{lem:Gronwall}, the second on
Corollary \ref{cor:cpType}, and the third is similar to
the second method, but restarts the
estimation after a series of short time-steps.
\subsection{First method}
This is based directly on the simple Gronwall Lemma \ref{lem:Gronwall}.
Assuming a poor bound to control the nonlinearity,
we prove a better error estimate.
\begin{theorem}\label{thm:method1}
Let $K^*=(8K)^{-1/8}=(2\times 7^7)^{-1/8}$. As long as
\begin{equation}\label{cond1}
\| d(0)\|_{H^1}^2 {\rm e}^{A(t)}
+ 2 \int_0^t \| \RES(s) \|_{H^{-1}}^2 {\rm e}^{ ( A(t) - A(s) )}
\; \mathrm{d}s \leq K^*,
\end{equation}
we have
\[
\|d(t)\|_{H^1}^2 \leq\| d(0)\|_{H^1}^2 {\rm e}^{A(t)}
+ 2 \int_0^t \| \RES(s) \|_{H^{-1}}^2
{\rm e}^{ ( A(t) - A(s) )} \; \mathrm{d}s,
\]
where $A(t) = - \frac{1}{4}t + 18 \int_0^t
\left\| \varphi_{xx}(\tau) \right\|_{L^\infty}^2
\mathrm{d}\tau.$
\end{theorem}
Note that the condition in (\ref{cond1}) involves only the numerical
solution $\varphi$.
\begin{proof}
It follows from the inequality (\ref{eqn:upperBound})
that as long as $\left\|d\right\|_{H^1}^8 \leq (8K)^{-1}$
we obtain
\begin{align*}
\partial_t \|d\|_{H^1}^2 \leq & - \frac{1}{4} \| d
\|_{H^1}^2 + 2 \| \RES \|_{H^{-1}}^2 + 18 \left\| d \right\|_{H^1}^2
\left\| \varphi_{xx} \right\|_{L^\infty}^2.
\end{align*}
Now we can apply Lemma \ref{lem:Gronwall} to deduce that
\begin{align*}
\|d(t)\|_{H^1}^2 \leq & \;\|d(0)\|_{H^1}^2 \exp
\Big\{ - \frac{t}{4} + 18 \int_0^t \|
\varphi_{xx}(\tau)\|_{L^\infty}^2 \mathrm{d}\tau \Big\} \\
& + 2 \int_0^t \| \RES(s) \|_{H^{-1}}^2 \exp
\Big\{ - \frac{t-s}{4} + 18 \int_s^t \| \varphi_{xx}(\tau)
\|_{L^\infty}^2 \; \mathrm{d}\tau \Big\} \mathrm{d}s.
\end{align*}
\end{proof}
Please note that if the bound exceeds $K^*$,
Theorem \ref{thm:method1} makes no assertions on $\|d(t)\|_{H^1}^2$.
\subsection{Second method}
Here we present the more sophisticated method based on direct
application of Corollary \ref{cor:cpType} (CP-Type II).
\begin{theorem}\label{thm:method2}
As long as the right-hand side is finite, the following
inequality holds for $d(t)$:
\begin{align*}
\|d(t)\|_{H^1}^2 &\leq {\rm e}^{A(t)}\Big( \|d(0)\|_{H^1}^2
+ \int_0^t \tilde{f}(s) \; \mathrm{d}s \Big) \\
& \qquad \times\Big\{ 1 - 4
\Big[ \|d(0)\|_{H^1}^2 + \int_0^t \tilde{f}(s) \; \mathrm{d}s \Big]^4
\int_0^t \tilde{b}(s) \;\mathrm{d}s \Big\}^{-1/4}
\end{align*}
with
\[
\tilde{b}(t) = K{\rm e}^{4 A(t)},\qquad
\tilde{f}(t) = {\rm e}^{-A(t)} \left\| \RES(t) \right\|_{H^{-1}}^2
\]
and
\[
A(t)
= -\frac{t}{4} + \int_0^t 9
\left\| \varphi_{xx}(s) \right\|_{L^\infty}^2 \; \mathrm{d}s.
\]
\end{theorem}
Again, the condition for regularity provided by the theorem depends
only on the numerical solution $\varphi$.
\begin{proof}
Apply Corollary \ref{cor:cpType} (CP-Type II) to our
inequality (\ref{eqn:upperBound}).
The corresponding functions are
\[
b(t) = \frac{7^7}{4}, \quad a(t)= 9
\left\| \varphi_{xx}(t) \right\|_{L^\infty}^2
- \frac{1}{4}, \quad f(t)= \left\| \RES(t) \right\|_{H^{-1}}^2,
\]
which immediately give us the statement of the theorem.
\end{proof}
\subsection{Second method with restarting}
The previous method can be further improved by introducing something
that can be best described as ``restarting". Instead of estimating
over the whole time interval $[0,T]$ at once, we estimate to
some smaller $t^*$ and use the resulting upper bound as the
new initial value.
\begin{theorem}\label{thm:method2R}
Given any arbitrary partition $\{t_i\}_{0 \leq i \leq n}$ of the
interval $[0,T]$ with $t_0 = 0$ and $t_n = T$, then by
Theorem \ref{thm:method2} we have for all $1 \leq i \leq n$
\begin{align*}
z(0) &:= \|d(0)\|_{H^1}^2 \\
\|d(t_i)\|_{H^1}^2 & \leq {\rm e}^{A(t_i)} \Big( z(t_{i-1}) +
\int_{t_{i-1}}^{t_i} \tilde{f}(s) \; \mathrm{d}s \Big) \\
& \; \times \Big\{ 1 - 4 \Big[ z(t_{i-1}) +
\int_{t_{i-1}}^{t_i} \tilde{f}(s) \; \mathrm{d}s \Big]^4
\int_{t_{i-1}}^{t_i} \tilde{b}(s) \;\mathrm{d}s \Big\}^{-1/4}\\
& =: z(t_i)
\end{align*}
as long as the right-hand side is finite, where for $t \in (t_{i-1},t_i]$
\[ \tilde{b}(t) = K {\rm e}^{4 A(t)}, \qquad
\tilde{f}(t) = {\rm e}^{-A(t)} \left\| \RES(t) \right\|_{H^{-1}}^2
\]
and
\[
A(t) = -\frac{1}{4} (t - t_{i-1}) +
\int_{t_{i-1}}^{t} 9 \| \varphi_{xx}(s) \|_{L^\infty}^2
\; \mathrm{d}s.
\]
\end{theorem}
\begin{proof}
Given some arbitrary partition $\{t_i\}_{0 \leq i \leq n}$ of the
interval $[0,T]$ with $t_0 = 0$ and $t_n = T$, we define our
new method as follows.
First, we apply Theorem \ref{thm:method2} to the interval $[0,t_1]$
\begin{align*}
z(0) &:= \|d(0)\|_{H^1}^2 \\
\|d(t_1)\|_{H^1}^2 & \leq {\rm e}^{A(t_1)}
\Big( z(0) + \int_0^{t_1} \tilde{f}(s) \; \mathrm{d}s \Big) \\
& \; \times \Big\{ 1 - 4 \Big[ z(0) +
\int_0^{t_1} \tilde{f}(s) \; \mathrm{d}s \Big]^4
\int_0^{t_1} \tilde{b}(s) \;\mathrm{d}s \Big\}^{-1/4} \\
&=: z(t_1)\\
\intertext{and define the upper bound for
$\|d(t_1)\|_{H^1}^2$ as $z(t_1)$. In the next step, $z(t_1)$ is
taken as the new ``initial value" when we apply
Theorem \ref{thm:method2} to the interval $[t_1,t_2]$.}
\|d(t_2)\|_{H^1}^2 & \leq {\rm e}^{A(t_2)}
\Big( z(t_1) + \int_{t_1}^{t_2} \tilde{f}(s)
\; \mathrm{d}s \Big) \\
& \; \times \Big\{( 1 - 4 \Big[ z(t_1) +
\int_{t_1}^{t_2} \tilde{f}(s) \; \mathrm{d}s \Big]^4
\int_{t_1}^{t_2} \tilde{b}(s)
\;\mathrm{d}s \Big\}^{-1/4} \\
\intertext{where $\tilde{b}(t),\tilde{f}(t)$ are defined as before,
only $A(t)$ for $t \in (t_{i-1},t_i]$ changes to}
A(t) &= -\frac{1}{4} (t - t_{i-1}) + \int_{t_{i-1}}^{t} 9
\left\| \varphi_{xx}(s) \right\|_{L^\infty}^2
\; \mathrm{d}s.
\end{align*}
This procedure is now repeated for every interval of the partition.
\end{proof}
Let us give an informal argument that this method converges to a
solution of the ODE as $h\to0$.
Let $z(t)$ be a smooth interpolation of the discrete
points $z(t_i)$, $i=1,2,\ldots$
and $h=t_{j+1}-t_j$.
Then
\[
\partial_t z(t_j) \approx \frac{z(t_{j+1})-z(t_j)}h
\]
Using, $\int_{t_j}^{t_{j+1}} g \; \mathrm{d}s \approx g(t_j)h $
and the abbreviations $z(t_j)=z_j$, $A_j = A(t_j)$ and
$\RES_j = \| \RES(t_j)\|_{H^{-1}}^2$,
we obtain from Theorem \ref{thm:method2R}
\[
\partial_t z(t_j) \approx \frac1h \Big[ \frac{{\rm e}^{A_{j+1}}
( z_j +h\tilde{f}_j)}{( 1 - 4[z_j +
h\tilde{f}_j]^4 h\tilde{b}_j)^{1/4}} - z_j\Big].
\]
Using $\tilde{b}_j=\frac{7^7}{4}{\rm e}^{4A_j} = \frac{7^7}{4} $ and
$\tilde{f}_j={\rm e}^{-A_j}\RES_j= \RES_j,$ as $A_j=0$ yields
\[
\begin{split}
\partial_t z(t_j) & \approx \frac1h
\Big[ \frac{ {\rm e}^{A_{j+1}}( z_j +h\RES_j)}{( 1 - [ z_j + h \RES_j]^4
\cdot h 7^7)^{1/4}} - z_j\Big]\\
& \approx \frac1h \Big[ \frac{ {\rm e}^{A_{j+1}}z_j +h{\rm e}^{A_{j+1}}\RES_j-
z_j\sqrt[4]{ 1 - 7^7h[ z_j + h\RES_j]^4} }
{\sqrt[4]{ 1 - 7^7h[ z_j + h\RES_j]^4}} \Big] \\
& \approx \frac1h \Big[{\rm e}^{A_{j+1}}z_j+
h{\rm e}^{A_{j+1}}\RES_j - z_j\sqrt[4]{ 1 - 7^7h[ z_j + h\RES_j]^4}\Big]
\end{split}
\]
Now using $\sqrt[4]{1-x}\approx 1 - \frac14 x + O(x^2)$ and
$A_{j+1}=O(h)$ leads to
\[
\begin{split}
\partial_t z(t_j) & \approx {\rm e}^{A_{j+1}} \RES_j + \frac1h({\rm e}^{A_{j+1}}-1)z_j
+ z_j \frac14 7^7[ z_j + h\RES_j]^4 \\
&\approx \RES_j+ A'(t_j) z_j + \frac14 7^7 z_j^5.
\end{split}
\]
Recall that $\RES_j= \| \RES(t_j)\|_{H^{-1}}^2$
and $A'(t_j)= - \frac14 + 9\| \varphi_{xx}(t_j) \|_{L^\infty}^2$,
and we recover that $z$ solves (\ref{eqn:upperBound})
with equality in the limit $h\to 0$.
\section{Numerical examples}
\label{sec:num}
To perform numerical verification rigorously,
upper bounds for the three methods
need to be calculated that include rounding errors
(e.g using interval arithmetic). However,
as our aim here is to illustrate the general behavior and
feasibility of the three methods, we neglected these rounding errors.
Although our methods allow $\varphi$ to be any arbitrary approximation,
that satisfies the boundary conditions, it
should be a reasonable choice, i.e.\ close to an expected solution,
for the methods to be successful.
For our simulations we calculate an approximate solution
using a spectral Galerkin scheme with $N$ Fourier modes in space
and a semi-implicit Euler
scheme with step-size $h$ in time, yielding the
values $\varphi(t)$ for $t=0,h,2h,...$.
To calculate the residual of $\varphi$, these values are interpolated
piecewise linearly in time.
There are two ways to show global regularity for $u(0)=u_0$ using
the numerical methods of the previous section:
\begin{itemize}
\item show that the solution exists until the time $T^*(u_0)$
(from Theorem \ref{thmtc}), since the solution is regular after
this time; or
\item show that
$\|\varphi(t)\|_{H^1} + \| d(t) \|_{H^1}<\varepsilon_0$
for some $t>0$, since then Theorem \ref{thm:smallness}
guarantees global regularity.
\end{itemize}
Note that the second criterion might be more strongly influenced by rounding
errors than the first one.
In all of our figures the maximum time
is always $T^*$, as defined by Theorem \ref{thmtc},
rounded to the first decimal digit $+ 0.1$,
which is enough to show global existence.
\begin{figure}[!ht]%
\begin{center}
\hspace*{\fill} %
\subfloat[Method 1]{\includegraphics[width=0.3\textwidth]
{img/fig1/method1.pdf}}
\hspace*{\fill} %
\subfloat[Method 2]{\includegraphics[width=0.3\textwidth]
{img/fig1/method2.pdf}}
\hspace*{\fill} %
\subfloat[Method 3]{\includegraphics[width=0.3\textwidth]
{img/fig1/method3.pdf}}
\hspace*{\fill} %
\\
\hspace*{\fill} %
\subfloat[Smallness Method 1]{\includegraphics[width=0.3\textwidth]
{img/fig1/boundsmethod1.pdf} \label{img:ex1smallnessM1}}
\hspace*{\fill} %
\subfloat[Smallness Method 2]{\includegraphics[width=0.3\textwidth]
{img/fig1/boundsmethod2.pdf} \label{img:ex1smallnessM2}}
\hspace*{\fill} %
\subfloat[Smallness Method 3]{\includegraphics[width=0.3\textwidth]
{img/fig1/boundsmethod3.pdf} \label{img:ex1smallnessM3}}
\hspace*{\fill} %
\\
\hspace*{\fill} %
\subfloat[$\| \varphi_{xx} \|_{L^\infty}$]
{\includegraphics[width=0.3\textwidth]{img/fig1/Linfphi.pdf}}
\hspace*{\fill} %
\subfloat[$\|\RES \|_{H^{-1}}$]{\includegraphics[width=0.3\textwidth]
{img/fig1/Res.pdf}}
\hspace*{\fill} %
\subfloat[$\varphi$]{\includegraphics[width=0.3\textwidth]
{img/fig1/phi.pdf}}
\hspace*{\fill} %
\\
\caption{Initial value $u(x,0) = \sin(x)$, $N=128$ Fourier modes and
step-size $h=10^{-5}$. Methods 2 and 3 show global existence as
they stay bounded until time $T^*$ and even fulfill the
smallness criterion before time $T^*$. Method 1 fails as it hits
its threshold at approximately $t=1.5 < T^*$ and also before the
smallness criterion is reached.}
\label{img:example1}
\end{center}
\end{figure}
In Figure \ref{img:example1} we have an initial value of $u(x,0) = \sin(x)$,
$N=128$ Fourier modes and a step-size of $h=10^{-5}$.
As Methods 2 and 3 stay bounded up to time $T^*$, we have global existence.
Method 1 fails because it hits its threshold at
approximately $t=1.5$, which is smaller than $T^*$. In the ``Smallness plots"
(Figures \ref{img:ex1smallnessM1}, \ref{img:ex1smallnessM2} and
\ref{img:ex1smallnessM3}) the grey area is the area
around $\|\varphi(t)\|_{H^1}$ (the solid line) with distance
$\|d(t)\|_{H^1}$, which is calculated by the respective method.
The dashed line is the critical value $\varepsilon_0 = \frac{1}{2}$.
As the order of $\|d(t)\|_{H^1}$ is $10^{-6}$ for Methods 2 and 3,
this area is not really visible in Figures \ref{img:ex1smallnessM2} and
\ref{img:ex1smallnessM3}.
An interesting detail
is, that in Figure \ref{img:ex1smallnessM1} it seems that the upper
bound reaches $\varepsilon_0$, but in fact it is still a tiny bit above
when Method 1 hits the threshold. To sum up, we have global regularity
for this initial value from Methods 2 and 3 by smallness and time criteria,
whereas both criteria fail for Method 1.
\begin{figure}[!ht]%
\begin{center}
\hspace*{\fill} %
\subfloat[Method 1]{\includegraphics[width=0.3\textwidth]
{img/fig2/method1.pdf}}
\hspace*{\fill} %
\subfloat[Method 2]{\includegraphics[width=0.3\textwidth]
{img/fig2/method2.pdf}}
\hspace*{\fill} %
\subfloat[Method 3]{\includegraphics[width=0.3\textwidth]
{img/fig2/method3.pdf}}
\hspace*{\fill} %
\\
\hspace*{\fill} %
\subfloat[Smallness Method 1]{\includegraphics[width=0.3\textwidth]
{img/fig2/boundsmethod1.pdf}}
\hspace*{\fill} %
\subfloat[Smallness Method 2]{\includegraphics[width=0.3\textwidth]
{img/fig2/boundsmethod2.pdf}}
\hspace*{\fill} %
\subfloat[Smallness Method 3]{\includegraphics[width=0.3\textwidth]
{img/fig2/boundsmethod3.pdf}}
\hspace*{\fill} %
\\
\hspace*{\fill} %
\subfloat[$\| \varphi_{xx} \|_{L^\infty}$]
{\includegraphics[width=0.3\textwidth]{img/fig1/Linfphi.pdf}}
\hspace*{\fill} %
\subfloat[$\|\RES \|_{H^{-1}}$]{\includegraphics[width=0.3\textwidth]
{img/fig2/Res.pdf}}
\hspace*{\fill} %
\subfloat[$\varphi$]{\includegraphics[width=0.3\textwidth]
{img/fig2/phi.pdf}}
\hspace*{\fill} %
\\
\caption{Same setting as in Figure \ref{img:example1}, but now with
smaller step-size $h=10^{-6}$. Now Method 1 shows global
existence, too. Note that the order of the Residual decreased
proportional to our step-size.}
\label{img:example2}
\end{center}
\end{figure}
In Figure \ref{img:example2} we get global existence also with Method 1
by decreasing the stepsize to $h=10^{-6}$. All other parameters
stay unchanged. Note that the order of the residual changed from
$10^{-4}$ to $10^{-5}$ by the same factor we decreased the step-size.
Figure \ref{img:example3} also suggests that Method 1 is inferior to the
other methods. With an already small step-size of $10^{-6}$ Method 1
does not get close to showing global existence, either by
the smallness condition or the time condition.
\begin{figure}[!ht]%
\begin{center}
\hspace*{\fill} %
\subfloat[Method 1]{\includegraphics[width=0.3\textwidth]
{img/fig3/method1.pdf}}
\hspace*{\fill} %
\subfloat[Method 2]{\includegraphics[width=0.3\textwidth]
{img/fig3/method2.pdf}}
\hspace*{\fill} %
\subfloat[Method 3]{\includegraphics[width=0.3\textwidth]
{img/fig3/method3.pdf}}
\hspace*{\fill} %
\\
\hspace*{\fill} %
\subfloat[Smallness Method 1]{\includegraphics[width=0.3\textwidth]
{img/fig3/boundsmethod1.pdf}}
\hspace*{\fill} %
\subfloat[Smallness Method 2]{\includegraphics[width=0.3\textwidth]
{img/fig3/boundsmethod2.pdf}}
\hspace*{\fill} %
\subfloat[Smallness Method 3]{\includegraphics[width=0.3\textwidth]
{img/fig3/boundsmethod3.pdf}}
\hspace*{\fill} %
\\
\hspace*{\fill} %
\subfloat[$\| \varphi_{xx} \|_{L^\infty}$]
{\includegraphics[width=0.3\textwidth]{img/fig3/Linfphi.pdf}}
\hspace*{\fill} %
\subfloat[$\|\RES \|_{H^{-1}}$]{\includegraphics[width=0.3\textwidth]
{img/fig3/Res.pdf}}
\hspace*{\fill} %
\subfloat[$\varphi$]{\includegraphics[width=0.3\textwidth]
{img/fig3/phi.pdf}}
\hspace*{\fill} %
\\
\caption{Initial value
$u(x,0) = \sin(x)+\frac{1}{2}\sin(2x)$, $N=128$
Fourier modes and step-size $h=10^{-6}$. Method 1 reaches its
threshold extremely fast. Methods 2 and 3 show global existence by
both criteria.}
\label{img:example3}
\end{center}
\end{figure}
In Table \ref{table:results} we have collected some more results
for different choices of initial data. One can see that if
we have global existence by
one criterion, then also by the other. The smallness criterion is
also reached significantly earlier than the time criterion.
\begin{table}[!htb]
\centering
\scalebox{0.8}{
\begin{tabular}{c|c|c|c|c|c|c|c|c|c}
\multicolumn{4}{c|}{\;} & \multicolumn{3}{c|}{Smallness}
& \multicolumn{3}{c}{Time} \\
$u(x,0)$ & $T^*\approx$ & N & h & M1 & M2 & M3 & M1 & M2 & M3 \\
\hline
$\sin(x)$ & $7.1$ & $128$ & $10^{-5}$ & -- & $1.17$ & $1.17$ &
$1.57$ & \checkmark & \checkmark \\
\hline
$\sin(x)$ & $7.1$ & $128$ & $10^{-6}$ & $1.2$ & $1.17$ & $1.17$ &
\checkmark & \checkmark & \checkmark \\
\hline
$\sin(x)+\frac{1}{2}\sin(2x)$ & $8$ & $128$ & $10^{-6}$ & -- &
$1.17$ & $1.17$ & $0.22$ & \checkmark & \checkmark \\
\hline
$\cos(x)-\frac{1}{2}\sin(2x)+\frac{1}{3}\cos(3x)$ & $8.3$ &
$128$ & $10^{-6}$ & -- & $1.2$ & $1.2$ & $0.1$ &
\checkmark & \checkmark \\
\hline
$\sin(2x)$ & $7.1$ & $128$ & $10^{-6}$ & -- & $0.12$ & $0.12$ &
$0.4$ & \checkmark & \checkmark \\
\hline
$2\sin(x)$ & $14.2$ & $128$ & $10^{-6}$ & -- & -- & -- &
$0.76$ & $0.14$ & $0.14$ \\
\hline
$\frac{3}{2}\cos(x)-\frac{1}{2}\sin(2x)+\frac{1}{3}\cos(3x)$ &
$11.5$ & $128$ & $10^{-6}$ & -- & -- & -- &
$0.03$ & $0.15$ & $0.15$ \\
\end{tabular}
}
\caption{All values are rounded to fit into the table. A "--" in the
"Smallness" columns means, that the smallness criterion was not met by
the respective method. Else there is the time when it was met. For
the "Time" columns this turns around. If the time criterion was met,
the respective method gets a "$\checkmark$", else the time of the
blowup / reaching the threshold. }
\label{table:results}
\end{table}
As the bounds from Methods 2 and 3 are in all of our simulations
almost identical, we need an artificial example to illustrate the
difference between those methods.
In Figure \ref{img:example4} we do this by artificially setting
a constant, relatively large $\| \RES(t) \|_{H^{-1}}$ and an also
constant, but smaller second derivative
$\| \varphi_{xx} \|_{L^\infty}$, without using any
numerical approximation.
In this case Method 3 delivers the largest time interval as it
stays finite up to $T \approx 0.16$, whereas Method 2 has a blowup at
$T \approx 0.11$ and Method 1 at $T \approx 0.03$.
\begin{figure}[!ht]%
\begin{center}
\hspace*{\fill} %
\subfloat[Method 1]{\includegraphics[width=0.3\textwidth]
{img/fig4/method1.pdf}}
\hspace*{\fill} %
\subfloat[Method 2]{\includegraphics[width=0.3\textwidth]
{img/fig4/method2.pdf}}
\hspace*{\fill} %
\subfloat[Method 3]{\includegraphics[width=0.3\textwidth]
{img/fig4/method3.pdf}}
\hspace*{\fill} %
\\
\hspace*{\fill} %
\subfloat[$\| \varphi_{xx} \|_{L^\infty}$]
{\includegraphics[width=0.3\textwidth]{img/fig4/phi.pdf}}
\hspace*{\fill} %
\subfloat[$\|\RES \|_{H^{-1}}$]{\includegraphics[width=0.3\textwidth]
{img/fig4/Res.pdf}}
\hspace*{\fill} %
\\
\caption{Artificial example for large fixed residual and
not too large fixed second derivative. Method 3 is superior,
but all methods do blow up relatively fast,
as the residual is large.}
\label{img:example4}
\end{center}
\end{figure}
\section{Conclusion}
We presented a method to verify global existence and uniqueness
by combining a-posteriori numerical data and a-priori estimates.
Therefore we prove a differential inequality for the error from the
data to the true solution,
having coefficients depending only on the numerical data.
Three methods are presented to evaluate rigorously
the analytic upper bounds for the error from these differential
inequalities.
The third method seems to be the best, as it provides rigorous upper bounds
and converges to a solution of the equality in the differential inequality.
Nevertheless, in all practical examples with our Galerkin
approximation, Methods 2 and 3 have shown nearly indistinguishable results.
While our proofs are rigorous, the implementation of the verification
methods are not
completely rigorous since we neglected rounding errors.
Our analysis and computations suggest that
numerical verification of regularity is feasible and can obtain
global existence for initial conditions that are not covered by current
analytical results.
We plan to perform a fully rigorous numerical verification
in a future paper,
using interval arithmetic to keep track of truncation errors.
Moreover, replacing the a-priori estimates of the linearized operator
by rigorous numerical estimates for its spectrum looks promising.
|
\section{Introduction}
The open charm production cross section $\sigma(c\bar c)$ at near threshold energy
in pA-interactions was given in our earlier works \cite{PAN73, PAN74} on
research of
$D^0$ mesons characteristics with SVD-2 setup \cite{IET56}.
In this work the results
of search and analysis of $D^+ \rightarrow K^-\pi^+\pi^+$
and $D^- \rightarrow K^+\pi^-\pi^-$ decays
in pA-interactions at 70 GeV are presented. Charged charm mesons production
inclusive cross sections were estimated and their properties were measured.
52 million of inelastic events were detected with three nuclear targets and
used in our analysis. Selection procedures for events with possible 3-prongs
decays of charged $D$ mesons were the following:\\
- Reconstruction of tracks and primary vertex using vertex detector data.\\
- Search of two prongs secondary vertices in track parameters space \verb|{a, b}|
\cite{Pr17}.
In this space each track is presented by a point and all points for tracks
from the same vertex lie on a straight line.\\
- Spatial reconstruction of charged particles tracks in the magnetic
spectrometer.\\
- Search of 3-prongs secondary vertices taking into account their charge signs
and spatial association with primary vertex.
After primary selection there were 16320 events with ($K^-\pi^+\pi^+$) decay hypothesis
and 8439 with ($K^+\pi^-\pi^-$) hypothesis. In fig. 1 raw experimental effective
mass spectra of two systems are presented. Signals from $D^\pm$ mesons
can be observed over the large background. To diminish this background additional cuts
and modeling were required.
\begin{figure*}[h]
\includegraphics[scale=0.35]{fig2a.eps
\hspace{0.2in}
\includegraphics[scale=0.35]{fig2b.eps
\centering
\caption{Raw experimental effective-mass spectra of ($K^-\pi^+\pi^+$) (a)
and ($K^+\pi^-\pi^-$) systems (b).}
\end{figure*}
\section{Modeling and optimization of selection criteria}
GEANT3.21 \cite{GEANT} program with the description of all SVD-2 components was used for modeling
and optimization of selection criteria for $D^\pm$ meson events. FRITIOF7.02 \cite{PYTHIA}
program was
used as the generator of pA-interactions.
At first
step the background under signals was simulated using 10 million Monte-Carlo (MC) events
without charm. 3-prongs secondary vertices were found in some
events because of detector noise and feature of procession algorithm. Distributions
of some characteristics (decay length ($L$), momentum ($P$) and
Feynman variable ($X_F$)) of 3-prongs systems for MC-events and experimental background
in interval of $D$ meson masses from fig. 1 ($M=1.86\pm3*0.02$ (GeV)) were compared.
The proper decay length $L$ was calculated from the observed $L_{lab}$ as $L=L_{lab}*M/P$.
All distributions really reproduce experimental background conditions. Momenta of
3-prongs systems lie above 7 GeV.
A half of million MC-events with $D^+ \rightarrow K^-\pi^+\pi^+$ decay were used
for optimization of selection criteria. At the first the Dalitz-plot for ($K^-\pi^+\pi^+$)
system was analyzed in $m_1=m(K^-\pi^+_1)$ and
$m_2=m(K^-\pi^+_2)$ coordinates.
All MC-events are grouped within an ellipse (fig. 2a).
In fig. 2b mass plot for experimental events with MC-events ellipse is presented.
For events in the ellipse, dependences of event densities on $\phi$ angle are shown
in fig. 2c. From the analysis of fig. 2c the following selection criteria were
taken: $\phi < 200^\circ$, or $\phi > 340^\circ$ and $R_{ell} <$ 1.
\begin{figure*}[h]
\includegraphics[scale=0.5]{fig4.eps
\includegraphics[scale=0.35]{fig5.eps
\includegraphics[scale=0.3]{fig6.eps
\centering
\caption{a) Dalitz-plot for ($K^-\pi^+\pi^+$) system. b) MC-ellipse imposed on
experimental Dalitz-plot.
c) Experimental and MC-events densities versus $\phi$.}
\end{figure*}
Another background occurs when charge track from primary vertex combined
with $K^0$ decay vertex.
To remove it another mass plot was considered:
in the same 3-prongs secondary vertex $K^-$ candidate was replaced with $\pi^-$ candidate.
In fig. 3a the plot with two pions mass hypotheses is presented.
The $K^0$ background lies in the lower part of the plot.
Events under the line
($M(\pi^+\pi^-)_{H1} + M(\pi^+\pi^-)_{H2} < C$)
have to be excluded. If $N_{cut}$ -- is the number of the rejected events and
$N_{tot}$-- total number of events, then the share of rejected events
$W=N_{cut}/N_{tot}$ depends on $C$.
From the analysis
of distributions in fig. 3b the cut parameter $C=$ 1.2 was taken for reduction of
$K^0$ background. The $K^0$ background became practically unseen.
\begin{figure*}[h]
\includegraphics[scale=0.35]{fig8.eps
\hspace{0.4in}
\includegraphics[scale=0.38]{fig9.eps
\centering
\caption{a) The plot for ($\pi\pi\pi$) system.
b) Dependence of $W$ value on parameter $C$.}
\end{figure*}
Distributions for MC-events
with $D^+$ and for experimental
events with 3-prongs secondary vertex vs. proper decay length $L$ of
($K^-\pi^+\pi^+$) system were got. From the analysis of these distributions the
$L >$ 0.12 mm cut was introduced.
\section{The analysis of experimental events}
As a result of simulation the selection criteria for $D^+$ decays with the minimum
background were taken:
1) $\phi(K^-\pi^+) < 200^\circ$, or $\phi(K^-\pi^+) > 340^\circ$ and $R_{ell} <$ 1;
2) $M(\pi^+\pi^-)_{H1} + M(\pi^+\pi^-)_{H2} <$ 1.2 GeV;
3) $L(K^-\pi^+\pi^+) >$ 0.12 mm.
The $D^+$ detection efficiency obtained after application of these criteria
to 500000 MC-events is 1.4\verb
500000 MC-events with $D^-$ decays. Detection efficiency for $D^-$ is 0.8\verb
In figs. 4a and 4b the experimental mass spectra of ($K^-\pi^+\pi^+$) and ($K^+\pi^-\pi^-$)
systems are presented. Signals of $D$ mesons were fitted by the sum of Gaussian
function and 6-order polynomial background. The parameters of the fits for $D^+$
were: $\chi^2/$NDF = 7.4 / 12, prob = 0.8; signal from $D^+$ = 15.4 events;
background under the signal = 16.6 events; $D^+$ mass = 1873$\pm$5 MeV;
standard deviation = 12 MeV. For $D^-$: $\chi^2/$NDF = 2.7 / 11,
prob = 0.99; signal from $D^-$ = 15.3 events; background under the signal = 8.7 events;
$D^-$ mass = 1863$\pm$8 MeV; standard deviation = 22 MeV.
The measured values of charged $D$ mesons masses are near to PDG value (1869.6 MeV)
within the errors. In the mass interval of $D$ mesons a $K^0$ background was not found.
\begin{figure*}[h]
\includegraphics[scale=0.35]{fig12.eps
\hspace{0.3in}
\includegraphics[scale=0.35]{fig13.eps
\centering
\caption{Effective-mass spectra of ($K^-\pi^+\pi^+$) (a) and
of ($K^+\pi^-\pi^-$) (b) systems.}
\end{figure*}
Measuring $D$ mesons lifetime we can get another proof of observing charm particles.
Events from signal slice ($M(D) \pm 2.5*\sigma$) were selected.
The decay length distributions were constructed for them taking into account detection
efficiency. The distributions were fitted by exponent. The background
under a signal was estimated with help of distributions for MC-events. The measured
values of the $c\tau$ parameters were 291 $\pm$ 75 $\mu$m for $D^+$ meson
and 341 $\pm$ 88 $\mu$m for $D^-$ meson. They coincide with PDG value
(311.8 $\mu$m) within the errors. Only statistical errors are here.
The lifetime for events outside of signal areas considerably differ from these values.
\section{Inclusive cross sections and A-dependence}
The formula for calculation of cross sections was taken as:
\begin{center}
$N_s = [N^0(\sigma_DA^\alpha)/(\sigma_{pp}A^{0.7})]\times[($B$\varepsilon)/K_{tr}]$, where
\end{center}
$N_s$ -- number of events in a signal;
$N_0$ -- number of events with pA-interactions in a target;
$\sigma_D$ -- charm cross section;
$A$ -- nuclear weight of target material (C, Si, Pb);
$\alpha$ -- parameter of A-dependence for charm cross section (= 0.7 for background);
$\sigma_{pp}$ -- inelastic pp-interactions cross section at 70 GeV (= 31440 mb);
B -- branching ratio of $D^\pm \rightarrow K\pi\pi$ decay (= 0.094);
$\varepsilon$ -- detection efficiency for $D$ mesons ($\varepsilon(D^+)$ = 0.014,
$\varepsilon(D^-)$ = 0.008);
$K_{tr}$ = 0.57 (trigger efficiency \cite{PAN73} after specification);
With the expressions:
$C_D = [N^0/(\sigma_{pp}A^{0.7})]\times[($B$\varepsilon)/K_{tr}]$ and
$ln(N_s / C_D) = \alpha \times ln(A) + ln(\sigma_D)$
A-dependence of cross sections was received. The slope parameters of linear fits
are: 1.02 $\pm$ 0.26 for $D^+$ and 1.04 $\pm$ 0.27 for $D^-$.
The average values of inclusive cross sections (weighed on target materials) are:
\begin{center}
$\sigma(D^+)$ = 1.2 $\pm$ 0.4(stat.) $\pm$ 0.2(syst.) ($\mu$b/nucleon),\\
$\sigma(D^-)$ = 1.9 $\pm$ 0.6(stat.) $\pm$ 0.4(syst.) ($\mu$b/nucleon).
\end{center}
The relative errors of the cross sections are: near 30\verb
and near 15\verb
calculations.
\section{The ratios of charm meson yields}
In earlier paper \cite{PAN73} the estimation of open charm total cross section neutral $D$ mesons
observations in pA-interactions at 70 GeV was obtained as:
\begin{center}
$\sigma(c\bar c)$ = 7.1 $\pm$ 2.4(stat.) $\pm$ 1.4(syst.) ($\mu$b/nucleon).\\
\end{center}
The cross sections of neutral charm mesons and anti-mesons were estimated as:
\begin{center}
$\sigma(D^0)$ = 2.5 $\pm$ 0.8(stat.) $\pm$ 0.5(syst.) ($\mu$b/nucleon),\\
$\sigma(\bar D^0)$ = 4.6 $\pm$ 1.6(stat.) $\pm$ 0.9(syst.) ($\mu$b/nucleon).
\end{center}
Table 1. Yields of D mesons and their ratios.
\begin{center}
\small{
\begin{tabular}{|l|c|c|c|c|c|c|c|} \hline
\multicolumn{1}{c}{Mesons}&\multicolumn{1}{c}{PYTHIA}&\multicolumn{3}{c}{FRITIOF}&
\multicolumn{1}{c}{SVD-2}&\multicolumn{2}{c}{Other experiments} \\
\multicolumn{1}{c}{}&\multicolumn{1}{c}{pp-int.}&\multicolumn{3}{c}{pA-interactions}&
\multicolumn{1}{c}{pA-int.}&\multicolumn{2}{c}{pA-interactions} \\ \hline
&&C&Si&Pb&&NA-27 \cite{PL189}&HERA-B \cite{EPJ52} \\ \hline
$D^0$ & 0.28& 0.48&0.51&0.55&0.35$\pm$0.16&0.57$\pm$0.08&0.44$\pm$0.18 \\ \hline
$\bar D^0$ & 0.74& 0.60&0.59&0.58&0.65$\pm$0.31&0.43$\pm$0.09&0.54$\pm$0.23 \\ \hline
$D^+$ & 0.13& 0.28&0.29&0.29&0.16$\pm$0.07&0.31$\pm$0.06&0.19$\pm$0.08 \\ \hline
$D^-$ & 0.24& 0.28&0.27&0.28&0.27$\pm$0.17&0.34$\pm$0.06&0.25$\pm$0.11 \\ \hline
$D^0 / \bar D^0$&0.38&0.80&0.86&0.95&0.54$\pm$0.25&1.33$\pm$0.25&0.81$\pm$0.23 \\ \hline
$D^+ / D^-$ & 0.54& 1.0&1.1&1.0&0.59$\pm$0.20&0.92$\pm$0.21&0.76$\pm$0.22 \\ \hline
$D^\pm/$&0.36&0.51&0.51&0.5&0.44$\pm$0.24&0.65$\pm$0.21&0.46$\pm$0.18 \\
$(D^0+\bar D^0)$&&&&&&& \\ \hline
$D^+ / D^0$ & 0.18& 0.56&0.52&0.46&0.59$\pm$0.21&0.54$\pm$0.11&0.44$\pm$0.12 \\ \hline
$D^- / \bar D^-$ & 0.32& 0.47&0.46&0.48&0.42$\pm$0.26&0.78$\pm$0.19&0.47$\pm$0.14 \\ \hline
\end{tabular}
}
\end{center}
\begin{figure*}[h]
\includegraphics[scale=0.45]{fig17.eps
\hspace{0.5in}
\includegraphics[scale=0.35]{fig18.eps
\centering
\caption{a) Yields of D-mesons. b) Ratios of ($D^++D^-$) and ($D^0+\bar D^0$) cross
sections. }
\end{figure*}
In table 1 yields of $D$ mesons and their ratios using SVD-2 results and data
of other experiments for pA-interactions are presented. The yields received from
the PYTHIA and FRITIOF programs at our energy are also given.
Fig. 5a shows that yields of mesons are decreasing with drop of energy, but
yields of anti-mesons are increasing. The difference in the yields of particles
and antiparticles was observed for the first time in a nA-interactions at average
neutron beam energy 43 GeV in BIS-2 experiment \cite{ZP37}.
In this experiment the decays of antiparticles ($\bar D^0$ and $D^-$) were detected, but
the decays of particles ($D^0$ and $D^+$) were not found. Cross sections of particles
production might appear below the sensitivity threshold in this experiment.
In fig. 5b the ratios of cross sections of charged and neutral $D$ mesons from
paper \cite{PR433}
and the present result are shown. The results are compared to the predictions of
statistical hadronization model \cite{Andronic}.
\section{Conclusion}
In SERP-E-184 experiment at SVD-2 setup (Protvino, Russia) $D^\pm$ mesons signals
were obtained in effective-mass spectra of 3-prongs ($K\pi\pi$) systems in
pA-interactions at 70 GeV. The selection criteria of events with open charm
production were optimized using detailed simulation with FRITIOF7.02 and GEANT3.21
programs. Inclusive cross sections of $D^\pm$ mesons production were estimated
at near threshold energy.
The SVD-2 active target with plates of different materials (C, Si, Pb) allowed to measure
the A-dependence parameters of cross sections for $D^\pm$ mesons production.
The yields of $D$ mesons and their ratios in comparison with data of other
experiments and theoretical predictions were estimated. Experimental data showed
the changes in $D$ mesons yields with a decrease of pA-interaction energy. These
results are close to the predictions of a statistical hadronization model.
|
\section{Introduction}
Smooth Calabi-Yau compactifications of the heterotic string and M-theory represent one of the classic and most-developed avenues from string theory to low energy physics \cite{Candelas:1985en}. Despite the great interest received over the years, however, this approach, until recently, has led to only a relatively small number of models that satisfy the most basic phenomenologically requirements, such as exhibiting the Standard Model particle content without any exotics \cite{Braun:2005ux, Braun:2005bw, Bouchard:2005ag, Anderson:2009mh, Braun:2011ni}. The scarceness of such models, owing to the considerable mathematical complexity involved in the analysis of the compactification geometry, has made it difficult to address more detailed phenomenological questions, such as proton decay or fermion masses, in a meaningful way.
Recently, this situation has changed. In a series of publications \cite{Anderson:2011ns, Anderson:2012yf, Anderson:2013xka, He:2013ofa}, a promising class of $E_8\times E_8$ heterotic compactifications, based on line bundle sums, has been proposed and analyzed. These heterotic line bundle models are based on rank four or five Abelian vector bundles over smooth Calabi-Yau three-folds with non-trivial fundamental group. Using line bundle sums, rather than bundles with non-Abelian structure groups, comes with a number of far-reaching advantages. For one, split bundles lead to the presence of additional, normally anomalous $U(1)$ symmetries which constrain the structure of the low-energy theory and can have interesting phenomenological implications. Further, the split nature of the bundle facilitates an efficient algorithmic search for physical models and large numbers of promising examples can be found in this way.
Specifically, in Refs.~\cite{Anderson:2011ns, Anderson:2012yf} over 200 $SU(5)$ GUT models were constructed on discrete quotients of complete intersection Calabi-Yau manifolds. All these have precisely three generations of GUT families, no anti-families, at least one $\mathbf{5} - \overline{\mathbf{5}}$ pair of Higgs fields and no other matter charged under the GUT group. After forming quotients by the freely-acting discrete symmetries and including Wilson lines, these models led to about 2000 models with the MSSM matter field spectrum plus one or more pair of Higgs doublets. In Ref.~\cite{Anderson:2013xka}, this preliminary search for standard-like models was extended to an exhaustive scan which led to some $35,000$ $SU(5)$ GUT models over the same class of Calabi-Yau manifolds. Finally, in \cite{He:2013ofa}, over $100$ $SU(5)$ models and about $29,000$ $SO(10)$ GUT models were constructed over $14$~Calabi-Yau hypersurfaces embedded in toric varieties. These large numbers are indicative of the huge potential of the line bundle construction.
The present paper will be studying heterotic line bundle models using a complementary approach. Rather than scanning over classes of Calabi-Yau manifolds and large numbers of bundles, we will be focusing on a specific Calabi-Yau manifold, the tetra-quadric hypersurface in $\left(\mathbb{C}\mathbb{P}^1\right)^{\!\times 4}$, for an in-depth analysis of some aspects of heterotic line bundle models. In particular, we will use the present case study to elaborate on three, related points. Firstly, it was noticed in \cite{Anderson:2013xka} that the number of viable models reaches a certain saturation limit after repeatedly increasing the range of integers defining the line bundles.
While this observation was made computationally, it is less clear how to prove finiteness of the class of relevant vector bundles -- essentially poly-stable line bundle sums with fixed total Chern class. We revisit this question here and present various partial results including a finiteness proof which relies on excluding the regions in K\"ahler moduli space close to the boundary; in physical terms this essentially corresponds to the supergravity approximation.
The second objective of the paper is to present in detail a line bundle model on the tetra-quadric manifold that has modestly attractive phenomenological properties. The model exhibits a superpotential that leads to a rank two up quark mass matrix and, while a rank one matrix may be preferably at this level, this means a perturbative and generically large top Yukawa coupling is present. The down quark and lepton Yukawa matrices are entirely zero at the perturbative level so, for a realistic model, they would have to be generated non-perturbatively. Further, all dimension four and five operators which can lead to fast proton decay are forbidden. This model serves as a starting point for the subsequent analysis of the bundle moduli space.
Finally, perhaps the most important issue is to understand how a given line bundle model is embedded into the larger moduli space of non-Abelian bundles and we will study this question for the aforementioned model on the tetra-quadric. Methodically, this problem can be approached in two complementary ways. From a fundamental point of view, one can attempt to construct non-Abelian bundles which split to the given line bundle sum at a particular locus in bundle moduli space. On the other hand, in the context of the four-dimensional effective theory, the analogous process can be carried out by giving VEVs to the bundle moduli. For our example model, we will consider both approaches as well as the relation between them. Similar studies have been undertaken in \cite{Anderson:2010ty, Kuriyama:2008pv, Anderson:2010tc}. An important question in this context concerns the fate of the Higgs doublet pair. As a vector-like pair, the Higgs is not automatically protected from acquiring a mass as one moves in moduli space. In fact, for most of the semi-realistic models constructed to date, the Higgs doublets are massless only on a specific sub-locus in moduli space and receive a typically super-heavy mass elsewhere. For the line bundle model, we have -- by construction -- a massless Higgs pair at the Abelian locus, but this is not guaranteed to remain massless as we move into the non-Abelian part of the moduli space. Clearly, the conditions under which Higgs doublets are part of the low-energy spectrum, in low-energy parlance referred to as the $\mu$-problem, is of vital importance for string model building and we will study this question for our example model.
These seemingly dissimilar questions run along a common thread and illustrate several problems which arise when studying the moduli space of heterotic line bundle models.
Coming from afar, we identify -- for our tetra-quadric example manifold -- a substantial, though finite, number of points in this moduli space representing Abelian bundles which lead to phenomenologically promising
low-energy theories. In the second step, Section~\ref{sec:model}, we focus on one of these points and, in the last part, Section~\ref{sec:monads}, we ``zoom out" again,
this time in order to explore the space of non-Abelian deformations around the chosen point.
The non-Abelian bundles discussed in Section~\ref{sec:monads} have structure group $SU(4) \times U_X(1)$ leading to $SU(5)$ GUT models with an additional global $U_X(1)$ symmetry which can be seen as a remnant of the four $U(1)$ symmetries present at the Abelian locus. Combined with hypercharge, this $U_X(1)$ leads to a $B-L$ symmetry. Furthermore, these models have a number of appealing phenomenological features such as a certain hierarchy of Yukawa couplings and the absence of dimension four and dimension five operators leading to a fast proton decay. In part, these operators are forbidden by the surviving $U_X(1)$ symmetry. However, it is well known that operators of the form ${\bf 10}\,{\bf 10}\,{\bf 10}\,\overline{\bf 5}$ are invariant under the $U_X(1)$ symmetry. Yet, in our $SU(5)\times U_X(1)$ models these operators are absent because of the presence of additional $U(1)$ symmetries which appear at the Abelian locus. Something similar happens with the Higgs multiplets. By construction we have a massless pair of Higgs doublets at the Abelian locus but, as we will see, the $U(1)$ symmetries also forbid a $\mu$-term with singlet insertions. This means that the Higgs doublets remain massless as we continue away from the Abelian locus to an $SU(5)\times U_X(1)$ model. These examples demonstrate the power and the advantage of the present approach: we start with an Abelian model which is easier to construct and analyse, but the symmetries which arise at the Abelian locus lead to some degree of control as to which couplings will or will not appear as we move into the non-Abelian part of the moduli space.
The discussion runs on two, even three, levels. On one hand, we have the high energy theory, described in terms of the compactification data: a Calabi-Yau three-fold supplied with a holomorphic, poly-stable bundle. On the other hand, this geometrical set-up leads to a four-dimensional, low-energy supersymmetric GUT, whose gauge group can further be broken to that of the Standard Model. Frequently, we switch back and forth between the high-energy and
the effective GUT. The breaking to the Standard Model is only discussed briefly in Section~\ref{sec:model}, simply to illustrate the virtues of the chosen line bundle model and will be presented in detail elsewhere.
\section{Heterotic line bundle models}\label{sec:lbs}
We begin by reviewing the construction of smooth heterotic compactifications with Abelian bundles. The line bundle construction has been extensively discussed in Refs.~\cite{Anderson:2011ns, Anderson:2012yf} and, below, we summarise its most important features, including the derivation of the GUT spectrum. Knowledge of the GUT spectrum will be crucial for the comparison between Abelian and non-Abelian compactifications in Section~\ref{sec:monads}.
\subsection{Construction}\label{sec:construction}
In the rest of the paper we will discuss $E_8\times E_8$ heterotic line bundle models leading to $SU(5)$ GUT models. For this purpose, we choose a Calabi-Yau three-fold, $X$, with a freely-acting discrete symmetry $\Gamma$ and a vector bundle $V\rightarrow X$ which is given by the sum of line bundles
\begin{equation}
V = \bigoplus_{a=1}^5 L_a\; .
\end{equation}
In order for $V$ to have structure group $S\left(U(1)^5\right)$, such that we are able to use the embedding $S\left(U(1)^5\right)\subset SU(5)\subset E_8$, we demand that $c_1(V)=0$. In practice, we can choose an integral basis, $\{J_i\}$ of the second cohomology of $X$ and characterise line bundles by their first Chern class, that is, write $L_a={\cal O}({\bf k}_a)$ if $c_1(L_a)=k_a^iJ_i$. Here the indices $i,j,\ldots $ run from $1$ to $h^{1,1}(X)$. The above line bundle sum is then specified by an $h^{1,1}(X)\times 5$ matrix, $(k_a^i)$, of integers and the vanishing of the first Chern class, $c_1(V)=0$, translates into the condition
\vspace{-4pt}
\begin{equation}
\sum_{a=1}^5{\bf k}_a=0\; . \label{c10}
\vspace{-4pt}
\end{equation}
The second Chern class and the index of such line bundle sums are given by
\vspace{-4pt}
\begin{equation}
c_2(V)=-\frac{1}{2}d_{ijk}\sum_{a=1}^5k_a^ik_a^j\;,\quad\quad
{\rm ind}(V)=\frac{1}{6}d_{ijk}\sum_{a=1}^5k_a^ik_a^jk_a^k\; ,\label{c2ind}
\vspace{-4pt}
\end{equation}
where the triple intersection numbers $d_{ijk}$ of $X$ are defined by
\begin{equation}
d_{ijk}=\int_XJ_i\wedge J_j\wedge J_k\; ,
\end{equation}
as usual. For a consistent heterotic compactification we have to ensure that the anomaly condition
\vspace{-4pt}
\begin{equation}
{\rm ch}_2(V) +{\rm ch}_2(V')-{\rm ch}(TX)=[W]\; ,\label{anomgen}
\vspace{-4pt}
\end{equation}
can be satisfied, where $V'$ is the bundle in the other (hidden) $E_8$ sector and $[W]$ is the (Poincar\'e dual of the) class of a holomorphic curve $W$ wrapped by a five-brane. A practical way to ensure this condition can be satisfied without having to explicitly construct the hidden sector is to demand that~\footnote{We are assuming here that $c_1(V)=0$, so that ${\rm ch}_2(V)=-c_2(V)$.}
\vspace{-4pt}
\begin{equation}
c_2(TX)-c_2(V)\in \mbox{Mori cone of }X\; . \label{anom}
\end{equation}
Then the anomaly condition~\eqref{anomgen} can always be satisfied by a suitable choice of five-brane (although other configurations which involve a non-trivial hidden bundle are normally possible as well).
Finally, we need to guarantee that the observable bundle is poly-stable with slope zero, so that supersymmetry is preserved by the gauge fields. Line bundles are automatically stable, so what remains to be checked is that all line bundles have vanishing slope
\vspace{-4pt}
\begin{equation}
\mu(L_a)\,\equiv\int_Xc_1(L_a)\wedge J\wedge J=d_{ijk}\,k_a^i\,t^j\,t^k\stackrel{!}{=}0\label{slope0}
\vspace{-8pt}
\end{equation}
for a common locus in K\"ahler moduli space, parametrized by $J=t^iJ_i$ with moduli $t^i$. The relative simplicity of this condition, as opposed to the condition of stability for non-Abelian bundles, is one of the major technical advantages of line bundle models.
We note that a line bundle $L$ (other than the trivial bundle) with vanishing slope, $\mu(L)=0$, has vanishing zeroth and third cohomology, $H^0(X,L)=H^3(X,L)=0$ so that
\begin{equation}
{\rm ind}(L)=-h^1(X,L)+h^1(X,L^{^*})\; .
\end{equation}
\subsection{The Spectrum}
For a bundle structure group $S(U(1)^5)\subset SU(5)\subset E_8$ the low-energy gauge group, given by the commutant of the structure group within $E_8$, is $SU(5)\times S\left(U(1)^5\right)$. The matter multiplets present in the four-dimensional theory can be obtained by decomposing the adjoint ${\bf 248}_{E_8}$ of $E_8$ under the $SU(5)\times S\left(U(1)^5\right)$ sub-group which leads to
\begin{equation}
{\bf 10}_a\; ,\quad \overline{\bf 10}_a\; ,\quad \overline{\bf 5}_{a,b}\;,\quad {\bf 5}_{a,b}\; ,\quad {\bf 1}_{a,b}\; ,
\end{equation}
Here the number indicates the $SU(5)$ representation and the indices $a,b,\ldots =1,\dots 5$ indicate which of the five $U(1)$ symmetries the multiplet is charged under. Specifically, the ${\bf 10}_a$ ($\overline{\bf 10}_a$) multiplets carry charge $+1$ ($-1$) under the $a^{\rm th}$ $U(1)$ symmetry while being uncharged under the others. The $\overline{\bf 5}_{a,b}$ (${\bf 5}_{a,b}$) multiplets carry charge $+1$ ($-1$) under $U(1)$ charges $a$ and $b$ while the singlets ${\bf 1}_{a,b}$ carry charge $+1$ under the $a^{\rm th}$ and charge $-1$ and the $b^{\rm th}$ $U(1)$. The multiplicity of each of these multiplets can be computed from line bundle cohomology, as summarised in Table~\ref{spectrum}.
\begin{table}[!h]
\vspace{12pt}
\begin{center}
\begin{tabular}{|l|l|l|l|l|l|l|l|}
\hline
\varstr{14pt}{9pt} repr. & cohomology & total number & required for MSSM \\ \hline\hline
\varstr{14pt}{9pt} $~{\bf 1}_{a,b}$ & $H^1(X, L_a \otimes L_b^{^*})$ & $\sum_{a,b} h^1(X, L_a \otimes L_b^{^*}) = h^1(X, V \otimes V^{^*})$ & \;\;\;\;\; - \\ \hline
\varstr{14pt}{9pt} $~{\bf 5}_{a,b}$ & $H^1(X, L_a^{^*} \otimes L_b^{^*})$ & $\sum_{a<b} h^1(X, L_a^{^*} \otimes L_b^{^*}) =h^1(X, \wedge^2 V^{^*}) $ & \;\;\;\;\;$n_h$\\ \hline
\varstr{14pt}{9pt} $~{\bf \overline{5}}_{a,b}$ & $H^1(X, L_a \otimes L_b)$ & $\sum_{a<b} h^1(X, L_a \otimes L_b) =h^1(X, \wedge^2 V) $ & \;\;\;\;\;$3 |\Gamma| + n_h$\\ \hline
\varstr{14pt}{9pt} $~{\bf 10}_{a}$ &$H^1(X, L_a)$ & $\sum_a h^1 (X,L_a) = h^1 (X,V)$& \;\;\;\;\;$3 | \Gamma|$\\ \hline
\varstr{14pt}{9pt} $~{\bf \overline{10}}_{a}$ & $H^1(X, L_a^{^*})$ & $\sum_a h^1(X,L_a^{^*}) = h^1(X,V^{^*})$&\;\;\;\;\; 0
\\ \hline
\end{tabular}
\vskip 0.4cm
\parbox{16.7cm}{\caption{\it\small The spectrum of $SU(5)$ GUT models derived from the heterotic line bundle construction. In the final column, $|\Gamma|$~stands for the order of the fundamental group of $X$ and $n_h$ represents the number of $\mathbf{5}-\overline{\mathbf{5}}$ Higgs fields.}\label{spectrum}}
\end{center}
\vspace{-14pt}
\end{table}
The cohomology of line bundles is usually easier to compute than that of non-Abelian bundles and this constitutes another major technical advantage of line bundle models. The phenomenological requirements on the GUT particle spectrum -- essentially the three-family constraint plus having an additional ${\bf 5} - \overline{\bf 5}$ pair to account for the Higgs doublets -- are summarized in the last column of Table~\ref{spectrum}.
In order to arrive at a standard-like model, we need a freely-acting symmetry $\Gamma$ on $X$, with order~$|\Gamma|$, which can be lifted to the bundle $V$, that is, the bundle $V$ needs to have a $\Gamma$-- equivariant structure. Then, performing the quotient by $\Gamma$ and including a Wilson line in the hypercharge direction will break the GUT group to the Standard Model group times $S(U(1)^5)$. These additional $U(1)$ symmetries are usually Green-Schwarz anomalous with super-heavy associated gauge bosons and, therefore, do not constitute a phenomenological problem. Upon quotienting by $\Gamma$, the number $3|\Gamma|$ of ${\bf 10}\oplus\overline{\bf 5}$ families which we have required for our GUT models, automatically become $3$ standard model families. From the additional $\overline{\bf 5}-{\bf 5}$ multiplets we should keep one pair of Higgs doublets and ensure that all Higgs triplets are projected out. This can frequently be achieved by a suitable choice of equivariant structure and Wilson line. Then, we have a standard model charged spectrum precisely as in the MSSM plus additional moduli fields -- the bundle moduli ${\bf 1}_{a,b}$ and gravitational moduli -- which are uncharged under the standard model group. Our experience is that many such models can be found relatively easily and in this paper we focus on the examples on the tetra-quadric.
{\setstretch{1.14}
\section{Heterotic line bundle models on the tetra-quadric}
In this section, we focus on the tetra-quadric manifold, discuss its specific properties and present the scan for phenomenologically interesting models on this manifold.
\subsection{The tetra-quadric}
A detailed discussion of the tetra-quadric, particularly of its K\"ahler cone, is provided in Appendix~\ref{app:KahlerCone}. Here we summarise the most important points. Tetra-quadric Calabi-Yau hypersurfaces are embedded in a product of four $\mathbb{C}\mathbb{P}^1$ spaces, defined as the zero locus of some homogeneous polynomial that is quadratic in the homogeneous coordinates of each $\mathbb{C}\mathbb{P}^1$ space. Manifolds in this class have Euler number $\eta = - 128$ and Hodge numbers $h^{1,1}(X)=4$ and $h^{2,1}(X)=68$. This information is summarised by the following configuration matrix:
\begin{equation}
X~=~~
\cicy{\mathbb{C}\mathbb{P}^1 \\ \mathbb{C}\mathbb{P}^1\\ \mathbb{C}\mathbb{P}^1\\ \mathbb{C}\mathbb{P}^1}
{ ~2 \!\!\!\!\\
~2\!\!\!\! & \\
~2\!\!\!\! & \\
~2\!\!\!\!}_{-128}^{4,68}\
\end{equation}
At certain loci in the complex structure moduli space, the tetraquadric hypersurface admits free actions of finite groups of orders $|\Gamma|=2, 4, 8,16$. Specifically, these groups are $\Gamma= \mathbb{Z}_2,\,\mathbb{Z}_2\times \mathbb{Z}_2,\, \mathbb{Z}_4$, $\mathbb{Z}_2\times \mathbb{Z}_4,\,\mathbb{Z}_8,\,\mathbb{H},\,\mathbb{Z}_4\times \mathbb{Z}_4,\,\mathbb{Z}_4 \rtimes \mathbb{Z}_4,\,\mathbb{Z}_8\times \mathbb{Z}_2,\,\mathbb{Z}_8\rtimes \mathbb{Z}_2,\,\mathbb{H}\times \mathbb{Z}_2$. Being at one or another of these special loci corresponds to different choices of coefficients for the monomials composing the defining polynomial, as discussed in Refs.~\cite{Candelas:2008wb, Candelas:2010ve}. In other words, saying that the tetraquadric manifold $X$ admits free quotients by a finite group $\Gamma$ implies a partial fixing of the complex structure of $X$. In due course, when we consider line bundle models on the tetra-quadric, some of the K\"ahler moduli will also be fixed by virtue of the slope zero conditions~\eqref{slope0}.
The tetra-quadric is ``favourable" in the sense that its entire second cohomology is spanned by the K\"ahler forms $J_1,\ldots ,J_4$ of the four $\mathbb{C}\mathbb{P}^1$ factors, restricted to the hypersurface. Its cone of K\"ahler forms $J=\sum_{i=1}^4t^iJ_i$ is given by
\begin{equation}
C_{{\bf t}} =\left\{ {\bf t} \in \mathbb{R}^4\ \left|\ t^i\geq 0,\, 1\leq i\leq 4\right. \right\} \label{kc}
\end{equation}
and the triple intersection numbers have the following simple form
\begin{equation}
d_{ijk} = \int_X J_i\wedge J_j\wedge J_k = \begin{cases} 2 & \mbox{ if } i\neq j, j\neq k \\ 0 &\mbox{ otherwise } \end{cases}\; .\label{tqisec}
\end{equation}
This leads to the volume form
\begin{equation}
\kappa = 12\left( t_1t_2t_3+t_1t_2t_4+t_1t_3t_4+t_2t_3t_4\right)\; .
\end{equation}
The second Chern class of the tangent bundle of the tetra-quadric, in the basis $\{\nu^i\}$ of four-forms dual to $J_i$, is given by
\begin{equation}
c_2(TX)=(24,24,24,24)\; . \label{tqc2}
\end{equation}
The Mori cone corresponds to all positive linear combinations of $\nu^i$.
\subsection{Line bundle models}
To construct line bundle models on the tetra-quadric we follow the general discussion in Section~\ref{sec:lbs}. A~line bundle sum is specified by a $4\times 5$ integer matrix $(k_a^i)$ subject to the constraint~\eqref{c10} to ensure the vanishing of the first Chern class of the bundle. In Appendix~\ref{app:lbtopology} we provide a more comprehensive account of topological identities of interest for line bundles on the tetra-quadric manifold. Here we focus on the most important quantities, starting with the second Chern class which, from Eqs.~\eqref{c2ind} and \eqref{tqisec}, is given by
\begin{equation}
c_{2i}(V)=-2\sum_{a=1}^5(k^2_ak^3_a+k^2_ak^4_a+k^3_ak^4_a,k^1_ak^3_a+k^1_ak^4_a+k^3_ak^4_a,k^1_ak^2_a+k^1_ak^4_a+k^2_ak^4_a,k^1_ak^2_a+k^1_ak^3_a+k^2_ak^3_a) \label{c2}
\end{equation}
relative to the basis $\{\nu^i\}$. Then, from Eq.~\eqref{tqc2}, the anomaly cancellation condition~\eqref{anom} becomes
\begin{equation}
c_{2i}(V)\leq 24\; . \label{tqanom}
\end{equation}
Again, from Eqs.~\eqref{c2ind} and \eqref{tqisec}, the index can be computed as
\begin{equation}
{\rm ind}(V)=2\sum_{a=1}^5\left(k^1_ak^2_ak^3_a+k^1_ak^2_ak^4_a+k^1_ak^3_ak^4_a+k^2_ak^3_ak^4_a\right)\; .
\end{equation}
Defining
\begin{equation}
(\kappa_i)=4(t_2t_3+t_2t_4+t_3t_4,t_1t_3+t_1t_4+t_3t_4,t_1t_1+t_1t_4+t_2t_4,t_1t_2+t_1t_3+t_2t_3)\label{kappai}
\end{equation}
the slope zero conditions~\eqref{slope0} translate into
\begin{equation}
\mu(L_a)=\kappa_ik^i_a\stackrel{!}{=}0\; . \label{tqslope0}
\end{equation}
These conditions need to be satisfied simultaneously for all $a=1,\ldots ,5$ somewhere in the interior of the K\"ahler moduli space, so for moduli values $t^i>0$. Of course, the vanishing of the first Chern class~\eqref{c10} ensures that at most four of these conditions are independent. Indeed, for a non-trivial solution of the Eqs.~\eqref{tqslope0} at most three of these condition can be independent, so a necessary condition for having a solution with vanishing slope is that the matrix $(k_a^i)$ of line bundle integers satsifies
\begin{equation}
{\rm rank}(k_a^i)\leq 3\; . \label{rkcons}
\end{equation}
Of the four additional $U(1)$ symmetries, $4-{\rm rank}(k^i_a)$ are non-anomalous and, hence, have massless gauge bosons at the line bundle locus. From~\eqref{rkcons} this means there is at least one such non-anomalous $U(1)$ symmetry present for models on the tetra-quadric. Such a $U(1)$ symmetry is phenomenologically unwanted and it can be spontaneously broken by giving VEVs to the singlets ${\bf 1}_{a,b}$. This corresponds to moving into the non-Abelian part of the moduli space, something we will explore in detail later on in the paper.
\begin{table}[!h]
\vspace{12pt}
\begin{center}
\begin{tabular}{| c || c | c | c | c | c |}
\hline
\myalign{| c||}{\varstr{21pt}{16pt}$\ \ \ \ |\Gamma| $ $\ \ \ \ $} &
\myalign{m{3.2cm}|}{$\ $ GUT models} &
\myalign{m{3.5cm}|}{ $\ \ \ \ $ no $ \overline{\mathbf{10}}$ multiplets$\ \ \ $ }&
\myalign{m{3.5cm}|}{$\ \ \ \ \ \ \ $ no $ \overline{\mathbf{10}}\,$s and $\ \ \ \ \ \ \ $ at least one $\mathbf{5}-\overline{\mathbf{5}}$ pair} &
$\ \ \ k_{\text{max}}\ \ \ $
\\ \hline\hline
\varstr{14pt}{8pt} 2 & 10 & 10 & 8 & 4 \\
\hline
\varstr{14pt}{8pt} 4 & 58 & 53 & 46 & 5 \\
\hline
\varstr{14pt}{8pt} 8 & 64 & 52 & 36 & 7 \\
\hline
\varstr{14pt}{8pt} 16 & 5 & 5 & 4 & 6 \\
\hline
\end{tabular}
\vskip 0.4cm
\parbox{16.7cm}{\caption{\it\small Statistics on the number of models on the tetraquadric manifold. The second column gives the number of viable GUT models consistent with the phenomenologically required values of the indices, for each group order $|\Gamma|$. The third and fourth columns give the number of such models which, in addition, satisfy the cohomology constraints specified. All models satisfy $|k_a^i|\leq k_{\rm max}$ with the smallest such value $k_{\rm max}$ given in the last column.}\label{tqmodels}}
\end{center}
\vspace{-12pt}
\end{table}
The most basic physical constraints on the spectrum are the ones which can be formulated in terms of the index. Most importantly, we have the three-family constraint ${\rm ind}(V)\stackrel{!}{=}-3|\Gamma|$ for the ${\bf 10}$ multiplets which, for $SU(5)$-bundles, implies that the three-family constraint ${\rm ind}(\wedge^2 V)=-3|\Gamma|$ for the $\overline{\bf 5}$ multiplets is automatically satisfied. In addition, we require that $-3|\Gamma|\leq {\rm ind}(L_a\otimes L_b)\leq 0$ for all $a<b$, where the lower limit is just as to not exceed the three-family bound and the upper limit is to avoid chiral ${\bf 5}_{a,b}$ multiplets.
At the more sophisticated level of cohomology, we should demand the absence of $\overline{\bf 10}$ multiplets, that is, $h^1(X,V^{^*})\stackrel{!}{=}0$ and the presence of at least one ${\bf 5}-\overline{\bf 5}$ pair, that is, $h^1(X,\wedge^2V^{^*})\stackrel{!}{>}0$. These constraints can be checked using the results from Appendix~\ref{app:tqcoh} where we present an explicit formula for computing line bundle cohomology on the tetra-quadric manifold.
In Ref.~\cite{Anderson:2013xka} a powerful algorithm for systematically generating all line bundle models $(k_a^i)$ with entries in the range $|k_a^i|\leq k_{\rm max}$ for a given upper bound $k_{\rm max}$ and selecting the models which satisfy all the above constraints has been outlined. In Ref.~\cite{Anderson:2013xka} this algorithm has been applied to the tetra-quadric, among other manifolds, and physically viable models have been extracted. The results are summarized in Table~\ref{tqmodels}. Altogether, $94$ viable GUT models for the available symmetry orders, $|\Gamma|=2,4,8,16$, are found. All these models correspond to consistent, anomaly-free and supersymmetric $SU(5)\times S(U(1)^5)$ GUT theories which satisfy the three-family constraint, have no $\overline{\bf 10}$ anti-families and at least one ${\bf 5}-\overline{\bf 5}$ pair to account for the Higgs doublets. Upon taking the quotient by $\Gamma$ and including a Wilson line, many of these become models with an MSSM spectrum and we will study one specific such example below. The complete dataset of viable line bundle sums can be accessed here \cite{lbdatabase}. The computational evidence that these models indeed represent the complete set of viable models on the tetra-quadric is presented in Fig.~\ref{SaturationPlots}, where the number of viable models with $|k_a^i|\leq k_{\rm max}$ is shown as a function of $k_{\rm max}$. For all symmetry orders, the number of models saturates at a value of $k_{\rm max}$ below $10$ and remains stable from thereon. In the next section, we will present various approaches to prove finiteness analytically.
\begin{figure}[h!]
\begin{flushright}
$${\includegraphics[width=3.1in]{NoModels2.pdf}}
\hskip 2pt \lower 0pt\hbox{\includegraphics[width=3.1in]{NoModels4.pdf}}
$$
\end{flushright}
\begin{flushright}
$${\includegraphics[width=3.1in]{NoModels8.pdf}}
\hskip 2pt \lower 0pt\hbox{\includegraphics[width=3.1in]{NoModels16.pdf}}
$$
\end{flushright}
\begin{center}
\captionsetup{width=16cm}
\caption{\itshape The plots show the number of line bundle models (before imposing the absence of $\overline{\mathbf{10}}$ multiplets and the existence of $\mathbf{5}-\overline{\mathbf{5}}$ pairs) on the tetraquadric manifold as a function of the maximal line bundle entry in modulus. The four plots correspond, in order from top-left to bottom-right, to $|\Gamma| = 2, 4, 8$ and $16$.}\label{SaturationPlots}
\end{center}
\end{figure}
}
{\setstretch{1.25}
\section{Finiteness of the Class of Line Bundle Models}\label{sec:finiteness}
In this section we would like to discuss the problem of finiteness of poly-stable line bundle sums with fixed total Chern class analytically. Specifically, we would like to consider, for the example of the tetra-quadric manifold, the following claim.\\[0.3cm]
{\bf Claim: } On a given Calabi-Yau three-fold $X$ and for a fixed rank $n$, the number of line bundle sums
\begin{equation}
V=\bigoplus_{a=1}^nL_a\; ,\quad L_a={\cal O}_X({\bf k}_a)\; , \label{lbsn}
\end{equation}
satisfying the following properties is finite:
\begin{itemize}
\item $c_1(V)=0$ or, equivalently, $\sum_{a=1}^n{\bf k}_a=0$.
\item The second Chern classes is constrained by $c_2(TX)-c_2(V)\in $ Mori cone of $X$.
\item All line bundles have vanishing slope, $\mu(L_a)=0$, simultaneously somewhere in the interior of the K\"ahler cone of $X$.
\item ${\rm ind}(V)=C$, where $C$ is a constant (here taken to be $C=-3|\Gamma|$, where $|\Gamma|$ is the order of a freely-acting symmetry $\Gamma$ on $X$).
\end{itemize}
The automated scan reported on in Ref.~\cite{Anderson:2013xka} and, for the case of the tetra-quadric, illustrated in Fig.~\ref{SaturationPlots} provides convincing evidence for the validity of this claim. In fact, as can be seen from those plots, all viable line bundle sums on the tetra-quadric satisfy $|k_a^i|\leq 10$.
Unfortunately, analysing in a straightforward manner the way in which the various above constraints conspire to produce a finite class is untidy. For the tetra-quadric we carry this out explicitly in Appendix~\ref{app:finiteness}. In the next sub-section, we will prove the above claim for the tetra-quadric and the special case of rank two line bundle sums ($n=2$). In Section~\ref{sec:physbound} we propose a transparent argument to derive a bound for line bundle sums of arbitrary rank; however, in order to make this argument feasible, we have to restrict the K\"ahler cone by imposing two constraints motivated from physics. We require that all K\"ahler moduli $t^i$ satisfy $t^i>1$, a constraint linked to the validity of the supergravity approximation, and finiteness of the Calabi-Yau volume, linked to the finiteness of low-energy coupling constants.
\subsection{A workable example: rank two line bundle sums}\label{sec:tqbound}
Rank two line bundle sums with vanishing first Chern class have the form $V = L\oplus L^{^*}$, where $L={\cal O}_X({\bf k})$, so they are described by a single integer vector ${\bf k}$ with four entries $k^i$. We would like to show that the above claim is valid for this specific class of line bundle sums on the tetra-quadric. We begin with the slope zero condition which represents the main technical difficulty and recall that it can be written as
\begin{equation}
\mu(L)=\kappa_i\,k^i\stackrel{!}{=}0\; ,
\end{equation}
where the quantities $\kappa_i=d_{ijk}\,t^j\,t^k$ have been explicitly given in Eq.~\eqref{kappai}. Here, the moduli $t^i$ have to be in the (interior of the) K\"ahler cone, $C_{\bf t}$, defined in Eq.~\ref{kc}. In order to avoid having to solve a quadric in $t^i$ we introduce the new coordinates ${\bf s}\in \mathbb{R}^4$, defined by ${\bf s}=f({\bf t}) = \left( \kappa_1, \kappa_2, \kappa_3, \kappa_4\right) /4$. It turns out, and it is explicitly shown in Appendix~\ref{app:KahlerCone}, that the K\"ahler cone $f(C_{\bf t})$ in these new variables is a dense subset of the cone
\begin{equation} \label{kcs}
C_{{\bf s}} = \left\{ {\bf s}\in\mathbb{R}^4 \ \left|\ {\bf n}_i\cdot {\bf s} \geq 0,\ {\bf e}_i\cdot {\bf s} \geq 0,\ 1\leq i\leq 4\right.\right\}\; .
\end{equation}
with ${\bf n}_i={\bf n}-{\bf e}_i$, ${\bf n}=(1,1,1,1)/2$ and the standard unit vectors ${\bf e}_i$ on $\mathbb{R}^4$.
It is also useful to introduce the cone $\check{C}_{\mathbf k}$, dual to $C_{\mathbf s}$ which is given by the standard definition $\check{C}_{\mathbf k} = \{\, {\mathbf k} \in \mathbb{R}^4 \left | {\mathbf k\cdot \mathbf s} \geq 0, \forall \mathbf s\in C_{\mathbf s} \right. \}$. A straightforward computation shows that it can also be written as
\begin{equation}
\check{C}_{\mathbf k} = \{\, {\mathbf k} \in \mathbb{R}^4 \left | {\mathbf k\cdot \mathbf e}_{ij} \geq 0, \forall i<j \right. \}\; ,
\end{equation}
where ${\mathbf e}_{ij} = {\mathbf e}_i + {\mathbf e}_j$. We can now say that the slope zero condition is satisfied iff the equation ${\bf s}\cdot{\bf k}=0$ has a non-trivial solution in the interior $\mathring{C}_{\mathbf s}$ of the cone ${C}_{\mathbf s}$ and this, in turn, is equivalent to the condition $\mathbf k \notin \check{C}_{\mathbf k} \cup \left( -\check{C}_{\mathbf k}\right)$. Given the structure of the cone $\check{C}_{\mathbf k}$, this means the slope zero condition can be satisfied somewhere in the interior of the K\"ahler cone precisely if the vector ${\bf k}$ has two components $k_i, k_j$ with $k_i + k_j> 0$ and two components $k_l, k_m$ with $k_l + k_m<0$. Thus, up to permutations of the components of ${\bf k}$, the slope zero condition can be satsified iff
\begin{equation}\label{ineq1}
\begin{aligned}
&k_1 + k_2 > 0 \text{ and } k_1 + k_3 < 0 \text{ or } \\
&k_1 + k_2 > 0 \text{ and } k_3 + k_4 < 0
\end{aligned}
\end{equation}
Further, the bound on the second Chern class becomes
\begin{eqnarray}
c_2^i(V)&=&4(k^2k^3+k^2k^4+k^3k^4,k^1k^3+k^1k^4+k^3k^4,k^1k^2+k^1k^4+k^2k^4,k^1k^2+k^1k^3+k^2k^3)\nonumber\\
&\stackrel{!}{\leq}&(24,24,24,24)\; .\label{ineq2}
\end{eqnarray}
It can be shown, for example using Mathematica, that the system of integer inequalities given by Eqs.~(\ref{ineq1}) and (\ref{ineq2}) has a finite number of solutions $\mathbf k = (k_1,k_2,k_3,k_4)$ which all satisfy $-7< k_i <7$. We note that, in arriving at this result, we have not even used the constraint on the index of $V$.
\subsection{A bound rooted in Physics}\label{sec:physbound}
Let us now present a more general finiteness proof which applies to line bundle sums of any rank on the tetra-quadric (and can indeed be applied to other Calabi-Yau manifolds) which, however, requires two additional, physically motivated assumptions.
First recall that the K\"ahler moduli space metric \cite{Candelas:1990pi} for a Calabi-Yau manifold can be written as
\begin{equation}
G_{ij} = \frac{1}{2\,\text{Vol}(X)} \int_X J_i\wedge \star J_j = -3 \left( \frac{\kappa_{ik}}{\kappa} - \frac{2\kappa_i\kappa_j}{3\kappa^2}\right)
\end{equation}
where $\text{Vol}(X)=\kappa/6$ is the Calabi-Yau volume with respect to the Ricci-flat metric, $\kappa = d_{ijk}\, t^i\,t^j\,t^k$, $\kappa_i = d_{ijk}\,t^j\,t^k$ and $\kappa_{ij}=d_{ijk}\,t^k$. The slope zero conditions for a line bundle sum of the form~\eqref{lbsn} can be written as
\begin{equation}
\kappa_ik_a^i=0\; .
\end{equation}
Now consider the sum
\begin{equation}\label{eq:bound1}
\sum_a {\bf k}_a^T G\, {\bf k}_a = -\frac{3}{\kappa} d_{ijk} \sum_a k_a^i\, k_a^j\, t^k = -\frac{6}{\kappa} t^i\, \text{ch}_{2i}(V) \leq \frac{6}{\kappa} |{\bf t}| |\text{ch}_{2i}(V) |\leq \frac{6}{\kappa}|{\bf t}||c_{2i}(TX)|\; .
\end{equation}
Introducing the modified moduli space metric $\widetilde G = \kappa\, G /( 6 |{\bf t}|)$ this means that
\begin{equation}
\sum_a {\bf k}_a^T \widetilde{G}\, {\bf k}_a\leq |c_{2i}(TX)|\; . \label{kcons}
\vspace{-9pt}
\end{equation}
For a fixed K\"ahler class ${\bf t}$ in the interior of the K\"ahler cone, the moduli space metric $G$ and indeed $\widetilde{G}$ are positive definite and, hence, the inequality~\eqref{kcons} constrains the available integer vectors ${\bf k}_a$ to a finite set. This statement applies to all Calabi-Yau three-folds. However, it has a limitation which is relevant for the physics application we are discussing. In physics, we are not interested in fixing the K\"ahler class, that is, different line bundle sums can satisfy the slope zero conditions for different loci in K\"ahler moduli space. In particular, we cannot, by the above argument, exclude a sequence of line bundle sums whose associated slope zero loci approach the boundary of the K\"ahler cone. In such a situation, the eigenvalues of $\widetilde G$ are no longer bounded from below and the above finiteness argument breaks down.
One way to resolve this difficulty is to restrict the ``allowed" region in K\"ahler moduli space, that is, in essence, exclude points close to the boundary. Specifically, what we require is that all $t^i>1$ (assuming the K\"ahler cone is given by $t^i\geq 0$, as is the case for the tetra-quadric) and that $\text{Vol}(X)\lesssim V_{\rm max}$, for a maximal volume $V_{\rm max}$. We are then asking about the number of line bundle sums satisfying all conditions listed in the above claim plus the additional requirement that the slope zero conditions hold in the so-defined portion of K\"ahler moduli space. The physical motivations for these two conditions are the validity of the supergravity approximation (which requires the internal space to be larger than one in string units) and the finiteness of the low-energy coupling constants (specifically finiteness of the gauge couplings and Newton's constant which are related to the volume).
\begin{figure}[h!]
\vspace{12pt}
\begin{center}
\includegraphics[width=4in]{BoundVsVolume.pdf}
\captionsetup{width=16cm}
\caption{\itshape The plot shows the dependence of the bound $48/\lambda_{\text{min}}$ (vertical axis) on the Calabi-Yau volume (horizontal axis). The red points represent the values obtained by numerical methods. The blue curve represents the best fit with a parabola passing through the origin. }\label{BoundVsVolume2}
\end{center}
\end{figure}
This method can be applied to any Calabi-Yau manifold, for which the eigenvalues of $\widetilde{G}$ are bounded from below over the region of K\"ahler moduli space defined above. Let us denote by $\lambda_{\rm min}$ the minimum eigenvalue assumed by $\widetilde{G}$ over the specified portion of K\"ahler moduli space. Clearly, the value of $\lambda_{\rm min}$ depends on the specific Calabi-Yau manifold and the maximal value, $V_{\rm max}$, of the volume. If indeed $\lambda_{\rm min}>0$ then Eq.~\eqref{kcons} leads to the bound
\begin{equation}
\sum_a |{\bf k}_a|^2\leq\frac{|c_2(TX)|}{\lambda_{\rm min}}\; . \label{kbound1}
\end{equation}
Let us carry this out explicitly for the tetra-quadric. In this case, the metric $\widetilde{G}$ is explicitly given by
\begin{equation}
\widetilde G_{ij} = \frac{1}{\sqrt{t_1^2+t_2^2+t_3^2+t_4^2}}\, \left({\displaystyle \sum_{a<b<c} t_a\,t_b\,t_c}\right)^{-1}\, {\displaystyle \sum_{\stackrel{a<b}{a,b\neq i,j}} t_a^2\, t_b^2}
\end{equation}
and the bound~\eqref{kbound1} specializes to
\begin{equation}
\sum_a |{\bf k}_a|^2\leq\frac{48}{\lambda_{\rm min}}\; . \label{kbound2}
\end{equation}
The value of $48/\lambda_{\text{min}}$ as a function of $V_{\rm max}$, determined by a straightforward numerical scan over the relevant portion of the moduli space, has been plotted in Figure~\ref{BoundVsVolume2}. For the explicit models in Table~\ref{tqmodels} whose slope zero locus intersects the region defined by $t^i>1$ and ${\rm Vol}(X)\leq 50$ we find that $\sum_a|{\bf k}_a|^2<50$. Comparing with Fig.~\ref{BoundVsVolume2} this means that, while the bound~\eqref{kbound2} holds, it is actually rather weak and presumably of limited practical use.
}
{\setstretch{1.15}
\section{A line bundle model on the tetra-quadric}\label{sec:model}
In this chapter, we present a specific example taken from the set of phenomenologically viable line bundle models on the tetra-quadric described in the previous section. It is in the context of this model that we will study the question of continuation into the non-Abelian part of the bundle moduli space and the implications for the mass of the Higgs doublets.
\subsection{Definition of the model}
The bundle $V$ for the model in question is given by the sum of the following five line bundles
\begin{equation}\label{lbs}
\begin{array}{lllllllllll}
L_1&=&{\cal O}_X(-1,0,0,1)&,&L_2&=&{\cal O}_X(-1,-3,2,2)&,&L_3&=&{\cal O}_X(0,1,-1,0)\\
L_4&=&{\cal O}_X(1,1,-1,-1)&,&L_5&=&{\cal O}_X(1,1,0,-2)
\end{array}
\end{equation}
so the associated matrix $(k_a^i)$ of line bundle integers reads
\begin{equation}\label{eq:example}
(k_a^i)=
\cicy{ \\ \\ \\ \\ }
{ -1 & -1 & ~~0 & ~~1 & ~~1~ \\
~~0 & -3 & ~~1 & ~~1 & ~~1~ \\
~~0 & ~~2 & -1 & -1 & ~~0 ~\\
~~1 & ~~2 &~~ 0 & -1 & -2 ~\\}\; .
\end{equation}
The rows of this matrix sum up to zero, so clearly we have $c_1(V)=0$, as required. From Eq.~\eqref{c2} we find
\begin{equation}
c_{2i}(V)=(24,8,20,12)
\end{equation}
so that the anomaly constraint~\eqref{tqanom} is satisfied. Further, with ${\rm rank}(k_a^i)=3$, the rank constraint is satisfied and all line bundle slopes~\eqref{tqslope0} are zero on the ray in K\"ahler moduli space where $\kappa_1=\kappa_2=\kappa_3=\kappa_4$ which corresponds to the diagonal $t_1=t_2=t_3=t_4$. Altogether this means we have defined a consistent, supersymmetric GUT model with symmetry $SU(5)\times S(U(1)^5)$. Since ${\rm rank}(k_a^i)=3$ one linear combination of the $U(1)$ symmetries is non-anomalous with a massless vector boson at the Abelian locus. This specific linear combination is $(0,1,2,0,1)$, the non-trivial vector in the kernel of the matrix $(k_a^i)$.
\subsection{The GUT spectrum at the Abelian locus}
The total dimensions of the relevant cohomology groups (computed, e.g.~using the formulae of Appendix~\ref{app:tqcoh}) are given by
\begin{equation}\label{cohV}
\begin{aligned}
h^{^{\!\bullet}}(X,V)\ \, & =\ (0,12,0,0) \\
h^{^{\!\bullet}} (X,\wedge^2V)&=\ (0, 15, 3, 0)\\
h^{^{\!\bullet}} (X,V\otimes V^{^*})&=\ (5, 60, 60, 5)
\end{aligned}
\end{equation}
Hence, we have a total of $12$ GUT families in ${\bf 10}\oplus\overline{\bf 5}$ plus three ${\bf 5}-\overline{\bf 5}$ pairs and a large number of singlet fields. In order to quotient this GUT model to a three-family standard model we need $|\Gamma|=4$ and with $\Gamma=\mathbb{Z}_2\times\mathbb{Z}_2$ we have such a symmetry available on the tetra-quadric. Before we discuss this in detail we should be more precise on how the GUT spectrum is split up into the various line bundle sectors. For this we compute the following relevant line bundle cohomologies
\begin{equation}
\begin{array}{lllllll}
h^{^{\!\bullet}}(X,L_2)&=&(0,8,0,0)&,&h^{^{\!\bullet}}(X,L_5)&=&(0,4,0,0)\\[4pt]
h^{^{\!\bullet}}(X,L_2\otimes L_4)&=&(0,4,0,0)&,&h^{^{\!\bullet}}(X,L_2\otimes L_5)&=&(0,3,3,0)\\[4pt]
h^{^{\!\bullet}}(X,L_4\otimes L_5)&=&(0,8,0,0)&,&h^{^{\!\bullet}}(X,L_1\otimes L_2^*)&=&(0,0,12,0)\\[4pt]
h^{^{\!\bullet}}(X,L_1\otimes L_5^*)&=&(0,0,12,0)&,&h^{^{\!\bullet}}(X,L_2\otimes L_3^*)&=&(0,20,0,0)\\[4pt]
h^{^{\!\bullet}}(X,L_2\otimes L_4^*)&=&(0,12,0,0)&,&h^{^{\!\bullet}}(X,L_3\otimes L_5^*)&=&(0,0,4,0)
\end{array}
\vspace{12pt}
\end{equation}
Here, we have dropped all entirely zero cohomologies. This gives rise to the following spectrum
\begin{equation}
8\, {\bf 10}_2\,,\; 4\,{\bf 10}_5\,,\;4\,\overline{\bf 5}_{2,4}\,,\;3\,\overline{\bf 5}_{2,5}\,,\;8\,\overline{\bf 5}_{4,5}\,,\; 3{\bf 5}_{2,5}\,,\;
12\,{\bf 1}_{2,1}\,,\;12\,{\bf 1}_{5,1}\,,\;20\,{\bf 1}_{2,3}\,,\;12\,{\bf 1}_{2,4}\,,\;4\,{\bf 1}_{5,3}\; .\label{gutspec}
\end{equation}
\subsection{The Standard Model spectrum at the Abelian locus}
The relevant $\Gamma=\mathbb{Z}_2\times\mathbb{Z}_2$ symmetry for the above model is the one whose generators are given by the action of the matrices
\vspace{12pt}
\begin{equation}\label{Z2Z2def}
\left(\begin{array}{rr}1&0\\0&-1\end{array}\right)\; ,\quad \left(\begin{array}{rr}0&1\\1&0\end{array}\right)
\vspace{12pt}
\end{equation}
simultaneously on the coordinates of all four $\mathbb{C}\mathbb{P}^1$ ambient space factors. For an appropriate choice of equivariant structure and Wilson line, forming the quotient by this symmetry leads to the downstairs spectrum~\footnote{Our notation is slightly judicious in that, strictly, we cannot decide at this stage which linear combination of the four available doublets is the down Higgs $H$ and which ones are the three lepton doublets.}
\begin{equation}
2\, {\bf 10}_2\,,\; {\bf 10}_5\,,\;\overline{\bf 5}_{2,4}\,,\;2\,\overline{\bf 5}_{4,5}\,,\; H_{2,5}\,,\;\overline{H}_{2,5}\,,\;
3\,{\bf 1}_{2,1}\,,\;3\,{\bf 1}_{5,1}\,,\;5{\bf 1}_{2,3}\,,\;3\,{\bf 1}_{2,4}\,,\;{\bf 1}_{5,3}\; . \label{smspec}
\end{equation}
Hence, we have precisely three standard model families (which we have listed in GUT notation but should be thought of as being broken up as ${\bf 10}_a\rightarrow (Q_a,u_a,e_a)$ and $\overline{\bf 5}_{a,b}\rightarrow (d_{a,b},L_{a,b})$ into standard model multiplets), one pair of Higgs doublets and $15$ bundle moduli singlets. We note that the $U(1)$ charges are the same for all standard model multiplets originating from the same GUT multiplet and, hence, for the purpose of discussing the implications of $S\left(U(1)^5\right)$ invariance, keeping the GUT notation is adequate.
\subsection{The superpotential}\label{sec:W}
The superpotential for the fields~\eqref{gutspec} is highly constrained by the $S(U(1)^5)$ charges. At the GUT level the only allowed terms, including operators with singlet insertions are
\begin{equation}
W=\lambda_{IJK} {\bf 5}_{2,5}^{(I)}{\bf 10}_2^{(J)}{\bf 10}_5^{(K)}+\rho_{IJK} {\bf 1}_{2,4}^{(I)}\overline{\bf 5}_{4,5}^{(J)}{\bf 5}_{2,5}^{(K)}\; ,
\label{Wgut}
\end{equation}
where the indices $I,J,K\ldots $ run over various ranges, as indicated by the multiplicities in the GUT spectrum~\eqref{gutspec} and $\lambda_{IJK}$ and $\rho_{IJK}$ are arbitrary couplings. At the standard model level, the analogous terms for the spectrum~\eqref{smspec} are
\begin{equation}
W=\lambda_i\overline{H}_{2,5}(Q_2^{(i)}u_5+Q_5u_2^{(i)})+\rho_{\alpha i} {\bf 1}_{2,4}^{(\alpha)}L_{4,5}^{(i)}\overline{H}_{2,5}\; , \label{Wsm}
\end{equation}
where $i=1,2$ labels the two ${\bf 10}_2$ families and the two lepton doublets $L_{4,5}$ from the two $\overline{\bf 5}_{4,5}$ multiplets and $\alpha=1,2,3$ labels the three singlets ${\bf 1}_{2,4}$.
These results have a number of important implications for the structure of the model and its phenomenology. To discuss this, let us focus on the standard model superpotential~\eqref{Wsm} for concreteness, although analogous statements follow for its GUT counterpart~\eqref{Wgut}. The presence of the Yukawa terms means that the up quark mass matrix has rank two and, while a rank one matrix may be preferably at this level, this means a perturbative and generically large top Yukawa coupling is present. The down quark and lepton Yukawa matrices are entirely zero at the perturbative level so, for a realistic model, they would have to be generated non-perturbatively. Further, all operators at dimension four and five which can lead to fast proton decay are forbidden. The point is that, while we certainly do not advertise this model as the one and only standard model from string theory, it does have modestly attractive phenomenological properties and provides a semi-realistic setting for the analysis of the bundle moduli space which we will carry out in the remaining part of the paper.
Specifically, our intention is to explore the moduli space of non-Abelian bundles for which the line bundle sum~\eqref{eq:example} arises as a special locus. From the viewpoint of the four-dimensional effective field theory, the Abelian locus is characterized by the vanishing VEVs of all singlet fields ${\bf 1}_{a,b}$, while switching on such VEVs corresponds to moving away from the Abelian locus into the non-Abelian part of the moduli space. From this point of view, the last term in the superpotential~\eqref{Wsm} for our example model is the most interesting one. At the Abelian locus where, in particular, $\langle {\bf 1}_{2,4}\rangle=0$ this term is simply a coupling. However, for $\langle {\bf 1}_{2,4}\rangle\neq 0$ this term will lead to a mass for the Higgs doublets (or rather for the up Higgs and one linear combination of what we have called lepton doublets) and essentially remove the Higgs from the low-energy spectrum. On the other hand, the spectrum~\eqref{smspec} contains many other singlets which do not appear in the superpotential. A continuation into the non-Abelian part of the moduli space along those singlet directions, while keeping $\langle {\bf 1}_{2,4}\rangle=0$, should leave the Higgs doublets massless. Phrased in terms of the GUT theory, the structure of the superpotential~\eqref{Wgut} suggests that three ${\bf 5}-\overline{\bf 5}$ pairs are removed from the low-energy spectrum if $\langle {\bf 1}_{2,4}\rangle\neq 0$ but that these states remain massless in all parts of the non-Abelian moduli space where $\langle {\bf 1}_{2,4}\rangle=0$. One goal for the remainder of the paper is to verify these statements from a more fundamental viewpoint, that is, by explicitly constructing families of non-Abelian bundles and computing their cohomology.
}
{\setstretch{1.15}
\section{Non-Abelian deformations}\label{sec:monads}
In this section, we would like to discuss two ways of constructing non-Abelian bundles which split into a given line-bundle sum at a specific locus in moduli space and apply these methods to the example presented in the previous section. We will focus on two main bundle constructions, namely extensions and monads.
\subsection{Extensions of line bundle sums}
Extension bundles provide a method of constructing non-Abelian bundles which split into a given set of line bundles
\begin{equation}
V=\bigoplus_{a=1}^nL_a\; .
\end{equation}
Group the line bundles into two sets, indexed by $I\subset \{1,\ldots ,n\}$ and $\bar{I}=\{1,\ldots ,n\}\backslash I$, define the sub-bundles $V_I=\bigoplus_{a\in I}L_a$ and $V_{\bar{I}}=\bigoplus_{a\in\bar{I}}L_a$, and write down the extension sequence
\begin{equation}
0\ \longrightarrow\ V_I\ \longrightarrow\ \widetilde{V}\ \longrightarrow\ V_{\bar{I}}\ \longrightarrow\ 0\; . \label{extseq}
\end{equation}
The moduli space of the extension bundles $\widetilde{V}$ defined by this short exact sequence is given by
\begin{equation}
{\rm Ext}^1(V_{\bar{I}},V_I)\cong H^1(X,V_I\otimes V_{\bar{I}}^{^*})=\bigoplus_{a\in I, b\in{\bar I}}H^1(X,L_a\otimes L_b^{^*})\; . \label{ext}
\end{equation}
The origin of this space corresponds to the split bundle $\widetilde{V}=V_I\oplus V_{\bar{I}}=V$, but away from it $\widetilde{V}$ becomes non-Abelian. Note that the tangent space to the bundle moduli space of $V$ is given by
\begin{equation}
H^1(X,V^{^*}\!\otimes V)=\bigoplus_{a\neq b}H^1(X,L_a\otimes L_b^{^*})
\end{equation}
and, in general, this is larger than~\eqref{ext}. However, if $H^1(X,L_a\otimes L_b^{^*})\neq 0$ only if $a\in I$ and $b\in \bar{I}$ for a suitable choice of $I$ and $\bar{I}$ then the two spaces are indeed the same and the extension sequence captures the full set of non-Abelian deformations. We note that the cohomology $H^1(X,L_a\otimes L_b^{^*})$ contains the low-energy singlets earlier denoted ${\bf 1}_{a,b}$. So, if the $U(1)$ charges $\{1,\ldots ,n\}$ can be split into two disjoint subsets $I$ and $\bar{I}$ such that only singlets ${\bf 1}_{a,b}$ with $a\in I$ and $b\in{\bar I}$ exist then the extension sequence~\eqref{extseq} is ``complete".
We recall from Eq.~\eqref{gutspec} that the singlet spectrum for our tetra-quadric example consists of ${\bf 1}_{2,1}$, ${\bf 1}_{5,1}$, ${\bf 1}_{2,3}$, ${\bf 1}_{2,4}$, ${\bf 1}_{5,3}$. Hence, for $I=\{2,5\}$ and $\bar{I}=\{1,3,4\}$ the above completeness condition is indeed satisfied and the relevant extension sequence for our example reads
\begin{equation}
0\ \longrightarrow\ V_I\ \longrightarrow\ \widetilde{V}\ \longrightarrow V_{\bar{I}}\ \longrightarrow\ 0\; ,\quad V_I=L_2\oplus L_5\; ,\quad V_{\bar{I}}=L_1\oplus L_3\oplus L_4\label{extV}
\end{equation}
The line bundles $L_1,\dots ,L_5$ have been defined in Eq.~\eqref{lbs}.
We would now like to compute the relevant cohomologies of the so-defined extension bundle $\widetilde{V}$ and compare these with the cohomologies of the line bundle sum $V$. For $\widetilde{V}$ we can simply consider the long exact sequence associated to the extension sequence~\eqref{extV}. With $h^{^{\!\bullet}}(V_I)=(0,12,0,0)$ and $h^{^{\!\bullet}}(V_{\bar{I}})=(0,0,0,0)$ this long exact sequence reads
\begin{equation}
\begin{array}{llllll}
&V_I&\longrightarrow& \widetilde{V}&\longrightarrow& V_{\bar{I}}\\[8pt]
h^0(X,\cdot\,)$\ \ \ \ $&0&&0&&0\\[3pt]
h^1(X,\cdot\,)$\ \ \ \ $&12&&12&&0\\[3pt]
h^2(X,\cdot\,)$\ \ \ \ $&0&&0&&0\\[3pt]
h^3(X,\cdot\,)$\ \ \ \ $&0&&0&&0
\end{array}\;
\end{equation}
so that $h^{^{\!\bullet}}(\widetilde{V})=(0,12,0,0)$. This coincides with the result for the cohomology of $V$ in Eq.~\eqref{cohV}. Since the index is unchanged, that is ${\rm ind}(\widetilde{V})={\rm ind}(V)=12$, (and we do not expect new vector-like states to appear in the non-Abelian region) this result is entirely expected. Physically, it means that all the ${\bf 10}$ multiplets which arise at the Abelian locus remain massless when moving into the non-Abelian part of the moduli space.
The situation is considerably more complicated for the ${\bf 5}$ and $\overline{\bf 5}$ multiplets which arise from the cohomology of $\wedge^2\widetilde{V}$. As before the index is unchanged, ${\rm ind}(\wedge^2\widetilde{V})={\rm ind}(\wedge^2V)=12$, so that the twelve $\overline{\bf 5}$ multiplets remain massless. However, the fate of the three vector-like ${\bf 5}-\overline{\bf 5}$ multiplets and, hence, the fate of the Higgs doublets is harder to decide. We begin with the second wedge power sequence
\begin{equation}
0\ \longrightarrow\ \wedge^2V_I\ \longrightarrow\ \wedge^2\widetilde{V}\ \longrightarrow\ \widetilde{V}\otimes V_{\bar{I}}\ \longrightarrow\ S^2V_{\bar{I}}\ \longrightarrow\ 0\; ,
\label{W2}
\end{equation}
associated to the extension sequence~\eqref{extV}. To determine the required cohomology of $\widetilde{V}\otimes V_{\bar{I}}$ we tensor the extension sequence~\eqref{extV} with $V_{\bar{I}}$ and, with $h^{^{\!\bullet}}(V_I\otimes V_{\bar{I}})=(0,12,0,0)$ and $h^{^{\!\bullet}}(V_{\bar{I}}\otimes V_{\bar{I}})=(0,6,6,0)$, this leads to
\begin{equation}
\begin{array}{cccccc}
&V_I\otimes V_{\bar{I}}&\longrightarrow& \widetilde{V}\otimes V_{\bar{I}}&\longrightarrow& V_{\bar{I}}\otimes V_{\bar{I}}\\[8pt]
h^0(X,\cdot\,)$\ \ \ \ $&0&&0&&0\\[3pt]
h^1(X,\cdot\,)$\ \ \ \ $&12&&18&&6\\[3pt]
h^2(X,\cdot\,)$\ \ \ \ $&0&&6&&6\\[3pt]
h^3(X,\cdot\,)$\ \ \ \ $&0&&0&&0
\end{array}\; ,
\end{equation}
This means that $h^{^{\!\bullet}}(\widetilde{V}\otimes V_{\bar{I}})=(0,18,6,0)$ and, in terms of spaces, that
\begin{eqnarray}
H^1(X,\widetilde{V}\otimes V_{\bar{I}})&\cong& H^1(X,V_I\otimes V_{\bar{I}})\oplus H^1(X,V_{\bar{I}}\otimes V_{\bar{I}})\label{coh1}\\[4pt]
H^2(X,\widetilde{V}\otimes V_{\bar{I}})&\cong&H^2(X,V_{\bar{I}}\otimes V_{\bar{I}})\label{coh2}\; .
\end{eqnarray}
We can now split the wedge power sequence~\eqref{W2} into two short exact sequences
\begin{equation}\label{W2split}
\begin{array}{cccccccccccccc}
&\wedge^2V_I&\longrightarrow& \wedge^2\widetilde{V}&\longrightarrow& K&\quad\quad\quad\quad&K&\longrightarrow&\widetilde{V}\otimes V_{\bar{I}}&\longrightarrow&S^2V_{\bar{I}}\\[8pt]
h^0(X,\cdot\,)$\ \ \ \ $&0&&0&&0& &0&&0&&0\\[3pt]
h^1(X,\cdot\,)$\ \ \ \ $&3&&12+c&&12& &12&&18&&6\\[3pt]
h^2(X,\cdot\,)$\ \ \ \ $&3&&c&&0& &0&&6&&6\\[3pt]
h^3(X,\cdot\,)$\ \ \ \ $&0&&0&&0& &0&&0&&0
\end{array}
\end{equation}
where $\delta:H^1(X,K)\rightarrow H^2(X,\wedge^2V_I)$ and $c={\rm dim}\, {\rm Coker}(\delta)$. The key is now to compute the Coker dimension $c$ of this map. For $c=3$ we have three massless ${\bf 5} - \overline{\bf 5}$ pairs while for $c=0$ they have disappeared from the low-energy spectrum. From Eqs.~\eqref{coh1}, \eqref{coh2}, the source and target space for the map $\delta$ can be written as
\begin{eqnarray}
H^1(X,K)&\cong& H^1(X,V_I\otimes V_{\bar{I}})=\bigoplus_{a\in I,b\in\bar{I}}H^1(X,L_a\otimes L_b)\nonumber\\
&=&H^1(X,L_2\otimes L_4)\oplus H^1(X,L_4\otimes L_5)\label{source}\\
H^2(X,\wedge^2V_I)&\cong& H^2(L_2\otimes L_5)
\end{eqnarray}
while the map itself resides in
\begin{equation}
\delta\in H^1(X,(V_I\otimes V_{\bar{I}})^{^*}\!\otimes \wedge^2V_I)=\bigoplus_{a\in I,b\in\bar{I}}H^1(X,L_a^{^*}\!\otimes L_b^{^*}\!\otimes L_2\otimes L_5)\; . \label{map}
\end{equation}
A comparison between Eqs.~\eqref{source} and \eqref{map} shows that only the components $(a,b)= (2,4),(5,4)$ of the map are relevant for the given source space. It turns out that the $(2,4)$ component of the map $H^1(L_5\otimes L_4^{^*})$ vanishes so that the non-trivial part of the map $\delta$ is characterized by
\begin{equation}
\delta:H^1(X,L_4\otimes L_5)\rightarrow H^2(X,L_2\otimes L_5)\quad\mbox{ where }\quad \delta\in H^1(X,L_2\otimes L_4^{^*})\; .
\end{equation}
The crucial observation is that the map only depends on $H^1(X,L_2\otimes L_4^{^*})$ which is precisely the cohomology containing the $12$ bundle moduli singlets earlier denoted by ${\bf 1}_{2,4}$. Hence, if the VEVs of those singlets vanish the map $\delta$ is trivial so that, independently of the other singlet field values, $c={\rm dim}\, {\rm Coker}(\delta)=3$. In this case, from \eqref{W2split}, the cohomology calculation leads to three massless ${\bf 5}-\overline{\bf 5}$ pairs, in accordance with the effective field theory expectation explained in Section~\ref{sec:W}. On the other hand, if the ${\bf 1}_{2,4}$ singlets have non-zero VEVs, the map $\delta$ becomes non-trivial and, as a result, $c<3$. For generic values of the ${\bf 1}_{2,4}$ VEVs the expectation is that $c=0$, so that $h^2(X,\wedge^2\widetilde{V})=3$ and all three ${\bf 5}-\overline{\bf 5}$ pairs are removed from the spectrum. Again, this conforms with the expectation from the four-dimensional effective field theory.
We would now like to explore non-Abelian continuations of line bundle sums using a different method - the monad construction.
}
{\setstretch{1.15}
\subsection{Monads from line bundle sums}
Monad bundles provide a relatively straightforward way to construct bundles with non-Abelian structure groups from line bundle sums. On a Calabi-Yau manifold $X$, define two line bundle sums
\begin{equation}
B = \bigoplus_{\alpha=1}^{\text{rk}(B)}{\cal O}_X({\bf b}_\alpha)\; ,\quad C = \bigoplus_{\mu=1}^{\text{rk}(C)} {\cal O}_X({\bf c}_\mu)
\end{equation}
and the bundle $\widetilde{V}$ by the short exact sequence
\begin{equation} \label{eq:monad}
0\ \longrightarrow\ \widetilde{V} \longrightarrow\ B\ \stackrel{f}{\longrightarrow}\ C\ \longrightarrow\ 0
\end{equation}
so that $\widetilde{V}\cong{\rm Ker}(f)$. Here, $f\in \text{Hom}(B,C)\cong H^0(X,C\otimes B^{^*})$ can be thought off as a ${\rm rk}(C)\times{\rm rk}(B)$ matrix with entries $f_{\mu\alpha}\in H^0(X,{\cal O}_X({\bf c}_\mu- {\bf b}_\alpha)$. Of course, it has to be checked that the above sequence is indeed exact for a given choice of map $f$ which amounts to verifying that ${\rm Im}(f)=C$.
\subsubsection{Monads with split loci}\label{sec:monadconstruction}
Typically, a family of monad maps $f$, parametrized by the coefficients of the polynomials $f_{ai}$, is available leading to a family of monad bundles $\widetilde{V}$. Our task is to construct such a family of monad bundles which splits into a given line bundle sum $V$ at a particular locus in moduli space, that is, for a specific sub-class of maps $f$.
An obvious way to proceed would be to define the line bundle sum $B$ in the monad sequence as $B = V\oplus \widetilde B$ where $V$ is the given line bundle sum and $\widetilde B$ is some other sum of line bundles. If the monad bundle $\widetilde{V}$ splits into $V$ at some locus, it follows that ${\rm ch}(\widetilde{V})={\rm ch}(V)$. This necessary property can be built into the construction by choosing the line bundle sum $C$ in the monad sequence such that $\text{ch}(\widetilde{B}) =\text{ch}(C)$. The monad map then has a block diagonal structure $f=(g,h)$, where $g$ corresponds to the $V$-part of $B$ and $h$ corresponds to $\widetilde B$. At the locus in bundle moduli space where $g=0$, we have $V\subset \text{Ker}(f)$ and this should correspond to the desired split locus. The problem with this construction is that $\text{Ker}(f)$ may not be a vector bundle when $f=(0,h)$. Indeed, for the case of the tetra-quadric, the (quadratic) matrix $h$ degenerates on a co-dimension one locus $\text{det} (h)=0$ in the ambient space, which, generically, intersects the tetra-quadric hypersurface.
\vspace{10pt}
To avoid this problem we will use the following alternative construction. We start with a monad realisation of the structure sheaf
\begin{equation}\label{eq:structure_sheaf}
0\ \longrightarrow\ {\cal O}_X \longrightarrow\ \widetilde{B}_a\ \stackrel{f_a}{\longrightarrow}\ \widetilde{C}_a\ \longrightarrow\ 0
\end{equation}
where $\widetilde{B}_a$ and $\widetilde{C}_a$ are sums of line bundles satisfying
\begin{equation}
\text{rk} (\widetilde{B}_a) = \text{rk}(\widetilde{C}_a) +1 \ \ \ \text{and}\ \ \ c_1(\widetilde{B}_a) = c_1(\widetilde{C}_a)\; .
\end{equation}
On the tetra-quadric, apart from the trivial realisation $\widetilde{B}_a = {\cal O}_X$ and $\widetilde{C}_a = 0$, one can also consider
\begin{equation}
\widetilde{B}_a = {\cal O}_X(0,0,0,p_a)\oplus {\cal O}_X(0,0,0,q_a) \ \ \ \text{and} \ \ \ \widetilde{C}_a = {\cal O}_X(0,0,0,p_a+q_a)
\end{equation}
where $p_a$ and $q_a$ are positive integers. For this choice, the map $f_a = (f_{1,a},f_{2,a})$ contains two polynomials of multi-degrees $(0,0,0,p_a)$ and $(0,0,0,q_a)$. For generic choices of the polynomials, this map has rank one generically. The rank reduces to zero at points in $\mathbb{C}\mathbb{P}^1$ where $f_{1,a}=f_{2,a}=0$ but for sufficiently generic polynomials these equations have no solution in $\mathbb{C}\mathbb{P}^1$. Hence, this indeed provides a monad representation of the structure sheaf. Of course, the integers $p_a$, $q_a$ in $\widetilde{B}_a$ and $\widetilde{C}_a$ can appear in any of the four entries, so that we have a large number of choices on how to represent the structure sheaf as a monad. We can choose the trivial representation or a non-trivial representation characterized by choosing one of the four line bundle components and two integers $p_a$, $q_a$.
\vspace{12pt}
Now consider a given line bundle sum $V=\bigoplus_{a=1}^n L_a$. We can obtain monad representations for the individual line bundles by simply twisting the monad sequence~\eqref{eq:structure_sheaf} with $L_a$. This leads to
\begin{equation}
0\ \longrightarrow\ L_a \longrightarrow\ L_a\otimes \widetilde{B}_a\ \stackrel{f_a}{\longrightarrow}\ L_a\otimes \widetilde{C}_a\ \longrightarrow\ 0\; .
\end{equation}
For the full line bundle sum $V$, we sum these sequences to obtain
\begin{equation}
0\ \longrightarrow\ V \longrightarrow\ B\ \stackrel{f}{\longrightarrow}\ C\ \longrightarrow\ 0
\vspace{-21pt}
\end{equation}
where
\begin{equation} \label{BtCt}
B = \bigoplus_{a=1}^n L_a\otimes \widetilde{B}_a\; ,\quad C = \bigoplus_{a=1}^n L_a\otimes \widetilde{C}_a \; ,\quad f=\text{diag}\left(f_1, \ldots f_n \right)\; .
\end{equation}
We note that for each line bundle, $a$, we can choose $\widetilde{B}_a$ and $\widetilde{C}_a$ independently, from the range of possibilities explained above, so there is significant flexibility in the construction. For the diagonal form of the monad map $f$, as given above, each such choice leads to a monad representation of the original line bundle sum $V$. However, the map $f$ may allow deformations away from this block-diagonal form and then defines a more general class of bundles
\begin{equation}
0\ \longrightarrow\ \widetilde{V} \longrightarrow\ B\ \stackrel{f}{\longrightarrow}\ C\ \longrightarrow\ 0 \label{m1}
\end{equation}
which split into $V$ at the locus where $f$ becomes block-diagonal. Since ${\rm ch}(\tilde{V})={\rm ch}(B)-{\rm ch}(C)={\rm ch}(V)$ the Chern character of the monad bundle $\tilde{V}$ is the same as that of the original line bundle sum $V$. Therefore, if $c_1(V)$ vanishes and the line bundle sum $V$ satisfies the anomaly constraint~\eqref{anom} then the same is true for the monad bundle $\tilde{V}$.
}
{\setstretch{1.15}
\subsubsection{Application to our example}
We would now like to apply the above procedure to our example on the tetra-quadric which was defined by a line bundle sum $V=\bigoplus_{a=1}^5L_a$ characterized by the integers
\begin{equation}\label{eq:example1}
(k_a^i)=
\cicy{ \\ \\ \\ \\ }
{ -1 & -1 & ~~0 & ~~1 & ~~1~ \\
~~0 & -3 & ~~1 & ~~1 & ~~1~ \\
~~0 & ~~2 & -1 & -1 & ~~0 ~\\
~~1 & ~~2 &~~ 0 & -1 & -2 ~\\}\; ,
\end{equation}
where the columns correspond to the line bundles $L_a$, $a=1,\ldots ,5$. To do this, we have to choose, for each $a=1,\ldots ,5$, the line bundle sums $\widetilde{B}_a$ and $\widetilde{C}_a$ which appear in the monad representation~\eqref{eq:structure_sheaf} of the structure sheaf. Our choice is
\begin{equation}
\begin{array}{lllllll}
\widetilde{B}_1&=&{\cal O}_X&,\quad&\widetilde{C}_1&=&0\\
\widetilde{B}_2&=&{\cal O}_X(0,2,0,0)^{\oplus 2}&,\quad&\widetilde{C}_2&=&{\cal O}_X(0,4,0,0)\\
\widetilde{B}_3&=&{\cal O}_X&,\quad&\widetilde{C}_3&=&0\\
\widetilde{B}_4&=&{\cal O}_X&,\quad&\widetilde{C}_4&=&0\\
\widetilde{B}_5&=&{\cal O}_X(0,0,0,2)^{\oplus 2}&,\quad&\widetilde{C}_2&=&{\cal O}_X(0,0,0,4)\\
\end{array}
\end{equation}
From Eq.~\eqref{BtCt} this leads to line bundle sums $B=\bigoplus_{\alpha=1}^7{\cal O}_X({\bf b}_\alpha)$ and $C=\bigoplus_{\mu=1}^2{\cal O}_X({\bf c}_\mu)$ in the monad sequence~\eqref{m1} characterized by the integers
\begin{equation}\label{eq:bandc}
(b_\alpha^i)~=~~
\cicy{ \\ \\ \\ \\ }
{ - 1 & -1 & -1 & ~~0 &~~1& ~~1& ~~1~ \\
~~0 & -1 & -1 & ~~1 &~~1& ~~1& ~~1~ \\
~~0 & ~~2 & ~~2 & -1 &-1 &~~0 &~~0~\\
~~1 & ~~2 &~~ 2 & ~~0 &-1& ~~0 &~~0~\\}\
\hskip0.35in
(c_\mu^i)~=~~
\cicy{ \\ \\ \\ \\ }
{ - 1 & ~~1~ \\
~~1 & ~~1~ \\
~~2 & ~~0 ~\\
~~2 & ~~2~\\}\; .
\end{equation}
The general structure of the monad map is
\begin{equation}\label{monadmap}
f\sim\left(\begin{array}{lllllll}f_{(0,1,2,1)}&f_{(0,2,0,0)}&f_{(0,2,0,0)}'&0&0&0&0\\
f_{(2,1,0,1)}&0&0&f_{(1,0,1,2)}&f_{(0,0,1,3)}&f_{(0,0,0,2)}&f_{(0,0,0,2)}'\end{array}\right)\; ,
\end{equation}
where the subscripts indicate the multi-degrees of the polynomials. For
\begin{equation}\label{splitcond}
f_{(0,1,2,1)}=f_{(2,1,0,1)}=f_{(1,0,1,2)}=f_{(0,0,1,3)}=0
\end{equation}
the map is block-diagonal and $\widetilde{V}$ splits into the original line bundle sums $V$ so the coefficients in those polynomials parametrize the deformations away from the split locus.
It is important to point out that, even though most of our discussion will be carried out on the cover manifold, the line bundle sums $B$, $C$ in Eq.~\eqref{eq:bandc} are equivariant under the $\mathbb{Z}_2\times\mathbb{Z}_2$ symmetry of the tetra-quadric which we have used for our line bundle model and which has been defined in \eqref{Z2Z2def}. This means that, subject to an appropriate restriction of the monad map $f$, the monad bundle $\widetilde{V}$ has a $\mathbb{Z}_2\times\mathbb{Z}_2$ equivariant structure and descends to the quotient manifold.
Before we discuss any further properties, it is crucial to check whether the so-defined bundle $\widetilde{V}$ is superysmmetric, that is, whether it is poly-stable with slope zero. This is certainly the case at the locus where $\widetilde{V}$ splits into the line bundle sum $V$ (provided the K\"ahler moduli are chosen along the diagonal $t_1=t_2=t_3=t_4$ where all line bundle slopes vanish) but away from the split locus this is no longer obvious and has to be checked.
\subsubsection{Stability of Monad Bundles}
Before we consider the monad bundle $\widetilde{V}$, we would like to review the general definition of poly-stability and outline the algorithm for checking it.
First, we define the slope of a coherent sheaf $F$ by
\begin{equation}
\mu_{\bf t}(F) = \frac{1}{\text{rk}(F)} \int_X c_1(F)\wedge J\wedge J = d_{ijk}\, c_1^i(F)\, t^j\, t^k = c_1^i(F)\,\kappa_i
\end{equation}
Note that this definition depends on the K\"ahler class $J=t^iJ_i$. A bundle $\widetilde{V}$ is said to be stable for a K\"ahler class ${\bf t}$ if $\mu_{\bf t}(F)<\mu_{\bf t}(\widetilde{V})$ for all sub-sheafs $F\subset \widetilde{V}$ with $0 < \text{rk}(F) < \text{rk}(\widetilde{V})$. Note that, due to the rank restriction on $F$, a line bundle is automatically stable. Further, a bundle $\widetilde{V}$ is poly-stable if it is a direct sum of stable bundles, $\widetilde{V}=U_1\oplus U_2\oplus\dots\oplus U_n$, all with the same slope, $\mu_{\bf t}(U_1)=\mu_{\bf t}(U_2)=\dots =\mu_{\bf t}(U_n)=\mu_{\bf t}(\widetilde{V})$. The bundle $\widetilde{V}$ is supersymmetric if it is poly-stable and has slope zero.
Frequently, and indeed for our present example, we are dealing with bundles $\widetilde{V}$ with vanishing first Chern class, so that $\mu_{\bf t}(\widetilde{V})=0$ automatically. In this case, poly-stability requires that $\mu_{\bf t}(U_r)=0$, for all $r=1,\ldots ,n$, and that $\mu_{\bf t}(F)<0$ for all sub-sheafs $F\subset U_r$ with $0<{\rm rk}(F)<{\rm rk}(U_r)$.
For a poly-stable bundle $\widetilde{V}$, its dual $\widetilde{V}^{^*}$ and $\widetilde{V} \otimes L$ for any line bundle $L$ are also poly-stable. A stable bundle $\widetilde{V}$ with $\text{rk}(\widetilde{V})>1$ and vanishing slope must satisfy $H^0(X, \widetilde{V}) = H^3(X,\widetilde{V}) = 0$ as well as $H^0(X,\wedge^k V) = H^3(X,\wedge^k V) = 0$.
While the slope of a bundle can be easily computed, finding all coherent sub-sheafs of a given bundle is difficult. A practical algorithm which deals with this complication is as follows. For a sub-sheaf $F$ of $\widetilde{V}$, define $k={\rm rk}(F)$ and the line bundle $L=\wedge^kF$. Then $L$ is a sub-sheaf of $\wedge^k\widetilde{V}$ and $\mu_{\bf t}(L)=k\mu_{\bf t}(F)$. So the region in the K\"ahler cone $C_{\bf t}$ where the bundle $\widetilde{V}$ (with $c_1(\widetilde{V})=0$) is supersymmetric is
\begin{equation}
C_{\widetilde{V}}=\{{\bf t}\in C_{\rm t}\,|\, \mu_{\bf t}(L)\leq 0\mbox{ for all line bundles }L\mbox{ which inject into }\wedge^k\widetilde{V}\; ,k=1,\ldots,{\rm rk}(\widetilde{V})-1\}\; . \label{CV}
\end{equation}
Note that, while stability requires $\mu_{\bf t}(L)<0$, we have used $\mu_{\bf t}(L)\leq 0$ in the above definition. In fact, at a locus in K\"ahler moduli space where $\mu_{\bf t}(L)=0$ the bundle splits into a direct sum and is no longer stable but poly-stable and, hence, still supersymmetric. In Ref.~\cite{Anderson:2009sw} this has been referred to as a stability wall. In the present context, this is of course precisely what we expect to happen when the bundle $\widetilde{V}$ splits into a line bundle sum $V$.
The detailed procedure to find the supersymmetric region based on Eq.~\eqref{CV} has been developed in \cite{Anderson:2008uw, Anderson:2008ex} and summarised in Appendix~\ref{sec:stab_criteria}.
\subsubsection{Checking stability for our example}\label{sec:stabex}
In order to compute the supersymmetric region~\eqref{CV} for our bundle $\widetilde{V}$ on the tetra-quadric, it is useful to describe the K\"ahler cone in terms of the variables $s_i=\kappa_i=d_{ijk}\,t^j\,t^k$. We have seen in Eq.~\eqref{kcs} that, in those variables, the K\"ahler cone is
\begin{equation}
C_{\bf s}=\{\mathbf{s}\in \mathbb R ^4 \ | \ \mathbf{s.e}_i\geq 0 \text{ and } \mathbf{s.n}_i\geq 0 \}\; .
\end{equation}
Then, for a line bundle $L={\cal O}_X(-{\bf k})$, the slope is simply given by the dot product $\mu_{\bf s}(L)=-{\bf s}\cdot{\bf k}$. The supersymmetric region, expressed in terms of the ${\bf s}$ coordinates, can, therefore be written as
\begin{equation}
C_{\widetilde{V}}=\{\mathbf{s}\in C_{\bf s}\, |\, \mathbf{s}.\mathbf{k} \geq 0 \text{ for any } L = \mathcal{O}(-\mathbf{k}),\, L \text{ injects into } \wedge^k \widetilde{V},\ k = 1, \ldots , \text{rk}(\widetilde{V} )-1 \} \; .
\end{equation}
Note that this is an intersection of hyperplanes so the supersymmetric region (written in the ${\bf s}$ coordinates) forms a cone.
We now need to find all line bundles, $L$, on the tetra-quadric which inject into some wedge power of the monad bundle $\widetilde{V}$ defined by Eqs.~\eqref{eq:bandc}. Appendix~\ref{sec:stab_criteria} sets out a number of simple sufficient conditions, based on computing cohomology dimensions only, for line bundles to inject or not to inject. Specifically, these criteria are given in (\ref{eq:cohcondiiton1}), (\ref{eq:cohcondiiton2}), (\ref{eq:cohcondiiton3}), (\ref{eq:cohcondiiton4}), (\ref{eq:cohcondiiton5}). Carrying out a scan over all line bundles with entries in the range from $-3$ to $3$ we collect all line bundles which definitely inject into some power of $\widetilde{V}$, according to our criteria. These line bundles reduce the supersymmetric region to the following:
\begin{equation}\label{eq:hyperplane}
C_{\widetilde{V}}=\{\mathbf{s}\in C_{\bf s}\, |\, \ \mathbf{s.}(1,1,-1,-1)= 0 \text{ and } \mathbf{s.}(-1,1,0,0)\geq 0 \text{ and } \mathbf{s.}(1,1,-2,0)\geq 0 \}\; .
\end{equation}
Increasing the range of line bundles integers further did not change this region of stability. For some line bundles the simple numerical criteria in Appendix~\ref{sec:stab_criteria} are not sufficient to decide whether or not they inject. In such cases, we have to compute ranks of relevant maps which is more involved. We will return to this problem in Section~\ref{sec:stability2} but we note here that, as it turns out, these line bundles do not change the result~\eqref{eq:hyperplane}.
The first condition in Eq.~\eqref{eq:hyperplane} comes from the fact that the line bundle $L_4={\cal O}_X(1,1,-1,-1)$ injects into $\widetilde{V}$, while its dual $L_4^{^*}$ injects into $\wedge^4 V\cong V^{^*}$. This means that we are confined to the hyperplace $s_1+s_2=s_3+s_4$ in K\"ahler moduli space and that we have a poly-stable split bundle $\widetilde{V}=U\oplus L_4$ which corresponds to the locus in bundle moduli space where the entry $f_{(0,0,1,3)}$ in the monad map~\eqref{monadmap} vanishes. The locus in bundle moduli space where equality holds for all three conditions in \eqref{eq:hyperplane}, which corresponds to $t_1=t_2=t_3=t_4$, is precisely where the conditions~\eqref{splitcond} have to be satisfied and the bundle fully splits into the line bundle sum~\eqref{lbs}. Note that $L_4$ is indeed one of the line bundles defining the original line bundle sum~\eqref{lbs}. Therefore, in its supersymmetric moduli space, $\widetilde{V}$ does not provide a fully non-Abelian version of this line bundle sum but only ``connects" four of the five line bundles. In terms of the effective field theory language, the locus in bundle moduli space where $L_4$ remains split off corresponds to vanishing VEVs for all ${\bf 1}_{a,b}$ which carry a $4$--index. From the spectrum~\eqref{gutspec}, these are precisely the singlets ${\bf 1}_{2,4}$. Our earlier arguments based on the effective field theory suggested that the three vector-like ${\bf 5}-\overline{\bf 5}$ multiplets (which give rise to the Higgs doublets) remain massless whenever $\langle {\bf 1}_{2,4}\rangle =0$ and we have verified this expectation explicitly by computing the cohomology of the extension bundles. Here we see that $\langle {\bf 1}_{2,4}\rangle =0$ in the entire supersymmetric moduli space of our monad bundle $\widetilde{V}$ so we expect three massless ${\bf 5}-\overline{\bf 5}$ pairs from the cohomology of~$\wedge^2\widetilde{V}$. We will now verify this expectation by an explicit compution.
\subsubsection{The Spectrum}\label{sec:monadspectrum}
In this subsection we will discuss the spectrum of the $S(U(4)\times U(1))$ compactification with the monad bundle $\widetilde{V}=U\oplus L_4$. The low-energy theory has a gauge group $SU(5)$ and an additional $U(1)$ symmetry, which is massive and, hence, global. As discussed in the next subsection, this additional symmetry, henceforth referred to as $U_X(1)$, combined with the hypercharge leads to the well-known $B-L$ symmetry.
\vspace{12pt}
We will now compute the cohomology of the bundle $\widetilde{V}$. We begin by writing down the long exact sequence
\begin{equation}
\begin{array}{cccccc}
&\widetilde{V}&\longrightarrow& B&\longrightarrow& C\\[8pt]
h^0(X,\cdot\,)$\ \ \ \ $&h^0(X,\widetilde{V})&&8&&12\\[3pt]
h^1(X,\cdot\,)$\ \ \ \ $&h^1(X,\widetilde{V})&&8&&0\\[3pt]
h^2(X,\cdot\,)$\ \ \ \ $&h^2(X,\widetilde{V})&&0&&0\\[3pt]
h^3(X,\cdot\,)$\ \ \ \ $&h^3(X,\widetilde{V})&&0&&0
\end{array}\;
\label{e1.1}
\vspace{12pt}
\end{equation}
associated to the monad sequence~\eqref{eq:monad}. Using an explicit representation for the cohomology groups $H^0(X,B)$ and $H^0(X,C)$, as well as for the map between these groups induced by the monad map, it follows that~\footnote{These results also follow from $h^0(X,\widetilde{V})=0$ which is a consequence of the poly-stability of $\widetilde{V}$.}
\begin{equation}
\begin{aligned}
\text{dim}\, \text{Ker} & \left( H^0(X,B) \longrightarrow H^0(X,C) \right) = 0\\
\text{dim} \, \text{Coker} & \left( H^0(X,B) \longrightarrow H^0(X,C) \right) = 4\; .
\end{aligned}
\label{e1.2}
\vspace{12pt}
\end{equation}
This implies
\begin{equation}
h^{^{\!\bullet}}(X,V)=(0,12,0,0)\;,
\label{e1.3}
\end{equation}
and, hence, a spectrum with $12$ ${\bf 10}$ multiplets and no $\overline{\bf 10}$ mutliplets, the same as we found for the original line bundle sum $V$ in Eq.~\eqref{cohV}. Of course, this does not come as a surprise, since these $12$ chiral multiplets are protected by the index. Also note that $h^1 (X, \widetilde{V})= h^1(X, U)+ h^1 (X, L_4)$. Since $h^1 (X, L_4)=0$ \footnote{In fact the line bundle $L_4= {\cal O}_X (1, 1, -1, -1)$ has entirely vanishing
cohomologies.} we get $h^1 (X, \widetilde{V})= h^1(X, U)=12$, so the multiplets transforming as ${\bf 10}$ come from the cohomology group $H^1(X, U)$.
\vspace{21pt}
It is considerably more interesting -- and difficult -- to analyze the spectrum of ${\bf 5}-\overline{\bf 5}$ multiplets which follows from $\wedge^2\widetilde{V}$. To do this, we explicitly focus on the supersymmetric moduli space where the bundle splits as $\widetilde{V}=U\oplus L_4$, and $U$ is described by the monad
\begin{equation}\label{eq:monadseqnew}
0\ \longrightarrow\ U \longrightarrow\ \widetilde{B}\ \stackrel{\tilde{f}}{\longrightarrow}\ C\ \longrightarrow\ 0 \; ,
\end{equation}
with the line bundle sums $\widetilde{B}$ and $C$ defined by the bundle integers
\vspace{12pt}
\begin{equation}\label{eq:bandc2}
(\tilde{b}_\alpha^i)~=~~
\cicy{ \\ \\ \\ \\ }
{ - 1 & -1 & -1 & ~~0 & ~~1& ~~1~ \\
~~0 & -1 & -1 & ~~1& ~~1& ~~1~ \\
~~0 & ~~2 & ~~2 & -1 &~~0 &~~0~\\
~~1 & ~~2 &~~ 2 & ~~0 & ~~0 &~~0~\\}\
\hskip0.35in
(c_\mu^i)~=~~
\cicy{ \\ \\ \\ \\ }
{ - 1 & ~~1~ \\
~~1 & ~~1~ \\
~~2 & ~~0 ~\\
~~2 & ~~2~\\}\; .
\vspace{14pt}
\end{equation}
Note these are the same line bundle sums as in the original monad~\eqref{eq:bandc} except that we have removed the fifth column from $(b_\alpha^i)$ which corresponds to the line bundle $L_4={\cal O}_X(1,1,-1,-1)$. The corresponding monad map $\tilde{f}$ is obtained by likewise removing the fifth column from the original monad map $f$ in Eq.~\eqref{monadmap} which results in
\begin{equation}\label{monadmap1}
f\sim\left(\begin{array}{llllll}f_{(0,1,2,1)}&f_{(0,2,0,0)}&f_{(0,2,0,0)}'&0&0&0\\
f_{(2,1,0,1)}&0&0&f_{(1,0,1,2)}&f_{(0,0,0,2)}&f_{(0,0,0,2)}'\end{array}\right)\; ,
\vspace{12pt}
\end{equation}
The locus where $U$ splits into the line bundle sum $L_1\oplus L_2\oplus L_3\oplus L_5$ is $t_1=t_2=t_3=t_4$ in K\"ahler moduli space and
\begin{equation}
f_{(0,1,2,1)}=f_{(2,1,0,1)}=f_{(1,0,1,2)}=0
\end{equation}
in bundle moduli space.
\vspace{31pt}
We need to compute the cohomology of $\wedge^2\widetilde{V} = \left(L_4\otimes U\right) \, \oplus\, \wedge^2 U$. The cohomology of $L_4\otimes U$ can be easily obtained by twisting the monad sequence (\ref{eq:monadseqnew}) with $L_4$:
\begin{equation}
0\ \longrightarrow\ L_4\otimes U \longrightarrow\ L_4\otimes \widetilde{B}\ \longrightarrow\ L_4\otimes C\ \longrightarrow\ 0 \,.
\label{e1.4}
\end{equation}
From the associated long exact sequence, together with $h^{^{\!\bullet}}(X,L_4\otimes\widetilde{B})=(8,8,0,0)$ and $h^{^{\!\bullet}}(X,L_4\otimes C)=(12,0,0,0)$, it follows that
\begin{equation}
h^{^{\!\bullet}}(X,L_4\otimes U)\ \, =\ (0,12,0,0)
\label{e1.5}
\vspace{12pt}
\end{equation}
so this part of $\wedge^2\widetilde{V}$ takes care of the chiral asymmetry. The fate of the Higgs doublets is determined entirely by $\wedge^2U$ and in order to compute this part we need to use the second exterior power of the monad sequence~(\ref{eq:monadseqnew}):
\vspace{10pt}
\begin{equation}
0\ \longrightarrow\ \wedge^2U \longrightarrow\ \wedge^2\widetilde{B}\ \longrightarrow\ \widetilde{B}\otimes C\ \longrightarrow\ S^2C\ \longrightarrow\ 0
\vspace{31pt}
\end{equation}
As usual, this can be split in two short exact sequences, whose associated long exact sequences read:
\vspace{10pt}
\begin{equation}\label{eq:2sequences}
\begin{array}{cccccccccccccc}
&\wedge^2U&\longrightarrow& \wedge^2\widetilde{B}&\longrightarrow& Q&\quad\quad\quad\quad&Q&\longrightarrow&\widetilde{B}\otimes C&\longrightarrow&S^2C\\[8pt]
h^0(X,\cdot\,)$\ \ \ \ $&h^0(\wedge^2U)&&53&&h^0(Q)&& h^0(Q)&&150&&96\\[3pt]
h^1(X,\cdot\,)$\ \ \ \ $& h^1(\wedge^2U) &&85&&h^1(Q)&&h^1(Q)&&134&&48\\[3pt]
h^2(X,\cdot\,)$\ \ \ \ $&h^2(\wedge^2U)&&0&&h^2(Q)&& h^2(Q)&&0&&0\\[3pt]
h^3(X,\cdot\,)$\ \ \ \ $&h^3(\wedge^2U)&&0&&h^3(Q)&& h^3(Q)&&0&&0
\end{array}
\end{equation}
The second short exact sequence implies that
\begin{equation}
\begin{aligned}
H^0(X,Q) & \cong \text{Ker}\left( H^0(X,\widetilde{B}\otimes C) \longrightarrow H^0(X,S^2C) \right)\\
H^1(X,Q) & \cong \text{Coker} \left( H^0(X,\widetilde{B}\otimes C) \longrightarrow H^0(X,S^2C) \right) \\
& \oplus \text{Ker}\left( H^1(X,\widetilde{B}\otimes C) \longrightarrow H^1(X,S^2C) \right) \\
H^2(X,Q) & \cong \text{Coker}\left( H^1(X,\widetilde{B}\otimes C) \longrightarrow H^1(X,S^2C) \right) \\
H^3(X,Q) & \cong 0
\end{aligned}
\end{equation}
The computation of these cohomology groups proceeds in several stages. In the first step, we need to find the map between the line bundle sums $\widetilde{B}\otimes C$ and $S^2C$ induced by the monad map (\ref{eq:monadseqnew}). In the second step, we construct the induced map between the various cohomology groups and compute their ranks using the CICY package \cite{CICYpackage}. A comprehensive exposition on the computation of cohomology groups and of ranks of maps goes beyond the scope of the present paper. The interested reader can find in Ref.~\cite{Anderson:2013qca} an outline of the basic techniques for computing line bundle cohomology on complete intersection Calabi-Yau manifolds in products of projective spaces.
\vspace{12pt}
We find that the map $H^0(X,\widetilde{B}\otimes C) \longrightarrow H^0(X,S^2C)$ has rank $94$, while the map between $H^1(X,\widetilde{B}\otimes C) \longrightarrow H^1(X,S^2C)$ has maximal rank $48$. This leads to
\begin{equation}
h^{^{\!\bullet}}(X,Q)\ \, =\ (56,88,0,0)
\end{equation}
The final step consists in determining the cohomology of $\wedge^2 U$. The first long exact sequence in cohomology in (\ref{eq:2sequences}) implies:
\begin{equation}
\begin{aligned}
H^0(X,\wedge^2 U) & \cong \text{Ker}\left( H^0(X,\wedge^2\widetilde{B}) \longrightarrow H^0(X,Q) \right)\\
H^1(X,\wedge^2 U) & \cong \text{Coker} \left( H^0(X,\wedge^2\widetilde{B}) \longrightarrow H^0(X,Q) \right) \\
& \oplus \text{Ker}\left( H^1(X,\wedge^2\widetilde{B}) \longrightarrow H^1(X,Q) \right) \\
H^2(X,\wedge^2 U) & \cong \text{Coker}\left( H^1(X,\wedge^2\widetilde{B}) \longrightarrow H^1(X,Q) \right) \\
H^3(X,\wedge^2U) & \cong 0
\end{aligned}
\end{equation}
Computing these maps involves several layers of additional complication. To start with, the map $H^0(X,\wedge^2\widetilde{B}) \longrightarrow H^0(X,Q)$ is induced by the bundle map $\wedge^2\widetilde{B} \longrightarrow Q$ which itself has to be determined from the monad map (\ref{eq:monadseqnew}). However, since
\begin{equation}
H^0(X,Q) \cong \text{Ker}\left( H^0(X,\widetilde{B}\otimes C) \longrightarrow H^0(X,S^2C) \right)
\end{equation}
it follows that $H^0(X,Q)$ is a subspace of $H^0(X,\widetilde{B}\otimes C)$ and thus the map $H^0(X,\wedge^2\widetilde{B}) \longrightarrow H^0(X,Q)$ is equivalent with the map $H^0(X,\wedge^2\widetilde{B}) \longrightarrow H^0(X,\widetilde{B}\otimes C)$, with the single difference that, for the latter, the target space is larger. Computing the rank of $H^0(X,\wedge^2\widetilde{B}) \longrightarrow H^0(X,Q)$, we obtain 53. This leaves us with the following tableaux of dimensions
\begin{equation}
\begin{array}{ccccccccc}
0 \ \ & \longrightarrow & \wedge^2 U & \longrightarrow &\wedge^2 \widetilde{B} & \longrightarrow& Q & \longrightarrow &0\ \ \ \ \\ \\[-12pt]
& & 0 &&53&& 56 && \\
& & 3+K &&85&& 88 & & \\
& & C &&0&&0&& \\
& & 0 &&0&&0&& \\
\end{array}
\end{equation}
where $K = \text{dim}\, \text{Ker}\left( H^1(X,\wedge^2\widetilde{B}) \longrightarrow H^1(X,Q) \right)$ and $C = \text{dim}\, \text{Coker}\left( H^1(X,\wedge^2\widetilde{B}) \longrightarrow H^1(X,Q) \right)$. From exactness, it follows that $C = 3 + K$. Computing the map $H^1(X,\wedge^2\widetilde{B}) \longrightarrow H^1(X,Q)$ would require the knowledge of a co-boundary map and we would have to deal with the additional complication that the target space $H^1(X,Q)$ is a direct sum of $H^0(X,S^2C)$ and $H^1(X, \widetilde{B}\otimes C)$. Fortunately, for the present case, the information acquired so far is enough to make an important statement. We have obtained that
\begin{equation}
h^{^{\!\bullet}}(X,\wedge^2 \widetilde{V}) \ \, = \ h^{^{\!\bullet}}(X, L_4\otimes U) \ +\ h^{^{\!\bullet}}(X,\wedge^2 U) \ \, =\ (0,15+K,3+K,0)
\end{equation}
where $K\geq 0$. We recall that the corresponding cohomology at the split locus (where $\widetilde{V}$ splits into the line bundle sum $V$ in \eqref{lbs}) is given by $h^{^{\!\bullet}}(X,V)=(0,15,3,0)$. This means that the three massless, vector-like ${\bf 5}-\overline{\bf 5}$ states present at the split locus remain massless throughout the supersymmetric moduli space of $\widetilde{V}$, as expected from our low-energy arguments.
In principle, the above cohomology calculation still allows $K>0$. However, from the viewpoint of the effective field theory, this is not expected. Indeed, as the move away from the split locus by switching on singlet VEVs we may generate mass terms for vector-like pairs, thereby reducing their number, but we do not expect this number to increase. Thus we conclude that $K=0$.
\vspace{12pt}
In summary, our result guarantees the presence of massless ${\bf 5}-\overline{\bf 5}$ multiplets (resulting in Higgs doublets in the downstairs theory) for non-Abelian deformations of the original line bundle sum $V$ which are of the form $\widetilde{V}=L_4\oplus U$. This result is in perfect agreement with the expectation from the four-dimensional effective theory and the earlier cohomology computation in the context of extension bundles.
Finally, the spectrum also contains singlets (vector bundle moduli), which correspond to elements of the cohomology group
\begin{equation}
H^1(X,\text{End}(\widetilde V)) = H^1(X, \widetilde V \otimes \widetilde V^{^*}) = H^1(X, U \otimes U^{^*}) \oplus H^1(X, L_4 \otimes U^{^*}) \oplus H^1(X, U \otimes L_4^{^*})
\end{equation}
The cohomology of $U\otimes L_4^{^*}$ can be easily obtained by twisting the monad sequence (\ref{eq:monadseqnew}) with $L_4^{^*}$. From the associated long exact sequence, together with the information $h^{^{\!\bullet}}(X,\widetilde{B}\otimes L_4^{^*})=(8,40,0,0)$ and $h^{^{\!\bullet}}(X, C\otimes L_4^{^*})=(8,28,0,0)$, it follows, after computing the necessary ranks of cohomology maps, that
\begin{equation}
h^{^{\!\bullet}}(X, U\otimes L_4^{^*})\ \, =\ (0,12,0,0)
\end{equation}
Thus the number of singlets is given by:
\begin{equation}
h^1(X, \widetilde V\otimes \widetilde V^{^*})\ \, =\ 12\, +\, h^1(X, U\otimes U^{^*})\label{VUrel}
\end{equation}
The cohomology of $U\otimes U^{^*}$ can be computed using a web of six short exact sequences. Unfortunately, this computation runs into the same difficulties as encountered in the computation of $\mathbf{5}-\overline{\mathbf{5}}$ pairs discussed above and we will not carry it out explicitly. However, there is an expectation for this cohomology from low-energy arguments. As we move away from the Abelian locus, three of the four $U(1)$ symmetries are broken spontaneously and, as a result, three bundle moduli should acquire a mass. With $60$ bundle moduli at the Abelian locus we, therefore, expect that $h^1(X, \widetilde V\otimes \widetilde V^{^*})=57$ and, from Eq.~\eqref{VUrel}, that $h^1(X, U\otimes U^{^*})=45$.
As explained in the next section, the 12 moduli coming from $ h^1(X, U\otimes L_4^{^*})$ are charged under the extra $U(1)$ symmetry. As such, a non-trivial vacuum value for these singlets breaks the $U_X(1)$ and corresponds to moving away from the $S(U(4) \times U_X(1))$ locus into the moduli space of generic $SU(5)$ bundles. These are the moduli ${\bf 1}_{2,4}$ in the spectrum~\eqref{gutspec}. The remaining singlets from $ H^1(X, U\otimes U^{^*})$ are uncharged under $U_X(1)$.
\subsubsection{The $U_X(1)$ symmetry and $B-L$}
In the present section we will show that the additional global $U_X(1)$ symmetry, combined with $U_Y(1)$ hypercharge, leads to the well-known $B-L$ symmetry. For this, we need to compute the $U_X(1)$ charges of the various multiplets in the low-energy GUT. We do this by considering the sequence of embeddings $S\left(U(4)\times U_X(1) \right) \subset SU(5)\subset E_8$. The matter multiplets can be obtained by decomposing the adjoint $\mathbf{248}_{E_8}$ of $E_8$ under the $SU(5)\times U_X(1)$ sub-group. The corresponding branching rule has been discussed in Ref.~\cite{Anderson:2012yf}, which we review below.
The $U_X(1)$ charges of the GUT multiplets can be represented by vectors $\mathbf{q}$. Due to the determinant condition, two such vectors $\mathbf{q}$ and $\mathbf{\widetilde q}$ have to be identified if $\mathbf{q}-\widetilde{\mathbf{q}} = \mathbb{Z} \mathbf{n}$, where $\mathbf{n}=(4,1)$. We summarise in Table~\ref{spectrum2} the resulting spectrum. The actual $U_X(1)$ charges can be recovered from this description by multiplying a charge vector $\mathbf{q}=(q_1,q_2)$ with $(-1,4)$.
\vspace{12pt}
\begin{table}[h]
\hspace{.25cm}
\parbox{.45\linewidth}{
\centering
\begin{tabular}{| c | c | r |}
\hline
\varstr{14pt}{9pt} repr. & cohomology & $U_X(1)$ charge \\ \hline\hline
\varstr{14pt}{9pt} $~{\bf 1}_{0}$ & $H^1(X, U \otimes U^{^*})$ & 0~~~~~~~ \\ \hline
\varstr{14pt}{9pt} $~{\bf 1}_{\mathbf{e}_1-\mathbf{e}_2}$ & $H^1(X, U \otimes L^{^*})$ & -5~~~~~~~ \\ \hline
\varstr{14pt}{9pt} $~{\bf 1}_{-\mathbf{e}_1+\mathbf{e}_2}$ & $H^1(X, L \otimes U^{^*})$ & 5~~~~~~~ \\ \hline
\varstr{14pt}{9pt} $~{\bf 5}_{-2\mathbf{e}_1}$ & $H^1(X, \wedge^2 U^{^*})$ & 2 ~~~~~~ \\ \hline
\varstr{14pt}{9pt} $~{\bf 5}_{-\mathbf{e}_1-\mathbf{e}_2}$ & $H^1(X, U^{^*}\otimes L^{^*})$ & -3 ~~~~~~ \\ \hline
\end{tabular}
}
\hspace{.5cm}
\parbox{.45\linewidth}{
\centering
\begin{tabular}{| c | c | r |}
\hline
\varstr{14pt}{9pt} $~~~{\bf \overline{5}}_{2\mathbf{e}_1}~~$ & $H^1(X, \wedge^2 U)$ & ~~~~~ -2~~~~~~ \\ \hline
\varstr{14pt}{9pt} $~{\bf \overline{5}}_{\mathbf{e}_1+\mathbf{e}_2}$ & $~H^1(X, U\otimes L)~$ & 3~~~~~~ \\ \hline
\varstr{14pt}{9pt} $~{\bf 10}_{\mathbf{e}_1}$ &$H^1(X, U)$ & -1~~~~~~ \\ \hline
\varstr{14pt}{9pt} $~{\bf 10}_{\mathbf{e}_2}$ &$H^1(X, L)$ & 4~~~~~~ \\ \hline
\varstr{14pt}{9pt} $~{\bf \overline{10}}_{-\mathbf{e}_1}$ & $H^1(X, U^{^*})$ & 1~~~~~~ \\ \hline
\varstr{14pt}{9pt} $~{\bf \overline{10}}_{-\mathbf{e}_2}$ & $H^1(X, L^{^*})$ & -4~~~~~~
\\ \hline
\end{tabular}
}\vspace{10pt}
\begin{center}
\parbox{15cm}{\caption{\it\small The spectrum of $SU(5)$ GUT models derived from heterotic models with $S\left( U(4)\times U(1)\right)$ bundles $U\oplus L$. Here $\mathbf{e}_1$ and $\mathbf{e}_2$ are the standard unit vectors in two dimensions. The third column has been obtained by projecting the charges $\mathbf{q}$ along the line $(-1,4)$.}\label{spectrum2}}
\end{center}
\vspace{-21pt}
\end{table}
Looking at the dimensions of the various cohomology groups, as computed in the previous section, we obtain the following spectrum:
\vspace{-4pt}
\begin{equation}
{12\,\bf{10}}_{-1}\; ,\quad 12\,\overline{\bf 5}_{3}\;,\quad 3\,\overline{\bf 5}_{-2}\;,\quad 3\, {\bf 5}_{2}\; ,\quad 12\,{\bf 1}_{-5}\; ,\quad h^1(X, U\otimes U^{^*}) \,{\bf 1}_{0}
\vspace{-4pt}
\end{equation}
It is well-known, in the context of regular $SO(10)$ GUTs, that the $U(1)$ factor in the standard decomposition $SO(10)\rightarrow SU(5)\times U(1)$ leads to $B-L$ with $SO(10)$ multiplets branching as
\vspace{-4pt}
\begin{equation}
{\bf 16}_{SO(10)}= {\bf 1}_{-5} + {\bar {\bf 5}}_3 + {\bf 10}_{-1}\; ,\quad {\bf 10}_{SO(10)}= {\bf 5}_{2} + {\overline {\bf 5}}_{-2}\,.
\label{e6}
\vspace{-4pt}
\end{equation}
The observation is that these $U(1)$ charges precisely coincide with the above $U_X(1)$ charges in our model. Explicitly, $B-L$ is given by the following combination
\vspace{-4pt}
\begin{equation}
B -L \ =\ -\frac{1}{5} X\, +\, \frac{2}{5} Y
\vspace{-4pt}
\end{equation}
of $U_X(1)$ and hypercharge. The absence of operators leading to fast proton decay in our $SU(5)\times U_X(1)$ model is in part, but not fully, explained by the presence of the $U_X(1)$ symmetry. For example, while dimension four operators of the form $\overline{\bf 5}\,\overline{\bf 5}\,{\bf 10}$ are clearly forbidden, dimension five operators $\overline{\bf 5}_3\,{\bf 10}_{-1}\,{\bf 10}_{-1}\,{\bf 10}_{-1}$ are allowed by the $U_X(1)$ symmetry. However, from Eq.~\eqref{Wgut}, these operators are forbidden, even at a general point in moduli space, by the additional $U(1)$ symmetries which arise at Abelian locus.
Let us finish this section with the following comment. The $12$ singlets in our model, charged under $U_X(1)$, can be viewed as right-handed neutrinos in analogy with conventional $SO(10)$ GUTs. As it happens, their number equals the number of families so that the charged spectrum forms $12$ copies of ${\bf 16}_{SO(10)}$. This means that the $U_X(1)$ is non-anomalous. This should be considered a coincidence in this particular model, as, in general, the number of singlets does not have to equal the number of families.
\vspace{-4pt}
\subsubsection{More on stability}\label{sec:stability2}
In Section~\ref{sec:stabex} we have argued that the bundle $\widetilde{V}$ is poly-stable in the region
\begin{equation}\label{eq:hyperplane1}
C_{\widetilde{V}}=\{\mathbf{s}\in C_{\bf s}\, |\, \ \mathbf{s.}(1,1,-1,-1)= 0 \text{ and } \mathbf{s.}(-1,1,0,0)\geq 0 \text{ and } \mathbf{s.}(1,1,-2,0)\geq 0 \}\; .
\end{equation}
in K\"ahler moduli space and that, due to the first condition above, it splits into a poly-stable bundle $\widetilde{V}=U\oplus L_4$. This result was based on computing cohomology dimensions only. However, for a complete stability analysis we need to consider line bundles injecting into all wedge powers $\wedge^k\widetilde{V}$, where $k=1,\ldots ,4$, by computing ranks of relevant maps. In this sub-section, we take up these computations and show that they do not change the above result for the poly-stable region.
Given that our bundle is already split, this task is somewhat simplified. All we need to consider is the line bundles injecting into powers of the rank four bundle $U$, that is, into $U$, $\wedge^2 U$ and $\wedge^3U$, which is the case for a line bundle $L$, precisely when $H^0(X,\wedge^kU\otimes L^{^*})$ is non-trivial. Starting with $U$ itself, a line bundle $L$ injects into it precisely if
\begin{equation}
H^0\left(X, U \otimes L^{^*} \right) \cong \text{Ker}\left( H^0\big(\widetilde{B}\otimes L^{^{\!*}}\big)\ \longrightarrow\ H^0\big(C\otimes L^{^{\!*}}\big)\right)
\end{equation}
is non-trivial. We have computed this kernel for all the line bundles with entries between $-3$ and $3$ and have obtained the following set of injecting line bundles:
\begin{table}[!h]
\vspace{4pt}
\begin{center}
\begin{tabular}{ c c c c c }
\varstr{14pt}{9pt} (-3, -3, -3, 2), &(-3, -3, -2, 2), &(-3, -3, -1, 2), &(-3, -3, 0, 1), &(-3, -3, 0, 2) \\
\varstr{14pt}{9pt}(-3, -2, -3, 1), &(-3, -2, -2, 1), &(-3, -1, -3, 1), &(-3, 1, -3, -1), &(-3, 1, -2, -1), \\
\varstr{14pt}{9pt}(-3, 1, -1, -3), &(-3, 1, -1, -2), &(-3, 1, -1, -1), &(-3, 1, 0, -3), &(-3, 1, 0, -2), \\
\varstr{14pt}{9pt}(-3, 2, -3, -3), &(-2, -3, -3, 2), &(-2, -3, -2, 2), &(-2, -3, -1, 2), &(-2, -3, 0, 1), \\
\varstr{14pt}{9pt}(-2, -3, 0, 2), &(-2, -2, -3, 1), &(-2, -1, -3, 1), &(-2, 1, -3, -1), &(-2, 1, -2, -1), \\
\varstr{14pt}{9pt}(-2, 1, -1, -3), &(-2, 1, -1, -2), &(-2, 1, -1, -1), &(-2, 1, 0, -3), &(-2, 1, 0, -2), \\
\varstr{14pt}{9pt}(-1, -3, -3, 1), &(-1, -3, -3, 2), &(-1, -3, -2, 1), &(-1, -3, -2, 2), &(-1, -3, -1, 1), \\
\varstr{14pt}{9pt}(-1, -3, -1, 2), &(-1, -3, 0, 1), &(-1, -3, 0, 2), &(-1, 1, -3, -1), &(-1, 1, -2, -1), \\
\varstr{14pt}{9pt}(-1, 1, -1, -3), &(-1, 1, -1, -2), &(-1, 1, -1, -1), &(-1, 1, 0, -3), &(-1, 1, 0, -2)
\end{tabular}
\end{center}
\vspace{-12pt}
\end{table}
\newline However, none of these line bundles de-stabilises the cone $C_{\widetilde{V}}$. To find the line bundles $L$ injecting into $\wedge^2 U$, we need to find the non-trivial cohomology groups
\begin{equation}
H^0\left(X, \wedge^2U \otimes L^{^*} \right) \cong \text{Ker}\left( H^0\big(\wedge^2\widetilde{B}\otimes L^{^{\!*}}\big)\! \longrightarrow \text{Ker}\big( H^0\big(\widetilde{B}\otimes C\otimes L^{^{\!*}}\big)\! \longrightarrow\! H^0\big(S^2C\otimes L^{^{\!*}}\big)\big)\!\right)\; .
\end{equation}
This computation is similar to that performed in Section~\ref{sec:monadspectrum} in order to decide the existence of $\mathbf{5}-\overline{\mathbf{5}}$ pairs. Although computationally challenging, we have computed the cohomology $H^0\left(X, \wedge^2 U \otimes L^{^*} \right)$ for all the line bundles with entries between $-1$ and $1$ and we find no injecting line bundles.
\vspace{20pt}
Finally, for the line bundles that potentially inject into $\wedge^3 U$ we need to compute cohomology groups of the type $H^0\left( X, \wedge^3 U\otimes L^{^*} \right)$. Using the equivalence $\wedge^3 U \cong U^{^*}\otimes L_4^{^*}$, and employing the dual monad sequence twisted up with $L_4^{^*}$ and $L^{^*}$ in a similar fashion to the discussion in Appendix~\ref{sec:stab_criteria}, it follows that the relevant cohomology group can be expressed as:
\begin{equation}
\begin{aligned}
H^0(L^{^*}\otimes \wedge^3 U) = &\
\text{Coker}\left(H^0\big( C^{^*}\!\otimes L_4^{^*}\otimes L^{^{\!*}} \big) \rightarrow H^0\big(B^{^*}\! \otimes L_4^{^*} \otimes L^{^{\!*}}\big) \right)\\ &\oplus \text{Ker}\!\left(H^1\big( C^{^*}\!\otimes L_4^{^*}\otimes L^{^{\!*}} \big) \rightarrow H^1\big(B^{^*}\!\otimes L_4^{^*} \otimes L^{^{\!*}}\big) \right)
\end{aligned}
\vspace{8pt}
\end{equation}
We have performed this computation for all the line bundles with entries between $-3$ and $3$, obtaining the following set of injecting line bundles:
\begin{table}[!h]
\begin{center}
\begin{tabular}{ c c c c c c}
\varstr{14pt}{9pt} (3, 0, 0, 3),\, &(3, -1, 3, 2),\, &(3, -1, 2, 3),\, &(2, 0, 0, 3),\, &(2, -1, 3, 2),\, &(1, -1, 3, 2)
\end{tabular}
\end{center}
\vspace{-20pt}
\end{table}
As before, none of these line bundles de-stabilises the cone $C_{\widetilde{V}}$.
}
\section{Conclusions and Outlook}
In this paper, we have presented an in-depth analysis of various aspects of heterotic line bundle models, in the context of a ``case-study" for the tetra-quadric Calabi-Yau hyper-surface in $(\mathbb{C}\mathbb{P}^1)^{\times 4}$. First, we have studied the question of finiteness of the class of heterotic line bundle models; mathematically this corresponds to the question of finiteness of poly-stable line bundles sums with fixed total Chern class. On the tetra-quadric, we have proved this finiteness result for the simple case of rank-two line bundle sums of the form $L\oplus L^{^*}$. For higher rank line bundle sums, we have seen that the complication arises at the boundary of K\"ahler moduli space. Indeed, restricting to the region in K\"ahler moduli space away from the boundaries (which corresponds to the supergravity approximation) and demanding a finite Calabi-Yau volume, we can show finiteness relatively easily.
From the $94$ phenomemologically promising models on the tetra-quadric which arise for the available group orders $|\Gamma|=2,4,8,16$ of freely-acting symmetries, we have chosen a particular example, exhibiting a symmetry $\Gamma=\mathbb{Z}_2\times\mathbb{Z}_2$, and we have used this for a more detailed discussion. At the GUT level, this model has gauge group $SU(5)\times U(1)^4$ (with three of the $U(1)$ anomalous) and twelve ${\bf 10}\oplus\overline{\bf 5}$ multiplets which lead to precisely three standard model families in the downstairs theory. There are three ${\bf 5}-\overline{\bf 5}$ pairs which, for appropriate Wilson line choices, lead to one pair of Higgs doublets downstairs. In addition, there are $60$ bundle moduli singlets, which reduce to $15$ downstairs. In summary, downstairs, this model has precisely the MSSM spectrum charged under the standard model group plus a number of standard model singlets.
For this model, we have studied the continuation into the non-Abelian part of the moduli space, using the four-dimensional effective theory as well as two different bundle constructions - extensions bundles and monads. Our particular focus was the fate of the Higgs doublets. We found that the only superpotential $\mu$-term allowed by the $U(1)^4$ symmetry is of the form ${\bf 1}_{2,4}\,L\,\bar{H}$, where ${\bf 1}_{2,4}$ is one type of bundle moduli singlets. Hence, in the part of bundle moduli space where $\langle {\bf 1}_{2,4}\rangle =0$ (which includes the line bundle locus, but also non-Abelian $S\left(U(4)\times U_X(1)\right)$-bundles) the Higgs doublets should remain massless. We have verified this prediction by explicitly constructing the associated non-Abelian bundles, both through extensions and monads, and computing their cohomologies. Hence, we have found a model where the Higgs doublets remain massless away from the purely Abelian line bundle locus, although not in the whole bundle moduli space. We have shown that the remaining $U_X(1)$ gives rise to a $B-L$ symmetry of the model.
The line bundle data base~\cite{lbdatabase} contains some models, defined on more complicated Calabi-Yau manifolds, where all $\mu$-terms are forbidden by the $U(1)^4$ symmetry. It would be interesting to explicitly study the non-Abelian continuations of these models. We hope that the results of the present paper will be helpful in this context and, more generally, for the task of finding a realistic standard model from string theory.
\section*{Acknowledgements}
The work of E.~I.~B.~is supported by the ARC Future Fellowship FT120100466.
A.~L.~is partially supported by the EPSRC network grant EP/l02784X/1 and by the STFC consolidated grant ST/L000474/1. A.~C.~wishes to thank the University College, Oxford and the STFC for supporting his graduate studies.
\newpage
|
\section{Introduction}
The \(\bar{\mbox{P}}\)ANDA detector at FAIR (Facility for Antiproton and Ion Research) in Darmstadt is ideally suited for the study of exclusive channels in proton-antiproton collisions.
Studying these reactions will open a new area of looking into the interior of hadrons and thus bring new perspectives to the question on how hadrons can be understood in terms of the fundamental degrees of freedom of QCD, i.e. in terms of quarks and gluons.
One possibility is to investigate the process \( p ~\antip \rightarrow \pip ~ \pim \).
Pion production is particularly interesting and illuminating because one has well established constraints at hand to find realistic light-cone wave functions (LCWFs) which serve as model input.
We adapt the description of heavy \( D\)-meson production within the handbag approach, cf. Ref.~\cite{gor1}, to the case of light meson production.
We advocate that for large transverse momentum transfer and for restricted parton virtualities and intrinsic transverse momenta the process amplitude factorizes into a (hard) partonic subprocess and (soft) hadronic transition matrix elements.
The hadronic transition matrix elements can be parameterized by transition distribution amplitudes (TDAs) which have been introduced in a series of papers, cf. Refs.~\cite{tda1}-\cite{tda4}.
To model the TDAs we extend the formalism of LCWF-overlap, developed in Ref.~\cite{kroll2}.
\section{The Handbag Mechanism}\label{sec_handbag}
\begin{figure}[h]
\centering
\includegraphics[width=.45\textwidth]{./dglap}
\hfill
\includegraphics[width=.45\textwidth]{./erbl}
\caption{The double-handbag mechanism for \( p ~\antip \rightarrow \pip ~ \pim \). Left: DGLAP-contribution. Right: ERBL-contribution (one of two). The blobs contain the non-perturbative dynamics of the process and are parameterized by transition distribution amplitudes.}
\label{fig_handbag}
\end{figure}
The assignments of the momenta of the incoming baryons and outgoing mesons can be read off from Fig.~\ref{fig_handbag}.
The reference frame (symmetric center-of-momentum system (CMS)) is aligned along the \( z \)-axis of the \( 3 \)-vector of the average momentum \( \bar{\mathbf{p}} \) defined by \( \bar{p} \equiv \frac{1}{2}(p + p^{\prime})\).
Symmetric CMS means that the transverse component of the momentum transfer, defined by \( \Delta \equiv p^\prime - p = q - q^\prime \), is symmetrically shared between the incoming and outgoing particles.
In light-cone (LC) coordinates the incoming proton and the outgoing \( \pi^{+} \) momenta are written as
\begin{equation}
\label{eq_mom}
p= \left[ \left( 1+ \xi \right)\bar{p}^+, \frac{M^2 + \mathbf{\Delta}_\perp^{2} /4}{2 \left(1 + \xi \right) \bar{p}^+}, - \frac{\mathbf{\Delta}_\perp}{2} \right] \hspace{0.5cm} \mbox{and} \hspace{0.5cm}
p^\prime = \left[ \left( 1 - \xi \right)\bar{p}^+, \frac{m^2 + \mathbf{\Delta}_\perp^{2} /4}{2 \left(1 - \xi \right) \bar{p}^+}, \frac{\mathbf{\Delta}_\perp}{2} \right].
\end{equation}
\noindent The, so called, ``skewness parameter'' \( \xi \) is defined by \( \xi \equiv -\frac{\Delta^+}{2 \bar{p}^+} \) and gives the relative momentum transfer between the proton and the \( \pi^{+} \) in the longitudinal plus direction\footnote{Note that in our reference frame and for the pion production \( \xi < 0 \).}.
The antiproton and the \( \pi^{-} \) momenta are parameterized in an analogous way.
Similar to wide-angle Compton scattering, cf. Ref.~\cite{kroll}, we argue that for large CMS scattering angles, i.e. large Mandelstam \( s,~ t \) and \( u \) (\( t,u \ge 1 ~\mbox{GeV}^2 \)), the process can be described by a double-handbag mechanism.
This double-handbag mechanism is shown in Fig.~\ref{fig_handbag}.
Given the large scattering angles, that provides us with a large energy scale, it is assumed that the \( p ~\antip \rightarrow \pip ~ \pim \) amplitude factorizes into a hard-scattering kernel on the partonic level and soft hadronic \( p \rightarrow \pi^{+} \) and \( \bar{p} \rightarrow \pi^{-} \) transition matrix elements.
We stress that there is no rigorous proof for factorization for our process.
It is, at most, a first step in this direction.
In the handbag mechanism only the minimal number of constituents which are required to transform the (initial) \( p \bar{p} \)- to the (final) \( \pi^{+} \pi^{-} \)-pair actively take part in the partonic subprocess.
Since we consider the proton to be a bound state of a scalar diquark, denoted by \( S[ud] \), and a \( u \)-quark the active constituents are the \( S[ud] \)-diquarks and \( d \)-quarks on the proton side and the corresponding antiparticles on the antiproton side.
Using the physically plausible assumptions of restricted parton virtualities and restricted intrinsic transverse momenta, cf. Refs.~\cite{gor1,kroll,gor2}, the hadronic amplitude can be written as
\begin{align}
\begin{aligned}
\label{eq_hadronic_amp}
\mathscr{M}_{\mu \nu} & =
\int \mathrm{d} \bar{x}~ \int \mathrm{d} \bar{x}^{\prime} ~ \, \widetilde{H} \left( \bar{x} ,\bar{x}^{\prime} \right) \\
&\times \bar{p}^{+} \int \frac{\mathrm{d} z^{-}}{2\pi} e^{i \bar{x} \bar{p}^{+} z^{-}} \braket{\pi^{+}:p^{\prime}|\Psi^{d} \left( -\frac{z^{-}}{2} \right) \phi^{S[ud]} \left( \frac{z^{-}}{2} \right) ~|p: p, \mu} \\
& \times \bar{q}^{-} \int \frac{\mathrm{d} z^{\prime +}}{2\pi} e^{i \bar{x}^{\prime} \bar{q}^{-} z^{\prime +}} \braket{\pi^{-}:q^{\prime}| ~\phi^{S[ud] \dagger} \left( \frac{z^{\prime+}}{2} \right) \bar{\Psi}^{d} \left( -\frac{z^{\prime+}}{2} \right) |\bar{p}:q,\nu}.
\end{aligned}
\end{align}
\noindent It is a convolution of a hard-scattering kernel \( \widetilde{H} \) and two hadronic matrix elements with respect to the average momentum fraction defined by
\begin{equation}
\bar{x}^{(\prime)} \equiv \frac{\bar{k}^{ + (\prime -) }}{\bar{p}^{+}(\bar{q}^-)} \hspace{0.5cm} \mbox{with} \hspace{0.5cm} \bar{k}^{+ (\prime -)}=\frac{k_{1}^{+ (\prime - )} + k_{2}^{+ (\prime -)}}{2}.
\end{equation}
\noindent Since we are working in the LC-quantization framework we want to work with the independent degrees of freedom of the \(d\)-quark field operators.
This can be done using the same projection techniques as in Refs.~\cite{gor1,kroll,gor2}.
The amplitude thus becomes
\begin{align}
\begin{aligned}
\label{eq_hadronic_amp2}
\mathscr{M}_{\mu \nu} & =
\frac{1}{4 \bar{p}^+ \bar{q}^-} \sum_{\lambda_1,\lambda_2 }\int \mathrm{d} \bar{x}~ \int \mathrm{d} \bar{x}^{\prime} ~ \, H_{\lambda_2,\lambda_2^\prime} \left( \bar{x} ,\bar{x}^{\prime} \right) \frac{1}{\bar{x} - \xi} \, \frac{1}{\bar{x}^{\prime} - \xi}\\
&\times \bar{v}\left( k_2, \lambda_2 \right) \gamma_+ \, \bar{p}^{+} \, \int \frac{\mathrm{d} z^{-}}{2\pi} e^{i \bar{x} \bar{p}^{+} z^{-}} \braket{\pi^{+}:p^{\prime}|\Psi^{d} \left( -\frac{z^{-}}{2} \right) \phi^{S[ud]} \left( \frac{z^{-}}{2} \right) ~|p: p, \mu} \\
& \times \bar{q}^{-} \int \frac{\mathrm{d} z^{\prime +}}{2\pi} e^{i \bar{x}^{\prime} \bar{q}^{-} z^{\prime +}} \braket{\pi^{-}:q^{\prime}| ~\phi^{S[ud] \dagger} \left( \frac{z^{\prime+}}{2} \right) \bar{\Psi}^{d} \left( -\frac{z^{\prime+}}{2} \right) |\bar{p}:q,\nu} \, \gamma^- u \left(k_2^\prime,\lambda_2^\prime \right),
\end{aligned}
\end{align}
\noindent where we have defined the hard-scattering amplitude as \( H_{\lambda_2,\lambda_2^\prime} \left( \bar{x},\bar{x}^\prime \right) \equiv \bar{u} \left( k_2^\prime , \lambda_2^\prime \right) \widetilde{H} \left( \bar{x} \bar{p}^+,\bar{x}^\prime \bar{p}^+ \right) \)
\( \times v \left(k_2, \lambda_2 \right)\).
\noindent Now, depending on the \( \bar{x}^{(\prime)} \)-region, three different partonic subprocesses contribute (see \\
Fig. \ref{fig_handbag}): \( S[ud] \, \overline{S[ud]} \rightarrow d \, \bar{d} \) (left), \( \overline{S[ud]} \rightarrow \overline{S[ud]} \, d \, \bar{d} \) (right) and \( S[ud] \rightarrow S[ud] \, d \, \bar{d} \) (not shown).
So far we have only taken a closer look on the contribution coming from the DGLAP-region alone, i.e. Fig.~\ref{fig_handbag} left. In Sec.~\ref{sec_dis} we discuss the possible influence of the ERBL-Region.
\section{Modeling the hadronic transitions}\label{sec_model}
The soft hadronic matrix elements are modeled in the spirit of an overlap of LCWFs (focusing on the \( p \rightarrow \pi^{+} \)-transition, since the \( \bar{p} \rightarrow \pi^{-} \)-transition can be treated along the same lines).
For doing that we take the Fourier representation of the field operators and the Fock-state decomposition of the hadron states in LC-quantum field theory.
We only take the valence Fock states of the hadrons into account, i.e. a \( \ket{p:S[ud],d} \) and a \( \ket{\pi^{+}: u,\bar{d} } \) state for the proton and \( \pi^{+} \), respectively.
Furthermore we only consider configurations with zero orbital angular momenta between the partons such that the parton helicities sum up to their parent-hadron helicity.
According to Ref.~\cite{kroll3} we choose for the proton
\begin{equation}
\label{proton_LCWF}
\Psi_p = N_p \; \tilde{x} \; \exp\left[- \frac{ a_p^{2}}{\tilde{x}(1-\tilde{x})} \, \tilde{\mathbf{k}}_{\perp}^2 \right]
\end{equation}
\noindent and according to Ref. \cite{kroll4} for the pion
\begin{equation}
\label{pion_LCWF}
\Psi_\pi = N_\pi \exp\left[- \frac{ a_\pi^{2}}{\hat{x}(1-\hat{x})} \, \hat{\mathbf{k}}_{\perp}^2 \right]
\end{equation}
\noindent as LCWFs.
\( \tilde{x} \) and \( \hat{x} \) denote the momentum fractions of the active \( S[ud] \)-diquark inside the proton and the active \( \bar{d} \)-quark inside the \( \pi^{+} \), respectively.
Each of the wave functions has two free parameters, a global normalization constant \( N_{p,\pi} \) and an oscillator parameter \( a_{p,\pi} \).
For the proton they are fixed by requiring a valence Fock state probability of \( P_p = 0.5 \) and a r.m.s of the intrinsic transverse momentum of the active quark of \( \sqrt{\braket{\mathbf{k}_\perp}^2} = 0.280 ~ \mbox{GeV} \).
This yields \( N_p = 61.8~\mbox{GeV}^{-2} \) and \(a_p = 1.1~\mbox{GeV}^{-1} \).
The pion parameters are fixed by demanding the r.m.s. of the active quark to be \( \sqrt{\braket{\mathbf{k}_\perp}^2} = 0.350 ~ \mbox{GeV} \) and the pion decay constant \( f_\pi = 0.093~ \mbox{GeV}\).
This gives \( N_\pi = 26.1 ~\mbox{GeV}^{-2} \) and \( a_\pi= 1.0 ~\mbox{GeV}^{-1}\).
The result for the LCWF overlap is shown in Fig.~\ref{fig_LCWF_overlap} for different values of the CMS scattering angle.
\vspace{-0.15cm}
\begin{figure}[h]
\centering
\includegraphics[width=.45\textwidth]{./DGLAPoverlap1}
\hfill
\includegraphics[width=.45\textwidth]{./DGLAPoverlap2}
\caption{The wave function overlap versus the average momentum \( \bar{x} \) for CMS scattering angles \( \theta = \pi/2\) (left panel) and \( \theta = 3\pi/2 \) (right panel) and different values of Mandelstam \( s\).}
\label{fig_LCWF_overlap}
\end{figure}
\vspace{-0.4cm}
\section{Hard-Scattering Amplitudes}\label{sec_hard}
Under the assumptions on the parton momenta discussed in Sec.~\ref{sec_handbag} one can show that in the hard-scattering amplitude the momentum fractions are approximately \( 1\), i.e. \( x_{1,2} = x_{1,2}^{\prime} \approx 1 \).
Since we treat the scalar diquarks and the \(d \)-quarks as being massless, the hard-scattering amplitude vanishes when \( \lambda_2 = \lambda_2^\prime \).
The non-vanishing contributions are
\begin{equation}
\label{eq_hard_amp}
H_{+,-} = H_{-,+}= - 4 \pi \frac{4}{9} \alpha_s (s) F_{s}(s) \sin \theta,
\end{equation}
\noindent where \( F_s(s) \), taken from Ref.~\cite{schweig}, takes care of the composite nature of the diquarks and the fact that they should dissolve into quarks if a large amount of momentum is transfered to the diquarks.
\section{Results and Discussions}\label{sec_dis}
The differential cross section for \( p ~\antip \rightarrow \pip ~ \pim \) reads
\begin{equation}
\label{eq_cross_sec}
\frac{d \sigma}{d \Omega} = \frac{1}{64 \pi^2} \, \frac{1}{s} \, \sqrt{\frac{1-4m^2/s}{1-4M^2/s}} \, \sigma_0 \hspace{0.5cm} \mbox{with} \hspace{0.5cm} \sigma_0 \equiv \frac{1}{4} \sum_{\mu,\nu} |\mathscr{M}_{\mu \nu}|^2 .
\end{equation}
\noindent The result for the differential cross section in the CMS-angle interval \( \frac{\pi}{2} \le \theta \le \frac{2 \pi}{3} \) is shown in Fig.~\ref{fig_diff_cs}.
Comparing our result at \( s = 12.9 ~\mbox{GeV}^2 \) for a CMS scattering angle of \( \pi/2 \) with the experimental data of Ref.~\cite{buran} reveals that our calculations in which we have only taken the DGLAP-region into account, predict a much smaller differential cross section. It is about three orders of magnitude smaller.
\begin{figure}[h]
\includegraphics[width=.45\textwidth]{./diffcsplot1}
\hfill
\includegraphics[width=.45\textwidth]{./diffcsplot3}
\caption{The differential cross section for \( s = 7 ~\mbox{GeV}^2~ \mbox{and}~ 10 ~\mbox{GeV}^2 \) as a function of the CMS scattering angle \( \theta \).}
\label{fig_diff_cs}
\end{figure}
\noindent For CMS energies around \( 10 ~\mbox{GeV}^2 \) (and larger) one may think of the ERBL-region to play a minor role, since there the skewness parameter is small\footnote{An increase of the CMS scattering angle \( \theta \) increases the skewness in magnitude in our kinematics.}, i.e. it is about \( 0.1 \) for a CMS angle of \( \theta = \frac{2\pi}{3} \) and most of the \( \bar{x}^{(\prime)}\) integration region \( 0 \le \bar{x}^{(\prime)} \le 1 \) is covered by the DGLAP-domain (\( \xi \le \bar{x}^{(\prime)} \le 1\)).
If this would be the case, the handbag mechanism, in its present form, would fail to give a reasonable description for \( p ~\antip \rightarrow \pip ~ \pim \).
However, such a conclusion could be premature.
In Ref. \cite{tda2} it is argued that the DGLAP-region dominates the cross section for large values of \( \xi \approx 1 \), while for small (and intermediate) values of \( \xi \) the dominant contribution is due to the ERBL-region, which is mainly determined by a baryon pole contribution.
Therefore it might be that the big discrepancy between our results and the experimental data can be attributed to the missing contributions from the ERBL-region.
In order to draw definite conclusions we thus have to include also the ERBL-region.
This is the subject of current investigations.
\ack
We want to thank the organizers of the ``Fairness 2013'' for providing an interesting workshop.
ATG and SK are supported by the Fonds zur F\"orderung der wissenschaftlichen Forschung in \"Osterreich via Grand No. J3163-N16 and FWF DK W1203-N16, respectively.
We further acknowledge helpful discussions with Bernard Pire from the \'{E}cole Polytechnique.
\section*{References}
\medskip
|
\section{Introduction}
A good understanding of (anti)neutrino cross sections is
crucial to reduce the systematic uncertainties
in oscillation experiments aiming at a precise determination of
neutrino properties~\cite{Formaggio:2013kya}. Our present knowledge
of neutrino-nucleus interactions
has been significantly improved by a new generation of oscillation and cross section
experiments. Quasielastic (QE) scattering measurements have been
published by MiniBooNE~\cite{AguilarArevalo:2010zc,AguilarArevalo:2010cx,AguilarArevalo:2013hm} at
neutrino energies $E_\nu \sim 1$~GeV, by MINER$\nu$A~\cite{Fields:2013zhk,Fiorentini:2013ezn}
at $E_\nu \sim 3.5$~GeV and by NOMAD at high (3-100~GeV) energies~\cite{Lyubushkin:2008pe}. Detailed
single pion production data have become available from
MiniBooNE~\cite{AguilarArevalo:2009ww,AguilarArevalo:2010xt,AguilarArevalo:2010bm} for different reaction
channels including the coherent one, which has also been studied by SciBooNE~\cite{Hiraide:2008eu,Kurimoto:2010rc} at $E_\nu \sim 1$~GeV
and NOMAD~\cite{Kullenberg:2009pu}. Finally, new inclusive cross section results have been reported by T2K~\cite{Abe:2013jth}, SciBooNE~\cite{Nakajima:2010fp}, MINOS~\cite{Adamson:2009ju} and NOMAD~\cite{Wu:2007ab} Collaborations. These results challenge our understanding of neutrino interactions with matter and have triggered a renewed theoretical interest~\cite{Morfin:2012kn}.
Quasielastic scattering has been investigated with a local Fermi gas~\cite{Graczyk:2003ru,Nieves:2004wx,Athar:2005hu,Martini:2009uj}, realistic spectral functions~\cite{Benhar:2005dj,Ankowski:2007uy}, different models to describe the interaction of the knocked-out nucleon with the residual nucleus~\cite{Martinez:2005xe,Butkevich:2007gm,Meucci:2011vd} and using the information from electron scattering data encoded in the scaling function~\cite{Caballero:2005sj}. The importance of two-nucleon contributions for the proper understanding of QE-like and inclusive cross sections has emerged in different studies~\cite{Martini:2009uj,Amaro:2010sd,Nieves:2011pp}, and their impact in the kinematic neutrino-energy reconstruction has been stressed~\cite{Martini:2012fa,Nieves:2012yz,Lalakulich:2012hs}. Incoherent pion production has also been scrutinized using microscopic models for the reaction mechanism on the nucleon~\cite{Sato:2003rq,Hernandez:2007qq,Hernandez:2010bx,Leitner:2008ue,Serot:2012rd,Hernandez:2013jka}, with special attention paid to pion final state interactions in
nuclei~\cite{Ahmad:2006cy,Leitner:2008wx,Golan:2012wx,Lalakulich:2012cj,Hernandez:2013jka}. New microscopic models have been
developed for coherent pion production~\cite{Singh:2006bm,AlvarezRuso:2007tt,Amaro:2008hd,Nakamura:2009iq,Zhang:2012xi} while traditional ones, based on the partial conservation of the axial current (PCAC), have been updated~\cite{Paschos:2005km,Berger:2008xs,Hernandez:2009vm,Kopeliovich:2012tu}.
One of the possible reaction channels is photon emission induced by
neutral current (NC) interactions (NC$\gamma$), which can occur on single nucleons
and on nuclear targets. Weak photon emission has a small cross
section compared, for example, with pion production, the most important
inelastic mechanism. In spite of this, NC photon emission turns out
to be one of the largest backgrounds in $\nu_\mu \to \nu_e$
$(\bar\nu_\mu \to \bar \nu_e)$ oscillation experiments where electromagnetic
showers instigated by electrons (positrons) and photons are not
distinguishable. Thus, NC events producing single photons become an
irreducible background to the charge-current (CC) QE signatures of
$\nu_e$ ($\bar \nu_e$) appearance. This is precisely the case of the MiniBooNE experiment
that was designed to test an earlier indication of a $\bar\nu_\mu \to \bar\nu_e$
oscillation signal observed at LSND~\cite{Athanassopoulos:1996jb,Athanassopoulos:1997pv}.
The MiniBooNE experiment finds an excess of
events with respect to the predicted background in both $\nu$ and
$\bar \nu$ modes. In the $\bar \nu$ mode, the data are found to be
consistent with $\bar\nu_\mu \to \bar\nu_e$ oscillations and have some
overlap with the LSND result~\cite{Aguilar-Arevalo:2013pmq}. MiniBooNE
data for $\nu_e$ appearance in the $\nu_\mu$ mode show a clear ($3 \sigma$)
excess of signal-like events at low reconstructed neutrino energies
($200 < E_\nu^{\mathrm{QE}} < 475$~MeV)~\cite{AguilarArevalo:2007it,Aguilar-Arevalo:2013pmq}.
However, the $E_\nu^{\mathrm{QE}}$ distribution of the events is only marginally compatible
with a simple two-neutrino oscillation model~\cite{Aguilar-Arevalo:2013pmq}. While
several exotic explanations for this excess have been proposed, it
could be related to unknown systematics or poorly understood backgrounds in the experimental analysis.
In a similar way, NC$\gamma$ is a source of misidentified electron-like events in the $\nu_e$
appearance measurements at T2K~\cite{Abe:2013xua}. Even if the NC$\gamma$ contribution to the background
is relatively small, it can be critical in measurements of the CP-violating phase.
It is therefore very important to have a robust theoretical understanding of the NC photon emission reaction,
which cannot be unambiguously constrained by data. This is the goal of the present work.
The first step forwards a realistic description of NC photon emission on nuclear targets of neutrino
detectors is the study of the corresponding process on the nucleon. Theoretical models for the
$\nu N \to \nu N \gamma$ reaction have been presented in Refs.~\cite{Hill:2009ek,Serot:2012rd}. They start
from Lorentz-covariant effective field theories with nucleon, pion, $\Delta(1232)$ but also
scalar ($\sigma$) and vector ($\rho$, $\omega$) mesons as the relevant degrees of freedom, and exhibit
a nonlinear realization of (approximate) $SU(2)_L \otimes SU(2)_R$ chiral symmetry. The single mechanism of
$\Delta(1232)$ excitation followed by its decay $\Delta \to N \gamma$ was considered in Ref.~\cite{Barbero:2012sb},
where a consistent treatment of the $\Delta$ vertices and propagator is adopted.
Several features of the previous studies,
in particular the approximate chiral symmetry and the dominance of the $\Delta(1232)$ mediated mechanism
are common to the model derived in our work. In Ref.~\cite{Serot:2012rd}, a special attention is paid
to the power counting, which is shown to be valid for neutrino energies below 550 MeV. However,
the neutrino fluxes of most neutrino experiments span to considerably higher energies.
Thus, in Ref.~\cite{Zhang:2012xn}, the power counting scheme was abandoned, and the
model of \cite{Serot:2012rd} was phenomenologically extended to the
intermediate energies ($E_\nu \sim 1 $ GeV) relevant for the MiniBooNE $\nu$ flux,
by including phenomenological form factors. Though the extension proposed for
the $\Delta$ and the nucleon Compton-like mechanisms seems reasonable, the one
for the contact terms notably increases the cross section above $\sim1$ GeV
(they are more significant for neutrinos than for antineutrinos).
Since the contact terms and the associated form factors
are not well understood so far, the model predictions for $E_\nu > 1$~GeV
should be taken cautiously, as explicitly acknowledged in Ref.~\cite{Zhang:2012xn}.
In nuclear targets, the reaction can be incoherent when the final nucleus is
excited (and fragmented) or coherent, when it remains in the ground state.
It is also possible that, after nucleon knockout, the residual excited nucleus decays emitting
low-energy $\gamma$ rays. This mechanism has been recently investigated~\cite{Ankowski:2011ei}
and shall not be discussed here. The model of Ref~\cite{Hill:2009ek} has been applied to incoherent
photon production in an impulse approximation that ignores nuclear corrections~\cite{Hill:2010zy}.
These are also neglected in the coherent case, which is calculated by treating the nucleus as a
scalar particle and introducing a form factor to ensure that the coherence is restricted to
low-momentum transfers~\cite{Hill:2009ek}. More robust is the approach of
Refs.~\cite{Zhang:2012aka,Zhang:2012xi} based on a chiral effective field theory for nuclei, again
extended phenomenologically to higher energies~\cite{Zhang:2012xn}.
In addition to Pauli blocking and Fermi motion, the $\Delta$ resonance
broadening in the nucleus, is also taken into account. The latter correction causes a very strong
reduction of the resonant contribution to the cross section, in variance with our results, as will
be shown below. The ratio of the $\Delta$ to photon and $\Delta$ to $\pi^0$ decay rates is enhanced
in the nuclear medium by an amount that depends on the resonance invariant mass, momentum and
also production
position inside the nucleus, as estimated with a transport model~\cite{Leitner:2008fg,Leitner:2009zz}.
The coherent channel has also been studied in Refs.~\cite{Gershtein:1980wu,Rein:1981ys} at high energies.
A discussion about these works can be found in Section V.E of Ref.~\cite{Hill:2009ek}.
It is worth mentioning that both the models of Ref.~\cite{Hill:2009ek} and
Refs~\cite{Serot:2012rd,Zhang:2012aka,Zhang:2012xi,Zhang:2012xn} have been used to calculate the
NC$\gamma$ events at MiniBooNE with contradicting conclusions~\cite{Hill:2010zy,Zhang:2012xn}.
While in Ref.~\cite{Hill:2010zy} the number of these events were calculated to be twice as many as expected
from the MiniBooNE in situ estimate, much closer values were predicted in Ref.~\cite{Zhang:2012xn}.
The result that NC$\gamma$ events give a significant contribution to the MiniBooNE low-energy excess~\cite{Hill:2009ek}
could have its origin in the lack of nuclear effects and rather strong detection efficiency correction.
Here we present a realistic model for NC photon emission in the $E_\nu \sim 1$~GeV
region that extends and improves certain relevant aspects of the existing descriptions.
The model is developed in the line of previous work on weak pion production on
nucleons~\cite{Hernandez:2007qq} and nuclei for both incoherent~\cite{Hernandez:2013jka}
and coherent~\cite{AlvarezRuso:2007it,Amaro:2008hd} processes.
The model for free nucleons satisfies the approximate chiral symmetry
incorporated in the non-resonant terms and includes the dominant
$\Delta(1232)$ excitation mechanism, with couplings and form factors taken from the available
phenomenology. Moreover, we have extended the validity of the
approach to higher energies by including intermediate excited states from the second
resonance region [$P_{11}(1440)$, $D_{13}(1520)$ and $S_{11}(1535)$]. Among them, we have
found a considerable contribution of the $D_{13}(1520)$ for (anti)neutrino energies above 1.5 GeV.
When the reaction takes place inside the nucleus, we have applied a series of standard
medium corrections that have been extensively confronted with experiment in similar
processes such as pion~\cite{Oset:1981ih,Nieves:1991ye},
photon~\cite{Carrasco:1989vq} and electron~\cite{Gil:1997bm,Gil:1997jg} scattering with nuclei, or
coherent pion photo~\cite{Carrasco:1991we} and electroproduction~\cite{Hirenzaki:1993jc}.
This paper is organized as follows. The model for NC production of
photons off nucleons is described in Sec.~\ref{sec:nucleon}. After
discussing the relevant kinematics, we evaluate the different
amplitudes in Subsec.~\ref{sec:gamma_amp}. In the first place, the
dominant $\Delta(1232)$ and non-resonant contributions are studied
(Subsec.~\ref{sec:delta-bkg}). Next, we examine the contributions
driven by $N^*$ resonances from the second resonance region
(Subsec.~\ref{sec:secondregion}). The relations between vector form
factors and helicity amplitudes, and the off-diagonal $N^*N$
Goldberger-Treiman (GT) relations are discussed in Appendices
\ref{sec:heliamp} and \ref{sec:axialcoupling}, respectively.
NC$\gamma$ reactions in nuclei are studied in
Sec.~\ref{sec:ncgnuclei}. First, in
Subsec.~\ref{sec:1p1h}, we pay attention to the incoherent channel driven by one
particle--one hole (1p1h) nuclear excitations. Next, in
Subsec.~\ref{sec:coherent}, the coherent channel is studied. We present our results in
Sec.~\ref{sec:results}, where we also compare some of our predictions
with the corresponding ones from Refs.~\cite{Hill:2009ek,Zhang:2012xn}. This Section
is split in two Subsections, where the results for NC$\gamma$ on
single nucleons (Subsec.~\ref{sec:res1}) and on nuclei
(Subsec.~\ref{sec:res2}) are discussed. Predictions for nuclear
incoherent and coherent reactions are presented in
Subsecs.~\ref{sec:res2a} and \ref{sec:res2b}, respectively. Finally
the main conclusions of this work are summarized in Sec.~\ref{sec:concl}.
\section{Neutral current photon emission off nucleons}
\label{sec:nucleon}
In this section, we describe the model for NC production of photons off nucleons,
\begin{equation}
\nu (k) N (p) \to \nu (k') N (p') \gamma (k_\gamma), \qquad
\bar\nu (k) N (p) \to \bar\nu (k') N (p') \gamma (k_\gamma) \label{eq:neu}
\end{equation}
\subsection{Kinematics and general definitions}
\begin{figure}[tbh]
\centerline{\includegraphics[width=0.55\textwidth]{plano.eps}}
\caption{Representation of the different LAB kinematical variables
used through this work.}\label{fig:coor}
\end{figure}
The unpolarized differential cross section with respect to the
photon kinematical variables (kinematics is sketched in Fig.~\ref{fig:coor}) is given in the
Laboratory (LAB) frame by
\begin{equation}
\frac{d^{\,3}\sigma_{(\nu,\bar\nu)}}{dE_\gamma d\Omega(\hat{k}_\gamma)} =
\frac{E_\gamma}{ |\vec{k}|}\frac{G^2}{16\pi^2}
\int \frac{d^3k'}{|\vec{k}^{\prime}\,|}
L_{\mu\sigma}^{(\nu,\bar\nu)}W^{\mu\sigma}_{{\rm NC}\gamma} \label{eq:sec} \,.
\end{equation}
As we neglect the neutrino masses, $E_\nu = |\vec{k}|$, $E^{\prime} = |\vec{k}^{\prime}\,|$ and $E_\gamma= |\vec{k}_\gamma|$, where $\vec{k}$, $\vec{k}^{\prime}$ and $\vec{k}_\gamma$ are the incoming neutrino, outgoing neutrino and outgoing photon momenta in LAB, in this order; $G=1.1664\times 10^{-11}$ MeV$^{-2}$ is the Fermi constant, while $L ^{(\nu,\bar\nu)}$ and $W_{{\rm NC}\gamma}$ stand for the leptonic and hadronic tensors, respectively. The leptonic tensor\footnote{Our conventions are such that $\epsilon_{0123}= +1$ and $g^{\mu\nu}=(+,-,-,-)$.}
\begin{eqnarray}
L_{\mu\sigma}^{(\nu,\bar\nu)}&=& (L_s)_{\mu\sigma}+ i
(L^{(\nu,\bar \nu)}_a)_{\mu\sigma} =
k^\prime_\mu k_\sigma +k^\prime_\sigma k_\mu
+ g_{\mu\sigma} \frac{q^2}2 \pm i
\epsilon_{\mu\sigma\alpha\beta}k^{\prime\alpha}k^\beta, \qquad \left(+\to
\nu, \, -\to \bar\nu \right) \label{eq:lep} \,,
\end{eqnarray}
is orthogonal to the four momentum transfer $q_\mu=k_\mu-k'_\mu$, with $q^2=-2k\cdot k'=-4EE'\sin^2\theta'/2$. The hadronic tensor
includes the non-leptonic vertices and reads
\begin{eqnarray}
W^{\mu\sigma}_{{\rm NC} \gamma} &=& \frac{1}{4M}\overline{\sum_{\rm
spins}} \int\frac{d^3p^\prime}{(2\pi)^3} \frac{1}{2E^\prime_N}
\delta^4(p^\prime+k_\gamma-q-p) \langle N \gamma |
j^\mu_{\rm NC\gamma}(0) | N \rangle \langle N \gamma | j^\sigma_{\rm NC \gamma}(0) | N
\rangle^* \label{eq:wmunu-nucleon} \,,
\end{eqnarray}
with $M$ the nucleon mass\footnote{We take the average of the neutron
and proton masses.} and $E^\prime_N$ the energy of the
outgoing nucleon. The bar over the sum of initial and final spins
denotes the average over the initial ones. The one particle
states are normalized as $\langle \vec{p}\, |
\vec{p}^{\,\prime} \rangle = (2\pi)^3 2p_0
\delta^3(\vec{p}-\vec{p}^{\,\prime})$. Then, the matrix element
$\langle N \gamma | j^\mu_{\rm NC\gamma}(0) | N \rangle$ is dimensionless.
For the sake of completeness, we notice that the NC, $j^\mu_{\rm NC}$ and electromagnetic (EM), $s^\mu_{\rm EM}$
currents at the quark level are given by
\begin{eqnarray}
j^\mu_{\rm NC} &=& \bar{\Psi}_u\gamma^\mu(1-\frac83 \sin^2\theta_W-
\gamma_5)\Psi_u - \bar{\Psi}_d\gamma^\mu(1-\frac43 \sin^2\theta_W-
\gamma_5)\Psi_d - \bar{\Psi}_s\gamma^\mu(1-\frac43 \sin^2\theta_W-
\gamma_5)\Psi_s \,, \nonumber\\
&=&\bar\Psi_q \gamma^\mu (1-\gamma_5) \tau^{(1)}_0 \Psi_q
- 4 \sin^2 \theta_W s^\mu_{\rm EM}
-\bar\Psi_s\gamma^\mu(1-\gamma_5)\Psi_s \,, \\
s^\mu_{\rm EM} &=& \frac23 \bar{\Psi}_u\gamma^\mu \Psi_u
- \frac13 \bar{\Psi}_d\gamma^\mu \Psi_d
- \frac13 \bar{\Psi}_s\gamma^\mu \Psi_s \,,
\end{eqnarray}
where $\Psi_u$, $\Psi_d$ and $\Psi_s$ are the quark fields and $\theta_W$
the weak angle ($\sin^2\theta_W= 0.231$). The zeroth spherical component of
the isovector operator $\tau^{(1)}$ is equal to the third component of the
isospin Pauli matrices $\vec{\tau}$.
By construction, the hadronic tensor accomplishes
\begin{eqnarray}
W^{\mu\sigma}_{{\rm NC} \gamma}=
W^{(s)\mu\sigma}_{{\rm NC} \gamma} + i
W^{(a)\mu\sigma}_{{\rm NC} \gamma} \,,
\end{eqnarray}
in terms of its real symmetric, $W^{(s)}_{{\rm NC} \gamma}$, and antisymmetric,
$W^{(a)}_{{\rm NC} \gamma}$, parts. Both lepton and hadron tensors are independent
of the neutrino flavor and, therefore, the cross section for the
reaction of Eq.~(\ref{eq:neu}) is the same for electron, muon or tau
incident (anti)neutrinos.
Let us define the amputated amplitudes $\Gamma^{\mu \rho}$, as
\begin{equation}
\langle N \gamma |
j^\mu_{\rm NC\gamma}(0) | N \rangle = \bar u(p') \Gamma^{\mu \rho} u (p) \epsilon^*_\rho(k_\gamma) \,,
\end{equation}
where the spin dependence of the Dirac spinors (normalized such that
$\bar uu= 2M$) for the nucleons is understood, and
$\epsilon(k_\gamma)$ is the polarization vector of the outgoing
photon. To keep the notation simple we do not specify the type of nucleon
($N=n$ or $p$) in $\Gamma^{\mu \rho}$. In terms of these
amputated amplitudes, and after performing the average (sum) over the
initial (final) spin states, we find
\begin{eqnarray}
W^{\mu\sigma}_{{\rm NC} \gamma} &=&- \frac{1}{8M} \int\frac{d^3p^\prime}{(2\pi)^3} \frac{1}{2E^\prime_N}
\delta^4(p^\prime+k_\gamma-q-p) {\rm Tr}\left[
(\slashchar{p}'+M)\Gamma^{\mu\rho}(\slashchar{p}+M) \gamma^0
(\Gamma^\sigma_{.\,\rho})^\dagger \gamma^0 \right] \,.
\label{eq:wmunu-nucleon-avg}
\end{eqnarray}
After performing the $d^3p^\prime$ integration,
there is still a $\delta(p^{\prime\,0}+E_\gamma-q^0-p^0) $
left in the hadronic tensor, which can be used to perform the integration over $| \vec{k}^\prime |$
in Eq.~(\ref{eq:sec}).
\subsection{Evaluation of the $\Gamma^{\mu \rho}$ amputated amplitudes}
\label{sec:gamma_amp}
\subsubsection{The $\Delta(1232)$ contribution, chiral symmetry and non-resonant terms}
\label{sec:delta-bkg}
Just as in pion production~\cite{Hernandez:2007qq}, one expects the NC$\gamma$ reaction to be dominated by
the excitation of the $\Delta(1232)$ supplemented with a non-resonant background. In our case, the leading
non-resonant contributions are nucleon-pole terms built out of $Z^0NN$ and $\gamma NN$ vertices that
respect chiral symmetry. The $q^2$ dependence of the amplitudes is introduced via phenomenological form factors.
We also take into account the subleading mechanism originated from the anomalous $Z^0\gamma\pi$ vertex, that involves
a pion exchange in the $t-$channel. Thus, in a first stage we consider the five diagrams depicted in Fig.~\ref{fig:five}.
The corresponding amputated amplitudes are
\begin{eqnarray}
\Gamma^{\mu \rho}_{N} = \Gamma^{\mu \rho}_{NP} + \Gamma^{\mu
\rho}_{CNP} &=& i e\,
J^\rho_{EM}(-k_\gamma)\frac{\slashchar{p}+\slashchar{q}+M}{(p+q)^2-M^2
+i\epsilon} J^\mu_{NC}(q)
+
i e\, J^\mu_{NC}(q)\frac{(\slashchar{p}'-\slashchar{q}+M)}{(p'-q)^2-M^2+i\epsilon}
J^\rho_{EM}(-k_\gamma) \label{eq:NP-CNP} \,, \\
\Gamma^{\mu \rho}_{\pi Ex} & = & eC_N\frac{g_A M}{4\pi^2 f_\pi^2}(1-4
\sin^2\theta_W) \frac{\epsilon^{\mu\rho\sigma\alpha}q_\sigma (k_{\gamma})_\alpha}{(q-k_\gamma)^2-m_\pi^2} \gamma_5, \qquad \left(C_N=+1\to
p, \, C_N=-1 \to n \right) \\
\Gamma^{\mu \rho}_{\Delta} = \Gamma^{\mu \rho}_{\Delta P} + \Gamma^{\mu
\rho}_{C\Delta P} &=& i e\,
\gamma^0 \left[ J^{\alpha \rho}_{EM}(p',k_\gamma)\right]^\dagger
\gamma^0\frac{P_{\alpha\beta}
(p+q)}{(p+q)^2-M_\Delta^2
+i M_\Delta \Gamma_\Delta} J^{\beta \mu}_{NC}(p,q) \nonumber \\
&+&
i e\, \gamma^0 \left[ J^{\alpha \mu}_{NC}(p',-q) \right]^\dagger
\gamma^0 \frac{P_{\alpha\beta}
(p^{\,\prime}-q)}{(p^{\,\prime}-q)^2-M_\Delta^2
+i \epsilon}J^{\beta \rho}_{EM}(p,-k_\gamma) \label{eq:DP-CDP} \,,
\end{eqnarray}
with $e > 0$ the electron charge, such that $\alpha=e^2/4\pi \approx
1/137$, $f_\pi = 92.4$ MeV the pion decay constant and $g_A=1.267$
the axial nucleon charge; $m_\pi$ and
$M_\Delta (\sim 1232\, {\rm MeV})$ are the pion and $\Delta$ masses, respectively.
\begin{figure}[ht!]
\begin{center}
\includegraphics[width=0.52\textwidth]{five.eps}
\caption{Model for photon emission off the
nucleon; direct and crossed nucleon-pole terms (a,b), direct and crossed
$\Delta(1232)$-pole terms (c,d) and the anomalous
$t-$channel pion exchange term (e). Throughout this work, we
denote these contributions as $NP$, $CNP$, $\Delta
P$, $C\Delta P$ and $\pi Ex$, respectively.
\label{fig:five}}
\end{center}
\end{figure}
As it will be clear in the following, each of the building blocks of the model
is gauge invariant by construction $ \bar u (p') \, \Gamma_{N,\Delta,\pi Ex}^{\mu \rho} \, u(p) (k_\gamma)_\rho = 0$.
The vector parts of these amplitudes are also conserved (CVC)
$ \bar u (p') \, V_{N,\Delta,\pi Ex}^{\mu \rho} \, u(p) q_\mu = 0$.
\paragraph{$NP$ and $CNP$ amplitudes:}
The nucleon NC and EM currents are given by
\begin{eqnarray}
J^\mu_{NC}(q)&=&\gamma^\mu \tilde{F}_1(q^2)+\frac{i}{2M}\sigma^{\mu\beta}q_\beta \tilde{F}_2(q^2) - \gamma^\mu \gamma_5 \tilde{F}_A(q^2), \\
J^\mu_{EM}(k_\gamma)&=& \gamma^\mu F_1(0) + \frac{i}{2M}\sigma^{\mu\nu}
(k_{\gamma})_\nu F_2(0),
\end{eqnarray}
where $\tilde{F}_{1,2}$ and $\tilde{F}_A$ are the NC vector and axial form factors\footnote{Note that pseudoscalar ($q^\mu \gamma_5$) terms do not contribute because $q^\mu L_{\mu\sigma}^{(\nu\bar\nu)}=0$ when neutrino masses are neglected.} while $F_{1,2}$ are the EM ones.
These form factors take different values for protons and neutrons. For $F_{1,2}$, we have that
\begin{eqnarray}
F^{(N)}_1 &=& \frac{G^N_E+\tau G^N_M}{1+\tau},\quad F^{(N)}_2 = \frac{G^N_M -
G^N_E}{1+\tau}, \qquad N=p,n
\end{eqnarray}
with
\begin{equation}
G_E^p = \frac{G_M^p}{\mu_p}=
\frac{G_M^n}{\mu_n} = -(1+ b \tau) \frac{G_E^n}{\mu_n a \tau} =
\left(\frac{1}{1-q^2/M^2_D}\right)^2 \,,
\end{equation}
where $\tau = -q^2/4M^2$, $M_D=0.84$ GeV, $\mu_p=2.793$, $\mu_n=-1.913$, $b=4.61$ and $a=0.942$~\cite{Krutov:2002tp}.
The NC vector form factors $\tilde{F}_{1,2}$ can be referred to the
EM ones thanks to isospin symmetry relationships,
\begin{eqnarray}
\tilde{F}^{(p)}_{1,2}&=&(1-4 \sin^2\theta_W) F^{(p)}_{1,2} -F^{(n)}_{1,2} -F^{(s)}_{1,2}\\
\tilde{F}^{(n)}_{1,2}&=&(1-4 \sin^2\theta_W) F^{(n)}_{1,2} -F^{(p)}_{1,2} -F^{(s)}_{1,2} \,,
\end{eqnarray}
where $F^{(s)}_{1,2}$ are the strange EM form factors. Furthermore, in the axial sector one has that
\begin{equation}
\tilde{F}^{(p,n)}_A = \pm F_A - F^{(s)}_A, \qquad \left(+\to
p, \, -\to n\right) \,,
\end{equation}
where $F_A$ is the axial form factor that appears in CCQE interactions, for which we adopt
a conventional dipole parametrization
\begin{equation}
F_A(q^2) = g_A \left(1-\frac{q^2}{M^2_A} \right)^{-2} \label{eq:gaxial}
\end{equation}
with an axial mass $M_A=1$~GeV~\cite{Bodek:2007ym}; $F^{(s)}_A$ is the strange axial form factor.
At present, the best determinations of the strange form factors are consistent with zero~\cite{Pate:2013wra}, thus
they have been neglected in the present study.
\paragraph{$\pi Ex$ amplitudes:}
The $t-$channel pion exchange contribution arises from the anomalous
($\pi^0\gamma Z^0$) Lagrangian~\cite{Hill:2009ek}
\begin{equation}
{\cal L}_{\pi^0\gamma Z^0}= \frac{eg}{4\cos\theta_W}
\frac{N_C}{12\pi^2f_\pi}(1-4 \sin^2\theta_W) \pi^0
\epsilon^{\mu\nu\alpha\beta}\partial_\mu Z_\nu\partial_\alpha
A_\beta \label{eq:WZW}
\end{equation}
together with the leading order $\pi^0NN$ interaction term
\begin{equation}
{\cal L}_ {\rm
\pi^0NN}= \frac{g_A}{f_\pi} \bar\Psi \gamma^\mu\gamma_5
\frac{\tau_3}{2}(\partial_\mu \pi^0)\Psi \,, \qquad
\Psi= \left (\begin{array}{c}p\cr n\end{array}\right ) \,,
\end{equation}
where $\Psi$, $\pi^0$, $A_\beta$, $Z_\nu$ are the nucleon, neutral pion,
photon and $Z^0$ boson fields, respectively.
Besides, $g = e/\sin \theta_W$ is related to the Fermi
constant $G$ and the $W$-boson mass as $G/\sqrt 2 = g^2/8M^2_W$;
$N_C$ is the number of colors.
The Lagrangian of Eq.~(\ref{eq:WZW})
arises from the Wess-Zumino-Witten
term~\cite{Wess:1971yu,Witten:1983tw}, which accounts for the axial anomaly
of QCD.
\paragraph{$\Delta P$ and $C \Delta P$ amplitudes:}
In the $\Delta-$driven amplitudes of Eq.~(\ref{eq:DP-CDP}), $P^{\mu\nu}$ is the spin
3/2 projection operator, which reads
\begin{equation}
P^{\mu\nu}(p_\Delta)= - (\slashchar{p}_\Delta + M_\Delta) \left [ g^{\mu\nu}-
\frac13 \gamma^\mu\gamma^\nu-\frac23\frac{p_\Delta^\mu
p_\Delta^\nu}{M_\Delta^2}+ \frac13\frac{p_\Delta^\mu
\gamma^\nu-p_\Delta^\nu \gamma^\mu }{M_\Delta}\right] \,; \label{eq:spin32}
\end{equation}
$\Gamma_\Delta$ is the resonance width in its rest frame, given by
\begin{equation}
\Gamma_\Delta(s) = \frac{1}{6\pi} \left ( \frac{f^*}{m_\pi}\right )^2
\frac{M}{\sqrt s} \left [ \frac{\lambda^\frac12
(s,m_\pi^2,M^2)}{2\sqrt s} \right ]^3 \Theta(\sqrt s
-M-m_\pi),\qquad s= p_\Delta^2 ,
\end{equation}
with $f^*=2.14$, the $\pi N \Delta$ coupling obtained from the empirical $\Delta \to N \pi$ decay width (see Table~\ref{tab:resonances});
$\lambda(x,y,z) = x^2+y^2+z^2-2xy-2xz-2yz$, and $\Theta$ is the step function.
The weak NC and EM currents for the nucleon to $\Delta$ transition are the
same for protons or neutrons and are given by
\begin{eqnarray}
\frac12 J^{\beta\mu}_{NC}(p,q) &=& \left[ \frac{\tilde{C}^V_3(q^2)}{M} (g^{\beta\mu} \slashchar{q}
-q^\beta \gamma^\mu ) +\frac{\tilde{C}^V_4(q^2)}{M^2} (g^{\beta\mu} q \cdot p_\Delta
-q^\beta p^\mu_\Delta )+\frac{\tilde{C}^V_5(q^2)}{M^2} (g^{\beta\mu} q \cdot p
-q^\beta p^\mu ) \right] \gamma_5 \nonumber\\
&+&\frac{\tilde{C}^A_3(q^2)}{M} (g^{\beta\mu}\slashchar{q}-q^\beta \gamma^\mu )
+ \frac{\tilde{C}^A_4(q^2)}{M^2} (g^{\beta\mu} q \cdot p_\Delta
-q^\beta p^\mu_\Delta ) + \tilde{C}^A_5(q^2) g^{\beta\mu}\label{eq:nc}\,,
\end{eqnarray}
\begin{eqnarray}
J^{\beta\rho}_{EM}(p,-k_\gamma) &=& -\left[ \frac{C^V_3(0)}{M}
(g^{\beta\rho} \slashchar{k}_\gamma-k_\gamma^{\beta} \gamma^\rho )
+\frac{C^V_4(0)}{M^2} (g^{\beta\rho} k_\gamma \cdot p_{\Delta c}
-k_\gamma^\beta p^\rho_{\Delta c} )
+ \frac{C^V_5(0)}{M^2} (g^{\beta\rho} k_\gamma \cdot p -k_\gamma^\beta p^\rho ) \right] \gamma_5 \,,
\end{eqnarray}
where $p_\Delta=p+q$ and $p_{\Delta c}= p-k_\gamma$; $\tilde{C}^V_i$, $\tilde{C}^A_i$ and $C^V_i$ are the NC vector, NC axial\footnote{There is another contribution to the axial current
$\tilde{C}^A_6(q^2) q^\beta q^\mu$, which does not contribute to the cross
section because $q^\mu L_{\mu\sigma}^{(\nu\bar\nu)}=0$ for
massless neutrinos.}
and EM transition form factors, respectively. As in the nucleon case, the NC vector form factors are related to the EM ones
%
\begin{equation}
\tilde{C}^V_i(q^2) = (1-2 \sin^2 \theta_W) C^V_i(q^2)
\end{equation}
according to the isovector character of the $N-\Delta$ transition. These EM form factors (and couplings) can be constrained
using experimental results on pion photo and electroproduction in the $\Delta$ resonance region. In particular,
they can be related to the helicity amplitudes $A_{1/2}$, $A_{3/2}$ and $S_{1/2}$~\cite{Lalakulich:2006sw,Leitner:2008ue}
commonly extracted in the analyses of meson electroproduction data. The explicit expressions are given in Appendix~\ref{sec:heliamp}. For the helicity amplitudes and their $q^2$ dependence we have taken the parametrizations of the MAID analysis~\cite{Drechsel:2007if,MAID}.~\footnote{The set of $N-\Delta(1232)$ vector form factors used in \cite{Hernandez:2007qq}, which were taken from Ref.~\cite{Lalakulich:2006sw}, lead to negligible changes in the results compared to those presented below.} In the axial sector, we adopt the Adler model~\cite{Adler:1968tw,Bijtebier:1970ku}
\begin{equation}
\tilde{C}^A_3(q^2) = 0,\qquad \tilde{C}^A_4(q^2) = -\frac{\tilde{C}^A_5(q^2)}{4}\,,
\end{equation}
for the subleading (in a $q^2$ expansion) form factors and assume a standard dipole for the dominant
\begin{equation}
\tilde{C}^A_5(q^2) = C^A_5(0) \left(1-\frac{q^2}{M^2_A} \right)^{-2} \,,
\label{eq:c5}
\end{equation}
with $C^A_5(0)=1.00\pm0.11$ and $M_A=0.93$~GeV fixed in a fit to $\nu_\mu d \to \mu^- \Delta^{++} n$ BNL and ANL data~\cite{Hernandez:2010bx}.
\subsubsection{The second resonance region}
\label{sec:secondregion}
Here, we extend the formalism to the second resonance region,
which includes three isospin 1/2 baryon resonances
$P_{11}(1440)$, $D_{13}(1520)$ and $S_{11}(1535)$ (see Table \ref{tab:resonances}).
In this way, we extend the validity of the model to higher energies.
A basic problem that has to be faced with resonances is the determination
of the transition form factors (coupling constants and $q^2$ dependence).
As for the $\Delta(1232)$, we obtain vector form factors
from the helicity amplitudes parametrized in Ref.~\cite{Drechsel:2007if}.
The equations relating helicity amplitudes and form factors are
compiled in Appendix~\ref{sec:heliamp}. Our knowledge of the axial transition
form factors is much poorer. Some constraints can
be imposed from PCAC and the
pion-pole dominance of the pseudoscalar form factors. These allow to
derive off-diagonal Goldberger-Treiman (GT) relations between the
leading axial couplings and the $N^* \to N \pi$ partial decay widths
(see Table~\ref{tab:resonances} and Appendix~\ref{sec:axialcoupling} for more details).
\begin{table}[htpb]
\caption{Properties of the resonances included in our model~\cite{Beringer:1900zz}. For each
state, we list the Breit-Wigner mass ($M_R$) , spin ($J$),
isospin ($I$), parity ($P$), total decay width
($\Gamma$), and axial coupling (denoted $F_A(0)$ for
spin 1/2 states and $C^A_5(0)$ for spin 3/2 states).}
\label{tab:resonances} \vspace{0.1cm}
\begin{center}
\begin{tabular}{c|ccccccc}
\hline\hline
& $M_R$ [MeV] &$J$ &$I$ &~$P$~ &$\Gamma$ [MeV] & $\Gamma(R \to N \pi)/\Gamma$ & $F_A(0)$ or $C^A_5(0)$ \\ \hline
$\Delta(1232)$ &1232 &3/2 &3/2 &+ &117 & 100\% &$1.00\pm
0.11$~\footnote{In the case of the
$\Delta$, we use a $C^A_5(0)$ value obtained
in a reanalysis~\cite{Hernandez:2010bx} of the $\nu_\mu p \to \mu^- p \pi^+$ ANL and
BNL bubble chamber data, which is smaller than
the corresponding GT relation by $\sim 20$\%.}
\\
$N(1440)$ &1440 &1/2 &1/2 &+ &300 & 65\% & $-$0.47 \\
$N(1520)$ &1520 &3/2 &1/2 &$-$ & 115 & 60\%&$-$2.14 \\
$N(1535)$ &1535 &1/2 &1/2 &$-$ & 150 & 45\%&$-$0.21 \\
\hline\hline
\end{tabular}\end{center}
\end{table}
For each of the three $P_{11}(1440)$, $D_{13}(1520)$ and
$S_{11}(1535)$ states, we have considered the contribution of direct ($RP$) and
crossed ($CRP$) resonance pole terms as depicted in Fig.~\ref{fig:highR}.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.52\textwidth]{excitFD.eps}
\caption{Direct (a) and crossed (b) $N^*$ pole contributions to the NC photon
emission process. We have considered the three resonances $\left[N(1440),
N(1535), N(1520)\right]$ right above the $\Delta(1232).$} \label{fig:highR}
\end{center}
\end{figure}
\paragraph{$N(1440)$ and $N(1535)$:}
The structure of the contribution of these two resonances to the amputated amplitudes is
similar to the one of the nucleon [Eq.~(\ref{eq:NP-CNP})]. We have
\begin{eqnarray}
\Gamma^{\mu \rho}_{R} = \Gamma^{\mu \rho}_{RP} + \Gamma^{\mu
\rho}_{CRP} &=& ie\,
J^\rho_{EM(R)}(-k_\gamma)\frac{\slashchar{p}+\slashchar{q}+M_R}{(p+q)^2-M_R^2
+iM_R \Gamma_R} J^\mu_{NC(R)}(q) \nonumber \\
& + &
ie\, J^\mu_{NC(R)}(q)\frac{(\slashchar{p}'-\slashchar{q}+M_R)}{(p'-q)^2-M^2_R+i \epsilon}
J^\rho_{EM(R)}(-k_\gamma) \,;
\end{eqnarray}
the resonance masses $M_R$ are listed in Table~\ref{tab:resonances} while the widths $\Gamma_R$
are discussed in Appendix~\ref{sec:width}. The EM and NC currents read
\begin{eqnarray}
\nonumber \\
J^\mu_{NC(P_{11})}(q) &=&
\frac{\tilde{F}_{1(P_{11})}(q^2)}{(2M)^2}(\slashchar{q} q^\mu-q^2\gamma^\mu) +
\frac{\tilde{F}_{2(P_{11})}(q^2)}{2M}i\sigma^{\mu\nu}q_{\nu} +\tilde{F}_{A(P_{11})} (q^2) \gamma^\mu \gamma_5 \,,
\\
J^\mu_{EM(P_{11})}(k_\gamma)&=&\frac{F_{1(P_{11})}(0)}{(2M)^2} \slashchar{k}_\gamma k_\gamma^\mu + \frac{F_{2(P_{11})}(0)}{2M}i\sigma^{\mu\nu}(k_{\gamma})_\nu
\end{eqnarray}
for the $N(1440)$ and
\begin{eqnarray}
J^\mu_{NC(S_{11})}(q) &=&\left[
\frac{\tilde{F}_{1(S_{11})}(q^2)}{(2M)^2}(\slashchar{q}
q^\mu-q^2\gamma^\mu) +
\frac{\tilde{F}_{2(S_{11})}(q^2)}{2M}i\sigma^{\mu\nu}q_{\nu} \right] \gamma_5
+\tilde{F}_{A(S_{11})} (q^2)\gamma^\mu \,,
\\
J^\mu_{EM (S_{11})}(k_\gamma)&=&\left[ \frac{F_{1(S_{11})}(0)}{(2M)^2} \slashchar{k}_\gamma k_\gamma^\mu + \frac{F_{2(S_{11})}(0)}{2M}i\sigma^{\mu\nu}(k_{\gamma})_\nu \right] \gamma_5
\end{eqnarray}
for the $N(1535)$.~\footnote{Note that by construction gauge
invariance and CVC are satisfied. This is also the case for the
$N(1520)$ amplitudes that will be discussed next.}
As in the nucleon case, isospin symmetry implies that
\begin{eqnarray}
\tilde{F}^{(p)}_{1,2(R)}&=&\left(1-4 \sin^2\theta_W \right)
F^{(p)}_{1,2(R)} -F^{(n)}_{1,2(R)} -F^{(s)}_{1,2(R)} \,, \nonumber \\
\tilde{F}^{(n)}_{1,2(R)}&=& \left(1 -4 \sin^2\theta_W \right)
F^{(n)}_{1,2(R)} -F^{(p)}_{1,2(R)}-F^{(s)}_{1,2(R)}\,,
\end{eqnarray}
with $F^{(N)}_{1,2 (P_{11},S_{11})}$ expressed in terms of the corresponding
helicity amplitudes (see Appendix~\ref{sec:heliamp}). The NC axial
form factors are
\begin{eqnarray}
\tilde{F}^{(p,n)}_{A(R)} &=& \pm F_{A(R)} + F^s_{A(R)}, \qquad \left(+\to
p, \, -\to n\right) \nonumber \\
F_{A(R)}(q^2) &=& F_{A(R)}(0) \left(1-\frac{q^2}{M^{* 2}_A} \right)^{-2}\label{eq:faj12t12} \,.
\end{eqnarray}
The couplings $F_{A(P_{11},S_{11})}(0)$ are obtained from the GT corresponding relations and have values
given in Table~\ref{tab:resonances}. The $q^2$ dependence of these form factors is unknown so we have
assumed a dipole ansatz with a natural value of $M^*_A=1.0$ GeV for the axial mass.
No information is available about the strange form factors $F^{(s)}_{1,2,A (P_{11},S_{11})}$
but they are likely to be small and to have a negligible impact on the
observables, so we set them to zero.
\paragraph{$N(1520)$:}
In this case, the structure of the contribution of this resonance to the amputated amplitudes
is similar to that of the $\Delta(1232)$, differing just in the definition
of the appropriate form factors and the isospin dependence. Thus, we have
\begin{eqnarray}
\Gamma^{\mu \rho}_{D_{13}} = \Gamma^{\mu \rho}_{D_{13} P} + \Gamma^{\mu
\rho}_{CD_{13} P} &=& ie\,
\gamma^0 \left[ J^{\alpha \rho}_{EM (D_{13})}(p',k_\gamma)\right]^\dagger
\gamma^0\frac{P_{\alpha\beta}^{D_{13}}
(p+q)}{(p+q)^2-M_{D_{13}}^2
+i M_{D_{13}} \Gamma_{D_{13}}} J^{\beta \mu}_{NC (D_{13})}(p,q) \nonumber \\
&+& ie\, \gamma^0 \left[ J^{\alpha \mu}_{NC (D_{13})}(p^{\,\prime},-q) \right]^\dagger
\gamma^0 \frac{P_{\alpha\beta}^{D_{13}}
(p^{\,\prime}-q)}{(p^{\,\prime}-q)^2-M_{D_{13}}^2
+i \epsilon} J^{\beta \rho}_{EM (D_{13})}(p,-k_\gamma)
\end{eqnarray}
where the resonance mass $M_{D_{13}}$ is given in Table~\ref{tab:resonances} and the width $\Gamma_{D_{13}}$
is discussed in Appendix~\ref{sec:width};
$P_{\mu\nu}^{D_{13}}$ is the spin 3/2 projection operator given also by
Eq.~(\ref{eq:spin32}), with the obvious replacement of $M_\Delta$ by $M_{D_{13}}$.
Besides, the $N-N(1520)$ EM and NC transition currents are given by
\begin{eqnarray}
J^{\beta\mu}_{NC(D_{13})}(p,q) &=& \frac{\tilde{C}^V_{3(D_{13})}(q^2)}{M} (g^{\beta\mu} \slashchar{q}
-q^\beta \gamma^\mu ) +\frac{\tilde{C}^V_{4(D_{13})}(q^2)}{M^2} (g^{\beta\mu} q \cdot p_{D_{13}}
-q^\beta p^\mu_{D_{13}} ) +\frac{\tilde{C}^V_{5(D_{13})}(q^2)}{M^2} (g^{\beta\mu} q \cdot p
-q^\beta p^\mu ) \nonumber \\
&+&\left[ \frac{\tilde{C}^A_{3(D_{13})}(q^2)}{M} (g^{\beta\mu}\slashchar{q}-q^\beta \gamma^\mu )
+ \frac{\tilde{C}^A_{4(D_{13})}(q^2)}{M^2} (g^{\beta\mu} q \cdot p_{D_{13}}
-q^\beta p^\mu_{D_{13}} ) + \tilde{C}^A_{5(D_{13})}(q^2) g^{\beta\mu}\right] \gamma_5
\end{eqnarray}
\begin{eqnarray}
J^{\beta\rho}_{EM(D_{13})}(p,-k_\gamma) &=& -\left[
\frac{C^V_{3(D_{13})}(0)}{M} (g^{\beta\rho}
\slashchar{k}_\gamma-k_\gamma^{\beta} \gamma^\rho )
+\frac{C^V_{4(D_{13})}(0)}{M^2} (g^{\beta\rho} k_\gamma \cdot
p_{D_{13}\, c} -k_\gamma^\beta p^\rho_{D_{13}\, c} ) \right. \nonumber \\
&+& \left. \frac{C^V_{5(D_{13})}(0)}{M^2} (g^{\beta\rho} k_\gamma \cdot p -k_\gamma^\beta p^\rho ) \right],
\end{eqnarray}
where $p_{D_{13}}=p+q$ and $p_{D_{13}\, c}= p-k_\gamma$;
$\tilde{C}^V_{i(D_{13})}$, $\tilde{C}^A_{i(D_{13})}$ and
$C^V_{i(D_{13})}$ are the NC vector, NC axial and EM form factors,
respectively. The NC vector form factors are related to the EM ones
in the same way as for the other isospin $1/2$ states considered above, namely
\begin{eqnarray}
\tilde{C}^{V(p)}_{i(D_{13})}&=&\left(1-4 \sin^2\theta_W
\right) C^{(p)}_{i(D_{13})} -C^{(n)}_{i(D_{13})}
-C^{(s)}_{i(D_{13})} \,,\nonumber \\
\tilde{C}^{V(n)}_{i(D_{13})}&=& \left( 1-4 \sin^2\theta_W
\right) C^{(n)}_{i(D_{13})} -C^{(p)}_{i(D_{13})} -C^{(s)}_{i(D_{13})} \,,
\end{eqnarray}
where $C^{(p,n)}_{3-5(D_{13})}$ are obtained from the helicity amplitudes
using Eqs.~(\ref{eq:a121520}-\ref{eq:s121520}).
For the axial form factors, one again has that
\begin{eqnarray}
\tilde{C}^{A(p,n)}_{i(D_{13})} &=& \pm C^A_{i(D_{13})} +
C^{sA}_{i(D_{13})} \,, \qquad \left(+\to
p, \, -\to n\right) \,.
\label{eq:c5N1520}
\end{eqnarray}
We take a standard dipole form for the dominant axial NC form factor
\begin{eqnarray}
C^A_{5(D_{13})}(q^2) &=& C^A_{5(D_{13})}(0) \left( 1-\frac{q^2}{M^{* 2}_A} \right)^{-2} \,,
\label{eq:c5q2N1520}
\end{eqnarray}
with $C^A_{5(D_{13})}(0)$ from the corresponding off diagonal GT relation
(see Appendix~\ref{sec:axialcoupling} and Table~\ref{tab:resonances}), and set
$M^*_A=1.0$ GeV as for the other $N^*$. The other axial form factors
$C^A_{3,4(D_{13})}$ are less important because their contribution to the amplitude
squared is proportional to $q^2$. We neglect them together with the
unknown strange vector and axial form factors.
\section{Neutral current photon emission in nuclei}
\label{sec:ncgnuclei}
In this section we outline the framework followed
to describe NC photon emission off nuclei.
Both incoherent and coherent reaction channels are considered.
\subsection{Incoherent neutral current photon emission}
\label{sec:1p1h}
To study the incoherent reactions
\begin{equation}
\nu_l (k) +\, A_Z \to \nu_l (k^\prime) + \,
\gamma(k_\gamma) +\, X,\qquad \bar\nu_l (k) +\, A_Z \to \bar\nu_l (k^\prime) + \,
\gamma(k_\gamma) +\, X,
\label{eq:reacincl}
\end{equation}
we pursue the many body scheme
derived in Refs.~\cite{Nieves:2004wx,Nieves:2005rq,Nieves:2011pp} for
the neutrino propagation in nuclear matter and adapted to
(semi)inclusive reactions on finite nuclei by means of
the local density approximation. With this formalism, the
photon emission cross section is
\begin{equation}
\left. \sigma_{(\nu,\bar\nu)}\right|_{\mathrm{incoh}} =
\frac{1}{ |\vec{k}~|}\frac{G^2}{16\pi^2}
\int \frac{d^3k'}{|\vec{k}^\prime|}
L_{\mu\sigma}^{(\nu,\bar\nu)}\left.W^{\mu\sigma}_{{\rm NC}\gamma}\right|_{\mathrm{incoh}} \label{eq:sec-incl}
\end{equation}
in terms of the leptonic tensor of Eq.~(\ref{eq:lep}) and the hadronic tensor $W^{\mu\sigma}_{{\rm NC} \gamma}|_{\mathrm{incoh}}= W^{(s)\mu\sigma}_{{\rm NC} \gamma}|_{\mathrm{incoh}} + i
W^{(a)\mu\sigma}_{{\rm NC} \gamma}|_{\mathrm{incoh}}$, which is determined by the contributions to the $Z^0$ selfenergy with a photon in the intermediate state $\Pi^{\mu\sigma}_{Z\gamma}(q)$
\begin{eqnarray}
\left. W^{(s)\mu\sigma}_{NC\gamma}\right|_{incoh} &=& - \Theta(q^0) \left (\frac{4 \cos\theta_W}{g} \right )^2
\int \frac{d^3 r}{2\pi}~ {\rm Im}\left [ \Pi_{Z\gamma}^{\mu\sigma}
+ \Pi_{Z\gamma}^{\sigma\mu} \right ] (q,r)\label{eq:zmunuselfs}\\
\left. W^{(a)\mu\sigma}_{NC\gamma}\right|_{incoh} &=& - \Theta(q^0) \left (\frac{4 \cos\theta_W}{g} \right )^2
\int \frac{d^3 r}{2\pi}~{\rm Re}\left [ \Pi_{Z\gamma}^{\mu\sigma}
- \Pi_{Z\gamma}^{\sigma\mu}\right] (q,r) \label{eq:zmunuselfa}.
\end{eqnarray}
In the density expansion proposed in Ref.~\cite{Nieves:2004wx}, the lowest order
contribution to $\Pi^{\mu\sigma}_{Z\gamma}$ is depicted in Fig.~\ref{fig:1p1h}.
The black dots stand for any of the
eleven terms ($NP$, $CNP$, $\pi Ex$, $RP$, $CRP$ with $R=\Delta(1232)$, $N(1440)$, $N(1520)$, $N(1535)$)
of the elementary $Z^0 N \to \gamma N$ amplitude derived in
Sec.~\ref{sec:nucleon}. The solid upwards and downwards oriented
lines represent nucleon particle and hole states in the Fermi sea.
\begin{figure}[h!]
\centerline{\includegraphics[height=7.cm]{1phgamma.eps}}
\caption{Diagrammatic representation of the one-particle-one-hole-photon (1p1h$\gamma$)
contributions to the $Z^0$ self-energy in nuclear matter. The black dots represent
$Z^0N \to \gamma N$ amplitudes.}\label{fig:1p1h}
\end{figure}
This $Z^0$ selfenergy diagram (actually 121
diagrams) is readily evaluated as
\footnote{In Eq.~(\ref{eq:W1pibis}), it is necessary to subtract the free space
contribution, i.e., the one that survives for vanishing nuclear
densities and renormalizes free space couplings and masses.
Actually, to obtain Eq.~(\ref{eq:Z1ga}), we have neglected the
contribution of the antiparticle pole ($p^0=- E(\vec{p}\,)-{\rm
i}\epsilon$) in the $p^0$ integration. This automatically removes
the unwanted vacuum part.}
\begin{equation}
-i \Pi_{Z\gamma;1{\rm p1h}\gamma}^{\mu\nu}(q,r) = i \left (\frac{g}{4\cos\theta_W} \right )^2
\sum_{N=p,n} \int \frac{ d^4 k_\gamma}{(2\pi)^4} \int\frac{d^4p}{(2\pi)^4}
\frac{1}{k_\gamma^2+i\epsilon}
{\rm Tr}\left[ S(p,\rho_N) \gamma^0\left(\Gamma_N^{ \mu \rho}
\right)^\dagger\gamma^0 S(p',\rho_N) (\Gamma_N)^\nu_{.\, \rho}
\right] \label{eq:W1pibis}
\end{equation}
where $p'=p+q-k_\gamma$ and $\Gamma_N^{\mu\rho}$ is the amputated amplitude for the
$Z^0N\rightarrow N \gamma $ process
\begin{gather}
\Gamma_N^{\mu\rho} =
\sum_a \Gamma^{\mu \rho}_{a;N},\,
\quad a=NP,\, CNP,\,\pi Ex,\, RP,\, CRP \,\,
\left [R=\Delta(1232), N(1440), N(1520), N(1535) \right ] \,.
\label{eq:defjA}
\end{gather}
The nucleon propagator in the medium reads
\begin{equation}
S(p,\rho_N) = (\slashchar{p}+M) G(p,\rho_N) \,
\end{equation}
with
\begin{eqnarray}
G(p\,; \rho_N) &=& \frac{1}{p^2-M^2+{\rm
i}\epsilon} + {\rm
i}\frac{\pi}{E(\vec{p}\,)}n_N(\vec{p}\,)\delta(p^0-E(\vec{p}\,))
\label{eq:Gpbis}\\
&=& \frac{1}{p^0+ E(\vec{p}\,) + {\rm
i}\epsilon} \left ( \frac{n_N(\vec{p}\,)}{p^0- E(\vec{p}\,) - {\rm
i}\epsilon} + \frac{1-n_N(\vec{p}\,)}{p^0- E(\vec{p}\,) + {\rm
i}\epsilon} \right ) \,.
\label{eq:Gp}
\end{eqnarray}
The occupation number in the local Fermi gas
$n_N (\vec{p}\,) = \Theta(k_F^N-|\vec{p}\,|)$ depends on the local density of nucleons (protons or neutrons)
in the nucleus via $k^N_F(r)= (3\pi^2\rho_N(r))^{1/3}$. The nucleon energy $E(\vec{p})$ is approximated by the
free one $\sqrt{\vec{p}\,^2 + M^2}$. Substituting the explicit expressions of $S(p,\rho_N)$ and $S(p',\rho_N)$ in
Eq.~(\ref{eq:W1pibis}) one obtains
\begin{eqnarray}
-i \Pi_{Z\gamma;1{\rm p1h}\gamma}^{\mu\nu}(q,r)
&=&-\left(\frac{g}{4\cos\theta_W} \right )^2\sum_{N=p,n} \int \frac{
d^4 k_\gamma}{(2\pi)^4} \int\frac{d^3p}{(2\pi)^3}\frac{1}{2E(\vec{p}\,)}
\frac{1}{2E(\vec{p}+\vec{q}-\vec{k}_\gamma)}\frac{n_N(\vec{p}\,)[1-n_{N}
(\vec{p}+\vec{q}-\vec{k}_\gamma)]}{q^{0}-k_\gamma^{0}+E(\vec{p}\,)-E(\vec{p}\,')+i\epsilon}
\times \nonumber \\
& \times & \frac{1}{k_\gamma^2+i\epsilon}
{\rm
Tr}\left [(\slashchar{p}+M)\gamma^0\left(\Gamma^{ \mu \rho}_{N}
\right)^\dagger\gamma^0(\slashchar{p}'+M)\left(\Gamma_{N}\right)^\nu_{.\,\rho}
\right]
+ \left[ (q-k_\gamma)\leftrightarrow -(q-k_\gamma) \right ]\,. \label{eq:Z1ga}
\end{eqnarray}
A convenient simplification can be made by evaluating the
$\Gamma^{\mu\rho}_{N}$ amplitudes at an average nucleon hole
four momentum $\langle p^\mu \rangle$. This
allows us to take the spin trace in Eq.~(\ref{eq:Z1ga}) out of the $d^3p$ integration,
which gives, up to constants, the Lindhard function, $\overline
{U}_R(q-k_\gamma,k_F^N,k_F^{N})$ (see Appendix B of
Ref.~\cite{Nieves:2004wx} for definition and explicit expressions). Therefore,
\begin{eqnarray}
-i \Pi_{Z\gamma;1{\rm p1h}\gamma}^{\mu\nu}(q,r)&=&-\left
(\frac{g}{4\cos\theta_W} \right )^2 \frac{1}{4M^2}\sum_N \int \frac{
d^4 k_\gamma}{(2\pi)^4}
\frac{1}{k_\gamma^2+i\epsilon}\overline {U}_R(q-k_\gamma,k_F^N,k_F^{N})
A^{\mu\nu}_{N=p,n}\left(\langle p\rangle,q,k_\gamma\right)\,, \label{eq:1ph1ga}
\\
A^{\mu\nu}_{N} &=& \frac12 {\rm
Tr}\left[\left(\langle\slashchar{p}\rangle+M\right)\gamma^0 \left( \left\langle \Gamma_{N} \right \rangle^{
\mu \rho} \right)^\dagger \gamma^0\left(\langle\slashchar{p}\rangle+\slashchar{q}
-\slashchar{k}_\gamma+M\right) \left\langle \Gamma_{N}\right \rangle^\nu_{.\,\rho} \right]
\end{eqnarray}
\noindent
where $\langle \Gamma_{N} \rangle^{\nu\rho}$ stands for $\Gamma^{\nu\rho}_{N}$ calculated at
the average hole four momentum $\langle p^\mu \rangle$.
To derive the $1{\rm p1h}\gamma$
contribution to the hadron tensor $W^{\mu\sigma} $, we remind that by
construction
\begin{equation}
A^{\mu\nu}_{N} = A^{(s) \mu\nu}_{N} + i
A^{(a) \mu\nu}_{N} \label{eq:sym}
\end{equation}
where $A^{(s) \mu\sigma}_{N}$
($A^{(a) \mu\sigma}_{N}$) is a real symmetric
(antisymmetric) tensor. Furthermore, it is easy to see that the
combinations of the $Z^0$ selfenergy present in
Eqs.~(\ref{eq:zmunuselfs}) and (\ref{eq:zmunuselfa}) fulfill
\begin{equation}
{\rm Im}\left [ \Pi_{Z\gamma;1{\rm p1h}\gamma}^{\mu\nu}
+ \Pi_{Z\gamma;1{\rm p1h}\gamma}^{\nu\mu} \right ] = 2 {\rm Im} \Pi_{Z\gamma;1{\rm
p1h}\gamma}^{(s) \mu\nu}, \qquad {\rm Re}\left [
\Pi_{Z\gamma;1{\rm p1h}\gamma }^{\mu\nu}
- \Pi_{Z\gamma;1{\rm p1h}\gamma}^{\nu\mu} \right ] = -2 {\rm Im} \Pi_{Z\gamma;1{\rm
p1h}\gamma}^{(a) \mu\nu}
\end{equation}
where $\Pi_{Z\gamma;1{\rm p1h}\gamma}^{(s,a) \mu\nu}$ are obtained by replacing
$A^{\mu\nu}_{N}$ in Eq.~(\ref{eq:1ph1ga}) by the corresponding
$A^{(s,a) \mu\nu}_{N}$ parts.
The imaginary
part of $\Pi_{ZW;1{\rm p1h}\gamma}^{\mu\nu}\Big|_{s(a)}$ can be obtained
following the Cutkosky rules. In this case we
cut the selfenergy diagram of Fig.~\ref{fig:1p1h} with a straight horizontal line.
The states intercepted by the line are placed on shell by taking the imaginary part of their
propagators. Technically, the rules to obtain ${\rm Im}
\Pi_{Z;1{\rm p1h}\gamma}^{\mu\nu}$ consist of the following substitutions:
\begin{eqnarray}
\Pi_{Z\gamma}^{\mu\nu}(q)&\rightarrow& 2i {\rm Im}\Pi^{\mu\nu}_{Z\gamma}(q)\Theta(q^0)\\
\frac{1}{k_\gamma^2+i\epsilon}
&\rightarrow& 2i {\rm Im} \frac{1}{k_\gamma^2+i\epsilon}
\Theta(k_\gamma^0) = -2\pi
i \delta(k_\gamma^2) \Theta(k_\gamma^0) \\
\overline {U}_R(q-k_\gamma,k_F^N,k_F^{N}) &\rightarrow& 2i {\rm
Im}\overline {U}_R(q-k_\gamma,k_F^N,k_F^{N}) \Theta(q^0-k_\gamma^0) \
\end{eqnarray}
Thus, taking into account that $A^{(s,a) \mu\nu}_{N}$ are real, we readily obtain
\begin{equation}
W^{\mu\nu}_{1{\rm p1h}\gamma}(q) = \Theta(q^0) \frac{1}{2M^2} \int
\frac{d^3r }{2\pi}
\sum_{N=p,n} \frac{d^3k_\gamma}{(2\pi)^3}
\frac{\Theta(q^0-E_\gamma)}{2E_\gamma} {\rm
Im}\overline {U}_R(q-k_\gamma,k_F^N,k_F^{N}) A^{\nu\mu}_{N\gamma} \label{eq:1p1hga-def}
\end{equation}
with $E_\gamma$ the photon on-shell energy.
The average nucleon hole momentum $\langle p^\mu \rangle$ is chosen as follows (see the
discussion after Eq.~(9) of Ref.~\cite{Hernandez:2013jka})
\begin{equation}
\langle p^0 \rangle = \frac{E_F^N + E_\mathrm{min}}2 \,, \qquad \langle
|\vec{p}\,| \rangle = \sqrt{\langle p^0 \rangle^2 -M^2} \label{eq:approx1}
\end{equation}
defined by the central value of the allowed energy region, with
\begin{equation}
E_\mathrm{min} = {\rm max} \left ( M, E_F^N-q^{\prime 0}, \frac{-q^{\prime
0}+|\vec{q}^{\,\prime}|\sqrt{1-4M^2/q^{\prime 2}}}2 \right ),
\end{equation}
where $q' = q-k_\gamma$ and $E_F^N=\sqrt{M^2+(k_F^N)^2}$. The corresponding
nucleon hole angle, in the LAB frame and with respect to $\vec{q}^{\,\prime}$,
is completely fixed by the kinematics to
\begin{equation}
\cos\theta_N = \frac{q^{\prime 2}+2\langle p^0 \rangle q^{\prime 0}}{2 \langle
|\vec{p}\,| \rangle |\vec{q}^{\,\prime}|}
\label{eq:approx3}
\end{equation}
while the azimuthal angle $\phi_N$ is fixed arbitrarily in the plane
perpendicular to $\vec{q}^{\,\prime}$. Similar approximations were performed,
and shown to be sufficiently accurate, in studies
of total inclusive and pion production in photo and electro-nuclear
reactions~\cite{Carrasco:1989vq,Carrasco:1991mb, Gil:1997bm, Gil:1997jg}. They were also used in
Ref.~\cite{Nieves:2011pp} to compute the total inclusive neutrino induced cross
section. We have checked that the approximation of Eqs.~(\ref{eq:approx1})--(\ref{eq:approx3})
induces uncertainties of at most 5\%, independently of $\phi_N$ values.
Furthermore, different choices of $\phi_N$
produce small variations of the order of 1-2\% in the results. This
approximation saves a considerable amount of computational time
because there are analytical expressions for ${\rm Im}\overline
{U}_R(q-k_\gamma,k_F^N,k_F^{N}) $ (see for instance Ref.~\cite{Nieves:2004wx}).
In the small density limit $ {\rm Im}\overline {U}_R(q^{\,\prime},k_F^N,k_F^{N})
\simeq - \pi \rho_N M\delta \left(q^{\,\prime \,0}+M -
\sqrt{M^2+\vec{q}^{\,\prime 2}}\right)/\sqrt{M^2+\vec{q}^{\,\prime 2}}$. Substituting
this expression in Eq.~(\ref{eq:1p1hga-def}) one obtains
\begin{eqnarray}
\lim_{\rho\to 0}W^{\mu\nu}_{1{\rm p1h}\gamma} \sim \int
d\Omega(\hat{k}_\gamma) dE_\gamma
E_\gamma \left (Z W^{\mu\nu}_{Z^0p \to
p\gamma} + N W^{\mu\nu}_{Z^0n \to
n\gamma}\right) \,,
\end{eqnarray}
where $Z$ and $N$ are the number of protons and neutrons in the nucleus,
and $W^{\mu\nu}_{Z^0N \to N\gamma}$ is the hadronic tensor for NC
photon production on the nucleon. In this way, the strict impulse
approximation is recovered. By performing the integral in
Eq.~(\ref{eq:1p1hga-def}), Pauli blocking and Fermi motion are taken into account.
\subsubsection{Further nuclear medium corrections}
\label{sec:delta-medium}
Given the dominant role played by the $\Delta P$ contribution and
since $\Delta$ properties are strongly modified in the nuclear
medium~\cite{Hirata:1978wp,Oset:1981ih,Freedman:1982yp,Oset:1987re,Nieves:1993ev,
Singh:1998ha,Lehr:1999zr} a proper treatment of the
$\Delta$ contribution is needed. Here, we follow
Ref.~\cite{AlvarezRuso:2007tt} and modify the $\Delta$ propagator in the
$\Delta P$ term as
\begin{equation}
\frac1{p_\Delta^2-M_\Delta^2+iM_\Delta\Gamma_\Delta}\to\frac1{\sqrt{p_\Delta^2}+M_\Delta}
\frac1{\sqrt{p_\Delta^2}-M_\Delta+i(\Gamma^{\rm Pauli}_\Delta/2-{\rm
Im}\Sigma_\Delta)}\,;
\end{equation}
$\Gamma^{\rm Pauli}_\Delta$, for which we take the expression in Eq. (15) of
Ref.~\cite{Nieves:1991ye},
is the free $\Delta$ width corrected by the Pauli blocking
of the final nucleon. The imaginary
part of the $\Delta$ self-energy in the medium ${\rm Im}\Sigma_\Delta$,
is parametrized as~\cite{Oset:1987re}
\begin{equation}
-\mathrm{Im}\Sigma_\Delta(\rho)=C_Q\left(\frac{\rho}{\rho_0}\right)^{\alpha}+C_{A2}\left(\frac{\rho}{\rho_0}\right)^{\beta}+
C_{A3}\left(\frac{\rho}{\rho_0}\right)^{\gamma}\,, \label{eq:delta-self}
\end{equation}
where the term proportional to $C_Q$ accounts for the QE part while
those with coefficients $C_{A2}$ and $C_{A3}$ correspond to the two-body ($\Delta
N\rightarrow NN$) and three-body ($\Delta NN\rightarrow NNN$)
absorption contributions, respectively. The parameters in Eq.~(\ref{eq:delta-self})
can be found in Eq.~(4.5) and Table 2 of Ref.~\cite{Oset:1987re}, given as
functions of the kinetic energy in the laboratory system of a pion that
would excite a $\Delta$ with the corresponding invariant mass.
These parametrizations are valid in the range 85 MeV
$< T_\pi<$315 MeV. Below
85 MeV, the contributions from $C_Q$ and $C_{A3}$ are rather small
and are taken from Ref~\cite{Nieves:1991ye}, where the model was extended
to low energies. The term with $C_{A2}$ shows a very mild energy
dependence and we still use the parametrization from Ref.~\cite{Oset:1987re}
even at low energies. For $T_\pi$ above 315 MeV we have kept
these self-energy terms constant and equal to their values at the
bound. The uncertainties in these pieces are not very relevant
there because the $\Delta \rightarrow N\pi$ decay becomes very large and
absolutely dominant.
For the $\Delta$ mass we shall keep its free value. While there are some
corrections arising from both the real part of the self-energy and random
phase approximation (RPA) sums, the net effect is smaller than the precision
achievable in current neutrino experiments,
and also smaller than the uncertainties due to our limited knowledge of the
nucleon to $\Delta$ transition form factor $C_5^A(q^2)$ (see
the related discussion in Sec. II.E of Ref.~\cite{Nieves:2011pp}).
\subsection{Coherent neutral current photon emission}
\label{sec:coherent}
The coherent reactions
\begin{equation}
\nu_l (k) +\, A_Z|_{gs}(p_A) \to \nu_l (k^\prime) +
A_Z|_{gs}(p^\prime_A) +\, \gamma(k_\gamma), \qquad
\bar\nu_l (k) +\, A_Z|_{gs}(p_A) \to \bar\nu_l (k^\prime) +
A_Z|_{gs}(p^\prime_A) +\, \gamma(k_\gamma)
\label{eq:reac}
\end{equation}
consist of a weak photon production where the nucleus
is left in its ground state, in contrast with the incoherent
production that we studied in the previous subsection, where the
nucleus is either broken or left in an excited state. Here, we
adopt the framework derived in Ref.~\cite{Amaro:2008hd} for neutrino-induced
coherent CC and NC pion production reactions.\footnote{The predictions of
Ref.~\cite{Amaro:2008hd} were updated in \cite{Hernandez:2010jf}
after the reanalysis of the $\nu_\mu p \to \mu^- p \pi^+$ old bubble
chamber data carried out in Ref.~\cite{Hernandez:2010bx}.} This work is,
in turn, based on previous studies of coherent pion production in
electromagnetic [$(\gamma, \pi^0)$~\cite{Carrasco:1991we}, $(e,
e'\pi^0)$~\cite{Hirenzaki:1993jc}] and hadronic reactions [$(^3{\rm
He}, ^3{\rm H}\, \pi^+)$~\cite{FernandezdeCordoba:1992ky},
$p(^4{\rm He},^4{\rm He})X$~\cite{FernandezdeCordoba:1993az}] in the
$\Delta (1232)$ region. More recently, the same scheme has been employed to
study charged kaon production by coherent scattering of neutrinos and
antineutrinos on nuclei~\cite{AlvarezRuso:2012fc}. The model for the
coherent process is built up from the coherent scattering
with each of the nucleons of the nucleus, producing an
outgoing $\gamma$. The nucleon state (wave function) remains unchanged
so that after summing over all nucleons, one
obtains the nuclear densities. In the elementary $Z^0 N \to N \gamma$
process, energy conservation is accomplished by imposing
$q^0=E_\gamma$, which is justified by the large nucleus mass,
while the transferred momentum
$\vec{q}-\vec{k}_\gamma$ has to be accommodated by the nucleon wave
functions. Therefore, the coherent production process is sensitive to the
Fourier transform of the nuclear density.
Following Ref.~\cite{Amaro:2008hd}, it is straightforward to find that
\begin{gather}
\left.\frac{d^{\,3}\sigma_{(\nu,\bar\nu)}}{dE_\gamma
d\Omega(\hat{k}_\gamma)}\right|_{\rm coh} =
\frac{E_\gamma}{ |\vec{k}~|}\frac{G^2}{16\pi^2}
\int \frac{d^3k'}{|\vec{k}^\prime|}
L_{\mu\sigma}^{(\nu,\bar\nu)}\left.W^{\mu\sigma}_{{\rm
NC}\gamma}\right|_{\rm coh} \,, \\[0.2cm]
\left. { W}^{\mu\sigma}_{{\rm NC}\gamma}\right|_{\rm coh} = -
\frac{\delta(E_\gamma-q^0)}{64\pi^3M^2} {\cal A}^{\mu\rho}(q,k_\gamma)
\left({\cal A}^\sigma_{.\,\rho}\right)^*(q,k_\gamma) \label{eq:zmunu} \,,
\\[0.2cm]
{\cal A}^{\mu\rho}(q,k_\gamma) =
\int d^3r\ e^{i\left(\vec{q}-\vec{k}_\gamma\right)\cdot\vec{r}}
\left\{\rho_{p}(r\,) {\hat \Gamma}^{\mu\rho}_{p}(r;q,k_\gamma)
+ \rho_{n}(r\,)
{\hat \Gamma}^\mu_{n}(r;q,k_\gamma) \right\}\label{eq:Jmunu2} \,.
\end{gather}
To evaluate the hadronic tensor, we use the model for the NC photon
production off the nucleon derived in Sect.~\ref{sec:nucleon} and thus
we have
\begin{gather}
{\hat \Gamma}^{\mu\rho}_{N}(r;q,k_\gamma) =
\sum_i {\hat \Gamma}^{\mu\rho}_{i;N}(r;q,k_\gamma),\,
\quad i=NP,\, CNP,\,\pi Ex,\, RP,\, CRP \,\,
\left [R=\Delta,N(1440), N(1535), N(1520)\right ]
\label{eq:cn-jcoh} \\
\left. {\hat \Gamma}^{\mu\rho}_{i;N}(r;q,k_\gamma) =
\frac12 {\rm Tr}\left[(\slashchar{p}+M)\gamma^0\,\Gamma_{i;
N}^{\mu\rho} \right]\frac{M}{p^0}
\right|_{p^\mu=\left (\sqrt{M^2+\frac{(\vec{k}_\gamma-\vec{q}\,)^2}{4}},
\frac12(\vec{k}_\gamma-\vec{q}\,)\right )}
\label{eq:cn-jcoh2}
\end{gather}
where the four-vector matrices $\Gamma_{i; N\gamma}^{\mu\rho}$ stand
for the amputated photon production amplitudes off
nucleons derived in Subsec.~\ref{sec:gamma_amp}.
We have also taken into account the modification of the $\Delta(1232)$
in the medium for the $\Delta P$ mechanism, as explained in
Subsec.~\ref{sec:delta-medium}.
Now we pay attention to the approximated treatment of nucleon momentum
distributions that has been adopted to obtain
Eqs.~(\ref{eq:zmunu})--(\ref{eq:cn-jcoh2}). The initial ($\vec{p}$)
and final ($\vec{p}^{\,\prime}$) nucleon three momenta are not
well defined. We take
\begin{equation}
p^\mu = \bigg( \sqrt{M^2+ \frac14{\left(\vec{k}_\gamma-\vec{q}\right)^2}} ,
\frac{\vec{k}_\gamma-\vec{q}}{2}\,\bigg) \label{eq:pmu} \, , \qquad
p^{\prime\,\mu} = q - k_\gamma + p
= \bigg( \sqrt{M^2+ \frac14{\left(\vec{k}_\gamma-\vec{q}\right)^2}} ,
-\frac{\vec{k}_\gamma-\vec{q}}{2}\,\bigg)\,,
\end{equation}
with both nucleons being on-shell. In this way, the
momentum transfer is equally shared between the initial and final
nucleons. This prescription, employed in
Refs.~\cite{AlvarezRuso:2007tt,AlvarezRuso:2007it,Amaro:2008hd,Zhang:2012xi},
for (anti)neutrino induced coherent pion production, was earlier applied to
$^{16}\mathrm{O}(\gamma,\pi^+)^{16}\mathrm{N}_\mathrm{bound}$~\cite{Eramzhian:1983pe}
and to coherent $\pi^0$ photo- and
electroproduction~\cite{Carrasco:1991we,Hirenzaki:1993jc,Drechsel:1999vh}.
The approximation is based on the fact that, for Gaussian nuclear wave
functions, it leads to an exact treatment of the terms in the elementary amplitude
that are linear in
momentum. In Ref.~\cite{Carrasco:1991we}
it was shown that in the case of $\pi^0$ photoproduction,
this prescription provided similar results as the
explicit sum over the nucleon momenta performed in
Ref.~\cite{Boffi:1991nh}. Thanks to the choice of Eq.~(\ref{eq:pmu}),
the sum over all nucleons is greatly simplified and cast in terms of
the neutron and proton densities [see Eq.~(\ref{eq:Jmunu2})].
Furthermore, the sum over nucleon helicities gives rise to the trace in
Eq.~(\ref{eq:cn-jcoh2}); more details can be found in the discussion after
Eq.~(6) of Ref.~\cite{Amaro:2008hd}. On the other hand,
this approximation eliminates some non-local
contributions to the amplitudes. In particular, the $\Delta$ momentum
turns out to be well defined once the the nucleon momenta are fixed.
In Ref.~\cite{Leitner:2009ph} this
constraint was relaxed for weak coherent pion
production via $\Delta(1232)$ excitation, while neglecting the modification of the
$\Delta$ properties in the nucleus and pion distortion.
It was found that non-localities in the $\Delta$
propagation cause a large reduction of the cross section at low
energies. In the more realistic description of Nakamura et
al.~\cite{Nakamura:2009iq}, the non-locality is preserved for the
$\Delta$ kinetic term in a linearized version of the $\Delta$ propagator
but, at the same time, a prescription similar to Eq.~(\ref{eq:pmu})
for the $WN\Delta$ and $\Delta N\pi$ vertices, and a local
ansatz for the in-medium $\Delta$ selfenergy have been taken. Nevertheless,
the mismatch between the non-local recoil effects and the local approximations
for vertices and selfenergy are likely to be minimized by the fact that the
parameters in the $\Delta$ selfenergy are adjusted to describe pion-nucleus
scattering data with the same model. Our point of view is that
the local approach adopted here and in
Refs.~\cite{AlvarezRuso:2007tt,AlvarezRuso:2007it,Amaro:2008hd,Zhang:2012xi},
together with the choice of the effective nucleon-nucleon interaction in the medium~\cite{Oset:1987re},
is internally consistent. The good agreement obtained for pion-nucleus
scattering~\cite{GarciaRecio:1989xa,Nieves:1991ye} and coherent pion photoproduction~\cite{Zhang:2012xi,JuanPC}
for medium and heavy nuclei seems to support this conjecture, although more detailed investigations are
necessary. In any case, for the present study, where the coherent contribution
is a small and not disentangled part of the total $NC\gamma$ cross section,
and in view of the uncertainty in the determination of the
$N\Delta$ axial coupling $C^A_5(0)$, it is safe to disregard possible non-local
corrections.
\section{Results}
\label{sec:results}
Before discussing our results an important remark is due.
The intermediate nucleon propagators in both the $NP$ and $CNP$ terms
of Eq.~(\ref{eq:NP-CNP}) can be put on the mass shell for $E_\gamma\to
0$ photons, leading to an infrared divergence. This divergence should be cancelled
by others present in the electromagnetic radiative corrections to
the elastic process $\nu N \to \nu N$ (without photon
emission). However, when the emitted photon is too soft, its energy
becomes smaller than the photon energy resolution of the detector.
Such an event would be recorded as an elastic one if at all.
For this reason, we have
implemented a cut in the available photon phase space, demanding
$E_\gamma \geq 140$ MeV, which corresponds to the MiniBooNE
detection threshold~\cite{AguilarArevalo:2007it}.
\subsection{Neutral current photon emission off nucleons}
\label{sec:res1}
\begin{figure}[h!]
\begin{center}
\makebox[0pt] {\includegraphics[width=0.5\textwidth]{cs_nu.eps}\includegraphics[width=0.5\textwidth]{cs_nb.eps}}
\caption{(color online) $\nu N \to \nu N \gamma$ (left) and $\bar\nu N
\to \bar\nu N \gamma$ (right) cross sections on protons and neutrons as a function of the
(anti)neutrino energy. A cut of $E_\gamma \geq 140$~MeV
in the phase space integrals has been applied.
Solid curves correspond to
the results from the full model, with error bands determined by
the uncertainty in the axial $N\Delta$ coupling $C^A_5(0)=1.0 \pm 0.11$
according to the determination of Ref.~\cite{Hernandez:2010bx}. The curves labeled as $\Delta$,
$N$ and $\pi$ stand for the partial contributions of the $(\Delta
P+C\Delta P)$, $(N P+CNP)$ and $\pi Ex$ mechanisms,
respectively. The $D_{13}$, $P_{11}$ and $S_{11}$ curves show the
contribution of the different $(RP+CRP)$ terms driven by the $N^*$ resonances.
Finally, the lines labeled as ``no $N^*$''
display the predicted cross section without the $N^*$ contributions. }
\label{fig:cs_nu}
\end{center}
\end{figure}
In Fig.~\ref{fig:cs_nu}, we show our results for the total NC photon
emission (anti)neutrino cross sections as a function of the
(anti)neutrino energy. As in other weak interaction processes,
the different helicities of $\nu$ and
$\bar \nu$ are responsible for different interference patterns, resulting in
smaller $\bar \nu$ cross sections with a more linear energy
dependence. The error bands on the full model results are
determined by the uncertainty in the
axial $N\Delta$ coupling $C^A_5(0)=1.00\pm0.11$~\cite{Hernandez:2010bx}.
This is the predominant source of uncertainty in the (anti)neutrino energy
range under consideration (see also the discussion of Fig.~\ref{fig:delta_term} below).
We also display the contributions from the different mechanisms considered in our model
(Figs.~\ref{fig:five} and \ref{fig:highR}). The $\Delta$ mechanism is
dominant and gives the same contribution for protons and neutrons, as
expected from the isovector nature of the electroweak $N-\Delta$ transition.
At $E_{\nu(\bar \nu)} \sim 1.5$ GeV, the cross section from nucleon-pole
terms is only about 2.5 smaller than the $\Delta$ one. Above $\sim 1.5$~GeV,
the $N(1520)$ contribution is sizable and comparable to that of the
sum of the $NP$ and $CNP$ mechanisms, specially for
$\bar \nu p$. However, the rest of $N^*$ contributions considered in the
model (with $N(1440)$ and $N(1535)$ intermediate states), together
with the $\pi Ex$ contribution of Fig.~\ref{fig:five}(e) can be safely
neglected in the whole range of (anti)neutrino energies considered in
this work. The fact that the $N(1520)$ resonance is the only one,
besides the $\Delta(1232)$, playing a significant role for $E_\nu < 2$~GeV
has also been observed in pion production~\cite{Hernandez:2013jka} and
for the inclusive cross section~\cite{Leitner:2008ue}.
\begin{figure}[h!]
\begin{center}
\makebox[0pt] {\includegraphics[width=0.5\textwidth]{dcs_nu_1gev.eps}\includegraphics[width=0.5\textwidth]{dcs_nb_1gev.eps}}
\caption{(color online) $\nu N \to \nu N \gamma$ (left) and $\bar\nu N
\to \bar\nu N \gamma$ (right) photon energy (top) and photon angular
(bottom) differential cross sections at $E_{\nu,\bar\nu}=1$ GeV on both
protons and neutrons. The angle
$\theta_\gamma$ is referred to the direction of the incoming
(anti)neutrino beam. A cut of $E_\gamma \geq 140$~MeV has been applied.
Solid curves are for the full model.
The curves labeled as $\Delta$, $N$
and $D_{13}$ stand for the partial contributions of the $(\Delta
P+C\Delta P)$, $(N P+CNP)$ and the $(N(1520)P+CN(1520)P)$ terms,
respectively. The lines labeled as ``no $N^*$'' display the
predictions neglecting the $N^*$ contributions. }
\label{fig:1gev}
\end{center}
\end{figure}
Photon angular and energy distributions on single nucleons, for
incoming (anti)neutrino energies of 1 and 2 GeV are shown in
Figs.~\ref{fig:1gev} and \ref{fig:2gev}. Solid curves
stand for the results from the full model. We also display the largest
contributions among the different mechanisms considered in our model.
As expected, the
$\Delta$ mechanisms are also dominant in the differential
cross sections, specially for reactions on neutrons and even more so
for the $\bar\nu n \to \bar\nu n \gamma$ process. Nucleon and $D_{13}$
direct and crossed pole-term contributions, though small, are not
negligible, particularly for protons. The $N(1520)$ terms become
more important for the largest (anti)neutrino energy.
At the lower energy the reaction is more forward-peaked for neutrinos
than for antineutrinos. In the later case, the maximum of the distribution
moves forward as the energy increases.
The photon energy differential cross sections always exhibit
a peak slightly above $E_\gamma=0.2$ GeV, mainly produced
by the interplay between the $\Delta-$ pole and the three-body phase
space photon energy distribution. The $\Delta$ propagator suppresses
not only the low photon energy contributions, but also the high photon
energy tail that would appear because of the boost to the LAB
frame.
\begin{figure}[h!]
\begin{center}
\makebox[0pt] {\includegraphics[width=0.5\textwidth]{dcs_nu_2gev.eps}\includegraphics[width=0.5\textwidth]{dcs_nb_2gev.eps}}
\caption{(color online) Same as Fig.~\ref{fig:1gev}, but for an
(anti)neutrino energy of 2 GeV.}
\label{fig:2gev}
\end{center}
\end{figure}
Next, we compare our predictions for the nucleon cross sections
with those obtained in Refs.~\cite{Hill:2009ek,Zhang:2012xn}. These two
models include the $NP+CNP$ and $\Delta P+C\Delta P$ mechanisms,
with dominance of $\Delta P$ like in our case.
The Compton-like contributions ($NP+CNP$) are determined by the
electromagnetic and axial nucleon form factors, which are reasonably
well constrained. The predictions of Ref.~\cite{Zhang:2012xn} for
these mechanisms are similar to ours. Instead, those in
Ref.~\cite{Hill:2009ek} exhibit a steeper energy dependence, because
of the higher nucleon axial mass, $M_A=1.2 $ GeV in $F_A$
[Eq.~(\ref{eq:gaxial})], used there. This choice was motivated by the
first phenomenological analysis of the MiniBooNE CCQE scattering data on
carbon using the relativistic Fermi gas model~\cite{AguilarArevalo:2007ab}\footnote{In the final MiniBooNE
analysis~\cite{AguilarArevalo:2010zc}, an even larger value of
$M_A\sim 1.35$ GeV was obtained. }. Later theoretical
studies~\cite{Martini:2009uj, Martini:2011wp,Nieves:2011pp,Nieves:2011yp}
have shown that such high values of $M_A$ encoded multi-nucleon contributions
that were not taken into account in the experimental analyses.
We use a lower value for $M_A = 1$~GeV, which is consistent with two
independent experimental
sources: bubble chamber neutrino/antineutrino induced QE reactions on
hydrogen and deuterium and pion electroproduction~\cite{Bodek:2007ym}.
In addition to the $NP+CNP$ and $\Delta P+C\Delta P$ mechanisms,
R. Hill~\cite{Hill:2009ek} also considers $t-$channel $\pi$, $\rho$
and $\omega$ exchanges. Only the latter one provides a non-negligible
cross section that, for antineutrinos, could become
comparable to the nucleon Compton-like contribution for incident
energies above 1.5 GeV. However, the size of the $\omega$ contribution
strongly depends on the mostly undetermined off-shell form factor and
is then affected by large uncertainties.
In the model of X. Zhang and B. Serot~\cite{Zhang:2012xn}, additional
contact terms allowed by symmetry were considered.
As pointed out in the Introduction, they
notably increase the cross section above $\sim 1$ GeV (see Fig. 3 of
that reference). In Ref.~\cite{Serot:2012rd}, it is argued that these
contact terms are the low-energy manifestation of anomalous $\rho$ and
$\omega$ interactions; their contributions below 550 MeV are
very small, as expected on the base of the power counting established there.
To extend these findings to higher energies, phenomenological form factors
are employed~\cite{Zhang:2012xn}, which are, however, not well understood.
Therefore, their cross section above $E_\nu \sim 1$ GeV should be taken cautiously
once contact terms are a source of uncontrolled systematics.
We now focus on the comparison for the dominant $\Delta$ contribution, which is
presented in Fig.~\ref{fig:delta_term}.
Different values of the axial $N\Delta$ coupling $C^A_5(0)$ and photon
energy cuts have been implemented in Refs.~\cite{Zhang:2012xn,Hill:2009ek},
as specified in the caption of
Fig.~\ref{fig:delta_term}. We have used these inputs and compared our
predictions with those found in these references, finding a good
agreement particularly with Ref.~\cite{Zhang:2012xn}.
In the case of Ref.~\cite{Hill:2009ek} the
agreement is better for antineutrinos than for neutrinos.
However, in the actual calculations, a
major difference arises from the fact that we are using a
substantially lower value of $C^A_5(0)=1.00$. Thus, our final
predictions for the dominant $\Delta$ contribution are about 30\%
or 45\% smaller than those of Refs.~\cite{Zhang:2012xn} and
Ref.~\cite{Hill:2009ek}, respectively. The error bands in our results
of Fig.~\ref{fig:cs_nu}, which are determined by the uncertainty
in $C^A_5(0)$, partially englobe these discrepancies. In this context,
it is worth reminding that the value of $C^A_5(0)=1.00 \pm 0.11$ used here
was determined in a
combined analysis of the neutrino induced pion production
ANL~\cite{Barish:1978pj,Radecky:1981fn} and
BNL~\cite{Kitagaki:1986ct,Kitagaki:1990vs} bubble chamber data. This was done
with a model closely resembling the present one i.e. including nonresonant
mechanisms, with the correct threshold behavior dictated by chiral symmetry,
the dominant $\Delta(1232)$ excitation and also deuteron effects~\cite{Hernandez:2010bx}.
Such a consistency with pion production data on the nucleon was not attempted in
Refs.~\cite{Hill:2009ek,Zhang:2012xn}. Actually, the ANL $\nu_\mu p \to \mu^- p \pi^+$ data are notably
overestimated in Ref.~\cite{Zhang:2012xn} as can be seen in Fig. 2 of that article.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.5\textwidth]{cs_delta.eps}
\caption{(color online) Top panel:
$\Delta P+C\Delta P$ cross sections obtained by us (solid lines)
and from Ref.~\cite{Hill:2009ek} (dashed lines), for $\nu N \to \nu N \gamma$
(red upper curves) and for $\bar\nu N \to \bar\nu N \gamma$ (blue lower
curves). For this comparison we have taken $C^A_5(0) = 1.2$ and no cut in $E_\gamma$
as in Ref.~\cite{Hill:2009ek}. Bottom panel: $\Delta P$ cross section obtained by us (solid lines)
and from Ref.~\cite{Zhang:2012xn} (dashed lines), for $\nu N \to \nu N \gamma$
(red upper curves) and for $\bar\nu N \to \bar\nu N \gamma$ (blue lower
curves). For this comparison we have adopted $C^A_5(0) = 1.14$ and an $E_\gamma \geq 0.2$ GeV cut,
as in Ref.~\cite{Zhang:2012xn}.}
\label{fig:delta_term}
\end{center}
\end{figure}
\subsection{Neutral current photon emission in nuclei}
\label{sec:res2}
For the present computations we take nuclear charge density distributions,
normalized to the number of protons in the nucleus, extracted from
electron scattering data~\cite{DeJager:1974dg}. The neutron matter density
profiles are parametrized in the same way as the charge densities
(but normalized to the number of neutrons) with small changes from
Hartree-Fock calculations~\cite{Negele:1972zp} and supported by pionic atom
data~\cite{GarciaRecio:1991wk}. The corresponding parameters are
compiled in Table I of Ref.~\cite{Nieves:1993ev}. Furthermore, these density
distributions have been deconvoluted to get center-point densities following the
procedure described in Ref.~\cite{Oset:1989ey}.
\subsubsection{Incoherent reaction: $1p1h\gamma$ contribution}
\label{sec:res2a}
In the left panels of Fig.~\ref{fig:incoh1}, we show our predictions
for the (anti)neutrino incoherent photon emission cross
sections on $^{12}$C as a function of the (anti)neutrino energy up to 2 GeV.
We observe that the neglect of nuclear medium corrections, as it
was done in the study of the NC$\gamma$ excess of events at MiniBooNE
of Ref.~\cite{Hill:2010zy}, is a quite poor approximation.
By taking into account Fermi
motion and Pauli blocking, the cross section already goes down by more
than 10\%. With the full model that also includes
the $\Delta$ resonance in-medium modification, the reduction is of the
order of 30\%. Furthermore, we corroborate the findings on nucleon targets
(Fig.~\ref{fig:cs_nu}) about the $N^*$ contributions
[mostly the $N(1520)$] being sizable above
$\sim 1.5$~GeV, specially for antineutrino cross sections.
\begin{figure}[h!]
\begin{center}
\makebox[0pt] {\includegraphics[width=0.5\textwidth]{csn.eps}
\includegraphics[width=0.5\textwidth]{csn_delta.eps}}
\caption{(color online) Left panel: Neutrino (top) and antineutrino (bottom)
incoherent photon emission cross sections on $^{12}$C.
All curves in this panel have been obtained with an $E_\gamma \geq 140$~MeV cut in the
phase space. Solid lines stand for
results from the complete model at the nucleon level,
while the dotted lines display
the predicted cross sections without the $N^*$ contributions.
Curves denoted as ``Free'' (upper blue curves) do
not include any nuclear correction:
the nuclear target is treated as a mere ensemble of
nucleons ($\sigma_A=Z\sigma_p+N\sigma_n)$.
Curves labeled as ``Full'' (lower red curves) take into account Pauli
blocking, Fermi motion and the in medium $\Delta$ resonance
broadening. The error bands show the uncertainty on the full model
that arises from the
determination of the axial $N\Delta$ coupling from data
($C^A_5(0) = 1.00 \pm 0.11$)~\cite{Hernandez:2010bx}.
Right panel: $\Delta P$ contribution to the neutrino
(top) and antineutrino (bottom) photon emission cross sections on
$^{12}$C from Ref.~\cite{Zhang:2012xn} compared to our predictions
for the same mechanism, adopting the same infrared photon energy cut $E_\gamma \geq 0.2$~GeV
and $C^A_5(0)=1.14$. The meaning of ``Free'' and ``Full'' labels is
the same as in the left plots.}
\label{fig:incoh1}
\end{center}
\end{figure}
In the right-hand plots of Fig.~\ref{fig:incoh1}, we compare our results
with the predictions of Ref.~\cite{Zhang:2012xn}. As in the nucleon
case (Fig.~\ref{fig:delta_term}), we
focus on the dominant $ \Delta P$ contribution and use the same $C_5^A(0)=1.14$
value and photon energy cut (200 MeV) as in Ref.~\cite{Zhang:2012xn}.
When all the nuclear corrections are neglected,
we certainly obtain the same curves as in Fig.~\ref{fig:delta_term},
but multiplied by the number of nucleons (12). As can be observed in the figure,
we find an excellent agreement both for neutrino and antineutrino cross sections.
However, nuclear medium effects turn out to be much more important, leading to a much
larger suppression ($\sim 50\%$), in the calculation of
Ref.~\cite{Zhang:2012xn} for neutrinos.
This seems surprising, first, because at this
moderately high neutrino energies, similar nuclear corrections should be obtained with
both models. In particular, one would not expect significant differences in
the $\Delta$ resonance broadening in the medium when calculated with Eq.~(\ref{eq:delta-self})
or with the spreading potential of Ref.~\cite{Zhang:2012xi}.
\footnote{We should mention that we agree better with the $\Delta P$ cross section of
Ref.~\cite{Zhang:2012xn} for neutrinos if we take an imaginary part of the
$\Delta$ selfenergy twice bigger than the one in Eq.~(\ref{eq:delta-self}).}
Because of the larger nuclear suppression, the $\Delta P$ cross section
found in Ref.~\cite{Zhang:2012xn} is smaller than the one obtained here
in spite of the 14\% larger $C_5^A(0)$. In the antineutrino cross sections, the
difference is not so large, and the medium effects shown in
Ref.~\cite{Zhang:2012xn} are only slightly greater than those found
in the present work. As a consequence of the large reduction of the $\Delta P$
contribution on $^{12}$C, the contact terms become relatively important
from $E_\nu =1$~GeV on, rapidly increasing and turning dominant
above 1.5 GeV (see Fig. 3 of Ref.~\cite{Zhang:2012xn}).
Indeed, contact terms compensate the suppression of the $ \Delta P$ mechanism,
so that the incoherent cross sections predicted in Ref.~\cite{Zhang:2012xn}
are comparable to ours in the 1~GeV region, but become about 40\% (70\%)
larger than our results for 2~GeV neutrinos (antineutrinos) even though
the contributions from resonances heavier that the $\Delta$ were not
taken into account.
\begin{figure}[htb]
\begin{center}
\makebox[0pt] {\includegraphics[width=0.33\textwidth]{incoherent_fivenuclear.eps}
\includegraphics[width=0.33\textwidth]{incoh_cos_fivenuclear.eps}\includegraphics[width=0.33\textwidth]{incoh_egamma_fivenuclear.eps}}
\caption{(color online) Neutrino (top) and antineutrino (bottom)
incoherent NC$\gamma$ total cross sections as a function of the
(anti)neutrino energy (left panels), photon angular (middle
panels) and photon energy (right panels)
differential distributions at $E_{\nu,\bar\nu}=$1 GeV. The angle
$\theta_\gamma$ is referred to the direction of the incoming
(anti)neutrino beam. Results for
different nuclei ($^{12}$C,$^{16}$O,$^{40}$Ar, $^{40}$Ca,$^{56}$Fe
and $^{208}$Pb) divided by the number of nucleons are shown.
All results are obtained with the full model, including nuclear effects and implementing an $E_\gamma \geq 140$~MeV cut.}
\label{fig:incoh2}
\end{center}
\end{figure}
In Fig.~\ref{fig:incoh2}, we show total NC$\gamma$ incoherent cross
sections for different nuclei (carbon, oxygen, argon, calcium, iron
and lead) as a function of the (anti)neutrino energy. We also display
photon angular and energy distributions for an incoming (anti)neutrino
energy of 1 GeV. We notice the approximated $A-$scaling present in the
results, which implies a mild $A$ dependence of nuclear effects.
Nevertheless, the cross section is smaller for heavier nuclei,
particularly $^{208}$Pb. We should stress that the observed deviation from scaling
cannot be explained only by neutron cross sections being smaller than proton ones
(around 15-20\% at $E_{\nu}\sim$ 1.5 GeV)\footnote{Note that the $\Delta$P contribution is the same on
protons and neutrons. Thus, this dominant mechanism does not contribute to such differences.}.
Concerning the kinematics of the emitted photons, the main
features are similar to those in Figs.~\ref{fig:1gev} and
\ref{fig:2gev} for scattering on single nucleons. As in that case,
the reaction is more forward for neutrinos than for
antineutrinos at $E_\nu =1$~GeV.
In the outgoing photon energy distributions (right panels),
the peak just above $E_\gamma=0.2$ GeV observed for nucleons is
reproduced here without any shift in the peak position but
with slightly larger width as the target mass increases.
\subsubsection{Coherent reaction}
\label{sec:res2b}
Total NC$\gamma$ coherent cross sections on
carbon as a function of the (anti)neutrino energy are presented in
Fig.~\ref{fig:coh1}.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.5\textwidth]{cs_coherent.eps}
\caption{(color online) Neutrino (top) and antineutrino (bottom) total
NC$\gamma$ coherent cross sections on $^{12}$C, as a function of the
(anti)neutrino energy. A photon energy cut of $E_\gamma \geq 140$~MeV has
been implemented. Red solid lines stand for
results from the complete model derived in
this work, including $\Delta$ resonance broadening, with error bands
determined by the uncertainty of $\pm 0.11$ in $C^A_5(0)$~\cite{Hernandez:2010bx}.
The solid blue lines below, labeled as ``no $N^*$'', display the
predicted cross sections without the $N^*$ amplitudes, while the magenta dotted ones
are the contributions from the $(\Delta P+ C\Delta P)$ mechanisms.
We also show the predictions of Ref.~\cite{Zhang:2012xn} for nucleon and $\Delta$ mechanisms
(red solid lines in Fig.4 of this reference).}
\label{fig:coh1}
\end{center}
\end{figure}
We display our results from the full calculation,
from $(\Delta P+ C\Delta P)$ alone, and without the mechanisms
from second $N^*$ resonance region.
The $N^*$ contributions
are quite small in the coherent channel, while the
$\Delta$ is absolutely dominant in both the neutrino and
the antineutrino modes. Nucleon-pole contributions are negligible
because the coherent kinematics favors a strong cancellation between
the direct and crossed terms of the amplitude. A similar effect has
been observed in weak coherent pion production~\cite{AlvarezRuso:2007it}.
For comparison, the predictions from the $(\Delta P+
C\Delta P+NP+CNP)$ part of the model of Ref.~\cite{Zhang:2012xn}
are also plotted. They are slightly above our corresponding results
(without $N^*$), and within the uncertainty band of our full-model curve,
up to (anti)neutrino energies of 1.4--1.5 GeV. Above these energies,
there is a a change of slope and a pronounced enhancement~\cite{Zhang:2012xn}.
Moreover, in the model of this reference,
the cross section above $E_{\nu,\bar\nu}=0.65$ GeV is not dominated by
the $(N+\Delta)$ mechanism, but by contact terms from higher order effective
Lagrangians whose extrapolation to higher energies is uncertain.
Indeed, for some choices of parameters, coherent cross sections as large
as $25\times 10^{-42}$ cm$^2$ were obtained
for $E_{\nu,\bar\nu}=1.5$ GeV~\cite{Zhang:2012xn}.
This amounts to a factor 3-4 larger
than our predictions. We should remind here that below 500 MeV, the
contact terms in the nucleon amplitudes are very small as
expected based on the power counting established in
Ref.~\cite{Serot:2012rd}. Because of the substantial reduction of the
$\Delta$ mechanisms, the contact terms in Ref.~\cite{Zhang:2012xn}
acquire further relevance when the processes take place in nuclei,
specially for the coherent reaction.
Our results for coherent NC$\gamma$ total and differential cross sections on
different nuclei are shown in Fig.~\ref{fig:coh2}.
\begin{figure}[h!]
\begin{center}
\makebox[0pt] {\includegraphics[width=0.34\textwidth]{coherent_fivenuclear.eps}
\includegraphics[width=0.34\textwidth]{coh_cos_fivenuclear.eps}\includegraphics[width=0.34\textwidth]{coh_egamma_fivenuclear.eps}}
\caption{(color online) Neutrino (top) and antineutrino (bottom) total
cross sections (left panels) photon angular (middle
panels) and photon energy (right panels) differential distributions
for the coherent NC$\gamma$ reaction, obtained with our full model.
The angle $\theta_\gamma$ is referred to the direction of the incoming
(anti)neutrino beam. The kinematic region of $E_\gamma < 140$~MeV has been cut out.
Results for different nuclei ($^{12}$C,$^{16}$O,$^{40}$Ar, $^{40}$Ca,$^{56}$Fe
and $^{208}$Pb) divided by the number of nucleons are shown.}
\label{fig:coh2}
\end{center}
\end{figure}
Neutrino (antineutrino) coherent cross sections are about a factor 15
(10) smaller than the incoherent ones given in Fig.~\ref{fig:incoh2}.
Thus, the relative relevance
of the coherent channel with respect to the incoherent channel is
comparable, if not greater than in the pion production reactions
induced by neutrinos and antineutrinos, where it is of the order of
few per cent~\cite{Hernandez:2010jf,Hernandez:2013jka}. Notice that in
these latter reactions the coherent cross section is
further reduced (by around a factor
of two) because of the strong distortion of the outgoing pion, which is
not present in photon production. It is also true that the incoherent cross section
is reduced ($\sim 20-30$\%) by final state interactions, again absent for photons.
The coherent cross sections neither scale with $A$, like the incoherent one approximately does, nor with
$A^2$ as one would expect from the coherence of the dominant isoscalar
$\Delta P$ mechanism (sum of neutron and proton amplitudes). This is due to the presence
of the nuclear form factor (Fourier transform of the nuclear density
for momentum $\vec{q}-\vec{k}_\gamma$), see the first paragraph of Sec.~\ref{sec:coherent}
and Eq.~(\ref{eq:Jmunu2}). The nuclear form
factor gets its maximum values when $\vec{q} = \vec{k}_\gamma$, which corresponds to $q^2 = 0$.
In this forward kinematics, the lepton tensor $L_{\mu\sigma}^{(\nu,\bar\nu)} \sim q_\mu q_\sigma$, and the
vector part of the amplitude squared is zero due to CVC. Furthermore, the axial contribution, which is purely
transverse $\sim (\vec{k}_\gamma \times \vec{q})$ also vanishes. Therefore, the largest differential
cross sections arise in kinematics that optimize the product of the amplitude squared of the elementary process
times the nuclear form factor. Such a balance also appears in
the $(^3{\rm He},^3{\rm H}\,\pi^+)$ reaction on nuclear targets~\cite{FernandezdeCordoba:1992ky} or
in electron and photon induced reactions, making the electromagnetic coherent pion
production cross section a rather small fraction of the total inclusive nuclear absorption
one~\cite{Carrasco:1991we,Hirenzaki:1993jc}.
The described pattern strongly influences the photon angular
dependence of this reaction shown in the middle panels of
Fig.~\ref{fig:coh2} although in a non-trivial way because the
$\theta_\gamma$ angle is given with respect to the direction of
the incoming (anti)neutrino beam; it is not the angle
formed by $\vec{q}$ and $\vec{k}_\gamma$, which is not observable.
Actually, for each value of $\theta_\gamma$, and integration over all possible $\vec{q}$ is
carried out. The details of the angular distributions are determined by interferences
between the dominant $\Delta P$ mechanism and the $C\Delta P$ and $N(1520)$ ones, enhanced by the
kinematic constrains imposed by the nuclear form factor. The impact of the latter is apparent
in the width of the angular distributions which are narrower for heavier nuclei.
Finally, in Fig.~\ref{fig:coh2} we display the outgoing photon energy
distributions (right panels). In the coherent NC$\gamma$ reaction,
there are two massless particles in the final state, and a third one (the nucleus)
which is very massive and has a small (negligible) kinetic energy
but can carry large momenta. The prominent peak observed for all nuclei is
due to the dominant $\Delta$
resonance~\footnote{The energy of the resonant photons in LAB can be estimated from
$M_R^2 \approx (k_\gamma + p^\prime)^2$. Taking $p^\prime$ from Eq.~(\ref{eq:pmu}) and for
the situation $\vec{k}_\gamma \approx \vec{q}$ favored by the nuclear form factor, one
finds that $k_{\gamma(R)}^0 \approx (M_R^2 - M^2)/(2 M)$. This gives 340~MeV for the $\Delta(1232)$
and 760~MeV for the $N(1520)$.}
shifted to slightly lower
invariant masses mostly by the energy dependence of the $\Delta$ width and the interference
with the $C\Delta P$ mechanism. The peak position does not change appreciably from nucleus to nucleus,
but it gets broader as $A$ increases. The second, smaller and broader peak that can be
discerned for neutrinos but not for antineutrinos corresponds to the excitation of the $D_{13}(1520)$
resonance.
\section{Conclusions}
\label{sec:concl}
Neutral current photon emission on nucleons and nuclei
at intermediate energies has been theoretically investigated.
We have developed a microscopic model for these reactions,
in line with previous work on weak
pion production~\cite{Hernandez:2007qq,AlvarezRuso:2007it, Amaro:2008hd,Hernandez:2013jka}.
We have critically reviewed previous models for the
NC$\gamma$ reaction on single nucleons~\cite{Zhang:2012xn,Hill:2009ek,Serot:2012rd} and
nuclei~\cite{Zhang:2012xn,Zhang:2012aka,Zhang:2012xi} and compared our
results with those found in these references. From such a comparison, we have identified
some aspects of the above studies that either needed to be improved or
that were sources of uncontrolled systematic corrections.
NC$\gamma$ processes are important backgrounds for $\nu_\mu \to \nu_e$
and $\bar\nu_\mu \to \bar \nu_e$ appearance oscillation experiments
when photons are misidentified as $e^\pm$ from CCQE
scattering of $\nu_e (\bar{\nu}_e)$. At the relevant energies for
MiniBooNE and T2K experiments, the reaction is dominated by the
weak excitation of the $\Delta(1232)$ resonance and its subsequent
decay into $N\gamma$. Besides, we have also considered
non-resonant amplitudes that, close to threshold,
are fully determined by chiral symmetry, and those driven by
nucleon excited states from the second resonance region. Among
the latter ones, we have found a sizable contribution of the $D_{13}(1520)$ state
for (anti)neutrino energies above 1.5 GeV.
The model on the nucleon is extended to nuclear targets taking into account Fermi motion,
Pauli blocking and the in-medium modifications of the $\Delta$ properties in a local
Fermi gas, with Fermi momenta determined from proton and neutron density distributions.
We have predicted different observables for several nuclei, including some of the
common ones in current and future neutrino detectors (carbon, oxygen, argon, iron).
The importance of nuclear corrections in both the coherent and
incoherent channels has been stressed. The $A$ dependence of the cross section, which is
different for the coherent and incoherent reactions, has also been discussed.
In the light of our results, a new analysis of the NC induced photon production at MiniBooNE with the
present model, aiming at the clarification of the role played by NC$\gamma$ events in the low-energy
excess observed in this experiment, looks timely and important. It will be the subject of future research.
|
\section{Introduction}
The study of Dirichlet series in several complex variables has seen
much development in recent years. Multiple Dirichlet series (MDS) can arise
from metaplectic Eisenstein series, zeta functions of prehomogeneous
vector spaces (PVS), height zeta functions or multiple zeta values. This
list is far from exhaustive. In general, completely different
techniques are involved in the study of different series arising in
these different contexts. Partially for this reason, it is of great
interest to identify examples of multiple Dirichlet series which lie
in two or more of the areas above.
This paper investigates one such example. We study in detail a
Shintani zeta function associated to a certain prehomogeneous vector
space and show that it coincides with a {\em Weyl group multiple Dirichlet
series} (WMDS) of the form introduced in BBCFH (\cite{BBCFH}). This latter series (in
three complex variables) is also
a Whittaker function of a Borel Eisenstein on the metaplectic double
cover of $\mathrm{GL}_4$.
The seminal work of Bhargava \cite{bhargava} generalizing Gauss's composition law on binary quadratic forms starts with investigating the rich structure of integer orbits of $2 \times 2 \times 2$-cubes acted on by $\mathrm{SL}_2(\mathbb{Z}) \times \mathrm{SL}_2(\mathbb{Z})\times \mathrm{SL}_2(\mathbb{Z})$, which refers to the $D_4$ case among the list of classification of PVS given in Sato-Kimura \cite{satokimura}. Bhargava shows that the set of projective integer orbits with given discriminant has a group structure and it is isomorphic to the square product of narrow class group of the quadratic order with that discriminant.
We begin with some definitions. Let $G_{\mathbb{C}}$ be a connected complex Lie group, usually we assume that $G_{\mathbb{C}}$ is the complexification of a real Lie group $G_{\mathbb{R}}$. A PVS $(G_{\mathbb{C}}, V_{\mathbb{C}})$ is a complex finite dimensional vector space $V_{\mathbb{C}}$ together with a holomorphic representation of $G_{\mathbb{C}}$ such that $G_{\mathbb{C}}$ has an open orbit in $V_{\mathbb{C}}$. One of the important properties is that if $V_{\mathbb{C}}$ is a PVS for $G_{\mathbb{C}}$ then there is just one open orbit, and that orbit is dense (see \cite[Chapter X]{knapp}). Let $P$ be a complex polynomial function on $V_{\mathbb{C}}$. We call it a relative invariant polynomial if $P(g v) = \chi(g) P(v)$ for some rational character $\chi$ of $G_{\mathbb{C}}$ and all $g \in G_{\mathbb{C}}$, $v\in V_{\mathbb{C}}$. We say the PVS has $n$ relative invariants if $n$ algebraically independent relative invariant polynomials generate the invariant ring. We define the set of semi-stable points $V_{\mathbb{C}}^{ss}$ to be the subset on which no relative invariant polynomial vanishes.
Let $V_{\mathbb{Z}} = \mathbb{Z}^2 \otimes \mathbb{Z}^2 \otimes \mathbb{Z}^2$ and consider the action of $ B'_2(\mathbb{Z}) \times B'_2(\mathbb{Z}) \times \mathrm{SL}_2(\mathbb{Z})$, where $B'_2(\mathbb{Z})$ is the subgroup of lower-triangular matrices in $\mathrm{SL}_2(\mathbb{Z})$ with positive diagonal elements. We denote $A \in V_{\mathbb{Z}}$ by $\left( \left( \begin{matrix}
a & b \\
c & d \end{matrix} \right), \left( \begin{matrix}
e & f \\
g & h \end{matrix} \right) \right)$ for simplicity. The complex group $ B'_2(\mathbb{C}) \times B'_2(\mathbb{C}) \times \mathrm{GL}_2(\mathbb{C})$ acting on $V_{\mathbb{C}}$ is again a PVS and will have three relative invariants. They are $ m(A) = ad- bc, n(A)= ag - ce$ and the discriminant $D(A) =\mathrm{disc}(A)$.
The primary object of study of PVS is the Shintani zeta function, see \cite{satoshintani} for the introduction and \cite{shintani} for the application to the average values of $h(d)$, the number of primitive inequivalent binary quadratic forms of discriminant $d$. The Shintani zeta function in three variables associated to the PVS of $2 \times 2 \times 2$-cubes is defined to be:
$$Z_{\rm{Shintani}}(s_1, s_2, w) = \sum_{ A \in \left( B'_2(\mathbb{Z}) \times B'_2(\mathbb{Z}) \times \mathrm{SL}_2(\mathbb{Z}) \right) \backslash V^{ss}_{\mathbb{Z}}} \frac{1}{|D(A)|^w |m(A)|^{s_1} |n(A)|^{s_2}} .$$
We denote the partial sum of Shintani zeta function by
\begin{align*}
Z_{\rm{Shintani}}^{\rm{odd}}(s_1, s_2, w) = \sum_{\substack{ A \in \left( B'_2(\mathbb{Z}) \times B'_2(\mathbb{Z}) \times \mathrm{SL}_2(\mathbb{Z}) \right) \backslash V^{ss}_{\mathbb{Z}} \\ D(A) \ \rm{odd}}} \frac{1}{|D(A)|^w |m(A)|^{s_1} |n(A)|^{s_2}} .
\end{align*}
We will show that it is closely related to another multiple Dirichlet series which arises in the Whittaker expansion of the Borel Eisenstein series on the metaplectic double cover of $\mathrm{GL}_4$, which is found to be the WMDS associated to the root system $A_3$ and it is of the form:
\begin{align*}
Z_{\rm{WMDS}} (s_1, s_2, w) = \sum_{ D \ \mathrm{odd} \ \mathrm{discriminant} } \frac{1}{|D|^w} \sum_{m, n>0} \frac{\chi_D(\hat{m}) \chi_D(\hat{n})}{m^{s_1} n^{s_2}}a(D, m, n)
\end{align*}
where $\hat{m}$ denotes the factor of $m$ that is prime to the square-free part of $D$ and $\chi_D$ is the quadratic character associated to the field extension $\mathbb{Q}(\sqrt{D})$ of $\mathbb{Q}$. A precise formula of the coefficients $a(D, m, n)$ will be given in section 4.
The main results of this paper, given in Theorems 1.1, 1.2 and 1.3 below, are as follows:
\begin{itemize}
\item[(1)]We give an explicit description of the Shintani zeta function associated to the PVS of $2 \times 2 \times 2$-cubes.
\item[(2)]We show how the series is related to the quadratic WMDS associated to the root system $A_3$.
\item[(3)]We give an arithmetic meaning to the semi-stable integer orbits of the PVS.
\end{itemize}
Denote by $A(d, a)$ the number of solutions to the congruence $x^2 =d \ (\text{mod} \ a) $.
\begin{theorem}
The Shintani zeta function associated to the $\mathrm{PVS}$ of $2\times 2 \times 2$-cubes is given by:
\begin{align*}
Z_{\rm{Shintani}}(s_1, s_2, w) &= \sum_{ A \in \left( B'_2(\mathbb{Z}) \times B'_2(\mathbb{Z}) \times \mathrm{SL}_2(\mathbb{Z}) \right) \backslash V^{ss}_{\mathbb{Z}}} \frac{1}{|D(A)|^w |m(A)|^{s_1} |n(A)|^{s_2}} \\
& =\sum_{D = D_0 D_1^2\neq 0} \frac{1}{|D|^w} \sum_{m, n > 0} \frac{B(D, m,n)}{m^{s_1}n^{s_2}}, \end{align*}
where
$$B(D, m, n) =\sum_{\substack{d|D_1 \\ d|m,\ d|n}} d \cdot A(D/ d^2, 4m/d) \cdot A(D/d^2, 4n/d).$$
\end{theorem}
\begin{theorem}
The Shintani zeta function can be related to the $A_3$-$\mathrm{WMDS}$ in the following way:
\begin{align*}
Z^{\mathrm{odd}}_{\rm{Shintani}}(s_1, s_2, w) = & 4 (1-2^{-s_1})(1+ 2^{-s_1})
(1-2^{-s_2})(1+ 2^{-s_2}) \zeta(s_1) \zeta(s_2) \\
& \cdot Z_{\rm{WMDS}} (s_1, s_2, w).
\end{align*}
where $\zeta(s)$ means the Riemann zeta function.
\end{theorem}
Lastly, we give the arithmetic meaning to the semi-stable integer orbits of the PVS by showing that:
\begin{theorem}
There is a natural map, which is a surjective and finite morphism,
$$\left( B_2'(\mathbb{Z}) \times B_2'(\mathbb{Z}) \times \mathrm{SL}_2(\mathbb{Z}) \right)\backslash V^{ss}_{\mathbb{Z}} \to \mathrm{Iso} \backslash \{(R;I_1, I_2): R/I_1 \cong \mathrm{N}(I_1)\mathbb{Z}, R/I_2 \cong \mathrm{N}(I_2)\mathbb{Z} \},$$
where $R$ is an oriented quadratic ring and $I_i's$ are the oriented ideals in $R$ with the norm $\mathrm{N}(I_i)$. The cardinality $ n( R; I_1, I_2)$ of the fiber is equal to
$$\sigma_1 \left( g.c.d.(D_1, a_1, a_2)\right),$$
where $D =D_0 D_1^2= \mathrm{disc}(R)$ with $D_0$ is square-free, and $a_i=\mathrm{N}(I_i)$.
It further satisfies
$$\sum _{\substack{(R; I_1, I_2)/\sim \\ \mathrm{N}(I_i)=a_i}} n (R; I_1, I_2) = B(D, a_1, a_2).$$
\end{theorem}
The organization of this paper is as follows. In section $2$, we will review a double Dirichlet series arising from three different approaches. We show the connection between Shintani's PVS approach and the $A_2$-WMDS approach. In section $3$, we will investigate the structure of integer orbits of the PVS of $2\times 2 \times2$-cubes, using the reduction theory, to derive the main formula in Thm 1.1. In section $4$, we will show its connection to the $A_3$-WMDS. The main idea of the proof is to consider the generating function for the $p$-parts of $a(D, m,n)$. From the construction of an $A_3$-WMDS, the $p$-parts of the rational function which is invariant under the Weyl group action is given explicitly by taking the residue of the convolution of two rational functions of $A_2$ root system. We show it coincides with the generating function of $a(D, m,n)$. In the section $5$, we will show the set of semi-stable integer orbits naturally encodes the arithmetic information by showing it maps finitely and surjectively to the moduli space of isomorphism classes of pairs $(R; I_1, I_2)$, where $R$ is an oriented quadratic ring and $I_i's$ are the oriented ideals in $R$. We will recall the definition of orientation of a quadratic ring and its ideal. We further show that each fiber is counted by a divisor function.
\subsection*{Acknowledgements}
The author would like to thank Gautam Chinta and Takashi Taniguchi for useful discussions. He is grateful to Gautam Chinta for making valuable comments. He also thanks the organizers of the ICERM Semester Program on Automorphic Forms, Combinatorial Representation Theory and Multiple Dirichlet Series from January to May of 2013, where some of this work was carried out.
\section{$A_2$ Weyl Group Dirichlet series}
In this section we will introduce three double Dirichlet series and show that they are all essentially the same. The results are not new, Diamantis and Goldfeld in \cite{diamantisgoldfeld} show their relations using a type of converse theorem, and Shintani in \cite[\S 2]{shintani} compares his series to Siegel's, but the computations in section will serve as a prototype for the more involved computations involving 3-variable Dirichlet series in the latter sections.
\subsection{Siegel Double Dirichlet Series}
The quadratic multiple Dirichlet series first appeared in the paper of Siegel \cite{siegel}, and its twisted Euler product with respect to one variable was given explicitly. Siegel constructed his series as the Mellin transform of an Eisenstein series of half integral weight on the congruence subgroup $\Gamma_0(4)$. In this section we will first recall the definition of Siegel's double Dirichlet series and then, using the multiplicative property of Euler products, prove it can be expressed as a sum formed from quadratic characters.
Denote by $A(d, a)$ the number of solutions to the quadratic congruence equation $x^2 = d \ ( \mathrm{mod} \ a)$.
Then Siegel's double Dirichlet series is defined to be:
$$ Z_{\text{Siegel}} (s, w) = \sum_{d\neq 0} \frac {1}{|d|^w}\sum_{a \neq 0}\frac{A(d, a)}{|a|^s}.$$
For any positive prime integer $p$ and any integer $d$, we define a generating series for the $p$-part of the inner summation of above series to be
\begin{align}
&f_p(d, s)= (1-p^{-s})\sum^{\infty}_{l=0}A(d,p^l) p^{-ls} \ (p \neq 2) ,\notag \\
&f_2(d, s)= (2^s-1) \sum^{\infty}_{l=1}A(d,2^l)2^{-ls}. \notag
\end{align}
Siegel shows in \cite[\S 4]{siegel} that
\begin{align*}
\frac{1- \chi_d(p)p^{-s}}{1-p^{-2s}}f_p(d, s)= \begin{cases} p^{\alpha(1-2s)}+ \left(1-\chi_d(p)p^{-s}\right) \sum^{\alpha-1}_{l=0}p^{l(1-2s)} \ &\text{($p\neq2$)}, \\
\frac{1+\chi_d(2)}{1+2^{-s}}+ \left(2^{1-s}-\chi_d(2)\right)\sum^{\alpha}_{l=0}2^{l(1-2s)} \ &\text{($p = 2$)},
\end{cases}
\end{align*}
where $\chi_d$ is the quadratic character associated to the field $\mathbb{Q}(\sqrt{d})$ and $ p^ {2\alpha}$ is the highest power of $p$ which divides $d/d^{*}$, $d^{*}$ is the fundamental discriminant of the field $\mathbb{Q}(\sqrt{d})$. Now the inner summation of Siegel's double Dirichlet series can be expressed as a normalized quadratic L-function.
\begin{proposition}
Fix an integer $d \neq 0$, then
$$\sum_{a < 0}\frac{ A(d,a)}{|a|^{s}} = \sum_{a > 0}\frac{ A(d,a)}{a^{s}} = \zeta(2s)^{-1} \zeta (s) L(s, \chi_d) P(d, s) $$
where the last term is
\begin{align}
P(d,s) =& 2^{-s} \prod_{p \neq 2} \left( p^{\alpha(1-2s)} + (1- \chi_d(p) p^{-s} )\sum^{\alpha-1}_{l=0}p^{l(1-2s)}\right) \notag \\
&\cdot \left( \frac{1+\chi_n(2)}{1+2^{-s}}+\frac{1-\chi_d(2) 2^{-s}}{1+2^{-2s}}(2^{s}-1)+
\left(2^{1-s}-\chi_n(2) \right)\sum^{\alpha}_{l=0}2^{l(1-2s)} \right). \notag
\end{align}
\end{proposition}
\begin{proof}
The first equality of the claim follows from the fact that
$$A(d, a) = A(d, -a).$$
For the second equality, first note that the multiplicative property holds
$$A(d,m) A(d,n) = A(d,mn)$$
for any pairs of coprime positive integers $m$ and $n$, by the Chinese remainder theorem.
Then the results follows from two equations
$$\frac{1- \chi_n(p)p^{-s}}{1-p^{-2s}}\left(1-p^{-s}\right) \sum_{l=0}^{\infty} A\left(d,p^l\right) p^{-ls} =p^{\alpha(1-2s)}+ \left(1-\chi_n\left(p\right)p^{-s}\right) \sum^{\alpha-1}_{l=0}p^{l\left(1-2s\right)} ,$$
$$\frac{1- \chi_n(2)2^{-s}}{1-2^{-2s}}\left(2^{s}-1\right) \sum_{l=1}^{\infty} A\left(d,2^l\right) 2^{-ls} =\frac{1+\chi_n(2)}{1+2^{-s}}+ \left(2^{1-s}-\chi_n(2)\right)\sum^{\alpha}_{l=0}2^{l(1-2s)} ,$$
as well as the multiplicative property.
\end{proof}
\subsection{Shintani Double Dirichlet Series}
Another approach to the theory of double Dirichlet series is based on the zeta function associated to the prehomogeneous vector space (PVS) developed by M. Sato and Shintani in \cite{satoshintani}. In \cite{shintani}, where the double Dirichlet series associated to the PVS of binary quadratic forms acted on by the Borel subgroup of $\mathrm{GL}_2(\mathbb{C})$ is studied in detail, the author obtains the mean values of class numbers of primitive and integral binary quadratic forms. It should be mentioned that in \cite{Hsaito} the essentially same double Dirichlet series is discovered as the zeta function associated to another PVS. In this section, we will recall the Shintani zeta function in two variables arising from the PVS approach.
Now we let $B'_2({\mathbb{C}})$ be the subgroup of lower-triangular matrices in $G_{\mathbb{C}} =\mathrm{GL}_2({\mathbb{C}})$ and let $\rho$ be the representation of $\mathrm{GL}_2({\mathbb{C}})$ acting on the three dimensional vector space $V_{\mathbb{C}}=\{Q(u, v)= a u^2 +b uv+ cv^2|(a, b, c )\in \mathbb{C}^3\}$ of binary quadratic forms
as follows
$$\rho(g) (Q)(u, v) = Q(au +cv , bu+dv)$$
where $ g =\left( \begin{array}{cc}
a & b \\
c & d \end{array} \right)$. It is well known that $(B'_2(\mathbb{C}), V_{\mathbb{C}})$ is a PVS and there are two relative invariants for the action of $ B'_2(\mathbb{C})$ on $V_{\mathbb{C}}$, namely, the discriminant $\mathrm{disc}(Q)=b^2 -4 ac$ of the quadratic form and $a= Q(1, 0)$. These two invariants freely generate the ring of relative invariants.
Define $B'_2(\mathbb{Z})$ to be the Borel subgroup in $\mathrm{SL}_2(\mathbb{Z})$ with positive diagonal elements. The Shintani zeta function associated to the PVS $(B'_2({\mathbb{C}}), V_{\mathbb{C}})$ is defined to be
\begin{align*}
Z_{\mathrm{Shintani}}( s, w;B'_2)& = \sum_{Q \in B'_2({\mathbb{Z}}) \backslash V^{ss}_{\mathbb{Z}}}\frac{1}{|Q(1, 0)|^s |\mathrm{disc}(Q)|^w} \notag \\
&=\sum_{a\neq 0} \frac{1}{|a|^{s}} \sum_{\substack{0 \leq b \leq 2a -1 \\ c: b^2 - 4a c \neq 0}} \frac{1}{|b^2 - 4 a c|^w }. \notag
\end{align*}
where $V^{ss}_{\mathbb{Z}}$ is the semi-stable subset of (not necessarily primitive) quadratic forms with $Q(1, 0) \neq 0$ and non-zero discriminant.
Alternatively, we can express the Shintani zeta function as
$$ \sum _{Q \in \mathrm{SL}_2({\mathbb{Z}}) \backslash V^{ss}_{\mathbb{Z}}} \frac{1}{|\mathrm{disc}(Q)|^{w}} \sum _{\gamma \in B'_2({\mathbb{Z}}) \backslash \mathrm{SL}_2({\mathbb{Z}}) /\mathrm{stab}_{Q}}\frac{1}{|\gamma \circ Q(1, 0)|^{s}}.$$
For the rest of this section we suppress the $B'_2$ from the notation.
We also define
\begin{align*}
Z_{\mathrm{Shintani}}^{\mathrm{odd}}( s, w)& = \sum_{\substack{Q \in B'_2(\mathbb{Z}) \backslash V^{ss}_{\mathbb{Z}} \\ \mathrm{disc}(Q) \ \mathrm{odd}}} \frac{1}{|Q(1, 0)|^s |\mathrm{disc}(Q)|^w} \notag \\
&=\sum_{a\neq 0} \frac{1}{|a|^{s}} \sum_{\substack{0 \leq b \leq 2a -1 \\ c: b^2 - 4a c \ \mathrm{odd}}} \frac{1}{|b^2 - 4 a c|^w }. \notag
\end{align*}
\begin{remark}
In section 5, we will show that the orbits in $B'_2({\mathbb{Z}}) \backslash V^{ss}_{\mathbb{Z}, \rm{primitive}}$ parameterize the isomorphism classes of the pairs $(R, I)$, where $R$ is an oriented quadratic ring and $I$ is an oriented ideal with cyclic quotient in $R$. We will call $(R_1, I_1)$ and $(R_2, I_2)$ isomorphic if there is a ring isomorphism from $R_1$ to $R_2$ preserving the orientation and sending $I_1$ to $I_2$.
\end{remark}
\begin{lemma}
The Shintani zeta function can be written as
$$Z_{\rm{Shintani}}( s, w)= \xi_1(s, w) + \xi_2(s, w),$$
where
$\xi_i(s, w)= \sum_{a, d > 0 } \frac{A( (-1)^{i-1}d,4a)}{a^s d^w} .$
\end{lemma}
\begin{proof}
First note that under the action of $g= \left( \begin{array}{cc} 1 & 0\\ m &1 \end{array} \right)$ on the quadratic form $Q(u, v) = a u^2 + buv + cv^2$, the middle coefficient $b$ is mapped to $b+ 2 a m$. Given non-zero integers $a$ and $d$, the number of the solutions to $b^2 - 4a c = d$ with $0 \leq d \leq 2a-1$ and $ c\in \mathbb{Z}$ is equal to $\frac{A(d, 4a)}{2}$. So we have the equality
$$Z_{\rm{Shintani}}( s, w) = \frac{1}{2} \sum_{a,d \neq 0} \frac{A(d, 4a) }{|a|^s |d|^w}.$$
On the other hand, $A(d, 4a) = A(d, -4a) $. Therefore,
$$Z_{\rm{Shintani}}( s, w) =\sum_{a,d > 0} \frac{A(d, 4a) }{a^s d^w} + \sum_{a,d > 0} \frac{A(-d, 4a) }{a^s d^w}.$$
The result follows.
\end{proof}
As a corollary to Proposition 2.1, we can express the inner summation of the Shintani zeta function in terms of a quadratic Dirichlet $L$-function.
\begin{corollary}
Fix $d \neq 0$. Then we have
$$\sum_{a >0}\frac{ A(d,4a)}{a^{s}} = \zeta(2s)^{-1} \zeta (s) L(s, \chi_d) P'(d, s), $$
where the last term is
\begin{align}
P'(d,s) =& 4^{s} \prod_{p \neq 2} \left( p^{\alpha(1-2s)} + \left(1- \chi_d(p) p^{-s} \right)\sum^{\alpha-1}_{l=0}p^{l(1-2s)} \right)\notag \\
&\cdot \left(\frac{1+\chi_n(2)}{1+2^{-s}}+ \left(2^{1-s}-\chi_n(2)\right)\sum^{\alpha}_{l=0}2^{l(1-2s)} -
\frac{1- \chi_n(2)2^{-s}}{1-2^{-2s}}\left(1-2^{-s}\right)\right) .\notag
\end{align}
\end{corollary}
\begin{proof}
As $\sum_{a>0} \frac{A(d, 4a)}{a^s} = 4^s \sum_{a >0}\frac{ A(d,4a)}{(4a)^{s}}$, by the proof of proposition 2.1, we need only to correct the generating function at prime $p=2$ in order to incorporate the factor $4$, in which case the generating function should be
$$ \sum_{l=0}^{\infty} A(d,4 \times 2^l) 2^{-ls}= 4^s \sum_{l=2}^{\infty} A(d, 2^l) 2^{-ls}.$$
While the function on the right hand side satisfies
\begin{align}
\frac{1- \chi_d(2)2^{-s}}{1-2^{-2s}}(2^{s}-1) \sum_{l=2}^{\infty} A(d,2^l) 2^{-ls}= & \frac{1- \chi_d(2)2^{-s}} {1- 2^{-2s}} f_2(d, s) \notag \\
& - \frac{1- \chi_n(2)2^{-s}}{1-2^{-2s}}(2^{s}-1) A(d, 2) 2^{-s} .\notag
\end{align}
Note that $A(d, 2) =1$, therefore, using the multiplicative property of $A(d, \cdot)$, we have
$$\sum_{a>0} \frac{A(d, 4a)}{a^s} = 4^s \sum_{a >0}\frac{ A(d,4a)}{(4a)^{s}} = \zeta(2s)^{-1} \zeta (s) L(s, \chi_d) P'(d, s).$$
The result follows.
\end{proof}
\subsection{$A_2$-Weyl Group Multiple Dirichlet Series}
Weyl group multiple Dirichlet series are a class of multiple Dirichlet series coming from Eisenstein series on metaplectic groups. The simplest example is the quadratic $A_2$ Weyl group double Dirichlet series (see \cite{chintagunnells}), it is defined as
$$ Z_{A_2}(s,w) =\sum_{\substack{ m > 0, \\ D \ \mathrm{odd} \ \mathrm{discriminant}}} \frac{\chi_D(\hat{m})}{m^s |D|^w}a(D, m), $$
where $\hat{m}$ is the factor of $m$ that is prime to the square-free part of $D$ and $\chi_D$ is the quadratic character associated to the field extension $\mathbb{Q}(\sqrt{D})$ of $\mathbb{Q}$. Moreover, the multiplicative factor $a(D, m)$ is defined by
$$a(D,m) = \prod_{ p^k||D,p^l||m}a(p^k, p^l)$$
and
$$a(p^k, p^l) = \begin{cases} \min(p^{k/2},p^{l/2}) & \mbox{if}\ \min(k,l)\ \mbox{is even}, \notag \\
0 & \mbox{otherwise}. \notag \end{cases}$$
Regarding to the relation between this quadratic $A_2$-WMDS and the Shintani zeta function of PVS of binary quadratic forms, we have
\begin{proposition}
Let $D$ be an odd integer. We can relate the inner sum of the Shintani zeta function $Z_{\mathrm{Shintani}}^{\mathrm{odd}}(s, w)$ with that of the quadratic $A_2$ Weyl group multiple Dirichlet series $Z_{A_2}(s, w)$ in the following way
$$ \sum_{m>0} \frac{A(D, 4m)}{ m^s} =\tilde{P}_2(D, s) \zeta(s) \sum_{m>0} \frac{\chi_D(\hat{m}) a(D, m) }{ m^s},$$
where
\begin{align*}
\tilde{P}_2(D, s) = \begin{cases} 2 (1 - 2^{-s}) ( 1+ 2^{-s}) & D \equiv 1 \ (\mathrm{mod} \ 4) ,\\
0 & \mathrm{otherwise}.
\end{cases}
\end{align*}
\end{proposition}
\begin{proof}
Analogous to the $p$-part formula of $a(p^k, p^l)$, we will show that for an odd prime $p$
$$A(p^k, p^l) = \begin{cases} 2 a(p^k, p^l) &\text{if} \ k<l, \\ p^{\lfloor l /2 \rfloor} &\mbox{otherwise}. \end{cases}$$
First consider the case when $ p \neq 2 $. If $k<l$ and $k$ is an odd integer, then the congruence equation $x^2= p^k (\text{mod} \ p^l) $ reduces to the equation of $x^2 = p (\text{mod} \ p^i) $ for some power $i$ of $p$ and there is no solution to it; while when $k$ is an even integer, the congruence equation $x^2 = p^k ( \text{mod} \ p^l)$ reduces to the equation of or $ x^2 = 1 (\text{mod} \ p^i)$ and there are two solutions to it. In both case, the number of solutions are both equal to the value of $2 a(p^k, p^l)$ by its definition.
If on the other hand $k \geq l$, then the set of solutions to the congruence equation $x^2 = p^k (\text{mod} \ p^l)$ is the set of multipliers of $p^{\left\lceil \frac{l}{2} \right\rceil}$, so the number of distinct solutions mod $p^l$ is $p^{\left\lfloor \frac{l}{2} \right\rfloor}$.
Therefore, for odd prime $p$,
$$ A(p^k,p^l ) =\chi_{p^k} (\hat{p}^l) a(p^k,p^l) + \chi_{p^k} (\hat{p}^{l-1}) a(p^k, p^{l-1}) =a(p^k,p^l) + a(p^k, p^{l-1}), $$
where we set the term $ a(p^k, p^{l-1})$ equal to $0$ when $l=0$.
Next, using Hensel's lemma, an integer $d$ relatively prime to an odd prime $p$ is a quadratic residue modulo any power of $p$ if and only if it is a quadratic residue modulo $p$. In fact, if an integer $d$ is prime to the odd prime $p$, as
$$A(d, p^l) =2 \iff \chi_d(p) =1 \iff A(d, p) =2, $$
so
$$A(d, p^l) = \chi_d(p^l) + \chi_d(p^{l-1}). $$
By the prime power modulus theory \cite{gauss}, if the modulus is $p^l$,
then $p^kd$
$$\begin{cases}
\text{is a quadratic residue modulo} \ p^l\ \text{if} \ k \geq l, \\
\text{is a non-quadratic residue modulo}\ p^l\ \text{if}\ k < l \ \text{is odd}, \\
\text{is a quadratic residue modulo} \ p^l \ \text{if}\ k < l \ \text{ is even and}\ d \ \text{is a quadratic residue},\\
\text{is a non-quadratic residue modulo} \ p^l\ \text{if}\ k < l \ \text{ is even and}\ d\ \text{is a non-quadratic residue}.
\end{cases}$$
Therefore, for an odd integer $d$ prime to $p\neq 2$, we have
$$A(dp^k, p^l) = \begin{cases} 0 & \chi_{d}(p^l) = \mbox{-1 and} \ k<l \ \text{even} ,\\ A(p^k,p^l) &\mbox{otherwise}. \end{cases}$$
In the former case, we have:
\begin{equation*}\tag{1}
A(dp^k, p^l) =0= \chi_{d p^k}(\hat{p}^l) a(dp^k, p^l) + \chi_{d p^k}(\hat{p}^{l-1}) a(dp^k, p^{l-1}).
\end{equation*}
In the latter case, we also have:
\begin{equation*} \tag{2}
A(dp^k, p^l) = A(p^k, p^l)= \chi_{d p^k}(\hat{p}^l) a(dp^k, p^l) + \chi_{d p^k}(\hat{p}^{l-1}) a(dp^k, p^{l-1}).
\end{equation*}
Now let $D$ be an arbitrary odd integer. Given an prime integer $p$, write $D= D_0 p^k$, where $D_0$ is prime to $p$. Then from the equality $(1)$ and $(2)$ with $d$ replaced by $D_0$,
it follows that
$$ \sum_{l=0}^{\infty} A(D, p^l) p^{-ls} =(1-p^{-s})^{-1} \sum_{l=0}^{\infty} \chi_{D}(\hat{p}^l) a(D, p^l) p^{-ls}.$$
For $p=2$, we define $\tilde{P}_2(D, s)$ by equating
\begin{align} \sum_{l =0}^{\infty} \frac{A(D, 2^{l+2})}{2^{ls} } &=\tilde{P}_2(D, s) (1-2^{-s})^{-1} \sum_{l=0}^{\infty}\frac{ \chi_D(2^l) }{2^{ls}} \notag \\
&=\tilde{P}_2(D, s) (1-2^{-s})^{-1} \sum_{l=0}^{\infty}\frac{ \chi_D(2^l) a(D, 2^l)}{ 2^{ls}}. \notag
\end{align}
By the multiplicative property of $A(D, \cdot)$ and that of $\chi_D(\cdot) a(D, \cdot)$,
$$ \sum_{m>0} \frac{A(D, 4m)}{ m^{s}} = \tilde{P}_2(D, s) \zeta(s) \sum_{m>0} \frac{\chi_D(\hat{m}) a(D,m)}{ m^{s}} .$$
It remains to compute $\tilde{P}_2$. We need the next lemma.
\begin{lemma}
Let $D$ be an odd integer. With $\tilde{P}_2(D, s)$ defined in the above proposition, we have
\begin{align*}
\tilde{P}_2(D, s) = \begin{cases} 2 (1 - 2^{-s}) ( 1+ 2^{-s}) & D \equiv 1 \ (\mathrm{mod} \ 4) ,\\
0 & \mathrm{otherwise}.
\end{cases}
\end{align*}
\end{lemma}
\begin{proof}
Write
$$ \sum_{l =0} \frac{A(D, 2^{l+2})}{2^{ls} } = A(D, 4) + \sum_{l =1} \frac{A(D, 2^{l+2})}{2^{ls} } ,$$
and note that
$$A(D, 4) = \begin{cases} 2 & D \equiv1 \ \text{or} \ 5 \ (\text{mod} \ 8), \\ 0 & \mathrm{otherwise}. \end{cases}$$
To simplifying the second term, note that if $D$ is an odd integer and $m = 8,16$, or some higher power of $2$, then $D$ is a quadratic residue modulo $m$ if and only if $D \equiv1 (\text{mod} \ 8)$, therefore for $l \geq 1$,
$$ A(D, 2^{l+2}) = \begin{cases} 4& D \equiv1 \ (\text{mod} \ 8) ,\\ 0 & \text{otherwise} .\end{cases}$$
Also note that
$$ \chi_D(2) = \begin{cases} 1& D \equiv1 \ (\text{mod} \ 8) , \\ -1 & D \equiv 5 \ (\mathrm{mod} \ 8),\\ 0 & \text{otherwise} .\end{cases}$$
Then direct computation gives the results.
\end{proof}
This finishes the proof of the proposition.
\end{proof}
\begin{theorem}
The Shintani zeta function $Z_{\mathrm{Shintani}}^{\mathrm{odd}}(s,w)$ and the quadratic $A_2$ Weyl group multiple Dirichlet series $Z_{A_2}(s, w)$ satisfy the relation:
\begin{align*}
Z_{\rm{Shintani}}^{\mathrm{odd}} (s, w) =& 2 (1-2^{-s})(1+ 2^{-s}) \zeta(s) Z_{A_2}(s, w).
\end{align*}
\end{theorem}
\section{Bhargava integer cubes and zeta functions}
\subsection{Prehomogeneous vector space of a parabolic subgroup.}
Let $V_{\mathbb{Z}}$ be the $\mathbb{Z}$-module of $2\times 2 \times 2$ integer matrices, which we also call Bhargava integer cubes. There are three ways to form pairs of matrices by taking the opposite sides out of $6$ sides. Denote them by
\begin{align}
A^{F} &=M_1=\left(\begin{array}{ccc}a &b \\
c &d
\end{array}\right); A^{B}=N_1= \left( \begin{array}{ccc}e &f \\
g &h
\end{array}\right),
\notag \\
A^{L} &=M_2=\left(\begin{array}{ccc}a &c \\
e &g
\end{array}\right); A^{R}=N_2 =\left( \begin{array}{ccc}b &d \\
f &h
\end{array}\right),
\notag \\
A^{U} &=M_3=\left(\begin{array}{ccc}a &e \\
b &f
\end{array}\right); A^{D}= N_3=\left( \begin{array}{ccc}c &g \\
d &h
\end{array}\right).
\notag
\end{align}
For each pair $(M_i, N_i)$ we can associate to it a binary quadratic form by taking
$$Q_i(u,v) = \det(M_i u - N_i v) .$$
Explicitly for $A$ as above,
\begin{align} Q_1(u,v) &= u^2(ad-bc) +uv ( -ah+bg+cf-de) +v^2(eh-fg), \notag \\
Q_2 (u,v)&= u^2(ag-ce)+uv(-ah-bg+cf+de) + v^2(bh-df), \notag \\
Q_3(u,v)&= u^2(af-be) +uv(-ah+bg-cf+de) +v^2(ch-dg).\notag \end{align}
Following Bhargava \cite{bhargava}, we call $A$ projective if the associated binary quadratic forms are all primitive. The action of $G_{\mathbb{Z}}= \mathrm{GL}_2( \mathbb{Z}) \times \mathrm{GL}_2( \mathbb{Z}) \times \mathrm{GL}_2(\mathbb{Z})$ on $V_{\mathbb{Z}}$ is defined by letting the $g_i$ in $\left( g_1, g_2, g_3 \right) \in G_{\mathbb{Z}}$ act on the matrix pair $(M_i, N_i)$. It is easy to check that the actions of the three components commute with each other, thereby giving an action of the product group. For example, if $g_1 = \left( \begin{matrix} g_{11} & g_{12} \\
g_{21} & g_{22} \end{matrix} \right) $, then it acts on the pair $(M_1, N_1)$ by
$$\left( \begin{matrix} g_{11} & g_{12} \\
g_{21} & g_{22} \end{matrix} \right) \cdot \left( \begin{matrix} M_1 \\ N_1 \end{matrix} \right).$$
The action extends to the complex group $G_{\mathbb{C}} = \mathrm{GL}_2(\mathbb{C}) \times \mathrm{GL}_2(\mathbb{C}) \times \mathrm{GL}_2(\mathbb{C})$ on the complex vector space $V_{\mathbb{C}}$. We denote by $(G_{\mathbb{C}}, V_{\mathbb{C}})$ a complex vector space acted on by a (connected) complex group. In our setting, the $\left(G_{\mathbb{C}}, V_{\mathbb{C}}\right)$ is a PVS which refers to the $D_4$ case discussed in \cite{wrightyukie}. Now we consider the Borel subgroup $B'_2( \mathbb{C}) \subset \mathrm{GL}_2(\mathbb{C})$ consisting of the lower-triangular matrices. The action of $B'_2(\mathbb{C}) \times B'_2(\mathbb{C}) \times \mathrm{GL}_2(\mathbb{C})$ induced from $G_{\mathbb{C}}$ on the vector space $V_{\mathbb{C}}$ has three relative invariants, explicitly for $A \in V_{\mathbb{C}}$, given as follows
\begin{align}
D(A) & =\mathrm{disc}(A) =(-ah+bg+cf-de)^2 - 4 (ad-bc)(eh-fg) , \notag \\
m(A) & =\det(A^{F}) =ad-bc ,\notag \\
n(A) &=\det(A^{L}) =ag-ce. \notag
\end{align}
Furthermore, we can show that
\begin{proposition}
The pair $(B'_2(\mathbb{C}) \times B'_2(\mathbb{C}) \times \mathrm{GL}_2(\mathbb{C}) , V_{\mathbb{C}})$ is a prehomogeneous vector space.
\end{proposition}
\begin{proof}
Let $\mathcal{H}$ be the hypersurface in $V_{\mathbb{C}}$ defined as the zero locus of the single equation
$$ \mathrm{disc}(A) \det(A^F) \det(A^L) =0.$$
Any $A \in V_{\mathbb{C}} \backslash \mathcal{H}$ is $B'_2(\mathbb{C}) \times B'_2(\mathbb{C}) \times \mathrm{GL}_2(\mathbb{C})$ equivalent to some element with the form
$$
(A^F, A^B)= \left(\left( \begin{array}{cc} a & 0 \notag \\ 0 & d \notag \end{array} \right) ,
\left( \begin{array}{cc} 0 & f \notag \\ g & h \notag \end{array} \right) \right),
$$
where $ a ,d, g ,f\neq 0 $.
Furthermore, we can show that they are all in the single orbit of
$$
(A^F, A^B) = \left( \left( \begin{array}{cc} 1 & 0 \notag \\ 0 & 1 \notag \end{array} \right) ,
\left( \begin{array}{cc} 0 & 1 \notag \\ 1 & 0 \notag \end{array} \right) \right) .
$$
This follows from finding solutions to the following system of equations, and then taking the proper scaling:
\begin{align}
\lambda_{12} f+ \lambda_{23} a+ \lambda_{32} d &= 0 ,\notag \\
\lambda_{13} a+ \lambda_{22} f+ \lambda_{32} g &= 0 ,\notag \\
\lambda_{12} g+ \lambda_{22} d+ \lambda_{33} a &= 0 ,\notag \\
\lambda_{13} d+ \lambda_{23} g+ \lambda_{33} d &= -h- \lambda_{14} h- \lambda_{24} h- \lambda_{34} h, \notag
\end{align}
where $\left(\begin{array}{ccc}\lambda_{i1} &\lambda_{i2} \\
\lambda_{i3} & \lambda_{i4}
\end{array}\right)$($ \lambda_{i2} =0 , \ \text{for} \ i =1, 2$) is in the $i$-th place of $B'_2(\mathbb{C}) \times B'_2(\mathbb{C}) \times \mathrm{GL}_2(\mathbb{C})$. So $\left(B'_2(\mathbb{C}) \times B'_2(\mathbb{C}) \times \mathrm{GL}_2(\mathbb{C}), V_{\mathbb{C}}\right)$ is a prehomogeneous vector space.
\end{proof}
Let $B'_2(\mathbb{Z})$ be the Borel subgroup of $B'_2(\mathbb{C})$ in $\mathrm{SL}_2(\mathbb{Z})$ with positive diagonal elements. The Shintani zeta function associated to the prehomogeneous vector space $(B'_2(\mathbb{C}) \times B'_2(\mathbb{C}) \times \mathrm{GL}_2(\mathbb{C}) , V_{\mathbb{C}})$ is defined to be
\begin{equation*} \tag{3}
Z_{\rm{Shintani}}(s_1, s_2, w) = \sum_{A \in \left( B'_2(\mathbb{Z}) \times B'_2(\mathbb{Z}) \times \mathrm{SL}_2(\mathbb{Z}) \right) \backslash V^{ss}_{\mathbb{Z}} } \frac{1}{|\mathrm{disc}(A)|^w |\det(A^F)|^{s_1} |\det(A^L)|^{s_2}} ,\notag
\end{equation*}
where $V^{ss}_{\mathbb{Z}}$ is the subset of semi-stable points of $V_{\mathbb{Z}}$ consisting of those orbits on which none of the three relative invariants vanishes. Denote by $\mathrm{Stab}(A)$ the stabilizer group in $ B'_2(\mathbb{Z}) \times B'_2(\mathbb{Z}) \times \mathrm{SL}_2(\mathbb{Z})$ of the cube $A$. We need to show that the order $|\mathrm{Stab}(A)|$ is finite.
\begin{proposition}
For any $A \in V^{ss}_{\mathbb{Z}}$, the stabilizer group $\mathrm{Stab}(A)$ in $B'_2(\mathbb{Z}) \times B'_2(\mathbb{Z}) \times \mathrm{SL}_2(\mathbb{Z})$ is:
$$\{ (I_2, I_2, I_2)\},$$
where $I_2$ is the $2\times 2$ identity matrix.
Therefore
$$|\mathrm{Stab}(A)| =1$$
for any $A \in V^{ss}_{\mathbb{Z}}$.
\end{proposition}
\begin{proof}
For a given $2\times 2\times 2$ integer cube $A \in V^{ss}_{\mathbb{Z}}$, we write
\begin{align}
Q_1(A) &= mu^2+ x uv + s v^2 , \notag \\
Q_2(A) &= n u^2 + y uv + t v^2 \notag
\end{align}
to be first two binary quadratic forms associated to it. Under the action of $ B'_2(\mathbb{Z}) \times B'_2(\mathbb{Z}) \times \mathrm{SL}_2(\mathbb{Z})$ we can change $A$ to another cube satisfying
$$ c(A) =0 \ \mathrm{and} \ 0 \leq x \leq 2 |m|-1 \ \mathrm{and} \ 0 \leq y \leq 2 |n|-1.$$
For such $A$, note that the entries $a(A), d(A), g(A) \neq 0$. Therefore the stabilizer group in $ B'_2(\mathbb{Z}) \times B'_2(\mathbb{Z}) \times \mathrm{SL}_2(\mathbb{Z})$ must have the form
$$
\left( \left( \begin{matrix} 1 & 0 \\
0 & 1 \end{matrix} \right),
\left( \begin{matrix} 1 & 0 \\
0 & 1 \end{matrix} \right),
\left( \begin{matrix} \pm 1 & 0 \\
0 & \pm 1 \end{matrix} \right) \right).
$$
Now it becomes obvious that the diagonal elements in the last matrix have to be both positive.
\end{proof}
We can rewrite the Shintani zeta function as
\begin{equation*} \tag{4}
Z_{\rm{Shintani}}(s_1, s_2, w) = \sum_{D \neq 0} \frac{1}{|D|^w} \sum_{m,n >0 } \frac{B(D, m,n)}{ m^{s_1} n^{s_2}} ,\notag
\end{equation*}
where
\begin{align}B(D, m, n) & = \sharp \{ A \in V^{ss}_{\mathbb{Z}} / \sim : \mathrm{disc}(A) =D, |\det(A^F)| =m,| \det(A^L)| =n \} \notag \\
& = 4 \cdot \sharp \{ A \in V^{ss}_{\mathbb{Z}} / \sim : \mathrm{disc}(A) =D, \det(A^F) =m, \det(A^L) =n \}. \notag
\end{align}
\subsection{Reduction theory}
In this section, we will consider the geometry of integer orbits of the PHVS under the parabolic group action. This will help to reduce the sum over semi-stable orbits of $2\times 2\times 2$ integer cubes in the Shintani zeta function $Z_{\mathrm{Shintani}}(s_1, s_2, w)$ to a sum over the tuples of integers $(D, m, n)$ satisfying certain relations. First we need a lemma which establishes the existence of a $2\times 2\times 2 $ integer cube with the required arithmetic invariants.
\begin{lemma}
Let $D$, $m$ and $n$ be non-zero integers. For each solution $(x, y)$ to the congruence equations
\begin{align}
x^2 &\equiv D \ ( \mathrm{mod} \ 4m) \ \text{for} \ 0\leq x \leq 2 |m| -1, \notag \\
y^2 &\equiv D \ ( \mathrm{mod} \ 4n\ ) \ \text{for} \ 0\leq y \leq 2 |n| -1, \notag
\end{align}
there exists a $2\times 2\times2$ integer cube $A$ such that
\begin{align}
\mathrm{disc}(A) &= D , \notag \\
Q_1(A)(u,v) & = mu^2+ x uv+ sv^2, \notag \\
Q_2(A)(u,v) &=n u^2+ y uv + t v^2. \notag
\end{align}
Moreover, the required $2 \times 2 \times 2$ integer cube $A$ can be chosen such that in the bottom side $A^D$ of $A$
\begin{align*}
c =0 \ \mathrm{and} \ g.c.d.( d, g, h) =1.
\end{align*}
\end{lemma}
\begin{proof}
If there are solutions to the congruence equations, we have
$$D =x^2 - 4 m s = y^2 - 4 n t$$
for some integers $s$ and $t$. It implies that $D$ is congruent to $0$ or $1 \ (\mathrm{mod} \ 4)$, and the integers $x, y$ have the same parity. Take
$$ a = |g.c.d.(m, n, \frac{x+y}{2})|,$$
from
\begin{equation*}\tag{5} \frac{x-y}{2} \cdot \frac{x+y}{2} = ms - nt, \end{equation*}
it follows that
\begin{equation*}\tag{6} g.c.d.(\frac{m}{a}, \frac{n}{a})\ | \ \frac{x-y}{2}.\end{equation*}
Set
$$d = \frac{m}{a}, \ g = \frac{n}{a}, \ \text{and} \ h=-\frac{x+y}{2 a}, $$
then
$$g.c.d.(d,g,h) =1,$$
and $(5)$ can be written as
\begin{equation*} \tag{7}\frac{x-y}{2} \cdot (-h) = ds- gt.\end{equation*}
We claim that there is an integer $f$, such that
\begin{align}
s+ f g & \equiv 0 \ (\text{mod}\ h), \notag \\
t + f d & \equiv 0 \ (\text{mod}\ h). \notag
\end{align}
This can be proved as follows: First if $h=0$, then $ds = gt$. As in this case $g.c.d.(d,g) =1$, we conclude that there exists such an integer $f$ such that $ s= -f g$ and $t = -f d$. If $h \neq 0$, for any prime divisor $p$ of $h$, we have
$$p\ | \ ds- gt \ \text{and} \ g.c.d.(d,g,p)=1,$$
it follows that there is a unique solution ($\text{mod} \ p$) to the congruences
\begin{align}
s+ f g & \equiv 0 \ (\text{mod}\ p), \notag \\
t + f d & \equiv 0 \ (\text{mod}\ p). \notag
\end{align}
Using the Chinese remainder theorem, we conclude that there are integers $b$ and $e$ such that
\begin{align}
s = eh - fg, \notag \\
t = bh - fd. \notag
\end{align}
It follows that
$$ ds- gt = (de- bg) \cdot h,$$
combined with $(7)$, we have
$$\frac{x-y}{2}= bg- de.$$
If in the case of $h=0$, from $(6)$, we know that there always exist integers $b$ and $e$ such that the above equation holds.
Now we define a $2 \times 2\times 2$ integer cube $A$ by
$$
(A^F, A^B)= \left(\left( \begin{array}{cc} a & b \notag \\ 0 & d \notag \end{array} \right) ,
\left( \begin{array}{cc} e & f \notag \\ g & h \notag \end{array} \right) \right).
$$
Then the two associated binary quadratic forms are
\begin{align}
Q_1(A)(u, v)&= ad u^2 +(-ah + bg-de)uv + (eh-fg) v^2 = m u^2+ x uv + s v^2, \notag \\
Q_2(A)(u,v) &= ag u^2 + ( -ah - bg+ de) uv+ (bh- df\ ) v^2 = nu^2+ y uv+ t v^2. \notag
\end{align}
So $A$ is the cube required. In particular, we have shown that $g.c.d.(d,g, h) =1$.
\end{proof}
We next want to show that under the assumption that $D$ is square-free, the data $(D, m, n, x, y)$ uniquely determine a $B'_2(\mathbb{Z}) \times B'_2(\mathbb{Z}) \times \mathrm{SL}_2(\mathbb{Z})$-orbit.
\begin{proposition}
Let $D$ be a non-zero square-free integer and $m$, $n$ be non-zero integers. Each solution to the congruence equations
\begin {align}
x^2 &\equiv D \ ( \mathrm{mod} \ 4m) \ \text{for} \ 0\leq x \leq 2 |m| -1, \notag \\
y^2 &\equiv D \ ( \mathrm{mod} \ 4n\ ) \ \text{for} \ 0\leq y \leq 2 |n| -1, \notag
\end{align}
determines a unique orbit of integer cubes in $\left(B'_{2}(\mathbb{Z}) \times B'_{2}(\mathbb{Z}) \times \mathrm{SL}_{2}(\mathbb{Z}) \right) \backslash V^{ss}_{\mathbb{Z}}$ such that
\begin{align}
\mathrm{disc}(A) &= D , \notag \\
Q_1(A)(u,v) &= mu^2+ x uv+ s v^2, \notag \\
Q_2(A)(u,v) &=nu^2 + yuv + t y^2. \notag
\end{align}
Moreover, two different solutions to the congruence equations correspond to two different orbits of integer cubes in $\left( B'_{2}(\mathbb{Z}) \times B'_{2}(\mathbb{Z}) \times \mathrm{SL}_{2}(\mathbb{Z}) \right) \backslash V^{ss}_{\mathbb{Z}}$.
\end{proposition}
\begin{proof}
From the above lemma, we know that there always exists such a $2 \times 2 \times 2$ integer cube $A$. We want to show that such an $2 \times 2 \times 2$ integer cube $A$ is uniquely determined by the data $(D, m, n, x, y)$ satisfying the congruence equations.
Denote $A$ by the pair of matrices
$$A = \left( \left( \begin{matrix}
a & b \\
c & d
\end{matrix} \right ) , \left( \begin{matrix}
e & f \\
g & h
\end{matrix} \right ) \right). $$
We first show that the $(a, d, g, h)$ is uniquely determined up to the sign of $a$. As $\{1\} \times \{1\} \times \mathrm{SL}_2(\mathbb{Z}) \subset B'_2(\mathbb{Z}) \times B'_2(\mathbb{Z}) \times \mathrm{SL}_2(\mathbb{Z}) $, we can assume that $c(A) =0$. Then we have equations
\begin{align}
bg- ah - ed &= x, \notag \\
ed-ah-bg&= y ,\notag \\
ad &= m ,\notag \\
ag &= n. \notag
\end{align}
after adding the first two equations,
\begin{align}
ah &= -(x+y) /2, \notag\\
ad &= m, \notag \\
ag &= n. \notag
\end{align}
Therefore we have
$$ a \times g.c.d.(h, d,g) = g.c.d.((x+y)/2, m, n).$$
As
$$D = (bg-ah -ed)^2- 4ad (eh-fg), $$
$D$ is square-free implies that $g.c.d.(h, d,g) =1$, therefore $a = |g.c.d.((x+y)/2, m, n)|$ as we can make $a$ to be positive.
We next show that for two $2\times 2\times 2$ integer cubes with $c =0$ and fixed discriminant $D$ such that they have the same tuples $(a, d, g ,h)$ and $(Q_1(u, v),Q_2(u, v))$, then they are equal up to the action of $\{1\} \times \{1\} \times B_2(\mathbb{Z})$, the third one is an upper-triangular matrix in $\mathrm{SL}_2(\mathbb{Z})$. Here in this statement we again require $D$ to be square-free.
We already showed that for a given $A$,
\begin{align}
Q_1(A)(u,v) &= m u^2 + x uv + sv^2 \notag = ad u^2 +(bg-ah-ed)uv+(eh-fg)v^2 ,\notag \\
Q_2(A)(u, v) &= n u^2 + y uv + tv^2 \notag = ag u^2 +(ed-ah-bg) uv+(bh-df) v^2 .\notag
\end{align}
As $(Q_1(A_i)(u, v),Q_2(A_i)(u, v))$ are the same for two integer cubes by assumption, we can make the following equations by setting the coefficients are equal
\begin{align}
x(A_1) = x(A_2): & b_1g- ah - e_1d= b_2 g - ah - e_2 d , \notag \\
y(A_1) = y(A_2): & e_1d -ah - b_1g = e_2 d- ah - b_2 g, \notag \\
s(A_1) =s(A_2) : & e_1 h-f_1 g = e_2 h- f_2g , \notag \\
t(A_1) = t(A_2): &b_1 h- df_1 = b_2 h - df_2 .\notag
\end{align}
Taking the subtraction of right side from the left side on each equation, we have:
\begin{align}
\Delta (b)g - \Delta( e) d =0, \notag \\
\Delta (e) h - \Delta (f )g = 0, \notag \\
\Delta (b) h- \Delta (f) d = 0, \notag
\end{align}
here $\Delta(b)= b_1-b_2$. Then we have solutions to the above equation system:
$$ \Delta(e); \Delta(b) = \Delta(e) d/g; \Delta(f) = \Delta(e)h /g.$$
Under the assumption of $D$ being square-free, we have shown that $g.c.d.(d, g, h) =1$. Therefore $g|\Delta(e)$, i.e., the solution has form
$$ \Delta(e)=gk; \Delta(b) = d k; \Delta(f) = h k.$$
This implies that we can make $A_1$ equivalent to $A_2$ under the action of $\{1\} \times \{1\} \times B_2(\mathbb{Z})$.
Now we prove the second part of the proposition by contradiction. Given a square-free integer $D$ and two non-zero integers $m$ and $n$, if $(x_1, y_1)$ and $(x_2, y_2)$ are two different solutions to the congruence equations in the proposition. Let $A_1$ and $A_2$ be the two corresponding $2 \times 2\times 2$ integer cubes satisfying $\mathrm{disc}(A_i)= D$, and
\begin{align}
Q_1(A_i)(u,v) &= m u^2 + x_i uv + s_i v^2,\notag \\
Q_2(A_i)(u,v) &= n u^2 + y_i uv + t_i v^2 .\notag
\end{align}
Suppose that $A_1$ and $A_2$ are $B'_2(\mathbb{Z}) \times B'_2(\mathbb{Z}) \times \mathrm{SL}_2(\mathbb{Z}) $ equivalent. We assume $c(A_1) = c(A_2) =0 $ as before. Since $0 \leq x_i \leq 2 m -1$ and $0 \leq y_i \leq 2 n -1$, the group action actually belongs to $\{1\} \times \{1\} \times B_2(\mathbb{Z})$, which implies that $(Q_1(A_i)(u,v), Q_2(A_i)(u,v))$ are the same. In particular $x_1 = x_2$ and $y_1 = y_2$, contradiction. So $A_1$ and $A_2$ are not $B'_2(\mathbb{Z}) \times B'_2(\mathbb{Z}) \times \mathrm{SL}_2(\mathbb{Z}) $ equivalent.
\end{proof}
Now we turn to the general case without assuming square-free discriminant. First we consider the non-zero integer $D = D_0 p^2$, where $D_0$ is square-free.
\begin{proposition}
Let $m$ and $n$ be non-zero integers, and $D= D_0 p^2$ where $D_0$ is square-free and $p$ is a prime integer. The coefficient $B(D, m,n)$ from $(4)$ is given by:
$$B(D, m, n) =A\left(D, 4m\right)A\left(D, 4n\right)+b(D_0, \frac{m}{p}, \frac{n}{p}), $$
where the second term
\begin{align*}
b(D_0, \frac{m}{p}, \frac{n}{p})= \begin{cases} p A(D_0, \frac{4m}{p})A(D_0, \frac{4n}{p}) & \mathrm{if} \ $p$ \ \mathrm{ divides} \ $g.c.d.(D,m,n)$, \\
0 & \mathrm{otherwise}.
\end{cases}
\end{align*}
\end{proposition}
\begin{proof}
By the existence Lemma 3.3, we can find an $2 \times 2 \times 2$ integer cube $A$ satisfying
\begin{align*}
c =0 \ \mathrm{and} \ g.c.d.(d,g,h) =1,
\end{align*}
such that
\begin{align*}
\mathrm{disc}(A) = D, \ \det(A^F) = m \ \mathrm{and} \ \det(A^L) =n.
\end{align*}
From the proof of the last proposition, the condition of $g.c.d.(d,g,h) =1$ implies that the integer cube $A$ determines a unique orbit in $\left( B'_2(\mathbb{Z}) \times B'_2(\mathbb{Z}) \times \mathrm{SL}_2(\mathbb{Z}) \right) \backslash V^{ss}_{\mathbb{Z}}(D)$. We denote by $V_{\mathbb{Z}}^{ss}(D)$ the semi-stable subset of $V_{\mathbb{Z}}$ with discriminant $D$.
If $p| g.c.d.(D, m, n)$, we write $m_0= \frac{m}{p}$, $n_0= \frac{n}{p}$. Applying the existence Lemma 3.3 to the non-zero integers $D_0, m_0, n_0$, we can find an $2 \times 2 \times 2$ integer cube $A_0$ satisfying
\begin{align*}
c_0 =0 \ \mathrm{and} \ g.c.d.(d_0,g_0,h_0) =1,
\end{align*}
such that
\begin{align*}
\mathrm{disc}(A_0) = D_0, \ \det(A_0^F) = m_0 \ \mathrm{and} \ \det(A_0^L) =n_0.
\end{align*}
The condition $g.c.d.(d_0, g_0, h_0)=1$ implies that $A_0$ determines a unique orbit in $ \left( B'_2(\mathbb{Z}) \times B'_2(\mathbb{Z}) \times \mathrm{SL}_2(\mathbb{Z}) \right) \backslash V^{ss}_{\mathbb{Z}}(D_0)$.
For each $g$ in $\{(1) \times (1) \times \left( \begin{array}{cc }1 & j \\
0 & p\end{array}\right): 0 \leq j \leq p-1\} $, the $2 \times 2 \times 2$ integer cube $g \cdot A_0$ determines uniquely the orbit in $ \left( B'_2(\mathbb{Z}) \times B'_2(\mathbb{Z}) \times \mathrm{SL}_2(\mathbb{Z}) \right) \backslash V^{ss}_{\mathbb{Z}}(D)$.
\end{proof}
\begin{corollary}
If $D$ is a fundamental discriminant, then we have
$$B(D, m,n)= A(D, 4m) A(D, 4n).$$
\end{corollary}
\begin{proof}
If $D$ is square-free then this follows from the above proposition we proved. Otherwise $D = 4 D_0$ where $D_0$ is square-free, using the above proposition applied to the case $p=2$, we have
$$B(D, m,n)= A\left(D, 4m\right) A\left(D, 4n\right)+ b(D_0, \frac{m}{2}, \frac{n}{2}) .$$
The second term is $0$ as the discriminant of a quadratic form should be $D_0= f^2 d_{K}$ where $d_K$ is a fundamental discriminant, but $D$ is already a fundamental discriminant.
\end{proof}
\begin{proposition}
Let $m$ and $n$ be non-zero integers, and $D= D_0 p^{2k}$ where $D_0$ is square-free and $p$ is a prime integer. We have
\begin{align}
B(D,m ,n) =& A(D, 4m)A(D,4n)+b(\frac{D}{p^2}, \frac{m}{p}, \frac{n}{p}) + \cdots \notag\\
& + b(\frac{D}{p^{2i}}, \frac{m}{p^i}, \frac{n}{p^i}) + \cdots + b(D_0, \frac{m}{p^k}, \frac{n}{p^k}) ,\notag
\end{align}
for some $1 \leq i \leq k$,
\begin{align*}
b(\frac{D}{p^{2i}}, \frac{m}{p^i}, \frac{n}{p^i})= \begin{cases} p^i A(\frac{D}{p^{2i}}, \frac{4m}{p^i})
A(\frac{D}{p^{2i}}, \frac{4n}{p^i}) & \mathrm{if} \ p^i \ \mathrm{ divides} \ g.c.d.(D, m,n) , \\
0 & \mathrm{otherwise}.
\end{cases}
\end{align*}
\end{proposition}
\begin{proof}
By the existence Lemma 3.3, we can find an $2 \times 2 \times 2$ integer cube $A$ satisfying
\begin{align*}
c =0 \ \mathrm{and} \ g.c.d.(d,g,h) =1,
\end{align*}
such that
\begin{align*}
\mathrm{disc}(A) = D, \ \det(A^F) = m \ \mathrm{and} \ \det(A^L) =n.
\end{align*}
The condition of $g.c.d.(d,g,h) =1$ implies that the integer cube $A$ determines a unique orbit in $\left( B'_2(\mathbb{Z}) \times B'_2(\mathbb{Z}) \times \mathrm{SL}_2(\mathbb{Z}) \right) \backslash V^{ss}_{\mathbb{Z}}(D)$.
If $p^i| g.c.d.(D, m,n)$, we write $m_i= \frac{m}{p^i}$, $n_i= \frac{n}{p^i}$. Applying the existence Lemma 3.3 to the non-zero integers $\frac{D}{p^{2i}}$, $\frac{m}{p^i}$, and $\frac{n}{p^i}$, we can find an $2 \times 2 \times 2$ integer cube $A_i$ satisfying
\begin{align*}
c_i =0 \ \mathrm{and} \ g.c.d.(d_i,g_i,h_i) =1,
\end{align*}
such that
\begin{align*}
\mathrm{disc}(A_i) = \frac{D}{p^{2i}}, \ \det(A_i^F) = \frac{m}{p^i} \ \mathrm{and} \ \det(A_i^L) =\frac{n}{p^i}.
\end{align*}
The condition $g.c.d.(d_i, g_i, h_i)=1$ implies that $A_i$ determines a unique orbit in $ \left( B'_2(\mathbb{Z}) \times B'_2(\mathbb{Z}) \times \mathrm{SL}_2(\mathbb{Z}) \right) \backslash V^{ss}_{\mathbb{Z}}(\frac{D}{p^{2i}})$.
For each $g$ in $\{(1) \times (1) \times \left( \begin{array}{cc }1 & j \\
0 & p^i\end{array}\right): 0 \leq j \leq p^i-1\} $, the $2 \times 2 \times 2$ integer cube $g \cdot A_i$ determines uniquely the orbit in $ \left( B'_2(\mathbb{Z}) \times B'_2(\mathbb{Z}) \times \mathrm{SL}_2(\mathbb{Z}) \right) \backslash V^{ss}_{\mathbb{Z}}(D)$.
\end{proof}
In summary we have the full general formula for the orbits counting number $B(D, m, n)$.
\begin{proposition}
Let $m$ and $n$ be non-zero integers, and $D= D_0 D_1^2$ where $D_0$ is square-free. We have
\begin{equation*}\tag{8}B(D,m,n) = \sum_{d|D_1} b(\frac{D}{d^2}, \frac{m}{d}, \frac{n}{d}),\end{equation*}
where
\begin{align*}
b(\frac{D}{d^2}, \frac{m}{d}, \frac{n}{d})= \begin{cases} d \cdot A(\frac{D}{d^2}, \frac{4m}{d}) \cdot
A(\frac{D}{d^2}, \frac{4n}{d}) & \mathrm{if} \ d \ \mathrm{ divides} \ g.c.d.(D_1, m,n) , \\
0 & \mathrm{otherwise}.
\end{cases}
\end{align*}
\end{proposition}
\begin{proof}
By the existence Lemma 3.3, we can find an $2 \times 2 \times 2$ integer cube $A$ satisfying
\begin{align*}
c =0 \ \mathrm{and} \ g.c.d.(d,g,h) =1,
\end{align*}
such that
\begin{align*}
\mathrm{disc}(A) = D, \ \det(A^F) = m \ \mathrm{and} \ \det(A^L) =n.
\end{align*}
The condition of $g.c.d.(d,g,h) =1$ implies that the integer cube $A$ determines a unique orbit in $\left( B'_2(\mathbb{Z}) \times B'_2(\mathbb{Z}) \times \mathrm{SL}_2(\mathbb{Z}) \right) \backslash V^{ss}_{\mathbb{Z}}(D)$.
If $d| g.c.d.(D_1, m,n)$, we write $m'= \frac{m}{d}$, $n'= \frac{n}{d}$. Applying the existence Lemma 3.3 to the non-zero integers $\frac{D}{d^2}$, $\frac{m}{d}$, and $\frac{n}{d}$, we can find an $2 \times 2 \times 2$ integer cube $A'$ satisfying
\begin{align*}
c' =0 \ \mathrm{and} \ g.c.d.(d',g',h') =1,
\end{align*}
such that
\begin{align*}
\mathrm{disc}(A') = \frac{D}{d^2}, \ \det\left((A')^F\right) = \frac{m}{d} \ \mathrm{and} \ \det\left((A')^L\right) =\frac{n}{d}.
\end{align*}
The condition $g.c.d.(d', g', h')=1$ implies that $A'$ determines a unique orbit in $ \left( B'_2(\mathbb{Z}) \times B'_2(\mathbb{Z}) \times \mathrm{SL}_2(\mathbb{Z}) \right) \backslash V^{ss}_{\mathbb{Z}}(\frac{D}{d^2})$.
For each $g$ in $\{(1) \times (1) \times \left( \begin{array}{cc }1 & j \\
0 & d\end{array}\right): 0 \leq j \leq d-1\} $, the $2 \times 2 \times 2$ integer cube $g \cdot A'$ determines uniquely the orbit in $\left( B'_2(\mathbb{Z}) \times B'_2(\mathbb{Z}) \times \mathrm{SL}_2(\mathbb{Z}) \right) \backslash V^{ss}_{\mathbb{Z}}(D)$.
\end{proof}
Finally we obtain the explicit formula for the Shintani zeta function associated to the PVS of $2\times 2 \times 2$ cubes.
\begin{theorem}
The Shintani zeta function $Z_{\rm{Shintani}}(s_1, s_2, w)$ can be expressed as
\begin{align}
& \sum_{A \in \left( B'_2(\mathbb{Z}) \times B'_2(\mathbb{Z}) \times \mathrm{SL}_2(\mathbb{Z}) \right) \backslash V^{ss}_{\mathbb{Z}} } \frac{1}{|\mathrm{disc}(A)|^w |\det(A^F)|^{s_1} |\det(A^L)|^{s_2}} \notag \\
& = \sum_{D=D_0D_1^2} \frac{1}{|D|^w} \sum_{m,n > 0} \frac{\sum_{\substack{d|D_1\\ d|m, d|n}}d\cdot A(\frac{D}{d^2}, \frac{4m}{d}) \cdot A(\frac{D}{d^2}, \frac{4n}{d})}{m^{s_1}n^{s_2}}. \notag
\end{align}
\end{theorem}
In particular, if $D$ is an odd integer, then $(8)$ becomes
\begin{equation*}\tag{9}B(D,m,n) =\sum_{\substack{d|D_1}}d\cdot A(\frac{D}{d^2}, \frac{4m}{d}) \cdot A(\frac{D}{d^2}, \frac{4n}{d}),\end{equation*}
as when $d$ is an odd integer it is a divisor of $m$ if and only if it is a divisor of $4m$. So we have the following
\begin{corollary}
The partial Shintani zeta function $Z^{\mathrm{odd}}_{\rm{Shintani}}(s_1, s_2, w)$ defined by
\begin{align*}
Z^{\mathrm{odd}}_{\rm{Shintani}}(s_1, s_2, w)=\sum_{\substack{ A \in \left( B'_2(\mathbb{Z}) \times B'_2(\mathbb{Z}) \times \mathrm{SL}_2(\mathbb{Z}) \right) \backslash V^{ss}_{\mathbb{Z}}\\ \mathrm{dics}(A) \ \mathrm{odd} }} \frac{1}{|\mathrm{disc}(A)|^w |\det(A^F)|^{s_1} |\det(A^L)|^{s_2}}
\end{align*}
can be written as
\begin{align*}
\sum_{D=D_0D_1^2} \frac{1}{|D|^w} \sum_{m,n > 0} \frac{\sum_{d|D_1}d\cdot A(\frac{D}{d^2}, \frac{4m}{d}) \cdot A(\frac{D}{d^2}, \frac{4n}{d})}{m^{s_1}n^{s_2}}.
\end{align*}
\end{corollary}
\section{$A_3$ Weyl Group Multiple Dirichlet Series}
In this section, we will relate the Shintani zeta function $Z_{\mathrm{Shintani}}(s_1, s_2, w) $ to the quadratic $A_3$-Weyl group multiple Dirichlet series. The idea is first to construct a multiple Dirichlet series $Z_{\rm{WMDS}}(s_1, s_2, w) $ and then show its relation to the Shintani zeta function of PVS of $2 \times 2 \times 2$ cubes, using the results we did for the relation between $Z_{A_2}(s, w)$ and $Z_{\mathrm{Shintani}}(s, w; B'_2)$. Finally we show the multiple Dirichlet series $Z_{\mathrm{WMDS}}(s_1, s_2, w)$ is the desired $A_3$ Weyl group multiple Dirichlet series by computing the generating function of its $p$-parts.
In the $A_2$-WMDS
\begin{align*}
Z_{A_2} (s, w) = \sum_{\substack{m >0, \\ D \ \mathrm{odd}\ \mathrm{discriminant}}} \frac{\chi_D(\hat{m}) a(D, m)}{ m^s |D|^w},
\end{align*}
we let $\tilde{A}(D, m) = \chi_D(\hat{m}) a(D, m)$. Write $D= D_0 D_1^2$ where $D_0$ stands for the square-free part of $D$. Define
\begin{align*}
Z_{\mathrm{WMDS}}(s_1,s_2, w)
= \sum_{ D \ \mathrm{odd} \ \mathrm{discriminant}} \frac{1}{|D|^w} \sum_{m,n>0} \frac{\chi_D(\hat{m}) \chi_D(\hat{n})}{m^{s_1}n^{s_2}}a(D,m,n) ,
\end{align*}
where
\begin{align*}
a(D, m,n) = \sum_{\substack{d|D_1\\ d|m,\ d|n}} d\cdot a(\frac{D}{d^2}, \frac{m}{d}) \cdot a(\frac{D}{d^2}, \frac{n}{d}).
\end{align*}
\begin{lemma}
The multiple Dirichlet series $Z_{\mathrm{WMDS}}(s_1, s_2, w)$ can be expressed as
\begin{align*}
Z_{\mathrm{WMDS}}(s_1,s_2, w)
&= \sum_{ D\ \mathrm{odd} \ \mathrm{discriminant}} \frac{1}{|D|^w} \sum_{m,n>0} \frac{\chi_D(\hat{m}) \chi_D(\hat{n})}{m^{s_1}n^{s_2}}a(D,m,n) \\
& = \sum_{\substack{D= D_0 D_1^2 \\ D\ \mathrm{odd} \ \mathrm{discriminant} }} \frac{1}{|D|^w}\frac{\sum_{\substack{d|D_1\\ d|m, d|n}}d \cdot \tilde{A}(\frac{D}{d^2},\frac{m}{d}) \cdot \tilde{A}
(\frac{D}{d^2}, \frac{n}{d})}{m^{s_1}n^{s_2}} .
\end{align*}
\end{lemma}
\begin{proof}
This follows from the fact that the quadratic character $\chi_{D}(\cdot)$ is the same as $\chi_{D/d^2}(\cdot)$ by definition, and the factor $m$ is prime to $D$ if and only if the factor $\frac{m}{d}$ is prime to $\frac{D}{d^2}$. Therefore $\chi_D(\hat{m}) \chi_D(\hat{n})$ is a common factor for fixed integers $D$, $m$ and $n$.
\end{proof}
As we did in the Proposition 2.5, we will show the relation between the inner sum of $Z^{\mathrm{odd}}_{\mathrm{Shintani}}(s_1, s_2, w)$ and $Z_{\rm{WMDS}}(s_1, s_2, w)$.
\begin{proposition}
Let $D$ be an odd integer. The inner sum of the Shintani zeta function $Z_{\mathrm{Shintani}}^{\mathrm{odd}} (s_1, s_2, w)$ can be expressed by
$$\sum_{m, n > 0} \frac{B(D, m, n)}{m^{s_1} n^{s_2}}= \tilde{P}_2(D,s_1,s_2) \zeta(s_1) \zeta(s_2) \sum_{m,n>0} \frac{\chi_D(\hat{m}) \chi_D(\hat{n})}{m^{s_1}n^{s_2}}a(D,m,n) ,$$
where
$$\tilde{P}_2(D,s_1,s_2) = \tilde{P}_2(D, s_1) \tilde{P}_2(D, s_2)$$
and $\tilde{P}_2(D, s_i)$ is defined in Proposition 2.5.
\end{proposition}
\begin{proof}
Recall that in the proof of Proposition 2.5, we have shown that
\begin{align*}
\sum_{m>0} \frac{A(D, 4m)}{m^{s} } = \tilde{P}_2(D, s) \zeta(s) \sum_{m>0} \frac{\chi_D(\hat{m})a(D,m)}{ m^{s}} .
\end{align*}
Therefore,
\begin{align*}
\sum_{m, n> 0} \frac{A(D, 4m) A(D, 4n)}{m^{s_1} n^{s_2}} & = \tilde{P}_2(D, s_1) \tilde{P}_2(D, s_2) \zeta(s_1) \zeta(s_2) \\
& \cdot \sum_{m, n>0} \frac{\chi_D(\hat{m})\chi_D(\hat{n})}{ m^{s_1} n^{s_2}} a(D, m) a(D, n).
\end{align*}
In particular for any $d^2|D$, replace $D$ by $\frac{D}{d^2}$, $m$ by $\frac{m}{d}$ and $n$ by $\frac{n}{d}$, we have
\begin{align*}
\sum_{m, n> 0} \frac{A(\frac{D}{d^2}, \frac{4m}{d}) A(\frac{D}{d^2}, \frac{4n}{d})}{m^{s_1} n^{s_2}} & = \tilde{P}_2(\frac{D}{d^2}, s_1) \tilde{P}_2(\frac{D}{d^2}, s_2) \zeta(s_1) \zeta(s_2) \\
& \cdot \sum_{m, n>0} \frac{\chi_D(\hat{m})\chi_D(\hat{n})}{ m^{s_1} n^{s_2}} a(\frac{D}{d^2}, \frac{m}{d}) a(\frac{D}{d^2}, \frac{n}{d}).
\end{align*}
As $D$ is odd and $D \equiv D/d^2 \ (\rm{mod}\ 8)$, by the definition of $\tilde{P}_2(D, s)$, it follows that
$$\tilde{P}_2(D, s) = \tilde{P}_2(\frac{D}{d^2}, s).$$
Finally, taking the sum over all $d$ with $d^2|D$, we have
\begin{align*}
\sum_{m,n > 0} \sum_{d^2|D} \frac{ d \cdot A(\frac{D}{d^2}, \frac{4m}{d}) \cdot A(\frac{D}{d^2}, \frac{4n}{d})}{m^{s_1}n^{s_2}} &= \tilde{P}_2(D,s_1) \tilde{P}_2(D,s_2) \zeta(s_1) \zeta(s_2) \\
& \cdot \sum_{m,n>0} \frac{\chi_D(\hat{m}) \chi_D(\hat{n})}{m^{s_1}n^{s_2}}a(D,m,n) .
\end{align*}
\end{proof}
Recall the definition of $\tilde{P}_2(D, s)$,
\begin{align*}
\tilde{P}_2(D, s) = \begin{cases} 2 (1 - 2^{-s}) ( 1+ 2^{-s}) & D \equiv 1 \ (\mathrm{mod} \ 4) ,\\
0 & \mathrm{otherwise}.
\end{cases}
\end{align*}
Now we can give an explicit relation between $Z_{\mathrm{Shintani}}(s_1, s_2, w)$ and $Z_{\rm{WMDS}}(s_1, s_2, w)$. We further relate the Shintani zeta function $Z_{\mathrm{Shintani}}(s_1, s_2, w)$ to a Weyl group multiple Dirichlet series by showing that $Z_{\rm{WMDS}}(s_1, s_2, w)$ is a quadratic $A_3$-WMDS.
\begin{theorem}
The Shintani zeta function of $\mathrm{PVS}$ of $2\times 2 \times 2$ cubes can be related to the multiple Dirichlet series $Z_{\mathrm{WMDS}}(s_1, s_2, w)$ by
\begin{align*}
Z^{\mathrm{odd}}_{\rm{Shintani}}(s_1, s_2, w) = & 4 (1-2^{-s_1})(1+ 2^{-s_1})
(1-2^{-s_2})(1+ 2^{-s_2}) \zeta(s_1) \zeta(s_2) \\
& \cdot Z_{\rm{WMDS}} (s_1, s_2, w).
\end{align*}
\end{theorem}
\begin{theorem}
$Z_{\rm{WMDS}}(s_1,s_2,w)$ is a quadratic $A3$ Weyl group multiple Dirichlet series.
\end{theorem}
\begin{proof} Consider the $p$-parts of our $a(D,m,n)$ defined by
$$ a_{klt}(p)=a(p^k, p^l, p^t).$$
Explicit from its definition,
$$ a_{klt}(p) = a(p^k, p^l) a(p^k,p^t) + p a(p^{k-2},p^{l-1}) a(p^{k-2}, p^{t-1})+ \cdots .$$
In \cite{chintagunnells} and \cite{chintagunnellsjams}, the authors developed a systematical way to construct the Weyl group multiple Dirichlet series. The idea is to construct a rational function invariant under the Weyl group action. In the case of root system of $A_3$ type, the $p$-parts of the rational function is given by
$$f_{A_3}(x,y,z)= \frac{(1 - xy - yz + xyz + pxy^2z - px^2y^2z - pxy^2z^2 +
px^2 y^3 z^2)}{(1 - x)(1 - y)(1 - z)(1 - py^2z^2) (1 -
py^2x^2) (1 - p^2x^2y^2z^2)}.$$
Write the expansion
$$f_{A_3}(x,y,z)= \sum b_{klt}(p) x^l y^k z^t,$$
then we can compare our $\{a_{klt}\}$ with $\{b_{klt}\}$. They coincide with each other as follows: we know that for $|x|, |y|, |z| < 1/p$ there is
$$f_{A_3}(x,y,z) = \frac{1}{1-pxy^2z} \int f_{A_2}(x, t)f_{A_2}(yt^{-1},z)\frac{dt}{t},$$
where
$$f_{A_2}(x_1,x_2) = \sum_{k,l \geq 0} a(p^k,p^l) x_1^k x_2^l$$
and the integral is taken over the circle $|t| =1/p$ (\cite[Example 3.7]{chintagunnells}). Substituting the above expansion of $f_{A_2}$ into $f_{A_3}$, note that $a(p^k, p^l) = a(p^l, p^k)$, we have
\begin{align}
f_{A_3}(x,y,z) & = \frac{1}{1-pxy^2 z} \sum_{k,l,t \geq 0} a(p^k,p^l) a(p^k,p^t) x^l y^k z^t \notag \\
&= \sum_{s=0}^{\infty} p^s x^s y^{2s} z^s \sum_{k,l,t \geq 0} a(p^k,p^l)a(p^k,p^t) x^ly^kz^t \notag \\
& =\sum_{k,l,t \geq 0} ( a(p^k, p^l) a(p^k,p^t) + p a(p^{k-2},p^{l-1}) a(p^{k-2}, p^{t-1})+ \cdots ) x^ly^kz^t \notag \\
& = \sum_{k.l,t \geq 0} a(p^k,p^l, p^t) x^l y^k z^t. \notag
\end{align}
Therefore, $a_{klt} = b_{klt}$.
\end{proof}
\section{moduli parameterizes ideals of a quadratic ring}
We first recall the classical results in the theory of binary quadratic forms \cite{cox}. Given a primitive integral binary quadratic form
$$Q(x, y) = ax^2 + bx y + cy^2$$
with discriminant $D= b^2 -4 ac$ such that $K =\mathbb{Q} ( \sqrt{D})$ a quadratic field, there is a canonical way to associate it an integral ideal $I$ in the quadratic order $R=R(D)$.
Let $\tau $ be one of the two roots of the quadratic function $Q(x, 1) =0$, then
\begin{align}
R = \langle 1, a \tau \rangle \notag \ \mathrm{and} \
I = \langle a, a \tau \rangle .\notag
\end{align}
If $f$ is the conductor of the quadratic order $R$, then
we can express $a\tau$ as:
$$a \tau = \frac{-b \mp f d_K }{2} \pm f w_K, $$
where $ w_ K = \frac{d_K+ \sqrt{d_K}} {2}$, and $d_K$ is the fundamental discriminant of the quadratic field $K$. It follows that $R$ has a $\mathbb{Z}$-basis $[1, f w_K]$, which only depends on the discriminant $D=f^2 d_K$ of $R$. Notice that $I$ is the ideal of $R$ satisfying $R/I \cong \mathrm{N}(I) \mathbb{Z}$, where the norm $\mathrm{N}(I) =|a|$.
In order to extend the above construction to an arbitrary integral binary quadratic form
\begin{equation*} \tag{10}Q(x, y) = ax^2 + bx y + cy^2 \end{equation*}
with discriminant $D= b^2 - 4 ac \neq 0$, we need to first recall the definition of oriented quadratic ring introduced in the paper \cite{bhargava}. A quadratic ring is the commutative ring with unity whose underlying additive group is $\mathbb{Z}^2$. There is a unique automorphism for a quadratic ring $R$. With the automorphism, we can define the trace of an element $x \in R$ by taking $\mathrm{Tr}(x) = x+ x'$, where $x'$ denotes the image of $x$ under the automorphism. Alternatively, the trace function $\mathrm{Tr} : R \to \mathbb{Z}$ is defined as the trace of the endomorphism $R \xrightarrow{ \times \alpha} R$. We also define the norm of an element $x \in R$ by taking $\mathrm{N}(x) = x \cdot x'$. The discriminant $\mathrm{disc}(R)$ of $R$ is defined to be the determinant $\det( \mathrm{Tr}(\alpha_i \alpha_j))$ where $\{ \alpha_i\}$ is any $\mathbb{Z}$-basis of $R$. As the $\mathbb{Z}$-basis of any quadratic $R$ has the form $[1, \tau]$, where $\tau$ satisfies the equation $\tau^2 + r \tau + s = 0$, the discriminant of $R$ is given explicitly by $\mathrm{disc}(R) = r^2 - 4 s$. Conversely, given any integer $D \equiv 0$ or $1 \ ( \rm{mod} \ 4)$, there exists a unique quadratic ring $R(D)$ with discriminant $D$. Canonically, $R(D)$ has a $\mathbb{Z}$-basis $[1, \tau_D]$, where $\tau_D$ is determined by
\begin{equation*}\tag{11}\tau_D^2 = \frac{D} {4} \ \mathrm{or} \ \tau_D^2 = \frac{D-1}{4} +\tau_D,\end{equation*}
in accordance to whether $D \equiv 0 \ ( \rm{mod} \ 4)$ or $D \equiv 1 \ ( \rm{mod} \ 4)$. We call $R$ non-degenerate if $\mathrm{disc}(R) \neq 0$. From now on, we only consider the case of non-degenerate quadratic rings.
For an integer $D \neq 0$, the quadratic ring $R(D)$ has a unique non-trivial automorphism. The quadratic ring $R(D)$ is oriented if we specify the choice of $\tau_{D}$. For an oriented quadratic ring $R(D)$, the specific choice of $\tau$ in any $\mathbb{Z}$-basis $[1, \tau]$ is made such that the change-of-basis matrix from the basis $[1, \tau]$ to the canonical basis $[1, \tau_D]$ has positive determinant. We call such a basis $[1, \tau]$ positively oriented. For the rest of the section, we always assume that the quadratic ring $R(D)$ is oriented with each $\mathbb{Z}$-basis $[1, \tau]$ positively oriented.
Finally, for a quadratic ring $R$ with non-zero discriminant $D$, we define for it the narrow class group $\mathrm{Cl}^{+}(R)$, the group of oriented ideal classes. Recall that an oriented ideal is the pair $(I, \epsilon)$, where $I$ is a (fractional) ideal of $R$ in $K(R) = R \otimes \mathbb{Q}$, and $ \epsilon = \pm 1$ gives the orientation of the ideal $I$. For an element $ k\in K(R)$, the product $ k \cdot (I, \epsilon)$ is defined to be the oriented ideal $(k I, \mathrm{sgn}\left(\mathrm{N}(k)\right) \epsilon)$. Two oriented ideal $(I_1, \epsilon_1)$ and $(I_2, \epsilon_2)$ belong to the same oriented ideal class if they satisfy $(I_1, \epsilon_1) = k \cdot(I_2, \epsilon_2)$. We will suppress $\epsilon$ for the rest of the section and assume $I$ always oriented. For an oriented ideal $I \subset R$, the unoriented norm of $I$ is defined to be $\mathrm{N}(I) = |R /I|$; while the oriented norm of $I$ is denoted by $\epsilon\cdot \mathrm{N}(I)$.
Now for the binary quadratic form $(10)$, the oriented quadratic ring $R$ is defined to be
$$R = \langle 1, \tau\rangle ,$$
where the choice of $\tau$ is specified and it satisfies
$$ \tau^2 + b \tau + ac =0.$$
Further, the oriented ideal $I$ is defined to be
$$I = \langle a, \tau \rangle$$
with the orientation given by the ordered basis $[a, \tau]$.
It is easy to see that $I$ is the ideal contained in $R$ with the norm $\mathrm{N}(I) = |a|$. If the binary quadratic form is primitive, i.e., $g.c.d.(a,b,c) =1$, then the ideal $I$ is proper, which means it has an inverse in the quadratic algebra $K(R)$.
Let $B'_2(\mathbb{Z}) \subset \mathrm{SL}_2(\mathbb{Z})$ be the subgroup of lower-triangular integer matrices with positive diagonal elements. Then we will show the following result which says the set of pairs $(R, I)$ with oriented ideal $I$ with cyclic quotient in $R$ can be parameterized by the integer orbits of the PVS of binary quadratic forms acted on by the Borel subgroup $B'_2(\mathbb{C})$.
\begin{proposition}
The natural map
\begin{align*}
&B'_2(\mathbb{Z}) \backslash \{Q(u, v) = au^2 + b uv + cv^2: b^2 -4ac \neq 0, a \neq 0 \} \\
& \to \mathrm{Iso} \backslash \{(R, I): R/I \cong \mathrm{N}(I) \mathbb{Z} \}
\end{align*}
defined above is a bijection. The isomorphism $f$ from the pair $(R_1, I_1)$ to another $(R_2, I_2)$ is defined to be the orientation-preserving isomorphism from $R_1$ to $R_2$ and sending $I_1$ to $I_2$.
\end{proposition}
\begin{proof}
From the construction above, the map is well defined. We first prove the surjectivity. Given an oriented quadratic ring $R$ with $\mathrm{disc}(R) = D$ and an oriented ideal $I \subset R$ defined by:
$$R = \langle 1, \tau_D \rangle \ \mathrm{and} \ I = \langle \alpha, \beta \rangle,$$
where $\tau_D$ is defined in $(11)$ and the orientation of $I$ is determined by the ordered basis $[\alpha, \beta]$, we can always assume that the norm $\mathrm{N}(\alpha) = \alpha \cdot \alpha' \neq 0$. This is trivial in the number field case. To prove it in the general case, we define a binary quadratic form by
\begin{equation*}\tag{12} (\alpha \cdot \alpha') u^2 - (\alpha' \cdot \beta + \alpha \cdot \beta') uv +( \beta \cdot \beta') v^2 , \end{equation*}
then it is easy to see that the discriminant of $(12)$ is exactly the discriminant of $I$ given by
$$\mathrm{disc}(I) = \left( \det \left( \begin{matrix} \alpha & \beta \\
\alpha' & \beta' \end{matrix} \right) \right)^2.$$
As
$$\mathrm{disc}(I) = (\mathrm{N}(I))^2 \cdot \mathrm{disc}(R),$$
and $R$ is non-degenerate, it implies that $\mathrm{disc}(I) \neq 0$. So at least one of the coefficients of $u^2$ and $v^2$ is non-zero. We can assume $\alpha \cdot \alpha' \neq 0$ by changing the order of $\alpha$ and $\beta$. As $\alpha, \beta \in I$, $\mathrm{N}(I) | \mathrm{N}(\alpha)$ and $\mathrm{N}(I) | \mathrm{N}(\beta)$, it follows that $\mathrm{N}(I)$ is the common factor of all coefficients of $(12)$. After canceling this common factor, we write it as
\begin{equation*} \tag{13} mu^2 + n uv + l v^2\end{equation*}
with $D = n^2- 4 m l $. So the quadratic ring $R(D)$ can be also written as
$$R(D) = [1, \tau_1],$$
where the choice of $\tau_1$ is specified and it satisfies
\begin{equation*}\tag{14} \tau_1^2 + n \tau_1 + m l =0. \end{equation*}
We write
$$ ( \alpha, \beta ) =(1, \tau_1) \left( \begin{array}{cc} p &r \notag \\ q &s\end{array} \right).$$
Substituting $u= \beta $ and $v= \alpha$ into $(13)$, it becomes zero, by comparing it with the defining equation $(14)$ of $\tau_1$, then
\begin{align*} p= ms + nq \ \mathrm{and} \ r = - lq; \quad \mathrm{or} \quad p= -ms \ \mathrm{and} \ r = ns + lq. \end{align*}
As $I$ is an ideal with cyclic quotient, we must have $g.c.d.(q, s) =1$. Then by the elementary divisor theorem, we can transform the matrix $\left( \begin{array}{cc} p &r \notag \\ q &s\end{array} \right)$ by a left multiplication in $B_2(\mathbb{Z})$ and a right multiplication in $\mathrm{SL}_2(\mathbb{Z})$ to the matrix has the form $\left( \begin{array}{cc} a & \ast \notag \\ 0 & 1 \notag \end{array} \right)$ with $a =\pm \mathrm{N}(I)$. Therefore under a certain basis, $R$ and $I$ can be written as
\begin{equation*}\tag{15}R(D) = \langle 1, \tau_2 \rangle \ \mathrm{and} \ I =\langle a, \tau_2 \rangle,\end{equation*}
where the choice of $\tau_2$ is made such that the basis $[1, \tau_2]$ is positively oriented. With $\alpha$ replaced by $a$ and $\beta$ replaced by $\tau_2$, by $(12)$ there is a binary quadratic form. From the discussion above, all coefficients of it are divisible by $|a| = \mathrm{N}(I)$. After canceling this common factor, we have
\begin{equation*} a u^2+ b u v + c v^2,\end{equation*}
with $D = b^2 - 4 ac$. This is the required binary quadratic form.
To prove the injectivity of the map, suppose that two binary quadratic forms
$$Q_i(u,v) = a_i u^2 + b_i u v+ c_i v^2,$$
where $a_1=a_2 =a \neq 0$ and $D = b_i^2 - 4 a_i c_i \neq 0$ for $i =1, 2$, have the same image. We want to prove that $b_2 =b_1+ 2n a$ and $c_2= n^2a+ b_1n +c_1$ for some integer $n$. From the definition of the map, the oriented quadratic ring and the oriented ideal can be written as
\begin{align}R_1 & = \langle 1, \tau_1\rangle , I_1= \langle a, \tau_1 \rangle ,\notag \\
R_2 &= \langle 1, \tau_2\rangle , I_2=\langle a, \tau_2 \rangle\notag \end{align}
respectively, where the choice of $\tau_i's$ are made such that both bases $[1, \tau_1]$ and $[1, \tau_2]$ are positively oriented, and they satisfy
$$\tau_i^2 + b_i \tau_i + a c_i =0.$$
As they have the same image, there exists an isomorphism $f$ from $R_1$ to $R_2$ preserving the orientation, so it has the form:
$$f(\tau_1) = \tau_2 +s ,$$
As it also satisfies $f(I_1 ) = I_2= \mathbb{Z} a + \mathbb{Z} \tau_2$, so we have
\begin{align}
s&= na, \notag \\
b_2 &= b_1+ 2 na, \notag \\
c_2 &= n^2a + b_1n + c_1, \notag
\end{align}
for some integer $n$.
\end{proof}
Now we return to the case of $2 \times 2\times 2$ integer cubes. Giver a $2\times 2\times 2$ integer cube $A$, suppose that the $D = \mathrm{disc}(A) \neq 0$, we consider the two binary quadratic forms $Q_1(A)(u,v)$ and $Q_2(A)(u,v)$ associated to $A$. Suppose that the coefficients $a_i$ of $u^2$ are not zero, then applying the map in the last proposition to each $Q_i(A)(u,v)$, we get the pairs $(R; I_1, I_2)$ where $I_i$ is the oriented ideal with $R / I_i \cong |a_i| \mathbb{Z}$. Explicitly, the map is given by
\begin{align*}\tag{16}
& \{A : Q_i(A)(u,v) = a_i u ^2 + b_i uv + c_i v^2 , i=1,2 \} \notag \\& \rightarrow \{(R; I_1, I_2): I_i = \langle a_i, \tau_i \rangle, \tau_i ^2 + b_i \tau + a_i c_i =0 \}. \notag
\end{align*}
\begin{theorem}
With the notation above. Then the natural map of $(16)$ defines a surjective and finite morphism
$$\left( B_2'(\mathbb{Z}) \times B_2'(\mathbb{Z}) \times \mathrm{SL}_2(\mathbb{Z}) \right) \backslash V^{ss}_{\mathbb{Z}} \to \mathrm{Iso} \backslash \{(R;I_1, I_2): R/I_1 \cong \mathrm{N}(I_1)\mathbb{Z}, R/I_2 \cong \mathrm{N}(I_2)\mathbb{Z} \}.$$
The cardinality $ n( R; I_1, I_2)$ of the fiber is equal to
$$\sigma_1( D_1, a_1, a_2),$$
where $D =D_0 D_1^2= \mathrm{disc}(R)$, and $D_0$ is square-free.
And it satisfies
$$\sum _{\substack{(R; I_1, I_2)/\sim \\ \mathrm{N}(I_i)=|a_i|}} n (R; I_1, I_2) = B(D, |a_1|, |a_2|).$$
\end{theorem}
\begin{proof}
The map is described above and it is easy to see well defined. We first prove the surjectivity of the map. Given a pair $(R; I_1, I_2)$ with $I_i$ an oriented ideal of $R$ and $R/I_i \cong \mathrm{N}(I_i) \mathbb{Z}$, by Proposition 5.1, we know that there are two binary quadratic forms
$$Q_i(u,v) = a_i u^2+ b_i uv+ c_i v^2 $$
for $i =1, 2$ with $D= \mathrm{disc}(R) = b_i^2 - 4 a_i c_i$, such that $R$ and $I_i$ are determined by
\begin{align}
R &= \langle 1, \tau_D \rangle = \langle 1, \tau_1 \rangle = \langle 1, \tau_2 \rangle ,\notag \\
I_1 &= \langle a_1, \tau_1 \rangle \ \mathrm{and} \ I_2 = \langle a_2, \tau_2 \rangle ,\notag
\end{align}
where $\tau_D$ is defined in $(11)$, and $\tau_i$ satisfies
$$ \tau_i^2 + b_i \tau_i +a_i c_i =0$$
for $ i =1,2$.
By the Lemma 3.3, we know that there exists a $2 \times 2 \times 2$ integer cube $A$ such that
$$Q_i(A)(u,v) = Q_i(u,v).$$
Under the correspondence of $(16)$, we conclude that the integer cube $A$ maps to the given pair $(R; I_1, I_2)$.
To prove the second part of the theorem, let $A$ and $A'$ be the two $2\times 2 \times2$ integer cubes which are in the fiber of $(R; I_1, I_2)$ with $\mathrm{N}(I_i)=|a_i|$, then they have the following arithmetic property: $\mathrm{disc}(R) =\mathrm{disc}(A) =\mathrm{disc}(A')=D $, and the two quadratic forms associated to them are the same $Q_i(A)(u,v) = Q_i (A')(u,v) = a_i u^2 + b_i uv + c_i v^2$, where $0 \leq b_i \leq 2 |a_i|-1 $ is assumed. Write $D = D_0 D_1^2$. From our general formula of $B(D, m,n)$ in the Shintani zeta function $Z_{\rm{Shintani}}(s_1, s_2, w)$, we know that the fiber counting function is equal to
$$n(R; I_1, I_2)= \sum_{\substack{d|D_1\\ d|a_1, d|a_2}} f(d), \mathrm{where} \ f(d) = \begin{cases} d \ \mathrm{if} \ (\frac{b_i}{d})^2 \equiv \frac{D}{d^2} \ (\mathrm{mod}\ \frac{4a_i}{d}), \notag \\ 0 \ \mathrm{otherwise} . \ \notag \end{cases}$$
Note that as $b_i^2 \equiv D \ ( \mathrm{mod} \ 4a_i ) $ already holds, so it automatically implies $(\frac{b_i}{d})^2 \equiv \frac{D}{d^2} \ (\mathrm{mod}\ \frac{4a_i}{d})$ if $d | g.c.d.(D_1, a_1, a_2)$.
It follows that
$$n(R; I_1, I_2) = \sigma_1\left( g.c.d.( D_1, |a_1|, |a_2|) \right).$$
Furthermore, the sum of cardinalities over the fixed norms of $\mathrm{N}(I_1) =| a_1|$ and $\mathrm{N}(I_2) =|a_2|$ is exactly
$$B(D, |a_1|, |a_2|) = \sharp \{ A \in V^{ss}_{\mathbb{Z}}/\sim : \mathrm{disc}(A) =D, |\det (A^F)| =|a_1|, |\det(A^L)| = |a_2|\}.$$
\end{proof}
|
\section{Introduction}
{\em Low-rank approximation} in linear algebra refers to the following optimization problem:
\begin{equation}
\label{eq:frobnorm1}
{\rm minimize} \,\,\,\,
||\, X-U\,||_\Lambda^2 \,\,\,= \,\,\,
\sum_{i=1}^m \sum_{j=1}^n \lambda_{ij} (x_{ij} - u_{ij})^2
\,\,\quad \hbox{subject to} \quad {\rm rank}(X) \leq r .
\end{equation}
Here, we are given a real {\em data matrix} $U = (u_{ij})$ of format $m \times n$,
and we wish to find a matrix $X = (x_{ij})$ of rank at most $r$ that is closest to $U$ in
a weighted Frobenius norm. The entries of the {\em weight matrix}
$\Lambda = (\lambda_{ij})$ are positive reals.
If $m \leq n$ and the weight matrix $\Lambda$ is the all-one matrix ${\bf 1}$ then
the solution to (\ref{eq:frobnorm1}) is given by the singular value decomposition
$$
U \,\,=\,\,\, T_1 \cdot {\rm diag}(\sigma_1, \sigma_2, \ldots, \sigma_m) \cdot T_2.
$$
Here $T_1,T_2$ are orthogonal matrices, and
$ \sigma_1 \geq \sigma_2 \geq \cdots \geq \sigma_m $ are the singular values of $U$.
By the Eckart-Young Theorem, the closest matrix of rank $\leq r$ to
the given matrix $U$ equals
\begin{equation}
\label{eq:EY}
U^* \,\, = \,\,\,
T_1 \cdot {\rm diag}(\sigma_1, \ldots, \sigma_r,0, \ldots, 0) \cdot T_2.
\end{equation}
For general weights $\Lambda$, the situation is more complicated,
as seen in the studies \cite{MMH, Rey, SJ}. In particular,
there can be many local minima.
We discuss a small instance in Example \ref{ex:rey}.
In {\em structured low-rank approximation} \cite{CFP, Mar},
we are also given a linear subspace $\mathcal{L} \subset
\mathbb{R}^{m \times n}$, typically containing the data matrix $U$,
and we wish to solve the restricted problem:
\begin{equation}
\label{eq:frobnorm2}
\! {\rm minimize} \,\, ||\, X-U\,||_\Lambda^2 \,=
\sum_{i=1}^m \sum_{j=1}^n \lambda_{ij} (x_{ij} - u_{ij})^2
\,\,\,\, \hbox{subject to} \,\,\, X \in \mathcal{L} \,\,\hbox{and} \,\, {\rm rank}(X) \leq r .
\end{equation}
A best-case scenario for $\Lambda = {\bf 1}$ is this: if $U $ lies in $\mathcal{L}$
then so does the matrix $U^*$ in (\ref{eq:EY}).
This happens for some subspaces $\mathcal{L}$, including
symmetric and circulant matrices, but
most subspaces $\mathcal{L}$
do not enjoy this property (cf.~\cite{CFP}).
Our problem is difficult even for $\Lambda = {\bf 1}$.
Most practitioners use {\em local methods} to solve (\ref{eq:frobnorm2}).
These methods return a local minimum. There are many heuristics
for ensuring that a local minimum is in fact
a global minimum, but there is never a guarantee
that this has been accomplished. Another
approach is to set up {\em sum of squares} relaxations,
which are then solved with semidefinite programming (cf.~\cite{BPT}).
These SOS methods furnish certificates of global optimality
whenever the relaxation is exact. While this does happen
in many instances, there is no a-priori guarantee either.
How then can one reliably find all global optima to
a polynomial optimization problem such as (\ref{eq:frobnorm2})?
Aside from interval arithmetic and domain decomposition techniques,
the only sure method we are aware of is to
list and examine all the critical points. Algorithms
that identify the critical points, notably
{\em Gr\"obner bases} \cite{Fau02}
and {\em numerical algebraic geometry} \cite{BHSW},
find all solutions over the complex numbers
and sort out the
real solutions after the fact.
The number of complex critical points is an intrinsic
invariant of an optimization problem,
and it is a good indicator of the running
time needed to solve that problem exactly.
The study of such algebraic degrees is an active
area of research, and well-developed results are now
available for semidefinite programming \cite{Ran} and
maximum likelihood estimation \cite{CHKS}.
The present paper applies this philosophy
to structured low-rank approximation.
A general degree theory for closest points on algebraic varieties
was introduced by Draisma {\it et al.}~in \cite{DHOST}.
Following their approach, our primary task is to compute the
number of complex critical points of (\ref{eq:frobnorm2}).
Thus, we seek to find the
{\em Euclidean distance degree} (ED degree)~of
$$ \mathcal{L}_{\leq r} \quad := \quad
\bigl\{\, X \in \mathcal{L} \,: \, {\rm rank}(X) \leq r \,\bigr\}.$$
This {\em determinantal variety} is always regarded as a subvariety
of the matrix space $\mathbb{R}^{m \times n}$, and we use the
$\Lambda$-weighted Euclidean distance coming from $\mathbb{R}^{m \times n}$.
We write ${\rm EDdegree}_\Lambda( \mathcal{L}_{\leq r} )$
for the $\Lambda$-weighted Euclidean distance degree
of the variety $\mathcal{L}_{\leq r}$. Thus
${\rm EDdegree}_\Lambda( \mathcal{L}_{\leq r} )$ is the number
of complex critical points of the problem (\ref{eq:frobnorm2})
for generic data matrices $U$.
The importance of keeping track of the
weights $\Lambda$ was highlighted in
\cite[Example 3.2]{DHOST}, for the
seemingly harmless situation when $\mathcal{L}$ is the
subspace of all symmetric matrices in $\mathbb{R}^{n \times n}$.
Our initial focus lies on the
{\em unit ED degree}, when $\Lambda = {\bf 1}$ is the all-one matrix,
and on the {\em generic ED degree}, denoted
${\rm EDdegree}_{\rm gen}( \mathcal{L}_{\leq r} )$,
when the weight matrix $\Lambda$ is generic.
Choosing generic weights $\lambda_{ij}$ ensures that the variety
${\mathcal L}_{\leq r}$ meets the isotropic quadric
transversally, and it hence allows us to apply
formulas from intersection theory such as \cite[Theorem 7.7]{DHOST}.
This paper is organized as follows.
In Section 2 we offer a computational
study of our optimization problem (\ref{eq:frobnorm2})
when the subspace $\mathcal{L}$ is generic of codimension $c$.
Two cases are to be distinguished: either
$\mathcal{L}$ is a vector space, defined by $c$
homogeneous linear equations in the matrix entries, or
$\mathcal{L}$ is an affine space, defined by $c$
inhomogeneous linear equations.
We refer to these as the {\em linear case}
and {\em affine case} respectively.
We present Gr\"obner basis methods for computing all
complex critical points, and we report on
their performance. From the complex critical points, one
identifies all real critical points and all local minima.
In Section 3 we derive some explicit formulas
for ${\rm EDdegree}_{\rm gen}(\mathcal{L}_{\leq r})$
when $\mathcal{L}$ is generic. We cover the four cases
that arise by pairing the affine case and the linear case
with either unit weights or generic weights.
Here we are using techniques from algebraic geometry,
including Chern classes and the analysis of singularities.
In Section 4, we shift gears and we focus on special matrices, namely
Hankel matrices and Sylvester matrices.
Those spaces $\mathcal{L}$ arise naturally
from symmetric tensor decompositions and
approximate {GCD} computations.
These applications require the use of
certain specific weight matrices $\Lambda$
other than ${\bf 1}$.
We close the introduction by presenting two examples
that illustrate the concepts above.
\begin{example} \label{eq:hankel33} \rm
Let $m=n=3$ and
$\mathcal{L} \subset \mathbb{R}^{3 \times 3}$ the
$5$-dimensional space of Hankel matrices:
$$ X \, = \, \begin{bmatrix} x_0 & x_1 & x_2 \\
x_1 & x_2 & x_3 \\
x_2 & x_3 & x_4 \end{bmatrix} ,
\quad
U \, = \, \begin{bmatrix} u_0 & u_1 & u_2 \\
u_1 & u_2 & u_3 \\
u_2 & u_3 & u_4 \end{bmatrix} \quad \hbox{and} \quad
\Lambda \, = \, \begin{bmatrix} \lambda_0 & \lambda_1 & \lambda_2 \\
\lambda_1 & \lambda_2 & \lambda_3 \\
\lambda_2 & \lambda_3 & \lambda_4 \end{bmatrix} .
$$
Our goal in (\ref{eq:frobnorm2})
is to solve the following constrained optimization problem
for $r = 1,2$:
$$ \begin{matrix} {\rm minimize} \,\,
\lambda_0 (x_0 - u_0)^2 +
2 \lambda_1 (x_1 - u_1)^2 +
3 \lambda_2 (x_2 - u_2)^2 +
2 \lambda_3 (x_3 - u_3)^2 +
\lambda_4 (x_4 - u_4)^2 \\
\, \hbox{subject to} \,\, {\rm rank}(X) \leq r .
\end{matrix}
$$
This can be rephrased as an unconstrained optimization problem.
For instance,
for rank $r = 1$, we get a one-to-one parametrization of
$\mathcal{L}_{\leq 1}$ by setting
$x_i = s t^i$, and our problem is to
$$ {\rm minimize} \,\,
\lambda_0 (s- u_0)^2 +
2 \lambda_1 (st - u_1)^2 +
3 \lambda_2 (st^2 - u_2)^2 +
2 \lambda_3 (st^3- u_3)^2 +
\lambda_4 (st^4- u_4)^2 .
$$
The ED degree is the number of critical points
with $t \not= 0$.
We consider three weight matrices:
$$
{\bf 1} \, = \,
\begin{bmatrix} 1 & 1 & 1 \\
1 & 1 & 1 \\
1 & 1 & 1
\end{bmatrix} , \qquad
\Omega \, = \,
\begin{bmatrix} 1 & 1/2 & 1/3 \\
1/2 & 1/3 & 1/2 \\
1/3 & 1/2 & 1
\end{bmatrix} , \qquad
\Theta \, = \,
\begin{bmatrix} 1 & 2 & 2 \\
2 & 2 & 2 \\
2 & 2 & 1
\end{bmatrix} .
$$
Here $\Omega$ gives the usual Euclidean metric when
$\mathcal{L}$ is identified with $\mathbb{R}^5$, and
$\Theta$ arises from identifying $\mathcal{L}$
with the space of symmetric $2 \times 2 \times 2 \times 2$-tensors,
as in Section 4. We compute
$$ \begin{matrix}
{\rm EDdegree}_{\bf 1}(\mathcal{L}_{\leq 1}) = 6, & \quad &
{\rm EDdegree}_\Omega (\mathcal{L}_{\leq 1}) = 10, & \quad &
{\rm EDdegree}_\Theta(\mathcal{L}_{\leq 1}) = 4 , \\
{\rm EDdegree}_{\bf 1}(\mathcal{L}_{\leq 2}) = 9, & \quad &
{\rm EDdegree}_\Omega (\mathcal{L}_{\leq 2}) = 13, & \quad &
{\rm EDdegree}_\Theta(\mathcal{L}_{\leq 2}) = 7 . \\
\end{matrix}
$$
In both cases, $\Omega$ exhibits the generic behavior, so
${\rm EDdegree}_{\rm gen}(\mathcal{L}_{\leq r}) = {\rm EDdegree}_\Omega(\mathcal{L}_{\leq r})$.
We refer to Sections 3 and 4 for larger Hankel matrices
and formulas for their ED degrees.
\hfill $\diamondsuit$
\end{example}
\begin{example} \label{ex:rey} \rm
Let $m=n=3, r = 1$ but now take $\mathcal{L} = \mathbb{R}^{3 \times 3}$,
so this is just the weighted rank-one approximation
problem for $3 \times 3$-matrices.
We know from \cite[Example 7.10]{DHOST} that
${\rm EDdegree}_{\rm gen}(\mathcal{L}_{\leq 1}) = 39$.
We take a circulant data matrix and a circulant weight matrix:
$$ U=\begin{bmatrix}
-59& \phantom{-}11& \phantom{-} 59 \, \\
\phantom{-}11& \phantom{-}59 & -59 \, \\
\phantom{-}59&-59& \phantom{-}11 \,
\end{bmatrix}
\quad \hbox{and} \quad
\Lambda =\begin{bmatrix}
9 & 6 & 1 \\
6 & 1 & 9 \\
1 & 9 & 6
\end{bmatrix}.
$$
This instance has $39$ complex critical points.
Of these, $19$ are real, and $7$ are local minima:
\begin{small} $$
\begin{bmatrix}
\phantom{-}0.0826 \! & \! \phantom{-}2.7921 \! & \! -1.5452 \\
\phantom{-}2.7921 \! & \! \phantom{-}94.3235 \! & \! -52.2007 \\
-1.5452 \! & \! -52.2007 \! & \! \phantom{-}28.8890
\end{bmatrix} \! ,
\begin{bmatrix}
-52.2007 \! & \! \phantom{-}28.8890 \! & \! -1.5452 \\
\phantom{-}2.7921 \! & \! -1.5452 \! & \! \phantom{-}0.0826 \\
\phantom{-}94.3235 \! & \! -52.2007 \! & \! \phantom{-}2.7921
\end{bmatrix} \! ,
\begin{bmatrix}
-52.2007 \! & \! \phantom{-}2.7921 \! & \! \phantom{-}94.3235 \\
\phantom{-}28.8890 \! & \! -1.5452 \! & \! -52.2007 \\
-1.5452 \! & \! \phantom{-}0.0826 \! & \! \phantom{-}2.7921
\end{bmatrix} \!,
$$
$$ \! \!
\begin{bmatrix}
-29.8794 \! & \! \phantom{-}36.2165 \! & \! -27.2599 \\
-32.7508 \! & \! \phantom{-}39.6968 \! & \! -29.8794 \\
\phantom{-}39.6968 \! & \! -48.1160 \! & \! \phantom{-}36.2165
\end{bmatrix} \! \! , \!
\begin{bmatrix}
-48.1160 \! & \! \phantom{-}36.2165 \! & \! \phantom{-}39.6968 \\
\phantom{-}36.2165 \! & \! -27.2599 \! & \! -29.8794 \\
\phantom{-}39.6968 \! & \! -29.8794 \! & \! -32.7508
\end{bmatrix} \!\! , \!
\begin{bmatrix}
-29.8794 \! & \! -32.7508 \! & \! \phantom{-}39.6968 \\
\phantom{-}36.2165 \! & \! \phantom{-}39.6968 \! & \! -48.1160 \\
-27.2599 \! & \! -29.8794 \! & \! \phantom{-}36.2165
\end{bmatrix} \!\!,
$$
\end{small}
$$
\begin{bmatrix}
-25.375 & -25.375 & -25.375 \\
-25.375 & -25.375 & -25.375 \\
-25.375 & -25.375 & -25.375
\end{bmatrix}.
$$
The first three are the global minima.
The last matrix is the local minimum where the objective function has the largest value:
note that each entry equals $-203/8$.
The entries of the first six matrices are algebraic numbers of
degree $10$ over $\mathbb{Q}$. For instance,
the two upper left entries $0.0826$ and $-48.1160$
are among the four real roots of the irreducible polynomial
$$
\begin{small}
\begin{matrix}
164466028468224 x^{10}+27858648335954688 x^9+1602205386689376672 x^8+
7285836260028875412 x^7 \\
-2198728936046680414272 x^6-14854532690380098143152 x^5
+
2688673091228371095762316 x^4 \\ +44612094455115888622678587 x^3
-
41350080445712457319337106 x^2\\
+27039129499043116889674775 x-1977632463563766878765625.
\end{matrix}
\end{small}
$$
Thus, the critical ideal in $\mathbb{Q}[x_{11}, x_{12}, \ldots, x_{33}]$ is not prime.
It is the intersection of six maximal ideals. Their
degrees over $\mathbb{Q}$ are $ 1,2,6,10,10,10$, for a total of $39
= {\rm EDdegree}_{\rm gen}(\mathcal{L}_{\leq 1})$.
\hfill $\diamondsuit $
\end{example}
William Rey \cite{Rey}
reports on numerical experiments with the optimization problem (\ref{eq:frobnorm1}),
and he asks whether the number of local minima is
bounded above by ${\rm min}(m,n)$.
Our Example \ref{ex:rey} gives a negative answer: the number of
local minima can exceed $\min(m, n)$.
This result highlights the value of
our exact algebraic methods for practitioners of optimization.
\section{Gr\"obner Bases}
The critical points of the low-rank approximation problem (\ref{eq:frobnorm2})
can be computed as the solution set of a system of polynomial equations.
In this section we derive these equations, and we demonstrate how
to solve a range of instances using current Gr\"obner basis techniques.
Here, our emphasis lies on the case when $\mathcal{L}$ is a generic
subspace, either linear of affine.
Starting with the linear case, let
$\{L_1,L_2,\ldots,L_s\}$ be a basis of $\mathcal{L}^\perp$, the space of linear forms on
$\mathbb{R}^{m \times n}$ that vanish on $\mathcal{L}$. Thus ${\rm codim}(\mathcal{L}) = s$,
each derivative $\partial L_k / \partial x_{ij}$ is a constant, and
$\mathcal{L} = \{X \in \mathbb{R}^{m \times n} : L_1(X) = \cdots = L_s(X) = 0 \}$.
The case when $\mathcal{L}$ is an affine space can be treated with the
same notation if we take
each $L_i$ to be a linear form plus a constant.
The following implicit formulation of the critical equations is a variation on \cite[(2.1)]{DHOST}.
We begin with the case
$m = n = r+1$. Let $D\in\mathbb{Z}[x_{11},\ldots,x_{nn}]$ denote the determinant of the
$n \times n$-matrix $X = (x_{ij})$. Given
a data matrix $\,U = (u_{ij}) \in\mathcal \mathbb{R}^{n\times n}$, the critical points of
$\sum_{i=1}^n\sum_{j=1}^n \lambda_{ij} (x_{ij}-u_{ij})^2$ on
the determinantal hypersurface
$\mathcal{L}_{\leq n-1} = \{X \in \mathcal{L} : D(X) = 0\}$ verify the following conditions.
The matrix on the right has $s+2$ rows and $n^2$ columns:
$$\begin{cases}D(X)=0\\
L_1(X)=0\\
\quad\quad\vdots\\
L_s(X)=0
\end{cases}
\quad\quad\quad\quad\mathsf{Rank}
\begin{bmatrix}
\partial D/ \partial x_{11}&\cdots& \partial D/ \partial x_{nn}\\
\partial L_1/\partial x_{11}& \cdots & \partial L_1 / \partial x_{nn} \\
\vdots&\ddots&\vdots\\
\partial L_s / \partial x_{11}&\cdots& \partial L_s / \partial x_{nn} \\
\lambda_{11} (x_{11}-u_{11}) &\cdots& \lambda_{nn}(x_{nn}-u_{nn})
\end{bmatrix}
\leq s+1.$$
Any singular point of $\mathcal L_{\leq n-1}$ also satisfies these conditions.
The rank condition on the Jacobian matrix can be modeled by
introducing Lagrange multipliers $z_0,z_1,\ldots,z_s$.
These are new variables.
We now consider the following polynomial system
in $n^2+s+1$ variables:
\begin{equation}\label{eq:EDdeterminant}
\! \begin{cases}D(X)=0\\
L_1(X)=0\\
\quad \vdots\\
L_s(X)=0
\end{cases}
\quad \begin{bmatrix} z_0 \! & \!
\cdots \! &\! z_s \! & \! \! 1
\end{bmatrix}
\! \cdot \!
\begin{bmatrix}
\partial D/ \partial x_{11}&\cdots& \partial D/ \partial x_{nn}\\
\partial L_1/\partial x_{11}& \cdots & \partial L_1 / \partial x_{nn} \\
\vdots&\ddots&\vdots\\
\partial L_s / \partial x_{11}&\cdots& \partial L_s / \partial x_{nn} \\
\lambda_{11}(x_{11}{-}u_{11}) \! &\cdots& \! \lambda_{nn}(x_{nn}{-}u_{nn})
\end{bmatrix}
\, = \, \begin{bmatrix} 0 \! &\! \cdots \! & \! 0 \end{bmatrix}.
\end{equation}
Table \ref{table:EDdeterminant} shows
the number of complex solutions to these equations.
These numbers are obtained from the general formulas in
Section 3, and we verified them using Gr\"obner bases.
\begin{table}
\centering
\begin{tabular}{|c|c|c|c|c|}
\hline
\multicolumn{5}{|c|}{linear, $\Lambda = {\bf 1}$}\\
\hline
$\,\quad n=$ &$2$&$3$&$4$&$5$\\
\hline\hline
$\! s=0 \!$&2&3&4&5\\
$s=1$&4&15&28&45\\
$s=2$&2&31&92&205\\
$s=3$&0&39&188&605\\
$s=4$&&33&260&1221\\
$s=5$&&21&284&1805\\
$s=6$&&9&284&2125\\
$s=7$&&3&284& 2205 \\
$s=8$&&0&284&2205 \\
$s=9$&&&264&2205 \\
$s=10$&&&204&2205 \\
$s=11$&&&120&2205 \\
$s=12$&&&52& 2205 \\
$s=13$&&&16& 2205\\
$s=14$&&&4& 2205 \\
$s=15$&&&0&2205\\
\hline
\end{tabular}
\begin{tabular}{|c|c|c|c|c|}
\hline
\multicolumn{4}{|c|}{affine, $\Lambda = {\bf 1}$}\\
\hline
$2$&$3$&$4$&$5$\\
\hline\hline
2&3&4&5\\
6&15&28&45\\
4&31&92&205\\
2&39&188&605\\
&39&260&1221\\
&33&284&1805\\
&21&284&2125\\
&9&284&2205 \\
&3&284&2205 \\
&&284&2205 \\
&&264&2205 \\
&&204& 2205 \\
&&120& 2205 \\
&&52&2205 \\
&&16&2205 \\
&&4&2205\\
\hline
\end{tabular}
\begin{tabular}{|c|c|c|c|c|}
\hline
\multicolumn{4}{|c|}{linear, $\Lambda$ gen.} \\
\hline
$2$&$3$&$4$&$5$\\
\hline\hline
6&39&284 & 2205 \\
4&39&284 & 2205 \\
2&39&284 & 2205 \\
0&39&284 & 2205 \\
&33&284 & 2205 \\
&21&284 & 2205 \\
&9&284 & 2205 \\
&3&284 & 2205 \\
&0&284 & 2205 \\
&&264 & 2205 \\
&&204 & 2205 \\
&&120 & 2205 \\
&&52 & 2205 \\
&&16 & 2205 \\
&&4 & 2205 \\
&&0 & 2205 \\
\hline
\end{tabular}
\begin{tabular}{|c|c|c|c|}
\hline
\multicolumn{3}{|c|}{affine, $\Lambda$ gen.}\\
\hline
$2$&$3$&$4$\\
\hline\hline
6&39&284 \\
6&39&284 \\
4&39&284 \\
2&39&284 \\
&39&284 \\
&33&284 \\
&21&284 \\
&9&284 \\
&3&284 \\
&&284 \\
&&264 \\
&&204 \\
&&120 \\
&&52 \\
&&16 \\
&&4 \\
\hline
\end{tabular}
\caption{\label{table:EDdeterminant}
The ED degree for the determinant of an $n \times n$-matrix
with linear or affine entries.}
\end{table}
We observe that Table~\ref{table:EDdeterminant} has the following remarkable properties:
\begin{itemize}
\item{} There is a shift between the ED degrees of affine and linear sections
for $s\geq (n-1)^2$. This phenomenon will be explained
in Proposition \ref{prop:affinelinear}.
\item{} The third column is constant for $s\le n(n-2)$ and the fourth column is constant for $s\le n(n-2)+1$.
This will be explained in Corollaries \ref{thm:sEDpolar} and \ref{cor:EDsequal2}.
\item{} The differences between the first and the third column
equal those
between the second and the fourth column. This gap
will be expressed (conjecturally) with formula (\ref{discrepancy}).
\end{itemize}
We now prove the correctness of the formulation (\ref{eq:EDdeterminant}),
and then we discuss our computations.
\begin{proposition} \label{prop:formulation1}
For a generic linear (or affine) space $\mathcal L$ of codimension $s$ and for a generic
data matrix $ U = (u_{ij})$ in $\mathcal L$, the solutions
$(X,z)$ of the polynomial system (\ref{eq:EDdeterminant}) correspond to the critical points
$X$ of the optimization problem (\ref{eq:frobnorm2})
for square matrices of corank one.
\end{proposition}
\begin{proof}
We prove this for linear spaces $\mathcal{L}$. The argument is similar when
$\mathcal L$ is an affine space.
Any solution of the system (\ref{eq:EDdeterminant}) corresponds to a point of $\mathcal L$ where the Jacobian matrix of $(D, L_1,\ldots, L_s, \lVert X-U \rVert_\Lambda^2)$ has a rank defect. There
are two types of such points:
the critical points of the distance function and singular points on the determinantal variety.
Hence it suffices to prove that no point in the singular locus corresponds to a solution of (\ref{eq:EDdeterminant}).
The matrix $U=(u_{ij})_{1\leq i,j\leq n}$ was assumed to be generic, so it has rank~$n$
since $\mathcal L$ is also generic.
If $X$ is a singular point of the linear section of the determinantal variety defined by $D(X)=L_1(X)=\dots=L_s(X)=0$, then there exists $(y_0,y_1,\ldots,y_s) $ with $y_0 \neq 0$ such that
$$\begin{bmatrix}y_0 & y_1 & \cdots & y_s
\end{bmatrix}\cdot
\begin{bmatrix} \partial D/ \partial x_{11}&\cdots& \partial D/ \partial x_{nn} \\
\partial L_1/\partial x_{11} &\dots& \partial L_1/\partial x_{nn} \\
\vdots&\ddots&\vdots\\
\partial L_s / \partial x_{11} &\cdots&
\partial L_s/\partial x_{nn}
\end{bmatrix}
\, = \, \begin{bmatrix} 0 \! &\! \cdots \! & \! 0 \end{bmatrix}. $$
Let us assume by contradiction that $X$ extends to a solution
$(X,z)$ of (\ref{eq:EDdeterminant}).
Then
$$
\begin{bmatrix}
0&
(z_1- y_1 z_0/y_0)
& \! \cdots\! &
(z_s- y_s z_0/y_0)
&1\,
\end{bmatrix}
\cdot
\begin{bmatrix}
\partial D/ \partial x_{11}&\! \cdots \!& \partial D/ \partial x_{nn}\\
\partial L_1/\partial x_{11}&\! \cdots \! & \partial L_1 / \partial x_{nn} \\
\vdots&\! \ddots \!&\vdots\\
\partial L_s / \partial x_{11}&\! \cdots \!& \partial L_s / \partial x_{nn} \\
\lambda_{11}(x_{11}{-}u_{11}) \! &\! \cdots\!& \! \lambda_{nn}(x_{nn}{-}u_{nn})
\end{bmatrix}
= \begin{bmatrix} 0 \! &\! \cdots \! & \! 0 \end{bmatrix}.
$$
This means that $X-U$ belongs to $\mathcal L$ and $\Lambda*(X-U)$ belongs to $\mathcal L^\perp$.
Here $*$ denotes the Hadamard (coordinatewise) product of two matrices.
The scalar product of $X-U$ and $\Lambda*(X-U)$ is zero.
Since all coordinates live in $\mathbb{R}$,
these conditions imply
$\lVert X-U \rVert_\Lambda^2=0$, and hence $X=U$. We get a contradiction since $U$ has full rank,
whereas $D(X) = 0$.
\end{proof}
The values of ${\rm EDdegree}_\Lambda(\mathcal{L}_{\leq n-1})$ in
Table~\ref{table:EDdeterminant} can be verified
computationally with the formulation
(\ref{eq:EDdeterminant}).
We used the implementation of Faug\`ere's Gr\"obner basis algorithm
$F_5$ \cite{Fau02} in the {\tt maple} package {\tt FGb}.
Computing Gr\"obner bases for (\ref{eq:EDdeterminant})
was fairly easy for $n \leq 4$, but
difficult already for $n=5$.
For each of the cases in Table~\ref{table:EDdeterminant},
we computed the ED degree by running {\tt FGb} over
the finite field with $65521$ elements.
However, due to substantial coefficient growth, this did not work over the field
$\mathbb{Q}$ of rational numbers.
Hence, to actually compute all critical points over $\mathbb{C}$
and hence all local minima over $\mathbb{R}$, even for $n = 4$,
a better formulation was required.
In what follows we shall present two such improved formulations.
\smallskip
Duality plays a key role in the computation of the critical points of
the Euclidean distance and was investigated in \cite[\S 5]{DHOST}.
In what follows, we show how to compute
efficiently the critical points of the weighted Euclidean distance of the
determinant by using this duality.
In the following statement we are using the standing
hypothesis that all $\lambda_{ij}$ are non-zero.
\begin{proposition}\label{prop:weightedDuality}
Let $U$ be a generic $m\times n$ matrix with $m\leq n$,
let $\Lambda$ be a weight matrix,
and fix an integer $r\leq \min(m,n)$. Then there is a bijection between the critical points of
\begin{enumerate}
\item[(1)] $Q(X)=\sum_{i,j} \lambda_{ij}(x_{ij}-u_{ij})^2$ on the variety $\mathbb{C}^{m\times n}_{\leq m-r}$ of corank $r$ matrices $X$, and
\item[(2)] $Q_{\rm dual}(Y)=\sum_{i,j}
(y_{ij}-\lambda_{ij} u_{ij})^2/\lambda_{ij}$ on the variety $\mathbb{C}^{m\times n}_{\leq r}$ of rank $r$ matrices $Y$.
\end{enumerate}
For each critical point $X$ of (1), the corresponding critical point $Y$ of (2)
equals $\,Y=\Lambda*U-\Lambda*X$, where $*$ denotes the
Hadamard product. In particular, if $U$ has real entries, then
the bijection sends real critical points of (1) to real critical points of (2) and conversely.
\end{proposition}
\begin{proof}
The critical points of (1) correspond to matrices $X$ such that
the Hadamard product $\Lambda*(U-X)$ is perpendicular to the tangent space
at $X$ of the variety $\mathbb{C}^{m\times n}_{\leq m-r}$ of corank $r$ matrices.
Recall, e.g.~from \cite[\S 5]{DHOST},
that the dual variety to $\mathbb{C}^{m\times n}_{\leq m-r}$ is the variety $\mathbb{C}^{m\times n}_{\leq r}$ of rank $r$ matrices.
Hence, the critical points in (1) can be found by solving
the linear equation $Y=\Lambda*(U-X)$
on the conormal variety.
That {\em conormal variety} is the set of all
pairs $(X,Y)$ such that
$X\in\mathbb{C}^{m\times n}_{\leq m-r}$, $Y\in\mathbb{C}^{m\times n}_{\leq r}$,
$X^{\rm t} \cdot Y = 0$, and
$X \cdot Y^{\rm t} = 0$.
We can now express $X$ in terms of $Y$ and the parameters
by writing $X=\Lambda^{*-1}*(\Lambda*U-Y)$,
where $\Lambda^{*-1}$ denotes the Hadamard (coordinatewise)
inverse of the weight matrix $\Lambda$.
Using biduality, this means that
$X=\Lambda^{*-1}*(\Lambda*U-Y)$ is
perpendicular to the tangent space at $Y$
of the variety $\mathbb{C}^{m\times n}_{\leq r}$.
This is equivalent to the statement that $Y$ is a critical point of (2) on $\mathbb{C}^{m\times n}_{\leq r}$.
\end{proof}
In both Propositions \ref{prop:formulation1} and \ref{prop:weightedDuality},
it is assumed that the given matrix $U$ is generic. Here the term {\em generic}
is meant in the usual sense of algebraic geometry: $U$ lies in the complement
of an algebraic hypersurface.
In particular, that complement is dense
in $\mathbb{R}^{m \times n}$, so $U$ will be generic
with probability one when drawn from a probability measure
supported on $\mathbb{R}^{m \times n}$.
However, an exact characterization of genericity is difficult.
The polynomial that defines the aforementioned hypersurface
is the {\em ED discriminant}. As can be seen in \cite[\S 7]{DHOST},
this is a very large polynomial of high degree, and we will rarely
be able to identify it in an explicit way.
\smallskip
Proposition \ref{prop:weightedDuality} shows that weighted low-rank approximation can be solved by the dual problem. We focus now on the corank 1 case (whose dual problem is rank 1 approximation).
In order to solve the dual problem, we use the parametrization of $n{\times} n$ matrices of rank $1$ by
$$(t_1,\ldots,t_n, z_1,\ldots, z_{n-1})
\,\,\mapsto \,\, \begin{bmatrix} t_1 & t_1 z_1 &\dots&t_1 z_{n-1}\\
\vdots&\vdots&\vdots&\vdots\\
t_n & t_n z_1 &\dots&t_n z_{n-1}
\end{bmatrix}. $$
This expresses the dual problem as an unconstrained optimization problem in $2n-1$ variables:
\begin{equation}\label{eq:unconstrOpt}
{\rm Maximize} \quad
Q_{\rm dual}\,=\sum_{1\leq i,j\leq n} \dfrac 1{\lambda_{ij}}(y_{ij}-\lambda_{ij}
u_{ij})^2, \quad \text{ where }y_{i1}=t_i \text{ and } y_{ij}=t_i z_{j-1}.
\end{equation}
Here, we use the word ``maximize'' in an unconventional way:
what we seek is the critical point furthest to $U$.
That critical point need not be a local maximum; see e.g.~\cite[Figure 4]{DHOST}.
We compute the critical points for (\ref{eq:unconstrOpt}) by
applying Gr\"obner bases to the equations
$$\partial Q_{\rm dual}/\partial t_i\,=\,\partial Q_{\rm dual}/\partial z_j\,=\,0
\qquad \hbox{ for } i\in\{1,\ldots,n\} \,\,\hbox{and} \, \, j\in\{1,\ldots, n-1\}.
$$
The critical points of the primal problem are then found by the formula
$Y=\Lambda*(U-X)$.
\medskip
This concludes our discussion of square matrices of rank $1$ or corank $1$.
We next consider the general case of rectangular matrices
of format $m \times n$ with general linear or affine entries.
We assume $r \leq m \leq n$ and $s \leq mn$.
Let $M$ be a complex $m \times n$-matrix of rank $r$.
Then $M$ is a smooth point in the variety
$\mathbb{C}^{m\times n}_{\leq r}$ of matrices of rank $\leq r$.
Let $\mathsf{Ker}_L(M)$ and $\mathsf{Ker}_R(M)$ denote the left and right kernels of $M$
respectively. The
normal space of $\mathbb{C}^{m\times n}_{\leq r}$ at $M$
has dimension $(m-r)(n-r)$, and it equals
$\mathsf{Ker}_L(M)\otimes \mathsf{Ker}_R(M)
\,\subset \, \mathbb{C}^{m \times n}$ \cite[Chapter 6]{GG}.
Its orthogonal complement is the tangent space at $M$,
which has dimension $rm+rn-r^2$.
In order to construct a polynomial system whose solutions are the critical points of
$X \mapsto || X-U||_\Lambda^2$
on the smooth locus of $\mathcal L_{\leq r}$, we introduce two matrices of unknowns:
$$Y=\begin{bmatrix}
1&\dots&0\\
\vdots&\ddots&\vdots\\
0&\dots&1\\
y_{1,1}&\dots&y_{1,m-r}\\
\vdots&\ddots&\vdots\\
y_{r,1}&\dots&y_{r,m-r}
\end{bmatrix} \quad\quad \hbox{and} \quad\quad Z=\begin{bmatrix}
1&\dots&0\\
\vdots&\ddots&\vdots\\
0&\dots&1\\
z_{1,1}&\dots&z_{1,n-r}\\
\vdots&\ddots&\vdots\\
z_{r,1}&\dots&z_{r,n-r}
\end{bmatrix}.$$
For $i\in\{1,\ldots, m-r\}$, $j\in\{1,\ldots, n-r\}$, let $N^{((m-r)(j-1)+i)}$ be the rank $1$
matrix which is the product of the $i$th column of $Y$ and of the $j$th row of $Z^\intercal$.
We consider the system
\begin{equation}\label{eq:EDdetgeneral}
\!\!\! \begin{cases}
Y^\intercal \cdot X = 0\\
X\cdot Z = 0\\
L_1(X) = 0\\
\quad\quad\vdots\\
L_s(X)=0
\end{cases}\,\, \begin{bmatrix}w_1 \! & \! \cdots \! & \! w_{(m-r)(n-r)+s} \! &\!\! 1
\end{bmatrix}
\begin{small} \!
\begin{bmatrix}
N^{(1)}_{11}&\! \dots \!& N^{(1)}_{mn}\\
\vdots&\! \ddots \!&\vdots\\
N^{((m-r)(n-r))}_{11}&\dots& N^{((m-r)(n-r))}_{mn}\\
\partial L_1 / \partial x_{11}& \! \cdots \! &
\partial L_1 / \partial x_{mn }\\
\vdots&\! \ddots \! &\vdots\\
\partial L_s / \partial x_{11 }&\! \cdots \! &
\partial L_s / \partial x_{mn }\\
\lambda_{11}(x_{11}-u_{11})&\! \dots \! &\lambda_{mn}(x_{mn}-u_{mn})
\end{bmatrix}
\end{small}
=0.
\end{equation}
The rank condition on the matrix in (\ref{eq:EDdetgeneral}) comes from the fact that $M\in \mathcal L_{\leq r}$ is a critical point if the gradient of the distance function at $M$ belongs to the normal space of $\mathcal L_{\leq r}$ at $M$. The first $(m-r)(n-r)+s$ rows of the matrix
span the normal space of $\mathcal L_{\leq r}$ at a smooth point.
This formulation avoids saturating by the singular locus, which is often too costly.
\begin{proposition}\label{prop:generalformulation}
For a generic affine space $\mathcal L$ of codimension $s$ and a generic matrix $U$ in $\mathcal L$, the polynomial system \ref{eq:EDdetgeneral} has finitely many complex solutions which correspond to the critical points of the weighted Euclidean distance function on the smooth locus of $\mathcal L_{\leq r}$.
\end{proposition}
\begin{proof}
This can be derived from \cite[Lemma 2.1]{DHOST}.
It is analogous to Proposition \ref{prop:formulation1}.
\end{proof}
We next discuss our computational experience with Gr\"obner bases.
In Table \ref{table:expeFormulation},
we compare the efficiency of the different approaches on a specific problem: computing the weighted rank $3$ approximation of a $4\times 4$ matrix. The experimental setting is the following: we consider a $4\times 4$ matrix $U$ with integer entries picked uniformly at random in $\{-100,\ldots, 100\}$ and a random weight matrix $\Lambda$ with positive integer entries chosen at random in $\{1,\ldots, 20\}$.
By Table \ref{table:EDdeterminant}, the generic ED degree is $284$ and the
ED degree for $\Lambda = {\bf 1}$ is $4$. We report in Table \ref{table:expeFormulation} the timings for computing a lexicographical Gr\"obner basis with the \texttt{maple} package \texttt{FGb} \cite{Fau02}. Once a Gr\"obner basis is known, isolation techniques may be used to obtain the real roots. The \texttt{maple} package \texttt{fgbrs} provides implementations of such methods.
\begin{table}[h]
\centering
\begin{tabular}{@{}|c|@{}c@{}|@{}c@{}|@{}c@{}|@{}c@{}|@{}}
\cline{2-5}
\multicolumn{1}{c|}{}&\begin{tabular}{c}Determinant\\
primal (\ref{eq:EDdeterminant})
\end{tabular}
&\begin{tabular}{c}Parametric\\
dual (\ref{eq:unconstrOpt})\end{tabular}
&\begin{tabular}{c} Normal space\\primal (\ref{eq:EDdetgeneral})\end{tabular}
&\begin{tabular}{c} Normal space\\dual (\ref{eq:EDdetgeneral}) \! \end{tabular}\\
\hline
$\Lambda $ generic, over $\mathsf{GF}(65521)$&5s&{\bf 1.3s}&6s&8.6s\\
$\Lambda $ generic, over $\mathbb{Q}$&$>1$ day&{\bf 891s}&1327s&927s\\
$\Lambda = {\bf 1}$, over $\mathbb{Q}$&0.3s&{\bf 0.2s}&0.4s&0.5s\\
\hline
\end{tabular}
\caption{\label{table:expeFormulation} Symbolic computation of the
weighted rank $3$ approximations of a $4\times 4$ matrix}
\end{table}
We examine three scenarios. In the first row, the computation is performed over a finite field.
This gives information about the algebraic difficulty of the problem: there is no
coefficient growth, and the timings
indicate the number of arithmetic operations in Gr\"obner bases algorithms. However,
finding local minima requires computing over $\mathbb{Q}$.
In rows 2 and 3 of Table \ref{table:expeFormulation}, we compare the
case of generic weights with the unweighted case
(\ref{eq:EY}) that corresponds to the singular value decomposition ($\Lambda=\mathbf 1$).
The dual problem is easiest to solve, in particular with the
unconstrained formulation (\ref{eq:unconstrOpt}). Note that, for
$s \geq 1$, such an unconstrained formulation is not available, since
$\mathcal{L}_{\leq r}$ is generally not a unirational variety.
\begin{table}[h] \centering
\begin{tabular}{|c|c|c|c|c|c|}
\hline
$(m,n,r)$&$s=0$&$s=1$&$s=2$&$s=3$&$s=4$\\
\hline\hline
$(4,4,2)$&
$\mathbf{4}$/0.42s/1.8s&
$\mathbf{54}$/1.93s/744s&
$\mathbf{230}$/52.7s/--&
$\mathbf{582}$/349.2s/--&
$\mathbf{998}$/1474s/--\\\hline
$(3,4,2)$&
$\mathbf{3}$/0.2s/0.3s&
$\mathbf{15}$/0.3s/7.4s&
$\mathbf{43}$/1.2s/132s&
$\mathbf{71}$/1.5s/1120s&
$\mathbf{83}$/2.2s/2696s\\\hline
$(3,5,2)$&
$\mathbf{3}$/0.3s/0.5s&
$\mathbf{15}$/0.5s/16s&
$\mathbf{43}$/2.1s/400s&
$\mathbf{87}$/7.1s/6038s&
$\mathbf{127}$/16s/59091s\\\hline
\end{tabular}
\medskip
\begin{tabular}{|c|c|c|c|c|}
\hline
$(m,n,r)$&$s=5$&$s=6$&$s=7$&$s=8$\\
\hline\hline
$(4,4,2)$&
$\mathbf{1250}$/2739s/--&
$\mathbf{1250}$/2961s/--&
$\mathbf{1074}$/1816s/--&
$\mathbf{818}$/821s/--\\\hline
$(3,4,2)$&
$\mathbf{83}$/2.3s/4846s&
$\mathbf{83}$/2.1s/5764s&
$\mathbf{73}$/2.2s/4570s&
$\mathbf{49}$/1.0s/1619s\\\hline
$(3,5,2)$&
$\mathbf{143}$/20s/160094s&
$\mathbf{143}$/20s/68164s&
$\mathbf{143}$/20s/99208s&
$\mathbf{143}$/20s/163532s\\\hline
\end{tabular}
\medskip
\begin{tabular}{|c|c|c|c|c|}
\hline
$(m,n,r)$&$s=9$&$s=10$&$s=11$&$s=12$\\
\hline\hline
$(4,4,2)$&
$\mathbf{532}$/349s/--&
$\mathbf{276}$/92s/--&
$\mathbf{100}$/42s/450988s&
$\mathbf{20}$/1.4s/1970s\\\hline
$(3,4,2)$&
$\mathbf{22}$/0.8s/350s&
$\mathbf{6}$/0.3s/6.4s&&\\\hline
$(3,5,2)$&
$\mathbf{128}$/18s/263586s&
$\mathbf{88}$/13s/67460s&
$\mathbf{40}$/1.9s/4568s&
$\mathbf{10}$/0.8s/114s\\\hline
\end{tabular}
\caption{\label{table:EDsecdet}
Symbolic computations for affine sections of determinantal varieties
with $\Lambda = {\bf 1}$.}
\end{table}
In Table \ref{table:EDsecdet}, we report on some Gr\"obner basis
computations with the \texttt{maple} package \texttt{FGb} for $\Lambda
= {\bf 1}$. Here we used the formulation (\ref{eq:EDdetgeneral}). The
ED degree, given in bold face, is followed by
the time, measured in seconds, for computing the graded reverse lexicographic
Gr\"obner basis. The first timing is obtained by
performing the computation over the finite field ${\sf
GF}(65521)$; the second one is obtained by computing over the field of rationals $\mathbb{Q}$. The symbol ``$-$'' means that we did not obtain the Gr\"obner basis after seven days of computation.
An important observation in Table \ref{table:EDsecdet} is the correlation
between the reported running times and the values of ${\rm EDdegree}_{\mathbf 1}$.
The former tell us how many arithmetic operations are needed to find a Gr\"obner basis.
This suggests that the ED degree is an accurate measure for
the complexity of solving low-rank approximation problems with symbolic algorithms,
and it serves as a key motivation for computing
ED degrees using advanced tools from algebraic geometry.
This will be carried out in the next section, both for
$\Lambda$ generic and for~$\Lambda = {\bf 1}$.
In particular, we shall arrive at theoretical explanations for the
ED degrees in Tables \ref{table:EDdeterminant}
and \ref{table:EDsecdet}.
\section{Algebraic Geometry}\label{sec:alggeo}
The study of ED degrees for
algebraic varieties was started in \cite{DHOST}.
This section builds on and further develops the geometric theory in that paper.
We focus on the low rank approximation problem (\ref{eq:frobnorm2}), and
we derive general formulas for the ED degrees in
Tables \ref{table:EDdeterminant} and \ref{table:EDsecdet}.
We recall that an affine variety $X\subset\mathbb{C}^{N+1}$ is an {\em affine cone}
if $x\in X$ implies $tx\in X$ for every $t\in\mathbb{C}$. The variety of $m
\times n$-matrices of rank $\leq r$ is an affine cone. If
$X\subset\mathbb{C}^{N+1}$ is an affine cone, then the corresponding
projective variety $\mathbb{P} X\subset\mathbb{P}^{N}$ is well defined. The ED degree of $\mathbb{P} X$ is,
by the definition given in \cite[\S 2]{DHOST}, the ED degree of its affine cone $X$.
The following proposition explains the shift between the third and fourth
column of Table~\ref{table:EDdeterminant}. More generally, it shows that we can restrict the analysis to linear sections, since the ED degree (for generic weights)
in the affine case can be deduced from the linear case.
\begin{proposition}\label{prop:affinelinear}
Let $X\subset\mathbb{C}^{N+1}$ be an affine cone,
let ${\mathcal A}^s$ (resp. ${\mathcal L}^{s}$) be a generic affine (resp. linear) subspace
of codimension $s \geq 1$ in $\mathbb{C}^{N+1}$. The following equality holds
\begin{equation}
\label{eq:confirm}
\mathrm{EDdegree}_{\rm gen}(X\cap {\mathcal A}^{s})
\,\,=\,\,
\mathrm{EDdegree}_{\rm gen}(X\cap {\mathcal L}^{s-1}).
\end{equation}
\end{proposition}
\begin{proof}
Let $\overline{X}\subset\mathbb{P}^{N+1}$ be the projective closure of $X$.
From \cite[Theorem 6.11]{DHOST}, we get that
$\mathrm{EDdegree}_{\rm gen}(X)=\mathrm{EDdegree}_{\rm gen}(\overline{X})$, since the transversality assumptions
in that result are satisfied for general weights.
From the equality $\overline{X\cap {\mathcal A}^s}=\overline{X}\cap {\mathcal L}^s$, we conclude that
$\mathrm{EDdegree}_{\rm gen}(X\cap {\mathcal A}^s)=\mathrm{EDdegree}_{\rm gen}(\overline{X}\cap {\mathcal L}^s)=
\mathrm{EDdegree}_{\rm gen}(\mathbb{P} X\cap {\mathcal L}^{s-1})=\mathrm{EDdegree}_{\rm gen}(X\cap {\mathcal L}^{s-1})$. Here, the second equality follows from
$\mathbb{P} X=\overline{X}\cap {\mathcal L}^1$.
\end{proof}
Consider a projective variety $X $ embedded in $ \mathbb{P}^N$ with a generic system of coordinates.
It was shown in \cite[Theorem 5.4]{DHOST} that
${\rm EDdegree}_{\rm gen}(X)$ is the sum of the degrees of the polar classes
$\delta_i(X)$. Here, $\delta_i(X)$ denotes the degree
of the {\em polar class} of $X$ in dimension $i$, as in \cite{Holme}. Moreover, if
${\mathcal L}^s$ is a generic linear subspace of codimension $s$ in $\mathbb{P}^N$ then
$\,\delta_i(X\cap {\mathcal L}^s)=\delta_{i+s}(X)\,$ by \cite[Corollary 6.4]{DHOST}.
We call {\em $s$-th sectional ED degree} of $X$ the number ${\rm EDdegree}_{\rm gen}(X\cap {\mathcal L}^s)$.
We denote by $X^*$ the dual variety of $X$,
as in \cite[\S 5]{DHOST}, and already seen
in the proof of Proposition \ref{prop:weightedDuality}.
\begin{corollary}\label{thm:sEDpolar}
The $s$-th {\em sectional ED degree} of $X$ is expressed
in terms of polar classes as
\begin{equation}
\label{eq:sumdelta}
\mathrm{EDdegree}_{\rm gen}(X\cap {\mathcal L}^s)\,\,=\,\sum_{\ell \ge s}\delta_\ell (X) .
\end{equation}
If $\,s \leq \mathrm{codim}(X^*)-1\,$ then the varieties
$X$ and $X \cap {\mathcal L}^s$ have the same generic ED degree.
\end{corollary}
\begin{proof}
This follows from results in Sections 5 and 6 in \cite{DHOST}.
To compute $\mathrm{ED degree\ }(X\cap {\mathcal L}^s)$ we have to sum
$\delta_i(X)$ for $i\ge s$. However, it is known that
$\delta_i(X)=0$ if $i\le \textrm{codim}(X^*)-2$.
\end{proof}
A special role in \cite{DHOST} is played by the
{\em isotropic quadric} $ Q = V(x_0^2 + x_1^2 + \cdots + x_N^2)$ in $\mathbb{P}^N$.
If $X$ is smooth and transversal to $Q$ then
\cite[Theorem 5.8]{DHOST} gives an explicit formula
for the ED degree in terms of Chern classes of $X$ $c_i(X)$.
A thorough treatment of Chern classes can be found in \cite{Fulton};
the reader interested in the applications in this paper can be referred to
the basics provided in \cite{DHOST}.
By combining \cite[Theorem 5.8]{DHOST}
with Corollary \ref{thm:sEDpolar}, we obtain
\begin{theorem} \label{thm:thmA}
Let $X \subset \mathbb{P}^N$ be a smooth projective variety of dimension $M$
and assume that $X$ is transversal to the isotropic quadric $Q$.
Then the $s$-th sectional ED degree of $X$ equals
$$ \mathrm{ED degree}_{\rm gen}(X\cap {\mathcal L}^s)\,\,\, = \,\,\,
\sum_{\ell = s}^{M}\sum_{k=\ell }^{M} (-1)^{M-k} \binom{k+1}{\ell+1} \deg(c_{M-k}(X)). $$
\end{theorem}
\begin{proof} The inner sum is equal to the polar class $\delta_i(X)$;
see the proof of \cite[Theorem 5.8]{DHOST}.
\end{proof}
We now apply Theorem \ref{thm:thmA} to the situation when
$M = m+n-2$, $N = mn-1$, and
$X = \mathbb{P}^{m-1} \times \mathbb{P}^{n-1}$ is the
Segre variety of $m\times n$ matrices of rank $1$ in $\mathbb{P}^N$.
The Chern polynomial of the tangent bundle of $X$ in the Chow ring $A^*(X)=\mathbb{Z}[s,t]/\langle s^m,t^n \rangle$
equals $(1+s)^m(1+t)^n$. By \cite[page 150]{Holme}, this implies
\begin{equation}
\label{eq:delta_l}
\delta_\ell(X) \,\,\, = \,\, \sum_{k=\ell}^{m+n-2} \! (-1)^{m+n-k} \binom{k+1}{\ell+1} V_k ,
\end{equation}
where $V_k = \deg(c_{M-k}(X)) $ is
the coefficient of $s^{m-1}t^{n-1}$ in the expansion of the polynomial
$(1+s)^m(1+t)^n(s+t)^{k}$. Toric geometers
may view $V_k$ as the sum of the normalized volumes of all $k$-dimensional
faces of the product of simplices $\Delta_{m-1} \times \Delta_{n-1}$; see
\cite[Corollary 5.11]{DHOST}.
\smallskip
The following result explains the ED degrees in the third column in Table \ref{table:EDdeterminant},
and it allows us to
determine this column for any desired value of $m$, $n$ and $s$:
\begin{theorem}
Let $m \leq n$ and $\mathcal{L}$ be a generic linear subspace
of codimension $s$ in $\mathbb{R}^{m \times n}$.
For matrices of rank $1$ or matrices of corank $1$, the generic ED degree
is given by
\begin{equation}
\label{eq:EDdegdual}
\begin{matrix}
\,{\rm EDdegree}_{\rm gen}(\mathcal{L}_{\leq 1}) & = & \delta_s(X) + \delta_{s+1}(X) + \cdots + \delta_{m+n-2}(X), \\
{\rm EDdegree}_{\rm gen}(\mathcal{L}_{\leq m-1}) & = & \delta_0(X) + \delta_1(X) + \cdots + \delta_{mn-2-s}(X),
\end{matrix}
\end{equation}
where $\delta_\ell(X)$ may be computed from (\ref{eq:delta_l}).
\end{theorem}
\begin{proof}
The dual in $\mathbb{P}^{mn-1}$ to the Segre variety $X= \mathbb{P}^{m-1} \times \mathbb{P}^{n-1}$
is the variety $X^*$ of matrices of rank $\leq m-1$.
By \cite[Theorem 2.3]{Holme}, we have $\delta_\ell(X)=\delta_{mn-2-\ell}(X^*)$ for all $\ell$.
With this duality of polar classes,
the result follows from Corollary \ref{thm:sEDpolar} and \cite[Theorem 5.4]{DHOST}.
\end{proof}
\begin{example} \rm
Fix $m=n=3$. For matrices of rank $1$, the formulas
(\ref{eq:delta_l}) and (\ref{eq:EDdegdual}) give
$$\begin{array}{c|c|c|c|c|c|c|c|c|}
s = {\rm codim}(\mathcal{L}) &0&1&2&3&4&5&6&7\\
\hline
V_s & 9 & 18 & 24 & 18 & 6 & 0 & 0 & 0 \\
\hline
\delta_s (X)&3&6&12&12&6&0&0&0\\
\hline
{\rm EDdegree}_{\rm gen}(\mathcal{L}_{\leq 1}) &39&36&30&18&6&0&0&0
\end{array}$$
Duality for polar classes now yields the formulas for
$3 \times 3$-matrices of rank $r=2$ in $\mathcal{L}$:
$$\begin{array}{c|c|c|c|c|c|c|c|c|}
s = {\rm codim}(\mathcal{L}) &0&1&2&3&4&5&6&7\\
\hline
\delta_s(X^*) = \delta_{7-s}(X) &0&0&0&6&12&12&6&3\\
\hline
{\rm EDdegree}_{\rm gen}(\mathcal{L}_{\leq 2})
&39&39&39&39&33&21&9&3
\end{array}$$
This is our theoretical derivation of the third column
in Table \ref{table:EDdeterminant} for $n=3$ and generic
$\Lambda$. \hfill $\diamondsuit $
\end{example}
Writing down closed formulas for intermediate values of $r$
is more difficult: it involves some Schubert calculus.
However, ${\rm EDdegree}_{\rm gen}(\mathcal{L}_{\leq r})$
can be conveniently computed with the following script in {\tt Macaulay2} \cite{M2}.
It is a slight generalization of that in~\cite[Example 7.10]{DHOST}:
\begin{verbatim}
loadPackage "Schubert2"
ED=(m,n,r,s)->
(G = flagBundle({r,m-r}); (S,Q) = G.Bundles;
X=projectiveBundle (S^n); (sx,qx)=X.Bundles;
d=dim X; T=tangentBundle X;
sum(toList(s..m*n-2),i->sum(toList(i..d),j -> (-1)^(d-j)*binomial(j+1,i+1)*
integral(chern(d-j,T)*(chern(1,dual(sx)))^(j)))))
\end{verbatim}
The function {\tt ED(m,n,r,s)} computes the ED degree of the variety
of $m\times n$ matrices of rank $\le r$, in general coordinates, cut with
a generic linear space
of codimension $s$ in $\mathbb{P}^{mn-1}$.
For $s=0$ this is precisely the function displayed in
\cite[Example 7.10]{DHOST}.
\begin{example} \label{ex:1350}
\rm
The bold face ED degrees in Table \ref{table:EDsecdet} were computed
for unit weights $\Lambda = {\bf 1}$. To find the analogous numbers
for generic weights $\Lambda$, we run our {\tt Macaulay2} code as follows:
\begin{verbatim}
apply(12,s->ED(4,4,2,s))
{1350, 1350, 1350, 1350, 1330, 1250, 1074, 818, 532, 276, 100, 20}
apply(12,s->ED(3,4,2,s))
{83, 83, 83, 83, 83, 83, 73, 49, 22, 6, 0, 0}
apply(12,s->ED(3,5,2,s))
{143, 143, 143, 143, 143, 143, 143, 143, 128, 88, 40, 10}
\end{verbatim}
\vskip -0.9cm
\hfill $\diamondsuit $
\end{example}
At this point, we wish to reiterate the main thesis of this paper, namely
that knowing the ED degree ahead of time is
useful for practitioners who seek to find and certify
the global minimum in the optimization problem (\ref{eq:frobnorm2}),
and to bound the number of local minima.
The following example illustrates this for one
of the numbers {\tt 83} in
the output in Example \ref{ex:1350}.
\begin{example} \rm We here solve the
generic weighted structured low-rank approximation problem over the reals
with parameters $m=3$, $n=4$, $r=2$ and $s=2$.
Consider the instance
$$\begin{array}{rl}
& U=\begin{bmatrix}-9& 4& 9 & -10\\
10 & 6 & 1 & -9\\
10 & 5 & 7 & 6\end{bmatrix}\qquad\qquad
\Lambda=\begin{bmatrix}
8 & 6 & 8 & 2\\
1 & 8 & 7 & 9\\
7 & 2 & 4 & 6\end{bmatrix}\medskip
\\
L_1(X) =& \! -10 x_{11} + 4 x_{12} + 6 x_{13} + 8 x_{14} + 4 x_{21} - 9 x_{22} +x_{23} - 10 x_{31}
{-} 10 x_{32} {-} 8 x_{33} {+} 2 x_{34} {-} 1,\medskip\\
L_2(X) = & 2 x_{11} + 7 x_{12} + 3 x_{13} - 7 x_{14}
- 4 x_{21} - 6 x_{22} - 7 x_{23} +5 x_{24} + 8 x_{31} +2 x_{33} + 3 x_{34} - 1.
\end{array}$$
We wish to find the matrix $X$ of rank at most $2$ that satisfies the
affine constraints $L_1(X)=L_2(X)=0$ and is
nearest to $U$.
Using Gr\"obner
bases computations and real isolation techniques via the {\tt
Maple} packages {\tt FGb} and {\tt fgbrs}, we find that the
weighted distance function has 83 complex critical points. This
matches the theoretical value $\mathrm{ED}(3,4,2,2)=83$ provided in Example~\ref{ex:1350}, so that we are guaranteed that there are no further critical points. Among them, seven are real and we obtain certified
numerical approximations of their values:
$$ \footnotesize
\begin{bmatrix}
\phantom{-}0.764&\!\! -1.457 \! &\!\! \phantom{-}2.436&\!\! \phantom{-}1.870\\
\phantom{-}0.753&\!\! -0.0154 \! &\!\! \phantom{-}0.030&\!\! -7.437\\
\phantom{-}2.020&\!\! -4.371\! &\!\! \phantom{-}7.308&\!\! \phantom{-}8.330
\end{bmatrix} \!
\begin{bmatrix}
-8.0341 \! &\!\! \phantom{-}4.127&\!\! \phantom{-}9.055&\!\! \phantom{-}5.364\\
\phantom{-}16.936 \! &\!\! \phantom{-}2.930&\!\! -1.330&\!\! -4.220\\
\phantom{-}9.429 \! &\!\! \phantom{-}7.525&\!\! \phantom{-}8.258&\!\! \phantom{-}1.242
\end{bmatrix} \!
\begin{bmatrix}
-8.215 \! &\!\! \phantom{-}5.033&\!\! \phantom{-}9.965&\!\! \phantom{-}1.647\\
\phantom{-}16.848 \! &\!\! \phantom{-}4.259&\!\! \phantom{-}0.423&\!\! -3.669\\
\phantom{-}9.070 \! &\!\! \phantom{-}6.218&\!\! \phantom{-}5.842&\!\! -2.054
\end{bmatrix}$$
$$ \footnotesize \begin{bmatrix}
-8.586&\!\! -1.743&\!\! \phantom{-}1.591&\!\! \phantom{-}2.436\\
\phantom{-}11.191&\!\! \phantom{-}2.985&\!\! -4.232&\!\! -7.159\\
\phantom{-}10.351&\!\! \phantom{-}0.292&\!\! \phantom{-}3.567&\!\! \phantom{-}7.185
\end{bmatrix}
\begin{bmatrix}
-4.853&\!\! \phantom{-}4.081&\!\! \phantom{-}6.301&\!\! -6.349\\
-6.067&\!\! \phantom{-}5.029&\!\! \phantom{-}8.600&\!\! -8.251\\
\phantom{-}2.616&\!\! -2.455&\!\! -0.878&\!\! \phantom{-}2.327
\end{bmatrix}
\begin{bmatrix}
-2.308&\!\! -4.584&\!\! \phantom{-}3.566&\!\! -5.484\\
-0.205&\!\! -2.210&\!\! \phantom{-}0.668&\!\! -3.178\\
-2.276&\!\! \phantom{-}0.983&\!\! \phantom{-}2.444&\!\! \phantom{-}2.810
\end{bmatrix}$$
$$ \footnotesize
\begin{bmatrix}
-9.664&\!\! \phantom{-}2.805&\!\! \phantom{-}7.113&\!\! -10.754\\
\phantom{-}14.942&\!\! \phantom{-}6.520&\!\! \phantom{-}3.149&\!\! -8.783\\
\phantom{-}8.344&\!\! \phantom{-}0.615&\!\! -2.185&\!\! \phantom{-}2.177
\end{bmatrix}
$$
The last matrix is the closest critical point on the manifold of rank
$2$ matrices satisfying $L_1 = L_2 = 0$.
This computation takes 1002 seconds and the most
time-consuming step is the computation of the Gr\"obner basis. In
order to certify that the global minimum is among these matrices, we
also solve the same low-rank approximation problem for rank $1$
matrices. Using the same method, this provides us with $11$ rank $1$
matrices with real entries in $79$ seconds. None of them is closer to
$U$ than the best rank $2$ approximation. Consequently, the global
minimum of the weighted distance is reached at the last matrix in the above list.
For comparison
purposes, with the same constraints $L_1,L_2$ and same data
matrix $U$ but by taking the Frobenius distance (\emph{i.e.} $\Lambda$
is the unit matrix), the number of complex critical points is $43$. Five
of them are real. Here, it takes only 27 seconds to find the global minimizer.
These computations have been performed on an {\tt Intel Xeon E7540/2.00GHz}.
\hfill $\diamondsuit $
\end{example}
In Table \ref{table:EDdeterminant} and Example \ref{ex:1350}
we observed that
the sectional ED degree for generic $\Lambda$ does not
depend on $s = {\rm codim}(\mathcal{L})$, provided $s$ is small.
The following corollary explains this.
\begin{corollary}\label{cor:EDsequal2}
For a generic linear subspace $\mathcal{L} $
of codimension $s < r(r+n-m)$ in $\mathbb{C}^{m \times n}$,
$$ {\rm EDdegree}_{\rm gen}(\mathcal{L}_{\le r})
\,\,=\,\, {\rm EDdegree}_{\rm gen}(\mathbb{C}^{m \times n}_{\le r}). $$
\end{corollary}
\begin{proof}
Let $X$ be the variety of matrices of rank $\leq r$.
Its dual $X^*$ is the variety of matrices of rank $\leq m-r$ and has codimension ${\rm codim}(X^*) = (r+n-m)r$. This implies
$\delta_\ell(X) = 0$ for $\ell < (r+n-m)r-1$ and
the assertion now follows directly from Corollary
\ref{thm:sEDpolar}.
\end{proof}
Corollary \ref{cor:EDsequal2} can be stated informally like this:
in the setting of generic weights and generic linear spaces of matrices with sufficiently
high dimension,
the algebraic complexity of structured low-rank approximation
agrees with that of ordinary low-rank approximation.
\smallskip
Shifting gears, we now consider the case of
unit weights $\Lambda = {\bf 1}$. Thus, we fix
$Q = V(\sum x_{ij}^2)$ as the isotropic
quadric in $\mathbb{P}^{mn-1}$.
Let $X= \mathbb{P}^{m-1} \times \mathbb{P}^{n-1}$ denote
the Segre variety of $m\times n$ matrices of rank~$1$ in $\mathbb{P}^{mn-1}$,
and let $Z = {\rm Sing}(X \cap Q)$ denote the
non-transversal locus of the intersection of $X$ with $Q$.
The dual variety $X^*$ consists of all matrices of rank $\leq m-1$ in $\mathbb{P}^{mn-1}$.
We conjecture that the following formula (put $m=n$) holds for
the gap between the third and the first column of Table \ref{table:EDdeterminant},
(or between the fourth and the second, as well),
\begin{equation}
\label{eq:3ED}
{\rm EDdegree}_{\rm gen}(X^* \cap {\mathcal L}^s)-{\rm EDdegree}_{\bf 1}(X^* \cap {\mathcal L}^s) \,\,= \,\,
{\rm EDdegree}_{\rm gen}(Z \cap {\mathcal L}^s ).
\end{equation}
To compute the right-hand side, and to test this conjecture,
we use the following lemma:
\begin{lemma} \label{lemma:nontrans}
The locus where $Q$ meets $X=\mathbb{P}^{m-1}\times\mathbb{P}^{n-1}$
non-transversally in $\mathbb{P}^{mn-1}$ is the product $\,Z=Q_{m-2}\times Q_{n-2}$,
where $Q_{i-2}$ denotes a general quadratic hypersurface in $\mathbb{P}^{i-1}$.
\end{lemma}
\begin{proof}
The Segre variety
$X$ meets $Q$ in the union of two irreducible components, $\mathbb{P}^{m-1}\times Q_{n-2}$
and $Q_{m-2}\times \mathbb{P}^{n-1}$. The non-transversality locus is the intersection of these components.
\end{proof}
\begin{example} \rm
Let $m=n=2$, so $X$ and $X^*$ represent $3 \times 3$-matrices of rank $1$ and rank $\le 2$ respectively.
Here $Z=Q_{1}\times Q_{1}$ corresponds to the Segre quadric $\mathbb{P}^1\times\mathbb{P}^1$,
embedded in $\mathbb{P}^8$ with the line bundle $\mathcal{O}(2,2)$. This is a toric surface
whose polygon $P$ is twice a regular square.
The facial volumes as in \cite[Corollary 5.1]{DHOST} are
$V_0 = 4$, $V_1 = 8 $ and $V_2 = 8$, and hence
$$ \delta_0(Z)=4-2\cdot 8+3\cdot 8=12 \,, \,\quad
\delta_1(Z)=-8+3\cdot 8=16 \, , \,\quad
\delta_2(Z)=8 . $$
We fill this into a table and, using Corollary \ref{thm:sEDpolar},
we compute the sectional ED degree of $Z$:
$$\begin{array}{c|c|c|c|c|c|c|c|c|}s&0&1&2&3&4&5&6&7\\
\hline
\delta_s(Z)&12&16&8&0& 0 & 0 & 0 & 0 \\
\hline
\mathrm{ED degree}_{\rm gen}(Z\cap {\mathcal L}^s)&36&24&8&0&0&0&0&0\\
\hline
\mathrm{ED degree}_{\rm gen}(X^*\cap {\mathcal L}^s)&39&39&39&39&33&21&9&3\\
\hline
\mathrm{ED degree}_{\bf 1}(X^*\cap {\mathcal L}^s) &3&15&31&39&33&21&9&3\\
\end{array}$$
The last two lines are taken from
Table~\ref{table:EDdeterminant}, and they confirm the formula (\ref{eq:3ED}).
\hfill $\diamondsuit$
\end{example}
Combining Lemma \ref{lemma:nontrans}, Corollary \ref{thm:sEDpolar} and the proof
of \cite[Theorem 5.8]{DHOST}, and abbreviating
$W_{j}=\deg(c_{m+n-4-j}(Q_{m-2}\times Q_{n-2}))$,
the right-hand side of (\ref{eq:3ED})
can be expressed~as
\begin{equation} \label{discrepancy} \begin{matrix}
\displaystyle \sum_{i=s}^{m+n-4}\,\,\, \sum_{j=i}^{m+ n-4}(-1)^{m+n-4-j}\binom{j+1}{i+1}W_{j}.
\end{matrix}
\end{equation}
Moreover, $W_{j}$ is equal to the coefficient of
$t^{m-2}s^{n-2}$ in the rational generating function
$$ 4\frac{(1+t)^m(1+s)^n}{(1+2t)(1+2s)}(t+s)^{j}. $$
This computation
allows us to
extend Table~\ref{table:EDdeterminant} to any desired value of $m$, $n$ and $s$.
\smallskip
Changing topics,
we now consider the
case when $\mathcal{L}$ is the space of
{\em Hankel matrices}.
The computation of low-rank approximation of Hankel matrices
will be our topic in Section~\ref{sec:HankelSylvester}.
In this section we focus on algebraic geometry and
derive a formula for the generic ED degree.
Set $d = p{+}q{-}2$ and
let $X_{d,r}$ denote the variety of
$p \times q$ Hankel matrices of rank $\leq r$.
See (\ref{ex:hankel56}) for examples.
This variety lives in the
projective space $\mathbb{P}^d = \mathbb{P}(S^d \mathbb{C}^2)$, whose points
represent binary forms of degree $d$.
Thus $X_{d,1}$ is the rational normal curve of degree $d$,
and $X_{d,r}$ is the $r$th secant variety of this curve.
We have $\,{\rm dim}(X_{d,r}) = 2r-1$ for $r+1\le\min(p,q)$.
\begin{theorem} \label{thm:EDsecant}
Let $d=p+q-2$ and $r+1\le\min(p,q)$. The generic ED degree of the variety $X_{d,r}$
of $p \times q$ Hankel matrices of rank $\leq r$
in $\mathbb{P}^d$ equals
\begin{equation}
\label{eq:2binomials}
{\rm EDdegree}_{\rm gen}(X_{d,r})\,\,\, = \,\,\,\sum_{i=0}^r \binom{d+1-r}{i} \binom{d-r-i}{r-i} 2^{r-i}.
\end{equation}
\end{theorem}
\begin{proof}
We note that the sum in (\ref{eq:2binomials})
is the coefficient of $z^r$~in the generating function
\begin{equation}
\label{eq:genfunc}
\frac{(1+z)^{d+1-r}}{(1-2z)^{d-2r+1}} .
\end{equation}
The conormal variety of $X_{d,r}$ is the closure~$\mathcal N_{X_{d,r}}$ of
the set
$$ \bigl\{\,(f,g)\,|\,\,\mathrm{rank}(f) = r\,\,
{\rm and} \,\,g\textrm{\ is tangent to\ }
X_{d,r}\textrm{\ at\ }f \bigr\} \,
\subset \,
\mathbb{P}(S^d \mathbb{C}^2 )\times\mathbb{P}(S^d(\mathbb{C}^2)^{*}). $$
The homology class of $\mathcal N_{X_{d,r}}$ is given by a binary form.
We will show that the sum $\sum_i \delta_i(X_{d,r})$
of its coefficients is the asserted coefficient of (\ref{eq:genfunc}).
By \cite[(5.3)]{DHOST}, this proves the~claim.
Let $p_1$, $p_2$ be the two projections.
The images of the conormal variety $\mathcal N_{X_{d,r}}$ are
$$ p_1(\mathcal N_{X_{d,r}})\,=\,X_{d,r}\quad
\hbox{and} \quad p_1(\mathcal N_{X_{d,r}})\,=\,X_{d,r}^*.$$
We desingularize $X_{d,r}$ by considering
$\mathrm{Sym}^r(\mathbb{P}^1)\simeq\mathbb{P}^r$. The desingularization map
is given by the scheme-theoretic intersection of
the rational normal curve of degree $r$ with a hyperplane.
A point in $\mathbb{P}^r$, identified with a hyperplane, gives $r$ points on $X_{d,1} \simeq \mathbb{P}^1$.
Their linear span in $\mathbb{P}^d$
defines a rank $r$ bundle on $\mathbb{P}^r$,
known as the {\em Schwarzenberger bundle} \cite[\S 6]{DK}.
This is the kernel of the bundle map $\mathcal{O}^{d+1}\to\mathcal{O}(1)^{d+1-r}$.
In the same way, we desingularize the conormal variety
$\mathcal N_{X_{d,r}}$ by the fiber product over $\mathbb{P}^r$ of the
projectivization of the
Schwarzenberger bundle $E_{d,r}={\rm kernel}(\mathcal{O}^{d+1}\to\mathcal{O}(1)^{d+1-r})$
and of the projective bundle of $\mathcal{O}(2)^{d-2r+1}$.
Exactly as in the proof of \cite[Proposition 4.1]{BR}, the degrees of the polar classes
of $X_{d,r}$ are
$$\delta_{r+i-1}(X_{d,r})\,\,=\,\int_{\mathbb{P}^r}s_{i}(E_{d,r})s_{r-i}(\mathcal{O}(2)^{d-2r+1}). $$
The total Segre class of $E_{d,r}$ is $(1+z)^{d+1-r}$.
The total Segre class of $\mathcal{O}(2)^{d-2r+1}$ is
$\frac{1}{(1-2z)^{d-2r+1}}$.
By multiplying them we obtain the degree sum
of the polar classes, thus proving (\ref{eq:genfunc}).
\end{proof}
\begin{corollary} \label{cor:powerof3}
The generic ED degree of the
hypersurface $X_{2r,r}$ defined by the
Hankel determinant
of format $(r+1)\times (r+1)$ is equal to
\begin{equation}
\frac{3^{r+1}-1}{2} \quad = \quad
\hbox{the coefficient of $z^r$ in} \,\,\,
\frac{\,(1+z)^{r+1}}{1-2z} .
\end{equation}
\end{corollary}
This corollary means that the ED degree of the
$(r{+}1) \times (r{+}1)$ Hankel determinant
agrees with the ED degree of the
general symmetric $(r{+}1) \times (r{+}1)$ determinant.
By ED duality \cite[Theorem 5.2]{DHOST}, this also
the ED degree of the second Veronese embedding of $\mathbb{P}^r$;
see \cite[Example 5.6]{DHOST}.
If we consider Hankel matrices of fixed rank $r$ then
we obtain polynomiality:
\begin{corollary} \label{cor:forfixed}
For fixed $r$, the generic ED degree of $X_{d,r}$ is a polynomial
of degree $r$ in $d$.
\end{corollary}
For example, we find the following explicit polynomials
when the rank $r$ is small:
$$ \begin{matrix}
{\rm EDdegree}_{\rm gen}(X_{d,1}) & = & 3d-2 ,\\
{\rm EDdegree}_{\rm gen}(X_{d,2}) & = & (9d^2-39d+38)/2 ,\\
{\rm EDdegree}_{\rm gen}(X_{d,3}) & = & ( 9 d^3-99 d^2+348d-388)/ 2 ,\\
{\rm EDdegree}_{\rm gen}(X_{d,4}) & = & ( 27d^4-558d^3+4221d^2-13818d+16472)/ 8.
\end{matrix}
$$
The values of these polynomials are the entries in
the columns on the left in Table~\ref{table:EDHankel} below.
\section{Hankel and Sylvester Matrices}\label{sec:HankelSylvester}
In this section we study the weighted low-rank approximation
problem for matrices with a special structure that is given by
equating some matrix entries and setting others to zero.
One such family consists of the Hurwitz matrices in \cite[Theorem 3.6]{DHOST}.
We here discuss Hankel matrices, then
catalecticants, and finally
Sylvester matrices.
The corresponding applications are
low-rank approximation of symmetric tensors
and approximate greatest common divisors.
The {\em Hankel matrix} $H[p,q]$ of format $p \times q$ has
the entry $x_{i+j-1}$ in row $i$ and column $j$.
So, the total number of unknowns is $n = p+q-1$.
We are most interested in the case when
this matrix is square or almost square.
The {\em Hankel matrix of order $n$}
is $H[(n{+}1)/2,(n{+}1)/2]$ if $n$ is odd,
and it is $H[(n/2,(n{+}2)/2]$ if $n$ is even.
We denote this matrix by $H_n$. For instance,
\begin{equation}
\label{ex:hankel56}
H_5 = \begin{bmatrix}
x_1 & x_2 & x_3 \\
x_2 & x_3 & x_4 \\
x_3 & x_4 & x_5
\end{bmatrix} \quad \hbox{and} \quad
H_6 = \begin{bmatrix}
x_1 & x_2 & x_3 & x_4 \\
x_2 & x_3 & x_4 & x_5 \\
x_3 & x_4 & x_5 & x_6
\end{bmatrix} .
\end{equation}
For approximations by low-rank Hankel matrices,
we consider three natural weight matrices:
\begin{itemize}
\item the matrix $\Omega_n$ has entry
$\,1/{\rm min}(i{+}j{-}1,n{-}i{-}j{+}2 )\, $ in row $i$ and column $j$;
\item the matrix ${\bf 1}_n$ has all entries equal to $1$;
\item the matrix $\Theta_n$ has entry
$\,\binom{n-1}{i+j-2}/{\rm min}(i{+}j{-}1,n{-}i{-}j{+}2 )\, $
in row $i$ and column $j$.
\end{itemize}
We encountered these matrices for $n=5$ in
Example \ref{eq:hankel33}. For $n = 6$ we have
$$ \Omega_6 \, = \,
\begin{bmatrix}
1 & 1/2 & 1/3 & 1/3 \\
1/2 & 1/3 & 1/3 & 1/2 \\
1/3 & 1/3 & 1/2 & 1
\end{bmatrix} \quad \hbox{and} \quad
\Theta_6 \, = \,
\begin{bmatrix}
1 & 5/2 & 10/3 & 10/3 \\
5/2 & 10/3 & 10/3 & 5/2 \\
10/3 & 10/3 & 5/2 & 1
\end{bmatrix} .
$$
The weights $\Omega_n$ represent
the usual Euclidean distance in $\mathbb{R}^n$,
the unit weights ${\bf 1}_n$ give the
Frobenius distance in the ambient matrix space,
and the weights $\Theta_n$ give
the natural metric in the space
of symmetric $2 {\times} 2 {\times} \cdots {\times} 2$-tensors.
Such a tensor corresponds to a binary form
$$ F(s,t) \,= \,
\sum_{i=1}^n \binom{n-1}{i-1} \cdot x_i \cdot s^{n-i} \cdot t^{i-1}. $$
The Hankel matrix $H_n$ has rank $1$ if and only if
$F(s,t)$ is the $(n{-}1)$st power of a linear form. More generally,
if $F(s,t)$ is the sum of $r$ powers of linear forms
then $H_n$ has rank $\leq r$.
As we saw in \S 3, this locus corresponds to the
$r$th secant variety of the rational normal curve in $\mathbb{P}^{n-1}$.
Various ED degrees for our three
weight matrices are displayed in Table \ref{table:EDHankel}.
\begin{table}
\centering
\begin{tabular}{|c|c|c|c|c|}
\hline
\multicolumn{5}{|c|}{$\Lambda = \Omega_n $}\\
\hline
$\! n \backslash r \! $ &$1$&$2$&$3$&$4$\\
\hline\hline
$3$& 4 & & & \\
$4$& 7 & & & \\
$5$& 10 & 13 & & \\
$6$& 13 & 34 & & \\
$7$& 16 & 64 & 40 &\\
$8$& 19 & 103 & 142 &\\
$9$& 22 & 151 & 334 & 121 \\
\hline
\end{tabular} \qquad
\begin{tabular}{|c|c|c|c|c|}
\hline
\multicolumn{5}{|c|}{$\Lambda = {\bf 1}_n $}\\
\hline
$ \! n \backslash r \! $ &$1$&$2$&$3$&$4$\\
\hline\hline
$3$&2 & & & \\
$4$&7 & & & \\
$5$&6 &9 & & \\
$6$&13 &34 & & \\
$7$&10 &38&34&\\
$8$&19 & 103& 142&\\
$9$&14 & 103 & 246&113\\
\hline
\end{tabular} \qquad
\begin{tabular}{|c|c|c|c|c|}
\hline
\multicolumn{5}{|c|}{$\Lambda = \Theta_n $}\\
\hline
$\! n \backslash r \! $ &$1$&$2$&$3$&$4$\\
\hline\hline
$3$& 2& & & \\
$4$& 3& & & \\
$5$& 4&7
& & \\
$6$& 5& 16 & & \\
$7$& 6& 28& 20&\\
$8$& 7& 43& 62&\\
$9$& 8& 61& 134 &53\\
\hline
\end{tabular}
\caption{\label{table:EDHankel}
Weighted ED degrees for Hankel matrices of order $n$ and rank $r$.}
\end{table}
The entries in the leftmost chart in Table \ref{table:EDHankel} come from the formula in
Theorem~\ref{thm:EDsecant}. Indeed, the variety of Hankel matrices $H_n$ of rank $\leq r$ is
precisely the secant variety $X_{n-1,r}$ we discussed in Section 3.
The weight matrix $\Lambda = \Omega_n$ exhibits the generic ED degree for that variety.
The columns on the left of Table \ref{table:EDHankel} are
the values of the polynomials in Corollary \ref{cor:forfixed},
and the diagonal entries $4, 13, 40, 121,\ldots$ are given by
Corollary \ref{cor:powerof3}.
All ED degrees in Table \ref{table:EDHankel}
were verified using Gr\"obner basis computations
over $\mathsf{GF}(65521)$ using the {\tt maple} package {\tt FGb} \cite{Fau02}.
The running times are closely tied to the valued of the ED degrees,
and they are similar to those reported in Table \ref{table:EDsecdet}.
Gr\"obner bases over $\mathbb{Q}$ can also be computed fairly easily
whenever the ED degree is below $100$, and for those cases we can locate
all real critical points using {\tt fgbrs}. However, for larger
instances, exact symbolic solving over $\mathbb{Q}$
becomes a considerable challenge due to the growth in coefficient size.
\smallskip
Hankel matrices of rank $r$ correspond to
symmetric $2 {\times} 2 {\times} \cdots {\times} 2$-tensors
of tensor rank~$r$, and these can be represented by binary forms
that are sums of $r$ powers of linear forms.
That is the point of the geometric discussion in Section \ref{sec:alggeo}.
This interpretation extends to symmetric tensors of arbitrary format,
with the rational normal curve replaced with the Veronese variety.
For a general study of
low-rank approximation of symmetric tensors
see Friedland and Stawiska \cite{FS}.
In general, there is no straightforward representation
of low rank tensors by low rank matrices with special structure.
However, there are some exceptions, notably for rank $r=2$
tensors, by the results of Raicu \cite{Rai} and others in the
recent tensor literature (cf.~\cite{Lan}).
The resulting generalized Hankel matrices are known as
{\em catalecticants} in the commutative algebra literature,
or as {\em moment matrices} in the optimization literature.
We now present a case study that arose from
a particular application in biomedical imaging.
We consider the following catalecticant matrix of format $6 \times 6$:
$$ X \quad = \quad \begin{bmatrix}x_{400}& x_{310}& x_{301}& x_{220}& x_{211}& x_{202}\\
x_{310}& x_{220}& x_{211}& x_{130}& x_{121}& x_{112}\\
x_{301}& x_{211}& x_{202}& x_{121}& x_{112}& x_{103}\\
x_{220}& x_{130}& x_{121}& x_{040}& x_{031}& x_{022}\\
x_{211}& x_{121}& x_{112}& x_{031}& x_{022}& x_{013}\\
x_{202}& x_{112}& x_{103}& x_{022}& x_{013}& x_{004}
\end{bmatrix}$$
The fifteen unknown entries are the coefficients of a ternary quartic
$$ \begin{matrix}
F(s,t,u) \, = \,\,
x_{400} s^4 {+}
x_{040} t^4 {+}
x_{004} u^4 +
6 x_{220} s^2 t^2 {+}
6 x_{202} s^2 u^2 {+}
6 x_{022} t^2 u^2 +
4 x_{310} s^3 t {+}
4 x_{301} s^3 u \\ \qquad \qquad \quad + \,
4 x_{130} s t^3 {+}
4 x_{031} t^3 u {+}
4 x_{103} s u^3 {+}
4 x_{013} t u^3 +
12 x_{211} s^2 t u {+}
12 x_{121} s t^2 u {+}
12 x_{112} s t u^2. \end{matrix} $$
The table $(x_{ijk})$ can be regarded as a
symmetric tensor of format $3 \times 3 \times 3 \times 3$.
The coefficients of the monomials in $F(s,t,u)$ indicate
the multiplicity with which the $15$ unknowns occur
among the $3^4 = 81$ coordinates of that tensor.
To model the invariant metric in the tensor space
$\mathbb{R}^{3 \times 3 \times 3 \times 3}$ in our matrix representation, we use the weight matrix
$$ \Theta \,= \,
\begin{bmatrix}
1 & 2 & 2 & 2 & 3 & 2 \\
2 & 2 & 3 & 2 & 3 & 3 \\
2 & 3 & 2 & 3 & 3 & 2 \\
2 & 2 & 3 & 1 & 2 & 2 \\
3 & 3 & 3 & 2 & 2 & 2 \\
2 & 3 & 2 & 2 & 2 & 1
\end{bmatrix} .
$$
The problem is to approximate a given catalecticant matrix
$U = (u_{ijk})$
by a rank $2$ matrix with respect to the weight matrix $\Theta$.
The expected number of critical points is as follows.
\begin{proposition} \label{prop:195}
Let $\mathcal{L}$ be the $15$-dimensional subspace of
catalecticants $X$ in
$\mathbb{R}^{6 \times 6}$.
Then
$${\rm EDdegree}_\Theta(\mathcal{L}_{\leq 2}) = 195
\quad \hbox{and} \quad
{\rm EDdegree}_{\rm gen}(\mathcal{L}_{\leq 2}) = 1813.
$$
\end{proposition}
The proof is a computation as explained below. We first discuss a numerical instance.
\begin{example} \label{ex:schultz1} \rm
We consider the following symmetric $3 \times 3 \times 3 \times 3$ tensor:
$$
\begin{array}{|c|c|}
\hline
u_{400}&0.1023\\
u_{220}&0.0039\\
u_{310}& -0.002\\
u_{103}&0.0196\\
u_{211}&-0.00032569\\
\hline
\end{array}\quad\quad
\begin{array}{|c|c|}
\hline
u_{040}&0.0197\\
u_{202}&0.0407\\
u_{301}&0.0581\\
u_{031}&0.0029\\
u_{121}&-0.0012\\
\hline
\end{array}\quad\quad
\begin{array}{|c|c|}
\hline
u_{004}&0.1869\\
u_{022}&-0.00017418\\
u_{130}&0.0107\\
u_{013}&-0.0021\\
u_{112}&-0.0011\\
\hline
\end{array}
$$
This tensor was given to us by Thomas Schultz, who heads the
Visualization and Medical Image Analysis Group at the University of Bonn.
It represents a fiber distribution function,
estimated from diffusion Magnetic Resonance Imaging.
See \cite{Sch} for more information.
\hfill $\diamondsuit$
\end{example}
We present an algebraic formulation of our
optimization problem which was found to be suitable for symbolic computation.
Introducing six unknowns $a,b,c,d,e,f$, we
parametrize the $6$-dimensional variety of
symmetric $3 \times 3 \times 3 \times 3$-tensors of rank $2$
by the ternary quartics
$$ \tilde F(s,t,u) \quad = \quad a \cdot (s+b t +c u)^4\,+\,d \cdot (s+e t+f u)^4. $$
Written out explicitly, this parametrization takes the form
$$
\begin{array}{rcl}
x_{400} &=& a+d \\
x_{220}&=& a b^2+d e^2\\
x_{310} &=& a b+d e \\
x_{103} &=& a c^3+d f^3\\
x_{211} &=& a b c+d e f \\
\end{array}\quad\quad
\begin{array}{rcl}
x_{040} &=& a b^4+d e^4\\
x_{202} &=& a c^2+d f^2 \\
x_{301} &=& a c+d f\\
x_{031} &=& a b^3 c+d e^3 f \\
x_{121}&=& a b^2 c+d e^2 f \\
\end{array}\quad\quad
\begin{array}{rcl}
x_{004} &=& a c^4+d f^4\\
x_{022} &=& a b^2 c^2+d e^2 f^2\\
x_{130} &=& a b^3+d e^3 \\
x_{013} &=& a b c^3+d e f^3\\
x_{112} &=& a b c^2+d e f^2 \\
\end{array}\quad\quad
$$
Note that our parametrization is $2$ to $1$: every rank $2$ catalecticant $X$
has two preimages, which are related by swapping the vectors
$(a,b,c)$ and $(d,e,f)$.
The fiber jumps in dimension over the singular locus,
which consists of matrices $X$ of rank $1$. Their
preimage in parameter space is given by the ideal
$\langle a d\rangle\cap \langle b-e, c-f\rangle$.
The chosen weight matrix $\Theta$ now specifies the following
unconstrained optimization problem. We seek to find the minimum in $\mathbb{R}^6$ of
$$\begin{matrix}
G(a,b,c,d,e,f) \,\,= \,\,
(u_{400}-a-d)^2+(u_{040}-a b^4-d e^4)^2+(u_{004}-a c^4-d f^4)^2 \qquad \qquad \\
+ 6 (u_{220}-a b^2-d e^2)^2+ 6 (u_{202}-a c^2-d f^2)^2
+6 (u_{022}-a b^2 c^2-d e^2 f^2)^2 \\
+4 (u_{310}-a b-d e)^2+4 (u_{301}-a c-d f)^2+4 (u_{130}-a b^3-d e^3)^2\\
+4 (u_{103}-a c^3-d f^3)^2+4 (u_{031}-a b^3 c-d e^3 f)^2+4 (u_{013}-a b c^3-d e f^3)^2 \\
\qquad + 12 (u_{112}-a b c^2-d e f^2)^2+12 (u_{211}-a b c-d e f)^2+12 (u_{121}-a b^2 c-d e^2 f)^2.
\end{matrix}
$$
The set of complex critical points is the zero locus of the ideal
$$I \quad = \quad \left\langle \frac{\partial G}{\partial a},\frac{\partial G}{\partial b},\frac{\partial G}{\partial c},\frac{\partial G}{\partial d},\frac{\partial G}{\partial e},\frac{\partial G}{\partial f}\right\rangle :
\bigl(\,\langle a d \rangle\cap \langle b-e, c-f\rangle \, \bigr)^\infty.$$
For applications, we are interested in the real points in this variety.
\begin{proof}[Computational proof of Proposition \ref{prop:195}]
As argued in \cite[\S 2]{DHOST},
the ideal $I$ is radical and zero-dimensional
when the $u_{ijk}$ are generic rational numbers.
The number of solutions is the degree of $I$,
and we found this to be $ 370 = 2 \cdot 195$.
This is twice the ED degree of $\mathcal{L}_{\leq 2}$
with respect to $\Lambda = \Theta$.
For this computation we used the {\tt FGb}
library in {\tt maple}. We used Gr\"obner bases over the finite field ${\sf
GF}(65521)$ to avoid the swelling of rational coefficients, the data
$u_{ijk}$ are chosen uniformly at random in this field, and we
saturate only by $\langle a d (b-e)\rangle$.
The computation took 90 seconds and returned 390 critical points of $G$.
Performing the same computation with the coefficients $1,6,4,12$ in $G(a,b,c,d,e,f)$
replaced with random field elements, we find $3626 = 2 \cdot 1813 $ critical points, and hence
${\rm EDdegree}_{\rm gen}(\mathcal{L}_{\leq 2}) = 1813$.
\end{proof}
\begin{example} \label{ex:schultz2} \rm
We return to the particular data set in
Example \ref{ex:schultz1}. Using the above parametrization, the best
rank 2 approximation can be obtained by solving a system of polynomial
equations. This can be achieved by using symbolic or numerical
methods.
A numerical computation conducted by Jose Rodriguez with the software
{\tt Bertini} indicates that, for Thomas Schultz' data, precisely $9$
of the $195$ critical points are real. These correspond to $2$ local
minima and $7$ saddle points of the Euclidean distance
function. The precomputation with generic data took 2 hours on
40 {\tt AMD Opteron 6276/2.3Ghz} cores. Then the computation
with the numerical data in Example \ref{ex:schultz2} was achieved in 1 minute.
These results were also computed by symbolic methods: a Gr\"obner basis
computation conducted by Jean-Charles Faug\`ere and Mohab Safey El
Din with the software {\tt FGb} returned an algebraic parametrization of the $195$ complex critical points
by the roots of a univariate polynomial of degree $195$.
This polynomial has $9$ real roots. Two of them correspond to the two
local minima. The average size of the integer coefficients of this univariate
polynomial is $11000$ digits.
For this computation, the above formulation as an unconstrained optimization problem was used.
It took 11 minutes on a 2.6GHz {\tt IntelCore i7}.
In general, for symbolic methods,
unconstrained formulations seem to be better
than the general implicit formulation in Proposition~\ref{prop:generalformulation}.
See the comparisons of timings in Table~\ref{table:EDsecdet}.
However, most instances of (\ref{eq:frobnorm2})
do not admit an unconstrained formulation,
because $\mathcal{L}_{\leq r}$ is usually not unirational.
\hfill $\diamondsuit$ \end{example}
Our last topic in this section is the study of Sylvester matrices.
We consider two arbitrary polynomials $F$ and $G$ in one variable $t$.
Suppose their degrees are $m$ and $n$ with $m \leq n$, so
$$ F(t) \,=\, \sum_{i=0}^m a_i t^i \quad \hbox{and} \quad
G(t) \,=\, \sum_{j=0}^n b_j t^j . $$
Fix $k $ with $ 1\leq k \leq m$.
The $k$-th Sylvester matrix of the pair $(F,G)$ equals
$$
{\rm Syl}_k(F,G) \quad = \quad
\begin{bmatrix}
a_0& 0 & \cdots & 0 & b_0& 0 & \cdots & 0 \\
\vdots & a_0 & \ddots& \vdots & \vdots & b_0 & \ddots& \vdots \\
a_m & \vdots & \ddots & 0 & b_n & \vdots & \ddots & 0 \\
0 & a_m & \vdots & a_0 & 0 & b_n & \vdots & b_0 \\
\vdots & \vdots & \ddots & \vdots & \vdots & \vdots & \ddots & \vdots \\
0 & 0 & \cdots & a_m & 0 & 0 & \cdots & b_n
\end{bmatrix}
$$
This matrix has $n+k$ rows and
$n-m+2k$ columns, so it is square for $k=m$,
and it has more rows than columns for $k < m$.
The maximal minors have size $n-m+2k$,
and they all vanish precisely when
${\rm Syl}_k(F,G)$ has a non-zero
vector in its kernel. Such a vector
corresponds to a polynomial of degree $m-k+1$ that is a common factor
of $F$ and $G$.
The {\em approximate gcd problem} in computer algebra \cite{KL, KYZ}
aims to approximate a given pair $(F,G)$
by a nearby pair $(F^*,G^*)$ whose
Sylvester matrix ${\rm Syl}_k(F^*,G^*)$
has linearly dependent columns.
Writing $\mathcal{L}$ for the subspace of
Sylvester matrices, this is precisely our
ED problem for $ \mathcal{L}_{\leq n-m+2k-1}$.
The following theorem furnishes a formula
for ${\rm EDdegree}_{\rm gen}(\mathcal{L}_{\leq n-m+2k-1})$.
\begin{theorem} \label{thm:FG}
For the variety of pairs $(F,G)$ of univariate polynomials
of degrees $(m,n)$ with a common factor of degree $m{-}k{+}1$,
the generic ED degree equals that of the
Segre variety of $(m{-}k{+}2) \times (n{-}m{+}2k)$-matrices of rank $1$.
It is given by setting $s = 0$ in (\ref{eq:EDdegdual}). Using the {\tt Macaulay2} function {\tt ED} in
Example \ref{ex:1350}, we can write this ED degree~as
$$\,{\rm EDdegree}_{\rm gen}(\mathcal{L}_{\leq n-m+2k-1}) \,\,= \,\,
{\tt ED(m-k+2,n-m+2*k,1,0)} . $$
\end{theorem}
\begin{proof}
A natural desingularization is given by multiplying with the
desired common factor:
\begin{equation}
\label{eq:segremap}
\begin{matrix}
\mathbb{P}^{m-k+1} \times \mathbb{P}^{n-m+2k-1} & \to & \mathcal{L}_{\leq n-m+2k-1}, \\
\left[\,A(t)\,,\,\left(B(t),C(t)\right)\,\right] & \mapsto & \left[\,A(t)B(t),A(t)C(t) \,\right].
\end{matrix}
\end{equation}
Here $A(t), B(t),C(t)$ are polynomials of degrees
$m-k+1, k-1,n-m+k-1$ respectively.
The map (\ref{eq:segremap})
lifts to a linear projection map
from the Segre embedding of $ \mathbb{P}^{m-k+1} \times \mathbb{P}^{n-m+2k-1}$.
Work of Piene \cite[\S 4]{Pie} implies that
the degrees of polar loci can be computed
on that Segre variety.
The ED degree is a sum of degrees of these,
by Corollary \ref{thm:sEDpolar}.
The result follows.
\end{proof}
For $m = k$, when the Sylvester matrix is square,
Theorem \ref{thm:FG} refers to
$2 \times (n{+}m)$-matrices of rank $1$.
Similarly to \cite[Example 5.12]{DHOST},
their ED degree is $4(m{+}n)-2$.
We conclude:
\begin{corollary}
The generic ED degree of the
Sylvester determinant ${\rm Syl}_m $ equals $4(m+n)-2$.
\end{corollary}
We consider three
natural choices of weight matrices for the
low-rank approximation of Sylvester matrices.
As before in Table \ref{table:EDHankel}, we write
$\Omega_{m,n}$
for the weight matrix
that represents the Euclidean distance on $\mathbb{R}^{m+n+2}$: it is the matrix which has the same pattern as ${\rm Syl}_k$ with $a_i$ and $b_j$ replaced respectively by $1/(n-m+k)$ and $1/k$.
We also write $\Theta_{m,n}$ for the weight matrix of
the rotation invariant quadratic form: $a_i$ is replaced by $1/((n-m+k)\binom{m}{i})$ and $b_j$ is replaced by $1/(k\binom{n}{j})$. In Table \ref{table:EDSylvester} we present the
ED degrees for these choices of weights.
The left table shows
the generic behavior predicted by Theorem \ref{thm:FG}.
At present, we do not
know a general formula for the entries of the two tables on the right side, but
we are hopeful that an approach like \eqref{eq:3ED}
will lead to such formulas.
Along the rightmost margins,
where the matrix ${\rm Syl}_m $ is square,
the formula seems to be
$\,{\rm EDdegree}_{\Theta}( \mathcal{L}_{\leq n+k-1})\,= \, 2n$.
\begin{table}[h]
\centering
\begin{tabular}{@{}|@{\,\,}c@{\,\,}|@{\,\,}c@{\,\,}|@{\,\,}c@{\,\,}|@{\,\,}c@{\,\,}|@{\,\,}c@{\,\,}|@{}}
\hline
\multicolumn{5}{|c|}
$\Lambda$ is generic
}\\
\hline
$(m,n) \backslash k \! $ &$1$&$2$&$3$&$4$\\
\hline\hline
$\!\! (2,2)$& 10 & 14 & & \\
$\!\! (2,3)$& 39 & 18 & & \\
$\!\! (2,4)$& 83 & 22 & & \\
$\!\! (2,5)$& 143 & 26 & & \\
$\!\! (3,3)$& 14 & 83 & 22 & \\
$\!\! (3,4)$& 83 & 143 & 26 &\\
$\!\! (3,5)$& 284 & 219 & 30 &\\
$\!\! (4,4)$& 18 & 284 & 219 & 30 \\
$\!\! (4,5)$& 143 & 676 & 311 & 34 \\
\hline
\end{tabular}
\begin{tabular}{@{}|@{\,\,}c@{\,\,}|@{\,\,}c@{\,\,}|@{\,\,}c@{\,\,}|@{\,\,}c@{\,\,}|@{\,\,}c@{\,\,}|@{}}
\hline
\multicolumn{5}{|c|}{$\Lambda = \Omega_{m,n}$
}\\
\hline
$(m,n) \backslash k \! $ &$1$&$2$&$3$&$4$\\
\hline\hline
$\!\! (2,2)$& 2 & 6 & & \\
$\!\! (2,3)$& 23 & 18 & & \\
$\!\! (2,4)$& 75 & 22 & & \\
$\!\! (2,5)$& 119 & 18 & & \\
$\!\! (3,3)$& 2 & 19 & 10 & \\
$\!\! (3,4)$& 35 & 95 & 26 &\\
$\!\! (3,5)$& 188 & 203 & 26 &\\
$\!\! (4,4)$& 2 & 36 & 59 & 14 \\
$\!\! (4,5)$& 47 & 276 & 215 & 34 \\
\hline
\end{tabular}
\begin{tabular}{@{}|@{\,\,}c@{\,\,}|@{\,\,}c@{\,\,}|@{\,\,}c@{\,\,}|@{\,\,}c@{\,\,}|@{\,\,}c@{\,\,}|@{}}
\hline
\multicolumn{5}{|c|}{$\Lambda = \Theta_{m,n}$}\\
\hline
$(m,n) \backslash k \! $ &$1$&$2$&$3$&$4$\\
\hline\hline
$\!\! (2,2)$&2 & 4 & & \\
$\!\! (2,3)$& 19 & 6 & & \\
$\!\! (2,4)$& 29 & 8 & & \\
$\!\! (2,5)$& 61 & 10 & & \\
$\!\! (3,3)$& 2 & 19 & 6 & \\
$\!\! (3,4)$& 41 & 53 & 8 &\\
$\!\! (3,5)$& 106 & 81 & 10 &\\
$\!\! (4,4)$& 2 & 50 & 45 & 8 \\
$\!\! (4,5)$& 71 & 256 & 101 & 10 \\
\hline
\end{tabular}
\caption{\label{table:EDSylvester}
Weighted ED degrees for Sylvester matrices ${\rm Syl}_k(F,G)$}
\end{table}
\bigskip
\medskip
\noindent
{\bf Acknowledgements.}\\
We thank the following colleagues for their help with this project:
Jean-Charles Faug\`ere, William Rey,
Ragni Piene, Jose Rodriguez,
Mohab Safey El Din, \'Eric
Schost, and Thomas Schultz.
Giorgio Ottaviani is a member of GNSAGA-INDAM.
Pierre-Jean Spaenlehauer and Bernd Sturmfels were hosted by
the Max-Planck Institute f\"ur Mathematik in Bonn, Germany.
Bernd Sturmfels was also supported
by the NSF (DMS-0968882).
|
\section{Introduction}
The history of the discovery of extrasolar planets is a story of challenges for theories of planet formation.
The first extrasolar planet orbiting a main sequence star\footnote{The first detection of two terrestrial-mass
exoplanets around a pulsar PSR B1257+12 was announced in 1992 \citep{Wolszczan-92}.}%
, 51 Peg b \citep{Mayor-95}, proved to be very much at odds with the theory of formation of
our own solar system (SS), our only reference at the time. As the number of planets increased and
several close-in (hot-) Jupiters were found, it became clear that a new formation mechanism was necessary.
An in situ formation of these planets was unlikely, because of the insufficient disk mass close to the star.
However, their presence at an orbit of $\sim$0.05 AU could be reconciled with a farther out formation by invoking migration
during or after the formation process \citep[e.g.][]{Lin-96}. The SS formation theories suggested that giant planets preferentially formed close to
the ice-line (especially in low-metallicity disks), where water is condensed into ice and the necessary building blocks for the formation of planets
could be found in large quantities \citep[e.g.][]{Ida-08}.
An important breakthrough was thus to explain the existence of hot-Jupiters by considering core-accretion
simultaneously with disk-driven migration \citep[e.g.][]{Pollack-96, Alibert-05, Mordasini-09a}
while reproducing several observational trends \citep{Mordasini-09b}.
It explained the correlation of the presence of giant planets with stellar metallicity
\citep[e.g.][]{Santos-01, Santos-04, Fischer-05, Sousa-11}, a correlation that could not be explained by the
alternative formation method, gravitational instability \citep[e.g.][]{Boss-98}.
However, when the formation paradigm seemed to be complete, an unexpected result emerged. The discovery of
hot-Jupiters whose orbital plane was misaligned with the stellar rotation axis
\citep[e.g.][]{Hebrard-08, Triaud-10, Brown-12} cast serious doubts on disk-driven migration as the mechanism responsible for
the hot-Jupiters. In particular, the occurrence of retrograde planets required an additional mechanism. Two main solutions put
forward were Kozai cycles plus tidal perturbations \citep[e.g.][]{Wu-03,Fabrycky-07,Correia-11} and planet-planet scattering
\citep[e.g.][]{Rasio-96,Beauge-12}. The implication was that planets first form in the disk at several AU,
then undergo gravitational perturbations that increase the eccentricity to very high values, and finally reduce their semi-major axis by tidal interactions
with the star that simultaneously dampen the eccentricity to zero by conservation of the orbital angular momentum\footnote{
However, there are several other mechanisms that can tilt a star relative to its protoplanetary disk: gravitational torques from massive distant bodies
\citep{Batygin-12} or angular momentum transport within the host star \citep{Rogers-12}.}.%
To study the main mechanism responsible for the presence of close-in Jupiters, \cite{Socrates-12} and \cite{Dawson-12} analyzed the eccentricity distribution of
proto-hot Jupiters. The paucity of super-eccentric proto-hot Jupiters observed in the Kepler sample allowed the latter authors to conclude that disk migration is
the dominant mechanism that produces hot-Jupiters, although some of these planets might be perturbed to high-eccentricity orbits by interactions with planetary companions.
It is interesting to note how the signatures of the formation and evolution mechanisms relate to the mass of
the planet. There is an ongoing debate on whether the core-accretion model can reproduce the properties of Earth-mass
planets \citep[][]{Fortier-13}. Recent works suggested that unlike for more massive planets, these systems might have been formed in situ
\citep[e.g.][]{Hansen-12, Chiang-13}. While these types of studies are still in their infancy, they already reproduce
some of the most common features of the low-mass planet systems, such as being dynamically
packed and showing low inclination and eccentricities \citep{Lovis-11, Figueira-12}. On the other hand, planets
with masses between 0.05$M_{Jup}$ and 20$M_{Jup}$ have been characterized by core-accretion and migration formation,
and evolution models and can be compared quite well with the observations.
Evidence for the migration processes involved in the formation of short-period planets has previously been discussed in the literature.
For instance, several works suggested discontinuities in observables inside this mass range.
\cite{Beauge-13} presented tentative evidence that the smallest ($<4R_{\oplus}$) Kepler planetary candidates
orbiting metal-poor stars show a period dependence (see Sect. 2); \cite{Dawson-13} argued that giant planets orbiting
metal-rich stars show signatures of planet interaction. The first result provides compelling evidence for
the importance of migration and accretion, the second for planet-planet interaction. For the upper end of the mass population,
composed of planets with masses above 4$M_{Jup}$, a long discussion has been ongoing on whether
this population resembles more closely lower-mass planets or shares properties with stellar binaries
\citep[e.g.][]{Udry-02, Halbwachs-05}. The eccentricity distribution of this planetary population has
been shown to be similar to that of binaries, and some doubts were cast on core-accretion as their
formation mechanism \citep[][]{Ribas-07}.
In this work we explore the relations and correlations between planetary mass and orbital parameters,
namely period and eccentricity, and how they relate to metallicity, for a wide range of planetary masses.
We start by extending the work of \cite{Beauge-13} for higher planetary masses in Sect. 2, and in Sect. 3
we explore the properties of planets with masses higher than 4$M_{Jup}$. We summarize our main results and their implications
on our understanding of planet formation in Sect. 4.
We note that in this work the masses of the planets (M$_{P}$) are indeed minimum masses, so that in some cases the ``true`` planet mass
may be significantly higher. Nevertheless, statistical analyses show that the distribution of M$_{P}$ sin \textit{i} values is similar to that of M$_{P}$
values (e.g. \cite{Jorissen-01}, but see the discussion by \cite{Lopez-12} for low-mass planets).
In fact, the average factor of overestimation is only 1.27, assuming a random distribution of the inclination.
We would like to say a word of caution regarding the significance estimates quoted in the next sections.
In general, if one uses the same data to test several hypotheses,
the results can be affected by the multiple-testing problem\footnote{See
\texttt{ http://en.wikipedia.org/wiki/Multiple\_comparisons}}
. However, since throughout our study we dealt with only a few parameters and the significance
levels of the results are very high, it can be shown by simple algebra that the current proposed trends are significant,
although we acknowledge the possibility that the quoted significance levels might be optimistic.
\section{Period-mass diagram and metallicity}
Recently, \cite{Beauge-13} analyzed the distribution of planets in the orbital period (P) versus planetary radius (R) diagram and P versus M$_{P}$ diagram.
They found a lack of small-size/low-mass planets (R $\lesssim$ 4$R_{\oplus}$, M$_{P}$ $\lesssim$ 0.05$M_{Jup}$) with periods P $<$ 5 days around metal-poor
stars. They also found a paucity of sub-Jovian size/mass planets around metal-poor stars with periods shorter than 100 days.
They explained these observed trends with a delayed formation and less planetary migration in metal-poor disks.
Interestingly, the authors found no significant correlation between metallicity and the position of giant planets in the (P,R) or (P,M$_{P}$) diagram.
In this section we extend the analysis of these authors to higher-masses using a large sample of FGK dwarf stars ($M_{*} > 0.5 M_{\odot}$)
with stellar parameters derived in a homogeneous way \citep[SWEET-Cat: a catalog of stellar parameters for stars with planets:][]{Santos-13}\footnote{
\texttt{https://www.astro.up.pt/resources/sweet-cat}}.
\begin{figure*}
\begin{center}
\begin{tabular}{cc}
\includegraphics[angle=0,width=0.5\linewidth]{per_mass4.ps}
\includegraphics[angle=0,width=0.5\linewidth]{per_mass_all.ps}
\end{tabular}
\end{center}
\caption{Planet distribution in the (P, M$_{P}$) diagram. In the top-left panel only the planets around evolved stars with logg $\leq$ 4 dex are shown.
In the bottom-left panel we plot only the planets detected with the transiting method, and in the right panel we present the planets orbiting FGK dwarfs (logg $>$ 4 dex).
The dotted line represents the approximate (empirical) detection limit for planets, and the dashed line is the linear fit for the full sample.}
\label{fig_p_m_feh}
\end{figure*}
Figure~\ref{fig_p_m_feh} shows the (P,M$_{P}$) diagram for planets detected with different techniques that orbit stars with different evolutionary states.
In the plot we separate the planets according to the metallicity of their host stars.
As one can see in the $bottom-left$ panel of the figure, most of the transiting planets have very short periods and their hosts are confined to a narrow range of
metallicity (their hosts are mostly metal-rich). It is also very clear that planets around evolved stars have a very small range of P and their hosts
show very different metallicity distributions compared with their unevolved counterparts (for example, about 40\% of planets around evolved stars
have [Fe/H] $<$ -0.1 dex, while fewer than 20\% of planets orbiting dwarf stars have similarly low metallicity).
This difference in metallicity distribution is probably due to selection biases in evolved stellar samples that are used for planets searches \citep{Mortier-13}.
We conclude therefore that these samples have such strong selection and detection biases that they are not suitable for our analysis.
We therefore concentrate the discussion on dwarf stars with planets detected by radial velocity when we discuss the impact of the planet host's metallicity.
The right panel of Fig.~\ref{fig_p_m_feh} shows RV-detected planets orbiting FGK dwarfs on the P-M$_{P}$ plane in which we separate low- and high-metallicity hosts
at a threshold of -0.1 dex, which is the average metallicity of the stars in the solar neighborhood \citep[e.g.][]{Adibekyan-12c}.
The same diagram, but for planets with masses lower than 0.1$M_{Jup}$ is shown in the $top$ panel of Fig.~\ref{fig_p_m_low}.
As one can see, for a fixed planetary mass, most of the planets around metal-poor stars are constrained to longer periods.
This tendency appears to be valid from about 0.03$M_{Jup}$ to about 4$M_{Jup}$ and confirms and extends the
results of \cite{Beauge-13} to higher masses.
To evaluate the statistical significance of the observed trend we performed a simple Monte Carlo (MC) test.
First, we counted the number of planets orbiting metal-poor stars in a given interval of mass from 0.03$M_{Jup}$ to $M_{P}$(max), and we
randomly drew the same number of points from the planetary sample around metal-rich stars.
Then we counted the number of planets orbiting metal-poor and metal-rich stars below the line of the linear fitting of the full data (P, M$_{P}$ of
all the planets in our initial sample, i.e., metal-poor and metal-rich planets with masses from 0.0036 to 17.4$M_{Jup}$).
We repeated the entire process 10$^5$ times. By comparing the average number (and using the standard deviation) of metal-rich planets found below
the fitted line with the number of metal-poor planets (again bellow the fitted line), we evaluated the significance (the z-score - n$\sigma$)
of the statement
that the two metal-poor and metal-rich populations have different distributions of (P,M$_{P}$)\footnote{
We note that the quoted significance levels are the result of a one-sided test, since the test by construction is one-sided.}%
.
The $M_{P}$(max) originally was 1$M_{Jup}$, which we later varied up to the highest value of the planet mass in the sample.
The metallicity boundary was also considered as variable (-0.10, -0.05, and 0.00 dex).
We found that the significance of the correlation decreases with the increase of $M_{P}$(max) and
interestingly, after $M_{P}$(max) $=$ 4$M_{Jup}$, the significance decreased faster.
From the $right$ panel of Fig.~\ref{fig_p_m_feh} one can see that almost all the planets with masses higher than 4$M_{Jup}$ orbiting metal-poor stars
lie below the dashed line. Depending on the combination of the parameters ([Fe/H] and $M_{P}$(max) $\leq$ 4$M_{Jup}$), the
significance of our results varies from \{4.4$\sigma$ to 6.1$\sigma$, which means that it is always significant.
To ensure that the observed trend is genuinely valid for the higher-mass domain and is not just influenced by the previously reported trend for low-mass
planets \citep{Beauge-13},
we repeated the test, selecting $M_{P}$(min) in the range from 0.1 to 0.3 Jupiter mass and excluding the planets with M$_{P}$ $>$ 4$M_{Jup}$.
We confirm the previous result at a 1.8$\sigma$ to 3.7$\sigma$ significance level (depending
on the combination of fixed parameters, i.e., [Fe/H], $M_{P}$(min)): the significance was lowest when $M_{P}$(min) is 0.3 Jupiter mass and
the $M_{P}$(max) is 1$M_{Jup}$ (when the number of planets in the subsample is the smallest).
Furthermore we made a similar test, but instead of counting the planets below the fitted line, we considered a horizontal line (constant period).
Varying this horizontal line from about 80 to 400 days, we again found that the result is always higher than 2.1$\sigma$ (reaching up to 3.2$\sigma$).
These tests show that the results do not depend significantly on the position of the separation (fitted line or constant period), and they suggest that the
findings by \cite{Beauge-13} about the lack of small-size planets around metal-poor stars at short periods extend up to 4 Jupiter masses.
\subsection{Earth-like planets}
All the planets with masses bellow 0.03$M_{Jup}$ (about 10$M_{\oplus}$) orbiting metal-rich stars have short periods of fewer than 18 days
(see $top$ panel of Fig. ~\ref{fig_p_m_low}). It is difficult to understand why these low-mass planets do not have longer periods, such
as we observe for planets in the metal-poor systems.
One of the reasons for this disposition can have a dynamical character. Most of the planets orbiting metal-poor stars belong to the multiple
systems where almost all the planets have low masses (M$_{P} <$ 0.03$M_{Jup}$), while planets in the multiple systems orbiting metal-rich stars have
higher-mass planetary companions with longer periods. However, in the $bottom$ panel of Fig. ~\ref{fig_p_m_low}, where we show the distribution of the planets
with the longest period in the system, one can see that there are six planets orbiting metal-rich stars without a (detected) higher-mass longer-period companion.
This means that at least for the planets presented in the plot, the shorter periods are not a result of interaction with higher-mass longer-period companions.
Another explanation of the lack of metal-rich low-mass planets with long periods might be a detection limit, since metal-rich planet hosts are on average slightly
more massive and hotter than their metal-poor counterparts. However, the mentioned differences are very small and probably cannot be responsible for the
observed short periods.
If there is no detection bias in the sample, then the observed distributions of metal-poor and metal-rich planets in the P-M$_{P}$ diagram could
mean that Earth-like planets orbiting metal-rich stars preferably migrate or form close to their parent stars, while planets in the metal-poor systems
form at a wider range of the semi-major axis or do not always migrate.
In the small range of periods (fewer than 18 days) where metal-rich planets orbit, there are also similar planets orbiting metal-poor stars,
and their period distribution seems similar (excluding the planets with longer periods orbiting metal-poor stars), although it is difficult
to evaluate the significance because of the small number of planets.
We note that the number of planets discussed in this subsection is small, therefore, the results and conclusions regarding them should be considered
with caution.
\begin{figure}
\begin{center}
\includegraphics[angle=0,width=1\linewidth]{per_mass_low_2.ps}
\end{center}
\caption{Same as Fig.~\ref{fig_p_m_feh} right panel, but only for planets with M$_{P}$ $<$ 0.1$M_{Jup}$ ($top-panel$). The (P, M$_{P}$) diagram for the
longest-period planet in the system with M$_{P} <$ 0.03M$_{Jup}$ ($bottom-panel$).}
\label{fig_p_m_low}
\end{figure}
\subsection{[Ref/H] vs. [Fe/H]}
Typically, [Fe/H] is used as a proxy of overall metallicity for stars, but if a star has peculiar chemistry or is enhanced by some individual elements
compared to iron, then the [Fe/H] index will differ from the total metallicity [M/H]. \cite{Adibekyan-12a,Adibekyan-12b} showed that most of the
planet-host stars with low-iron content are enhanced by $\alpha$-elements, including Mg and Si which are fairly abundant and have condensation
temperatures similar to iron \citep{Lodders-03,Lodders-09}. \cite{Beauge-13} recently questioned how their result would be affected
if one used the [Ref/H] index\footnote{This index was proposed by \cite{Gonzalez-09} and quantifies the mass abundances of refractory elements (Mg, Si and Fe)
relevant for planet formation.}
instead of [Fe/H].
To quantify the effect of [Ref/H] on our result we used the HARPS sample of 135 planet-host stars with detailed and precise chemical abundances derived in
\cite{Adibekyan-12c}. In this sample 32 planet hosts have [Fe/H] lower than -0.1 dex. We found that in the mentioned metallicity region
the [Ref/H] is on average higher than [Fe/H] by $\approx$ 0.05 dex: about 0.00 dex for stars with [Fe/H] $\approx$ -0.1 dex and up to 0.15 dex for the stars with
[Fe/H] $\approx$ -0.6 dex.
Only two out of the mentioned 32 planet hosts with [Fe/H] $<$ -0.1 dex show [Ref/H] $>$ -0.1 dex, but both cases are very close to the boundary
(-0.09 and -0.08 dex).
Recalling that our results do not depend on the metallicity boundary, and assuming that the HARPS planet host sample is representative for the whole sample
discussed above, we can conclude that using the [Ref/H] index instead of [Fe/H] will not affect our results.
\subsection{Role of [Fe/H] on the position of planets in the P-M$_{P}$ diagram. Discussion}
The most important insights from our tests and results come from the continuity or discontinuity of the extrasolar
planet population as a function of key parameters. The correlation between period, mass, and metallicity presented in
Fig.~\ref{fig_p_m_feh} suggests the existence of a mechanism that affects a wide range of masses, from 10$M_{\oplus}$ to 4$M_{Jup}$,
and that depends on metallicity. \cite{Dawson-13} showed that high-eccentricity Jupiter-mass planets with semi-major axes ($a$) between 0.1 and 1 AU were
preferably found orbiting metal-rich stars; the authors interpreted this finding as evidence for planet-planet
scattering on a population created by smooth migration.
The main assumption was that [Fe/H] will not significantly affect the
P-M$_{P}$ relationship as created by core-accretion+migration \citep[see][for details]{Mordasini-12}, and
thus the observed difference was created exclusively by dynamical interactions.
They also noted that beyond 1 AU, the metal-rich and metal-poor samples have a similar eccentricity distribution with the
explanation that planets with $a >$ 1AU may have formed where we observe them.
We applied a Kolmogorov-Smirnov (K-S) test for the samples of giant planets (M$_{P}$ $>$ 0.1$M_{Jup}$ and with P $>$ 10 days) orbiting metal-poor ([Fe/H] $<$ -0.1 dex)
and metal-rich ([Fe/H] $\geq$ -0.1 dex) dwarf stars and found that their eccentricity distributions are similar (K-S probability $\approx$ 0.63).
To test whether these distributions are similar for different period regimes we made two subsamples of planets
with periods 10 $<$ P $\leq$ 300 days (300 days corresponds to $\sim$1 AU, which was the upper limit of $a$ in \cite{Dawson-13}),
and with P $>$ 300 days. Since the number of planets orbiting around
metal-poor stars with [Fe/H] $\geq$ -0.1 dex is small (only seven stars), we changed the metallicity boundary from -0.1 to 0.0 dex. We note that this was also
the metallicity boundary chosen by \cite{Dawson-13}. The applied K-S test
predicts $P_{KS}$ = 0.10 and $P_{KS}$ = 0.97 that metal-poor and metal-rich populations of planets with short and long periods have the same eccentricity
distribution, respectively. Thus, we confirm the results of \cite{Dawson-13} for ''close-in`` giants. We note that for the whole range of periods (P $>$ 10 days)
the K-S statistics delivers $P_{KS} \approx$ 0.66 for the similarity of eccentricity distribution of planets orbiting stars with [Fe/H] $<$ 0.0 dex and
[Fe/H] $\geq$ 0.0 dex.
The test implies that the P-M$_{P}$ distribution of the metal-poor population (at least for planets with periods longer than 300 days)
results from formation and migration and has no dynamical character.
The fact that massive planets at low-[Fe/H] are found farther out implies that these planets probably are only able to form beyond the
ice-line for low-metallicities \citep{Mordasini-10, Mordasini-12}\footnote{This would in turn imply that migration does not significantly change
the semi-major axis of low-metallicity planets. This last point seems to be in line with what is expected from the models,
but depends on the definition of the ice-line, among other aspects.}.
The similarity of the period distribution of short-period Earth-like planets (M$_{P} < 10M_{\oplus}$) around metal-rich and metal-poor stars probably means that
enough low-mass planets can be formed at small semi-major axes even at low metallicities because a sufficient amount of protoplanetary mass is locally available.
At the same time, the fact that planets orbiting metal-rich stars are observed only at short periods might imply that these planets form close to their central stars
or ordinarily migrate.
However, we note that the statistical significance of the last results are not evaluated because of small number of planets.
It is also interesting to note that if we compare metal-poor stars with their metal-rich counterparts, there is a
dearth of planets with mass around 0.1$M_{Jup}$ orbiting the metal-poor stars \citep[see also][]{Beauge-13}.
In the paradigm of \cite{Dawson-13}, these Neptune-mass planets
are preferentially formed farther out, (again probably around or beyond the ice-line), and move into smaller orbits
because of planet interaction. Interestingly, this is qualitatively in line with results from microlensing surveys,
which state that cold Neptunes are indeed common \citep{Sumi-10, Gaudi-12}.
However, a quantitative analysis would require an assessment and correction for detection bias, which is far beyond the scope of this paper.
\begin{figure}
\begin{center}
\includegraphics[angle=270,width=1\linewidth]{per_mass_ecc_rv.ps}
\end{center}
\caption{(P,M$_{P}$) diagram for the all RV detected planets with M$_{P}$ $>$ 1$M_{Jup}$ orbiting FGK stars from \texttt{exoplanet.eu}.
Planets with high- and low-eccentricity orbits are presented by crosses and circles, respectively.}
\label{fig_p_m_ecc}
\end{figure}
\section{Orbital and physical properties of very massive planets}
In the previous section we noted that the planets with very high masses show different dependencies on metallicity in the (P,M$_{P}$) diagram from
their less massive counterparts. This may be related both to formation and to post-formation processes (interaction with disk or interaction between planets).
To form a very high planetary mass, a critical core must grow, a process that takes longer in a low-metallicity environment \citep{Mordasini-12}.
The time scale of the orbital evolution of planets also depends on the properties of the protoplanetary disk and on other initial conditions
\citep[e.g.][]{Xiang-Gruess-13, Bitsch-13}. To determine whether this difference between high- and low-mass planets in the (P,M$_{P}$,[Fe/H]) diagram is reflected
on other parameters as well, we compared the physical and orbital parameters of planets with masses from 1 to 4$M_{Jup}$ to those with masses
higher than 4$M_{Jup}$\footnote{
We chose to establish a low-mass cut-off because one can assume that at least the mechanism of formation of the planets in this two planetary mass regimes
is similar.}. For this analysis we used all the RV-detected planets orbiting FGK stars from the Extrasolar Planets Encyclopaedia \citep{Schneider-11}\footnote{
http://www.exoplanet.eu}.
\subsection{Orbital eccentricity}
The first significant correlation we found is that very massive planets are more eccentric than their lower-mass counterparts.
This correlation is illustrated in Fig.~\ref{fig_p_m_ecc}. Sixty-one out of 90 planets (66.3$\pm$5.0\%) with M$_{P}$ $>$ 4$M_{Jup}$ have $e$ $>$ 0.2, while only 94 out of 173
(54.3$\pm$3.8\%) planets with masses between 1 and 4$M_{Jup}$ are similarly eccentric. To assess the statistical significance of the difference
in observed eccentricity distributions between low- and high mass planets we performed a K-S test. The K-S statistics predict a 0.0026 ($\approx$3$\sigma$)
probability ($P_{KS}$) that the two subsamples come from the same underlying distribution of $e$. We note that when we consider only the planets orbiting FGK dwarf stars
the difference remains significant; 80.4$\pm$5.8 and 62.1$\pm$4.6 percent of eccentric planets with masses higher than 4$M_{Jup}$ and with
1$M_{Jup}$ $<$ M$_{P}$ $<$ 4$M_{Jup}$, respectively. To test whether our result is affected by close-in planets, which mostly have circular
orbits due to tide effects \citep[see e.g.][]{Ford-06}, we established a minimum period cut-off at 10, 50, and 100 days (although from Fig.~\ref{fig_p_m_ecc} it is
clear that most of the RV-detected planets have P $>$ 100 days). The K-S probabilities for the three cases
were 0.004, 0.0036, and 0.0022 respectively, which allows us to conclude firmly that this is not the case.
\cite{Udry-02} and \cite{Ribas-07} had already found a marginal tendency for low-mass planets (M$_{P}$ $<$ 4$M_{Jup}$) to be less eccentric than more massive
planets and binary stars.
We note that in these two studies the authors did not separate very small planets (Earth-like or Neptune-like) from the massive gaseous planets\footnote{
It is worth to note, that if we compare eccentricities of low- and high-mass planets considering all the planets with M $<$ 4$M_{Jup}$ and M $>$ 4$M_{Jup}$
then the $P_{KS}$ decreases to 3$\times10^{-5}$.}.
However, \cite{Udry-02} noted that when one restricts the sample to periods longer than 50 days (avoiding the circularization through tidal interactions), the
difference in eccentricities between low- and high-mass planets disappears.
It has been suggested that the eccentricity of planets can be increased through the planet-disk interaction under favorable conditions
and especially if the planetary mass is very high \citep{Papaloizou-01, Kley-06, D'Angelo-06}. Very recently, \cite{Bitsch-13}, performing isothermal 3D simulations,
showed that the eccentricity of planets in single systems with masses between 1 and 5$M_{Jup}$ is generally damped due to planet-disk interaction, while
for very massive planets with masses above $\sim$ 5$M_{Jup}$ the eccentricity can increase for low orbital inclinations relative to the disk.
At the same time, N-body simulations of multiple giant planets performed by \cite{Raymond-10} and population synthesis models by \cite{Ida-13}, and before that
by \cite{Thommes-08}, showed that
the eccentricity increases with planetary mass. The latter authors explained this with a scenario in which multiple giant planets are mainly formed in relatively
massive disks where dynamical instabilities, cohesive collisions, and orbit crossings are more common and can result in excitation of higher eccentricities.
Summarizing, \cite{Bitsch-13} predicted higher eccentricities for very high mass planets due to the interaction with disk (note that they performed
their simulation for single-planet systems), and \cite{Ida-13} predicted correlation between $e$ and M$_{P}$ for a wider range of planetary mass because of
close scatterings and interaction of gas giants. We again applied a K-S statistics to test whether the eccentricity distribution of planets with
masses between 1 to 5$M_{Jup}$ and planets with M$_{P} >$ 5$M_{Jup}$ are similar or not, but now we considered only planets in single systems.
Our test predicts $P_{KS} \approx$ 0.004 that the two populations have the same underlying eccentricity distribution. We note that in the last test
we considered only planets with P $>$ 10 days. If one consider planets with P $>$ 50 or P $>$ 100 days, or changes the M$_{P}$ boundary to 4$M_{Jup}$, the
$P_{KS}$ remains always smaller than 0.005 throughout.
Based on the observed trends, we can conclude that in general the eccentricities of high-mass planets, even in single-planet systems (although there is
a possibility that in the systems there are more planets that are undetectable by current instrumentation and surveys or that some planets have been ejected), are
higher than those for lower-mass planets, as predicted from the numerical simulation \citep{Papaloizou-01, Bitsch-13}.
Very recently, high-resolution near-infrared observation of HD100546 showed a variable ro-vibration of CO and OH emission lines
\citep{Liskowsky-12, Brittain-13}, which the authors explained by postulating the presence of an eccentric massive planetary companion. This is an direct
observational support of our results obtained statistically for a large sample.
\subsection{Orbital period}
The second correlation we observe is that very massive planets with eccentric orbits have longer periods than those with more circular orbits with $e <$ 0.2
(see Fig.~\ref{fig_p_m_ecc}).
The K-S statistics predict 0.6\% probability that the two families of planets have the same period distribution.
This difference in periods is probably related to the interaction and migration processes in the disk that the planets suffered.
\cite{Cumming-04} suggested that planets with long periods are on average easier to detect if their orbits are eccentric. However, our result cannot
be explained with this observational selection effect because there is no similar correlation observed for lower-mass planets.
The K-S statistics do not report a significant difference in period for planets with eccentric and circular orbits and
with masses 1$M_{Jup}$ $<$ M$_{P}$ $<$ 4$M_{Jup}$.
The $P_{KS}$ is 0.15 that low- and high-eccentricity planets with a mass in this interval have the same period distribution .
This probability even increases from 0.15 to 0.76 when restricting the sample to periods longer than 10 to 100 days.
Since the period distribution of planets with low (1$M_{Jup}$ $<$ M$_{P}$ $<$ 4$M_{Jup}$) and high mass is similar ($P_{KS} \approx 0.65$ for planets with
P $> 10$ days), one could expect a more prominent correlation between eccentricity and period for lower-mass planets if there is a selection bias.
This allows us to conclude that the correlation between $e$ and P for very high mass planets has a physical meaning, and does not come
from an observational bias.
\subsection{Metallicity}
Interestingly, \cite{Ribas-07} found that the metallicity of planet host stars decreases with planet mass. In particular, they showed that the average metallicity
of stars with planets of mass lower than 4$M_{Jup}$ is different from that of higher mass planet hosts at 3-$\sigma$ level, which is lower by about 0.15 dex.
We again note that in their comparison
they did not separate super-Earths and Neptunes, and in their sample almost all of the very low mass planets (M$_{P}$ $<$ 0.3$M_{Jup}$) have [Fe/H] $>$ 0,
while the majority of recently detected low-mass planets have subsolar or solar metallicities \citep[e.g.][]{Mayor-11, Buchhave-12}.
Figure~\ref{fig_mass_feh} shows the metallicities of FGK dwarf stars hosting RV-detected planets as a function of the planet mass. The metallicities were
taken from the SWEET-Cat. On average, the metallicities of high-mass planets are similar. The average values and the standard error
of the means are presented in the Table 1.
\begin{figure}
\begin{center}
\includegraphics[angle=0,width=1\linewidth]{mass_feh.ps}
\end{center}
\caption{Planetary mass against metallicty a) for all the planets in the system and b) for the most massive planets in each planetary systems.
All the planets are detected with the radial velocity technique and planet hosts are FGK dwarf stars.
The small circles correspond to individual planets and the large circles represent the average metallicty for the four mass bins
M$_{P} < 0.1M_{Jup}$; $0.1M_{Jup} \leq M_{P} < 1M_{Jup}$; $1M_{Jup} \leq M_{P} < 4M_{Jup}$; and M$_{P}$ $>$ 4$M_{Jup}$. The error bars indicate
the standard deviation of the [Fe/H] in each bin.}
\label{fig_mass_feh}
\end{figure}
We tested the results obtained by \cite{Ribas-07} by applying a K-S test for the RV-detected sample of planets with masses between 1 and 4$M_{Jup}$ and
with M$_{P}$ $>$ 4$M_{Jup}$. We obtained $P_{KS}$ $\approx$ 5.5\% that the two subsamples have the same underlying metallicity distribution.
If we consider only the planet with highest mass in the system, the probability increases to 15.3\%. We note that
the metallicities of these stars were taken from the literature, are not homogeneous, and include stars with different evolutionary stages.
We again applied the same statistics, but taking only RV-detected planet-host FGK dwarf stars for which SWEET-Cat provides homogeneous stellar parameters,
including metallicity. In this case the difference in metallicity distribution between the two subsamples becomes even less significant with $P_{KS}$ $\approx$ 15\%
and $\approx$ 34.5\% if one consider only the planet with the highest mass in each system.
We therefore conclude that a simple separation in mass at 4$M_{Jup}$ does not reveal two different populations in metallicity,
as hypothesized in \cite{Ribas-07}. This result is expected from the core-accretion theory, since the
metallicity mostly acts as a threshold for giant planet formation, but is not correlated with the mass of giant planets, except for planets with
M$_{P}$ $\gtrsim$ 10-20$M_{Jup}$, which are harder to form at clearly subsolar metallicities \citep{Mordasini-12}.
\begin{table}
\centering
\caption{Average and standard errors of [Fe/H] for planets with different masses: taking into account all the planets in the system (a) and
only the most massive planet in the systems (b).}
\label{ks}
\begin{tabular}{lcc}
\hline
\hline
Planet mass & $<$[Fe/H]$>_a$ & $<$[Fe/H]$>_b$\tabularnewline
$M_{Jup}$ & dex & dex\tabularnewline
\hline
M$_{P} < 0.1$ & -0.061$\pm$0.025 & -0.082$\pm$0.036\tabularnewline
$0.1 \leq M_{P} < 1$ & 0.088$\pm$0.023 & 0.096$\pm$0.025\tabularnewline
$1 \leq M_{P} < 4$ & 0.149$\pm$0.016 & 0.146$\pm$0.018\tabularnewline
M$_{P}$ $>$ 4 & 0.083$\pm$0.032 & 0.084$\pm$0.033\tabularnewline
\hline
\end{tabular}
\end{table}
\section{Summary}
We analyzed the possible relations between orbital and physical properties of planets with different mass and physical
properties of their host stars. The main findings and results are itemized below .
\begin{itemize}
\item
We found that for a fixed maximum planetary mass between 10M$_{\oplus}$ and 4$M_{Jup}$, planets orbiting metal-poor stars are constrained
to longer periods than those orbiting metal-rich stars. Applied MC tests show that the obtained correlation is statistically significant.
This result suggests that the mechanism responsible for the ''separation'' of planets in the P-M$_{P}$ is operational for a wide range of planetary mass.
The observed dependence can be explained by assuming that planets in a metal-poor disk form farther out from their central stars and/or do not migrate
as far as planets in metal-rich systems because they form later.
Our result confirms the \cite{Beauge-13} findings and extends their conclusions to higher planetary masses.
\item
The Earth-like planets (M$_{P} < 10M_{\oplus}$) orbiting metal-rich stars have shorter periods than those orbiting metal-poor stars.
If there is no detection bias, this could imply that the low-mass planets in metal-rich systems ordinarily migrate or that they always form close
to their parent stars.
The presence of Earth-like planets with short-periods orbiting metal-poor stars probably means that these planets can be formed close to
their parent stars even at low metallicities because enough amount of protoplanetary mass is locally available.
\item
By applying a K-S test for a sample of FGK RV-detected dwarf planet-hosts we obtained $P_{KS}$ up to 35\% that the low- and high-mass
planet hosts have the same underlying metallicity distribution. This result is expected in the core-accretion models \citep[e.g.][]{Mordasini-12},
but it contradicts the results of \cite{Ribas-07}, who found that the metallicity of planet-host stars decrease with planet mass.
\item
Analyzing the eccentricity distribution of high-mass planets, we found statistically significant evidence
that planets with masses higher than 4$M_{Jup}$ have on average more highly eccentric orbits than the giant planets with masses between 1$M_{Jup}$ and 4$M_{Jup}$.
These trends and dependencies agree with the core-accretion and formation in a disk models and allow us to conclude that
planet-disk interaction is a very important and orbit-shaping mechanism when one moves towards higher planetary masses.
In addition to these observational results, we found that less eccentric very high mass planets (M$_{P}$ $>$ 4$M_{Jup}$) have shorter periods than those with
similar mass that orbit their host stars with more highly eccentric orbits.
This difference in periods is probably related to the interaction and migration processes in the disk that the planets underwent.
\end{itemize}
The dependencies and trends presented in this work can provide new constraints for the models and numerical simulations of planet formation and evolution.
In particular the fact that almost all the giant planets orbiting metal-poor stars show long periods (i.e. longer than 100 days) shows that
migration is less rapid than assumed in core-accretion planet formation models \citep[e.g.][]{Mordasini-12}.
\begin{acknowledgements}
{This work was supported by the European Research Council/European Community under the FP7 through Starting Grant agreement
number 239953. V.Zh.A., S.G.S., E.D.M., and M.O. are supported by grants SFRH/BPD/70574/2010,
SFRH/BPD/47611/2008, SFRH/BPD/76606/2011, and SFRH/BD/51981/2012 from the FCT (Portugal), respectively.
A.C. is supported by grant PEst-C/CTM/LA0025/2011 from the FCT.
C.M. acknowledges the support of the MPG through the Reimar-L\"ust Fellowship.
G.I. acknowledges financial support from the Spanish Ministry project MINECO AYA2011-29060.
We gratefully acknowledge the anonymous referee for the constructive comments and suggestions, and Astrid Peter for
the help concerning English.}
\end{acknowledgements}
|
\section{INTRODUCTION}
\subsection{Vehicular platooning}
The field of \emph{vehicular platooning} was active as early as the 1960's and remains so until now. The task is to safely and effectively control several vehicles driving behind each other, for example on a highway lane. It is motivated by higher throughput, lower fuel consumption, increase of traffic safety etc.
Regarding control strategies; among the first treatments of vehicular platooning were papers by \cite{Levine1966} and \cite{Melzer1971}. They examined a centralized control approach with a single global controller governing all vehicles. However, \cite{Jovanovic2005a} later showed that one has to be careful about the stabilizability of the system, since it might degrade with increasing number of vehicles. Nevertheless, more attention is paid to fully or partially distributed control, wherein each vehicle is controlled by its own on-board controller with only limited knowledge about the platoon. Among the first papers dealing with the distributed control was work by \cite{chu_decentralized_1974}. Basic questions about the feasibility and performance of such systems was introduced by \cite{Cosgriff1969} and later formalized by \cite{Swaroop1996} under the term \emph{string stability}. String stability, or more precisely string instability, is a~phenomenon that causes higher control demands on the members of a~vehicular platoon that are further from the source of regulation error. Although string stability does not guarantee that the vehicles do not crash into each other, it is a useful analysis tool. A way how a regulation error or a disturbance propagates in a platoon of vehicles controlled by various distributed control strategies was examined in several papers, see for instance \cite{Seiler2004a}, \cite{Barooah2005a} and \cite{Shaw2007}. A fundamental limitation of many distributed algorithms with only local information about the platoon is inability to maintain coherence in a~large-scale platoon subjected to stochastic disturbances \cite{Bamieh2012b}. Though, the coherence can be improved by introducing optimal non-symmetric localized feedback \cite{Lin2011}.
A common goal of each platooning algorithm is to drive the platoon with a reference velocity and inter-vehicle distances. Many distributed algorithms have been introduced in the platooning field. The most simple algorithm relying only on the measurement of the distance to the immediately preceding vehicle is the so-called \emph{predecessor following algorithm}. A straightforward extension is the so-called \emph{bidirectional control algorithm}, which additionally measures the distance to the immediate follower. Depending on the weight between these two distance measurements, we distinguish either \emph{symmetric} or \emph{asymmetric} bidirectional control. Although, the asymmetric version improves the stability in terms of the least stable closed-loop eigenvalue as proved by \cite{Barooah2009}, we let our in-platoon vehicles to be controlled by the symmetric version, analysed for instance in \cite{Lestas2007} or \cite{Middleton2010}. The reason for doing this will be clear after Section~\ref{sec:reflections}.
\subsection{Wave-based control concept}
The origins of the control based on travelling waves lies in the 1960's in mathematical modeling and analysis of flexible structures. Paper of \cite{vaughan_application_1968} was one of the first treatments analysing simpler instances of flexible structures such as beams and plates. Analysis and control of a more complex flexible structures from the viewpoint of travelling wave-modes was investigated in a series of papers by von Flotow and his colleagues in \cite{Flotow1985} and \cite{Flotow1986}.
Recently, the concept was revisited in a series of papers by O'Connor in \cite{OConnor2006} and \cite{OConnor2007} for vibrationless positioning of lumped multi-link flexible mechanical systems. It was named \emph{wave-based control} and it is based on the so-called \emph{wave transfer function}, which describes how the traveling wave propagates in the lumped system. Simultaneously with O'Connor, the wave concept was also revisited for a control of continuous flexible structures by \cite{Halevi2005} under the name \emph{absolute vibration suppression}. It relies on the transfer function as well, though in this case, the time delay plays a key role. Surprisingly, it was shown by the joint paper of the last two mentioned authors in \cite{Peled2012}, that both the wave-based control and the absolute vibration suppression are just a feedback version of the input shaping control. It was also shown that the wave-based control can be generalized even for continuous flexible systems, e.g. a steel rod, and then it coincides with the absolute vibration suppression.
The key idea of the wave-based control is to generate a wave at the actuated front end of the interconnected system and let it propagate to the opposite end of the system, where it reflects and returns back to the front-end actuator. When it reaches the front again, it is absorbed by the front-end actuator by means of the wave transfer function. A both interesting and troublesome property of the wave transfer function is the presence of the square root of polynomial in the function. This makes its implementation in the time domain very challenging. To be able to run numerical simulations, we therefore introduce a convergent recursive algorithm that approximates the wave transfer function for an arbitrary dynamics of the local system.
There are other viewpoints on the wave-based control. One was introduced by \cite{Ojima2001} in terms of the characteristic impedance for a mass-spring system. Other possible viewpoint introduced \cite{Nagase2005} for wave control of ladder electric networks.
\subsection{Objective of the paper}
In this paper, a finite one-dimensional platoon of vehicles moving in a highway lane is considered. Each individual vehicle in the platoon is locally controlled by a bidirectional controller, which plays the role of string-damper connection in mechanical structures and hence enables a wave to propagate back and forth. One or both of the platoon ends are controlled by the \emph{wave-absorbing controller} allowing active absorption of the traveling wave. The similarity of bidirectional control with continuous wave equation was described in \cite{Herman2013}
The key objective of the paper is to generalize the principle of the wave-based control used in the field of mechanics for vehicular platooning control in such a way that the distances between vehicles are additionally considered. In this regard, the presented concept offers a symmetric version of bidirectional control enhanced by the feedback control of one or both platoon ends. Thus, it significantly decreases long transient oscillations during platoon manoeuvres such as acceleration/deceleration or changing the distances between vehicles. In addition, the paper contributes the following: a) It generalizes the wave transfer function description for the arbitrary dynamics of the local system, b) it offers the convergent recursive algorithm that approximates the wave transfer function, c) it presents an alternative way of deriving the wave transfer function using a continued fraction approach, and d) it provides a mathematical derivation of the transfer functions describing reflections on the platoon ends.
The paper is structured as follows. Section 2 gives a mathematical model of the vehicle. Section 3 describes the wave transfer function as a requisite tool for the wave description. A mathematical description of wave reflections on forced and free ends is given in Section 4. Section 5 introduces the wave-absorbing controller as an addon for the bidirectional control. The new controllers are analyzed by numerical simulations in section 6. The necessary mathematical derivations are given in the three appendices.
\section{LOCAL CONTROL OF THE PLATOON VEHICLES}
A vehicle in a platoon indexed by $n$ is in the Laplace domain modelled as
\begin{align}
X_{n}(s) = P(s) U_n(s),
\label{eq:system_car}
\end{align}
where $s$ is the Laplace variable, $X_n(s)$ is a position of the $n$th vehicle in the Laplace domain, $P(s)$ represents the transfer function of system dynamics and $U_n(s)$ is the system input which is generated by the local controller of the vehicle specified in the following.
Except for the leader indexed $n=0$ and the rear-end vehicle, each vehicle in the platoon is equipped with a symmetric bidirectional controller $C(s)$ with the task of equalizing the distances to its immediate predecessor and successor, giving
\begin{align}
U_n(s) = C(s)(D_{n-1}(s) - D_{n}(s)),
\label{eq:orig_diff_eq_long}
\end{align}
where $D_n(s)$ is the distance between vehicles indexed by $n$ and $n+1$, hence $D_n(s)= X_{n}(s) - X_{n+1}(s)$. Substituting (\ref{eq:orig_diff_eq_long}) into (\ref{eq:system_car}) yields the resulting model of the in-platoon vehicle with the bidirectional control for the inter-vehicle distances,
\begin{align}
X_{n}(s) = P(s)C(s) (X_{n-1}(s)-2X_{n}(s)+X_{n+1}(s)).
\label{eq:orig_diff_eq_simp}
\end{align}
Using the notation,
\begin{equation}
\alpha(s) = \frac{1}{P(s)C(s)}+2,
\label{eq:alpha_generalized}
\end{equation}
equation (\ref{eq:orig_diff_eq_simp}) is thus rewritten as
\begin{align}
X_{n}(s) &= \frac{1}{\alpha(s)}(X_{n-1}+X_{n+1}).
\label{eq:bidir_sys_laplace}
\end{align}
The vehicle at the rear end of the platoon is driven by the predecessor following algorithm and is supposed to equalize the distance to its immediate predecessor and reference distance $D_{\text{ref}}$,
\begin{align}
X_{N}(s) = \frac{1}{\alpha(s)-1}(X_{N-1}(s)-D_{\text{ref}}(s)),
\label{eq:last_sys_laplace}
\end{align}
where $X_{N}(s)$ is the position of the last vehicle in the platoon.
To carry out numerical simulations, we will use the model that is often used in theoretical studies. The vehicle is described by a double integrator model with a simple (linear) model of friction, $\xi$, and controlled by a PI controller. Hence, $P(s) = 1/(s^2+\xi s)$ and $C(s) = (k_{\text{p}} s +k_{\text{i}})(s)$, where $k_{\text{p}}$ and $k_{\text{i}}$ are proportional and integral gains of the PI controller, respectively. Such a model was also used in experimental studies in \cite{Martinec2012}.
\section{WAVE TRANSFER FUNCTION}
The bidirectional property of locally controlled systems causes any change in the movement of the leading vehicle to propagate through the platoon as a \emph{wave} up to the last vehicle. To describe this wave, we need to find out how the position of a vehicle is influenced by the position of its immediate neighbours. For a moment, let us assume that the length of the platoon is infinite, so that there is no platoon end where the wave can reflect. A generalization for platoon with a one platoon end, i.e. a semi-infinite platoon, is done in the next section.
\subsection{Mathematical model of the wave transfer function}
Following the standard arguments for \emph{wave equation} found for instance in \cite{Asmar2004}, the solution to the wave equation can be decomposed into two components: $A_{n}(s)$ and $B_n(s)$ (also called \emph{wave variables} in the literature), which represent two waves propagating along a platoon in the forward and backward directions, respectively.
To find a transfer function describing the wave propagation, we are searching for two linearly independent recurrence relations that satisfy (\ref{eq:bidir_sys_laplace}). We first recursively apply (\ref{eq:bidir_sys_laplace}) and (\ref{eq:last_sys_laplace}) with $D_{\text{ref}}(s) = 0$, for a platoon with an increasing number of vehicles. The transfer function for a platoon with two vehicles is $A_1 / A_0 = (\alpha-1)^{-1}$, for a platoon with three vehicles is $A_1/A_0 = \left(\alpha-(\alpha-1)^{-1} \right)^{-1}$, for a platoon with four vehicles is $A_1/A_0 = \left(\alpha - \left( \alpha-(\alpha-1)^{-1} \right)^{-1} \right)^{-1}$ and so on. Continuing recursively, $A_1/A_0$ is expressed by the continued fraction
\begin{equation}
\dfrac{A_1}{A_0} = \dfrac{1}{\alpha-\dfrac{1}{\alpha-\dfrac{1}{\alpha-\dfrac{1}{\ddots}}}}.
\label{eq:trans_continued_fraction}
\end{equation}
The continued-fraction expansion of a square root is given by \cite{jones1984}
\begin{equation}
\sqrt{z^2+y} = z+\dfrac{y}{2z+\dfrac{y}{2z+\dfrac{y}{2z+\dfrac{y}{\ddots}}}}.
\label{eq:sqrt_continued_fraction}
\end{equation}
Letting the number of vehicles approach infinity, the right-hand sides of (\ref{eq:trans_continued_fraction}) and (\ref{eq:sqrt_continued_fraction}) are equal, provided that $y=-1$ and $z=\alpha/2$. Hence,
\begin{equation}
\frac{A_{1}}{A_0} = \frac{\alpha}{2}-\frac{1}{2}\sqrt{\alpha^2-4}.
\label{eq:G1_continued_fraction}
\end{equation}
Likewise, the transfer function $A_2/A_1$ can be expressed from (\ref{eq:bidir_sys_laplace}) and (\ref{eq:last_sys_laplace}) for $n = 2$ as
\begin{align}
\alpha A_1 &= A_0+A_2
\end{align}
Substituting for $A_0$ from the previous recursive step (\ref{eq:G1_continued_fraction}) gives
\begin{align}
\alpha A_1 &= A_1\left(\frac{\alpha}{2}+\frac{1}{2}\sqrt{\alpha^2-4}\right) + A_2,
\end{align}
which provides
\begin{align}
\frac{A_2}{A_1} &= \frac{\alpha}{2}-\frac{1}{2}\sqrt{\alpha^2-4}.
\label{eq:G1_continued_fraction2}
\end{align}
Continuing recursively, we can find that the transfer function $A_{n+1}/A_{n}$ is again equal to (\ref{eq:G1_continued_fraction}) or (\ref{eq:G1_continued_fraction2}). We can conclude that the transfer function from the $n$th to $(n+1)$th vehicle is the same for each vehicle, and is equal to
\begin{equation}
G_1(s) = \frac{\alpha}{2}-\frac{1}{2}\sqrt{\alpha^2-4}. \label{eq:G1_WTF}
\end{equation}
Analogously, the second linearly independent recurrence relation of (\ref{eq:bidir_sys_laplace}) and (\ref{eq:last_sys_laplace}) is searched for by their recursive application with a decreasing index of vehicles. After similar algebraic manipulations as for $A_n$, we find
\begin{equation}
\dfrac{B_n}{B_{n-1}} = \alpha- \dfrac{1}{\alpha-\dfrac{1}{\alpha-\dfrac{1}{\alpha-\dfrac{1}{\ddots}}}}.
\label{eq:trans_continued_fraction_B}
\end{equation}
Letting the number of vehicles approach infinity, the right-hand sides of (\ref{eq:trans_continued_fraction_B}) and (\ref{eq:sqrt_continued_fraction}) are equal provided that $y=-1$ and $z=\alpha/2$. Hence,
\begin{equation}
\frac{B_n}{B_{n-1}} = \frac{\alpha}{2}+\frac{1}{2}\sqrt{\alpha^2-4}.
\label{eq:G2_continued_fraction}
\end{equation}
The transfer function from $n$th to $(n-1)$th vehicle is the same for each vehicle, and is equal to
\begin{equation}
G_2(s) = \frac{\alpha}{2}+\frac{1}{2}\sqrt{\alpha^2-4}. \label{eq:G2_WTF}
\end{equation}
The resulting model of the vehicular platoon with an infinite number of vehicles is therefore described as follows:
\begin{align}
X_n &= A_n + B_n,
\label{eq:pos_decomp}\\
A_{n+1} &= G_1A_n,
\label{eq:anp1}\\
B_{n} &= G_2B_{n-1},
\label{eq:bnp1}\\
G_1 &= G_2^{-1},
\label{eq:Ginv}
\end{align}
where (\ref{eq:Ginv}) follows from the multiplication of (\ref{eq:G1_WTF}) and (\ref{eq:G2_WTF}). Equations (\ref{eq:anp1})-(\ref{eq:bnp1}) express the \emph{rheological property} of the platoon, that is, they define the form of how these two components propagate through the platoon. Equation (\ref{eq:Ginv}) expresses the \emph{principle of reciprocity}, that is, if $A(s)$ propagates with the help of $G_1(s)$ to higher indexes of vehicles, then $B(s)$ propagates with the help of $G_1(s)$ to lower indexes of vehicles. The function $G_1(s)$ is hereafter referred to as the wave transfer function.
It should be noted that if there is a boundary in the system, e.g., if the length of platoon is finite, where the rheology property for wave propagation changes abruptly, the principles must be supplemented by boundary conditions. We discuss this case in the following section.
\subsection{Verification of the wave transfer function}
We now outline an alternative way to derive the wave transfer function. Let the model of the vehicular platoon (\ref{eq:pos_decomp})-(\ref{eq:Ginv}) hold and now search for the transfer functions $G_1(s)$ and $G_2(s)$ that satisfy these four equations. Substituting (\ref{eq:pos_decomp}) into (\ref{eq:bidir_sys_laplace}) yields
\begin{align}
\alpha(A_n+B_n) = A_{n-1}+B_{n-1} + A_{n+1}+B_{n+1},
\end{align}
which, in view of (\ref{eq:anp1}) and (\ref{eq:bnp1}), is
\begin{align}
\alpha(s) = G_1(s) + G_2(s).
\end{align}
We can substitute either for $G_{1}(s)$ or $G_{2}(s)$ from (\ref{eq:Ginv}). Either possibility leads to the same quadratic equation ($m=1,2$),
\begin{align}
G_m^2(s) - \alpha(s)G_m(s) +1=0,
\label{eq:quadratic_equation}
\end{align}
with two linearly independent solutions,
\begin{align}
G_{m}(s) = \frac{\alpha}{2} \mp \frac{1}{2}\sqrt{\alpha^2-4}.
\label{eq:transferG_alpha}
\end{align}
Let $G_1(s)$ be chosen as the solution with the negative sign in front of the square root. Then (\ref{eq:Ginv}) only allows $G_2(s)$ to be the solution with the positive sign in front of the square root. Hence, $G_1(s)$ and $G_2(s)$ are identical to those derived in the previous section. The quadratic equation (\ref{eq:quadratic_equation}) can be employed as a starting model for the positioning of multi-link flexible mechanical systems \cite{OConnor2006}.
\subsection{Approximation of the wave transfer function}
It will be shown later in the paper that to be able to implement the wave-absorbing controller advertised at the beginning of the paper, we need to find the impulse response of the wave transfer function, i.e. the inverse Laplace transform of $G_1(s)$. Due to the presence of the square root in the function it is very challenging to find exact impulse response of $G_1(s)$. However, we can approximate the impulse response with a finite impulse response (FIR) filter. Therefore, we first approximate the wave transfer function in the Laplace domain, then transform this approximate form to the time domain and finally truncate and sample the approximate impulse response to obtain FIR filter coefficients.
The square root function in (\ref{eq:transferG_alpha}) can be approximated by various ways, e.g., Newton's method, the binomial theorem, or continued fraction expansion (\ref{eq:trans_continued_fraction}). We employ the last option since it guarantees the convergence of iterative approximations and is applicable to an arbitrary dynamics of the local system with a generalized parameter $\alpha(s)$ as in (\ref{eq:alpha_generalized}). The recursive formula (\ref{eq:trans_continued_fraction}) immediately provides the iterative approximation of $G_1(s)$,
\begin{equation}
G_1^l(s) = \frac{1}{\alpha(s)-G_1^{l-1}(s)},
\label{eq:G_recursive_approx}
\end{equation}
where $l = 1,2,\ldots $, and the initial value $G_1^0(s) = 1$. The approximate $G_1^l(s)$ can be transformed to the time domain by Matlab or Mathematica. Our experience with the inverse Laplace solvers for the Fractional Calculus \emph{invlap} \cite{dehoog1982}, \emph{weeks} \cite{Weeks1966} and \emph{nilt} \cite{brancik1999} in Matlab is that, while they were not capable of performing the inverse Laplace transform of (\ref{eq:transferG_alpha}) due to the square root function, they carried out the inverse Laplace transform of $G_{1}^l(s)$ without complications since (\ref{eq:G_recursive_approx}) is~a rational function.
The approximate $G_1^l(s)$ can interpreted as follows. Equation (\ref{eq:G_recursive_approx}) represents the transfer function from the position of the leader to the position of the first follower in a platoon of $l$ vehicles. Increasing the number of iterations (\ref{eq:G_recursive_approx}) means that the length of a platoon grows and the effect of the rear-end vehicle on $G_1(s)$ weakens. The approximation of $G_1(s)$ therefore successively improves. Figs. \ref{fig:bode_comparison_only_exact} and \ref{fig:impulse_comparison_only_exact} show the Bode characteristics $G_{1}^l(s)$ and the associated impulse responses for various number of iterations, respectively. Increasing the numbers of iterations makes the peak in the Bode characteristic sharper, more localized and moves it towards lower frequencies, eventually disappearing entirely. The basic characteristic of the impulse response is fitted after a few iterations while small differences occur at longer times. To obtain the FIR filter coefficients, we truncate the approximate impulse response at a few seconds and sample it with an appropriate frequency. In our numerical simulations it was sufficient to stop the iterative procedure after $20$ iterations, to truncate the impulse response at $15$ seconds and sample it at a frequency of $100\,\text{Hz}$.
\begin{figure}[htb
\centering
\includegraphics[width=0.48\textwidth]{G1_iters_bodemag.pdf}
\caption{The Bode characteristics of $G_{1}(s)$ approximations after several iterations by (\ref{eq:G_recursive_approx}) for $k_{\text{p}}=k_{\text{i}}=\xi=4$.}
\label{fig:bode_comparison_only_exact}
\end{figure}
\begin{figure}[htb
\centering
\includegraphics[width=0.48\textwidth]{impulse_comparison_CF.pdf}
\caption{The impulse responses of $G_{1}(s)$ after several iterations by (\ref{eq:G_recursive_approx}) for $k_{\text{p}}=k_{\text{i}}=\xi=4$.}
\label{fig:impulse_comparison_only_exact}
\end{figure}
\section{REFLECTION OF THE WAVE ON PLATOON ENDS}
\label{sec:reflections}
To be able to design a wave-absorbing controller for the platoon end, we first need to mathematically describe the wave reflection.
In the previous section an infinite platoon is considered, whereas here we assume a semi-infinite platoon having one end that is either externally controlled (forced end) or allowed to move freely (free end). When a wave propagates along a platoon and reaches its free end, it is reflected with the same polarity, i.e., the same sign of amplitude, but with the opposite polarity at the fixed/forced end. This phenomenon, known from basic wave physics \cite{french2003}, is discussed in the following in terms of the wave transfer function. The necessary mathematical derivations are given in Appendix A and B.
\subsection{The forced-end boundary}
\label{sec:forced_end_boundary}
We call the forced-end boundary such a vehicle that is externally controlled and is not to the other vehicles. However, the neighbouring vehicle is one-directionally linked with this forced boundary. The platoon leader therefore represents the forced-end boundary.
\begin{figure}[htb
\centering
\includegraphics[width=0.49\textwidth]{platoon_fixed_end.pdf}
\caption{Scheme of wave reflection on the leader, i.e. reflection on the forced-end boundary, described by (\ref{eq:refl_fixed_end}).}
\label{fig:fixed_end}
\end{figure}
The reflection on the forced-end boundary is sketched in Fig. \ref{fig:fixed_end}. Changing the position of the forced end, $X_0$, generates the outgoing wave as a first contribution to $A_1$. Moreover, the incoming wave ($B_1$) is reflected on the forced end and transformed to the outgoing wave as the second contribution to $A_1$. The force-end reflection is derived in \ref{sec:app_forced_end} and summarized by (\ref{eq:forced_end_appendix}),
\begin{equation}
A_1 = G_1 X_{0}-G_1^2B_1.
\label{eq:refl_fixed_end}
\end{equation}
This first shows that changing the position of the forced end is translated to $A_{1}$ through $G_{1}$. Second, since the DC gain of $G_1$ is equal to plus one (see Fig. \ref{fig:bode_comparison_only_exact}), the minus sign in front of $G_1^2$ causes the wave to be reflected with the opposite sign.
\subsection{The free-end boundary}
\label{sec:free_end_boundary}
A free-end boundary is a boundary where a vehicle is two-directionally linked with one neighbour only and, additionally, it is aware about steady state of the link. The rear-end vehicle described by (\ref{eq:last_sys_laplace}) represents the free-end boundary.
\begin{figure}[ht]
\centering
\includegraphics[width=0.49\textwidth]{platoon_free_end.pdf}
\caption{Scheme of wave reflection on the rear-end vehicle, i.e. reflection on the free-end boundary, described by (\ref{eq:refl_free_end}).}
\label{fig:free_end}
\end{figure}
The reflection on the free-end boundary is outlined in Fig. \ref{fig:free_end}. The wave travelling from the free-end boundary ($B_N$) is composed of two parts, the incoming wave ($A_N$) which is reflected back through $G_1$ and the component due to adjusting the reference distance $D_{\text{ref}}$. The free-end reflection is derived in \ref{sec:app_free_end} and summarized by (\ref{eq:free_end_appendix}),
\begin{align}
B_N = G_1 A_N + \frac{G_1 -1}{\alpha-2} D_{\text{ref}}.
\label{eq:refl_free_end}
\end{align}
The reflection from the free-end boundary does not change the sign that is expressed by the plus sign in front of $G_1 A_N$. Moreover, the signal reflected from the free-end is delayed as a linear function of $G_{1}(s)$, while as a quadratic function when it is reflected from the forced-end boundary, as shown by (\ref{eq:refl_fixed_end}).
It should be noted that the verification of the above wave-based model was done in \cite{OConnor2007}. The transfer function
\begin{equation}
\frac{X_N}{X_0} = G_1^N \frac{1+G_1}{1+G_1^{2N+1}},
\end{equation}
where $N$ is index of the last vehicle, was shown to be identical to the transfer function derived by the state space description. This result is valid not only for a double integrator with P controller, but for an arbitrary dynamics of the local system.
\section{WAVE-ABSORBING CONTROLLER}
\label{sec:wave-based_platoon_control}
The three main control requirements are: i) to travel the platoon at reference velocity $v_{\text{ref}}$, ii) to keep inter-vehicle distances $d_{\text{ref}}$, iii) to actively absorb the wave travelling towards the platoon's end.
This section introduces three possible configurations of the platoon with the wave-absorbing controller. First, we will describe the configuration where the wave-absorbing controller is implemented at the platoon leader.
\subsection{Front-sided wave-absorbing controller}
\label{sec:driving_ref_vel}
\subsubsection{Absorption of the wave}
\label{subsec:wave_absorption}
To absorb the incoming wave at the platoon front, the transfer function from $B_1$ to $A_1$ in (\ref{eq:refl_fixed_end}) has to be equal to zero. In other words, we are searching for $X_0$ to satisfy the equation $G_1 X_0/B_1 - G_1^2 = 0$. The only solution is
\begin{equation}
X_0 = G_1B_1.
\label{eq:leader_absorb}
\end{equation}
To be consistent with the model (\ref{eq:pos_decomp})-(\ref{eq:Ginv}), we denote $B_0 = G_1B_1$ and $A_0 = X_0-B_0$, then (\ref{eq:refl_fixed_end}) is expressed as $A_1 = G_1X_0-G_1B_0 = G_1 A_0$. Summarizing this yields the wave components of the leader
\begin{align}
B_0 &= G_1X_1-G_1^2A_0, \label{eq:B0_leader}\\
A_0 &= X_0-B_0.
\end{align}
This means that if one component of the position of the leader is equal to $B_0$, then the leader absorbs the incoming wave. We can imagine that if the leader is pushed/pulled by its followers, thus it manoeuvres like one of the in-platoon vehicles.
\subsubsection{Acceleration to the reference velocity}
\label{subsec:acceleration_platoon}
The previous algorithm actively absorbs the incoming wave to the platoon leader. To change the platoon's velocity and inter-vehicle distances are other tasks that need to be solved.
To accelerate the platoon, we need to add an external/reference input, $X_{\text{ref}}$, for the leader. This changes (\ref{eq:leader_absorb}) to $X_0 = B_0+X_{\text{ref}}$. The rear-end vehicle represents the free-end boundary, therefore, $B_0$ is expressed by the combination of (\ref{eq:anp1}), (\ref{eq:bnp1}), (\ref{eq:refl_fixed_end}) and (\ref{eq:B0_leader}) as $B_0 = G_1^{2N+1}X_{\text{ref}}$. This leads the transfer function from $X_{\text{ref}}$ to $X_0$ to be
\begin{equation}
\frac{X_{\text{0}}}{X_{\text{ref}}} = 1+G_1^{2N+1}.
\label{eq:leader_accel_trans}
\end{equation}
Fig. \ref{fig:bode_comparison_only_exact} showed that the DC gain of $G_1$ is equal to one, therefore, the DC gain of ($1+G_1^{2N+1}$) is equal to two. This means that to accelerate the platoon to reference velocity $v_{\text{ref}}$, the leader has to be commanded to accelerate to velocity $v_{\text{ref}}/2$ at the beginning of the manoeuver, as shown in Fig. \ref{fig:velocity_N10}.
\begin{figure*}[htb
\centering
\includegraphics[width=0.99\textwidth]{components_comp_N10.pdf}
\caption{Simulation of the velocity wave propagating in the platoon with the Front-sided wave-absorbing controller at several time instances. At the beginning, $t=0\,{\text{s}}$, all platoon vehicles are standing still except for the leader which accelerates to a velocity $0.5\,\text{ms}^{-1}$. At intermediate times, the wave travels to the rear vehicle, where it is reflected and travels back to the leader to be completely absorbed. By propagating, it forces platoon vehicles to accelerate by another $0.5\,\text{ms}^{-1}$ to a velocity $1\,\text{ms}^{-1}$. At the final stage, $t=30\,{\text{s}}$, the leader is the last one reaching the velocity $1\,\text{ms}^{-1}$ and the whole platoon moves with $1\,\text{ms}^{-1}$. The red crosses represent the derivation of $A+B$ positional components computed by the wave transfer function approach, the green plus signs are the velocities simulated by the Matlab Simulink.}
\label{fig:velocity_N10}
\end{figure*}
Fig. \ref{fig:velocity_N10} additionally shows an independent validation of the wave transfer function approach. The derivation of the sum of $A+B$ velocity components (red crosses) of the wave travelling through the platoon are compared against the velocities simulated by the Matlab Simulink (green plus signs). We can see an agreement between the wave-transfer-function-derived and independently-simulated velocities.
\subsubsection{Changing of the inter-vehicle distances}
\label{subsec:changing_distances}
Increasing the inter-vehicle distances poses a more difficult task than merely accelerating the platoon. The reason is that the rear-end vehicle reacts to the change of reference distance $d_{\text{ref}}$ by acceleration/deceleration. This creates a velocity wave propagating towards the leader who absorbs it by changing its velocity. This means, however, that when all vehicles reach the desired inter-vehicle distance $d_{\text{ref}}$, the whole platoon travels with a new velocity different from the original. Only by an additional action of the leader, see the next paragraph, will the original velocity be reestablished.
Although the platoon has a finite number of vehicles, it behaves like a semi-infinite platoon because no wave reflects from the platoon leader, who is equipped with the wave absorber. Since (\ref{eq:refl_free_end}) holds for a semi-infinite platoon, it can be now used to determine the transfer function from $D_{\text{ref}}$ to velocity of the leader, $V_0(s)$, that is
\begin{equation}
\frac{V_0}{D_{\text{ref}}} = G_1^N \frac{s(G_1-1)}{\alpha-2}.
\label{eq:tf_Dref_V0}
\end{equation}
The DC gain of (\ref{eq:tf_Dref_V0}) reads as
\begin{equation}
\kappa_{\text{f}} = \lim_{s\rightarrow 0} \left(G_1^N \frac{s(G_1-1)}{\alpha-2} \right).
\end{equation}
In the case where the reference distance is changed and the leader does not accelerate, the velocity of the platoon changes by $(\kappa_{\text{f}} d_{\text{ref}})$. This means that the platoon slows down or even moves backwards. To compensate for this undesirable velocity change, the leader is commanded to accelerate to the velocity
$(-\kappa_{\text{f}} d_{\text{ref}})/2$. The platoon will consequently travel with the original velocity, hence compensating for the acceleration/deceleration of the rear-end vehicle.
The DC gain of (\ref{eq:tf_Dref_V0}) for the PI controller case is equal to $(-\sqrt{k_{\text{i}}/\xi})$.
\subsubsection{Overall control of the leader}
Let us now assume that the leader has a positional controller with input $X_{\text{f}}$. Summarizing preceding subsections yields the resulting control law of the leader,
\begin{align}
X_{\text{f}}(s) &= X_{\text{ref}}(s) + B_0(s),
\label{eq:external_input_leader}
\end{align}
From the above discussion, $X_{\text{ref}}(s)$ must be represented by a ramp signal with slope $w_{0}$,
\begin{equation}
w_{0} = \frac{1}{2}\left(v_{\text{ref}}-\kappa_{\text{f}} d_{\text{ref}}\right),
\label{eq:ramp_leader}
\end{equation}
to ensure that the platoon travels with a reference velocity $v_{\text{ref}}$ and inter-vehicle distances $d_{\text{ref}}$. In case of the PI controller, $w_0 = \left(v_{\text{ref}} +\sqrt{k_{\text{i}}/\xi} d_{\text{ref}}\right)/2$. The Front-sided wave-absorbing controller is summarized in Fig. \ref{fig:platoon_actuators_one}.
\begin{figure*}[htb
\centering
\includegraphics[width=0.8\textwidth]{platoon_actuators_one6.pdf}
\caption{Scheme of the Front-sided wave-absorbing vehicular platoon controller.}
\label{fig:platoon_actuators_one}
\end{figure*}
\subsection{Rear-sided wave-absorbing controller}
\label{sec:rear_sided_WBC}
Instead of placing the wave-absorbing controller at the platoon's front, it can be placed at the platoon's rear. In this case, the platoon has one leader in the front and one wave-absorbing controller at the rear. However, the absence of the predecessor follower in the platoon has an important consequence. Any velocity change of the leader, $V_{0}(s)$, causes a change in the distance to the first follower, $D_1(s)$, as shown in (\ref{eq:forced_end_dist_vel}). Consequently, all other distances between vehicles are changed. This negative effect is to be compensated by an acceleration/deceleration of the rear-end vehicle. We denote $\kappa_{\text{r}}$ to be the DC gain of the transfer function from $V_0(s)$ to $D_1(s)$.
Having specified the DC gain, a certain reference signal needs to be sent to the platoon end to set up a desired inter-vehicle distance $d_{\text{ref}}$. The input to the positional controller of the rear-end vehicle, $X_{\text{r}}(s)$, is expressed, analogous to (\ref{eq:external_input_leader}), as
\begin{align}
X_{\text{r}}(s) &= X_{\text{ref,rear}}(s) + G_1(s)A_{N-1}(s),
\label{eq:external_input_last}
\end{align}
where $X_{\text{ref,rear}}(s)$ is a reference ramp signal with slope $w_{r}$,
\begin{equation}
w_{r} = \frac{1}{2}\left(v_{\text{ref}}-\kappa_{\text{r}} d_{\text{ref}}\right).
\label{eq:ramp_last}
\end{equation}
In other words, the platoon leader drives the platoon to travel with velocity $v_{\text{ref}}$, while the rear-end vehicle makes the platoon travel with inter-vehicle distances $d_{\text{ref}}$. For the PI controller case $\kappa_{\text{r}} = k_{\text{i}}/\xi$.
\subsection{Two-sided wave-absorbing controller}
\label{sec:two_sided_WBC}
The Front-sided and Rear-sided wave-absorbing controllers can be combined by implementing wave absorbers to both the platoon leader and the rear-end vehicle. In this case, no wave is reflected back from neither of platoon ends.
The input to the positional controller of the leader is given by (\ref{eq:external_input_leader}) with the ramp signal (\ref{eq:ramp_leader}), while the input to the positional controller of the rear-end vehicle is (\ref{eq:external_input_last}) with the ramp signal (\ref{eq:ramp_last}). In this way, each platoon end generates a velocity wave propagating towards the opposite end. Likewise, as for the Front-sided and Rear-sided wave-absorbing controllers (Section \ref{sec:driving_ref_vel} and \ref{sec:rear_sided_WBC}), the amplitudes of the two waves are summed up to $v_{\text{ref}}$, meaning that the platoon travels with velocity $v_{\text{ref}}$ and inter-vehicle distances $d_{\text{ref}}$.
\subsection{Asymptotic and string stability}
Using the same technique as in \cite{Herman2013}, it can be shown that a platoon with the symmetric bidirectional controller is asymptotically stable. Since $G_1(s)$ can be represented by such a platoon, it is asymptotically stable as well. The truncated approximate of $g_1(t)$ is BIBO (bounded-input bounded-output) stable, which is a well known fact about FIR filters. Therefore, a platoon with the wave-absorbing controller on one or both platoon ends remains asymptotically stable.
We follow the $L_2$ string stability definition from \cite{Eyre1998a} that can be formulated as:\emph{ The system is called $L_2$ string stable if there is an upper bound on the $L_2$-induced system norm of $T_{0,n}$ that does not depend on the number of vehicles, where $T_{0,n}$ is the transfer function from position of the leader to the position of the vehicle indexed $n$.}
In the case of the platoon with the Front-sided wave-absorbing controller, the position of the $n$th vehicle is described as
\begin{equation}
X_{n} = (G_1^{n}+G_1^{2N+1-n}) X_{0}.
\end{equation}
Due to the triangle inequality and the fact that $||G_1||_{\infty} \leq 1$, which is shown in \ref{app:string_stability}, we obtain
\begin{equation}
||G_1^{n}+G_1^{2N+1-n}||_{\infty} \leq ||G_1^{n}||_{\infty}+||G_1^{2N+1-n}||_{\infty} \leq 2.
\end{equation}
This means that the magnitude of the maximum peak in the frequency response of the transfer function from the position of the leader to the position of the $n$th vehicle is smaller or equal to $2$. Since the $L_2$-induced norm and $H_{\infty}$ coincide, we can state that the platoon with the Front-sided wave-absorbing controller is $L_2$ string stable.
The position of the $n$th vehicle with an absorber placed at the rear-end vehicle is
\begin{equation}
X_{n} = G_1^{n}X_0 + (G_1^{N-n}-G_1^{N+n})X_{N}.
\end{equation}
We apply the same idea and state that $H_{\infty}$ norm of both $G_1^{n}$ and $(G_1^{N-n}-G_1^{N+n})$ are bounded regardless of the number of vehicles. Therefore, the platoon with the Rear-sided wave-absorbing control is $L_2$ string stable.
The position of the $n$th vehicle in a platoon with absorbers on both ends is expressed as
\begin{equation}
X_n = G_1^n X_{0} +G_1^{N-n}X_{N},
\end{equation}
which immediately shows that the platoon with the Two-sided wave-absorbing controller is $L_2$ string stable as well.
\section{NUMERICAL SIMULATIONS}
We consider the linear friction of our system to be $\xi = 4$ and search for the parameters of the PI controller such that oscillations of the impulse response of $G_1(s)$ are minimized. The parameters $k_{\text{p}} = k_{\text{i}} = 4$ satisfy this requirement. All numerical simulations are run for a platoon of $50$ vehicles to demonstrate that the wave-absorbing controllers are capable of controlling large platoons.
To demonstrate the advantages of the wave-absorbing controllers, we will compare their performance against a pure bidirectional control without any wave-absorbing controller. This means that the leader travels with a constant velocity $v_{\text{ref}}$ for the whole time of the simulation. Fig. \ref{fig:accel_of_platoon_noWBC} shows outcomes of numerical simulation when the leader without wave-absorbing controller increases its velocity. We can see significant limitations of the bidirectional control. The oscillatory behaviour in the movement of the platoon is caused by numerous wave reflections from both platoon ends. Eventually, the platoon settles at a desired velocity after many velocity oscillations. These oscillations not only significantly prolong the settling time, but they could lead to accidents within the platoon.
\begin{figure*}[!htbp]
\centering
\includegraphics[width=0.99\textwidth]{platoon_noWBC2.pdf}
\caption{Simulation of the platoon without the wave-absorbing controller when the leader accelerates to velocity $v_{\text{ref}} = 1\,\text{ms}^{-1}$. The reference distance is kept fixed, $d_{\text{ref}} = 1\,\text{m}$, for the whole time.}
\label{fig:accel_of_platoon_noWBC}
\end{figure*}
The performance of the Front-sided wave-absorbing controller during two platoon manoeuvrers is shown in Fig. \ref{fig:accel_of_platoon_WBC}. In the first $150\,\text{s}$ manoeuver, the platoon accelerates (not necessarily from zero velocity) to reach a desired velocity. In comparison with the pure bidirectional control, see Fig. \ref{fig:accel_of_platoon_noWBC}, the settling time is now significantly shorter. Moreover, under some circumstances, it can be guaranteed that vehicles do not crash into each other during the platoon acceleration. In fact, the distances between vehicles are increased at the beginning of the acceleration as suggested by (\ref{eq:forced_end_dist_vel}) and shown in the middle panel of the Fig. \ref{fig:accel_of_platoon_WBC}. However, the distances may undershoot the initial inter-vehicles distances in the second part of the acceleration manoeuver. If the impulse response of the wave transfer function is tuned such that it does not undershoot the zero value, then the distances between vehicles can not become less than the initial inter-vehicle distances. In the opposite case (not shown here), where the platoon travels with a constant velocity and starts to decelerate, the distances between vehicles are temporarily decreased and a collision may occur.
\begin{figure*}[!htbp]
\centering
\includegraphics[width=0.99\textwidth]{platoon_WBC3.pdf}
\caption{Simulation of two platoon manoeuvrers with the Front-sided wave-absorbing controller. At the beginning, the vehicles are standing still separated by one meter. For the first manoeuver, the platoon is commanded to accelerate to $v_{\text{ref}}=1\,\text{ms}^{-1}$ with $d_{\text{ref}} = 1\,\text{m}$ starting at time $t=0\,\text{s}$. At time $t = 150\,\text{s}$, the platoon is commanded to perform the second manoeuver such that the reference distance is increased to $d_{\text{ref}} = 1.5\,\text{m}$ without changing the reference velocity.}
\label{fig:accel_of_platoon_WBC}
\end{figure*}
At time $t = 150\,\text{s}$ in Fig. \ref{fig:accel_of_platoon_WBC}, the platoon is commanded to perform the second manoeuver such that the reference distance is increased, but the reference velocity is kept unchanged. The rear-end vehicle reacts to this command at the same time as the leader since it is controlled by the reference distance that is now changing. However, the end vehicles differ in action; the leader accelerates, while the rear-end vehicle decelerates. This behaviour creates an undesirable overshoot in distances.
A numerical simulation of the two manoeuvrers for the platoon controlled by the Rear-sided wave-absorbing controller is shown in Fig. \ref{fig:rear_sided_WBC}. During the acceleration manoeuver the inter-vehicle distances between vehicles closer to the rear end are temporarily decreased while those for vehicles near the leader are temporarily increased. During the changing-distance manoeuver, on the other hand, no overshoot in distances occurs.
\begin{figure*}[!htbp]
\centering
\includegraphics[width=0.99\textwidth]{rear_sided_WBC.pdf}
\caption{As in Fig. \ref{fig:accel_of_platoon_WBC} but with the Rear-sided wave-absorbing controller.}
\label{fig:rear_sided_WBC}
\end{figure*}
In Fig. \ref{fig:two_sided_WBC}, the acceleration and changing-distance manoeuvrers carried out for the one-sided wave-absorbing controllers are now performed for the two-sided wave-absorbing controller. Since both platoon ends are fully controlled, the settling time is only half of that for the one-sided wave-absorbing controllers. The middle panel in Fig. \ref{fig:two_sided_WBC} shows that there is no overshoot in distances during the second manoeuver. On the other hand, there is no guarantee that the vehicles will not collide during the acceleration manoeuver.
\begin{figure*}[!htbp]
\centering
\includegraphics[width=0.99\textwidth]{platoon_WBC_two_sides2.pdf}
\caption{As in Fig. \ref{fig:accel_of_platoon_WBC} but with the Two-sided wave-absorbing controller. The second command to increase $d_{\text{ref}}$ comes at $t = 100\,{\text{s}}$.}
\label{fig:two_sided_WBC}
\end{figure*}
\subsection{Evaluation of the performance}
We now evaluate the performance of the acceleration manoeuver described in the previous section with the help of the mean squared error (MSE) criterion,
\begin{equation}
\text{MSE} = \frac{1}{N+1} \sum_{n=0}^{N} \frac{1}{T}\sum_{t=0}^{T} (v_{\text{ref}}(t) -v_n(t))^2,
\label{eq:MSE}
\end{equation}
where $T$ is the simulation time (in our case $T=500\,\text{s}$), $v_{\text{ref}}(t)$ is the reference velocity of the platoon at time $t$ and $v_n(t)$ is the actual velocity of the $n$th vehicle at time $t$.
The comparison in performance of the four controllers for various platoon lengths is depicted in Fig. \ref{fig:stats_accel}. We can see that the MSE increases linearly for all wave-absorbing controllers, but quadratically for the pure bidirectional control without wave absorber. Moreover, a linear increase in MSE for the Two-sided controller is only about half of that for the Front-sided controller. The linear increase of MSE for the Rear-sided controller lies between these two cases. Evidently, the wave-absorbing controller qualitatively improves the performance of the bidirectional control.
\begin{figure}[ht]
\centering
\includegraphics[width=0.49\textwidth]{stats_accel5.pdf}
\caption{MSE performance evaluation of the acceleration manoeuver from Figs. \ref{fig:accel_of_platoon_noWBC}, \ref{fig:accel_of_platoon_WBC} and \ref{fig:two_sided_WBC}. All the four controllers are evaluated; pure bidirectional without wave absorber (left panel), Front-sided wave-absorbing controller (solid line in the middle panel), Rear-sided wave-absorbing controller (dashed line in the middle panel) and Two-sided wave-absorbing controller (right panel) for various platoon lengths according to (\ref{eq:MSE}).}
\label{fig:stats_accel}
\end{figure}
The settling time of the acceleration manoeuver arising from the four types of controllers are compared in Table \ref{table_settling_time}. We can see that the settling time increases quadratically with the platoon length for a platoon without wave-absorbing controller, but approximately linearly for a platoon with wave-absorbing controllers.
\begin{table}[!h]
\renewcommand{\arraystretch}{1.1}
\caption{The time required the platoons of various lengths to accelerate and stay within a range of $5\%$ of $v_{\text{ref}}$.}
\label{table_settling_time}
\centering
\small
\begin{tabular}{|c|c|c|c|c|}
\hline
& No wave & Front-sided & Rear-sided & Two-sided\\
& abs.& wave abs. & wave abs. & wave abs. \\
\hline
5 veh. & $70\,\text{s}$ & $12\,\text{s}$ & $11\,\text{s}$ & $7.5\,\text{s}$\\
\hline
10 veh. & $322\,\text{s}$ & $24\,\text{s}$ & $23\,\text{s}$ & $14\,\text{s}$\\
\hline
20 veh. & $1365\,\text{s}$ & $46\,\text{s}$ & $45\,\text{s}$ & $26\,\text{s}$\\
\hline
40 veh. & $5460\,\text{s}$ & $90\,\text{s}$ &$88\,\text{s}$& $49\,\text{s}$\\
\hline
\end{tabular}
\end{table}
\normalsize
\subsection{Effect of noise in the platoon}
This subsection examines the performance of the four controllers when noise is present in the system. The reference commands for a platoon of $20$ vehicles are $v_{\text{ref}}= 0\,\text{ms}^{-1}$ and $d_{\text{ref}} = 0\,\text{m}$, that is, the platoon is commanded not to move. Normally distributed noise is simulated for $2000$ seconds and added to distance measurements of each vehicle, except for the leader. Different realizations of a normally distributed noise with the mean value $\mu = 0$ and variance $\sigma^2 = 1$ are applied to each vehicle.
Table \ref{table_noise} assesses quantitatively the effect of noise on the performance of the four controllers. The mean squared error of positions, $\text{MSE}_{\text{pos}}$, and the arithmetic mean of positions, $\text{Mean}_{\text{pos}}$, show that the platoon without any absorber and with the Rear-sided wave-absorbing controller perform significantly better than with the other two controllers. This is due to the fact that, at least, one of the platoon ends is anchored at position $0$, meaning that the platoon does not drift away from position $0$, which is not the case for the Front-sided and Two-sided wave-absorbing controllers. Despite the disturbances by noise, all wave-absorbing controllers are better at maintaining in the coherence of the platoon than the pure bidirectional controller, as indicated by the mean squared error of inter-vehicle distances, $\text{MSE}_{\text{dist}}$, and the maximum distance between the leader and the rear end, $\text{MAX}_{\text{dist}}$.
\begin{table}[!h]
\renewcommand{\arraystretch}{1.1}
\caption{Performance of the four controllers when considering normally distributed noise affecting distance measurement of vehicles. Four criterions used for evaluation are introduced in the text.}
\label{table_noise}
\centering
\small
\begin{tabular}{|c|c|c|c|c|}
\hline
& No wave & Front-sided & Rear-sided & Two-sided\\
& abs. & wave abs. & wave abs. & wave abs. \\
\hline
$\text{MSE}_{\text{pos}}$ & $2.7\times 10^7$ & $8.4\times 10^7$ & $7.1\times 10^5$ & $1.3\times 10^8$\\
\hline
$\text{Mean}_{\text{pos}}$ & $2\times 10^{-3}$ & $-3.9$ & $3.2\times 10^{-3}$ & $-3.1$\\
\hline
$\text{MSE}_{\text{dist}}$ & $1.9\times 10^5$ & $2.4\times 10^4$ & $2.5\times 10^4$ & $1.8\times 10^4$\\
\hline
$\text{MAX}_{\text{dist}}$ & $5.75$ & $1.37$ & $1.15$ & $0.64$\\
\hline
\end{tabular}
\end{table}
\normalsize
\section{CONCLUSIONS}
This paper introduces novel concepts for the control of a vehicular platoon, which significantly improve the popular bidirectional control. The main idea is to control the front end or both ends of a platoon to actively damp the waves of positional changes arriving from the opposite platoon end. The absorbing-end vehicle is assumed to i) measure the distance to its neighbour, ii) know its own position and iii) represent the dynamics of a vehicle in terms of the wave transfer function.
The new schemes allow us to control the platoon velocity and the inter-vehicle distances without long-lasting transient and oscillatory behaviour. The velocity errors during the platoon manoeuvres with the traditional bidirectional control grows quadratically with number of vehicles in the platoon, while errors grows only linearly for the bidirectional control enhanced with the wave-absorbing controller. Moreover, the platoon with the wave-absorbing controller is string stable.
Additionally, the wave-absorbing controller preserves advantages of the bidirectional control such as: i) The lack of a need for vehicle-to-vehicle communication, ii) none of the vehicles needs to know the number of vehicles in the platoon, iii) an in-platoon vehicle does not need to know its relative position in the platoon, and iv) an in-platoon vehicle does not need to know the reference velocity and the reference distance for the platoon.
However, a considerable mathematical difficulty in the wave-absorbing control lies in finding the impulse response of the wave transfer function. In this paper, we proposed the iterative approach of constructing an approximation of the wave transfer function that is based on a continued fraction representation. Even for a small number of iterative steps, when the wave transfer function is rather roughly approximated, the wave-absorbing control still performs efficiently to damper oscillations in the platoon's characteristics (i.e. velocity, inter-vehicle distances).
It should be noted that the absorbing-end vehicle is assumed to be equipped with the positional controller since the differences in positions between vehicles are controlled. Alternatively, when the absorbing-end vehicle is equipped with a velocity controller, the commanded position of the vehicle derived using (\ref{eq:external_input_leader}) or (\ref{eq:external_input_last}) can be numerically differentiated to obtain the velocity commanded to the absorbing-end vehicle.
Undesirable overshoots in the velocities or inter-vehicle distances of the wave-absorbing control can be eliminated by introducing time delays in the reference signal applied to one of the platoon ends. An appropriate value of this time delay is dependent upon the platoon length and thus requires the extension of the wave-absorbing control. This topic warrants further investigation.
This paper extends \cite{Martinec2014} submitted on 15.~October 2013 in the following way: i) It presents mathematical derivation of the approximating formula for the wave transfer function and derivation of the transfer functions describing wave reflection on platoon ends, ii) it generalizes the result from double integrator model with linear friction and PI controller for an arbitrary local system dynamics, iii) it introduces two additional modifications of the wave-absorbing controller for the vehicular platoon, iv) it analyses asymptotic and string stability of a platoon with the wave-absorbing controller and v) it more thoroughly evaluates performance of the wave-absorbing controller.
\section{ACKNOWLEDGEMENTS}
This work was supported by Grant Agency of the Czech Republic within the project GACR P103-12-1794.
The authors thank Kevin Fleming for his comments on the manuscript.
|
\section{Existentially closed groups}
A group $G$ is called {\em existentially closed}, if any finite
consistent system of equations and in-equations with coefficients
from $G$ has a solution in $G$. A system
$$
S=\{ w_i(x_1, \ldots, x_n)=1; (1\leq i\leq r), w_j(x_1, \ldots,
x_n)\neq 1; (r+1\leq j\leq s)\}
$$
with coefficients in $G$ is called consistent, if there is a group
$K$ containing $G$, such that $S$ has a solution in $K$. One can
generalize this definition to an arbitrary class of groups: Let
$\mathfrak{X}$ be a class of groups. A group $G\in \mathfrak{X}$ is
called existentially closed in the class $\mathfrak{X}$, if every
$\mathfrak{X}$-consistent system $S$ has a solution in $G$. Here,
$\mathfrak{X}$-consistency means that there exists a group $K\in
\mathfrak{X}$ which contains $G$ and $S$ has a solution in $K$. As
we said in the introduction, if $\mathfrak{X}$ is the class of all
groups, then existentially closedness is the same as algebraically
closedness.
The next lemma and theorem are proved in \cite{Scott} for the class of all groups, but we give here the proofs again for the sake of completeness.
Recall that a class of groups is called {\em inductive}, if it contains the union of any chain of its elements. \\
\begin{lemma}
Let $\mathfrak{X}$ be an inductive class of groups which is closed
under the operation of taking subgroups. Let $G\in \mathfrak{X}$.
Then there is a group $H\in \mathfrak{X}$ which contains $G$ and its
order is at most $\max \{ \aleph_0, |G|\}$. Further, for any finite
system $S$ of equations and in-equations over $G$, we have either of
the following assertions:
1- $S$ has a solution in $H$.
2- For any extension $H\subseteq E\in \mathfrak{X}$, the system $S$ has no solution in $E$.
\end{lemma}
\begin{proof}
Suppose $\kappa=\max \{ \aleph_0, |G|\}$. Clearly the cardinality of
the set of all finite systems of equations and inequations over $G$,
is also at most $\kappa$. We suppose that this set is well-ordered
as $\{ S_{\alpha}\}_{\alpha}$, by the ordinals $0\leq \alpha \leq
\kappa$. Let $G_0=G$ and suppose that $G_{\gamma}\in \mathfrak{X}$
is already defined in such a way that $|G_{\gamma}|\leq \kappa$ and
$\beta<\gamma$ implies $G_{\beta}\subseteq G_{\gamma}$. Let
$$
K_{\alpha}=\bigcup_{\gamma<\alpha}G_{\gamma}.
$$
Clearly, $K_{\alpha}\in \mathfrak{X}$ and $|K_{\alpha}|\leq \kappa^2=\kappa$. If $S_{\alpha}$ has no solution in any extension of $K_{\alpha}$,
then we set $G_{\alpha}=K_{\alpha}$, otherwise there is a $E\in \mathfrak{X}$ which contains $K_{\alpha}$ and $S_{\alpha}$ has a solution,
say $\bar{u}=(u_1, \ldots, u_n)$ in $E$ ($n$ is the number of indeterminate in $S_{\alpha}$). Let
$$
G_{\alpha}=\langle K_{\alpha}, u_1, \ldots, u_n\rangle \leq E.
$$
Then $G_{\alpha}\in \mathfrak{X}$ and $|G_{\alpha}|\leq \kappa$. So, for any $\alpha<\kappa$, we have defined a $G_{\alpha}$. Note that, we have also
$$
\beta<\alpha \Rightarrow G_{\beta}\subseteq G_{\alpha}.
$$
Now, the group $H=\cup G_{\alpha}\in \mathfrak{X}$ has the required properties. This is because, any system $S=S_{\alpha}$ which has no solution in $H$,
has no solution in $K_{\alpha}$ as well. So it has no solution in any extension of $H$.
\end{proof}
\begin{theorem}
Let $\mathfrak{X}$ be an inductive class of groups which is closed under the operation of taking subgroups. Let $G\in \mathfrak{X}$. Then,
there exists a group $G^{\ast}\in \mathfrak{X}$, with the following properties,
1- $G$ is a subgroup of $G^{\ast}$.
2- $G^{\ast}$ is existentially closed in the class $\mathfrak{X}$.
3- $|G^{\ast}|\leq \max \{ \aleph_0, |G|\}$.
\end{theorem}
\begin{proof}
Let $G^0=G$ and $G^1=H$, where $H$ is the group constructed in the lemma. Suppose $G^m$ is already defined and $G^{m+1}$ is the group which is
proved to does exist for $G^m$ in the lemma. Let $G^{\ast}=\cup G^m$. Therefore, $G^{\ast}\in \mathfrak{X}$, satisfies conditions 1 and 3.
To prove 2, suppose $S$ is a consistent system, with coefficients from $G^{\ast}$. Since $S$ is finite, so there is an $m$ such that all of the
coefficients of $S$ belong to $G^m$. So, $S$ has a solution in $G^{m+1}\subseteq G^{\ast}$.
\end{proof}
Note that in the number 3 of the above theorem, we may have
$|G^{\ast}|< \max \{ \aleph_0, |G|\}$. For example, suppose $G$ is
an arbitrary group and $\mathfrak{X}$ is the set of all subgroups of
$G$. Then trivially, $G^{\ast}=G$ and so, if $G$ is finite, then the
inequality 3 is strict. Note that also, there are many inductive
classes of groups, which are closed under subgroup: any variety,
quasi-variety and in general, any universal class of groups has this
property. As the first application of the above theorem, we show
that every torsion free group can be embedded in a torsion free
group with exactly two conjugacy classes. Note that we can use a
similar arguments to prove the existence of torsion free groups of
any infinite cardinality with just two conjugacy classes. Also note
that the type of the group we are giving here, is not new, it is known from works of Higman, Neumann and Neumann \cite{Lynd}.\\
\begin{corollary}
Every torsion free group $G$ can be embedded in a torsion free group
$G^{\ast}$ which has just two conjugacy classes and its cardinality
is the same as the cardinality of $G$.
\end{corollary}
\begin{proof}
Suppose $\mathfrak{X}$ is the class of all torsion free groups.
Hence $\mathfrak{X}$ is inductive and closed under the operation of
taking subgroups. We, begin with the group $G\in \mathfrak{X}$.
Suppose $G^{\ast}\in \mathfrak{X}$ is the existentially closed group
relative to $\mathfrak{X}$, which is constructed for $G$ in the
theorem. We show that $G^{\ast}$ is the required group. Let $a, b\in
G^{\ast}$ be two non-identity elements. Consider the equation
$xax^{-1}=b$. Let
$$
G^{\ast}_{a, b}=\langle G^{\ast}, t:\ tat^{-1}=b\rangle
$$
be an HNN-extension of $G^{\ast}$. We know that every torsion
element of this HNN-extension is conjugate to a torsion element of
$G^{\ast}$, so $G^{\ast}_{a, b}$ is torsion free. It also contains
$G^{\ast}$ as a subgroup and clearly $t$ is a solution for
$xax^{-1}=b$ in $G^{\ast}_{a, b}$. Therefor, there is already a
solution in $G^{\ast}$. Hence $G^{\ast}$ is a torsion free group,
containing $G$, with just two conjugacy classes. Further
$|G^{\ast}|=|G|$.
\end{proof}
\section{An embedding theorem for T$_{\pi}$-groups}
It is easy to see that the simple group $G^{\ast}$ obtained in the
corollary 1.3 is divisible. In this section, we prove a similar
result for T$_{\pi}$-groups. Let $\pi$ be a set of primes and $G$ be
a group, such that its torsion elements have $\pi$-orders, i.e. the
equality $x^m=1$ implies that $x=1$ or $m$ is a $\pi$-number. Then
we say that $G$ is a T$_{\pi}$-group. If $\pi=\emptyset$, then
T$_{\pi}$-groups are exactly torsion free groups. Note that a finite
T$_{\pi}$-group is just a finite $\pi$-group. A group $G$ is also
said to be a $\pi^{\prime}$-group, if for any $\pi^{\prime}$-number
$m$, every element of $G$ has an $m$-th root.
\begin{theorem}
Let $G$ be a T$_{\pi}$-group. Then there exists a
$\pi^{\prime}$-divisible simple T$_{\pi}$-group $G^{\ast}$
containing $G$ and with the cardinality $\max \{ \aleph_0, |G|\}$,
such that
1- element of the same order in $G^{\ast}$ are conjugate,
2- $G^{\ast}$ is not finitely generated,
3- every finite $\pi$-group embeds in $G^{\ast}$.
4- every finitely presented $\pi$-group can be residually embedded
in $G^{\ast}$,
5- for any finite $\pi$-group $A$, we have $Aut(A)\cong
\frac{N_{G^{\ast}}(A)}{C_{G^{\ast}}(A)}$.
\end{theorem}
\begin{proof}
Let $\mathfrak{X}_{\pi}$ be the class of all T$_{\pi}$-groups. This
class is inductive and closed under subgroup. Since $G\in
\mathfrak{X}_{\pi}$, so there exists a group $G^{\ast}\in
\mathfrak{X}_{\pi}$ containing $G$, which is existentially closed in
the class $\mathfrak{X}_{\pi}$. Suppose $1\neq a, b\in G^{\ast}$
have the same orders. We show that $a$ and $b$ are conjugate.
Suppose the HNN-extension
$$
G^{\ast}_{a,b}=\langle G^{\ast}, t: tat^{-1}=b\rangle.
$$
It is enough to show that $G^{\ast}_{a,b}$ belongs to
$\mathfrak{X}_{\pi}$. Let $x$ be an element of finite order in
$G^{\ast}_{a, b}$. We know that $x$ is conjugate to an element of
$G^{\ast}$, see \cite{Lynd}. So the order of $x$ is a $\pi$-number,
showing that $G^{\ast}_{a,b}\in \mathfrak{X}_{\pi}$. This shows
that elements of the same order in $G^{\ast}$ are conjugate. To show
that $G^{\ast}$ is $\pi^{\prime}$-divisible, let $x\in G^{\ast}$ and
$m$ be a $\pi^{\prime}$-number. We know that the orders of $x$ and
$x^m$ are equal. Hence there is a $z\in G^{\ast}$ such that
$x=zx^mz^{-1}$. Suppose $u=zxz^{-1}$. Then $u^m=x$ and hence
$G^{\ast}$ is $\pi^{\prime}$-divisible.
We show that $G^{\ast}$ is simple. Note that $G^{\ast}\ast \langle
x\rangle$ is a T$_{\pi}$-group. This follows from the fact that
$G^{\ast}$ is an T$_{\pi}$-group and the the reduced form of
elements in free products is unique. Let $1\neq a, w\in G^{\ast}$ be
arbitrary elements and suppose $u=wxw^{-1}x^{-1}$ and
$v=axw^{-1}x^{-1}$. Then $u$ and $v$ are reduced in the free product
and so they have infinite orders. Hence we can consider the
HNN-extension
$$
M=\langle G^{\ast}\ast \langle x\rangle, t: tut^{-1}=v\rangle.
$$
With the same argument (as for $G^{\ast}_{a,b}$), we see that $M$ is
also an T$_{\pi}$-group. The equation
$$
twxw^{-1}x^{-1}t^{-1}=axw^{-1}x^{-1}
$$
has a solution for $t$ and $x$ in $M$, therefore it has already a solution in $G^{\ast}$. We have
$$
a=(twt^{-1})(txw^{-1}x^{-1}t^{-1})(xw^{-1}x^{-1}),
$$
so $a\in \langle w^{G^{\ast}}\rangle$. Hence, for all $1\neq w\in G^{\ast}$, we have $G^{\ast}=\langle w^{G^{\ast}}\rangle$.
This proves that $G^{\ast}$ is simple.
Now, since $G^{\ast}$ is non-abelian simple group, we have $Z(G^{\ast})=1$. On the other hand, for any finite subset
$a_1, \ldots, a_m\in G^{\ast}$, the system
$$
a_1x=xa_1, \ldots, a_mx=xa_m, x\neq 1,
$$
has a solution in the T$_{\pi}$-group $G^{\ast}\times \langle
x\rangle$, so it has a solution in $G^{\ast}$. Therefore
$$
C_{G^{\ast}}(\langle a_1, \ldots, a_m\rangle)\neq 1.
$$
This proves that $G^{\ast}$ is not finitely generated.
Now, suppose that $A$ is a finite $\pi$-group and
$$
A=\{ 1=g_0, g_1, \ldots, g_m\}.
$$
For any $1\leq i, j\leq m$ there exists a unique $0\leq k(i, j)\leq
m$, such that $g_ig_j=g_{k(i,j)}$. The group $G^{\ast}\times A$ is
clearly an T$_{\pi}$-group and the system
$$
x_ix_j=x_{k(i.j)}; (0\leq i, j\leq m); x_i\neq x_j; (0\leq i<j\leq m)
$$
has a solution in $G^{\ast}\times A$ and hence it has a solution in
$G^{\ast}$. This shows that $A$ is embedded in $G^{\ast}$. To prove
$4$, let $A$ be a finitely presented $\pi$-group, with a finite
presentation
$$
A=\langle x_1, \ldots, x_n: r_1, \ldots, r_m\rangle.
$$
Suppose $1\neq w\in A$ and consider the system
$$
r_i(x_1, \ldots, x_n)=1; (1\leq i\leq m), w(x_1, \ldots,x_n)\neq 1.
$$
Clearly this equation has a solution in $G^{\ast}\times A\in
\mathfrak{X}_{\pi}$, and hence it has a solution in $G^{\ast}$. This
proves that there is a homomorphism $\varphi: A\to G^{\ast}$ with
$\varphi(w)\neq 1$. Hence $A$ embeds in $G^{\ast}$ residually. We
now prove $5$. Suppose again that $A$ is a finite $\pi$-group and
$\alpha\in Aut(A)$. This time, we consider the system
$$
xax^{-1}=\alpha(a);\ (a\in A).
$$
This system has a solution in the HNN-extension
$$
G^{\ast}_{\alpha}=\langle G^{\ast}, t: tat^{-1}=\alpha(a); (a\in A)\rangle.
$$
Therefore, there exists an element $x\in G^{\ast}$ such that $\alpha$ is equal to the restriction of
the inner automorphism $T_x$ to $A$. This also shows that $x\in
N_{G^{\ast}}(A)$ and so the map which sends $x$ to the restriction
of $T_x$ on $A$ is an epimorphism from $N_{G^{\ast}}(A)$ to $Aut(A)$
with the kernel $C_{G^{\ast}}(A)$. This completes the proof of $5$.
\end{proof}
We can generalize the above theorem for HNN-classes of groups. A
class $\mathfrak{X}$ of groups is called an HNN-class if it is
inductive, closed under subgroup, and HNN-extensions of the
elements of $\mathfrak{X}$ belong to $\mathfrak{X}$. The class of
all T$_{\pi}$-groups is an example of HNN-classes. By a similar
argument as in the above theorem, we can prove the next result.
\begin{theorem}
Let $\mathfrak{X}$ be an HNN-class of groups. Then every element $G$
of $\mathfrak{X}$ embeds in a simple $G^{\ast}\in \mathfrak{X}$,
with the cardinality at most $\max \{ \aleph_0, |G|\}$. Further
elements of the same orders in $G^{\ast}$ are conjugates.
\end{theorem}
\section{Some Olshanskii like groups}
In mid twenties, Alfred Tarski asked about the existence of
infinite groups all proper non-trivial subgroups of which are of
fixed prime order $p$. In 1982, A. Yu. Olshanskii \cite{Olshan},
constructed an uncountable family of such groups using his geometric
method of graded diagrams over groups, for all primes $p>10^{75}$.
The groups constructed are called Tarskii monsters since then. These
groups are two-generator simple groups and hence are countable. In
this section, for any fixed prime $p$, we give a quite elementary
proof for existence of countable non-abelian simple groups with the
property that their all non-trivial {\em finite subgroups} are
cyclic of order $p$.
We will consider two special classes of groups in this section. The
first one consists of groups all finite subgroups in which are
cyclic. We will denote this class by $\mathfrak{X}_{fc}$. The second
class which will be denoted by $\mathfrak{X}_p$, is the class of all
groups in which their non-trivial finite subgroups are of order $p$,
for a fixed prime $p$. Note that both classes are HNN-classes.
Clearly the Monsters constructed by Olshanskii satisfy the
requirements of the next theorem, but we don't use that monsters,
since we have a very elementary proof for our claims. What we need
is the theorem 1.2 and some facts about finite subgroups of
HNN-extensions (and also those of free products). It is known that
(see \cite{Lynd}, page 212) every finite subgroup of any
HNN-extension
$$
G=\langle A, t: tFt^{-1}=\phi(F)\rangle
$$
is contained in some conjugate of $A$. Also, every finite subgroup
of any free product $A\ast B$ is contained in some conjugate of $A$
or some conjugate of $B$.
\begin{theorem}
There exists a countable non-abelian simple group $M$ such that all
finite subgroups of $M$ are cyclic and for any prime $p$, the group
$M$ has an element of order $p$.
\end{theorem}
\begin{proof}
We know that the class $\mathfrak{X}_{fc}$ is inductive and closed
under subgroup, so we can apply 1.2. Note that, we have many groups
which belong to $\mathfrak{X}_{fc}$, for example any finite cyclic
group, $\mathbb{Z}_{p^{\infty}}$, and any torsion free group. So,
let $G\in \mathfrak{X}_{fc}$ be a countable arbitrary element.
Therefore, there exists a group $G^{\ast}$ satisfying requirements
of 1.2. Let $M=G^{\ast}$ and $p$ be a prime. We show that $M$ has an
element of order $p$. Suppose this is not true. Hence the system
$$
x^p=1, \ x\neq 1
$$
has no solutions in $M$. Let $H\leq M\times \mathbb{Z}_p$ be a
finite subgroup. We have $H\subseteq \pi_1(H)\times \pi_2(H)$, where
$\pi_1$ and $\pi_2$ are projections. Since $M\in \mathfrak{X}_{fc}$,
so $\pi_1(H)$ is cyclic and by our assumption $p$ is co-prime to the
order of $\pi_1(H)$. Therefore $\pi_1(H)\times \pi_2(H)$ is cyclic
and so is $H$. Hence $M\times \mathbb{Z}_p$ belongs to
$\mathfrak{X}_{fc}$. But the above system has a solution in $M\times
\mathbb{Z}_p$, a contradiction.
Note that, for any cyclic group $\langle x\rangle$ the group $M\ast
\langle x\rangle$ belongs to $\mathfrak{X}_{fc}$. To see this, let
$H$ be a finite subgroup of this free product. Then $H$ must be
contained in a conjugate of $M$ and hence $H$ is cyclic. Now,
suppose $1\neq a, b\in M$ and consider two elements
$$
u=axa^{-1}x^{-1},\ v=bxa^{-1}x^{-1}
$$
in $M\ast \langle x\rangle$. Clearly $u$ and $v$ have infinite
orders and so we can consider the HNN-extension
$$
M^{\ast}=\langle M\ast \langle x\rangle, t:\ tut^{-1}=v\rangle.
$$
Let $H$ be a finite subgroup of this HNN-extension. Then $H$ is
contained in some conjugate of $M$ and therefore it is cyclic. Hence
$M^{\ast}$ is also an element of $\mathfrak{X}_{fc}$. So, the
equation
$$
taxa^{-1}x^{-1}t^{-1}=bxa^{-1}x^{-1}
$$
has a solution for $t$ and $x$ in $M^{\ast}$, therefore it has
already a solution in $M$. We have
$$
b=(tat^{-1})(txa^{-1}x^{-1}t^{-1})(xa^{-1}x^{-1}),
$$
so $b\in \langle a^{M}\rangle$. Hence, for all $1\neq a\in M$, we
have $M=\langle a^{M}\rangle$. This proves that $M$ is simple.
Finally note that since for any $a\in M$, the in-equation $ax\neq
xa$ has a solution in $M\ast\langle x\rangle$, so the center of $M$
is trivial and hence $M$ is non-abelian.
\end{proof}
Now, a similar argument on the class $\mathfrak{X}_p$ proves the
next theorem.
\begin{theorem}
Let $p$ be a fixed prime. Then there exists a countable non-abelian
simple group $M$ (which is not torsion free) such that any finite
non-trivial subgroup of $M$ is cyclic of order $p$.
\end{theorem}
The following two corollaries can be deduced instantly.
\begin{corollary}
There exists a non-abelian two generator group such that the
orders of the generators are distinct primes and its every finite
subgroup is cyclic.
\end{corollary}
\begin{corollary}
There exists a non-abelian $p$-group with two generators such that
its every finite subgroup has order $p$.
\end{corollary}
\section{Existentially closed algebras}
We are going now to find similar embedding theorems for Lie
algebras. But there are two major differences between groups and Lie
algebras. First, we don't have any suitable definition of {\em
torsion} in the case of Lie algebras so, in advance, we don't have a
parallel concept of T$_{\pi}$-Lie algebra and so on. Instead, we can
express our theorems in terms of arbitrary Lie algebras. The second
main difference is related to HNN-extensions of Lie algebras. Here,
an HNN-extension comes from a Lie algebra and a derivation of some
subalgebra, despite groups where HNN-extensions are always defined
by groups and isomorphisms between subgroups. We will give a brief
summary of HNN-extensions of Lie algebras in the next section. In
this section, we give the analogue of Theorem 1.2 for Lie algebras,
in fact since it can be formulated for arbitrary non-associative
algebras, we prove it in the most general form.
\begin{lemma}
Let $\mathfrak{X}$ be an inductive class of (not necessarily associative) algebras over a field $K$. Suppose $\mathfrak{X}$ is closed under
subalgebra and $L\in \mathfrak{X}$. Then there exists an algebra $H\in \mathfrak{X}$ containing $L$ such that its dimension is at most
$$
\max\{ \aleph_0, \dim L, |K|\}.
$$
Further, for any system $S$ of equations and in-equations over $L$,
we have either of the following assertions:
1- $S$ has a solution in $H$.
2- For any extension $H\subseteq E\in \mathfrak{X}$, the system $S$ has no solution in $E$.
\end{lemma}
\begin{proof}
We assume that $X$ is a countable set of variables and
$$
\eta=\max \{\aleph_0, \dim L, |K|\}.
$$
Any equation over $L$ consists of finitely many elements of $L$ and
$X$, so $\kappa$, the number of systems of equations and
in-equations over $L$, is $|L\cup X|$. Note that $|L|=\max \{ \dim
L, |K|\}$, hence $\kappa=|L|+\aleph_0=\eta$.
We well-order the set of all systems as $\{ S_{\alpha}\}_{\alpha}$, using ordinals $0\leq \alpha\leq \kappa$. Suppose $L_0=L$.
Let for any $0\leq \gamma\leq \alpha$, the algebra $L_{\gamma}\in \mathfrak{X}$ is defined in such a way that $|L_{\gamma}|\leq \kappa$ and
$$
\beta\leq \gamma \Rightarrow L_{\beta}\subseteq L_{\gamma}.
$$
We put
$$
E_{\alpha}=\bigcup_{\gamma\leq \alpha}L_{\gamma},
$$
so $E_{\alpha}\in \mathfrak{X}$ and further
$$
|E_{\alpha}|\leq \alpha |L_{\gamma}|\leq \kappa^2=\kappa.
$$
Suppose $S_{\alpha}$ has not solution in any extension of $E_{\alpha}$. Then we set $L_{\alpha}=E_{\alpha}$. If there is an extension
$E_{\alpha}\subseteq E\in \mathfrak{X}$ such that $S_{\alpha}$ has a solution $(u_1, \ldots, u_n)$ in $E$ ($n$ is the number of indeterminate
in $S_{\alpha}$), then we set
$$
L_{\alpha}=\langle E_{\alpha}, u_1, \ldots, u_n\rangle \subseteq E.
$$
Since $\mathfrak{X}$ is closed under subalgebra, so $L_{\alpha}\in \mathfrak{X}$ and also
$$
|L_{\alpha}|=|E_{\alpha}|\leq \kappa.
$$
Now, we define
$$
H=\bigcup_{0\leq \alpha\leq \kappa}L_{\alpha}
$$
which is an element of $\mathfrak{X}$. We have $|H|\leq \kappa^2=\kappa$ and hence
$$
\max\{ \dim H, |K|\}\leq \max \{\aleph_0, \dim L, |K|\},
$$
therefore we have
$$
\dim H\leq \max \{\aleph_0, \dim L, |K|\}.
$$
\end{proof}
\begin{theorem}
Let $\mathfrak{X}$ be an inductive class of (not necessarily associative) algebras over a field $K$. Suppose $\mathfrak{X}$ is closed under
subalgebra and $L\in \mathfrak{X}$. Then there exists an algebra $L^{\ast}\in\mathfrak{X}$ with the following properties,
1- $L$ is a subalgebra of $L^{\ast}$.
2- $L^{\ast}$ is existentially closed in the class $\mathfrak{X}$.
3- $\dim L^{\ast}\leq \max \{\aleph_0, \dim L, |K|\}$.
\end{theorem}
\begin{proof}
Let $H^0=L$ and $H^1=H$ be an algebra satisfying requirements of the previous lemma. Suppose $H^m$ is defined and let $H^{m+1}$ be an algebra
obtained by the lemma from $H^m$. We have
\begin{eqnarray*}
\dim H^{m+1}&\leq& \max \{ \aleph_0, \dim H^m, |K|\}\\
&=&\max \{\aleph_0, \dim L, |K|\}.
\end{eqnarray*}
Now, suppose
$$
L^{\ast}=\bigcup_mH^m.
$$
This is an algebra having the properties 1-2-3.
\end{proof}
\section{Embedding of Lie algebras}
In \cite{shirvani} and \cite{Alon}, the concept of the HNN-extension
is defined for Lie algebras. Suppose $L$ is a Lie algebra over a
filed $K$ and $A$ is a subalgebra. Let $\delta:A\to L$ be a
derivation. Define a Lie algebra $L_{\delta}$ with the presentation
$$
L_{\delta}=\langle L, t:\ [t, a]=\delta(a); (a\in A)\rangle.
$$
The properties of this HNN-extension is studied in \cite{shirvani}
and \cite{Alon}. It is proved that $L$ is a subalgebra of
$L_{\delta}$. Similar constructions are also introduced for Lie
$p$-algebras and rings in \cite{shirvani}. In this section, using
this HNN-extension and the notion of existentially closed Lie
algebras, we obtain a new embedding theorem.
\begin{theorem}
Let $L$ be a Lie algebra over a field $K$. Then there exists a Lie algebra $L^{\ast}$ having the following properties,
1- $L$ is a subalgebra of $L^{\ast}$,
2- for any non-zero $a, b\in L^{\ast}$, there exists $x\in L^{\ast}$
such that $[x,a]=b$, and so $L^{\ast}$ is simple.
3- $\dim L^{\ast}\leq \max \{ \aleph_0, \dim L, |K|\}$,
4- $L^{\ast}$ is not finitely generated,
5- every finite dimensional simple Lie algebra over $K$ embeds in
$L^{\ast}$,
6- every finitely presented Lie algebra over $K$ embeds residually
in $L^{\ast}$,
7- if $K$ is finite, then every finite dimensional Lie algebra over
$K$ embeds in $L^{\ast}$,
8- if $K$ is finite and $A$ is finite dimensional Lie algebra over
$K$, then we have
$$
Der(A)\cong \frac{N_{L^{\ast}}(A)}{C_{L^{\ast}}(A)}.
$$
\end{theorem}
\begin{proof}
We suppose that $\mathfrak{X}$ is the class of all Lie algebras and
then we apply the theorem 3.2. Hence, there exists an existentially
closed Lie algebra $L^{\ast}$ containing $L$ such that
$$
\dim L^{\ast}\leq \max \{ \aleph_0, \dim L, |K|\}.
$$
Let $0\neq a, b\in L^{\ast}$ and $\delta:\langle a\rangle \to L$ be the derivation $\delta(a)=b$. Consider the HNN-extension
$$
L^{\ast}_{\delta}=\langle L^{\ast}, t: [t, a]=b\rangle.
$$
We know that $L^{\ast}\leq L^{\ast}_{\delta}$ and the equation
$[x,a]=b$ has a solution in $L^{\ast}_{\delta}$, so 2 is proved.
This implies also that $L^{\ast}$ is simple. To prove 4, suppose
$x_1, \ldots, x_n$ is a finite set of elements of $L^{\ast}$.
Consider the system
$$
[x, x_i]=0; (1\leq i\leq n), x\neq 0.
$$
This system has a solution in the Lie algebra $L^{\ast}\times \langle x\rangle$, and so we have $C_{L^{\ast}}(\langle x_1, \ldots, x_n\rangle)\neq 0$,
while $Z(L^{\ast})=0$. Therefore $L^{\ast}$ is not finitely generated.
Suppose $E$ is a finite dimensional simple Lie algebra with a basis $u_1, \ldots, u_n$ and suppose $[u_i, u_j]=\sum_r \lambda_{ij}^r u_r$.
Consider the system
$$
[x_i, x_j]=\sum_r\lambda_{ij}^r x_r; (1\leq i, j\leq n), x_i\neq 0; (1\leq i\leq n).
$$
This system has a solution in $L^{\ast}\times E$ and so there is a
non-zero homomorphism $E\to L^{\ast}$. Therefore $E$ embeds in
$L^{\ast}$. The proof of 6 is similar, so we prove 7. Suppose $K$ is
finite and let $E$ be a finite dimensional Lie algebra with a basis
$u_1, \ldots, u_n$ and structural constants $\lambda_{ij}^r$, i.e.
$[u_i,u_j]=\sum_r\lambda_{ij}^r u_r$. Let $T$ be the set of all
$n$-tuples $(a_1, \ldots, a_n)\in K^n$ with $a_i\neq 0$ for some
$i$. Consider the system
$$
[x_i, x_j]=\sum_r\lambda_{ij}^r x_r; (1\leq i,j\leq n)
$$
$$
\sum_r a_rx_r\neq 0; ( (a_1, \ldots,a_n)\in T).
$$
This system has a solution in $L^{\ast}\times E$, and so $E$ embeds in $L^{\ast}$.
Finally, to prove 8, let $K$ be finite and $A$ be finite dimensional
(subalgebra of $L^{\ast}$). Let $\delta\in Der(A)$ and consider the
HNN-extension
$$
L^{\ast}_{\delta}=\langle L^{\ast}, t: [t,a]=\delta(a); (a\in A)\rangle
$$
in which the system
$$
[x, a]=\delta(a); (a\in A)
$$
has a solution. So, there is an $x\in L^{\ast}$ such that $\delta(a)=[x,a]$ for all $a\in A$. Clearly $x\in N_{L^{\ast}}(A)$.
So there is an epimorphism $N_{L^{\ast}}(A)\to Der(A)$ with the kernel $C_{L^{\ast}}(A)$. So we have
$$
Der(A)\cong \frac{N_{L^{\ast}}(A)}{C_{L^{\ast}}(A)}.
$$
\end{proof}
As an interpretation of the above theorem, define a Lie algebra to
be {\em division Lie algebra}, if for all $a$ and $b$ with $a\neq
0$, there exists an $x$ such that $[a, x]=b$. Then the above theorem
shows that every Lie algebra can be embed in a simple division Lie
algebra. However, that division Lie algebra is not finitely
generated. It can be asked that if there exists a finitely generated
division Lie algebra.
In \cite{shah}, we proved that if a Lie algebra $L$ over a filed of
characteristics zero has a finite dimensional ideal $U$ such that
$U^n\subseteq [A, U]$ for some abelian subalgebra $A$ and $n\geq 2$,
then $U$ is solvable. Here $[A, U]$ is the linear span of the set
$\{ [a, x]:\ a\in A, x\in U\}$. The Lie algebra $L^{\ast}$ which we
obtained in 5.1, shows that the assumption on the dimension of $U$
is essential, for example of we let $U=L^{\ast}$ and if we assume
that $A$ is any 1-dimensional subalgebra, then always $U^n\subseteq
[A, U]$, but clearly $U$ is not solvable ($L^{\ast}$ is simple).
\section{Existentially complete groups}
In this section, we define the notion of {\em existentially
complete} groups. It can be define the general frame of the
universal algebra. Existentially complete groups are special kind of
existentially closed groups, where in their definition, we use some
special sentences of the second order language of groups.
Let $G$ be a group and $A\leq G$ be a finitely generated subgroup.
Suppose $w(x_1, \ldots, x_n)\in G\ast F(x_1, \ldots, x_n)$. The
expression
$$
w(x_1, \ldots, x_n)\in A
$$
is called a {\em membership}. The negation of a membership is a {\em
non-membership}. We say that a finite system of memberships and
non-memberships (briefly, a membership system), is consistent, if a
there is a group $K$ containing $G$, such that the system has a
solution in $K$. A group $G$ with the property that every consistent
membership system has a solution in $G$, will be called
existentially complete. Clearly, such a group is in also
existentially closed. We can also define the notion of existentially
complete group in a given class. In some classes, the two notions
are equivalent, for example, in the class of locally finite groups,
there is no difference between two notions. We can define similarly,
existentially complete Lie algebras and also it is easy to see that
over finite fields, two properties of existentially closedness and
existentially completeness are equivalent.
A completely similar argument as in the section 1, the reader can
prove the next theorem. As a result, it shows that there exists
existentially complete groups.
\begin{theorem}
Let $\mathfrak{X}$ be an inductive class of groups, closed under
subgroup, such that $G\in \mathfrak{X}$. Then there exists an
existentially complete group $G^{\ast}$, containing $G$, such that
$$
|G^{\ast}|\leq \max\{ \aleph_0, |G|\}.
$$
\end{theorem}
The following embedding theorem, gives an interesting application of
the notion of existentially complete groups.
\begin{theorem}
Let $M$ be an existentially complete group and $A$ be a finitely
generated subgroup. Then, every extension of $A$ by any finite
group embeds in $M$.
\end{theorem}
\begin{proof}
Let $A=\langle a_1, \ldots, a_m\rangle$ and $G$ be a group, having a
normal subgroup $B$, isomorphic to $A$. Let $\phi$ is the
isomorphism from $A$ to $B$ and $b_i=\phi(a_i)$. Suppose also that
$$
\frac{G}{B}=\{ g_1B, \ldots, g_nB\}.
$$
Let
$$
(g_iB)(g_jB)=g_{r(i,j)}B,
$$
where $r(i,j)$ is a function of $i$ and $j$. So, for any $i$ and
$j$, there exists an element $q_{ij}\in B$ such that
$$
g_{r(i,j)}^{-1}g_ig_j=q_{ij}.
$$
We also have $g_ib_jg_i^{-1}\in B$, so $g_ib_jg_i^{-1}=p_{ij}$, for
some $p_{ij}\in B$. Now, consider the next membership system
\begin{eqnarray}
x_{r(i,j)}^{-1}x_ix_j&=&\phi^{-1}(q_{ij}),\ \ \ \ \ \ \ (1\leq i,j\leq n)\\
x_i^{-1}x_j&\not\in& A,\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ (1\leq i<j\leq n)\\
x_ia_jx_i^{-1}&=& \phi^{-1}(p_{ij}),\ \ \ \ \ \ \ (1\leq i\leq n, 1\leq
j\leq m)
\end{eqnarray}
Let $M_1=\langle M\ast G: A=B, \phi\rangle$ be the amalgamated free
product of $M$ and $G$. We can verify easily that $(g_1, \ldots,
g_n)$ is a solution of the above system in $M_1$. Hence, the system
has a solution $(u_1, \ldots, u_n)$ in $M$. Let
$$
K=\langle a_1, \ldots, a_m, u_1, \ldots, u_n\rangle,
$$
which is a subgroup of $M$, containing $A$ as a normal subgroup, by
(3). This shows that every element of $K$ has the form $w(u_1,
\ldots, u_n)a$, such that $a\in A$ and $w$ is a group word in $u_1,
\ldots, u_n$. Now, by (1), we can replace every $u_iu_j$ by
$u_{r(i,j)}\phi^{-1}(q_{ij})$, and collecting all factors by this
method, we finally see that any element of $K$ has the unique form
$u_ia$ for some $1\leq i\leq n$ and $a\in A$. Note that the same
observation is also true for the elements of $G$. Now, we prove that
$K\cong G$. Let $\psi:K\to G$ be define as $\psi(u_ia)=g_i\phi(a)$.
Clearly, $\psi$ is a bijection, so we show that it is a
homomorphism.
Let $k, k^{\prime}\in K$. We have $k=u_ia$ and
$k^{\prime}=u_ja^{\prime}$, for some $1\leq i, j\leq n$ and $a,
a^{\prime}\in A$. We have
$$
\psi(kk^{\prime})=\psi(u_iu_j(u_j^{-1}au_j)a^{\prime}).
$$
Note that by (3), we can write $u_j^{-1}au_j$ as $\phi^{-1}(p)$, for
some suitable $p\in B$. Also by (1), we have
$u_iu_j=u_{r(i,j)}\phi^{-1}(q_{ij})$, hence
\begin{eqnarray*}
\psi(kk^{\prime})&=&\psi(u_{r(i,j)}\phi^{-1}(q_{ij})\phi^{-1}(p)a^{\prime})\\
&=&g_{r(i,j)}q_{ij}p\phi(a^{\prime}).
\end{eqnarray*}
On the other hand, we have
\begin{eqnarray*}
\psi(k)\psi(k^{\prime})&=&g_i\phi(a)g_j\phi(a^{\prime})\\
&=&g_ig_j(g_j^{-1}\phi(a)g_j)\phi(a^{\prime})\\
&=&g_ig_jp\phi(a^{\prime})\\
&=&g_{r(i,j)}q_{ij}p\phi(a^{\prime}).
\end{eqnarray*}
This completes the proof.
\end{proof}
As a final note, we should say that the same result also can be
proved for Lie algebras.
{\bf Acknowledgement} The author would like to thank M. Kuzucuoglu
for comments and suggestions.
|
\section{Introduction}
In this paper, $K$ is a
field of characteristic zero and $K^*$ is its group of units, and the following notation is fixed:
\begin{itemize}
\item $P_n:= K[x_1, \ldots , x_n] $
is a polynomial algebra and
$S_n:= K[[x_1, \ldots , x_n]]$ is the algebra of power series over $K$, $\gm := (x_1, \ldots , x_n)$, and $S_n^*$ is the group of units of $S_n$,
\item
$\mS_n:=\AutKalgc (S_n)$ is the group of continuous (with respect to the $\gm$-adic topology) automorphisms of $S_n$ and $\mS_n^c:=\{ \s \in \mS_n\, | \, \CJ (\s ) \in K\}$ where $\CJ (\s )$ is the Jacobian of $\s$,
\item $\der_1:=\frac{\der}{\der x_1}, \ldots , \der_n:=\frac{\der}{\der
x_n}$ are the partial derivatives ($K$-linear derivations) of
$S_n$,
\item
$\gsn:=\Der_K(S_n) =\bigoplus_{i=1}^nS_n\der_i$
is the Lie
algebras of $K$-derivations of $S_n$ where $[\der , \d ]:= \der \d -\d \der $, and $D_n:= \Der_K(P_n) =\bigoplus_{i=1}^n P_n\der_i$,
\item $\CD_n:=\bigoplus_{i=1}^n K\der_i$,
\item
$H_1:=x_1\der_1, \ldots , H_n:=x_n\der_n$,
\item for a derivation $\der = \sum_{i=1}^n a_i\der_i\in \gsn$, $\divv (\der ) := \sum_{i=1}^n\frac{\der a_i}{\der x_i}$ is the {\em divergence} of $\der$,
\item $\divn0 :=\{ \der \in D_n \, | \, \divv (\der ) =0\}$ and $\Divn0 :=\{ \der \in \gsn \, | \, \divv (\der ) =0\}$ are the Lie algebras of polynomial, respectively, formally analytic vector fields (derivations) with zero divergence,
\item $\mfGn :=\Aut_{{\rm Lie}}(\divn0 )$ and $\mfGhn :=\Aut_{{\rm Lie}}(\Divn0 )$,
\item $\divnc :=\{ \der \in D_n \, | \, \divv (\der ) \in K\}$ and $\Divnc :=\{ \der \in \gsn \, | \, \divv (\der ) \in K\}$ are the Lie algebras of polynomial, respectively, formally analytic vector fields (derivations) with constant divergence,
\item $\mfGnc :=\Aut_{{\rm Lie}}(\divnc )$ and $\mfGhnc :=\Aut_{{\rm Lie}}(\Divnc )$,
\end{itemize}
{\bf The groups of automorphisms of the Lie algebras $\divn0$ and $\divnc$}. Let $\Sh_1:=\{ s_\mu \in \AutKalg (K[x])\, | \, s_\mu (x) = x+\mu, \; \mu \in K\}$.
\begin{theorem}\label{16Mar13
\cite{Rudakov-1986}, \cite{Bav-Aut-Div} $\mfGn \simeq \begin{cases}
G_1/\Sh_1\simeq K^*& \text{if }n=1,\\
G_n& \text{if }n\geq 2.\\
\end{cases}$
\end{theorem}
Theorem \ref{16Mar13} was announced in \cite{Rudakov-1986} where a sketch of the proof is given based on a study of certain Lie subalgebras of $\divn0$ of finite codimension. The proof in \cite{Bav-Aut-Div} is based on completely different ideas.
\begin{theorem}\label{A16Mar13
\cite{Bav-Aut-Div} $\mfGnc \simeq G_n$.
\end{theorem}
{\bf The groups of automorphisms of the Lie algebras $\Divn0$ and $\Divnc$}.
\begin{theorem}\label{FD16Mar13
\cite {Rudakov-1969}, \cite{Rudakov-1986} $\mfGhn \simeq \mS_n^c $ for $n\geq 2$.
\end{theorem}
The aim of the paper is to prove the following theorem.
\begin{theorem}\label{FDA16Mar13
$\mfGhnc \simeq \begin{cases}
G_1& \text{if }n=1,\\
\mS_n^c& \text{if }n\geq 2.\\
\end{cases}$
\end{theorem}
{\em Proof}. For $n=1$, $\mathfrak{Div}_1^c = K\der_1\oplus KH_1= \mathfrak{div}_1^c$ and so
$\widehat{\mathbf{G}}_1^c=\mathbf{G}_1^c=G_1$, by Theorem \ref{A16Mar13}. So, let $n\geq 2$.
$\noindent $
(i) $\mS_n^c \subseteq \mfGhnc$ via the group monomorphism (Theorem \ref{FD16Mar13} and Theorem \ref{FB16Mar13}):
$$\mS_n^c\ra \mfGhnc, \;\; \s \mapsto \s : \der \mapsto \s (\der ):=\s \der \s^{-1}.$$
(ii) $\Divn0 = [\Divnc , \Divnc ]$: The equality follows from the fact that $\Divn0$ is a simple Lie algebra which is an ideal of the Lie algebra $\Divnc$ and $\Divnc = \Divn0 \oplus KH_1$.
$\noindent $
(iii) The short exact sequence of group homomorphisms
$$ 1\ra F:=\Fix_{\mfGhnc}(\Divn0 ) \ra \mfGhnc
\stackrel{{\rm res}}{\ra}\mfGhn\ra 1$$
is exact (by (i) and Theorem \ref{FD16Mar13}) where res $:\s \mapsto \s|_{\Divn0}$ is the restriction map, see (ii).
$\noindent $
(iv) Since $\mfGhn = \mS_n^c$ (Theorem \ref{FD16Mar13}) and $\mS_n^c\subseteq \mfGhnc$ (by (i)), the short exact sequence splits
\begin{equation}\label{FDG=GF}
\mfGhnc \simeq \mfGhn\ltimes F.
\end{equation}
(v) $ F=\{ e\}$ (Lemma \ref{Fa18Mar13}). Therefore, $\mfGhnc \simeq \mS_n^c$. $\Box $
$\noindent $
The Lie algebra $\Divnc$ is a topological Lie algebra with respect to the $\gm$-adic topology, i.e., the set $\{ \gm^i \Divnc\}_{i\in \N}$ is a base of open neighbourhoods of zero. Let
$\widehat{\mathbf{G}}_{n,top}^c$ be group of automorphisms of the topological Lie algebra $\mathfrak{Div}_n^c$. Clearly, $\widehat{\mathbf{G}}_{n,top}^c\subseteq\mfGhnc$. The inverse inclusion follows from Theorem \ref{FDA16Mar13}.
\begin{corollary}\label{aFDA16Mar13
$\widehat{\mathbf{G}}_{n,top}^c=\mfGhnc$.
\end{corollary}
$\noindent $
{\bf The group $\mS_n$}.
Every continuous automorphism $\s \in \mS_n$ is uniquely determined by the elements
$$x_1':=\s (x_1), \; \ldots , \; x_n':=\s (x_n)$$
that necessarily (by the continuity of $\s $) belong to the maximal ideal $\gm$ of the algebra $S_n$, and for all series $f=f(x_1, \ldots , x_n)\in S_n$, $\s (f) = f(x_1', \ldots , x_n')$.
Let $M_n(S_n)$ be the algebra of $n\times n$ matrices over $S_n$. The matrix $J(\s) := (J(\s )_{ij}) \in M_n(S_n)$, where $J(\s )_{ij} =\frac{\der x_j'}{\der x_i}$, is called the {\em Jacobian matrix} of $\s$ and its determinant $\CJ (\s ) :=\det \, J(\s)$ is called the {\em Jacobian} of $\s$. So, the $j$'th column of $J(\s )$ is the {\em gradient} $\grad \, x_j':=(\frac{\der x_j'}{\der x_1}, \ldots , \frac{\der x_j'}{\der x_n})^T$ of the series $x_j'$. Then the derivations
$$\der_1':= \s \der_1\s^{-1}, \; \ldots , \; \der_n':= \s\der_n\s^{-1}$$ are the partial derivatives of $S_n$ with respect to the variables $x_1', \ldots , x_n'$,
\begin{equation}\label{ddp=dxi}
\der_1'=\frac{\der}{\der x_1'}, \; \ldots , \; \der_n'=\frac{\der}{\der x_n'}.
\end{equation}
Every derivation $\der \in \gsn$ is a unique sum $\der = \sum_{i=1}^n a_i\der_i$ where $a_i = \der *x_i\in S_n$. Let $\der := (\der_1, \ldots , \der_n)^T$ and $ \der' := (\der_1', \ldots , \der_n')^T$ where $T$ stands for the transposition. Then
\begin{equation}\label{dp=Jnd}
\der'=J(\s )^{-1}\der , \;\; {\rm i.e.}\;\; \der_i'=\sum_{j=1}^n (J(\s )^{-1})_{ij} \der_j\;\; {\rm for }\;\; i=1, \ldots , n.
\end{equation}
In more detail, if $\der'=A\der $ where $A= (a_{ij})\in M_n(S_n)$, i.e. $\der_i=\sum_{j=1}^n a_{ij}\der_j$. Then for all $i,j=1, \ldots , n$,
$$\d_{ij}= \der_i'*x_j'=\sum_{k=1}^na_{ik}\frac{\der x_j'}{\der x_k}$$
where $\d_{ij}$ is the Kronecker delta function. The equalities above can be written in the matrix form as $AJ(\s) = 1$ where $1$ is the identity matrix. Therefore, $A= J(\s )^{-1}$.
For all $\s, \tau \in \mS_n$,
\begin{equation}\label{SJst=JsJ}
J(\s \tau ) = J(\s )\cdot \s (J(\tau )).
\end{equation}
By taking the determinants of both sides of (\ref{SJst=JsJ}), we have a similar equality of the Jacobians: for all $\s , \tau \in \mS_n$.
\begin{equation}\label{SJst=JsJ1}
\CJ (\s \tau ) = \CJ(\s )\cdot \s (\CJ(\tau )).
\end{equation}
By putting $\tau = \s^{-1}$ in (\ref{SJst=JsJ}) and (\ref{SJst=JsJ1}) we see that $\ J(\s ) \in \GL_n(S_n)$, $\CJ (\s ) \in S_n^*$, and
\begin{equation}\label{SJsm}
J(\s^{-1}) = \s^{-1} (J(\s )^{-1}),
\end{equation}
\begin{equation}\label{SJsm1}
\CJ (\s^{-1} ) = \s^{-1} (\CJ(\s )^{-1}).
\end{equation}
\begin{eqnarray*}
\mS_n &=& \{ \s \in \End_{K-{\rm alg},c}(S_n)\, | \, \CJ (\s ) \in S_n^* \} \\
&=& \{ \s \in \End_{K-{\rm alg},c}(S_n)\, | \, \s (x) = Ax+\cdots , A= (a_{ij}) \in \GL_n(K)\} ,
\end{eqnarray*}
that is $\s (x_i) = \sum_{j=1}^n a_{ij} x_j+\cdots $ where the three dots mean smaller terms ($\cdots \in \gm^2$).
$\noindent $
\begin{lemma}\label{a15Oct13
For all $\s \in \mS_n^c$, $$ \sum_{j=1}^n \der_j*(J(\s )^{-1})_{ij}=0 \;\; {\rm for } \;\; i=1, \ldots , n.$$
\end{lemma}
{\it Proof}. By (\ref{dp=Jnd}), $\der_i' = \sum_{i=1}^n (J(\s )^{-1})_{ij}\der_j$. By Theorem \ref{FD16Mar13}, we have the result.
$\Box $
$\noindent $
{\bf The divergence commutes with automorphisms $\mS_n^c$}. The following theorem shows that the divergence commute with automorphisms $\mS_n^c$, i.e. the divergence map $\divv : \gsn \ra S_n$ is an $\mS^c_n$-module homomorphism.
\begin{theorem}\label{FB16Mar13
For all $\s \in \mS^c_n$ and $\der \in \gsn$,
$$ \divv (\s (\der ))= \s (\divv (\der )).$$
\end{theorem}
{\it Proof}. Let $\der = \sum_{i=1}^n a_i\der_i$ where $a_i\in S_n$. Then
$ \der' = \s \der \s^{-1} =\sum_{i=1}^n \s (a_i) \der_i'$ where, by (\ref{dp=Jnd}), $\der_i' = \sum_{j=1}^n (J(\s )^{-1})_{ij}\der_j$. Now, by Lemma \ref{a15Oct13},
\begin{eqnarray*}
\divv (\der') &=& \sum_{i,j=1}^n \der_j*((J(\s)^{-1})_{ij}\s (a_i)) = \sum_{i=1}^n (\sum_{j=1}^n \der_j *(J(\s )^{-1})_{ij})\cdot \s (a_i))+ \sum_{i=1}^n \sum_{j=1}^n (J(\s)^{-1})_{ij} \der_j*\s (a_i) \\
&=& \sum_{i=1}^n \der_i'*\s (a_i) = \sum_{i=1}^n \s \der_i \s^{-1}\s (a_i) = \s ( \sum_{i=1}^n \der_i (a_i)) = \s (\divv (\der )). \;\; \Box
\end{eqnarray*}
\begin{lemma}\label{Fa18Mar13
$ \Fix_{\mfGhnc}(\Divn0 ) =\{ e\}$ for $n\geq 2$.
\end{lemma}
{\it Proof}. Let $\s \in F:= \Fix_{\mfGhnc}(\Divn0 )$, $H_1':= \s (H_1) , \ldots , H_n':= \s (H_n)$. Since $\Divnc= \Divn0\oplus KH_i$, $i=1, \ldots , n$, it suffices to show that $ \s (H_i) = H_i$ for $i=1, \ldots , n$. For $i\neq j$, $\s (H_i-H_j) = H_i-H_j$, and so $d:= H_i'-H_i= H_j'-H_j$. For all $i=1, \ldots , n$,
$$ [\der_i , d] = \s ([\der_i, H_i]) - [\der_i, H_i]= \s (\der_i ) - \der_i = \der_i - \der_i =0.$$ So, $d\in C_{\Divnc}(\CD_n ) = \CD_n$ (since $C_{\gsn}(\CD_n ) = \CD_n$) and $d= \sum_{i=1}^n \l_i\der_i$ for some $\l_i\in K$ where $C_\CG (\CH ):= \{ g\in \CG \, | \, [g, \CH ]=0\}$ is the centralizer of a subset $\CH$ of a Lie algebra $\CG$. The elements $H_1'=H_1+d, \ldots , H_n'=H_n+d$ commute hence $d=0$. Therefore, $\s = e$. $\Box $
$\noindent $
$${\bf Acknowledgements}$$
The work is partly supported by the Royal Society and EPSRC.
\small{ |
\section{Introduction}
The Standard-Model (SM) value of the muon anomaly can be calculated with
sub-parts-per-million precision. The comparison between the measured and
the SM prediction provides a test of the completeness of the
Standard Model. At present, there appears to be a three- to four-standard
deviation between these two values, which has motivated extensive theoretical
and experimental work on the hadronic contributions to the muon anomaly.
A lepton ($\ell = e,\,\mu,\,\tau$) has a magnetic moment which is along its
spin, given by the relationship
\begin{equation} \vec \mu_{\ell} =g_{\ell}
\frac{Qe}{2m_{\ell}}\vec{s}\,,\qquad \underbrace{g_{\ell} =2}_{\mbox{\rm
\small Dirac}}(1+a_\ell), \qquad a_\ell = \frac{g_\ell -2}{2}
\end{equation}
where $Q = \pm 1$, $e>0 $ and $m_\ell$ is the lepton mass.
Dirac theory predicts
that $g \equiv 2$, but experimentally, it is known to be greater than 2.
The small number $a$, the anomaly, arises from quantum fluctuations,
with the largest contribution coming from the mass-independent
single-loop diagram in
Fig.~\ref{fg:schwinger}(a). With his famous calculation
that obtained $a = (\alpha/2 \pi) = 0 .00116\cdots$,
Schwinger~\cite{Schwinger:1948} started an ``industry'',
which required Aoyama, Hayakawa, Kinoshita and Nio
to calculate more than 12,000 diagrams to evaluate the tenth-order
(five loop) contribution~\cite{Aoyama1:2012}.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.7\textwidth,angle=0]{sch-vp-X}
\caption{The Feynman graphs for: (a) The lowest-order (Schwinger)
contribution to the lepton anomaly ; (b) The vacuum polarization
contribution, which is one of five fourth-order,
$( \alpha/\pi)^2$, terms; (c) The schematic contribution of new particles
$X$ and $Y$ that couple to the muon.
\label{fg:schwinger}}
\end{center}
\end{figure}
The interaction shown in Fig.~\ref{fg:schwinger}(a) is a
chiral-changing, flavor-conserving process, which gives it a special
sensitivity to possible new physics~\cite{Miller:2012,Stoeckinger2010}.
Of course heavier particles can also contribute, as indicated by the
diagram in Fig.~\ref{fg:schwinger}(c). For example,
$X = W^\pm$ and $Y = \nu_\mu$, along with
$X = \mu$ and $Y = Z^0$, are the lowest-order weak contributions.
In the
Standard-Model, $a_\mu$ gets measureable
contributions from QED, the strong interaction, and
from the electroweak interaction,
\begin{equation}
a^{SM} = a^{QED} + a^{Had} + a^{Weak} .
\end{equation}
In this document we present the latest evaluations of the SM
value of $a_\mu$, and then discuss expected improvements
that will become available over the next five to seven years.
The uncertainty in this evaluation is dominated by the
contribution of virtual hadrons in loops. A worldwide effort is under way to
improve on these hadronic contributions. By the time that
the Fermilab muon $(g-2)$ experiment, E989, reports a result later in this
decade, the uncertainty should be significantly reduced. We emphasize that
the existence of E821 at Brookhaven motivated significant work over the past
thirty years that permitted more than an order of magnitude improvement in the
knowledge of the hadronic contribution. Motivated by
Fermilab E989 this work continues, and another factor of two
improvement could be possible.
Both the electron~\cite{Hanneke08}
and muon~\cite{Bennett:2006I} anomalies have been measured very precisely:
\begin{eqnarray}
a_{e}^{exp} &=& 1\,159\,652\,180.73 \,(28)\times 10^{-12} \ \ {\pm 0.24\,
\rm{ppb}} \\
a_{\mu}^{exp} &=& 1\,165\,920\,89\, (63) \times 10^{-11} \ \ {\pm 0.54\,
\rm{ppm}}
\end{eqnarray}
While the electron anomaly has been measured to $\simeq 0.3$~ppb (parts per
billion)~\cite{Hanneke08}, it is significantly less
sensitive to heavier physics,
because the relative
contribution of heavier virtual particles to the muon anomaly goes
as $(m_\mu/m_e)^2 \simeq 43000$. Thus the lowest-order hadronic contribution
to $a_e$ is~\cite{Davier-LM}:
$ a_e^{\rm had,LO} = (1.875\pm0.017)~10^{-12}$, 1.5~ppb
of $a_e$. For the muon the hadronic contribution is $\simeq 60$~ppm
(parts per million). So with much less precision, when
compared with the electron, the measured muon anomaly is sensitive to mass
scales in the several hundred GeV region. This not only includes
the contribution of the $W$ and $Z$ bosons, but perhaps
contributions from new, as yet undiscovered, particles such as the
supersymmetric partners of the electroweak gauge bosons (see
Fig.~\ref{fg:schwinger}(c)).
\section{Summary of the Standard-Model Value of $a_\mu$}
\subsection{QED Contribution}
The QED contribution to $a_\mu$ is well
understood. Recently the four-loop QED contribution has been updated
and the full five-loop contribution has been calculated~\cite{Aoyama1:2012}.
The present QED value is
\begin{equation}
a_{\mu}^{\rm QED} = 116~584~718.951~(0.009)( 0.019 )(0.007)(.077)
\times 10^{-11}
\end{equation}
where the uncertainties are from the lepton mass ratios, the eight-order
term, the tenth-order term, and the value of
$\alpha$ taken from the $^{87}$Rb atom
$\alpha^{-1}(Rb) = 137.035\,999\,049(90)$ [0.66 ppb].~\cite{Bouchendira:2011}.
\subsection{Weak contributions}
The electroweak contribution (shown in Fig.~\ref{fg:weak})
is now calculated through two
loops~\cite{CKM96,PPdR95,CKM95,CMV03,CM-LM,Gnendiger:2013}.
The one loop result
\begin{eqnarray}
\label{EW1}
a_{\mu}^{\mbox{\rm\tiny
EW(1)}} &\!\! =\!\! &
\frac{G_{\mbox\tiny F}}{\sqrt{2}}\frac{m_{\mu}^2}{8\pi^2}
\left\{\underbrace{\frac{10}{3}}_{W} +
\underbrace{\frac{1}{3}(1\!-\!4\sin^{2}\theta_{W})^2-\frac{5}{3}}_{Z} \right.
\nonumber
\\
& + & \!\!\left. {\mathcal O} \!\left(\!
\frac{m_{\mu}^2}{M_{Z}^2}\log\frac{M_{Z}^2}{m_{\mu}^2}\!\right)
\!+\!\frac{m_{\mu}^2}{M_{H}^2}
\int_{0}^{1}\!\! dx
\frac{2x^2
(2-x)}{1-x+{\frac{m_{\mu}^2}{M_{H}^2}}x^2}
\!\right\} \nonumber \\
& = & 194.8\times 10^{-11}\,,
\end{eqnarray}
was calculated by five separate groups~\cite{weak} shortly after the
Glashow-Salam-Weinberg theory was shown by 't Hooft to be
renormalizable. Due to the small Yukawa coupling of the Higgs boson to the
muon, only the $W$ and $Z$ bosons contribute at a measurable level
in the lowest-order electroweak term.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.8\textwidth]{weak_line}
\end{center}
\caption{Weak contributions to the muon anomalous magnetic moment.
Single-loop contributions from
(a) virtual $W$ and (b) virtual $Z$ gauge bosons.
These two contributions enter with
opposite sign, and there is a partial cancellation.
The two-loop contributions fall into three categories:
(c) fermionic loops which involve the coupling of the gauge bosons to quarks,
(d) bosonic loops which appear as corrections to the one-loop diagrams, and
(e) a new class of diagrams involving the Higgs
where {\sl G} is the longitudinal component of the
gauge bosons. See Ref. \cite{Miller:07I} for details.
The $\times$
indicates the photon from the magnetic field.}
\label{fg:weak}
\end{figure}
The two-loop electroweak contribution (see Figs.~\ref{fg:weak}(c-e)),
which is negative~\cite{CKM95,PPdR95,CKM96,CMV03},
has been re-evaluated using the
LHC value of the Higgs mass~\cite{Gnendiger:2013}.
The total electroweak contribution is
\begin{equation}
a_{\mu}^{\rm EW} = (153.6\pm 1.0) \times 10^{-11}
\label{eq:ew}
\end{equation}
where the error comes from hadronic effects in the second-order
electroweak diagrams with quark triangle loops, along with unknown three-loop
contributions\cite{CMV03,KPPdeR02,Vain03,KPPdeR04}. The leading
logs for the next-order term have been shown to be
small~\cite{CMV03,Gnendiger:2013}. The weak contribution is about 1.3~ppm of the
anomaly, so the experimental uncertainty on $a_\mu$ of $\pm
0.54$~ppm now probes the weak scale of the Standard Model.
\subsubsection{Hadronic contribution}
The hadronic contribution to $a_{\mu}$ is about 60 ppm of the total
value.
The lowest-order diagram shown in Fig.~\ref{fg:had}(a)
dominates this contribution and its error, but the hadronic light-by-light
contribution Fig.~\ref{fg:had}(e) is also important. We discuss both of
these contributions below.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=32pc]{had_line}
\end{center}
\caption{The hadronic contribution to the muon anomaly, where the
dominant contribution comes from the lowest-order diagram (a). The
hadronic light-by-light contribution is shown in (e).}
\label{fg:had}
\end{figure}
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.54\textwidth]{hadpro_ee_ISR_bw}
\end{center}
\caption{(a) The ``cut'' hadronic vacuum polarization diagram; (b) The
$e^+ e^-$ annihilation into hadrons; (c) Initial state radiation
accompanied by the production of hadrons.}
\label{fg:hadpro}
\end{figure}
The energy scale for the virtual hadrons is of order $ m_{\mu} c^2$,
well below the perturbative region of QCD. However it can be
calculated from the
dispersion relation shown pictorially in Fig.~\ref{fg:hadpro},
\begin{equation}
a_{\mu}^{\rm had;LO}=\left({\alpha m_{\mu}\over 3\pi}\right)^2
\int^{\infty} _{m_{\pi}^2} {ds \over s^2}K(s)R(s), \quad
{\rm where} \quad
R\equiv{ {\sigma_{\rm tot}(e^+e^-\to{\rm hadrons})} \over
\sigma_{\rm }(e^+e^-\to\mu^+\mu^-)}\, ,
\label{eq:dispersion}
\end{equation}
using
the measured cross sections for $e^+ e^- \rightarrow {\rm hadrons}$
as input, where $K(s)$ is a kinematic factor ranging from { 0.4} at
$s= m_\pi^2$ to $0$ at $s = \infty$ (see Ref.~\cite{Miller:07I}).
This dispersion relation relates the bare cross
section for $e^+e^- $ annihilation into hadrons
to the hadronic vacuum polarization contribution
to $a_{\mu}$.
Because the integrand
contains a factor of $s^{-2}$, the values
of $R(s)$ at low energies (the $\rho$ resonance) dominate the determination of
$a_{\mu}^{\rm had;LO}$, however at the level of precision needed, the
data up to 2~GeV are very important.
This is shown in Fig.~\ref{fg:had-cont}, where the
left-hand chart gives the relative contribution to the integral for the
different energy regions, and the right-hand gives the contribution to the
error squared on the integral.
The contribution is dominated by the two-pion final
state, but other low-energy multi-hadron cross sections are also important.
These data for $e^+e^-$ annihilation to hadrons are also important as
input into the determination of $\alpha_{QED}(M_Z)$ and other electroweak
precision measurements.
\begin{figure}[h!]
\vskip0.2in
\begin{center}
\includegraphics[width=0.5\textwidth]{a-had-LO-pie-small}
\end{center}
\caption{Contributions to the dispersion integral for different
energy regions, and to the associated error
(squared)
on the dispersion integral in that energy region. Taken from
Hagiwara et al.~\cite{Hagiwara:2011}.}
\label{fg:had-cont}
\end{figure}
Two recent analyses~\cite{Davier11,Hagiwara:2011}
using the $e^+e^- \to hadrons$ data obtained:
\begin{eqnarray}
a_{\mu}^{\rm had;LO}&=&\left(6~923\pm 42\right)\times
10^{-11}\,, \label{eq:hvp1-pub}\\
a_{\mu}^{\rm had;LO}&=&\left(6~949\pm 43\right)\times
10^{-11}\,,
\label{eq:hvp2-pub}
\end{eqnarray}
respectively.
Important earlier global analyses include
those of Hagiwara et al.~\cite{HMNT07}, Davier, et al.,~\cite{Davier07},
Jegerlehner and Nyffler~\cite{Jegerlehner:2009ry}.
In the past, hadronic $\tau$ spectral functions and CVC, together with
isospin breaking corrections have been used to calculate the hadronic
contribution~\cite{Alemany:1998,Davier11}.
While the original predictions showed a discrepancy between
$e^+e^-$ and $\tau$ based evaluations, it
has been shown that after $\gamma$-$\rho$ mixing is taken into account, the
two are compatible~\cite{Jegerlehner:2011}. Recent evaluations based on a
combined $e^+e^-$ and $\tau$ data fit using the Hidden Local Symmetry (HLS)
model have come to similar conclusions and result in values for $a_\mu^{\rm
HVP}$ that are smaller than the direct evaluation without the HLS
fit~\cite{Benayoun:2012a,Benayoun:2012b}.
The most recent evaluation of the next-to-leading
order hadronic contribution shown in
Fig.~\ref{fg:had}(b-d), which can also be determined
from a dispersion relation, is~\cite{Hagiwara:2011}
\begin{equation}
\label{eq:exphvpnlo}
a_{\mu}^{\rm had;NLO}=(-98.4\pm 0.6_{\exp}\pm0.4_{\rm rad}\ )\times 10^{-11}\,.
\end{equation}
\subsubsection{Hadronic light-by-light contribution}
The hadronic light-by-light contribution (HLbL)
cannot at present be determined from data, but rather must be
calculated using hadronic models that correctly reproduce
properties of QCD. This contribution is shown below in
Fig.~\ref{fg:HLBL}(a). It is dominated by the long-distance contribution
shown in Fig.~~\ref{fg:HLBL}(b). In fact, in the so called chiral limit
where the mass gap between the pseudoscalars ( Goldstone-like) particles and
the other hadronic particles (the $\rho$ being the lowest vector
state in Nature) is
considered to be large, and to leading order in the $1/N_c$--expansion ($N_c$
the number of colors), this contribution has been calculated
analytically~\cite{KNPdeR02} and provides a long-distance constraint to model
calculations. There is also a short-distance constraint from the operator
product expansion (OPE) of two electromagnetic currents which, in specific
kinematic conditions, relates the light-by-light scattering amplitude to an
Axial-Vector-Vector triangle amplitude for which one has a good theoretical
understanding~\cite{MV04}.
Unfortunately, the two asymptotic QCD constraints mentioned above are not
sufficient for a full model independent evaluation of the HLbL
contribution. Most of the last decade calculations found in the literature
are compatible with the QCD chiral and large-$N_c$ limits. They all
incorporate the $\pi^0$-exchange contribution modulated by $\pi^0 \gamma^*
\gamma^*$ form factors correctly normalized to the Adler, Bell-Jackiw
point-like coupling.
They differ, however, on whether or not they satisfy the particular OPE
constraint mentioned above, and
in the shape of the vertex form factors which follow from the different
models.
\begin{figure}[h!]
\begin{center}
\subfigure[]
{ \includegraphics[width=0.35\textwidth]{HLL}}
\subfigure[]
{ \includegraphics[width=0.35\textwidth]{HLBL-pi0}}
\end{center}
\caption{(a)The Hadronic Light-by contribution. (b) The pseudoscalar meson contribution. }
\label{fg:HLBL}
\end{figure}
A synthesis of the model contributions, which was agreed to by
authors from each of the leading groups that have been working in this field,
can be found in ref.~\cite{Prades:2010}\footnote{This compilation is
generally referred
to as the ``Glasgow Consensus'' since it grew out of a workshop in
Glasgow in 2007. http://www.ippp.dur.ac.uk/old/MuonMDM/ }. They
obtained
\begin{equation} a_\mu^{\rm HLbL} = (105 \pm 26) \times 10^{-11} \, .
\label{eq:HLbL} \end{equation}
An alternate evaluation~\cite{Jegerlehner:2009ry,Nyffeler:2009tw}
obtained, $a_{\mu}^{\rm HLbL} =
(116\pm 40)\times 10^{-11}$, which agrees well with the Glasgow
Consensus~\cite{Prades:2010}.
Additional work on this contribution is underway on a number of fronts,
including on the lattice. A workshop was held in March 2011 at the
Institute for Nuclear Theory in Seattle~\cite{INT:2011} which brought
together almost all of the interested experts. This will be followed by a
workshop at the Mainz Institute for Theoretical Physics in April 2014.
One important point should be stressed here. The main physics of the hadronic
light-by-light scattering contribution is well understood. In fact, but for
the sign error unraveled in 2002, the theoretical predictions for
$a_\mu^{\rm HLbL}$ have been relatively stable for more than ten
years\footnote{A calculation using a Dyson-Schwinger
approach~\cite{Goecke:2011} initially
reported a much larger value for the HLbL contribution. Subsequently a
numerical mistake was found. These authors are continuing this work, but
the calculation is still incomplete.}.
\subsection{Summary of the Standard-Model Value and Comparison with
Experiment}
We determine the SM value using the new QED calculation from
Aoyama~\cite{Aoyama1:2012}; the electroweak from Ref.~\cite{Miller:2012}, the
hadronic light-by-light contribution from the ``Glasgow
Consensus''~\cite{Prades:2010}; and lowest-order hadronic contribution from
Davier, et al.,~\cite{Davier11}, or Hagiwara et al.,~\cite{Hagiwara:2011},
and the higher-order hadronic contribution from Ref.~\cite{Hagiwara:2011}.
A summary of these values is given in
Table~\ref{tb:SMvalue}.
\begin{table}[h!]
\begin{center}
\caption{Summary of the Standard-Model contributions to the muon anomaly. Two
values are quoted because of the two recent evaluations of the lowest-order
hadronic vacuum polarization.}
{\small
\begin{tabular}{lr}
\hline \hline {\sc\small } &
{\sc\small Value $(\times \, 10^{-11})$ units }
\\ \hline
QED ($\gamma + \ell$) & $116\,584\,718.951\pm 0.009\pm 0.019 \pm 0.007 \pm
0.077_{\alpha}$ \\
HVP(lo)~\cite{Davier11} &
$6\,923 \pm
42$\\
HVP(lo)~\cite{Hagiwara:2011} &
$6\,949 \pm
43$\\
HVP(ho)~\cite{Hagiwara:2011} & $-98.4\pm 0.7 $ \\
HLbL& $105\pm 26 $ \\
EW & $154\pm 1 $ \\
\hline
Total SM~\cite{Davier11} & $116\,591\,802 \pm 42_{\mbox{\rm \tiny H-LO}}
\pm 26_{\mbox{\rm \tiny H-HO}} \pm 2_{\mbox{\rm \tiny other}} \, (\pm 49_{\mbox{\rm \tiny tot}}) $\\
Total SM~\cite{Hagiwara:2011} & $116\,591\,828 \pm 43_{\mbox{\rm \tiny H-LO}}
\pm 26_{\mbox{\rm \tiny H-HO}} \pm 2_{\mbox{\rm \tiny other}} \, (\pm 50_{\mbox{\rm \tiny tot}}) $\\
\hline\hline\
\end{tabular}}
\label{tb:SMvalue}
\end{center}
\end{table}
This SM value is to be compared with the combined $a_\mu^+$ and
$a_\mu^-$ values from E821~\cite{Bennett:2006I} corrected for the
revised value of $\lambda= \mu_\mu/\mu_p$ from Ref~\cite{CODATA08I},
\begin{equation}
a_\mu^{\rm E821} = (116\,592\,089 \pm 63)\times 10^{-11}\quad
{\rm (0.54\,ppm)},\\
\end{equation}
which give a difference of
\begin{eqnarray}
\Delta a_{\mu}({\rm E821-SM}) &=& (287 \pm 80 ) \times 10^{-11} \
\cite{Davier11} \\
&=& (261 \pm 78 ) \times 10^{-11} \ \cite{Hagiwara:2011}
\label{eq:Delta}
\end{eqnarray}
depending on which evaluation of the lowest-order hadronic contribution that
is used~\cite{Davier11,Hagiwara:2011}.
\begin{figure}[h!]
\begin{center}
\subfigure[]
{\includegraphics[width=0.4\textwidth]{ans10-final-e+e-DHMYZ}}
\subfigure[]
{\includegraphics[width=0.55\textwidth]{Citations-15Oct2013.png}}
\end{center}
\caption{(a)Measurements of $a_\mu$ from CERN and BNL E821.
The vertical band is the SM value using the hadronic contribution from
Ref.~\cite{Davier11} (see Table~\ref{tb:SMvalue}). (b) Citations to the E821
papers by year.
}
\label{fg:SM-Exp}
\end{figure}
This comparison between the experimental values and the present
Standard-Model value is shown
graphically in Fig. \ref{fg:SM-Exp}.
The lowest-order hadronic evaluation of Ref.~\cite{Benayoun:2012b}
using the hidden local symmetry model results in a
difference between experiment and theory that ranges
between 4.1 to 4.7$\sigma$.
This difference of { 3.3} to 3.6 standard deviations is tantalizing, but we
emphasize that whatever the final agreement between the measured and
SM value turns out to be, it will have significant implications on
the interpretation of new phenomena that might be found at the LHC
and elsewhere. Because of the power of $a_\mu$ to constrain, or point to,
speculative models of New Physics, the E821 results have been highly cited,
with more than 2450 citations to date.
\section{Expected Improvements in the Standard-Model Value}
The present uncertainty on the theoretical value is dominated by the hadronic
contributions~\cite{Davier11,Hagiwara:2011} (see Table~\ref{tb:SMvalue}).
The lowest-order contribution determined from $e^+e^-
\rightarrow {\rm hadrons}$ data using a dispersion relation is theoretically
relatively straightforward.
It does require the combination of data sets from different
experiments.
The only significant theoretical uncertainty comes from radiative
corrections, such as vacuum polarization (running $\alpha$), along with
initial and
final state radiation effects, which are needed to obtain the correct hadronic
cross section at the required level of precision.
This was a problem for the older data sets. In the analysis of the
data collected over the past 15 years, which now dominate the determination
of the hadronic contribution, the
treatment of radiative corrections has been significantly
improved. Nevertheless, an additional uncertainty due to the treatment of these
radiative corrections in the older data sets has been estimated to be of
the order of $20\times 10^{-11}$~\cite{Hagiwara:2011}. As more data
become available, this uncertainty will be significantly reduced.
There are two methods that have
been used to measure the hadronic cross sections:
The energy scan (see Fig.~\ref{fg:hadpro}(b)), and
using initial state radiation with a fixed beam
energy to measure the cross section for energies below the total
center-of-mass energy of the colliding beams (see Fig.~\ref{fg:hadpro}(c)).
Both are being employed in the next round of measurements.
The data from the new experiments that are now underway at VEPP-2000 in
Novosibirsk
and BESIII in Beijing, when combined with the analysis of
existing multi-hadron final-state data from BaBar and Belle,
should significantly reduce the uncertainty on the
lowest-order hadronic contribution.
The hadronic-light-by-light contribution does not lend itself to
determination by a dispersion relation. Nevertheless there are some
experimental data that can help to pin down related amplitudes and to
constrain form factors used in the model calculations.
\subsection{Lowest-order Hadronic Contribution}
Much experimental and theoretical work is going on worldwide to
refine the hadronic contribution. The theory of $(g-2)$, relevant
experiments to determine the hadronic contribution, including work on the
lattice, have featured prominently
in the series of tau-lepton workshops and PHIPSI workshops which are
held in alternate years.
Over the development period of Fermilab E989, we expect further
improvements in the SM-theory evaluation. This projection is based
on the following developments:
\subsubsection {Novosibirsk}
The VEPP2M machine has been upgraded to VEPP-2000. The maximum
energy has been increased from $\sqrt{s} =1.4$ GeV to 2.0 GeV.
Additionally, the SND detector has been upgraded and the CMD2
detector was replaced by the much-improved CMD3 detector. The
cross section will be measured from threshold to 2.0~GeV using an
energy scan, filling in the energy region between 1.4~GeV, where the
previous scan ended, up to 2.0~GeV, the lowest energy point reached
by the BES collaboration in their measurements. See
Fig.~\ref{fg:had-cont} for the present contribution to the overall
error from this region. Engineering runs began in 2009, and data
collection started in 2011. So far two independent energy scans between
1.0 and 2.0~GeV were performed in 2011 and 2012. The peak luminosity
of $3\times 10^{31} {\rm cm}^{-2} {\rm s}^{-1}$
was achieved, which was limited by the positron
production rate. The new injection facility, scheduled
to be commissioned during
the 2013-2014 upgrade, should permit the
luminosity to reach $10^{32} {\rm cm}^{-2} {\rm s}^{-1}$ .
Data collection resumed
in late 2012 with a new energy scan over
energies below 1.0~GeV. The goal of experiments at VEPP-2000 is to
achieve a systematic error 0.3-0.5\% in the $\pi^+\pi^-$ channel,
with negligible
statistical error in the integral. The high statistics, expected at VEPP-2000,
should allow a detailed comparison of the measured cross-sections
with ISR results at BaBar and DA$\Phi$NE. After the upgrade, experiments at
VEPP-2000 plan to take a large amount
of data at 1.8-2~GeV, around the $N\bar{N}$ threshold.
This will permit ISR data with the beam energy of 2~GeV, which is
between the PEP2 energy at
the $\Upsilon (4S)$ and the 1~GeV $\phi$ energy at the DA$\Phi$NE
facility in Frascati. The dual ISR and scan approach will provide an
important cross check on the two central methods used to determine
the HVP.
\subsubsection{The BESIII Experiment}
The BESIII experiment at the Beijing tau-charm
factory BEPC-II has already collected several femtobarns of integrated
luminosity
at various centre-of-mass energies in the range 3 - 4.5 GeV.
The ISR program includes cross section measurements of:
$e^+e^- \to \pi^+\pi^-$, $e^+e^- \to \pi^+\pi^-\pi^0$,
$e^+e^- \to \pi^+\pi^-\pi^0\pi^0$ --
the final states most relevant to $(g-2)_\mu$.
Presently, a data sample of 2.9 fb$^{-1}$
at $\sqrt{s}=3.77$ GeV is being analyzed, but new data at $\sqrt{s} > 4$ GeV can be used for
ISR physics as well and will double the statistics.
Using these data, hadronic invariant masses from threshold up to
approximately 3.5 GeV can be accessed at BESIII.
Although the integrated luminosities
are orders of magnitude lower compared to the $B$-factory
experiments BaBar and BELLE, the ISR method at BESIII still
provides competitive statistics. This is due to the fact that the
most interesting mass range for the HVP contribution of $(g-2)_\mu$, which is
below approximately 3 GeV,
is very close to the centre-of-mass energy of the collider BEPC-II and hence
leads
to a configuration where
only relatively low-energetic ISR photons need to be emitted, providing
a high ISR cross section.
Furthermore, in contrast to the $B$ factories, small angle ISR photons can be
included in the event selection for kinematic reasons which leads to a very
high overall geometrical acceptance.
Compared to the KLOE experiment, background from final state radiation (FSR)
is reduced significantly as
this background decreases with increasing center of mass energies of the collider.
BESIII is aiming for a precision measurement of the ISR $R$-ratio
$R_{\rm ISR}= N(\pi\pi\gamma)/N(\mu\mu\gamma)$ with a precision of
about 1\%. This requires an excellent pion-muon separation, which
is achieved by training a multi-variate neural network.
As a preliminary result, an absolute cross section
measurement of the reaction $e^+e^- \rightarrow \mu^+\mu^- \gamma$
has been achieved, which
agrees with the QED prediction within 1\% precision.
Moreover, at BESIII a new energy scan campaign is planned
to measure the inclusive $R$ ratio in the energy range between 2.0 and 4.6 GeV. Thanks to
the good performance of the BEPC-II accelerator and the BESIII detector
a significant improvement upon the existing BESII measurement can be expected.
The goal is to arrive at an inclusive
$R$ ratio measurement with about 1\% statistical and 3\% systematic precision
per scan point.
\subsubsection{Summary of the Lowest-Order Improvements from Data}
A substantial amount of new $e^+e^-$ cross section
data will become available over the next
few years. These data have the potential to significantly reduce the
error on the lowest-order hadronic contribution.
These improvements
can be obtained by reducing the uncertainties of the hadronic
cross-sections from 0.7\% to 0.4\% in the region below 1 GeV and from 6\%
to 2\% in the region between 1 and 2 GeV as shown in
Table~\ref{tb:LO-impr}.
\begin{table}[h]
\begin{center}
\renewcommand{\arraystretch}{1.4}
\setlength{\tabcolsep}{1.6mm}
{\footnotesize
\begin{tabular}{|c|c|c|c|c|}
\hline
& $\delta (\sigma)/\sigma$ present &$\delta a_{\mu}$present
& $\delta (\sigma)/\sigma$ future &$\delta a_{\mu}$future \\
\hline
$\sqrt{s}<1$~GeV & 0.7\% & 33 & 0.4\% & 19 \\
$1<\sqrt{s}<2$~GeV & 6\% & 39 & 2\% & 13 \\
$\sqrt{s}>2$~GeV & & 12 & & 12 \\
\hline
total & & 53 & & 26 \\
\hline
\end{tabular}
}
\caption{\label{tb:LO-impr} Overall uncertainty of the cross-section
measurement required to get the reduction of uncertainty on $a_{\mu}$ in units
$10^{-11}$ for
three regions of $\sqrt{s}$ (from Ref.~\cite{Jegerlehner:2008zz}).}
\end{center}
\end{table}
\subsubsection {Lattice calculation of the Lowest-Order HVP:}
With computer power presently available, it is possible for lattice QCD calculations to make important
contributions to our knowledge of the lowest-order hadronic
contribution. Using several different discretizations for QCD, lattice groups
around the world are computing the
HVP~\cite{Burger:2013jya,DellaMorte:2012cf,Blum:2013qu,Boyle:2011hu,Aubin:2006xv}
(see also several recent talks at Lattice 2013 (Mainz). The varied
techniques have different systematic errors, but in the continuum limit
$a\to0$ they should all agree. Many independent calculations provide a
powerful check on the lattice results, and ultimately the dispersive ones
too.
Several groups are now performing simulations with physical light quark
masses on large boxes, eliminating significant systematic errors. So called
quark-disconnected diagrams are also being calculated, and several recent
theory advances will help to reduce systematic errors associated with fitting
and the small $q^{2}$
regime~\cite{deDivitiis:2012vs,DellaMorte:2012cf,Blum:2012uh,Aubin:2012me,Feng:2013xsa,Aubin:2013daa}. While
the HVP systematic errors are well understood, significant computational
resources are needed to control them at the $\sim1\%$ level, or
better. Taking into account current resources and those expected in the next
few years, the lattice-QCD uncertainty on $a_\mu$(HVP), currently at the
$\sim 5$\%-level, can be reduced to 1 or 2\%
within the next few years. This is already
interesting as a wholly independent check of the dispersive
results for $a_\mu$(HVP). With increasing experience and computer power, it
should
be possible to compete with the $e^+e^-$ determination of
$a_\mu$(HVP) by the end of the decade, perhaps sooner with additional
technical advances.
\subsection{The Hadronic Light-by-Light contribution}
There are two approaches to improving the HLbL contribution: Measurements of
$\gamma^*$ physics at BESIII and KLOE; Calculations on the lattice.
In addition to the theoretical work on the HLbL,
the KLOE detector at DA$\Phi$NE has been upgraded with a tagging system to
detect the final leptons in the reaction $e^+ e^- \to e^+ e^-
\gamma^*\gamma^*$.
Thus a coincidence between the scattered electrons and a $\pi^0$
would provide information
on $\gamma^* \gamma^* \rightarrow \pi^0$~\cite{KLOE-2},
and will provide experimental constraints on the
models used to calculate the hadronic light-by-light
contribution~\cite{Babusci:2011bg}.
Any experimental information on the neutral pion lifetime and the
transition form factor is important in order to constrain the models
used for calculating the pion-exchange contribution (see
Fig.~\ref{fg:HLBL}(b)). However, having
a good description, e.g.\ for the transition form factor, is only
necessary, not sufficient, in order to uniquely determine
$a_\mu^{\mathrm{HLbL};\pi^0}$. As stressed in
Ref.~\cite{Jegerlehner_off-shell}, what enters in the calculation of
$a_\mu^{{\rm HLbL}; \pi^0}$ is the fully off-shell form factor ${\cal
F}_{{\pi^0}^*\gamma^*\gamma^*}((q_1 + q_2)^2, q_1^2, q_2^2)$ (vertex
function), where also the pion is off-shell with 4-momentum $(q_1 +
q_2)$. Such a (model dependent) form factor can for instance be
defined via the QCD Green's function $\langle VVP \rangle$, see
Ref.~\cite{Nyffeler:2009tw} for details. The form factor with on-shell
pions is then given by ${\cal F}_{\pi^0\gamma^*\gamma^*}(q_1^2, q_2^2)
\equiv {\cal F}_{{\pi^0}^*\gamma^*\gamma^*}(m_\pi^2, q_1^2, q_2^2)$.
Measurements of the transition form factor ${\cal
F}_{\pi^0\gamma^\ast\gamma}(Q^2) \equiv {\cal
F}_{{\pi^0}^\ast\gamma^\ast\gamma^\ast}(m_{\pi}^2, -Q^2, 0)$ are in
general only sensitive to a subset of the model parameters and do not
permit the reconstruction the full off-shell form factor.
For different models, the effects of the off-shell pion can vary a
lot. In Ref.~\cite{Nyffeler:2009tw} the off-shell lowest meson dominance
(LMD) plus vector meson dominance (LMD+V) form factor was
proposed and the estimate $a_{\mu; {\rm LMD+V}}^{{\rm HLbL}; \pi^0} =
(72 \pm 12) \times 10^{-11}$ was obtained (see also Ref.~\cite{KN_EPJC_2001}). The error estimate comes
from the variation of all model parameters, where the uncertainty of
the parameters related to the off-shellness of the pion completely
dominates the total error. In contrast to the off-shell LMD+V model,
many other models, e.g.\ the VMD model or constituent quark models, do
not have these additional sources of uncertainty related to the
off-shellness of the pion. These models often have only very few
parameters, which can all be fixed by measurements of the transition
form factor or from other observables. Therefore, for such models, the
precision of the KLOE-2 measurement can dominate the total precision of
$a_\mu^{\mathrm{HLbL};\pi^0}$.
Essentially all evaluations of the pion-exchange contribution use for
the normalization of the form factor, ${\cal
F}_{{\pi^0}^*\gamma^*\gamma^*}(m_\pi^2, 0, 0) = 1 / (4 \pi^2
F_\pi)$, as derived from the Wess-Zumino-Witten (WZW) term. Then the
value $F_\pi = 92.4~\mbox{MeV}$ is used without any error attached to it,
i.e. a value close to $F_\pi = (92.2 \pm 0.14)~\mbox{MeV}$, obtained
from $\pi^+ \to \mu^+ \nu_\mu(\gamma)$~\cite{Nakamura:2010zzi}. If one
uses the decay width $\Gamma_{\pi^0 \to \gamma\gamma}$ for the
normalization of the form factor, an additional source of uncertainty
enters, which has not been taken into account in most
evaluations~\cite{Nyffeler:2009uw}.
Until recently, the experimental world average of
$\Gamma^{PDG}_{\pi^0 \to \gamma\gamma}= 7.74\pm 0.48$ eV~\cite{Nakamura:2010zzi} was only known
to 6.2\% precision.
Due to the poor agreement between the existing data, the PDG error of the
width average is inflated (scale factor of $2.6$) and it gives an additional motivation for new precise
measurements. The PrimEx Collaboration, using a Primakoff effect
experiment at JLab, has achieved 2.8\% fractional
precision~\cite{Larin:2010kq}.
There are plans to further reduce the
uncertainty to the percent level.
Though theory and experiment are in a fair agreement, a better experimental precision is needed to really test the theory predictions.
\subsubsection{Impact of KLOE-2 measurements on $a_\mu^{\mathrm{HLbL};\pi^0}$}
For the new data taking of the KLOE-2 detector,
which is expected to start by the end of 2013,
new small angle
tagging detectors have been installed along DA$\Phi$NE beam line.
These ``High Energy Tagger'' detectors \cite{het}
offer the possibility to study
a program of $\gamma\gamma$ physics through the process $e^+e^-\to
e^+\gamma^* e^-\gamma^* \to e^+e^- X$.
In Ref.~\cite{KLOE-2_impact} it was shown that planned measurements at
KLOE-2 could determine the $\pi^0\to\gamma\gamma$ decay width to 1\%
statistical precision and the $\gamma^\ast\gamma\to\pi^0$ transition
form factor ${\cal F}_{\pi^0\gamma^\ast\gamma}(Q^2)$ for small
space-like momenta, $0.01~\mbox{GeV}^2 \leq Q^2 \leq
0.1~\mbox{GeV}^2$, to 6\% statistical precision in each bin. The
simulations have been performed with the Monte-Carlo program
EKHARA~\cite{EKHARA} for the process $e^+ e^- \to e^+ e^- \gamma^*
\gamma^* \to e^+ e^- \pi^0$, followed by the decay $\pi^0 \to
\gamma\gamma$ and combined with a detailed detector simulation. The
results of the simulations are shown in
Figure~\ref{simulation_FF_data}. The KLOE-2 measurements will allow to
almost directly measure the slope of the form factor at the origin and
check the consistency of models which have been used to extrapolate
the data from larger values of $Q^2$ down to the origin.
With the decay width $\Gamma_{\pi^0\to\gamma\gamma}^{\rm PDG}$
$[\Gamma_{\pi^0\to\gamma\gamma}^{\rm PrimEx}]$ and current data for
the transition form factor ${\cal F}_{\pi^0\gamma^\ast\gamma}(Q^2)$,
the error on $a_\mu^{{\rm HLbL}; \pi^0}$ is $\pm 4 \times 10^{-11}$
[$\pm 2 \times 10^{-11}$], not taking into account the uncertainty
related to the off-shellness of the pion. Including the simulated
KLOE-2 data reduces the error to $\pm (0.7 - 1.1) \times
10^{-11}$.
\begin{figure}[h]
\centerline{\includegraphics[width=.7\textwidth]{plot_FF_rebinned}}
\caption{Simulation of KLOE-2 measurement of $F(Q^2)$ (red triangles)
with statistical errors for $5$~fb$^{-1}$, corresponding to one year
of data taking. The dashed line is the $F(Q^2)$ form factor
according to the LMD+V model~\cite{Nyffeler:2009tw,KN_EPJC_2001}, the solid
line is $F(0) = 1/(4\pi^2 F_\pi)$ given by the Wess-Zumino-Witten term.
Data~\cite{TFF_data}
from CELLO (black crosses) and CLEO (blue stars) at high $Q^2$ are
also shown for illustration.}
\label{simulation_FF_data}
\end{figure}
\subsubsection{BESIII Hadronic light-by-light contribution}
Presently, data taken at $\sqrt{s}=$3.77 GeV are being analyzed
to measure the form factors of the
reactions $\gamma^*\gamma \to X$, where
$X=\pi^0, \eta, \eta^\prime, 2\pi$.
BESIII has launched a program of two-photon interactions with
the primary goal to measure the transition form factors (TFF)
of pseudoscalar mesons as well as of the two-pion system in the
spacelike domain. These measurements are carried out in the
single-tag mode, i.e. by tagging one of the two beam leptons at large
polar angles
and by requiring that the second lepton is scattered at small
polar angles. With these kinematics
the form factor, which in general depends on the
virtualities of the two photons, reduces to $F(Q^2)$,
where $Q^2$ is the negative momentum transfer of
the tagged lepton. At BESIII, the process $\gamma \gamma^* \to \pi^0$, which
is known to play
a leading contribution in the HLbL correction to $(g-2)$, can be measured
with unprecedented
precision in the
$Q^2$ range between 0.3 GeV$^2$ and 4 GeV$^2$. In the future
BESIII will also embark on untagged as well as double-tag measurements, in
which either
both photons are quasi-real or feature a high virtuality. The goal is to
carry out
this program for the final states $\pi^0, \eta, \eta^\prime, \pi\pi$.
It still needs to be proven that the
small angle detector,
which recently has been installed close to the BESIII beamline,
can be used for the two-photon program.
\subsubsection {Lattice calculation of Hadronic Light-by-Light Scattering:}
Model calculations show that the hadronic light-by-light (HLbL) contribution
is roughly $(105
\pm 26)\times 10^{-11}$, $\sim 1$ ppm of $a_{\mu}$. Since the error
attributed to this estimate is difficult to reduce, a modest, but
first principles
calculation on the lattice would have a large impact.
Recent progress towards this goal has been reported~\cite{Blum:2013qu}, where
a non-zero signal (statistically speaking) for a part of the amplitude
emerged in the same ball-park as the model estimate. The result was computed
at non-physical quark mass, with other systematic errors mostly
uncontrolled. Work on this method, which treats both QED and QCD interactions
non-perturbatively, is continuing. The next step is to repeat the calculation
on an ensemble of gauge configurations that has been generated with
electrically charged sea quarks (see the poster by Blum presented at Lattice
2013). The charged sea quarks automatically include the quark disconnected
diagrams that were omitted in the original calculation and yield the complete
amplitude. As for the HVP, the computation of the HLbL contribution requires
significant resources which are becoming available. While only one group has
so far attempted the calculation, given the recent interest in the HVP
contribution computed in lattice QCD and electromagnetic corrections to
hadronic observables in general, it seems likely that others will soon enter
the game. And while the ultimate goal is to compute the HLbL contribution to
10\% accuracy, or better,
we emphasize that a lattice calculation with even a solid 30\% error would
already be very interesting. Such a result, while not guaranteed, is not out
of the question during the next 3-5 years.
\section{Summary}
The muon and electron anomalous magnetic moments are among, if not the most
precisely measured and calculated quantities in all of physics. The
theoretical uncertainty on the Standard-Model contribution
to $a_\mu$ is $\simeq 0.4$~ppm,
slightly smaller than the experimental error from BNL821. The new Fermilab
experiment, E989, will achieve a precision of 0.14~ppm. While the hadronic
corrections will most likely not reach that level of precision, their
uncertainty will be significantly decreased. The lowest-order contribution
will be improved by new data from Novosibirsk and BESIII. On the timescale of
the first results from E989, the lattice will also become relevant.
The hadronic light-by-light contribution will also see significant
improvement. The measurements at Frascati and at BESIII will provide
valuable experimental input to constrain the model calculations. There is
hope that the lattice could produce a meaningful result by 2018.
We summarize possible near-future improvements in the table below.
Since it is difficult to project the improvements in the hadronic
light-by-light contribution, we assume a conservative improvement: That
the large amount of work
that is underway
to understand this contribution, both
experimentally and on the lattice, will support the level of
uncertainty assigned in the ``Glasgow Consensus''.
With these improvements, the overall uncertainty on $\Delta a_{\mu}$ could be
reduced by a factor 2. In case the central value would remain the same, the
statistical significance would become 7-8 standard deviations, as it can be
seen in Fig.~\ref{tab:g-2a}.
\begin{figure}[h!]
\begin{center}
\begin{minipage}{\textwidth-.2in}
\begin{minipage}{\textwidth/2}
{
\begin{tabular}{|c|c c |c|}
\hline
Error & \cite{Davier11} & \cite{Hagiwara:2011} & Future \\
\hline
$\delta a_{\mu}^{\rm
SM}$ & 49 & 50 & 35 \\
\hline
$\delta a_\mu^{\rm HLO}$ & 42 & 43 & 26 \\
$\delta a_\mu^{\rm HLbL}$ & 26 &26 & 25 \\
\hline
$\delta (a_\mu^{\rm EXP} - a_\mu^{\rm SM})$ & 80 & 80 & 40 \\
\hline
\end{tabular}
}
\end{minipage}
\hskip0.2in
\begin{minipage}{\textwidth/2}
\begin{center}
\includegraphics[width=.9\textwidth]{g-2new-v3}
\end{center}
\end{minipage}
\end{minipage}
\end{center}
\caption{\label{tab:g-2a} Estimated uncertainties $\delta a_{\mu}$ in units
of $10^{-11}$ according to Refs.~\cite{Davier11,Hagiwara:2011} and
(last column) prospects for improved precision in the $e^+e^-$
hadronic cross-section measurements.
The final row projects the uncertainty on
the difference with the Standard Model, $\Delta a_{\mu}$.
The figure give the comparison between $a_{\mu}^{\rm SM}$ and $a_{\mu}^{\rm EXP}$.
DHMZ is Ref.~\cite{Davier11}, HLMNT is Ref.~\cite{Hagiwara:2011};
``SMXX'' is the same central value with a
reduced error as
expected by the improvement on the hadronic cross section measurement (see
text); ``BNL-E821 04 ave.'' is the current experimental value
of $a_{\mu}$; ``New (g-2) exp.'' is the same central value with a fourfold
improved precision as planned by the future (g-2)
experiments at Fermilab and J-PARC.
}
\end{figure}
Thus the prognosis is excellent that the results from E989 will clarify
whether the measured value of $a_\mu$ contains contributions from outside of
the Standard Model. Even if there
is no improvement on the hadronic error, but
the central theory and experimental values remain the same, the significance
of the difference would be over $5\sigma$. However, with the worldwide
effort to improve on the Standard-Model value, it is most likely that the
comparison will be even more convincing.
\input{SMWhitepaper-bib.tex}
\end{document}
|
\section{Introduction}
\label{s:intro}
Although interstellar clouds have ionization fractions of only 10$^{-4}$--10$^{-8}$, ionic species are very useful to probe physical conditions in such clouds \citepads{2012RPPh...75f6901L}.
In diffuse clouds ($A_V < 1$), the main ion source is UV photoionization of carbon, while in dense clouds ($A_V > 1$), cosmic-ray ionization of hydrogen is the dominant ionization mechanism \citepads{2007ARA&A..45..339B}.
Proton transfer reactions of interstellar H$_3^+$\ with abundant species such as CO and N$_2$\ lead to HCO$^+$\ and N$_2$H$^+$, which are widely observed in the interstellar medium.
Such stable ionic species are useful as tracers of the interaction of interstellar gas with magnetic fields
(Houde et al \citeyearads{2004ApJ...604..717H}, Schmid-Burgk et al \citeyearads{2004A&A...419..949S}), whereas ions which react rapidly with H$_2$\ trace other parameters such as the gas density and the ionization rate.
At temperatures ${_<\atop{^\sim}}$250~K, the formation of interstellar H$_2$O\ in the gas phase proceeds through a series of ion-molecule reactions.
After charge transfer of H$^+$\ or H$_3^+$\ to O, repeated reactions of O$^+$\ with H$_2$\ produce OH$^+$, H$_2$O$^+$, and finally H$_3$O$^+$, which upon dissociative recombination with a free electron produces H$_2$O.
The OH$^+$\ and H$_2$O$^+$\ ions are well known from the spectra of comets where they appear as photodissociation products of H$_2$O\ \citepads{1950ApJ...111..530S,1974A&A....31..123H}.
In the interstellar medium, H$_2$O\ and H$_3$O$^+$\ have been known for decades \citepads{1992ApJ...399..533P}, but observation of the intermediate products OH$^+$\ and H$_2$O$^+$\ had to await the launch of ESA's {\it Herschel} Space Observatory \citepads{2010A&A...518L...1P}.
Strong absorption in rotational lines of interstellar OH$^+$\ and H$_2$O$^+$\ is seen with Herschel on many lines of sight in our Galaxy (Gerin et al \citeyearads{2010A&A...518L.110G}, Ossenkopf et al \citeyearads{2010A&A...518L.111O}) and even some beyond \citepads{2010A&A...521L...1W}.
In addition, electronic absorption lines of OH$^+$\ have been reported in sensitive near-UV spectra of several diffuse interstellar clouds \citepads{2010ApJ...719L..20K}.
The Herschel data, as well as the single OH$^+$\ line observed from the ground toward the Galactic Center source SgrB2 \citepads{2010A&A...518A..26W}, imply large column densities of OH$^+$\ and H$_2$O$^+$.
The hydrogen in the absorbing clouds thus cannot be purely in atomic form, because no OH$^+$\ and H$_2$O$^+$\ would be produced, nor in purely molecular form, because all OH$^+$\ and H$_2$O$^+$\ would react into H$_3$O$^+$\ and H$_2$O.
Using models of UV-irradiated interstellar clouds (PDRs), the observed abundances of OH$^+$\ and H$_2$O$^+$\ can be used to infer the relative fractions of hydrogen in atomic and molecular forms \citepads{2010A&A...521L..10N}, which itself traces the ionization rates of the clouds \citepads{2012ApJ...754..105H}.
While the interpretation of interstellar H$_n$O$^+$\ absorption is reasonably well understood, lines of OH$^+$\ and H$_2$O$^+$\ have also been observed in emission from the nuclei of several active galaxies, most famously Mrk~231 \citepads{2010A&A...518L..42V}.
The large dipole moments and small reduced masses of the H$_n$O$^+$\ ions imply high line frequencies and large radiative decay rates, so that collisional excitation of their rotational levels requires extremely high densities and line emission is not expected to be observable.
Understanding this phenomenon benefits from finding a Galactic source of H$_n$O$^+$\ line emission, which can be studied in more detail than extragalactic nuclei.
This paper presents the first observation of OH$^+$\ line emission toward a source within our Galaxy: the Orion Bar.
Due to its brightness and nearly edge-on geometry, this PDR is well-suited to observe physical and chemical changes in the gas as a function of depth into the cloud, as the intensity of UV irradiation by the Trapezium stars decreases \citepads[e.g.,]{2009A&A...498..161V}.
The Orion Bar is also notable as the only known Galactic source of \new{interstellar} HF line emission \citepads{2012A&A...537L..10V}.
In this case, the proximity of this region (420~pc: Menten et al \citeyearads{2007A&A...474..515M}, Hirota et al \citeyearads{2007PASJ...59..897H}) allows us to resolve the H$_n$O$^+$\ line emission both spatially and spectrally.
We use non-LTE radiative transfer models and PDR thermo-chemical models to interpret our results.
\section{Observations}
\label{s:obs}
\begin{table}
\caption{Observed lines.}
\label{t:lines}
\begin{tabular}{ccccc}
\hline \hline
\noalign{\smallskip}
\multicolumn{2}{c}{Molecule / Transition} & Frequency & $E_{\rm up}$ & $A_{\rm ul}$ \\
& & GHz & K & s$^{-1}$ \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
OH$^+$\ & $1_0$--$0_1$ $F$=1/2--3/2 & \phantom{1}909.159 & 43.6 & 0.011 \\
OH$^+$\ & $1_2$--$0_1$ $F$=5/2--3/2\tablefootmark{a} & \phantom{1}971.804 & 46.7 & 0.033 \\
OH$^+$\ & $1_1$--$0_1$ $F$=3/2--3/2 & 1033.119 & 49.6 & 0.018 \\
p-H$_2$O$^+$\ & $1_{10}$--$1_{01}$ & \phantom{1}607.227 & 59.2 & 0.006 \\
o-H$_2$O$^+$\ & $1_{11}$--$0_{00}$ & 1115.186 & 53.6 & 0.027 \\
H$_3$O$^+$\ & $0_0^-$--$1_0^+$ & \phantom{1}984.709 & 54.7 & 0.023 \\
H$_3$O$^+$\ & $1_1^-$--$1_1^+$ & 1655.831 & 79.5 & 0.055 \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
\end{tabular}
\tablefoot{The strongest hyperfine component is listed, unless otherwise noted. \\
\tablefoottext{a}{Blend with the $F = 3/2 -1/2$ hyperfine component. The frequency and $E_{\rm up}$ are averages, while the $A_{\rm ul}$ is the sum.}}
\end{table}
The CO$^+$ peak ($\alpha_\mathrm{J2000}=\rm{05^h35^m20.6^s}$, $\delta_\mathrm{J2000}=-05^\circ 25'14''$) in the Orion Bar \citepads{1995A&A...296L...9S} has been observed as a spectral scan over the full HIFI range as part of the HEXOS (Herschel observations of EXtra-Ordinary Sources) guaranteed-time key program \citepads{2010A&A...521L..20B} using the HIFI instrument \citepads{2010A&A...518L...6D} of the Herschel Space Observatory \citepads{2010A&A...518L...1P}.
This paper uses data from HIFI bands 1b (H$_2$O$^+$ $1_{10}$--$1_{01}$), 3b (OH$^+$ $1_0$--$0_1$), 4a (OH$^+$ $1_2$--$0_1$ and $1_1$--$0_1$), 5a (H$_2$O$^+$ $1_{11}$--$0_{00}$), and 6b (H$_3$O$^+$\ $1_1^-$--$1_1^+$).
These observations were carried out in 2011 March and April in load chop mode with a redundancy of 4, except that frequency switching was used in band 5.
On-source integration times are $\approx$50~s for most spectra except OH$^+$\ $1_0$--$0_1$ (20~s) and H$_2$O$^+$\ $1_{11}$--$0_{00}$ (190~s).
\new{The ObsIDs of the spectra, without the leading 1342, are 215923 for Band 1b, 216380 for Band 3b, 218628 for Band 4a, 194666 for Band 5a and 218426 for Band 6b}.
Table~\ref{t:lines} lists the frequencies of the lines as well as other spectroscopic parameters, which have been taken from the CDMS database \citepads{2001A&A...370L..49M}\footnote{\tt http://www.astro.uni-koeln.de/cdms/}.
The size of the telescope beam for these observations is 19--23$''$ FWHM, corresponding to 9000 AU or 0.04 pc, except for the H$_2$O$^+$\ $1_{10}$--$1_{01}$ line where it is 35$''$, and for the H$_3$O$^+$\ $1_1^-$--$1_1^+$ line where it is 15$''$.
The WBS (Wide-Band Spectrometer) was used as backend, covering 4 GHz bandwidth in four 1140~MHz subbands at 1.1~MHz resolution.
The velocity calibration of HIFI data is accurate to $\sim$0.5 kms$^{-1}$ or better.
The data were reduced with HIPE \citepads{2010ASPC..434..139O} pipeline version 6.0, using the task \textit{doDeconvolution} for the sideband deconvolution, while further analysis was done in the CLASS package.
In addition to the HIFI spectral scans, a $115''\times65''$ area centered on $\alpha$ = 05:35:20.81, $\delta$ = --05:25:17.1 with a position angle of 145$^\circ$ was mapped in the OH$^+$ $1_2-0_1$ transition with HIFI, in On-The-Fly (OTF) mapping mode with position-switch reference, using a total integration time of 20 min.
\new{The noise level of the map is 0.22~K per 0.7~km\,s$^{-1}$\ channel and its ObsID is 218216.}
The \new{fully sampled} map was reduced with HIPE 6.0 and exported to CLASS for further analysis.
\new{We compare these data to a map of the same area in the CO 10-9 line at $\nu$ = 1151.985 GHz ($E_{\rm up}$ = 304~K), also observed within the HEXOS program under ObsID 217736, with a noise level of 0.56~K per 1.0~km\,s$^{-1}$\ channel.}
\section{Results}
\label{s:res}
\begin{figure}[t]
\centering
\includegraphics[width=9cm,angle=0]{aa22164f1.eps}
\caption{Line profiles of the OH$^+$\ $1_2$--$0_1$ (top) and $1_1$--$0_1$ (bottom) transitions, observed with HIFI toward the \new{CO$^+$\ peak of the} Orion Bar, \new{after smoothing to 0.6~km\,s$^{-1}$\ resolution.}}
\label{f:profi}
\end{figure}
\begin{table}
\caption{Line parameters measured from the HIFI spectra.}
\label{t:pars}
\begin{tabular}{ccccccc}
\hline \hline
\noalign{\smallskip}
Line & $\int$ $T_A^*$\ $\Delta${\it V}\ & $V_{\rm LSR}$\ & $\Delta${\it V}\ & rms \\
& K\,km\,s$^{-1}$\ & km\,s$^{-1}$\ & km\,s$^{-1}$\ & mK \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
OH$^+$\ $1_0$--$0_1$ & $<$0.80 & ... & ... & \phantom{1}58 \\
OH$^+$\ $1_2$--$0_1$ & 2.3(3) & 10.9(3) & 5.1(6) & 133 \\
OH$^+$\ $1_1$--$0_1$ & 1.2(1) & 10.1(3) & 4.3(5) & \phantom{1}88 \\
H$_2$O$^+$\ $1_{10}$--$1_{01}$ & $<$0.18 & ... & ... & \phantom{1}20 \\
H$_2$O$^+$\ $1_{11}$--$0_{00}$ & $<$1.90 & ... & ... & 112 \\
H$_3$O$^+$\ $0_0^-$--$1_0^+$ & $<$1.17 & ... & ... & \phantom{1}78 \\
H$_3$O$^+$\ $1_1^-$--$1_1^+$ & $<$2.27 & ... & ... & \phantom{1}90 \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
\end{tabular}
\tablefoot{Numbers in parentheses are error bars in units of the last decimal. Noise levels in the last column are for a channel width of 1~MHz. Upper limits in column 2 are for $\Delta${\it V}\ = 4.3\,km\,s$^{-1}$.
}
\end{table}
\begin{figure}[t]
\centering
\includegraphics[width=6cm,angle=-90]{aa22164f2a.eps}
\bigskip
\includegraphics[width=8cm,angle=0]{aa22164f2b.eps}
\smallskip
\includegraphics[width=6.5cm,angle=-90]{aa22164f2c.eps}
\caption{Spatial distribution of the OH$^+$\ $1_2$--$0_1$ emission, integrated between $V_{\rm LSR}$\ = +6 and +10 km\,s$^{-1}$\ (top), between +8 and +12 km\,s$^{-1}$\ (middle), and between +10 and +14 km\,s$^{-1}$\ (bottom). Greyscale levels start at 0.6 K\,km\,s$^{-1}$\ and increase by 0.2 K\,km\,s$^{-1}$. Contours of CO 10--9 emission are at 5, 15, ...95\% of the peak intensity in the respective velocity channel. \newest{Vibrationally excited H$_2$\ peaks near the North-West edge of the CO emission (cf. Fig. 1 of Nagy et al 2013)}.}
\label{f:maps}
\end{figure}
\subsection{Line profiles}
\label{ss:prof}
Figure~\ref{f:profi} presents the line profiles of the OH$^+$\ lines at 971 and 1033 GHz, as observed with HIFI toward the Orion Bar.
Although hints of OH$^+$\ and H$_2$O$^+$\ emission have been seen toward the high-mass protostar W3 IRS5 \citepads{2010A&A...521L..35B}, this is the first time that a pure emission profile is seen toward a Galactic source.
As the line profiles appear single-peaked at our sensitivity and spectral resolution, we have fitted a Gaussian model to extract the line parameters.
Table~\ref{t:pars} presents the results of these Gaussian fits, and also reports our upper limits to other lines of OH$^+$, H$_2$O$^+$\ and H$_3$O$^+$\ from our spectral survey of the Orion Bar.
We have also searched for the O$^+$ excited fine structure ($^2$D $J$ = 5/2 $\to$ 3/2) line near 593.7 GHz \citepads{2004ApJ...610..813B,2004ApJ...606..605S}, leading to an upper limit on $T_A^*$\ of 30~mK rms per 0.5 km\,s$^{-1}$\ channel.
No continuum emission is detected in the spectra, down to upper limits ranging from $T_{mb}^{SSB}$ = 0.1\,K at 607~GHz to 0.3\,K at 1655~GHz, which is consistent with the SPIRE measurements of \citetads{2012A&A...541A..19A}.
The difference between the measured central velocities and FWHM widths of the two detected OH$^+$\ lines is mostly due to hyperfine blending of the $1_2$--$0_1$ line.
We thus adopt the measured parameters of the $1_1$--$0_1$ line as the best estimate of the central velocity and FWHM width of the OH$^+$\ emitting gas in the Orion Bar.
While the central velocity of this line of 10.1\,km\,s$^{-1}$\ is in good agreement with the value of 10.0$\pm$0.2 km\,s$^{-1}$\ from ground-based observations of other molecular species toward this source, its width of 4.3\,km\,s$^{-1}$\ is much larger than the value of 1.7$\pm$0.3 km\,s$^{-1}$\ measured for the dense gas deep inside ($A_V {_>\atop{^\sim}} 1$) the Bar (Hogerheijde et al \citeyearads{1995A&A...294..792H}, Leurini et al \citeyearads{2006A&A...454L..47L}).
On the other hand, its width is smaller than the value of $\approx$5~km\,s$^{-1}$\ measured in species such as HF which trace the surface ($A_V {_<\atop{^\sim}} 0.1$) of the PDR \citepads{2012A&A...537L..10V}, and more similar to the width of the [C II] line \citep{2013A&A...550A..57O}.
Based on the line width, the OH$^+$\ emission thus seems to originate from close to the PDR surface ($0.3 < A_V < 0.5$) where CH$^+$\ and SH$^+$\ peak as well in the chemical model of \citetads{2013A&A...550A..96N}.
\new{In contrast, the non-hydride reactive molecular ions CO$^+$, SO$^+$ and HOC$^+$ have smaller line widths \citepads[2--3\,km\,s$^{-1}$,]{2003A&A...406..899F}. These species have a different formation channel and thus a different sensitivity to the abundances of atomic H and electrons.}
\subsection{Spatial distribution}
\label{ss:map}
Figure~\ref{f:maps} presents our maps of the OH$^+$\ 971 GHz line emission, as observed with HIFI toward the Orion Bar, integrated over three velocity ranges, shown in greyscale.
The contours show the CO 10--9 line \new{emission}, integrated over the same velocity ranges.
The OH$^+$\ emission around the central velocity (middle panel) is seen to be extended over at least an arc minute (25,000~AU or 0.12~pc) on the sky, and to roughly follow the structure of the Bar seen in CO 10--9 and other molecular tracers \citepads[e.g.,]{2009A&A...498..161V}.
The emission from the Bar is concentrated in two clumps: one peaking near the map center at $V_{\rm LSR}$\ $\approx$10 km\,s$^{-1}$, and another peaking 20--25$''$ to the East near $V_{\rm LSR}$\ = 12 km\,s$^{-1}$, which is seen most pronounced in the figure's bottom panel.
\new{This second peak lies $\approx$10$''$ South-East of the CO 10--9 peak at the same velocity, which coincides with H$^{13}$CN clumps 2 and 3 from \citetads{2003ApJ...597L.145L}. Since OH$^+$\ is not expected deeper into the PDR than CO and HCN, we suggest that the location of the second OH$^+$\ clump marks a deviation from pure plane-parallel geometry.}
The map of low-velocity OH$^+$\ in the top panel of Figure~\ref{f:maps} shows a third clump which lies to the North-West of the Bar, and is connected to it by a bridge of fainter OH$^+$\ emission.
This `perpendicular' emission is also seen in CO 10--9 and other tracers of the Bar surface such as \new{OH 119\,$\mu$m \citepads{2011A&A...530L..16G}}, [CII] 158\,$\mu$m \citepads{2013A&A...550A..57O}, and CH$^+$\ 3--2 \citepads{2013A&A...550A..96N}.
The feature corresponds to the Southern tip of the Orion Ridge facing the Trapezium cluster, which confines the HII region, as visible in large-scale maps of CN line emission \citepads{1998A&A...329.1097R} and $^{13}$CO\ 3--2 emission (Lis \& Schilke \citeyearads{2003ApJ...597L.145L}; Melnick et al \citeyearads{2012ApJ...752...26M}).
From multi-line CN observations, \citetads{2001ApJ...559..985R} derive H$_2$\ densities as high as 10$^6$\,cm$^{-3}$\ for this ridge, which may be favourable to excite OH$^+$\ line emission.
However, this density is probably an overestimate, as collisions of CN with electrons were not taken into account, which are known to be important for CN \citepads{1991ApJ...369L...9B,2013MNRAS.435.3541H}.
Furthermore, we note that the OH$^+$\ line is not detected toward the Orion~S clump further up the Orion Ridge, possibly because emission and absorption from different layers cancel out each other.
\begin{table}
\caption{Optically thin estimates of H$_n$O$^+$\ column densities \newest{toward the CO$^+$\ peak of the Orion Bar}, in units of 10$^{12}$ cm$^{-2}$, as a function of assumed excitation temperature.}
\label{t:cold}
\begin{tabular}{cccc}
\hline \hline
\noalign{\smallskip}
$T_{\rm ex}$\ (K) & OH$^+$\ & H$_2$O$^+$\ & H$_3$O$^+$\ \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
10 & 82.1 & $<$2.85 & $<$7.21 \\
20 & 8.61 & $<$0.21 & $<$0.64 \\
40 & 4.01 & $<$0.10 & $<$0.27 \\
80 & 3.90 & $<$0.13 & $<$0.32 \\
160 & 5.45 & $<$0.25 & $<$0.64 \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
\end{tabular}
\tablefoot{The limits for H$_2$O$^+$\ and H$_3$O$^+$\ are based on the 607~and 985~GHz lines; limits from the other lines are significantly higher.}
\end{table}
\begin{figure}[t]
\centering
\includegraphics[width=6cm,angle=-90]{aa22164f3.ps}
\caption{Abundances of OH$^+$, H$_2$O$^+$, H$_3$O$^+$, \newest{and C$^+$}, predicted by the Meudon PDR model as a function of visual extinction for a radiation field of $\chi$=10,000 \newest{in units of the local interstellar radiation field according to \citetads{1978ApJS...36..595D}} and a pressure of $P$=10$^8$ K\,cm$^{-3}$, as applicable to the Orion Bar.}
\label{f:pdr}
\end{figure}
\subsection{Column densities}
\label{ss:cold}
The variations in OH$^+$\ emission level by factors of 2--3 across the maps in Figure~\ref{f:maps} likely correspond to variations in the total column density or the OH$^+$\ abundance with position, although excitation conditions (volume density, kinetic temperature) may also play a role.
{Since the emission does not appear to be strongly peaked,}
We focus in the following analysis on the CO$^+$\ peak where we have limits on the other H$_n$O$^+$\ ions.
To estimate the column densities of these ions, we use the line fluxes from Table~\ref{t:pars} and apply a main beam efficiency of 76\% at 607~GHz, 70\% at 1655~GHz, and 74\% at our other line frequencies \citepads{2012A&A...537A..17R}. The column density $N$ depends on the excitation temperature $T_{\rm ex}$ through
$$ N_{\rm tot} = \frac{8\pi k \nu^2 }{hc^3} \frac{Q(T_{\rm ex})}{g_u A_{\rm ul}} e^{E_u / kT_{\rm ex}} \int T_{\rm mb} dV$$
where $\nu$ is the line frequency, $Q$ is the partition function, $g_u$ is the upper state degeneracy, and $A_{ul}$ is the spontaneous decay rate.
This equation assumes optically thin emission and negligible background radiation ($T_{\rm bg} << T_{\rm ex}$), but does not assume the Rayleigh-Jeans limit ($h\nu << kT_{\rm ex}$) because our measurements are at high frequency.
The optically thin assumption is reasonable given the low expected abundances of the H$_n$O$^+$\ ions, and neglecting background radiation is justified given the low level of continuum radiation in our observations.
Adopting a background brightness temperature of 9~K increases the column density estimates by 5\% for $T_{\rm ex}$\ = 20~K, and by $<$1\% for higher values of $T_{\rm ex}$.
To evaluate the above expression for the column density, we use the spectroscopic parameters of the lines in Table~\ref{t:lines}.
Table~\ref{t:cold} presents estimates of the column densities of OH$^+$, H$_2$O$^+$\ and H$_3$O$^+$\ for values of $T_{\rm ex}$\ between 10~and 160~K, which is the expected range for the Orion Bar.
If the excitation of H$_n$O$^+$\ is close to LTE, $T_{\rm ex}$\ would be close to the kinetic temperature of the gas, which ranges from $\approx$85~K for the dense gas \citepads{1995A&A...294..792H} to $\approx$150~K near the cloud surface \citepads{2003A&A...408..231B,2011A&A...530L..16G}.
If collisional excitation of the lines cannot compete with their radiative decay, $T_{\rm ex}$\ will drop below $T_{\rm kin}$, while $T_{\rm ex}$\ may exceed $T_{\rm kin}$\ if radiative or chemical pumping plays a role.
Section~\ref{s:nonlte} discusses these processes in more detail.
For now, the table shows that the derived column densities of H$_n$O$^+$\ vary by factors of 2--3 for $T_{\rm ex}$\ ${_>\atop{^\sim}}$20~K, but that the estimate increases by an order of magnitude if $T_{\rm ex}$\ is as low as 10~K.
The OH$^+$\ column density in Table~\ref{t:cold} for $T_{\rm ex}$\ = 10~K is comparable to the values toward other Galactic sources, while the estimates for $T_{\rm ex}$\ ${_>\atop{^\sim}}$20~K are $\sim$10$\times$ lower.
Absorption line studies indicate $N$(OH$^+$) values between a few 10$^{13}$ and a few 10$^{14}$~cm$^{-2}$, both for the diffuse foreground clouds toward G10.6, W49N, W51 and OMC-2 \citepads{2010A&A...518L.110G,2010A&A...521L..10N,2012ApJ...758...83I,2013A&A...549A.114L} and the dense gas around the protostars AFGL 2591, W3 IRS5, and Orion-KL \citepads{2010A&A...521L..44B,2010A&A...521L..35B,2010A&A...521L..47G}.
Our OH$^+$/H$_2$O$^+$\ ratio of $>$40 and OH$^+$/H$_3$O$^+$\ ratio of $>$15 are larger than in previous observations of diffuse clouds, suggesting an origin of the observed emission in very diffuse low-extinction layers of the PDR.
\section{PDR models}
\label{s:pdr}
\begin{figure}[t]
\centering
\includegraphics[width=6cm,angle=0]{aa22164f4.ps}
\caption{\newest{Face-on} column densities of OH$^+$, H$_2$O$^+$\ and H$_3$O$^+$, predicted by the Meudon PDR model as a function of radiation field (in Draine units) for a pressure of $P$=10$^8$ K\,cm$^{-3}$, as applicable to the Orion Bar. \new{The integrated $A_{\rm V}$ of the model is 10 mag.}}
\label{f:chi}
\end{figure}
\begin{table}[t]
\begin{minipage}[t]{\linewidth}\centering
\caption{Rate coefficients of the main formation and destruction paths of H$_n$O$^+$. }
\label{table:rates}
\renewcommand{\footnoterule}{}
\begin{tabular}{lll}
\hline \hline
\noalign{\smallskip}
Reaction& $k$ (cm$^{3}$\,s$^{-1}$) & $\Delta T$ (K) \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
$k_1$& $1.70\times10^{-9}$& 10--41000\\
$k_3$& $2.10\times10^{-9}$& 10--41000\\
$k_4$& $1.01\times10^{-9}$& 10--41000\\
$k_5$& $6.40\times10^{-10}$& 10--41000\\
$k_7$& $6.30\times10^{-9}~(T/300)^{-0.48}$& -- \tablefootmark{a} \\
$k_9$& $3.05\times10^{-7}~(T/300)^{-0.50}$& 10--1000\\
$k_{10}$& $8.60\times10^{-8}~(T/300)^{-0.50}$& 10--1000\\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
Reaction& $k$ (s$^{-1}$) & $\Delta T$ (K) \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
$k_2$& $1.60\times10^{-12}~\exp{(-3.1 A_{\rm{V}})}$& 10--41000\\
$k_6$& $1.00\times10^{-9}~\exp{(-1.7 A_{\rm{V}})}$& -- \tablefootmark{b} \\
$k_8$& $1.10\times10^{-11}~\exp{(-3.5 A_{\rm{V}})}$& 10--41000\\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
\end{tabular}
\end{minipage}
\tablefoot{Rates are based on the UMIST database \citepads{2007A&A...466.1197W} unless otherwise noted. The photoreaction rates at the bottom are given for a standard Draine radiation field. The last column gives the temperature range over which the $k$-values are valid.}
\tablefoottext{a}{Rate from the OSU database (version March 2008).} \\
\tablefoottext{b}{Assumed value.}
\end{table}
\new{To understand the chemistry of H$_n$O$^+$\ in more detail, we model the Orion Bar with version 1.4.4 of the Meudon PDR model \citepads{2006ApJS..164..506L,2007A&A...467....1G,2008A&A...485..127G}.
We describe the object with an isobaric model, where the pressure is kept constant and the program solves for the thermal and chemical balance as a function of depth.
Isobaric models are the simplest description of gas in steady-state, stationary molecular clouds where gravitation is negligible.
The model includes a gas-phase chemical network based on the UMIST database, as well as H$_2$\ formation on grains and neutralization of ions on grains and PAHs.}
Figure~\ref{f:pdr} shows the predicted abundances of OH$^+$, H$_2$O$^+$\ and H$_3$O$^+$\ as a function of visual extinction, for a radiation field of $\chi$=10,000 Draine (\citeyearads{1978ApJS...36..595D}) units ($\chi_0$ = \pow{2.7}{-3} erg\,s$^{-1}$\,cm$^{-2}$), \newest{a cosmic-ray ionization rate of \pow{2}{-16}\,s$^{-1}$, } and a pressure of $P$=10$^8$ K\,cm$^{-3}$.
\new{The radiation field is taken from \citet{2011A&A...530L..16G}}
and the pressure is chosen to match the intensities of the CH$^+$\ $J$ = 1--0 to 6--5 transitions observed toward the Orion Bar CO$^+$ peak with Herschel HIFI and PACS \citepads{2013A&A...550A..96N}.
For these parameters, the total gas density at the depth where the H$_n$O$^+$ abundances peak ($A_{\rm{V}}\sim0.3-1.0$) is in the range between $5.6\times10^4$ and $1.8\times10^5$ cm$^{-3}$.
\new{The temperatures at these depths in the model (Table~\ref{table:rates_meudon}) agree well with estimates for the PDR surface from lines of H$_2$\ \citepads[400--700\,K,]{2005ApJ...630..368A} and H \citepads[$\sim$550\,K,]{2013ApJ...762..101V}.}
The column densities of our observed species, integrated up to $A_V$=1~mag, are $N$(OH$^+$) = \pow{1.6}{13}~cm$^{-2}$, $N$(H$_2$O$^+$) = \pow{4.4}{12}~cm$^{-2}$, and $N$(H$_3$O$^+$) = \pow{7.5}{12}~cm$^{-2}$.
When integrating up to $A_V$=10~mag, the values for OH$^+$\ and H$_2$O$^+$\ remain the same to 1\%, and $N$(H$_3$O$^+$) increases by just 6\% which is insignificant.
These predictions are consistent with the estimates in Table~\ref{t:cold} for $T_{\rm ex}$\ = 10--20~K, except that the observed OH$^+$/H$_2$O$^+$\ ratio is larger than in the models.
Note that the predictions correspond to a face-on model; to compare to the observations, they must be multiplied by a factor of 1/sin~$i$ $\approx$4, where $i\approx15^\circ$ for the Orion Bar \newer{\citepads{2013A&A...550A..96N}}.
As shown in Figure~\ref{f:pdr}, H$_n$O$^+$ abundances in the high UV illumination Orion Bar PDR peak near the cloud surface, at depths of $A_{\rm{V}}\sim0.3-0.4$, and decrease rapidly beyond $A_{\rm{V}}\sim1$.
\newest{Given the clumpy structure of the Orion Bar, this narrow range in $A_{\rm{V}}$ is spatially rather extended, as indicated by the maps in Fig.~\ref{f:maps}.}
To understand the chemistry of H$_n$O$^+$\ in the Orion Bar, we study the main formation and destruction channels of these ions at depths of $A_{\rm{V}}$ = 0.3, 0.4 and 1.0 mag.
The corresponding reaction coefficients and rates are listed in Tables~\ref{table:rates} and~\ref{table:rates_meudon}.
At a depth of $A_{\rm{V}}$ = 0.3 -- 0.4, the dominant pathways for OH$^+$ formation are:
\begin{equation}
\label{k1}
\mathrm{H_2 + O^+ \xrightarrow{k_1} OH^+ + H}
\end{equation}
\begin{equation}
\label{k2}
\mathrm{OH + h\nu \xrightarrow{k_2} OH^+ + e^-}
\end{equation}
\begin{equation}
\label{k3}
\mathrm{OH + H^+ \xrightarrow{k_3} OH^+ + H}
\end{equation}
The dominant OH$^+$ formation path is via H$_2$ and O$^+$, which accounts for \new{70}\% of the total OH$^+$ production at $A_{\rm{V}} = 0.3$ and \new{85}\% at $A_{\rm{V}} = 0.4$. Photoionization of OH accounts for \new{18}\% of OH$^+$ formed at $A_{\rm{V}} = 0.3$ and \new{12}\% at $A_{\rm{V}} = 0.4$.
Charge exchange between OH and H$^+$ is mostly significant at $A_{\rm{V}}\sim0.3$, resulting in \new{12}\% of OH$^+$ production which decreases to \new{3}\% at $A_{\rm{V}}\sim0.4$.
This reaction is more important in X-ray dominated regions (XDRs), which is a key difference between such regions and high-illumination PDRs \citep{2010A&A...518L..42V}.
At a depth of $A_{\rm{V}}\sim1$, the path via H$_2$ and O$^+$ is still the dominant OH$^+$ formation path (72.1\%).
However, the role of OH photoionization increases, as it accounts for producing \new{27}\% of OH$^+$ at this depth.
Once OH$^+$ is formed, H$_2$O$^+$ and H$_3$O$^+$ formation rapidly follows via similar reactions with H$_2$:
\begin{equation}
\label{k4}
\mathrm{OH^+ + H_2 \xrightarrow{k_4} H_2O^+ + H}
\end{equation}
\begin{equation}
\label{k5}
\mathrm{H_2O^+ + H_2 \xrightarrow{k_5} H_3O^+ + H}
\end{equation}
At a depth of $A_{\rm{V}} = 0.3$, \new{100\% of} H$_2$O$^+$ forms from OH$^+$ in the reaction with H$_2$.
At a depth of $A_{\rm{V}} = 0.4$, as the H$_3$O$^+$ abundance is increased compared to that at $A_{\rm{V}}$ = 0.3, a secondary reaction apart from the reaction from H$_2$ (83\%) produces \new{17}\% of the H$_2$O$^+$ formed at this depth:
\begin{equation}
\label{k6}
\mathrm{H_3O^+ + h\nu \xrightarrow{k_6} H_2O^+ + H}
\end{equation}
At a depth of $A_{\rm{V}}$ = 1.0, the role of H$_2$O$^+$ production from H$_3$O$^+$ decreases to \new{9}\%, compared to the dominant reaction from H$_2$ (\new{88}\%).
At each of these depths in the model,100\% of H$_3$O$^+$ is produced in the reaction of H$_2$O$^+$ with H$_2$.
The destruction of OH$^+$ is mainly via reactions with electrons, photons and H$_2$.
The reaction with H$_2$ (Eqn. \ref{k4}) is the most important H$_2$O$^+$ formation path, as mentioned above.
It is also the most important destruction path for OH$^+$ at depths of $A_{\rm{V}}$ = 0.4 and beyond.
At smaller depths, both dissociative recombination with electrons and UV photodissociation contribute significantly to the OH$^+$ destruction: at $A_{\rm{V}}\sim~0.3$,
\begin{equation}
\label{k7}
\mathrm{OH^+ + e^- \xrightarrow{k_7} H + O}
\end{equation}
\noindent
contributes \new{52}\%, and
\begin{equation}
\label{k8}
\mathrm{OH^+ + h\nu \xrightarrow{k_8} H + O^+}
\end{equation}
\noindent
contributes \new{18}\%.
Destruction of H$_2$O$^+$ occurs mainly via dissociative recombination with electrons at low depths ($A_{\rm{V}}$ = 0.3 -- 0.4):
\begin{equation}
\label{k9}
\mathrm{H_2O^+ + e^- \xrightarrow{k_9} H + O + H}
\end{equation}
\begin{equation}
\label{k10}
\mathrm{H_2O^+ + e^- \xrightarrow{k_{10}} H + OH}
\end{equation}
Reaction \ref{k9} accounts for \new{71}\% of H$_2$O$^+$ destruction at $A_{\rm{V}}$ = 0.3, while reaction \ref{k10} contributes \new{20}\%.
At a depth of $A_{\rm{V}}$ = 0.4, reactions \ref{k9} and \ref{k10} contribute \new{44} and \new{12}\% (respectively) to the H$_2$O$^+$ destruction.
\new{As H$_2$O$^+$\ does not have photodissociation channels longward of 13.6 eV, the model does not include this process} \newer{\citepads{2012ApJ...754..105H}}.
Destruction by H$_2$, which controls H$_3$O$^+$ production is only significant at $A_{\rm{V}}\sim1$ (\new{87}\%).
Therefore, the high electron density and UV radiation field \new{in the Orion Bar may} explain the H$_2$O$^+$ and H$_3$O$^+$ non-detections, as a significant fraction of OH$^+$ and H$_2$O$^+$ is directly destroyed by UV photons or by recombination with electrons, limiting the H$_2$O$^+$ and H$_3$O$^+$ formation.
Our non-detections of H$_2$O$^+$\ and H$_3$O$^+$\ are thus consistent with an origin of the OH$^+$\ emission in diffuse gas where C$^+$\ is abundant.
Alternatively, our observed H$_2$O$^+$/OH$^+$\ ratio may be due to a low molecular fraction (H$_2$/H ratio) in the gas probed by our observations.
Gerin et al. (2010) derive an analytic expression for the OH$^+$/H$_2$O$^+$ abundance ratio as a function of the gas temperature, electron density, and H$_2$ volume density:
\begin{equation}
\nonumber
n({\rm{OH^+}})/n({\rm{H_2O^+}})=0.64 + 430 \times (T/300)^{-0.5} \times [n({\rm{e}}^-)/n({\rm{H}}_2)]
\end{equation}
As most OH$^+$ forms in the outermost layers of the PDR ($A_{\rm{V}}<0.4$) at very low molecular fractions (0.01\%-0.4\% for $A_{\rm{V}}=0.3$ and $A_{\rm{V}}=0.4$, respectively), this formula is consistent with the observed abundance ratio of $N$(OH$^+$)/$N$(H$_2$O$^+$)$>$2.
This suggests that, apart from the effect of the high electron density mentioned above, the low molecular fraction in the surface of the PDR also contributes to the observed OH$^+$/H$_2$O$^+$ column density ratios.
We conclude that most of our observed \newer{OH$^+$}\ emission originates at extinctions below $A_V = 0.4$.
Fig.~\ref{f:pdr} shows a second peak in the H$_3$O$^+$ abundance at a depth of $A_{\rm{V}}\sim7-8$.
This second peak is expected in the abundance of H$_n$O$^+$ in interstellar clouds with a large range of physical conditions \citepads[e.g.,]{2012ApJ...754..105H} and does not significantly contribute to the total H$_3$O$^+$ column density.
The sequence of H$_n$O$^+$ formation at this depth is initiated by the cosmic-ray ionization of H$_2$, followed by a reaction of H$_2^+$ with H$_2$ producing H$_3^+$.
At this depth, OH$^+$ is formed from H$_3^+$, whose abundance is increased due to the lower electron abundance.
The second abundance peak disappears toward higher $A_V$ as oxygen freezes out on the grain surfaces as water.
Fig.~\ref{f:chi} shows the effect of changing the radiation field on the H$_n$O$^+$\ column densities.
The model corresponds to $P$=10$^8$ cm$^{-3}$\,K and the figure considers radiation fields between $\chi$=5000 and 50000 Draine units.
The predicted column densities of the H$_n$O$^+$\ species are seen to peak for $\chi$=20,000--30,000 and to drop for higher radiation fields.
\new{The H$_n$O$^+$\ column densities in our model are similar to those by \citetads{2012ApJ...754..105H} under Orion Bar conditions (their Figs.~8 and~9).}
\section{Excitation of OH$^+$}
\label{s:nonlte}
\begin{table}[t]
\begin{minipage}[!h]{\linewidth}\centering
\caption{Rates (in cm$^{-3}$ s$^{-1}$) corresponding to the main formation and destruction paths of H$_n$O$^+$ in the model at $A_{\rm{V}}=0.3$, 0.4, and 1.0.}
\label{table:rates_meudon}
\renewcommand{\footnoterule}{}
\begin{tabular}{lccc}
\hline \hline
\noalign{\smallskip}
$A_{\rm{V}}$ & 0.3 & 0.4 & 1.0 \\
$T_{\rm gas}$ & 1783 K & 1013 K & 569 K \\
$n_{\rm gas}$ & \pow{5.6}{4}\,cm$^{-3}$ & \pow{9.9}{4}\,cm$^{-3}$ & \pow{1.8}{5}\,cm$^{-3}$ \\
$n$(H$_2$) & \pow{1.9}{1}\,cm$^{-3}$ & \pow{3.9}{3}\,cm$^{-3}$ & \pow{1.0}{5}\,cm$^{-3}$ \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
Reaction& \multicolumn{3}{c}{Rate}\\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
$k_1$& $1.31\times10^{-10}$& $6.25\times10^{-9}$& $1.53\times10^{-9}$\\
$k_2$& $3.38\times10^{-11}$& $8.46\times10^{-10}$& $5.77\times10^{-10}$\\
$k_3$& $2.15\times10^{-11}$& $2.25\times10^{-10}$& $1.22\times10^{-11}$\\
$k_4$& $3.11\times10^{-11}$& $7.14\times10^{-9}$& $2.11\times10^{-9}$\\
$k_5$& $9.52\times10^{-14}$& $3.29\times10^{-9}$& $2.08\times10^{-9}$\\
$k_6$& $4.84\times10^{-14}$& $1.43\times10^{-9}$& $2.03\times10^{-10}$\\
$k_7$& $9.75\times10^{-11}$& $1.09\times10^{-10}$& $2.93\times10^{-12}$\\
$k_8$& $3.27\times10^{-11}$& $2.50\times10^{-11}$& $3.54\times10^{-14}$ \\
$k_9$& $2.20\times10^{-11}$& $3.76\times10^{-9}$& $2.18\times10^{-10}$\\
$k_{10}$& $6.21\times10^{-12}$& $1.06\times10^{-9}$& $6.15\times10^{-11}$\\
\noalign{\smallskip}
\hline
\end{tabular}
\end{minipage}
\end{table}
Our estimated column densities of OH$^+$, H$_2$O$^+$\ and H$_3$O$^+$\ are in reasonable agreement with the predictions from the Meudon PDR model, but the appearance of the lines in emission \newest{is unlike other Galactic sources observed in OH$^+$\ so far}.
To understand this behaviour we perform a non-LTE analysis of the excitation of OH$^+$, which takes both reactive and inelastic collisions into account, as well as excitation by the background continuum radiation field.
\subsection{Collisional and radiative excitation}
\label{ss:radex}
The calculations in Appendix~\ref{ss:crc} indicate inelastic electron collision rates for the OH$^+$\ 971 GHz line of $\approx$\pow{6}{-7} cm$^3$\,s$^{-1}$, which is $\sim$10$\times$ higher than the dissociative recombination rate at $T \approx$100~K (Table~\ref{table:rates}).
At this temperature, $h \nu / k T \approx 0.5$ for this transition, so that by detailed balance, the upward and downward rates differ by only a few percent.
For collisions with H$_2$, inelastic collision rates are not known, but it may be reasonable to assume that reactive collisions dominate.
In contrast, the reaction of OH$^+$\ with H is endothermic, and inelastic collisions with H could influence the excitation of OH$^+$.
\new{Test calculations assuming a collisional rate coefficient of $10^{-10}$\,cm$^3$\,s$^{-1}$\ (cf. Andersson et al \citeyearads{2008ApJ...678.1042A} for the case of CO$^+$) for all radiatively allowed transitions indicate that this effect is comparable with electron collisions.
\newest{Very recent calculations on the OH$^+$-He system confirm this estimate, at least to order of magnitude (F. Lique, priv. comm.).}
We therefore first calculate the excitation of OH$^+$\ assuming that inelastic collisions with H and electrons dominate, and consider the effect of reactive collisions with H$_2$\ in \S\ref{ss:form-dest}.}
We have used the non-LTE radiative transfer program Radex \citepads{2007A&A...468..627V} to calculate the excitation of OH$^+$\ in the Orion Bar, assuming steady-state conditions and using the inelastic collision rates from Appendix~\ref{ss:crc}.
Observations of C recombination lines toward the Orion Bar surface indicate hydrogen densities of \pow{5}{4}--\pow{2.5}{5} cm$^{-3}$\ \citepads{1997ApJ...487L.171W}, which for all carbon in C$^+$ and C/H = \pow{1.4}{-4} translates into an electron density of $n$(e) $\approx$10 cm$^{-3}$, as we have used before for HF \citepads{2012A&A...537L..10V}.
We use an electron temperature of $T_e$ = 300 K, but our results are insensitive to variations in $T_e$ between 100 and 1000~K.
The adopted line width is the observed 4.3 km\,s$^{-1}$\ and for the background radiation field we adopt a \newest{modified blackbody} distribution with a dust temperature of $T_d$=50~K and a dust emissivity index of $\beta$=1.6, as found by \citet{2012A&A...541A..19A} for the interior of the Bar, so that $\tau_d$=0.21 at 971~GHz.
The model is insensitive to the details of this radiation field; in particular, the results are unchanged when adopting $T_d$=70~K and $\beta$=1.2 as found by Arab et al for the Bar's surface.
The model predicts an excitation temperature of $\approx$10~K for the OH$^+$\ lines near 1~THz.
The excitation is due to the combination of collisions and radiation: a model with the \newest{modified blackbody} replaced by the 2.73~K cosmic microwave background results in $T_{\rm ex}$\ $\approx$7~K.
The effect on the emerging line intensities is large, since the Planck function at 1~THz increases by a factor of $\sim$4 from $T_{\rm ex}$\ = 7~K to 9~K.
The calculated excitation temperature of $\approx$10~K is above the background radiation temperature of 9.2~K, which explains the appearance of the line in emission, but
the models require $N$(OH$^+$) $\approx$\pow{5}{14}\,cm$^{-2}$\ to match our observed 971~GHz line intensity.
At this high column density, the OH$^+$\ lines become optically thick which drives the three $N$=1--0 lines to approximately the same peak brightness, inconsistent with our non-detection of the $1_0-0_1$ line of OH$^+$.
We therefore regard this model as untenable, conclude that the observed line emission is optically thin, and search for a model with a higher excitation temperature in the next section.
\subsection{Effect of reactive collisions}
\label{ss:form-dest}
The above calculation indicates that \new{inelastic} collisions are sufficient to make the OH$^+$\ line appear in emission given the low background intensity of the Orion Bar.
A fully self-consistent model should however include formation and destruction terms in the rate equations.
We use Equation 12 from \citetads{2007A&A...468..627V}, where the number density of OH$^+$\ molecules (in cm$^{-3}$) is the ratio of its formation rate (in cm$^{-3}$\,s$^{-1}$) and its destruction rate (in s$^{-1}$).
We adopt a destruction rate of OH$^+$\ in the Orion Bar of 10$^{-4}$~s$^{-1}$, which is due to reactions with H$_2$, electron recombination, and photodissociation.
In contrast, the formation of OH$^+$\ through ion-molecule reactions should proceed at a rate at or below the Langevin rate of $\approx$10$^{-9}$~cm$^3$\,s$^{-1}$\ and is limited by the supply of hydrogen or oxygen ions.
Lacking state-to-state formation rates, we approximate the distribution of the newly-formed OH$^+$\ over its energy levels by a thermal distribution at a temperature $T_f$.
The value of $T_f$ depends on the dominant formation route and is likely to be a significant fraction of its excess energy.
The reactions of O$^+$\ with H$_2$\ and of O with H$_3^+$\ have exothermicities of 0.55 and 0.66~eV \citepads{1984PhRvL..52.2084F,2000CPL...319..482M}, so $T_f$ should be in the range 2000--3000~K.
For OH$^+$\ formation rates between 10$^{-12}$ and 10$^{-9}$~cm$^3$\,s$^{-1}$\ and formation temperatures of 2000--3000~K, the model predicts excitation temperatures of $\approx$12~K for the OH$^+$\ lines near 1~THz.
Compared to models with a negligible formation rate, the line brightness increases by a factor of $\approx$5, almost independent of $T_f$, for the same column density.
The observed \newest{intensities of the $1_2--0_1$ and $1_1--0_1$ lines are} matched for $N$(OH$^+$) $\approx${\pow{1}{14}~cm$^{-2}$, which is considerably lower than for the steady-state excitation model.
The difference with the estimates in Table~\ref{t:cold} is the presence of background radiation.
\newest{This model predicts some emission in the $1_0--0_1$ line, but only at the 2$\sigma$ level of our observations.}
We conclude that formation pumping plays an important role for OH$^+$\ in the Orion Bar.
\section{Discussion}
\label{s:disc}
\subsection{Effect of X-ray ionization}
\label{ss:zeta}
\newest{Our derived OH$^+$\ column density of $\approx$\pow{1}{14}\,cm$^{-2}$\ is similar to that in diffuse clouds, as observed in absorption toward W49N and other sources (see references in \S\ref{ss:cold}), but $\sim$1.6$\times$ higher than the model prediction of \pow{6.4}{13}\,cm$^{-2}$\ in \S\ref{s:pdr}.}
The calculations in \S\ref{s:pdr} assume a cosmic-ray ionization rate of \pow{2}{-16}~s$^{-1}$, which is a typical value for diffuse interstellar clouds in the Solar neighbourhood \citepads{2012ApJ...745...91I} and an order of magnitude higher than the value for dense clouds \citepads{2000A&A...358L..79V}.
However, the actual ionization rate of the Orion Bar may be atypically high because of its proximity to the Trapezium stars.
The effects of X-rays from these stars on the chemistry are similar to those of cosmic rays \citep{2006ApJ...650L.103M}.
In particular, \citetads{2010A&A...521L..47G} estimate an ionization rate of \pow{3}{-15} s$^{-1}$\ for the Orion KL region, dominated by X-rays from the star $\theta^1$C Ori.
Calculations using the Meudon code with an ionization rate of \pow{2}{-15}~s$^{-1}$ result in very similar H$_n$O$^+$\ column densities as before, though, presumably because of the high gas density in the Orion Bar.
\subsection{Possible ion sources}
\label{ss:ion-src}
\newest{Alternatively, the discrepancy between our observed and modeled column densities of OH$^+$\ may arise because the assumption of a stationary medium is inappropriate.}
Regarding the Orion Bar as a molecular cloud under external illumination and heating, the main source of OH$^+$\ may be the reaction of H$_3^+$\ with O, where the H$_3^+$\ is due to cosmic-ray ionization as well as leakage of UV photons from the ionized region.
However, interaction of the PDR with the neighbouring photoionized nebula may supply O$^+$\ ions which produce extra OH$^+$\ in their reaction with H$_2$.
The supply of O$^+$\ ions may be due to leakage of hydrogen- and oxygen-ionizing photons, but also to advection of H$^+$\ and O$^+$\ ions from the ionized nebula into the mostly neutral PDR, as in some planetary nebulae \citepads{1983IAUS..103...91B}\footnote{\newest{Very recently, OH$^+$\ emission has been detected towards the Helix nebula (Van Hoof et al, in prep).}}.
The concentration of H$^+$\ with depth into the PDR is then important, because charge transfer reactions (H$^+$\ + O $\leftrightarrow$ O$^+$\ + H) will rapidly couple the O$^+$/O ratio to the H$^+$/H ratio.
Observations of the [O{\sc III}] 88~$\mu$m\ line toward the Orion Bar with Herschel/PACS show that ionized oxygen is widespread in the region (C. Joblin \& J. Goicoechea, priv. comm.);
the emission extends well into the mostly-neutral gas traced by the [O{\sc I}] 63~$\mu$m\ line (M. Gerin, priv. comm.), making the O$^+$\ + H$_2$\ channel a likely source of OH$^+$\ \new{at low depths}.
\new{Detailed comparison of the optical and near-IR (forbidden and/or permitted) lines of O and O$^+$\ would be a stronger test of this scenario \citepads{2000A&A...364..301W,2011MNRAS.417..420M}}.
In addition, simulations of the full molecular + atomic + ionized gas in the Orion PDR with the Cloudy program \citepads{2013RMxAA..49..137F} would be useful to constrain the role of possible ion sources in the Orion Bar.
\subsection{Radiative pumping}
\label{ss:pump}
Besides changing the chemistry of OH$^+$, the strong infrared and ultraviolet radiation fields in the Orion Bar may change its excitation.
Section~\ref{ss:radex} already showed that far-infrared continuum radiation raises the OH$^+$\ excitation temperature significantly above the level due to \new{inelastic} collisions alone.
In addition, mid-infrared pumping \newest{through the $v$=1--0 band at 3.38~$\mu$m} may contribute if the radiative excitation rate of OH$^+$\ in the Bar exceeds the collisional excitation rate by electrons, which is $n$(e) $\times C_{\rm lu}$ = 10~cm$^{-3}$\ $\times$\pow{6}{-7} cm$^3$\,s$^{-1}$\ = \pow{6}{-6}~s$^{-1}$.
The radiative rate is $B_{\rm lu} U_{\rm rad}$, the Einstein absorption coefficient times the radiative energy density, which is approximately $A_{\rm vib} \epsilon f / (e^{(h\nu / (kT_d))} -1)$, where $\epsilon$ is the dust emissivity, $f$ the dust filling factor, $A_{\rm vib}$ the spontaneous decay rate of 265~s$^{-1}$, and $T_d$ the dust temperature.
For the Orion Bar, the filling factor should be close to unity, and for the emissivity we assume 1 at the short wavelength of the fundamental vibrational band of OH$^+$.
Equating the radiative rate to the electron collision rate indicates a minimum temperature of $\approx$240~K, which is reasonable for the gas at the Bar's surface, but too much for the dust, as the PACS and SPIRE data show \citep{2012A&A...541A..19A}.
Only PAHs and small grains would reach such high temperatures, but with very low opacities,
even though the OH$^+$\ vibrational fundamental is close to the PAH 3.3~$\mu$m\ and aliphatic 3.4~$\mu$m\ emission features.
Besides continuum radiation from dust in the Bar itself, pumping by infrared starlight from the Trapezium may influence the excitation of OH$^+$.
The brightest of these stars is $\theta^1$C~Ori, which is 127$''$ away from our observing position \citepads{2007A&A...474..653V}.
If both objects lie at the same distance from the Sun, the stellar continuum flux at the Orion Bar is 2.96 million times stronger than at the Earth.
The stellar temperature of 37,000~K implies a radiative intensity of 0.2~Jy/nsr or \pow{2}{-15} erg\,s$^{-1}$\,cm$^{-2}$\,Hz$^{-1}$\,sr$^{-1}$.
Setting this equal to the Planck function at a radiation temperature $T_R$, we obtain $T_R = 185$\,K at $\lambda$ = 3.38\,$\mu$m.
The corresponding pumping rate in the OH$^+$\ vibrational fundamental is of order $A_{\rm vib} / \{ \exp(h\nu / kT_R) -1 \}$ = \pow{5.7}{-8}\,s$^{-1}$, which is much less than the collisional excitation rate by electrons.
While infrared pumping does not seem play a role for OH$^+$\ in the Orion Bar, the absorption rate through electronic transitions in the near ultraviolet is rather higher.
The radiative intensity of $\theta^1$C~Ori at the wavelength of the A$^3\Pi_i$ -- X$^3\Sigma^-$ $v$=0--0 band of OH$^+$\ at 27949\,cm$^{-1}$ (3577\,\AA) is $\approx$1.8\,Jy/nsr.
The A-value for this band is \pow{8.01}{5}\,s$^{-1}$\ \citepads{1981A&A....95..383D}, so that the absorption rate in this band alone is $\approx$\pow{3.3}{-6}\,s$^{-1}$, which is only slightly less than the collisional excitation rate.
\subsection{Comparison with extragalactic systems}
\label{ss:xgal}
The Orion Bar is the first and so far only position within our Galaxy where lines of OH$^+$\ appear purely in emission, \new{although a mix of emission and absorption is seen in W3 IRS5 \citep{2010A&A...521L..35B} and NGC 3603 (Makai et al, in prep.) and possibly many more star-forming regions.}
Detections of extragalactic OH$^+$\ and H$_2$O$^+$\ emission have been made with Herschel-SPIRE by \citetads{2010A&A...518L..42V} toward the active nucleus of the galaxy Mrk~231, with Herschel-PACS toward the ultraluminous merger Arp~220 by \citetads{2011ApJ...743...94R} and with Herschel-SPIRE toward the Seyfert nucleus NGC~1068 by \citetads{2012ApJ...758..108S}.
In addition, detections of extragalactic H$_n$O$^+$\ absorption exist toward M~82 using HIFI \citepads{2010A&A...521L...1W} and SPIRE \citepads{2012ApJ...753...70K}.
Recently, PACS observations of excited OH$^+$, H$_2$O$^+$\ and H$_3$O$^+$\ toward NGC 4418 and Arp 220 have been discussed by \citetads{2013A&A...550A..25G}.
{We suspect that the nuclei where H$_n$O$^+$\ lines appear in emission have an enhanced electron density, far-infrared continuum, and/or ionizing (UV/X-ray) continuum.}
\new{The Orion Bar is special in our Galaxy for its large column density of warm, mostly-atomic gas and its weak far-infrared continuum; we suspect that extragalactic nuclei where H$_n$O$^+$\ lines appear in emission have similar conditions}. \newest{In addition, supernova remnants may contribute, as the recent detection of OH$^+$\ emission towards the Crab nebula suggests (Barlow et al, in prep.).}
\section{Conclusions}
\label{s:conc}
We have presented maps and spectra of OH$^+$\ line emission toward the Orion Bar, and limits on lines of H$_2$O$^+$\ and H$_3$O$^+$.
The OH$^+$\ line emission is extended over $\sim$1$'$ (=25,000~AU = 0.12~pc) and traces the Bar itself as well as the Southern tip of the Orion Ridge.
\new{Analysis of the chemistry and the excitation of OH$^+$\ } suggests an origin of the emission at a depth of $A_V$=0.3--0.5, similar to CH$^+$\ and SH$^+$.
The OH$^+$\ column density of $\approx$\pow{1.0}{14}~cm$^{-2}$, derived using a non-LTE model including both inelastic and reactive collisions and radiative pumping, is similar to that in previous absorption line studies, while our limits on the OH$^+$/H$_2$O$^+$\ and OH$^+$/H$_3$O$^+$\ ratios are higher than seen before.
Non-LTE models of the excitation of OH$^+$\ show that the unusual appearance of the OH$^+$\ lines in emission is the combined result of inelastic {electron} collisions, far-infrared radiative pumping by dust, and chemical pumping through the O$^+$\ + H$_2$\ and O + H$_3^+$\ channels.
The same conditions may apply to extragalactic sources of H$_n$O$^+$\ line emission.
In the future, high-resolution maps of Galactic and extragalactic OH$^+$\ line emission with ALMA will shed further light on the chemistry of this reactive ionic species.
Our observed H$_n$O$^+$\ column densities are qualitatively reproduced by a model of the Orion Bar (using the Meudon PDR code) using a radiation field of $\chi$=10$^4$~$\chi_0$ and a pressure of $P$=10$^8$~K\,cm$^{-3}$\ as suggested by previous observations.
Analysis of the main formation and destruction paths of the ions indicates that our high OH$^+$/H$_2$O$^+$\ and OH$^+$/H$_3$O$^+$\ ratios are due to the high UV radiation field and electron density in the Orion Bar.
Destruction of OH$^+$\ and H$_2$O$^+$\ by photodissociation and electron recombination limits the formation of H$_2$O$^+$\ and H$_3$O$^+$.
\new{In addition, the low molecular fraction at the PDR surface limits the production of H$_2$O$^+$\ and H$_3$O$^+$.}
Quantitatively, the Meudon PDR models underpredict the absolute OH$^+$\ column density by a factor of \newest{$\sim$1.6}.
To match the observed line intensity with an OH$^+$\ column density similar to that in the PDR model, the electron density would have to be $\approx$100\,cm$^{-3}$, which is much higher than the PDR model predicts at the depth where the H$_n$O$^+$\ ions are abundant, as seen in Fig.~10 of \citetads{2013A&A...550A..96N}.
However, raising the pressure in the PDR model by $\approx$50\% would increase the predicted $N$(OH$^+$) to the value suggested by the \new{non-LTE} models.
Such an increase is consistent with the CH$^+$\ and SH$^+$\ observations, and is also suggested by observations of high-$J$ CO lines with PACS (Joblin et al, in prep.).
Furthermore, the Meudon PDR model uses a scaling of the average interstellar radiation field, while realistic models should use direct observations of the dominant hot star $\theta^1$C~Ori for the H-ionizing part of the spectrum and the wavelength range where OH photoionization occurs, which could contribute significantly to the formation of OH$^+$.
\begin{acknowledgements}
{The authors thank S\'ebastien Bardeau and J\'er\^ome Pety (IRAM) for help with the data reduction, Franck le Petit (Paris-Meudon) for assistance with the PDR model calculations, Arturo Rodr\'{\i}guez-Franco (Madrid) for sending his CN map in electronic form, Xander Tielens (Leiden) for useful discussions, Inga Kamp (Groningen) for a careful reading of the manuscript, Christine Joblin (Toulouse) \& Javier Goecoechea (Madrid) for sharing their PACS observations of the [O{\sc III}] line, \new{and the referee for a useful report}.
AF acknowledges support by the Agence Nationale de la Recherche (ANR-HYDRIDES), contract ANR-12-BS05-0011-01.}
\par
{HIFI has been designed and built by a consortium of institutes and university departments from across Europe, Canada and the US under the leadership of SRON Netherlands Institute for Space Research, Groningen, The Netherlands with major contributions from Germany, France and the US. Consortium members are: Canada: CSA, U.Waterloo; France: CESR, LAB, LERMA, IRAM; Germany: KOSMA, MPIfR, MPS; Ireland, NUI Maynooth; Italy: ASI, IFSI-INAF, Arcetri-INAF; Netherlands: SRON, TUD; Poland: CAMK, CBK; Spain: Observatorio Astron\'omico Nacional (IGN), Centro de Astrobiolog\'{\i}a (CSIC-INTA); Sweden: Chalmers University of Technology - MC2, RSS \& GARD, Onsala Space Observatory, Swedish National Space Board, Stockholm University - Stockholm Observatory; Switzerland: ETH Z\"urich, FHNW; USA: Caltech, JPL, NHSC.}
\end{acknowledgements}
\bibliographystyle{aa}
|
\section{Introduction}
Shelah has defined unidimensional theories as (stable) theories in which any two sufficiently large
$\vert T\vert^+$-saturated models of the same cardinality are isomorphic. For stable theories this
definition is equivalent to the requirement that any two non-algebraic types are non-orthogonal.
This requirement serves as the definition of unidimensionality for the larger class of simple
theories. A problem posed by Shelah was whether any unidimensional stable theory is superstable.
Around 1986 Hrushovski has solved the problem by answering it in the affirmative [H1].
Several years after the discovery of Kim [K] that the algebraic properties of forking (symmetry and
transitivity) can be proved for simple theories (1996) and the development of the basic machinery
[K,KP,HKP], there were several attempts to generalize the above result of Hrushovski to the simple
case. A generalization of this proof along the the same lines seems very problematic because of the
lack of definability of types, and so many of the results on definable groups in stable theories do
not seem to generalize to simple theories in a direct way.
In 2003, we observed that any small simple unidimensional theory is supersimple [S3]. A bit later,
Pillay [P] has proved that any countable hypersimple theory (i.e. a simple theory that eliminates
hyperimaginaries) with the wnfcp (the weak non finite cover property) is supersimple; this proof
builds on ideas from Hrushovski's old proof of the result for countable stable theories [H0] and
some machinery from the theory of lovely-pairs [BPV]. This has been extended by Pillay [P1] to any
countable low hypersimple theory using the result on elimination of the "there exists infinitely
many" quantifier [S2]. In 2008, it has been proved that any countable hypersimple unidimensional
theory is supersimple [S1]. An important notion that used in [S1] is the forking topology (or the
$\tau^f$-topology); this is a variant of the topology used in [H0] and [P]: for variables $x$ and
set $A$ the forking topology on $S_x(A)$ is defined as the topology whose basis is the collection
of all sets of the form $\UU=\{a \vert \phi(a,y)$ forks over $A \}$, where $\phi(x,y)\in L(A)$.
The goal of this paper is to reduce the problem on supersimplicity of general hypersimple
unidimensional theories (possibly uncountable) to the case where the theory is s-essentially
1-based by means of the forking topology, namely, any type internal in a SU-rank 1 type is
s-essentially 1-based (a strong version of the notion "essentially 1-based" from [S1]) by means of
the forking-topology. We do this by generalizing the dichotomy theorem from [S1] to any hypersimple
theory (rather than a countable one) equipped with a projection-closed family of topologies, while
its conclusion is strengthened to get that any type internal in a SU-rank 1 type is s-essentially
1-based (in [S1] we got only "essentially 1-based" in the conclusion), provided that no unbounded
open supersimple is interpreted. This will ensure the existence of many stable formulas that
witness forking. In [S1] we dealt with the remaining case by the development of a model theoretic
Baire category theorem in which we analyze more complicated "forking sets" that are related to the
forking topology. This theorem made an essential use of the existence of many stable formulas and
the assumption that the language is countable.
We assume basic knowledge of simple theories; a good textbook on simple theories is [W]. Throughout
this paper we work in a $\kappa$-saturated and $\kappa$-strongly saturated model $\CC$, for some
large $\kappa$, of a complete first order theory $T$.
\section{The dichotomy}
In this section we assume $T=T^{eq}$ is a hypersimple theory and we work in $\CC=\CC^{eq}$. First
recall the definition of a projection-closed family of topologies.
\begin{definition}\em
A family $$\Upsilon=\{\Upsilon_{x,A} \vert\ x \mbox{ is a finite sequence of variables and }
A\subset \CC \mbox{ is small}\}$$ is said to be \em a projection-closed family of topologies \em if
each $\Upsilon_{x,A}$ is a topology on $S_x(A)$ that refines the Stone-topology on $S_x(A)$, this
family is invariant under automorphisms of $\CC$ and change of variables by variables of the same
sort, the family is closed under product by the full Stone spaces $S_y(A)$ (where $y$ is a disjoint
tuple of variables) and closed by projections, namely whenever $\UU(x,y)\in \Upsilon_{xy,A}$,
$\exists y\UU(x,y)\in\Upsilon_{x,A}$.
\end{definition}
From now on $\Upsilon$ denotes a projection-closed family of topologies.
\begin{definition}\label {def ess-1-based}\em
1) A type $p\in S(A)$ is said to be \em s-essentially 1-based over $A_0\subseteq A$ by means of
$\Upsilon$ \em if for every finite tuple $\bar c$ from $p$ and for every $\Upsilon$-open set $\UU$
over $A\bar c$, with the property that $a$ is independent from $A$ over $A_0$ for every $a\in \UU$,
the set $\{a\in \UU \vert\ Cb(a/A\bar c)\not\in bdd(aA_0)\}$ is nowhere dense in the Stone-topology
of $\UU$. We say $p\in S(A)$ is \em s-essentially 1-based by means of $\Upsilon$ \em if $p$ is
s-essentially 1-based over $A$ by means of $\Upsilon$.\\ 2) Let $V$ be an $A_0$-invariant set and
let $p\in S(A_0)$. We say that $p$ is \em analyzable in $V$ by s-essentially 1-based types by means
of $\Upsilon$ \em if there exists $a\models p$ and there exists a sequence $(a_i\vert\
i\leq\alpha)\subseteq dcl(A_0a)$ with $a_\alpha=a$ such that $tp(a_i/A_0\cup\{a_j\vert j<i\})$ is
$V$-internal and s-essentially 1-based over $A_0$ by means of $\Upsilon$ for all $i\leq\alpha$.
\end{definition}
In [S1] we said that \em $p\in S(A)$ is essentially 1-based with respect to $\Upsilon$, \em if 1)
in Definition \ref{def ess-1-based} holds with the additional requirement that $\UU$ is
type-definable. Before stating the main theorem, recall that for an $A$-invariant set $\UU$ and a
type $p$ over $A$, we say that \em $\UU$ is almost $p$-internal (over $A$) \em if $tp(a/A)$ is
almost $p$-internal for every $a\in\UU$. Also, $\UU$ is said to be \em unbounded \em if it contains
the solution set of some non-algebraic type (equivalently, its cardinality is $\geq\kappa$). We can
now phrase the dichotomy.
\begin{theorem}\label{dichotomy thm}
Let $T$ be any hypersimple theory. Let $\Upsilon$ be a projection-closed family of topologies. Let
$p_0$ be a partial type over $\emptyset$ of $SU$-rank 1. Then, either there exists an unbounded
$\Upsilon$-open set (over some small set $A$) that is almost $p_0$-internal (and in particular has
finite $SU$-rank ), or every complete type $p\in S(A)$ that is internal in $p_0$ is s-essentially
1-based over $\emptyset$ by means of $\Upsilon$. In particular, either there exists an unbounded
$\Upsilon$-open set that is almost $p_0$-internal, or whenever $p\in S(A)$ and every non-algebraic
extension of $p$ is non-foreign to $p_0$, $p$ is analyzable in $p_0$ by s-essentially 1-based types
by means of $\Upsilon$.
\end{theorem}
Before proving the dichotomy, note the following easy generalization of [S1, Proposition 4.4]
(recall the domination notation: $b\unrhd_a c$ iff for any $d$ if $d$ independent from $b$ over $a$
then $d$ is independent from $c$ over $a$.)
\begin{proposition}\label {open Cb}
Let $q(x,y)\in S(\emptyset)$ and let $\chi(x,y,z)$ be an $\emptyset$-invariant set such that for
all $(c,b,a)\models \chi(x,y,z)$ we have $b\unrhd_a bc$. Then the set $$\UU=\{(e,c,b,a) \vert\ e\in
acl(Cb(cb/a))\}$$ is relatively Stone-open inside the set
$$F=\{(e,c,b,a)\vert\ \nonforkempty{b}{a}, \models\chi(c,b,a), tp(cb)=q\}.$$ (where
$e$ is taken from a fixed sort too).
\end{proposition}
The proof of Proposition \ref{open Cb} is the same as in [S1], we write it for completeness. Let us
recall the basic notion and fact that are needed for the proof. Recall that a set $\UU$ is said to
be a basic $\tau^f_*$-open set over $C$ if there exists $\psi(x,y,C)\in L(C)$ such that $\UU=\{a
\vert \ \psi(x,aC) \mbox{\ forks\ over } a\}$.
\begin{fact}\label {tau_star}[S1, Lemma 4.3]
Let $C$ be any set and let $\WW=\{(e,a) \vert\ e\in acl(Cb(C/a))\}$ (where $e,a$ are taken from
fixed sorts). Then $\WW$ is a $\tau^f_*$-open set over $C$.
\end{fact}
\proof Note that since $q\in S(\emptyset)$, it is enough to show that for any fixed $c^*b^*\models
q$ the set $\UU^*=\{(e,a) \vert\ e\in acl(Cb(c^*b^*/a))\}$ is relatively Stone-open inside
$$F^*=\{(e,a)\vert\ \nonforkempty{b^*}{a}, \models\chi(c^*,b^*,a)\}.$$ Now, by Fact \ref{tau_star},
we know $\UU^*$ is a $\tau_*^f$-open set over $b^*c^*$. Thus, for some $\psi_i(t_i;w,z,c^*b^*)\in
L(c^*b^*)$ ($i\in I$) we have $\UU^*=\bigcup_i \UU^*_{\psi_i}$ where $$\UU^*_{\psi_i}=\{(e,a)\vert\
\psi_i(t_i;e,a,c^*b^*) \mbox{\ forks over } ea\}.$$
\begin{subclaim}\label{subclaim 1}
For every $(e,a)\in F^*$ we have $(e,a)\in \UU^*_{\psi_i}$ iff $$\forall d
(\psi_i(d;e,a,c^*b^*)\rightarrow \forkempty{da}{b^*})\wedge e\in acl(a).$$
\end{subclaim}
\proof Let $(e,a)\in F^*$. Assuming the left hand side we know $e\in acl(Cb(c^*b^*/a))$, hence
$e\in acl(a)$. Let $d\models\psi_i(z;e,a,c^*b^*)$. If $\nonforkempty{da}{b^*}$, then
$\nonfork{d}{b^*}{a}$. Since $(e,a)\in F^*$, $b^*\unrhd_a b^*c^*$ implies
$\nonfork{d}{b^*c^*}{ea}$, contradicting $(e,a)\in \UU^*_{\psi_i}$. Assume now the right hand side.
By a way of contradiction assume there exists $d\models\psi_i(t_i;e,a,c^*b^*)$ such that
$\nonfork{d}{b^*c^*}{ea}$. Since $e\in acl(a)$, this is equivalent to $\nonfork{d}{b^*c^*}{a}$.
Since
$(e,a)\in F^*$ this is equivalent to $\nonforkempty{da}{b^*}$, contradiction.$\ \ \ \ \Box$\\
\noindent By Subclaim \ref{subclaim 1} we see that each of $\UU^*_{\psi_i}$ and hence $\UU^*$ is
Stone-open relatively inside $F^*$ (since dependence in $b^*$ is a Stone-open condition over $b^*$).$\ \ \ \ \Box$\\
\noindent\textbf{Proof of Theorem \ref{dichotomy thm}} $\Upsilon$ will be fixed and we'll freely
omit the phrase "by means of $\Upsilon$". To see the "In particular" part, work over $A$ and assume
that every $p'\in S(A')$, with $A'\supseteq A$, that is internal in $p_0$, is s-essentially 1-based
over $A$. Moreover, assume $p\in S(A)$ is non-algebraic and every non-algebraic extension of $p$ is
non-foreign to $p_0$. Then, for $a\models p$ there exists $a'\in dcl(Aa)\backslash acl(A)$ such
that $tp(a'/A)$ is $p_0$-internal and thus s-essentially 1-based over $A$ by our assumption. Thus,
by repeating this process we get that $p$ is analyzable in $p_0$ by s-essentially 1-based types. We
now prove the main part. Assume there exists $p\in S(A)$ that is internal in $p_0$, and $p$ is not
s-essentially 1-based over $\emptyset$. By the definition, there exist a finite tuple $d$ of
realizations of $p$ and $b$ that is independent from $d$ over $A$, and a finite tuple $\bar
c\subseteq p_0$ such that $d\in dcl(Ab\bar c)$, and there exists a $\Upsilon$-open set $\UU$ over
$Ad$ such that $a$ is independent from $A$ for all $a\in \UU$ and $\{a\in \UU \vert
Cb(a/Ad)\not\subseteq acl(a)\}$ is not nowhere dense in the Stone-topology of $\UU$. So, since
$\Upsilon$ refines the Stone-topology, by intersecting $\UU$ with a definable set, we may assume
that $\{a\in \UU \vert Cb(a/Ad)\not\subseteq acl(a)\}$ is dense in the Stone-topology of $\UU$.
\noindent Now, for each (finite) subsequence $\bar c_0$ of $\bar c$, let $$F_{\bar c_0}=\{ a\in \UU
\vert\ \exists b',\bar c'_0,\bar c'_1\ \mbox{s.t.}\ tp(b'\bar c'_0\bar c'_1/Ad)=tp(b\bar c_0(\bar
c\backslash \bar c_0)/Ad)\ \mbox{and} \nonforkempty{a}{Ab'\bar c'_0}\}.$$ Note that since $d$ is
independent from $b$ over $A$, any $a\in\UU$ is independent from $Ab'$ whenever
$tp(b'/Ad)=tp(b/Ad)$ and $\nonfork{a}{b'}{Ad}$. Thus $F_{\langle\rangle}=\UU$. Let $\bar c^*_0$ be
a maximal subsequence (with respect to inclusion) of $\bar c$ such that $F_{\bar c^*_0}$ has
non-empty Stone-interior in $\UU$ over $Ad$ (note that $F_{\bar c}$ has no Stone-interior
relatively in $\UU$). Let $\UU^*=\bigcap_{\bar c^*_0\subset\bar c'\subseteq\bar c} \UU\backslash
F_{\bar c'}$. Note that each $F_{\bar c'}$ is Stone closed relatively in $\UU$. Thus $\UU^*$ is
Stone-dense and open in $\UU$ and therefore there exists a
non-empty relatively Stone-open in $\UU$ set $W^*\subseteq F_{\bar c_0^*}\cap \UU^*$.
\begin{subclaim}\label{subclaim0_main}
$W^*$ is a non-empty $\Upsilon$-open set over $Ad$ such that $\{a\in W^* \vert\
Cb(a/Ad)\not\subseteq acl(a)\}$ is dense in the Stone-topology of $W^*$ and for every $a\in W^*$ we
have: there exists $b'\bar c'_0\bar c'_1\models tp(b\bar c^*_0(\bar c\backslash \bar c^*_0)/Ad)$
such that $a$ is independent from $Ab'\bar c'_0$ over $\emptyset$ and moreover, for every $b'\bar
c'_0\bar c'_1\models tp(b\bar c^*_0(\bar c\backslash \bar c^*_0)/Ad)$ such that $a$ is independent
from $Ab'\bar c'_0$ we necessarily have $\bar c'_1\in acl(aAb'\bar c'_0)$.
\end{subclaim}
\proof As $p_0$ has $SU$-rank 1, this is a conclusion of our construction.$\ \ \ \ \Box$\\
\noindent Let us now define a set $V$ over $Ad$ by\\ $$V=\{(e',b',\bar c'_0,\bar c'_1,a') \vert \
\mbox{if}\ tp(b'\bar c'_0\bar c'_1/Ad)=tp(b\bar c^*_0(\bar c\backslash \bar c_0^*)/Ad)\ \mbox{and}
\nonforkempty{a'}{Ab'\bar c'_0}$$ $$\mbox{then}\ e'\in acl(Cb(Ab'\bar c'_0\bar c'_1/a'))\}.$$
\noindent Let $V^*=\{e' \vert \exists a'\in W^*\ \forall b',\bar c'_0,\bar c'_1\ V(e',b',\bar
c'_0,\bar c'_1,a')\}.$
\begin{subclaim}\label{subclaim1_main}
$V^*$ is a $\Upsilon$-open set over $Ad$.
\end{subclaim}
\proof By Proposition \ref{open Cb} and Subclaim \ref{subclaim0_main}, there exists a Stone-open set $V'$
over $Ad$ such that for all $a'\in W^*$ and for all $e',b',\bar c'_0,\bar c'_1$ we have
$V'(e',b',\bar c'_0,\bar c'_1,a')$ if and and only if $V(e',b',\bar c'_0,\bar c'_1,a')$. Thus, we
may replace $V$ by $V'$ in the definition of $V^*$. As Stone-open sets are closed under the
$\forall$ quantifier, the $\Upsilon$ topology refines the Stone-topology and closed under product
by a full Stone-space and closed under projections, we conclude that $V^*$ is a $\Upsilon$-open
set.$\ \ \ \ \Box$
\begin{subclaim}\label{subclaim2_main}
For appropriate sort for $e'$, the set $V^*$ is unbounded and is almost $p_0$-internal (over $Ad$)
and thus has finite $SU$-rank over $Ad$.
\end{subclaim}
\proof First, note the following general observation.
\begin{fact}\label{dcl_cb remark}\em
Assume $d\in dcl(c)$. Then $Cb(d/a)\in dcl(Cb(c/a))$ for all $a$.
\end{fact}
\noindent Let $a^*\in W^*$ be such that $Cb(a^*/Ad)\not\subseteq acl(a^*)$. Then
$Cb(Ad/a^*)\not\subseteq acl(Ad)$. By Fact \ref{dcl_cb remark}, there exists $e^*\not\in acl(Ad)$
such that $e^*\in acl(Cb(Ab'\bar c'_0\bar c'_1/a^*))$ for all $b'\bar c'_0\bar c'_1\models tp(b\bar
c^*_0(\bar c\backslash \bar c^*_0)/Ad)$. In particular, $e^*\in V^*$. Thus, if we fix the sort for
$e'$ in the definition of $V^*$ to be the sort of $e^*$, then $V^*$ is unbounded. Now, let $e'\in
V^*$. Then for some $a'\in W^*$, $\models V(e',\bar c'_0,\bar c'_1,b',a')$ for all $b',\bar
c'_0,\bar c'_1$. By Subclaim \ref{subclaim0_main}, there exists $b'\bar c'_0\bar c'_1\models
tp(b\bar c^*_0(\bar c\backslash \bar c^*_0)/Ad)$ such that $a'\ \mbox{is\ independent\ from}\
Ab'\bar c'_0\ \mbox{over}\ \emptyset$. Thus, by the definition of $V^*$ and $V$, $e'\in
acl(Cb(Ab'\bar c'_0\bar c'_1/a'))$. Since $Ab'$ is independent from $a'$ over $\emptyset$, $tp(e')$
is almost-$p_0$-internal (as $Cb(Ab'\bar c'_0\bar c'_1/a')$ is in the definable closure of any
Morley sequence of $Lstp(Ab'\bar c'_0\bar c'_1/a')$ ), and in particular $tp(e'/Ad)$ is almost
$p_0$-internal (note that, in general, whenever $q=tp(a/A)$ is internal in an $\emptyset$-invariant
set $\RR$
then any extension of $q$ is almost $\RR$-internal) and therefore $tp(e'/Ad)$ has finite $SU$-rank. $\ \ \ \ \Box$\\
\noindent Thus $V^*$ is the required set. $\ \ \ \ \Box$\\
We now draw some consequences of the above dichotomy for countable languages.
\begin{theorem}\label{cor1}
Let $T$ be any countable hypersimple theory. Let $\Upsilon$ be a projection-closed family of
topologies such that $\{ a\in \CC^x \vert a\not\in acl(A)\}\in \Upsilon_{x,A}$ for all $x$ and set
$A$ . Let $p_0$ be a partial type over $\emptyset$ of $SU$-rank 1. Then, either there exists an
unbounded type-definable $\Upsilon$-open set over some small set that is almost $p_0$-internal and
has \textbf{bounded} finite $SU$-rank, or every complete type $p\in S(A)$ that is internal in $p_0$
is essentially 1-based over $\emptyset$ by means of $\Upsilon$. In particular, either there exists
an unbounded $\Upsilon$-open set that is almost $p_0$-internal and has \textbf{bounded} finite
$SU$-rank, or whenever $p\in S(A)$, where $A$ is countable, and every non-algebraic extension of
$p$ is non-foreign to $p_0$, $p$ is analyzable in $p_0$ by essentially 1-based types by means of
$\Upsilon$.
\end{theorem}
\proof We go back to the proof of Theorem \ref{dichotomy thm} (the main part); we start with $p\in
S(A)$ that is $p_0$-internal and not essentially 1-based over $\emptyset$ and apply the same proof
(but note that in the proof of Theorem \ref{dichotomy thm} we assumed $p$ is not s-essentially
1-based). So, now $\UU$ is assumed to be a \textbf{type-definable} $\Upsilon$-open set over $Ad$.
\begin{subclaim}
We may assume $W^*$ is type-definable and $\Upsilon$-open over $Ad$ and there exists
$V^{**}\subseteq V^*$ that is unbounded, type-definable and $\Upsilon$-open over $Ad$.
\end{subclaim}
\proof In the proof of Theorem \ref{dichotomy thm} the set $W^*$ is chosen to be a non-empty
intersection of $\UU$ with a Stone-open set over $Ad$, so we could instead take it to be a
non-empty intersection of $\UU$ with a definable subset of this Stone-open set (and still
$W^*\subseteq F_{\bar c_0^*}\cap \UU^*$). Since $\UU$ is $\Upsilon$-open and type-definable, $W^*$
is type-definable and $\Upsilon$-open over $Ad$. Now, by the definition of $V^*$ and the proof of
Subclaim \ref{subclaim1_main} there exist a Stone open set $V_0$ over $Ad$ such that $V^*=\{ e'
\vert \exists a'\in W^*\ (V_0(e',a'))\}$. From this we easily get the required set $V^{**}$ (by
replacing $V_0$ by a definable set and using the fact that $W^*$ is type-definable and that $\Upsilon$ is a
projection-closed family of topologies). $\ \ \ \ \Box$\\
\noindent By the proof of Subclaim \ref{subclaim2_main} we know that for all $e'\in V^{**}$ we have
$e'\in acl(Cb(Ab'\bar c'_0\bar c'_1/a'))$ for some $a'\in W^*$ and some $b',\bar c'_0,\bar c'_1$
such that $a'$ is independent from $Ab'\bar c'_0\ \mbox{over}\ \emptyset$ and $b'\bar c'_0\bar
c'_1\models tp(b\bar c^*_0(\bar c\backslash \bar c^*_0)/Ad)$. Let $q=tp(Ab\bar c^*_0)$. For every
$\chi=\chi(x,y_0,...,y_n,\bar z)\in L$ (for some $n<\omega$) such that $\forall y_0 y_1 ...y_n \bar
z\ \exists^{<\infty}x\ \chi(x,y_0,y_1,...y_n,\bar z)$, and $m<\omega$ let
$$F_{\chi,m}=\{e\in V^{**} \vert\ \models\chi(e,C_0,C_1,..C_n,\bar c)\ \mbox{for\ some\ } \bar c\in
p_0^m\ \mbox{and\ some}\ \emptyset-\mbox{independent\ sequence}$$ $\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \
\ \ \ \ \ \ \ \ \ (C_i\vert i\leq n)\ \mbox{of\ realization of}\ q\ \mbox{with}
\nonforkempty{e}{(C_i\vert i\leq n)} \}.$\\
\noindent By the aforementioned, we get that $V^{**}\subseteq \bigcup_{m,\chi} F_{m,\chi}$ (the
union is over each $m,\chi$ as above). By the Baire category theorem applied to the Stone-topology
of the Stone-closed set $V^{**}\backslash acl(Ad)$, there exists $\theta\in L(Ad)$ such that
$$\tilde V\equiv \theta^\CC\cap(V^{**}\backslash acl(Ad))\neq\emptyset\ \mbox{and\ } \tilde
V\subseteq F_{m^*,\chi^*}$$ for some $m^*,\chi^*$ as above. Clearly, $\tilde V$ is unbounded,
type-definable and $\Upsilon$-open (by the assumptions on $\Upsilon$). Now, for every $a\in \tilde
V$, $SU(a/Ad)\leq m^*$ and $tp(a/Ad)$ is almost $p_0$-internal (as $tp(a)$ is almost
$p_0$-internal, and $SU(a)\leq m^*$ by the definition of $F_{m^*,\chi^*}$). This completes the
proof of the first part of the theorem. The rest follow easily by repeated applications of the
first part (when working over $A$).$\ \ \ \ \Box$
$\\$
Recall that $T$ is PCFT if its forking-topologies is a projection-closed family of topologies, that
is, whenever $\UU(x,y)$ is a $\tau^f$-open set over a small set $A$, $\exists y\UU(x,y)$ is a
$\tau^f$-open set over $A$. Applying Theorem \ref{cor1} for the special case of the
forking-topologies we conclude the following.
\begin{corollary}
Let $T$ be any countable hypersimple theory with PCFT. Let $p_0$ be a partial type over $\emptyset$
of $SU$-rank 1. Then, either there exists a weakly-minimal formula that is almost $p_0$-internal,
or every complete type $p\in S(A)$ that is internal in $p_0$ is essentially 1-based over
$\emptyset$ by means of $\tau^f$. In particular, either there exists a weakly-minimal formula that
is almost $p_0$-internal, or whenever $p\in S(A)$, where $A$ is countable, and every non-algebraic
extension of $p$ is non-foreign to $p_0$, $p$ is analyzable in $p_0$ by essentially 1-based types
by means of $\tau^f$.
\end{corollary}
\proof Our assumptions are clearly a special case of the assumptions of Theorem \ref{cor1}, thus
we only need to prove the first part. By the conclusion of Theorem \ref{cor1}, we may assume that there
exists a $\tau^f$-open set $\UU$ of bounded finite $SU$-rank over some small set $A$ that is almost
$p_0$-internal. Recall now [S0, Proposition 2.13]:
\begin{fact}\label{tau bounded SU}
Let $\UU$ be an unbounded $\tau^f$-open set over some set $A$. Assume $\UU$ has bounded finite
$SU$-rank. Then there exists a set $B\supseteq A$ and $\theta(x)\in L(B)$ of $SU$-rank 1 such that
$\theta^\CC\subseteq \UU\cup acl(B)$.
\end{fact}
By Fact \ref{tau bounded SU}, there exists exists a weakly-minimal $\theta(x,b)\in L(B)$ for some
small set $B\supseteq A$, such that $\theta^\CC\subseteq \UU\cup acl(B)$. Now, $tp(a/B)$ is almost
$p_0$-internal for every $a\in \theta^\CC$, and so $tp(a/b)$ ($b$ is the parameter of $\theta(x,b)$
) is almost $p_0$-internal over $b$ for every $a\in \theta^\CC$ (by taking non-forking
extensions).$\ \ \ \ \Box$\\
We now state the main conclusion for uncountable hypersimple unidimensional theories.
\begin{definition}
We say that $T$ is s-essentially 1-based if for every $SU$-rank 1 partial type $p_0$ over some $A$,
every $p\in S(A)$ that is internal in $p_0$ is s-essentially 1-based by means of $\tau^f$.
\end{definition}
\begin{corollary}\em
Let $T$ be a hypersimple unidimensional theory that is not s-essentially 1-based. Then $T$ is
supersimple.
\end{corollary}
\proof First, recall the following fact [S1, Corollary 3.15] (an $A$-invariant set $\UU$ is called
\em supersimple \em if $SU(a/A)<\infty$ for every $a\in\UU$).
\begin{fact}\label {fact1}
Let $T$ be a hypersimple unidimensional theory and work in $\CC=\CC^{eq}$. Let $p\in S(A)$ and let
$\UU$ be an unbounded $\tau^f$-open set over $A$. Then $p$ is analyzable in $\UU$ in finitely many
steps. In particular, for such $T$ the existence of an unbounded supersimple $\tau^f$-open set over
some small set $A$ implies $T$ is supersimple.
\end{fact}
\noindent Now, assume $T$ is a hypersimple unidimensional theory that is not s-essentially 1-based.
By Theorem \ref{dichotomy thm}, there exists an unbounded $\tau^f$-open set of finite $SU$-rank
over some small set. By Fact \ref{fact1}, every complete type has finite $SU$-rank.$\ \ \ \ \Box$
|
\section{Introduction}
Let $\Omega\subset\rz^3$ be a bounded and convex domain.
It is well known that, e.g., by Rellich's selection theorem
using standard indirect arguments,
the Poincar\'e\footnote{The estimate \eqref{poincarehoc}
is often called Friedrichs'/Steklov inequality as well.} inequalities
\begin{align}
\mylabel{poincarehoc}
\exists\,c_{\mathtt{p},\circ}&>0&
\forall\,u&\in\Hgen{1}{}{\circ}&
\norm{u}&\leqc_{\mathtt{p},\circ}\norm{\nabla u},\\
\mylabel{poincareho}
\exists\,c_{\mathtt{p}}&>0&
\forall\,u&\in\Hgen{1}{}{}\cap\mathbb{R}^{\bot}
&\norm{u}&\leqc_{\mathtt{p}}\norm{\nabla u}
\end{align}
hold. Here, $c_{\mathtt{p},\circ}$ and $c_{\mathtt{p}}$ are the Poincar\'e constants, which satisfy
$$0<c_{\mathtt{p},\circ}=1/\sqrt{\lambda_{1}}<1/\sqrt{\mu_{2}}=c_{\mathtt{p}},$$
where $\lambda_{1}$ is the first Dirichlet
and $\mu_{2}$ the second Neumann eigenvalue of the Laplacian.
By $\scp{\,\cdot\,}{\,\cdot\,}$ and $\norm{\,\cdot\,}$
we denote the standard inner product and induced norm
in $\Lgen{2}{}$ and we will write the usual $\Lgen{2}{}$-Sobolev spaces
as $\Hgen{1}{}{}$ and $\Hgen{1}{}{\circ}$, the latter is defined as the closure in $\Hgen{1}{}{}$
of smooth and compactly supported test functions.
All spaces and norms are defined on $\Omega$.
Moreover, we introduce the standard Sobolev spaces
for the rotation and divergence by $\r$ and $\d$.
As before, we will denote the closures of test vector fields
in the respective graph norms by $\color{red}$ and $\divgen{}{\circ}$.
An index zero at the lower right corner of the latter spaces
indicates a vanishing derivative, e.g.,
$$\mathbb{R}:=\set{E\in\r}{\rot E=0},\quad\divgen{0}{\circ}:=\set{E\in\divgen{}{\circ}}{\div E=0}.$$
As $\Omega$ is convex, it is especially simply connected
and has got a connected boundary. Hence, the Neumann and Dirichlet fields
of $\Omega$ vanish, i.e., $\mathbb{R}\cap\divgen{0}{\circ}=\rotgen{0}{\circ}\cap\mathbb{R}=\{0\}$.
By the Maxwell compactness properties, i.e.,
the compactness of the two embeddings
$$\color{red}\cap\d\hookrightarrow\Lgen{2}{},\quad\r\cap\divgen{}{\circ}\hookrightarrow\Lgen{2}{},$$
(and again by a standard indirect argument) the Maxwell inequalities
\begin{align}
\mylabel{maxestelecconv}
\exists\,c_{\mathtt{m,t}}&>0&
\forall\,E&\in\color{red}\cap\d&
\norm{E}&\leqc_{\mathtt{m,t}}\big(\norm{\rot E}^2+\norm{\div E}^2\big)^{1/2},\\
\mylabel{maxestmagconv}
\exists\,c_{\mathtt{m,n}}&>0&
\forall\,H&\in\r\cap\divgen{}{\circ}&
\norm{H}&\leqc_{\mathtt{m,n}}\big(\norm{\rot H}^2+\norm{\div H}^2\big)^{1/2}
\end{align}
hold. To the best of the author's knowledge,
general bounds for the Maxwell constants $c_{\mathtt{m,t}}$ and $c_{\mathtt{m,n}}$ are missing.
On the other hand, at least estimates for $c_{\mathtt{m,t}}$ and $c_{\mathtt{m,n}}$ from above are very important
from the point of view of applications, such as preconditioning
or a priori and a posteriori error estimation for numerical methods.
In the paper at hand we will prove that
\begin{align}
\mylabel{cmcpintro}
c_{\mathtt{p},\circ}\leqc_{\mathtt{m,t}}\leqc_{\mathtt{m,n}}=c_{\mathtt{p}}\leq\diam(\om)/\pi
\end{align}
holds true. We note that \eqref{cmcpintro}
is already well known in two dimensions,
even for general Lipschitz domains $\Omega\subset\mathbb{R}^{2}$
(except of the last inequality),
but new in three dimensions.
Furthermore, the last inequality in \eqref{cmcpintro}
has been proved in the famous paper of Payne and Weinberger \cite{payneweinbergerpoincareconvex},
where also the optimality of the estimate was shown.
This paper contains a small mistake, which has been corrected in \cite{bebendorfpoincareconvex}.
\section{Results and Proofs}
We start with an inequality for irrotational fields.
\begin{lem}
\mylabel{lemNarbdiv}
For all $E\in\nabla\Hgen{1}{}{\circ}\cap\d$ and all $H\in\nabla\Hgen{1}{}{}\cap\divgen{}{\circ}$
$$\norm{E}\leqc_{\mathtt{p},\circ}\norm{\div E},\quad\norm{H}\leqc_{\mathtt{p}}\norm{\div H}.$$
\end{lem}
\begin{proof}
Let $\varphi\in\Hgen{1}{}{\circ}$ with $E=\nabla\varphi$. By \eqref{poincarehoc} we get
$$\norm{E}^2
=\scp{E}{\nabla\varphi}
=-\scp{\div E}{\varphi}
\leq\norm{\div E}\norm{\varphi}
\leqc_{\mathtt{p},\circ}\norm{\div E}\norm{\nabla\varphi}
=c_{\mathtt{p},\circ}\norm{\div E}\norm{E}.$$
Let $\varphi\in\Hgen{1}{}{}$ with $H=\nabla\varphi$ and $\varphi\bot\mathbb{R}$.
Since $H\in\divgen{}{\circ}$ and by \eqref{poincareho} we obtain
$$\norm{H}^2
=\scp{H}{\nabla\varphi}
=-\scp{\div H}{\varphi}
\leq\norm{\div H}\norm{\varphi}
\leqc_{\mathtt{p}}\norm{\div H}\norm{\nabla\varphi}
=c_{\mathtt{p}}\norm{\div H}\norm{H},$$
completing the proof.
\end{proof}
\begin{rem}
\mylabel{remNarbdiv}
Clearly, Lemma \ref{lemNarbdiv} extends to arbitrary
Lipschitz domains $\Omega\subset\mathbb{R}^{N}$, $N\in\mathbb{N}$.
\end{rem}
As usual in the theory of Maxwell's equations, we need another crucial tool,
the Helmholtz decompositions of vector fields into irrotational and solenoidal vector fields.
For convex domains, these decompositions are very simple. We have
\begin{align}
\mylabel{helmdecoconvex}
\Lgen{2}{}=\nabla\Hgen{1}{}{\circ}\oplus\rot\r,\quad\Lgen{2}{}=\nabla\Hgen{1}{}{}\oplus\rot\color{red},
\end{align}
where $\oplus$ denotes the orthogonal sum in $\Lgen{2}{}$.
We note
$$\rotgen{0}{\circ}=\nabla\Hgen{1}{}{\circ},\quad\mathbb{R}=\nabla\Hgen{1}{}{},\quad\divgen{0}{}=\rot\r,\quad\divgen{0}{\circ}=\rot\color{red}.$$
Moreover, with
$$\overset{\circ}{\cR}:=\color{red}\cap\rot\r,\quad\cR:=\r\cap\rot\color{red}$$
we have
\begin{align}
\mylabel{helmdecoconvexaugmented}
\color{red}=\nabla\Hgen{1}{}{\circ}\oplus\overset{\circ}{\cR},\quad\r=\nabla\Hgen{1}{}{}\oplus\cR
\end{align}
and see
$$\rot\color{red}=\rot\overset{\circ}{\cR},\quad\rot\r=\rot\cR.$$
We note that all occurring spaces of range-type are closed subspaces of $\Lgen{2}{}$,
which follows immediately by the estimates \eqref{poincarehoc}-\eqref{maxestmagconv}.
More details about the Helmholtz decompositions can be found e.g. in \cite{leisbook}.
To get similar inequalities for solenoidal vector fields as in Lemma \ref{lemNarbdiv} we
need a crucial lemma from \cite[Theorem 2.17]{amrouchebernardidaugegiraultvectorpot},
see also \cite{saranenineqfried,grisvardbook,giraultraviartbook,costabelcoercbilinMax}
for related partial results.
\begin{lem}
\mylabel{french}
Let $E$ belong to $\color{red}\cap\d$ or $\r\cap\divgen{}{\circ}$. Then $E\in\Hgen{1}{}{}$ and
\begin{align}
\mylabel{frenchformula}
\norm{\nabla E}^2\leq\norm{\rot E}^2+\norm{\div E}^2.
\end{align}
\end{lem}
We emphasize that for $E\in\Hgen{1}{}{\circ}$ and any domain $\Omega\subset\rz^3$
\begin{align}
\mylabel{frenchformulaequal}
\norm{\nabla E}^2=\norm{\rot E}^2+\norm{\div E}^2
\end{align}
holds since $-\Delta=\rot\rot-\nabla\div$.
This formula is no longer valid if $E$ has just the tangential
or normal boundary condition but for convex domains
the inequality \eqref{frenchformula} remains true.
\begin{lem}
\mylabel{lemNThreerot}
For all vector fields $E$ in $\color{red}\cap\rot\r$ or $\r\cap\rot\color{red}$
$$\norm{E}\leqc_{\mathtt{p}}\norm{\rot E}.$$
\end{lem}
\begin{proof}
Let $E\in\rot\r=\rot\cR$ and $\Phi\in\cR$ with $\rot\Phi=E$.
Then $\Phi\in\Hgen{1}{}{}$ by Lemma \ref{french} since $\cR=\r\cap\divgen{0}{\circ}$.
Moreover, $\Phi=\rot\Psi$ can be represented by some $\Psi\in\color{red}$.
Hence, for any constant vector $a\in\rz^3$ we have $\scp{\Phi}{a}=0$.
Thus, $\Phi$ belongs to $\Hgen{1}{}{}\cap(\rz^3)^{\bot}$.
Then, since $E\in\color{red}$ and by Lemma \ref{french} we get
$$\norm{E}^2
=\scp{E}{\rot\Phi}
=\scp{\rot E}{\Phi}
\leq\norm{\rot E}\norm{\Phi}
\leqc_{\mathtt{p}}\norm{\rot E}\norm{\nabla\Phi}
\leqc_{\mathtt{p}}\norm{\rot E}\norm{\underbrace{\rot\Phi}_{=E}}.$$
If $E\in\rot\color{red}$ there exists $\Phi\in\color{red}$ with $\rot\Phi=E$.
Using \eqref{helmdecoconvexaugmented} we decompose
$$E=E_{0}+E_{\rot}\in\mathbb{R}\oplus\cR.$$
Then, $\rot E_{\rot}=\rot E$ and
again by Lemma \ref{french} we see $E_{\rot}\in\Hgen{1}{}{}$.
Let $a\in\rz^3$ such that $E_{\rot}-a\in\Hgen{1}{}{}\cap(\rz^3)^{\bot}$.
Since $\Phi\in\color{red}$, $\scp{\rot\Phi}{H_{0}}$ and $\scp{\rot\Phi}{a}$ vanish.
By Lemma \ref{french}
$$\norm{E}^2
=\scp{\rot\Phi}{E}
=\scp{\underbrace{\rot\Phi}_{=E}}{E_{\rot}-a}
\leq\norm{E}\norm{E_{\rot}-a}\\
\leqc_{\mathtt{p}}\norm{E}\norm{\nabla E_{\rot}}
\leqc_{\mathtt{p}}\norm{E}\norm{\underbrace{\rot E_{\rot}}_{=\rot E}}$$
holds, which completes the proof.
\end{proof}
\begin{rem}
\mylabel{lemNThreerottwod}
It is well known that Lemma \ref{lemNThreerot}
holds in two dimensions for any Lipschitz domain $\Omega\subset\mathbb{R}^2$.
This follows immediately from Lemma \ref{lemNarbdiv}
if we take into account that in two dimensions the rotation $\rot$
is given by the divergence $\div$ after $90^{\circ}$-rotation
of the vector field to which it is applied.
\end{rem}
\begin{theo}
\mylabel{maintheo}
For all vector fields $E\in\color{red}\cap\d$ and $H\in\r\cap\divgen{}{\circ}$
$$\norm{E}^2
\leqc_{\mathtt{p},\circ}^2\norm{\div E}^2+c_{\mathtt{p}}^2\norm{\rot E}^2,\quad
\norm{H}^2
\leqc_{\mathtt{p}}^2\norm{\div H}^2+c_{\mathtt{p}}^2\norm{\rot H}^2$$
hold, i.e., $c_{\mathtt{m,t}},c_{\mathtt{m,n}}\leqc_{\mathtt{p}}$.
Moreover, $c_{\mathtt{p},\circ}\leqc_{\mathtt{m,t}}\leqc_{\mathtt{m,n}}=c_{\mathtt{p}}\leq\diam(\Omega)/\pi$.
\end{theo}
\begin{proof}
By the Helmholtz decomposition \eqref{helmdecoconvex} we have
$$\color{red}\cap\d\ni E=E_{\nabla}+E_{\rot}\in\nabla\Hgen{1}{}{\circ}\oplus\rot\r$$
with $E_{\nabla}\in\nabla\Hgen{1}{}{\circ}\cap\d$ and $E_{\rot}\in\color{red}\cap\rot\r$ as well as
$\div E_{\nabla}=\div E$ and $\rot E_{\rot}=\rot E$.
By Lemma \ref{lemNarbdiv} and Lemma \ref{lemNThreerot} and orthogonality we obtain
$$\norm{E}^2=\norm{E_{\nabla}}^2+\norm{E_{\rot}}^2
\leqc_{\mathtt{p},\circ}^2\norm{\div E}^2+c_{\mathtt{p}}^2\norm{\rot E}^2.$$
Similarly we have
$$\r\cap\divgen{}{\circ}\ni H=H_{\nabla}+H_{\rot}\in\nabla\Hgen{1}{}{}\oplus\rot\color{red}$$
with $H_{\nabla}\in\nabla\Hgen{1}{}{}\cap\divgen{}{\circ}$ and $H_{\rot}\in\r\cap\rot\color{red}$ as well as
$\div H_{\nabla}=\div H$ and $\rot H_{\rot}=\rot H$.
As before,
$$\norm{H}^2=\norm{H_{\nabla}}^2+\norm{H_{\rot}}^2
\leqc_{\mathtt{p}}^2\norm{\div H}^2+c_{\mathtt{p}}^2\norm{\rot H}^2.$$
This shows the upper bounds.
For the lower bounds, let $\lambda_{1}$ be the first Dirichlet eigenvalue
of the negative Laplacian $-\Delta$, i.e.,
$$\frac{1}{c_{\mathtt{p},\circ}^2}
=\lambda_{1}
=\inf_{0\neq u\in\Hgen{1}{}{\circ}}\frac{\norm{\nabla u}^2}{\norm{u}^2},$$
and let $u\in\Hgen{1}{}{\circ}$ be an eigenfunction to $\lambda_{1}$.
Note that $u$ satisfies
$$\forall\,\varphi\in\Hgen{1}{}{\circ}\quad\scp{\nabla u}{\nabla\varphi}=\lambda_{1}\scp{u}{\varphi}.$$
Then $0\neq E:=\nabla u\in\nabla\Hgen{1}{}{\circ}\cap\d=\rotgen{0}{\circ}\cap\d$
and $-\div E=-\div\nabla u=\lambda_{1}u$.
By \eqref{maxestelecconv} and \eqref{poincarehoc} we have
\begin{align*}
\norm{E}
&\leqc_{\mathtt{m,t}}\norm{\div E}
=c_{\mathtt{m,t}}\lambda_{1}\norm{u}
\leqc_{\mathtt{m,t}}\lambda_{1}c_{\mathtt{p},\circ}\norm{\nabla u}
=\frac{c_{\mathtt{m,t}}}{c_{\mathtt{p},\circ}}\norm{E},
\end{align*}
yielding $c_{\mathtt{p},\circ}\leqc_{\mathtt{m,t}}$.
Now, let $\mu_{2}$ be the second Neumann eigenvalue
of the negative Laplacian $-\Delta$, i.e.,
$$\frac{1}{c_{\mathtt{p}}^2}
=\mu_{2}
=\inf_{0\neq u\in\Hgen{1}{}{}\cap\mathbb{R}^{\bot}}\frac{\norm{\nabla u}^2}{\norm{u}^2},$$
and let $u\in\Hgen{1}{}{}\cap\mathbb{R}^{\bot}$ be an eigenfunction to $\mu_{2}$.
Note that $u$ satisfies
$$\forall\,\varphi\in\Hgen{1}{}{}\cap\mathbb{R}^{\bot}\quad\scp{\nabla u}{\nabla\varphi}=\mu_{2}\scp{u}{\varphi}$$
and that this relation holds even for all $\varphi\in\Hgen{1}{}{}$.
Then $0\neq H:=\nabla u\in\nabla\Hgen{1}{}{}\cap\divgen{}{\circ}=\mathbb{R}\cap\divgen{}{\circ}$
and satisfies $-\div H=-\div\nabla u=\mu_{2}u$.
By \eqref{maxestmagconv} and \eqref{poincareho} we have
\begin{align*}
\norm{H}
&\leqc_{\mathtt{m,n}}\norm{\div H}
=c_{\mathtt{m,n}}\mu_{2}\norm{u}
\leqc_{\mathtt{m,n}}\mu_{2}c_{\mathtt{p}}\norm{\nabla u}
=\frac{c_{\mathtt{m,n}}}{c_{\mathtt{p}}}\norm{H},
\end{align*}
yielding $c_{\mathtt{p}}\leqc_{\mathtt{m,n}}$ and completing the proof.
\end{proof}
\begin{rem}
\mylabel{eigenvalues}
\begin{itemize}
\item[\bf(i)]
It is unclear but most probable that $c_{\mathtt{p},\circ}<c_{\mathtt{m,t}}<c_{\mathtt{m,n}}=c_{\mathtt{p}}$ holds.
In forthcoming publications \cite{paulymaxconst1,paulymaxconst2}
we will show more and sharper estimates on the Maxwell constants,
showing additional and sharp relations
between the Maxwell and the Poincar\'e/Friedrichs/Steklov constants.
\item[\bf(ii)]
Our results extend also to all polyhedra which allow the
$\Hgen{1}{}{}$-regularity of the Maxwell spaces $\color{red}\cap\d$ and $\r\cap\divgen{}{\circ}$
or to domains whose boundaries consist of combinations of convex boundary parts
and polygonal parts which allow the $\Hgen{1}{}{}$-regularity.
Is is shown in \cite[Theorem 4.1]{costabelcoercbilinMax}
that \eqref{frenchformula}, even \eqref{frenchformulaequal},
still holds for all $E\in\Hgen{1}{}{}\cap\color{red}$ or $E\in\Hgen{1}{}{}\cap\divgen{}{\circ}$
if $\Omega$ is a polyhedron\footnote{The crucial point is
that the unit normal is piecewise constant and hence the curvature is zero.}.
We note that even some non-convex polyhedra admit the $\Hgen{1}{}{}$-regularity of the Maxwell spaces
depending on the angle of the corners, which are not allowed to by too pointy.
\item[\bf(iii)]
Looking at the proof, the lower bounds $c_{\mathtt{p},\circ}\leqc_{\mathtt{m,t}}$ and $c_{\mathtt{p}}\leqc_{\mathtt{m,n}}$
remain true in more general situations, i.e.,
for bounded Lipschitz\footnote{The Lipschitz assumption
can also be weakened. It is sufficient that $\Omega$ admits the Maxwell compactness properties.}
domains $\Omega\subset\rz^3$.
\end{itemize}
\end{rem}
\begin{acknow}
The author is deeply indebted to Sergey Repin not only for bringing his attention
to the problem of the Maxwell constants in 3D.
Moreover, the author wants to thank Sebastian Bauer und Karl-Josef Witsch
for nice and deep discussions.
\end{acknow}
\bibliographystyle{plain}
|
\section{Introduction}
E. Cartan \cite{Car} introduced the axiom of $n$-planes as: A
Riemannian manifold $M$ of dimension $m\geq3$ is said to satisfy the
axiom of $n$-planes, where $n$ is a fixed integer $2\leq n\leq m-1$,
if for each point $p\in M$ and each $n$-dimensional subspace
$\sigma$ of the tangent space $T_p(M)$, there exists an
$n$-dimensional totally geodesic submanifold $N$ such that $p\in N$
and $T_p(N)=\sigma.$ He also gave a criterion for constancy of
sectional curvature for any Riemannian manifold of dimension
$m\geq3$ in the following theorem.
\begin{theorem} A Riemannian manifold of dimension $m\geq3$ with the axiom of $n$-planes is a real space form.
\end{theorem}
D.S. Leung and K. Nomizu \cite{Leu} introduced \emph{the axiom of
$n$-spheres} by using totally umbilical submanifold $N$ with
parallel mean curvature vector field instead of
totally geodesic submanifold $N$ in the axiom of $n$-planes. They proved a generalization of Theorem 1.1.\\
Later on, Cartan's idea was applied to almost Hermitian manifolds in
various studies. K\"{a}hlerian manifolds were studied in
\cite{BChe,Gol,Ha,Kassa,No,Yama,Yano}. The articles \cite{Vanh} and
\cite{Yam} discussed nearly K\"{a}hlerian (almost Tachibana)
manifolds. The results concerning larger classes of almost Hermitian
manifolds can be found in \cite{Kas,Kass,Tas,Van,Vanh}. Here, we
shall call the criterions used in all of the above papers as
\emph{Cartan-type criterions}.
\section{Preliminaries}
\subsection{Some classes of almost Hermitian manifolds}
Let $M$ be an almost Hermitian manifold with an almost complex
structure $J$ in its tangent bundle and a Riemannian metric $g$ such
that $g(JX,JY)=g(X,Y)$ for all $X,Y\in\chi(M)$, where $\chi(M)$ is
the Lie algebra of $C^{\infty}$ vector fields on $M$. Let $\nabla$
be the Riemannian connection on $M$. The Riemannian curvature
tensor $R$ associated with $\nabla$ is defined by
$R(X,Y)=\nabla_{[X,Y]}-[\nabla_{X},\nabla_{Y}].$ We denote
$g(R(X,Y)Z,U)$ by $R(X,Y,Z,U)$. Curvature identities are of
fundamental importance for understanding the geometry of almost
Hermitian manifolds. The following curvature identities are used in
various studies e.g.(\cite{Graya,Graylv}):
\begin{enumerate}
\item $R(X,Y,Z,U)=R(X,Y,JZ,JU),$
\item $R(X,Y,Z,U)=R(JX,JY,Z,U)+R(JX,Y,JZ,U)$
$\quad\quad\quad\quad\quad\quad$$+R(JX,Y,Z,JU),$
\item $R(X,Y,Z,U)=R(JX,JY,JZ,JU)$.
\end{enumerate}
Let $AH_{i}$ denote the subclass of the class $AH$ of almost
Hermitian manifolds satisfying the curvature identity $(i).$,
$i=1,2,3.$ We know that
$$AH_{1}\subset AH_{2}\subset AH_{3}\subset AH,$$
from \cite{Graya}. Some authors call $AH_{1}$-manifold as a
\emph{para-K\"{a}hlerian manifold} and call $AH_{3}$-manifold as an
$RK$-\emph{manifold} (\cite{Van}). An almost Hermitian manifold $M$
is called \emph{K\"{a}hlerian} if $\nabla_{X}J=0$ for all
$X\in\chi(M)$ and \emph{nearly K\"{a}hlerian (almost Tachibana)} if
$(\nabla_{X}J)X=0$ for all $X\in\chi(M).$ It is well-known that a
K\"{a}hlerian manifold is $AH_{1}$-manifold and a nearly
K\"{a}hlerian manifold (non para-K\"{a}hlerian) manifold
is $AH_{2}$-manifold, see (\cite{Graylv,Van}).\\
A two-dimensional linear subspace of a tangent space $T_{p}(M)$ is
called a \emph{plane section}. A plane section $\sigma$ is said to
be \emph{holomorphic} (resp.\emph{ anti-holomorphic} or
\emph{totally real}) if $J\sigma=\sigma$ (resp. $J\sigma\bot\sigma$)
(\cite{BChe}, \cite{Yama}). The sectional curvature $K$ of $M$ which
is determined by orthonormal vector fields $X$ and $Y$ is given by
$K(X,Y)=R(X,Y,X,Y).$ The sectional curvature of $M$ restricted to a
holomorphic (resp. an anti-holomorphic) plane $\sigma$ is called
\emph{holomorphic} (resp. \emph{anti-holomorphic}) \emph{sectional
curvature}. If the holomorphic (resp. anti-holomorphic) sectional
curvature at each point $p\in M$ does not depend on $\sigma$, then
$M$ is said to be \emph{ pointwise constant holomorphic} (resp.
\emph{pointwise constant anti-holomorphic) sectional curvature}. A
connected Riemannian (resp. K\"{a}hlerian) manifold of (global)
constant sectional curvature (resp. of constant holomorphic
sectional curvature) is called a \emph{real space form} (resp. a
\emph{complex space form}) (\cite{BChe,Vanh,Kas,Yan}). The following
useful notion was defined by A. Gray in \cite{Gray}.
\begin{definition} Let $M$ be an almost Hermitian manifold. Then $M$ is said
to be of \emph{constant type} at $p\in M$ provided that for all
$X\in T_{p}(M)$, we have $\lambda(X,Y)=\lambda(X,Z)$ whenever the
planes $span\{X,Y\}$ and $span\{X,Z\}$ are anti-holomorphic and
$g(Y,Y)=g(Z,Z)$, where the function $\lambda$ is defined by
$\lambda(X,Y)=R(X,Y,X,Y)-R(X,Y,JX,JY)$. If this holds for all $p\in
M$, then we say that $M$ has \emph{(pointwise) constant type}.
Finally, if for $X,Y\in\chi(M)$ with $g(X,Y)=g(JX,Y)=0,$ the value
$\lambda(X,Y)$ is constant whenever $g(X,X)=g(Y,Y)=1$,
then we say that $M$ has \emph{global constant type.}
\end{definition}
It follows that any $AH_{1}$-manifold has global vanishing constant type from Definition 2.1.\\
Let $M$ be a $2m$-dimensional K\"{a}hlerian manifold, for all $X,Y,Z,U\in T_{p}(M)$ and $p\in M,$
the Bochner curvature tensor $B$ \cite{Kassab} is defined by\\
$B(X,Y,Z,U)=R(X,Y,Z,U)-L(Y,Z)g(X,U)+L(Y,U)g(X,Z)$
$\quad\quad\quad\quad\quad\quad\quad$$-L(X,U)g(Y,Z)+L(X,Z)g(Y,U)-L(Y,JZ)g(X,JU)$
$\quad\quad\quad\quad\quad\quad\quad+L(Y,JU)g(X,JZ)-L(X,JU)g(Y,JZ)+L(X,JZ)g(Y,JU)$
$\quad\quad\quad\quad\quad\quad\quad+2L(Z,JU)g(X,JY)+2L(X,JY)g(Z,JU),$\\
where $L=\frac{\varrho}{2(m+2)} -\frac{\tau}{8(m+1)(m+2)}g,$
$\varrho$ is the \emph{Ricci tensor} and $\tau$ is the \emph{scalar
curvature} of $M.$ It is well-known that the Bochner curvature
tensor is a complex analogue of the Weyl conformal curvature tensor
\cite{Yan} on a Riemannian manifold.
\subsection{Submanifolds of a Riemannian manifold}
Let $N$ be a submanifold of a Riemannian manifold $M$ with a
Riemannian metric $g.$ Then Gauss and Weingarten formulas are
respectively given by $\nabla_{X}Y=\hat{\nabla}_{X}Y+h(X,Y)$ and
$\nabla_{X}\xi=-A_{\xi}X+\nabla_{X}^{\bot}\xi$ for all
$X,Y\in\chi(N)$ and $\xi\in\chi^{\bot}(N)$. Here $\nabla,
\hat{\nabla},$ and $\nabla^{\bot}$ are respectively the Riemannian,
induced Riemannian, and induced normal connection in $M, N,$ and the
normal bundle $\chi^{\bot}(N)$ of $N$, and $h$ is the second
fundamental form related to shape operator $A$ corresponding to the
normal vector field $\xi$ by $g(h(X,Y),\xi)=g(A_{\xi}X,Y).$ A
submanifold $N$ is said to be \emph{totally geodesic} if its second
fundamental form identically vanishes: $h=0,$ or equivalently
$A_{\xi}=0.$ We say that $N$ is \emph{totally umbilical} submanifold
in $M$ if for all $X,Y\in\chi(N),$ we have
\begin{equation}
\label{e1}
\begin{array}{c}
h(X,Y)=g(X,Y)\eta
\end{array},
\end{equation}
where $\eta\in\chi^{\bot}(N)$ is the mean curvature vector field of
$N$ in $M$. A vector field $\xi\in\chi^{\bot}(N)$ is said to be
\emph{parallel} if $ \nabla^{\bot}_{X}\xi=0$ for each $X\in\chi(N)$.
The Codazzi equation is given by
\begin{equation}
\label{e1}
\begin{array}{c}
(R(X,Y)Z)^{\bot}=(\nabla_{X}h)(Y,Z)-(\nabla_{Y}h)(X,Z)
\end{array},
\end{equation}
for all $X,Y,Z\in\chi(N),$ where $^{\bot}$ denotes the normal
component and the covariant derivative of $h$ denoted by
$\nabla_{X}h$ is defined by
\begin{equation}
\label{e1}
\begin{array}{c}
(\nabla_{X}h)(Y,Z)=\nabla^{\bot}_{X}(h(Y,Z))-h(\hat{\nabla}_{X}Y,Z)-h(Y,\hat{\nabla}_{X}Z)
\end{array},
\end{equation}
for all $X,Y,Z\in\chi(N)$ (\cite{BChe,Gol,Ha,Yama}).
\subsection{Anti-invariant submanifolds of an almost Hermitian manifold}
Let $M$ be a $2m$-dimensional almost Hermitian manifold endowed with
an almost complex structure $J$ and a Hermitian metric $g.$ An
$n$-dimensional Riemannian manifold $N$ isometrically immersed in
$M$ is called an \emph{anti-invariant submanifold} of $M$ (or
\emph{totally real submanifold} of $M$) if $JT_{p}(N)\subset
T_{p}(N)^{\bot}$ for each point $p$ of $N$. Then we have $m\geq n$
(\cite{Yan}).
\section{The axiom of anti-invariant 2-spheres}
S. Yamaguchi and M. Kon \cite{Yama} introduced the axiom of
anti-invariant 2-spheres as: An almost Hermitian manifold $M$ is
said to satisfy the axiom of anti-invariant 2-spheres, if for each
point $p\in M$ and each anti-holomorphic $2$-plane $\sigma$ of the
tangent space $T_{p}(M)$, there exists a $2$-dimensional totally
umbilical anti-invariant submanifold $N$ such that $p\in N$ and
$T_{p}(N)=\sigma.$ They proved in Theorem 1(\cite{Yama}) that a
K\"{a}hlerian manifold with the axiom of anti-invariant 2-spheres is
a complex space form. Here, we give a generalization of Theorem
1(\cite{Yama}). From now on, we shall assume that all manifolds are
connected throughout this study.\\
We shall need the following Lemma for the proof of the main Theorem
3.1 below.
\begin{lemma}(\cite{Ho}) Let $N$ be an anti-invariant submanifold of a nearly
K\"{a}hlerian manifold $M$. Then
\begin{equation}
\label{e1}
\begin{array}{c}
A_{JZ}X=A_{JX}Z
\end{array}
\end{equation}
holds for any two vectors $X$ and $Z$ tangent to $N$, where $A$ is
the shape operator of $N$.
\end{lemma}
Let $N$ be given as in Lemma 3.1, then
\begin{equation}
\label{e1}
\begin{array}{c}
g(A_{JZ}X,Y)=g(A_{JX}Z,Y)=g(A_{JY}X,Z)
\end{array},
\end{equation}
for any vectors $X,Y$ and $Z$ tangent to $N$, from (3.1). This
equation is equivalent to
\begin{equation}
\label{e1}
\begin{array}{c}
g(h(X,Y),JZ)=g(h(Z,Y),JX)=g(h(X,Z),JY)
\end{array},
\end{equation}
where $h$ is the second fundamental form of $N$.
\begin{theorem}
Let $M$ be a nearly K\"{a}hlerian manifold of dimension $2m\geq6.$
If $M$ satisfies the axiom of anti-invariant 2-spheres, then $M$ is
a space of constant holomorphic sectional curvature.
\end{theorem}
\begin{proof}
Let $p$ be any point of $M$. Let $X$ and $Y$ be any orthonormal
vectors of $T_{p}(M)$ spanning an anti-holomorphic 2-plane $\sigma$.
By the axiom of anti-invariant 2-spheres, there exists a
two-dimensional totally umbilical anti-invariant submanifold $N$
such that $p\in N$ and $T_{p}(N)=\sigma.$ Then, we have
\begin{equation}
\label{e1}
\begin{array}{c}
(R(X,Y)Y)^{\bot}=\nabla^{\bot}_{X}\eta
\end{array},
\end{equation}
with the help of (2.1) and (2.3) from (2.2), where $\eta$ is the
mean curvature vector field of $N$ in $M$. We find
\begin{equation}
\label{e1}
\begin{array}{c}
R(X,Y,Y,JX)=g(\nabla^{\bot}_{X}\eta,JX)
\end{array},
\end{equation}
from (3.4), since $JX$ is normal to $N.$ If we put $Y=Z$ into (3.3),
we obtain
\begin{equation}
\label{e1}
\begin{array}{c}
g(\eta,JX)=0
\end{array},
\end{equation}
by using (2.1). Similarly, we also have $g(\eta,JY)=g(\eta,JZ)=0$, for any vectors $X,Y,$ and $Z$ tangent to $N.$
If we differentiate the equation (3.6) with respect to $X$, then we
have
\begin{equation}
\label{e1}
\begin{array}{c}
0=X[g(\eta,JX)]=g(\nabla^{\bot}_{X}\eta,JX)+g(\eta,\nabla^{\bot}_{X}JX)
\end{array}.
\end{equation}
Upon straightforward calculation, we see that
$g(\eta,\nabla^{\bot}_{X}JX)=0.$ Thus, we obtain
\begin{equation}
\label{e1}
\begin{array}{c}
g(\nabla^{\bot}_{X}\eta,JX)=0
\end{array},
\end{equation}
from (3.7). We get
\begin{equation}
\label{e1}
\begin{array}{c}
R(X,Y,Y,JX)=0
\end{array}
\end{equation}
by combining (3.8) with (3.5). For all orthonormal vectors $X,Y\in
T_{p}(M)$ with $g(X,JY)=0,$ we derive
\begin{equation}
\label{e1}
\begin{array}{c}
R(JY,Y,Y,JX)=0,
\end{array}
\end{equation}
by replacing $X$ by $\frac{1}{\sqrt{2}}(X+JY)$ in (3.9). In this
case, it follows that $M$ has pointwise constant holomorphic
sectional curvature, using (3.10) and Lemma 1(\cite{Kas}). We
conclude that $M$ is a space of constant holomorphic sectional
curvature by Theorem 6(\cite{He}).
\end{proof}
\section{The axiom of co-holomorphic $(2n+1)$-spheres}
Let $M$ be a $2m$-dimensional almost Hermitian manifold. L. Vanhecke
\cite{Vanh} defined a co-holomorphic $(2n+1)$-plane as a
$(2n+1)$-plane containing a holomorphic $2n$-plane for the manifold
$M$. It is not difficult to see that a co-holomorphic $(2n+1)$-plane
contains an anti-holomorphic $(n+1)$-plane and that $1\leq n\leq
m-1$. He also gave the axiom of co-holomorphic $(2n+1)$-spheres as:
A $2m$-dimensional almost Hermitian manifold $M$ which is said to
satisfy the axiom of co-holomorphic $(2n+1)$-spheres, if for each
point $p\in M$ and each co-holomorphic $(2n+1)$-plane $\sigma$ of
the tangent space $T_{p}(M)$, there exists a $(2n+1)$-dimensional
totally umbilical submanifold $N$ such that $p\in N$ and
$T_{p}(N)=\sigma.$
He studied this axiom for $AH_{3}$-manifolds and obtained several results.\\
Now, we study this axiom for larger classes of almost Hermitian
manifolds.
\begin{lemma}
Let $M$ be an almost Hermitian manifold of dimension $2m\geq4.$ If
$M$ satisfies the axiom of co-holomorphic $(2n+1)$-spheres, then we
have
\begin{equation}
\label{e1}
\begin{array}{c}
\lambda(X,Y)=K(X,Y)
\end{array},
\end{equation}
for all orthonormal vectors $X,Y\in T_{p}(M)$ with $g(X,JY)=0,$
where $\lambda(X,Y)=R(X,Y,X,Y)-R(X,Y,JX,JY)$ and $K$ denotes
anti-holomorphic sectional curvature.
\end{lemma}
\begin{proof}
Let $p$ be an arbitrary point of $M$. Let $X$ and $Y$ be any
orthonormal vectors in $T_{p}(M)$ with $g(X,JY)=0,$ that is, they
span an anti-holomorphic plane. Consider the co-holomorphic
$(2n+1)$-plane $\sigma$ containing $X,JX,$ and $Y$ such that $JY$ is
normal to $\sigma$. By the axiom of co-holomorphic $(2n+1)$-spheres,
there exists a $(2n+1)$-dimensional totally umbilical submanifold
$N$ such that $p\in N$ and $T_{p}(N)=\sigma.$ Then, we have
\begin{equation}
\label{e1}
\begin{array}{c}
(R(X,Y)JX)^{\bot}=0
\end{array}.
\end{equation}
with the help of
(2.1) and (2.3), from (2.2). We get
\begin{equation}
\label{e1}
\begin{array}{c}
R(X,Y,JX,JY)=0
\end{array},
\end{equation}
from (4.2), since $JY$ is normal to $N$. Thus, our assertion follows
from Definition 2.1 and (4.3).
\end{proof}
Now, we are ready to prove our second main result.
\begin{theorem}
Let $M$ be an almost Hermitian manifold of dimension $2m\geq6.$ If
$M$ satisfies the axiom of co-holomorphic $(2n+1)$-spheres for some
$n$, then $M$ has pointwise constant type $\alpha$ if and only if
$M$ has pointwise constant anti-holomorphic sectional curvature
$\alpha$.
\end{theorem}
\begin{proof}
Let $M$ be an almost Hermitian manifold of dimension $2m\geq6$
satisfying the axiom of co-holomorphic $(2n+1)$-spheres for some
$n$. If $M$ has pointwise constant type, that is, $M$ has constant
type at $p$, for all $p\in M$. Then, for all $X,Y,Z\in T_{p}(M)$
whenever the planes $span\{X,Y\}$ and $span\{X,Z\}$ are
anti-holomorphic and $g(Y,Y)=g(Z,Z),$ we have
\begin{equation}
\label{e1}
\begin{array}{c}
\lambda(X,Y)=\lambda(X,Z)
\end{array}.
\end{equation}
Here, we can assume that $g(Y,Y)=g(Z,Z)=1.$ Thus, for all
orthonormal vectors $X,Y,Z\in T_{p}(M)$ with $g(X,JY)=g(X,JZ)=0$, we
get
\begin{equation}
\label{e1}
\begin{array}{c}
K(X,Y)=K(X,Z)
\end{array},
\end{equation}
from Lemma 4.1.\\
On the other hand, we can choose a unit vector $U$ in
$(span\{X,JX\})^{\bot}\cap(span\{Z,JZ\})^{\bot},$ since the
dimension of $M$ is greater than 6. Then, we have
\begin{equation}
\label{e1}
\begin{array}{c}
K(X,U)=K(X,Z)
\end{array},
\end{equation}
from (4.5). This implies that the sectional curvature is same for
all anti-holomorphic sections which contain any given vector $X.$
Hence, we write
\begin{equation}
\label{e1}
\begin{array}{c}
K(X,Y)=K(Y,Z)=K(Z,U)
\end{array}.
\end{equation}
Therefore, we find
\begin{equation}
\label{e1}
\begin{array}{c}
K(X,Y)=K(Z,U)
\end{array},
\end{equation}
for all $X,Y,Z,U\in T_{p}(M)$ whenever the planes $span\{X,Y\}$ and
$span\{Z,U\}$ are anti-holomorphic. It follows that the sectional
curvature is same for all anti-holomorphic sections at $p\in M.$
Namely, $M$ has pointwise constant
anti-holomorphic sectional curvature.\\
Conversely, let $M$ be of pointwise constant anti-holomorphic
sectional curvature and let $p$ be any point of $M$. Then for all
orthonormal vectors $X,Y,Z\in T_{p}(M)$ with $g(X,JY)=g(X,JZ)=0,$
($span\{X,Y\}$ and $span\{X,Z\}$ are anti-holomorphic planes and
$g(X,X)=g(Y,Y)=g(Z,Z)=1$), we have
\begin{equation}
\label{e1}
\begin{array}{c}
K(X,Y)=K(X,Z)
\end{array}.
\end{equation}
By Lemma 4.1, we get
\begin{equation}
\label{e1}
\begin{array}{c}
\lambda(X,Y)=\lambda(X,Z)
\end{array},
\end{equation}
for all orthonormal vectors $X,Y,Z\in T_{p}(M)$ whenever the planes
$span\{X,Y\}$ and $span\{X,Z\}$ are anti-holomorphic. It is not
difficult to see that (4.10) also holds in the case
$g(Y,Y)=g(Z,Z)\neq1.$ It follows that $M$ has constant type at $p.$
Additionally, if the constant value of $\lambda(X,Y)$ equals
$\alpha$, then the pointwise constant anti-holomorphic sectional
curvature $K$ must be $\alpha$, because of Lemma 4.1.
\end{proof}
We remark that above technical method was used also in Theorem 3.4(\cite{Tas}).\\
Next, we give some applications of Theorem 4.1.
\begin{theorem}
Let $M$ be a $2m$-dimensional almost Hermitian manifold with
pointwise constant type
$\alpha$ and $m\geq3.$ If $M$ satisfies the axiom of co-holomorphic $(2n+1)$-spheres for some $n$, then\\
\textbf{i)} $M$ is a space of constant curvature $\alpha$ and $M$ has global constant type $\alpha$,\\
\textbf{ii)} $M$ is an $AH_{2}$-manifold.
\end{theorem}
\begin{proof} Let $p$ be any point of $M$. Let $X$ and $Y$ be any orthonormal
vectors in $T_{p}(M)$ with $g(X,JY)=0.$ Consider the co-holomorphic
$(2n+1)$-plane $\sigma$ containing $X,JX,$ and $JY$ such that $Y$ is
normal to $\sigma$. By the axiom of co-holomorphic $(2n+1)$-spheres,
there exists a $(2n+1)$-dimensional totally umbilical submanifold
$N$ such that $p\in N$ and $T_{p}(N)=\sigma.$ Then, we have
\begin{equation}
\label{e1}
\begin{array}{c}
(R(X,JX)JX)^{\bot}=0
\end{array}.
\end{equation}
with the help of (2.1) and (2.3), from (2.2). Hence, we get
\begin{equation}
\label{e1}
\begin{array}{c}
R(X,JX,JX,Y)=0
\end{array},
\end{equation}
for all orthonormal vectors $X,Y\in T_{p}(M)$ with $g(X,JY)=0,$
since $Y$ is normal to $N$. It follows that $M$ is an
$AH_{3}$-manifold with pointwise constant holomorphic sectional
curvature using (4.12) and Lemma 1(\cite{Kas}) together with Lemma
3(\cite{Kas}). On the other hand, we have that $M$ has constant
anti-holomorphic sectional curvature $\alpha$ at $p,$ from Theorem
4.1. In this case, we obtain that the constant holomorphic sectional
curvature $H$ of $M$ is $\alpha$ at $p$ from Theorem 5(\cite{Van}).
By using Theorem 4(\cite{Van})in Theorem 2(\cite{Van}), we obtain
\begin{equation}
\label{e1}
\begin{array}{c}
K(X,Y)=\alpha
\end{array},
\end{equation}
for all orthonormal vectors $X,Y\in T_{p}(M),$ where
$K(X,Y)=R(X,Y,X,Y)$ is sectional curvature. It is not difficult to
see that (4.13) is also true for all $X,Y\in T_{p}(M).$ By the
well-known Schur's theorem(\cite{Yan})
it follows that $M$ is a space of constant curvature $\alpha$ and $M$ has global constant type.\\
Now, we prove the part \textbf{ii)}. From the part \textbf{i)},
automatically, both holomorphic and anti-holomorphic sectional
curvature equal to $\alpha$. In which case, it follows from Theorem
3(\cite{Gan}) that $M$ is an $AH_{2}$-manifold.
\end{proof}
\begin{remark}
Theorem 4.2 without part \textbf{ii)} was also obtained by O.T.
Kassabov in (\cite{Kass}) with different approach. It is a
generalization of Theorem 1(\cite{Vanh}) concerning
$AH_{3}$-manifolds.
\end{remark}
By Theorem 4.1, Theorem 4.2 and Theorem 5(\cite{Van}), we have the
following result which is a generalization of Corollary
1(\cite{Vanh}) concerning $AH_{3}$-manifolds.
\begin{corollary} Let $M$ be a $2m$-dimensional almost Hermitian manifold with
vanishing constant type and $m\geq3.$ If $M$ satisfies the axiom of
co-holomorphic $(2n+1)$-spheres for some $n$, then $M$ is a flat
$AH_{2}$-manifold.
\end{corollary}
We end this paper by giving a result related to the Bochner
curvature tensor of a K\"{a}hlerian manifold satisfying the axiom of
co-holomorphic $(2n+1)$-spheres.
\begin{theorem}
Let $M$ be a K\"{a}hlerian manifold of dimension $2m\geq6.$ If $M$
satisfies the axiom of co-holomorphic $(2n+1)$-spheres for some $n$,
then $M$ has a vanishing Bochner curvature tensor.
\end{theorem}
\begin{proof} Let $p$ be any point of $M$. Let $X,Y,$ and $Z$ be any unit vectors
of $T_{p}(M)$, which span an anti-holomorphic 3-plane, that is,
$g(X,Y)=g(X,Z)=g(Y,Z)=0,$ and $g(X,JY)=g(X,JZ)=g(Y,JZ)=0.$ Consider
the co-holomorphic $(2n+1)$-plane $\sigma$ containing $X,JX,$ and
$Y$ such that $Z$ is normal to $\sigma$. By the axiom of
co-holomorphic $(2n+1)$-spheres, there exists a
$(2n+1)$-dimensional totally umbilical submanifold $N$ such that
$p\in N$ and $T_{p}(N)=\sigma.$ Then, we have
\begin{equation}
\label{e1}
\begin{array}{c}
(R(X,JX)Y)^{\bot}=0
\end{array},
\end{equation}
and
\begin{equation}
\label{e1}
\begin{array}{c}
(R(X,Y)JX)^{\bot}=0
\end{array}.
\end{equation}
with the help of (2.1) and
(2.3) from (2.2). For all unit vectors $X,Y,Z\in T_{p}(M)$, which
span an anti-holomorphic 3-plane, we respectively get
\begin{equation}
\label{e1}
\begin{array}{c}
R(X,JX,Y,Z)=0
\end{array}
\end{equation}
and
\begin{equation}
\label{e1}
\begin{array}{c}
R(X,Y,JX,Z)=0
\end{array},
\end{equation}
from (4.14) and (4.15), since $Z$ is normal to $N$. Thus, our
assertion follows from (4.16), (4.17) and Lemma(\cite{Kassa}).
\end{proof}
|
\section{Properties of bounded degree graphs}\label{sec:bounded}
The following three lemmas are proved by Cygan and Pilipczuk~\cite{CP2013}. We slightly extend the statements and provide the proofs for the sake of completeness.
The lemma below allows to find in a graph a set of vertices of high degree with a better upper bound on its size than given by the standard averaging argument.
\begin{lemma}[\cite{CP2013}, Lemma~3.2]\label{lemma:coresize}
For any graph $G=(V,E)$ of average degree at most $d$, any integer $m\ge 1$ and any $\alpha \ge 1$ there exists
$m\le D\le M$ such that $ |V_{>D}| \le \frac{nd}{\alpha D}$
where $M=\lfloor me^{\alpha+1}+1\rfloor$.
\end{lemma}
\begin{proof}\label{page:averageproofs}
Clearly,
\[\sum_{i=0}^{\infty}|V_{>i}|=\sum_{i=0}^{\infty}i|V_{=i}| \le nd \,.\]
If, on the other hand, $|V_{>i}|>\frac{nd}{\alpha i}$ for all $m\le i\le M$ then
\[\sum_{i=0}^{\infty}|V_{>i}| \ge \sum_{i=m}^{M}|V_{>i}|>\frac{nd}{\alpha}\sum_{i=m}^{M}\frac{1}{i}=
\frac{nd}{\alpha}\left(\sum_{i=1}^{M}\frac{1}{i}-\sum_{i=1}^{m-1}\frac{1}{i}\right) \ge
\frac{nd}{\alpha}\left(\ln M-\ln(em)\right) \ge
nd \, ,\]
where the next to last inequality uses the well-known estimate for the harmonic series:
\[ \ln(i+1) \le 1+\frac{1}{2}+\ldots +\frac{1}{i} \le \ln i + 1=\ln(ie)\,. \]
\end{proof}
\begin{remark}
Such $D$ can be easily found in polynomial time by going through all the values $D=m,\ldots, M$.
\end{remark}
The next lemma shows that one can find
a subset of vertices of linear size that is independent in the square of a graph.
\begin{lemma}[\cite{CP2013}, Lemma~3.1]\label{lemma:indsetsize}
For any graph $G=(V,E)$ of average degree at most $d$ and maximum degree at most $\Delta$
one can find in polynomial time a subset of vertices $B \subseteq V$ such that
\begin{enumerate}
\item the size of $B$ is linear: $|B| \ge \frac{n}{6\Delta d}$;
\item degrees of vertices from $B$ are small: for any $v \in B$, $\operatorname{deg}_G(v) \le 2d$;
\item $B$ is independent in $G^2$: for any $u \neq v \in B$, $N_G[u] \cap N_G[v] = \emptyset$.
\end{enumerate}
\end{lemma}
\begin{proof}
Clearly, $|V_{\le 2d}| \ge \frac{nd}{2d}=\frac{n}{2}$. The required set $B$ can be constructed by a straightforward greedy algorithm: while $V_{\le 2d}$ is not empty, take any $v \in V_{\le 2d}$, add it to $B$, and remove from $V_{\le 2d}$ the vertex $v$ together with all its neighbors in~$G^2$.
The number of such neighbors is at most $2d+2d(\Delta-1)=2d\Delta$. Hence
at each iteration at most $2d\Delta+1$ vertices are removed and the total number of iterations is at least
\[\frac{|V_{\le 2d}|}{2d\Delta+1} \ge \frac{n}{4d\Delta+2} \ge \frac{n}{6d\Delta} \,.\]
\end{proof}
The following lemma allows us to find efficiently in a graph $G$ of average degree $d=O(1)$
a subset of vertices $Y$ of high degree such that $(G\setminus Y)^2$ contains a large enough
independent set. The last inequality in the statement
can be seen as exponential savings in the running time.
\begin{lemma}[\cite{CP2013}, Lemma~3.4]\label{lemma:ugly}
For any constants $\nu \ge 1, \mu <1, a \ge 0, 0< c < 1$ there exists $\beta>0$ such that for any graph $G=(V,E)$ of average degree~$d=O(1)$ one can find in polynomial time subsets
$A,Y \subseteq V$ such that:
\begin{enumerate}
\item $A \cap Y = \emptyset$;
\item $A$ is an independent set in $(G\setminus Y)^2$;
\item $2|Y|\le |A| \le cn$;
\item each vertex from $A$ has at most $2d$ neighbors in $G\setminus Y$:
$\forall v \in A,\, |\{u \in V \setminus Y \colon (u,v) \in E\}| \le 2d\,;$
\item
\begin{equation}\label{eq:ugly1}
\binom{|A|}{|Y|}^a\nu^{|Y|}\mu^{|A|} < 2^{-\beta n}\,.
\end{equation}
\end{enumerate}
\end{lemma}
\begin{proof}
Let $\alpha=\alpha(d, \nu, \mu, a,c)$ be a large enough constant to be defined later.
Using Lemma~\ref{lemma:coresize} we can find
\begin{equation}\label{eq:cd}
\frac{1}{12dc} \le D \le \frac{1}{12dc}{e^{\alpha+1}+1},
\end{equation}
such that $|V_{>D}| < \frac{nd}{\alpha D}$. Let $Y=V_{>D}$. Note that the graph $G\setminus Y$ has average degree at most $d$ and maximum degree at most~$D$. Lemma~\ref{lemma:indsetsize} allows us to find
a subset $A \subseteq V \setminus Y$ such that $A$ is independent in $(G \setminus Y)^2$,
for all $v \in A$, $\operatorname{deg}_{G \setminus Y}(v) \le 2d$ and
\[|A| \ge \frac{n-|Y|}{6dD} \ge \frac{n}{12dD} \, ,\]
where the last inequality is true when $\alpha \ge 2d$, i.e., $\alpha$ is large enough.
Remove from $A$ arbitrary vertices such that $|A|=\frac{n}{12dD}$.
Because of~(\ref{eq:cd}), $\frac{n}{12dD} \le nc$.
To guarantee that $|A|>2|Y|$ it is enough to take $\alpha \ge 24d^2$.
We now show how to choose $\alpha$ such that the last inequality from the statement is satisfied. Using the well known estimates $\binom nk \le \left(\frac{en}{k}\right)^k$
and $\binom nk \le \binom{n}{k'}$ for $k \le k' \le \frac n2$ we get
\[ \binom{|A|}{|Y|}^a \le \left(\frac{en\alpha D}{12dDnd}\right)^\frac{nda}{\alpha D}=\left(\frac{e\alpha}{12d^2}\right)^\frac{nda}{\alpha D}=(\gamma \alpha)^\frac{nda}{\alpha D} \, ,\]
where $\gamma=\frac{e}{12d^2}$ is a constant. Thus we can upper-bound
(\ref{eq:ugly1}) as follows:
\begin{equation}\label{eq:ugly2}
\left(\left(\gamma \alpha\right)^\frac{da}{\alpha}\left(\nu^d \right)^\frac{1}{\alpha}\mu^\frac{1}{12d}\right)^\frac{n}{D} \, .
\end{equation}
Recall now that $\mu<1$ and note that $(\gamma\alpha)^\frac{da}{\alpha} \to 1$ and
$\left(\nu^d\right)^\frac{1}{\alpha} \to 1$ with $\alpha \to +\infty$. Thus for a large enough $\alpha$, (\ref{eq:ugly2}) is $\left(2^{-\beta'}\right)^\frac{n}{D}$ for a constant $\beta'>0$. It remains to recall that $D < e^\alpha$ and take $\beta=\beta'e^{-\alpha}$.
\end{proof}
\section{The chromatic number problem}
\begin{definition}
In the \emph{list coloring problem} each vertex $v$ of input graph is assigned a list $L_v$ of allowed colors and the task is to properly color a graph such that each vertex
is given a color from its list.
\end{definition}
To reduce the search space in the list coloring problem we introduce the following problem.
\begin{definition}
In the \emph{coloring with preferences problem} each vertex $v$ together with a list $L_v$ is given a \emph{preferred color} $p_v \in L_v$ and the task is to color the graph properly such that each vertex $v$ is given a color from $L_v$ and moreover at
least one vertex from $N[v]$ is given the color~$p_v$.
\end{definition}
\begin{lemma}
Let $G$ be a graph, $\{L_v\}_{v \in V}$ be a set of list colors for its vertices, and
$\{p_v \in L_v\}_{v \in V}$ be a set of preferred colors. Then
there is a solution for an instance $(G, \{L_v\})$ of the list coloring problem iff
there is a solution for an instance $(G, \{L_v\}, \{p_v\})$ of the coloring with preferences problem.
\end{lemma}
\begin{proof}
Obviously, if there is a coloring satisfying preferences then it is also a list-coloring.
For the reverse direction, consider a proper list coloring such that for some vertex
$v$ neither of the vertices from $N[v]$ is given the color $p_v$. One can then change the color of $v$ to $p_v$. This clearly does not violate any coloring constraints and moreover it strictly increases the number of vertices that are given their preferred color.
Thus, after repeating this operation at most $n$ times one arrives at a valid coloring
with preferences.
\end{proof}
Checking whether a graph $G$ has a proper $k$-coloring is a $(V,\mathcal{F},k)$-partition problem where for all $i=1,\ldots,k$, $\mathcal{F}_i=IS(G)$, the set of all independent sets of $G$. Note that Theorem~\ref{thm:simple} already implies
a $O^*(2^n)$ time and space algorithm for the Chromatic Number problem. An algorithm with the same time and space bounds was given recently by Bj\"{o}rklund et al.~\cite{BHK2009}.
For $k$-coloring with preferences the families $\mathcal{F}_i$ are defined differently. Namely, $\mathcal{F}_i$
consists of all independent sets $I$ of $G$ that can be assigned the color $i$
without violating any list constraints and preferred color constraints:
\begin{itemize}
\item (list constraints): $\forall v \in I$, $ i \in L_v$;
\item (preferred color constraints): $\forall v \in V$ such that $p_v=i$,
$N_G[v] \cap I \neq \emptyset$.
\end{itemize}
Using this interpretation of the coloring with preferences problem,
we give an algorithm solving the Chromatic Number problem
on graphs of bounded average degree in time $O^*((2-\varepsilon)^n)$ and space.
\begin{theorem}
There is an algorithm checking whether for a given graph $G$ of average degree $d=O(1)$ there exists a proper $k$-coloring in time $O^*((2-\varepsilon(d))^n)$
and exponential space.
\end{theorem}
\begin{proof}
First consider the case $k \ge 2d$. Note that $|V_{\ge k}| \le \frac{nd}{k} \le n/2$.
Then one can find a proper $k$-coloring
of a graph $G[V_{\ge k}]$ in time $O^*(2^{n/2})$. Such a coloring can be easily extended to the whole graph (since there always exists a vacant color for a vertex of degree at most $k-1$). Thus, in the following we assume that $k < 2d=O(1)$.
Let $\nu \ge 1, \mu<1,a \ge 0, 0<c<1$ be constants to be defined later and let $A,Y \subseteq V$ be as provided by Lemma~\ref{lemma:ugly}. For $Y$ we try all possible colorings in time $k^{|Y|}$. A~fixed coloring of $Y$ produces a list coloring problem for~$G \setminus Y$.
Let $L$ be one of the most frequent color lists of vertices from~$A$ and assume w.l.o.g. that $L=\{1, \dots, l\}$ (hence $l \le k$). Let
$C=\{v \in A \colon L_v = L\}$. Since there are at most $2^k$ different lists,
$|C| \ge |A|/2^k$. Remove arbitrary vertices from $C$ so that $|C|=\frac{|A|}{2^k}$.
Now, first remove a few vertices
from $C$ such that $|C|$ is divisible by
$l+1$ and partition $C$ arbitrarily into $l+1$ subsets $C_1, \dots, C_{l+1}$
of size $\frac{|C|}{l+1}$.
For all $i=1, \ldots,l$, assign the preferred color $p_v=i$ for all vertices $v$ from $C_i$
and assign preferred colors arbitrarily for all the remaining vertices ($V\setminus Y \setminus \cup_{i=1}^{l}C_i$).
We are ready to construct sets with infants for $V\setminus Y$.
For $j=1, \dots, \frac{|C|}{(l+1)}$, let
\begin{equation}\label{eq:ri}
R_j = \bigcup_{i=1}^{l+1}N_G[C_i[j]] \text{ and } r_j=C_{l+1}[j] \, .
\end{equation}
See Figure~\ref{fig:ris} for a visual explanation.
First of all note that all $R_j$'s are disjoint since they consist of
the closed neighbourhoods of vertices from $A$ and according to Lemma~\ref{lemma:ugly} such neighbourhoods are disjoint. To show that the sets $R_j$'s are indeed sets with infants,
consider a set $F \in \mathcal{F}_i$ and assume that $r_j \in F$. Recall that $\mathcal{F}_i$ consists of the sets that can be assigned the color~$i$.
Our goal is to show that $F$ also contains at least one other element of~$R_j$. Although we do not know the color of $r_j$, we do know that its color is from $L$. Note that each family must have at list one element of each color, hence $r_j$ has a relative in~$F$. Below we formally prove that the constructed system satisfies all properties of families with infants.
\begin{figure}
\begin{center}
\begin{tikzpicture}
\draw (0,0) rectangle (8,4);
\foreach \y/\i in {3.75/C_1, 3.25/C_2, 2.5/\vdots, 1.75/C_i, 1/\vdots, 0.25/C_{l+1}}
\node at (4,\y) {$\i$};
\foreach \y in {0.5, 1.5, 2, 3, 3.5}
\draw (0,\y) -- (8,\y);
\foreach \y/\i in {3.75/1, 3.25/2, 1.75/i, 0.25/l+1}
\node[anchor=east] at (-0.1,\y) {$\i$};
\foreach \x/\i in {0.25/1, 0.75/2, 6.25/j, 7.75/\frac{|C|}{l+1}}
\node[anchor=north] at (\x,-0.2) {$\i$};
\node at (0.5,4.5) {$C$};
\node[rectangle,anchor=west,text width=50mm] at (8.2,1.75) {the preferred color of all vertices from $C_i$ is $i$\\ (for all $1 \le i \le l$)};
\draw (6,-0.1) rectangle (6.5,4.1);
\node[anchor=south west,rectangle,text width=80mm] at (6.25,4.2) {$R_j = \bigcup_{i=1}^{l+1}N_G[C_i[j]] \text{ and } r_j=C_{l+1}[j]$};
\end{tikzpicture}
\end{center}
\caption{Schematic explanation for $R_i$'s.}
\label{fig:ris}
\end{figure}
The fact
that $r_j$ belongs to $F \in \mathcal{F}_i$ means that the vertex $r_j$ can be assigned the color~$i$.
In other words, $i \in L$. Recall from the definition (\ref{eq:ri}) of $R_j$ that
it contains, in particular, $N_G[C_i[j]]$. Since the preferred color of the vertex $C_i[j]$ is~$i$, at least one of $N_G[C_i[j]]$ must be given the color~$i$. Thus, besides of $r_j$, the set $F$ must contain also at least one element from $N_G[C_i[j]] \subseteq R_j$.
By choosing a small enough value for $c$ one can guarantee that $pq \le n$.
We conclude that $((R_1, r_1), \dots, (R_p,r_p))$ is
a $(p,q)$-family of sets with infants for $p=\frac{|C|}{l+1}=\frac{|A|}{2^k(l+1)}$ and $q=(2d+1)(l+1) \le (2d+1)(k+1) \le (2d+1)^2$.
Theorem~\ref{thm:enc} implies that the running time of the resulting algorithm is at most
\[k^{|Y|}\cdot \left( 2^{|V \setminus Y|} \cdot
\left(\frac{2^q-1}{2^q}\right)^p\right) \le 2^n \cdot \left( k^{|Y|} \cdot
\left(\frac{2^q-1}{2^q}\right)^p\right) \, .\]
We now choose the constants $\nu,\mu,a$ so that (\ref{eq:ugly1}) implies that the expression in parentheses is at most $2^{-\beta n}$ for a constant $\beta > 0$. Let $a=0$, $\nu=k$, and
\[\mu=\left(\frac{2^q-1}{2^q}\right)^{\frac{1}{2^k(l+1)}}\]
(recall that $k<2d=O(1)$ and $l \le k=O(1)$).
It is easy to see that (\ref{eq:ugly1}) then implies that the total running time is $2^{(1-\beta)n}$ for a constant $\beta>0$.
\end{proof}
\begin{remark}
It is not difficult to see that the presented algorithm actually solves a more general problem known as maximum $k$-cut. In this problem the goal is to partition the vertices into $k$ parts such that the number of edges joining different parts is maximal possible.
\end{remark}
\section{Introduction}
In this paper we consider algorithms for three classical hard problems: the traveling salesman problem, the chromatic number problem, and the problem of counting
perfect matchings.
$O^*(2^n)$ algorithms for the traveling salesman problem by Bellman~\cite{B1962} and Held and Karp~\cite{HK1971} are known for more than 50 years already ($n$~is the number of vertices of an input graph). The current record upper bound $O^*(2^n)$ for the chromatic number problem
is proved by Bj\"{o}rklund et al.~\cite{BHK2009} in~2006. The number of perfect matchings in an $n$-vertex graph can be computed in time $O^*(2^{n/2})$ as shown by Bj\"{o}rklund~\cite{B2012} in 2012 (this matches the bound of Ryser's algorithm~\cite{R1963} for bipartite graphs).
These problems (and many others) can be seen as partition problems:
in the chromatic number problem the goal is to partition the vertices into
independent sets; in the traveling salesman problem one needs to partition
the vertices into several Hamiltonian sets with minimal total weight; in counting perfect matchings the goal is to partition the vertices into adjacent pairs.
We show a general way to solve such partition problems in $O^*(2^n)$ time
where $n$ is the size of a set to be partitioned. The method is in some sense a rephrasing of the method from~\cite{BHK2009} where instead of the inclusion-exclusion principle we use
the fast Fourier transform (FFT). See, e.g.,~\cite{DPV2006} for an introduction to FFT. In particular, the method immediately implies an $O^*(2^n)$
upper bound for such partition problems as chromatic number, maximum $k$-cut, domatic number, bin packing.
For all three problems mentioned above (chromatic number, traveling salesman, counting perfect matchings),
improving
the known bound for the general case
is a major open problem in the field of algorithms for NP-hard problems. However in recent years it was shown that the bound can be improved for various special cases.
In~\cite{BHKK2008, BHKK2010, CP2013}
better upper bounds are proved for graphs of bounded degree.
We further extend our approach to get bounds of the form
$O^*((2-\epsilon)^n)$ in various special cases. Namely we show that
such a bound follows almost immediately if the corresponding partition problem possesses a certain structure. Informally, this structure can be
described as follows. Assume that a group of people is going to an excursion and our task is to seat them into buses with several constraints
saying that a pair of people does not want to see each other in the same bus.
This is the coloring problem and it can be solved in $O^*(2^n)$ time. Assume now that we have additional constraints: the group of people
contains several infants and these infants should be accompanied by their relatives in a bus. Roughly, we prove that if the number of infants is linear
then the problem can be solved in $O^*((2-\varepsilon)^n)$ time.
Using this approach we unify several known results of this kind. An additional advantage of the approach is that it is particularly easy to use it as a black box. Namely, all one needs to do is to reveal the corresponding structure of families with infants. This way, some of the known results follow just in a few lines. By using additional combinatorial ideas we also prove the following new results.
For the chromatic number problem, Bj\"{o}rklund et al.~\cite{BHKK2010} presented an algorithm working in time
$O^*((2-\varepsilon(\Delta))^n)$ on graphs of bounded maximum degree~$\Delta=O(1)$. The algorithm is based on Yate's
algorithm and Moebios inversion and thus uses exponential space. We extend this result to a wider class
of bounded average degree graphs. This closes an open problem concerning the existence of such an algorithm stated
by Cygan and Pilipczuk~\cite{CP2013}.
For the traveling salesman problem on graphs of maximum degree~$\Delta=O(1)$, Bj\"{o}rklund et al.~\cite{BHKK2008} presented an algorithm working in time $O^*((2-\varepsilon(\Delta))^n)$
and exponential space. Cygan and Pilipczuk~\cite{CP2013} extended the result to graphs of bounded average (instead of maximum) degree. Both algorithms are based on dynamic programming and the savings in the running time comes from an observation that in case of bounded degree graphs an algorithm does
not need to go through all possible subsets of vertices (e.g., a disconnected subset of vertices is not Hamiltonian for sure). It is also because of the dynamic programming technique that both mentioned algorithms use exponential space. We further improve these results by showing an algorithm working in time
$O^*((2-\varepsilon(d))^n)$ and polynomial space on graphs of average degree~$d$.
Cygan and Pilipczuk~\cite{CP2013} developed an algorithm with running time $O^*((2-\varepsilon(d))^{n/2})$ and exponential space for counting perfect matching in graphs of average degree~$d$. We present an algorithm solving this problem in $O^*((2-\varepsilon(d))^{n/2})$ time and polynomial space. Several bounds of this kind are already known for bipartite graphs~\cite{ALS1991,BF2002,SW2005,RBR2009,IW2012,CP2013}.
\section{Notation}
Let $G=(V,E)$ be a simple undirected graph. Throughout the paper we implicitly assume that the set of vertices of a graph under consideration is $V=\{1,2, \dots, n\}$. For simplicity, we consider undirected graphs only (whether a graph is directed or not is only important for the traveling salesman problem; the presented algorithm works for both undirected and directed graphs).
By $d(G)$ and $\Delta(G)$ we denote the average and the maximum degree of $G$ (we omit $G$ if it is clear from the context).
$N_G(v)$ is a \emph{neighborhood} of $v$ in~$G$, i.e., all the neighbors of $v$ in~$G$ and $N_G[v]=N_G(v) \cup \{v\}$ is its \emph{closed neighborhood}.
For $S \subseteq V$, by $G[S]$ we denote a subgraph of $G$ induced by $S$. We use $G\setminus S$ as a shortcut for $G[V \setminus S]$.
We often exploit the following simple fact: one can find in $G$ an independent set of size at least $\frac{n}{\Delta(G)+1}$ in polynomial time (this is done by a straightforward greedy algorithm).
The \emph{square} of $G=(V,E)$ is a graph $G^2=(V,E')$ where $E' \supseteq E$ is
\[E'=\{(u,v) \colon \text{ there is a path of length at most $2$ from $u$ to $v$ in $G$}\} \,.\] Note that $\Delta(G^2) \le (\Delta(G))^2$ and hence one can easily find an independent set of size $\frac{n}{(\Delta(G))^2+1}$
in~$G^2$.
Following~\cite{CP2013}, by
$V_{>c}$ we denote a subset of vertices $V$ of degree greater than~$c$.
$V_{<c}$, $V_{=c}$, $V_{\le c}$, $V_{\ge c}$ are defined similarly.
\section{Partition problems}\label{sec:partitions}
\begin{definition}
Let $V=\{1, \dots, n\}$, $1\le k \le n$ be an integer and $\mathcal{F}=\{\mathcal{F}_1, \ldots, \mathcal{F}_k\}$, where each $\mathcal{F}_i \subseteq 2^V$ is a family of subsets of~$V$. A~\emph{$(V,k,\mathcal{F})$-partition problem} is to represent $V$ as a disjoint union of $k$ sets from~$\mathcal{F}_i$'s:
\[V=F_1 \sqcup \ldots \sqcup F_k \text{, where } F_i \in \mathcal{F}_i, \forall 1 \le i \le k \, .\]
\end{definition}
The brute force search algorithm for this problem takes time $O^*(\max_{1 \le i \le k}|\mathcal{F}_i|^k)$.
Using FFT one can easily prove an upper bound $O^*(2^n)$ which beats the previously mentioned bound in many interesting cases.
\begin{theorem}\label{thm:simple}
A $(V,k,\mathcal{F})$-partition problem can be solved in $O^*(2^n)$ time and space.
\end{theorem}
Note that Theorem~\ref{thm:simple} immediately implies $O^*(2^n)$ upper bound for such problems as domatic number, chromatic number, maximum $k$-cut. These bounds were proved relatively recently by Bj\"{o}rklund, Husfeldt and Koivisto~\cite{BHK2009} using the inclusion-exclusion method.
Below, we formally define a combinatorial structure called families with infants that
allow to get smaller than $O^*(2^n)$ upper bounds for partition problems.
\begin{definition}
$\mathcal{R}=((R_1,r_1), \dots, (R_p,r_p))$ is a called a \emph{$(p,q)$-system of families with infants} for a $(V,k,\mathcal{F})$-partition problem if all of the following conditions are satisfied:
\begin{enumerate}
\item $pq \le n$;
\item for all $i=1,\dots,p$, $r_i \in R_i$; $r_i$ is called \emph{an infant} and all the elements of $R_i\setminus \{r_i\}$ are called \emph{relatives} of $r_i$; the sets $R_i$ are called \emph{families};
\item the size of each family $R_i$ is at most $q$;
\item all families $R_i$'s are pairwise disjoint;
\item in any valid partition each infant is accompanied by at least one of its relatives:
\begin{equation}\label{eq:mainprop}
\text{for all $1 \le i \le p, 1 \le j \le k$ and all $F \in \mathcal{F}_j$, if $r_i \in F$ then $|F \cap R_i| \ge 2$.}
\end{equation}
\end{enumerate}
\end{definition}
\begin{theorem}\label{thm:enc}
Let $\mathcal{R}=((R_1,r_1), \dots, (R_p,r_p))$ be a {$(p,q)$-system of families with infants}
for a $(V,k,\mathcal{F})$-partition problem.
Then the problem can be solved in time and space
\begin{equation}
O^*\left(2^{n} \cdot \left(\frac{2^q-1}{2^q}\right)^p \cdot 2^q\right) \, .
\end{equation}
\end{theorem}
Roughly, the savings in the running time comes from the fact that
while looking for a valid partition of $V$ one can avoid the case $F \cup R_i=\{r_i\}$
(i.e., instead of considering all $2^q$ possibilities of $F \cap R_i$ one considers
$2^q-1$ of them).
We always use Theorem~\ref{thm:enc} with $q=O(1)$ and $p=\Omega(n)$, which makes the running time less than $O^*(2^n)$.
\begin{corollary}\label{cor:enc}
Let $\mathcal{R}=((R_1,r_1), \dots, (R_p,r_p))$ be a {$(p,q)$-system of families with infants}
for a $(V,k,\mathcal{F})$-partition problem. If $q=O(1)$ and $p=\Omega(n)$,
then the problem can be solved in time and space $O^*((2-\varepsilon)^n)$.
\end{corollary}
As an illustration of Theorem~\ref{thm:enc} we replicate a result from~\cite{BHKK2010}. In the (decision version of) domatic number problem the question is to partition the set of vertices into $k$ dominating sets.
\begin{lemma}\label{lemma:domatic}
The domatic number problem in a graph of maximum degree $\Delta=O(1)$ can be solved in time and space $O^*((2-\varepsilon(\Delta))^n)$.
\end{lemma}
\begin{proof}\label{page:domaticproof}
The domatic number problem is a $(V,k,\mathcal{F})$-problem where each $\mathcal{F}_i$ is just the set of all dominating sets of $G$. By definition, for any $v \in V$ and any dominating set $U \subseteq V$, $N_G[v] \cap U \neq \emptyset$. This gives a straightforward construction of families with infants.
Find greedily an independent set $I \subseteq V$ of size $p=\frac{n}{\Delta^2+2}$ in $G^2$. Assume w.l.o.g. that $I=\{1, \dots, p\}$. For each $1 \le i \le p$,
let $R_i = N_G[i]$.
At this point we have at least $n-p(\Delta+1) \ge p$ remaining vertices in $V \setminus \cup_{i=1}^{p}R_i$. So, we can extend each $R_i$ by one vertex
and declare this one additional vertex as the infant of $R_i$.
All $R_i$ have size at most $q=\Delta+2=O(1)$, the total number of $R_i$'s is $p=\frac{n}{\Delta^2+2}=\Omega(n)$. Clearly $pq \le n$. The constructed sets satisfy the property~(\ref{eq:mainprop}) by an obvious reason: each $R_i$ is a superset of $N_G[v]$ for some $v \in V$ and none of these elements is the infant of the family~$R_i$. And $U \cap N_G[v] \neq \emptyset$ for any dominating set~$U$ and any vertex~$v$, i.e.,
any dominating set always contains at least one relative of $r_i$
(even if it does not contain $r_i$).
The upper bound now follows from Theorem~\ref{thm:enc} and Corollary~\ref{cor:enc}.
\end{proof}
Another example is a faster than $O^*(2^n)$ algorithm for finding
a Hamiltonian cycle in a graph of bounded degree. This result
was given in~\cite{BHKK2008}.
\begin{lemma}\label{lemma:tspmax}
The Hamiltonian cycle problem on a graph of maximum degree $\Delta=O(1)$ can be solved in time and space $O^*((2-\varepsilon(\Delta))^n)$.
\end{lemma}
\begin{proof}\label{page:tspmax}
Guess three vertices $v_0,v_1,v_2 \in V$.
Let $\mathcal{F}=(\mathcal{F}_0,\mathcal{F}_1,\mathcal{F}_2)$ where $\mathcal{F}_i \subseteq 2^V$ consists of all
subsets $S \subseteq V$ of size $|S|=n/3$ for which $G$
contains a path $P$ such that
\begin{enumerate}
\item $P$ starts in $v_i$ and ends in $u_i$ such that $(u_i,v_{(i+1) \bmod 3}) \in E$;
\item $P$ goes through all the vertices in $S$ exactly once.
\end{enumerate}
The family $\mathcal{F}$ can be computed in time $O^*(2^{H(1/3)n})=O^*(1.99^n)$ (where $H(x)=-x\log_2-(1-x)\log_2(1-x)$ is the binary entropy function) by dynamic programming.
We now need to construct a system of families with infants for the resulting $(V,3,\mathcal{F})$-partition problem.
We construct the required family for $p=\frac{n}{\Delta^2+1}$ and~$q=\Delta+1$.
Find greedily an independent set $I$ of size $p=\frac{n}{\Delta^2+1}$ in~$G^2$.
Assume that $I=\{1, \dots, p\}$ and let $R_i=N_G[i]$, $r_i=i$. Clearly, if $F \in \mathcal{F}_j$
contains an infant $r_i \in R_i$ then this infant is necessarily accompanied by one of its relatives since $F$ contains a Hamiltonian path.
\end{proof}
The corresponding algorithms allow also to solve weighted partition problems. In such problems, each subset $F$ of $\mathcal{F}_i$ is assigned a non-negative integer weight $w(F)$ and the goal is to find
a partition of minimum total weight.
\begin{theorem}\label{thm:weighted}
If in Theorems~\ref{thm:simple} and \ref{thm:enc} one is given a weighted partition problem then the upper bounds on the running time and space are multiplied by $W$ where $W$ is the maximum weight of a subset.
\end{theorem}
Also, one can turn the algorithm to use polynomial space by providing an algorithm that enumerates the sets $\mathcal{F}_i$.
\begin{theorem}\label{thm:encenum}
Let $\mathcal{F}=(\mathcal{F}_1, \ldots, \mathcal{F}_k)$ and assume that there exists an algorithm that for any $i=1,\ldots, k$ enumerates the set $\mathcal{F}_i$ in time $T$ and polynomial space. Then one can turn
algorithms from Theorems~\ref{thm:simple}, \ref{thm:enc}, \ref{thm:weighted} into polynomial space algorithms at the cost of multiplying the running time by~$T$.
\end{theorem}
As a corollary we get a polynomial-space algorithm for the case when each $\mathcal{F}_i$ is of polynomial size.
\begin{corollary}\label{cor:encpolyspace}
Let $\mathcal{R}$ and $(V,k,\mathcal{F})$ be as in Theorem~\ref{thm:encenum}.
If for all $i=1,\ldots,k$, $\mathcal{F}_i$ is enumerable in polynomial time (in particular, $|\mathcal{F}_i|=\poly(n)$) then the corresponding algorithm uses polynomial space.
\end{corollary}
We conclude the section by noting that the same bounds hold also
for the case when instead of partition one looks for a cover of $V$
by $k$ subsets from $\mathcal{F}_1, \ldots, \mathcal{F}_k$. We call the corresponding problem a \emph{$(V,k,\mathcal{F})$-covering problem}.
\begin{theorem}\label{thm:covering}
Theorems~\ref{thm:simple}, \ref{thm:enc}, \ref{thm:weighted}, \ref{thm:encenum} hold for $(V,k,\mathcal{F})$-covering problems.
\end{theorem}
\input bounded
\input proofs
\input coloring
\input tsp
\input matchings
\section{Further directions}
As a conclusion we would like to pose some open questions which arise from the suggested algorithms. The first question is to obtain an efficient polynomial-space algorithm for the chromatic number problem. For example, $O^*(2^n)$-time algorithm for coloring in general graphs or $O^*((2-\varepsilon)^n)$-algorithm for coloring in bounded maximum/average degree graphs. A long-standing open question is to solve the traveling salesman problem in time $O^*((2-\varepsilon)^n)$. Any evidence of hardness of TSP would be of great interest as well.
\bibliographystyle{plain}
\section{Counting perfect matchings}
The algorithm for counting perfect matchings uses many ideas from the algorithm for the traveling salesman problem presented in Theorem~\ref{thm:tsp}.
\begin{lemma}\label{lm:antimatching}
Let $G$ be a graph of average degree $d=O(1)$. Then $\bar{G}$ (the complement of~$G$) contains a matching consisting of $\frac{n}{2}-O(1)$
edges.
\end{lemma}
\begin{proof}
Clearly $|V_{\ge \frac{n}{3}}| \le \frac{nd}{n/3}=3d$. After removing all these vertices from $G$ we get a graph with at least $n-3d$ vertices such that the degree of each vertex is at most $\frac{n}{3}$. This implies that the degree of each vertex in the complement of this graph is at least $n-3d-1-\frac{n}{3}$. This is at least $\frac{n}{2}$ for large enough $n$. By Dirac's theorem~\cite{D1952} this graph is Hamiltonian and hence contains a perfect matching.
\end{proof}
\begin{theorem}
The number of perfect matchings in a graph $G$ with $2n$ vertices of average degree $d=O(1)$ can be found in time $O^*((2-\varepsilon(d))^n)$ and polynomial space.
\end{theorem}
\begin{proof}
Assume that the vertices of $V=\{1, \ldots,2n\}$ are numbered in such a way
that
\begin{equation}\label{eq:loops}
(1,n+1),(2,n+2), \ldots, (m,n+m) \not \in E
\end{equation}
where $m=n-O(1)$. Such a numbering exists due to Lemma~\ref{lm:antimatching} (and can be found efficiently since we can find a maximum matching in $\bar{G}$ in polynomial time). Following~\cite{B2012} and \cite{CP2013} we reduce the problem of counting perfect matchings to a problem of counting cycle covers of a special type. We construct an auxiliary multigraph $G'=(V',E')$ where $V'=\{1,\ldots,n\}$
and each edge $(i,j) \in E$ is transformed into an edge
$e=((i \mod n)+1, (j \mod n)+1) \in E'$
with the label $l(e)=\{i,j\}$. In other words, we contract each pair of vertices
$(1, n+1), \ldots, (n,2n)$ and on each edge we keep a label showing where it originates from.
Any two vertices in $G'$ are joined by at most $4$ edges.
The average degree of $G'$ is at most~$2d$.
Recall that a cycle cover of a multigraph is a collection of cycles such that each vertex belongs to exactly one cycle. In other words, this is a subset of edges such that each vertex is adjacent to exactly two of these edges (and a self-loop is thought to be adjacent to its vertex two times).
The important property of the graph $G'$ is the following:
each perfect matching in $G$ corresponds to a cycle cover $C \subseteq E'$ in $G'$
such that $\cup_{e \in C}l(e)=V$ and vice versa. Indeed, each vertex $i$ in $G'$ is adjacent to exactly two edges. These two edges have different labels so they correspond to edges in the original graph $G$ that match both $i$ and $i+n$.
We have reduced the problem to counting cycle covers with disjoint labels in~$G'$.
We further reduce the problem to counting cycle covers without self-loops.
Note that by (\ref{eq:loops}) $G'$ has at most $s=O(1)$ self-loops. For each such loop $e=(i,i)$ we can consider two cases: to count the number of cycle covers with $e$ we
count the number of cycle covers in $G'$ without the vertex~$i$; to count the number
of cycle covers without $e$ we can just remove the loop $e$ from $G'$ and count the number
of cycle covers. This way, we reduce the problem to $2^s=O(1)$ problems of counting cycle
covers in a multigraph without self-loops.
Note that we only need to ensure that the labels on adjacent edges are disjoint.
Namely, one of the edges adjacent to a vertex $i$ in $G'$ must contain $i$ in its label while the other one must contain $i+n$.
Recall that Remark~\ref{rem:counting} claims that we can use Theorem~\ref{thm:enc} to find the number of valid partitions. We consider two important special cases.
\begin{enumerate}
\item
{\em Counting the number of covers by $k$ cycles of length at most $\sqrt{n}$.}
The good thing with covers by short cycles is that all such cycles can be easily enumerated in subexponential time and polynomial space. Indeed, to enumerate all possible short cycles one first goes through all possible length values $l$ (at most $\sqrt n$ choices), then through all subsets of vertices of a cycle (at most $\binom nl \le n^{\sqrt{n}}=2^{o(n)}$ choices), then through all orders of the vertices (at most $l! \le \sqrt{n}! \le \sqrt{n}^{\sqrt{n}}=2^{o(n)}$ choices), and for each
pair of consecutive vertices go through all possible
edges that join them (at most $4^{\sqrt{n}}=2^{o(n)}$ choices). To guarantee that each cycle is counted only once instead of considering all $l!$ orderings we consider $\frac{l!}{2l}$
of them.
We then proceed as in Theorem~\ref{thm:tsp}. Namely, let $A,Y \subseteq V'$ be provided by Lemma~\ref{lemma:ugly}. We then go through all possible sets $Y' \subseteq A$ such that $|Y'| \le |Y|$. The set $Y'$ is thought as predecessors of the vertices from $Y$ in cycle covers that lie in $A$ (since we consider cycle covers without self-loops each vertex has a predecessor). Then, each vertex from $A'=A \setminus Y'$ has a predecessor in $V' \setminus Y'$ and we can find a system of families with infants in exactly the same way as in Theorem~\ref{thm:tsp}. Note that for each $Y'$ we enumerate only short cycles where $Y'$ is indeed the set of predecessors of~$Y$.
Thus, the number of covers by cycles with disjoint labels of length at most $\sqrt{n}$ can be counted in time $O^*((2-\varepsilon(d))^n)$ and polynomial space.
Note that in fact we count ordered cycle covers
(where the order of cycles does matter), so we should also divide the resulting number
by~$k!$.
\item
{\em Counting the number of covers by $k$ cycles of length greater than $\sqrt{n}$.}
Long cycles cannot be quickly enumerated, but we know for sure that if all the cycles
have length at least $\sqrt{n}$ then the number $k$ of cycles is at most $\sqrt{n}$.
We go through all possible cases of lengths of $k$ cycles
(at most $n^{\sqrt n}=2^{o(n)}$ choices). Then we proceed as in Theorem~\ref{thm:tsp}.
Namely,
for each cycle of length $l$
we go through all possible tuples of $t=\frac{l}{\sqrt{n}}$ vertices $v_1, \ldots, v_t$ (call them pivot vertices)
such that:
\begin{itemize}
\item $v_1$ is the minimal vertex of the considered cycle (this ensures that the corresponding cycle is counted only once);
\item the length of a subpath of the considered cycle between $v_i$ and $v_{i+1}$
is $\sqrt{n}$ for all $i=1, \ldots, t-1$
\end{itemize}
(at most
$n^{2\sqrt{n}}=2^{o(n)}$ choices since for long cycles there are at most
$2\sqrt{n}$ such vertices in total).
For each vertex from the tuples we also go through all possible labels of two edges adjacent to it (at most $(2n)^{2\sqrt{n}}$ choices since each vertex of $G'$ has at most $2n$ adjacent edges).
The subpaths of length at most $\sqrt{n}$ can be enumerated in subexponential time and polynomial space. Then we can solve the corresponding partition problem
in time $O^*((2-\varepsilon(d))^n)$ and polynomial space as shown in Theorem~\ref{thm:tsp}.
As in the previous case, we divide the resulting number by $k!2^k$ since we counted ordered cycle covers and each cycle was counted twice (in clockwise and anticlockwise directions).
\end{enumerate}
We proceed to the general case. To count the total number of cycle covers we first go through
all $O(n^2)$ pairs $(k_1,k_2)$ where $k_1,k_2$ is the number of short and long cycles, respectively. We then go through all possible values of lengths of all long cycles
and all tuples of pivot vertices. Given the lengths of long cycles, by $k'_2$ we denote the total number of subpaths of length $\sqrt{n}$ which are needed to cover the long cycles. When these objects are fixed we take subsets $A,Y \subseteq V'$ as given by Lemma~\ref{lemma:ugly}. We then go through all possible $Y' \subseteq A$ such that $|Y'| \le |Y|$ and $Y'$ are thought as predecessors of $Y$ in $A$.
We define the corresponding $(V,k_1+k'_2,\mathcal F)$-partition problem as follows. For $1\le i\le k_1$, ${\mathcal F}_i$ contains all cycles of length $\le\sqrt{n}$ where adjacent edges have disjoint labels. For $k_1+1\le i\le k_1+k'_2$, ${\mathcal F}_i$ is the set of subpaths of length $\sqrt{n}$ (note that some of ${\mathcal F}_i$'s correspond
to the last part of a long cycle and hence consist of subpaths of length
smaller than~$\sqrt{n}$) where adjacent edges have disjoint labels. In order to get the number of unordered cycle covers, we need to divide the resulting value by $k_1!k_2!$.
Let then $A'$ be $A \setminus Y'$ with all pivot vertices removed. Using $A'$ we can find a system of families with infants exactly as it is done in Theorem~\ref{thm:tsp}.
\end{proof}
\section{Solving Partition Problems}
For a subset $U \subseteq V$, let $b(U) \in \{0,1\}^n$ denote the characteristic
vector of the set $U$ (i.e., $b(U)[i]=1$ iff $i \in U$).
In the analysis below we sometimes identify a bit vector $b(U)$ with a non-negative integer between $0$ and $2^n-1$ that it represents.
For a bit vector $b$, we denote the Hamming weight of $b$ by $w(b)$, i.e., the number of $1$'s in $b$. Note the following simple fact: for any two non-negative integers
$a$ and $b$,
\begin{equation}\label{eq:carries}
w(\operatorname{bin}(a))+w(\operatorname{bin}(b)) \ge w(\operatorname{bin}(a+b))
\end{equation}
and the equality holds iff there are no carries in $a+b$.
\subsection{Proof of Theorem~\ref{thm:simple}}
\input proofofthm1
\subsection{Proof of Theorem~\ref{thm:enc}}
\begin{definition}
For a matrix $M=(M[i,j])_{0\le i \le p-1, 0 \le j \le q-1} \in \mathbb{Z}_{\ge 0}^{p \times q}$ let
\begin{eqnarray*}
\operatorname{colweight}(M,j)&=&\sum_{i=0}^{p-1}M[i,j]\,,\quad
\operatorname{weight}(M)=\sum_{i=0}^{p-1}\sum_{j=0}^{q-1}M[i,j]=\sum_{j=0}^{q-1}\operatorname{colweight}(M,j)\,,\\
\operatorname{rowcode}(M,i)&=& -M[i,0]+\sum_{j=1}^{q-1}2^j\cdot M[i,j]\,,\quad
\operatorname{rowsum}(M)=\sum_{i=0}^{p-1}\operatorname{rowcode}(M,i)\,,\\
\operatorname{code}(M)&=&\sum_{i=0}^{p-1}(2^q - 1)^i\cdot \operatorname{rowcode}(M,i)\,.
\end{eqnarray*}
\end{definition}
\begin{definition}
A matrix $M \in \mathbb{Z}_{\ge 0}^{p \times q}$ is called \emph{row-normalized} if $\operatorname{rowcode}(M,i) \ge 0$ for all $0 \le i \le p-1$.
\end{definition}
\begin{remark}
In the analysis below we will need the following simple estimates. Let $E \in \{0,1\}^{p \times q}$. Then
\begin{enumerate}
\item $\operatorname{rowcode}(E,i)<0$ iff $E[i]=[1,0,0,\ldots,0],$
\item $\operatorname{rowcode}(E,i) \le 2^q-2 $.
\item $\operatorname{code}(E) \le (2^q-2)\cdot \sum_{i=0}^{p-1}(2^q-1)^i<(2^q-1)^p$ (assuming $q \ge 2$).
\end{enumerate}
\end{remark}
\begin{lemma}
\label{lemma:system}
The expansion of $X\in\mathbb{Z}_{\ge 0}$ in powers of $b>1$ as $X=\sum_{i=0}^{\infty}x_i\cdot b^i, x_i\ge0$ has the minimal value of the sum of digits $\sum_{i=0}^{\infty}x_i$ iff $\forall i: 0\le x_i<b$ (i.e., $X$ is written in the numeral system of base $b$).
\end{lemma}
\begin{proof}
If $x_i\ge b$, then we can reduce $x_i$ by $b$ and increase $x_{i+1}$ by $1$, decreasing the sum of digits by $b-1>0$. The other direction follows from the uniqueness of representation in base~$b$.
\end{proof}
\begin{lemma}
\label{lemma:coding}
Let $q\ge 2$ and $E \in \{0,1\}^{p \times q}$
and $M \in \mathbb{Z}_{\ge 0}^{p \times q}$ be row-normalized matrices.
If
$\operatorname{colweight}(M,0)=\operatorname{colweight}(E,0)$,
$\operatorname{weight}(M)=\operatorname{weight}(E)$,
$\operatorname{rowsum}(M)=\operatorname{rowsum}(E)$,
$\operatorname{code}(M)=\operatorname{code}(E)$,
then $M=E$.
\end{lemma}
\begin{proof}
The claim follows from Lemma~\ref{lemma:system}.
Since for all $i$, $\operatorname{rowcode}(E,i) \le 2^q-2 $, $\operatorname{code}(E)$ has the minimal sum of digits in base $(2^q-1)$ system. This in turn implies that for each $i$, $\operatorname{rowcode} (E,i)=\operatorname{rowcode}(M,i)$. Then the first columns of matrices $E$ and $M$ are equal modulo $2$, because parities of rowcodes depend only on the first column. Since $\operatorname{colweight}(M,0)=\operatorname{colweight}(E,0)$ we conclude that the first columns of $M$ and $E$ are equal.
Now each $\operatorname{rowcode}(E,i)$ has the minimal sum of digits in the system of base $2$, which means that $\operatorname{weight}(E)$ has the minimal possible value for these $\operatorname{rowcodes}$. It follows from Lemma~\ref{lemma:system} that each $M[i,j]$ must be equal to $E[i,j]$.
\end{proof}
\begin{definition}
An injective function $\alpha \colon U \to \{0,\dots,p-1\} \times \{0,\dots, q-1\}$ is called a \emph{matrix representation} of a set $U$. For such $\alpha$ and $S \subseteq U$, a~\emph{characteristic matrix} $M_\alpha(S) \in \{0,1\}^{p \times q}$ is defined as follows: $i \in U$ iff $M_\alpha(S)[\alpha(i)]=1$.
\end{definition}
\begin{proof}[Proof of Theorem~\ref{thm:enc}]
Let $\mathcal{R}=((R_1,r_1), \dots, (R_p,r_p))$ be a {$(p,q)$-system of families with infants}
for a $(V,k,\mathcal{F})$-partition problem. Append arbitrary elements from $V$ to families so that the size of each family equals $q$ and the families are still disjoint (this is possible since $pq \le n$). Denote the union of families by $R$ and the rest of $V$ by~$L$.
For each family~$R_i$, fix an order of its elements such that
the $0$th element is~$r_i$. Now consider a matrix representation $\alpha \colon V \to \{0,\dots,p-1\} \times \{0,\dots, q-1\}$ defined as follows. If $v$ is the $j$th element of $R_i$, then $\alpha(v)=(i,j)$.
We encode each set $F\in\mathcal{F}_i$ by parts. We encode vertices from $F\cap L$ using the standard technique from Theorem~\ref{thm:simple}. To encode vertices from $F\cap R$ we use the characteristic matrix $M_\alpha(F\cap R)$. Note that $M_\alpha(F\cap R)$ is a {row-normalized matrix}, because if $F$ contains an infant $r_i$ of a family~$R_i$, then it must contain at least one other element from the same row.
Consider the following polynomials for $1 \le i \le k$:
\[P_i(x,y,z,s,t,u)=\sum_{F \in \mathcal{F}_i}x^{|F\cap L|}\cdot y^{b(F\cap L)} \cdot z^{\operatorname{colweight}(M,0)} \cdot s^{\operatorname{weight}(M)} \cdot t^{\operatorname{rowsum}(M)} \cdot u^{\operatorname{code}(M)}\, ,\]
where $M=M_\alpha(F\cap R)$ and $b(F\cap L)$ is an integer from $0$ to $2^{|L|}-1$.
There is a solution to a $(V,\mathcal{F},k)$-partition problem iff
the monomial \[x^{|L|}y^{b(L)}z^{\operatorname{colweight}(R,0)} s^{\operatorname{weight}(R)} t^{\operatorname{rowsum}(R)} u^{\operatorname{code}(R)}\] has a non-zero coefficient in $\prod_{i=1}^{k}P_i(x,y,z,s,t,u)$. Indeed, as it was shown in Theorem~\ref{thm:simple}, $x^{|L|}y^{b(L)}$ corresponds to partitions of $L$. Lemma~\ref{lemma:coding} implies that only partitions of $R$ may have the term $z^{\operatorname{colweight}(R,0)} s^{\operatorname{weight}(R)} t^{\operatorname{rowsum}(R)} u^{\operatorname{code}(R)}$. Note that the degrees of $x,z,s$ in $\prod_{i=1}^{k}P_i(x,y,z,s,t,u)$ are bounded from above by $n^2$, the degree of $y$~--- by $n\cdot2^{|L|}$, the degree of $t$~--- by $n^2\cdot2^q$, the degree of $u$~--- by $n\cdot(2^q-1)^p$. Now we can apply Kronecker substitution: $y=x^{(n+1)^2}, z=x^{(n+1)^32^{|L|}}, s=x^{(n+1)^52^{|L|}}, t=x^{(n+1)^72^{|L|}}, u=x^{(n+1)^92^{|L|}2^q}$. The running time of FFT is bounded by the degree of the resulting univariate polynomial, i.e.
\[O^*(\poly(n)2^{|L|}2^q(2^q-1)^p)=O^*(2^{n-pq}(2^q-1)^p2^q)=O^*\left(2^n\cdot\left(\frac{2^q-1}{2^q}\right)^p \cdot 2^q\right)\,.\]
\end{proof}
\subsection{Proofs of Theorems~\ref{thm:weighted}, \ref{thm:encenum}, \ref{thm:covering}}
\begin{proof}[Proof of Theorem~\ref{thm:weighted}]
Following the proofs of Theorem~\ref{thm:simple} and \ref{thm:enc}, we introduce polynomials corresponding to $\mathcal{F}_i$'s. But in the weighted partition problem we multiply each monomial by $z^w$, where $z$ is a new variable and $w$ is the weight of the corresponding set. For example, in Theorem~\ref{thm:simple} the new polynomials would look as follows:
\[P_i(x,y,z)=\sum_{F \in \mathcal{F}_i}x^{|F|}y^{b(F)}z^{w(F)} \, ,\]
where $w(F)$ is the weight of $F$. Now it is clear that there exists a partition of total weight $w$ iff
the monomial $x^ny^{b(V)}z^w$ has a non-zero coefficient in $\prod_{i=1}^{k}P_i(x,y,z)$. We apply Kronecker substitution $y=x^{n^2+1}, z=x^{(n^2+1)n2^n}$ and use FFT to find all the coefficients of the product. Now we just need to find the smallest $w$, s.t. the coefficient of $x^{n+(n^2+1)\cdot b(V)+(n^2+1)n2^nw}$ does not equal $0$. Since the degree of the polynomial is at most $O^*(2^nW)$, the running time of the algorithm is $O^*(2^nW)$.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{thm:encenum}]
Assume that we need to find the coefficient of the monomial $x^m$ in $P=\prod_{i=1}^{k}P_i(x)$, where $\deg(\Pi)\le d$.
The theorem assumption claims that we can evaluate each $P_i$ at any point in time $O^*(T)$. Note that in order to get one specific coefficient of $P(x)$, we just need to list all the Fourier coefficients of $P$. Indeed, the coefficient of $x^m$ in $P$ equals
$$\frac{1}{d}\sum_{i=0}^{d-1}\omega_d^{-im}P(\omega_d^k)=\frac{1}{d}\sum_{i=0}^{d-1}\omega_d^{-im}\prod_{i=1}^{k}P_i(\omega_d^k).$$
Since each $P_i(x)$ can be evaluated in time $O^*(T)$, we need $O^*(d\cdot T)$ steps and only polynomial space to find one coefficient of the product.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{thm:covering}]
The proofs are very similar to the proofs for partition problems, the only difference is that now we consider polynomials
\[P'_i(x,y)=\sum_{F \in \mathcal{F}_i} \prod_{v\in F}({1+x\cdot y^{b(\{v\})}}) \, .\]
Let $\mathcal{F}'_i=\bigcup_{F\in\mathcal{F}_i}{2^F}$. Clearly, a $(V,\mathcal{F},k)$-covering problem has a solution iff a $(V,\mathcal{F}',k)$-partition problem has one. The polynomial $P'_i(x,y)$ corresponds to the polynomial $P_i(x,y)$ for $\mathcal{F}'$ from the proof of Theorem~\ref{thm:enc}. Namely, $x^my^{b(U)}$ has a non-zero coefficient in $P'(x,y)$ iff $U\subseteq\mathcal{F}_i$ and $|U|=m$. Again, the degree of $x$ in $\prod_{i=1}^{k}P'_i(x,y)$ is less than $n^2+1$, so we can apply Kronecker substitution $y=x^{n^2+1}$. Thus, there exists a valid cover iff the coefficient of the monomial $x^{n+(n^2+1)\cdot b(v)}$ does not equal $0$. From now on we can follow the proofs of Theorems~\ref{thm:simple}, \ref{thm:enc}, \ref{thm:weighted}, \ref{thm:encenum}.
\end{proof}
\section{The traveling salesman problem}
\begin{theorem}\label{thm:tsp}
The traveling salesman problem on graphs of average degree $d=O(1)$ can be solved
in time $O^*(W \cdot (2-\varepsilon(d))^n)$ and polynomial space $O(n^2\log W)$ ($W$ being the maximal edge weight).
\end{theorem}
\begin{proof}
Let $k$ be a parameter to be defined later. Guess $k$ vertices $v_1, \ldots, v_{k} \in V$ that split an optimal Hamiltonian cycle into $t$ paths of length $n/k$. All such $k$-tuples
can be enumerated in time~$n^k$.
Then the corresponding weighted $(V,k,\mathcal{F})$-partition problem is defined as follows.
$\mathcal{F}=(\mathcal{F}_1, \ldots, \mathcal{F}_k)$ where $\mathcal{F}_i$ consists
of all subsets $S \subseteq V$ of size $|S|=\frac{n}{k}$ for which $G$ contains a Hamiltonian path $P$ such that
$P$ starts in $v_i$, goes through all vertices from $S$, and ends in $u_i$ such that $(u_i,v_{(i \bmod k) +1}) \in E$. The weight $w(S)$ of such a set $S$ is equal to the minimal possible weight of a path $P$ (including the weight of the edge $(u_i,v_{(i \bmod k) +1})$). It is not difficult to see that solving the traveling salesman problem is equivalent to solving the $(V,k,\mathcal{F})$-partition problem if the vertices $v_1,\dots,v_k$ are guessed correctly.
Note that any $\mathcal{F}_i$ can be enumerated in time
$\binom{n}{\frac{n}{k}} \cdot \left(\frac{n}{k}\right)! $
(the first term is for guessing the subset of vertices $S$, the second one
is for guessing the order of these vertices in an optimal path~$P$).
Recall also that guessing the vertices $v_1, \ldots, v_k$ requires $O(n^k)$ time. By choosing $k=\sqrt{n}$ we turn both these estimates into subexponential $2^{o(n)}$.
Now we turn to constructing the corresponding system of families with infants. Let $\nu \ge 1, \mu<1,a \ge 0, 0<c<1$ be constants to be defined later. Let $A,Y \subseteq V$ be provided by Lemma~\ref{lemma:ugly}.
Consider an optimal Hamiltonian cycle $C$ in the graph. Let $Y' \subseteq V$ be the successors of the vertices from $Y$ in the cycle $C$ and let $A'=A \setminus Y'\setminus\{v_1,\ldots,v_k\}$.
Note that $|A'| \ge \frac{|A|}{2}-k$ (since $|A| \ge 2|Y|$) and for each vertex $v \in A'$ its predecessor $u$ in the cycle $C$ belongs to $V \setminus Y$.
Let $A'=\{1, \ldots, p\}$. Then for all $i=1,\ldots, p$, $R_i=N_{G\setminus Y}[i]$ and $r_i=i$. Clearly, $|R_i| \le q=2d+1$. By choosing $c<\frac{1}{2d+1}$ we can guarantee that $pq \le n$. All $R_i$'s are disjoint since $A'$ is an independent set in $(G \setminus Y)^2$. Finally, if the set $Y'$ is guessed correctly (i.e., $Y'$ are indeed successors of $Y$ in the optimal cycle~$C$) then $((R_1,r_1), \dots, (R_p,r_p))$ is a $(p,q)$-system of families with infants. Indeed, if $r_i \in F$ for some $F \in \mathcal{F}_j$, then $r_i$'s predecessor in $C$ must lie in $V\setminus Y$, i.e., in $R_i$
(and $r_i$ must have a predecessor since $r_i \neq v_1, \ldots, v_k$).
By Theorems~\ref{thm:weighted} and \ref{thm:encenum} the total running time does not exceed
\[2^{o(n)} \cdot \binom{|A|}{|Y|} \cdot 2^n \cdot \left(\frac{2^q-1}{2^q}\right)^p\cdot W.\]
Recall that $p=|A'|-k \ge |A|/2-|A|/6=|A|/3$ for large enough~$n$ (since $|A|=cn$ and $k=\sqrt{n}$). Choose $a=1$, $\nu=1$, $\mu=\left( \frac{2^{2d+1}-1}{2^{2d+1}}\right)^{1/3}$. Then (\ref{eq:ugly1}) implies that
\[\binom{|A|}{|Y|}\cdot \left(\frac{2^q-1}{2^q}\right)^p < 2^{-\beta n}\]
for a constant $\beta > 0$. Thus the total running time is $O^*(W\cdot (2-\varepsilon)^n)$.
\end{proof}
|
\section{Introduction and statement of results}
A bi-harmonic operator with first order perturbation is a differential operator of the form
$$\mathcal{L}_{A,q}(x,D):=\Delta^{2}+A(x)\cdot D+q(x)$$
with $D=\frac{1}{i}\nabla$. Here $A$ is a complex-valued vector field called the magnetic potential, $q$ is a complex-valued function called the electric potential. This type of operators arise in physics when considering the equilibrium configuration of an elastic plate hinged along the boundary. It is also widely used in other physical models, see \cite{GGS}. In this paper we study the identifiability of the first order perturbation of a bi-harmonic operator from partial boundary measurements in two types of open subsets of $\mathbb{R}^{n}$, the first type is an infinite slab, and the second type is a bounded domain with $C^{\infty}$ boundary.
First we consider an infinite slab $\Sigma$. The geometry of an infinite slab arises in many applications, for instance, in the study of wave propagation in marine acoustics. It is also a simple geometric setting in medical imaging. By choosing appropriate coordinates, we may assume that
$$\Sigma:=\{x=(x',x_{n})\in\mathbb{R}^{n}:x'=(x_{1},\dots,x_{n-1})\in\mathbb{R}^{n-1}, 0<x_{n}<L\},\quad L>0.$$
Its boundary consists of two parallel hyperplanes
$$\Gamma_{1}:=\{x\in\mathbb{R}^{n}:x_{n}=L\} \quad\quad \Gamma_{2}:=\{x\in\mathbb{R}^{n}:x_{n}=0\}.$$
Given $(f_{1},f_{2})\in H^{\frac{7}{2}}(\Gamma_{1})\times H^{\frac{3}{2}}(\Gamma_{1})$ with $f_{1}, f_{2}$ compactly supported on $\Gamma_{1}$, we are interested in the following Dirichlet problem
\begin{equation}\label{Dirichlet1}
\left\{
\begin{array}{rll}\vspace{1ex}
\mathcal{L}_{A,q}u= & 0 &\quad\textrm{ in } \Sigma \\ \vspace{1ex}
u=f_{1} \quad \Delta u= & f_{2} & \quad\textrm{ on } \Gamma_{1} \\ \vspace{1ex}
u=0 \quad \Delta u= & 0 & \quad\textrm{ on } \Gamma_{2}. \\
\end{array}
\right.
\end{equation}
In Appendix A we show that problem \eqref{Dirichlet1} has a unique solution in $H^{4}(\Sigma)$, where $H^{4}(\Sigma)$ is the standard Sobolev space on $\Sigma$. We define the Dirichlet-to-Neumann map for the above boundary value problem by
$$
\begin{array}{rcl}\vspace{1ex}
\Lambda_{A,q}: \quad (H^{\frac{7}{2}}(\Gamma_{1})\cap\mathcal{E}'(\Gamma_{1}))\times (H^{\frac{3}{2}}(\Gamma_{1})\cap\mathcal{E}'(\Gamma_{1})) & \rightarrow & H^{\frac{5}{2}}_{loc}(\partial\Sigma)\times H^{\frac{1}{2}}_{loc}(\partial\Sigma)\\ \vspace{1ex}
(f_{1},f_{2}) & \mapsto & (\partial_{\nu}u|_{\partial\Sigma},\partial_{\nu}(\Delta u)|_{\partial\Sigma}),
\end{array}
$$
where $u$ is the solution of \eqref{Dirichlet1}, $\mathcal{E}'(\Gamma_{1})$ is the set of compactly supported distributions on $\Gamma_{1}$, $\nu$ is the unit outer normal vector field to $\partial\Sigma=\Gamma_{1}\cup\Gamma_{2}$. The inverse problem we will study is as follows. Let $\gamma_{1}\subset\Gamma_{1}$,$\gamma_{2}\subset\partial\Sigma$ be non-empty open subsets of the boundary, assuming that
$$\Lambda_{A^{(1)},q^{(1)}}(f_{1},f_{2})|_{\gamma_{2}\times\gamma_{2}}=\Lambda_{A^{(2)},q^{(2)}}(f_{1},f_{2})|_{\gamma_{2}\times\gamma_{2}}$$
for all $(f_{1},f_{2})\in (H^{\frac{7}{2}}(\Gamma_{1})\cap\mathcal{E}'(\Gamma_{1}))\times (H^{\frac{3}{2}}(\Gamma_{1})\cap\mathcal{E}'(\Gamma_{1}))$ with $supp(f_{1})\subset\gamma_{1}$, $supp(f_{2})\subset\gamma_{1}$, can we conclude that $A^{(1)}=A^{(2)}$ and $q^{(1)}=q^{(2)}$ in $\Sigma$?
We will show this is valid for some open subsets $\gamma_{1},\gamma_{2}$ assuming that $A^{(j)},q^{(j)},j=1,2$ are compactly supported in $\bar{\Sigma}$. Our first result considers the case when the data and the measurements are on different boundary hyperplanes.
\begin{thm}
Let $\Sigma\subset\mathbb{R}^{n}(n\geq 3)$ be an infinite slab with boundary hyperplanes $\Gamma_{1}$ and $\Gamma_{2}$. Let $A^{(j)}\in W^{1,\infty}({\Sigma};\mathbb{C}^{n})\cap\mathcal{E}'(\bar{\Sigma};\mathbb{C}^{n})$, $q^{(j)}\in L^{\infty}(\Sigma;\mathbb{C})\cap\mathcal{E}'(\bar{\Sigma};\mathbb{C}), j=1,2.$ Denote by $B\subset\mathbb{R}^{n}$ an open ball containing the supports of $A^{(j)},q^{(j)},j=1,2$. Let $\gamma_{j}\subset\Gamma_{j}$ be open sets such that
$\Gamma_{j}\cap\bar{B}\subset\gamma_{j},\;j=1,2.$ If
$$\Lambda_{A^{(1)},q^{(1)}}(f_{1},f_{2})|_{\gamma_{2}\times\gamma_{2}}=\Lambda_{A^{(2)},q^{(2)}}(f_{1},f_{2})|_{\gamma_{2}\times\gamma_{2}}$$
for all $(f_{1},f_{2})\in (H^{\frac{7}{2}}(\Gamma_{1})\cap\mathcal{E}'(\Gamma_{1}))\times (H^{\frac{3}{2}}(\Gamma_{1})\cap\mathcal{E}'(\Gamma_{1}))$ with $supp(f_{1})\subset\gamma_{1}$ and $supp(f_{2})\subset\gamma_{1}$,
then $A^{(1)}=A^{(2)}$ and $q^{(1)}=q^{(2)}$.
\end{thm}
We would like to remark that when the supports of $A^{(j)},q^{(j)}$ are strictly contained in the interior of the slab, then $\gamma_{1}$ and $\gamma_{2}$ in the above theorem can be chosen to be arbitrarily small.
Our next result considers the case when the data and the measurements are on the same boundary hyperplane.
\begin{thm}
Let $\Sigma\subset\mathbb{R}^{n}(n\geq 3)$ be an infinite slab between two parallel hyperplanes $\Gamma_{1}$ and $\Gamma_{2}$. Let $A^{(j)}\in W^{1,\infty}({\Sigma};\mathbb{C}^{n})\cap\mathcal{E}'(\bar{\Sigma};\mathbb{C}^{n})$, $q^{(j)}\in L^{\infty}(\Sigma;\mathbb{C})\cap\mathcal{E}'(\bar{\Sigma};\mathbb{C}), j=1,2.$ Denote by $B\subset\mathbb{R}^{n}$ an open ball containing the supports of $A^{(j)},q^{(j)},j=1,2$. Let $\gamma_{1},\gamma'_{1}\subset\Gamma_{1}$ be open sets such that $\Gamma_{1}\cap\bar{B}\subset\gamma_{1}$ amd $\Gamma_{1}\cap\bar{B}\subset\gamma'_{1}.$ If
$$\Lambda_{A^{(1)},q^{(1)}}(f_{1},f_{2})|_{\gamma'_{1}\times\gamma'_{1}}=\Lambda_{A^{(2)},q^{(2)}}(f_{1},f_{2})|_{\gamma'_{1}\times\gamma'_{1}}$$
for all $(f_{1},f_{2})\in (H^{\frac{7}{2}}(\Gamma_{1})\cap\mathcal{E}'(\Gamma_{1}))\times (H^{\frac{3}{2}}(\Gamma_{1})\cap\mathcal{E}'(\Gamma_{1}))$ with $supp(f_{1})\subset\gamma_{1}$ and $supp(f_{2})\subset\gamma_{1}$,
then $A^{(1)}=A^{(2)}$ and $q^{(1)}=q^{(2)}$.
\end{thm}
Proofs of Theorem 1.1 and Theorem 1.2 are based on the construction of a special class of complex geometric optics (CGO) solutions which vanish on appropriate boundary hyperplanes, using a reflection argument. The idea of constructing such solutions for the Schr\"{o}dinger operator goes back to \cite{SU}. Constructing complex geometric optics solutions using a reflection argument was initiated in \cite{I}.
Inverse problems of identifying an embedded object in a slab have been studied by many authors in \cite{Ik,KLU,LU,SW}. In \cite{LU} the authors considered the Schr\"{o}dinger operator $\Delta+q$ in a slab and showed that the electric potential $q$ can be uniquely determined from partial boundary measurements. In \cite{KLU} the authors considered the magnetic Schr\"{o}dinger operator $\Delta+A(x)\cdot D+q$ and showed that the magnetic field $dA$ and the electric potential $q$ can be uniquely determined from partial boundary measurements. Here $dA$ is the exterior differentiation of the magnetic potential vector field $A$, and notice that determining $dA$ is equivalent to determining the equivalence class $\{\tilde{A}:\tilde{A}=A+\nabla\Phi \textrm{ for some }\Phi\in C^{1,1}(\overline{\Sigma})\}$. It was also pointed out in \cite{KLU} that, by only looking at the Dirichlet-to-Neumann map, such a gauge transformation obstruction always exists, so the best one can hope for the magnetic Schr\"{o}dinger operator is to determine $dA$. However, for the perturbed bi-harmonic operator, our results indicate that this type of obstruction can be overcome and one therefore determines not only $dA$, but also $A$ itself. This is due to the fact that in our proof we are able to construct more CGO solutions than for the magnetic Schr\"{o}dinger operator thanks to the higher order of the bi-harmonic operator.
\vspace{1ex}
In the remaining part of this section we shall discuss an inverse boundary value problem for the perturbed bi-harmonic operator on a bounded domain. Let $\Omega\subset\mathbb{R}^{n}(n\geq 3)$ be a bounded open subset with $C^{\infty}$ boundary. Consider the Dirichlet problem
\begin{equation}\label{Dirichlet2}
\left\{
\begin{array}{rll}\vspace{1ex}
\mathcal{L}_{A,q}u= & 0 &\quad\textrm{ in } \Omega \\ \vspace{1ex}
u= & f_{1} & \quad\textrm{ on } \partial\Omega \\ \vspace{1ex}
\Delta u= & f_{2} & \quad\textrm{ on } \partial\Omega \\
\end{array}
\right.
\end{equation}
with $A\in W^{1,\infty}(\Omega;\mathbb{C}^{n}),q\in L^{\infty}(\Omega;\mathbb{C})$ and $(f_{1},f_{2})\in H^{\frac{7}{2}}(\partial\Omega)\times H^{\frac{3}{2}}(\partial\Omega)$. The operator $\mathcal{L}_{A,q}$, equipped with the domain
$$\mathcal{D}(\mathcal{L}_{A,q}):=\{u\in H^{4}(\Omega):u|_{\partial\Omega}=(\Delta u)|_{\partial\Omega}=0\}$$
is an unbounded closed operator on $L^{2}(\Omega)$ with purely discrete spectrum, see \cite{G}. We make the following assumption
\begin{center}
\textbf{(A)}: $0$ is not an eigenvalue of the perturbed bi-harmonic operator $\mathcal{L}_{A,q}:\mathcal{D}(\mathcal{L}_{A,q})\rightarrow L^{2}(\Omega)$.
\end{center}
Under the assumption \textbf{(A)}, the Dirichlet problem \eqref{Dirichlet2} has a unique solution $u\in H^{4}(\Omega)$, Let $\nu$ be the unit outer normal vector field to $\partial\Omega$, we define the Dirichlet-to-Neumann map to \eqref{Dirichlet2} as
$$
\begin{array}{rl}\vspace{1ex}
\Lambda_{A,q}:H^{\frac{7}{2}}(\partial\Omega)\times H^{\frac{3}{2}}(\partial\Omega) & \rightarrow H^{\frac{5}{2}}(\partial\Omega)\times H^{\frac{1}{2}}(\partial\Omega)\\ \vspace{1ex}
(f_{1},f_{2}) & \mapsto (\partial_{\nu}u|_{\partial\Omega},\partial_{\nu}(\Delta u)|_{\partial\Omega})
\end{array}
$$
where $u\in H^{4}(\Omega)$ is the solution to the problem \eqref{Dirichlet2}. We can also introduce the Cauchy data set $\mathcal{C}_{A,q}$ for the operator $\mathcal{L}_{A,q}$ defined by
$$\mathcal{C}_{A,q}:=\{(u|_{\partial\Omega},(\Delta u)|_{\partial\Omega},\partial_{\nu}u|_{\partial\Omega},\partial_{\nu}(\Delta u)|_{\partial\Omega}):u\in H^{4}(\Omega), \mathcal{L}_{A,q}u=0 \textrm{ in } \Omega\}.$$
When the assumption \textbf{(A)} holds, the Cauchy data set $\mathcal{C}_{A,q}$ is the graph of the Dirichlet-to-Neumann map $\Lambda_{A,q}$.
Let $\gamma_{1},\gamma_{2}\subset\partial\Omega$ be non-empty open subsets of the boundary. In this paper we are interested in the inverse boundary value problem for the operator $\mathcal{L}_{A,q}$ with partial boundary measurements: assuming that
$$\Lambda_{A^{(1)},q^{(1)}}(f_{1},f_{2})|_{\gamma_{2}\times\gamma_{2}}=\Lambda_{A^{(2)},q^{(2)}}(f_{1},f_{2})|_{\gamma_{2}\times\gamma_{2}}$$
for all $(f_{1},f_{2})\in H^{\frac{7}{2}}(\partial\Omega)\times H^{\frac{3}{2}}(\partial\Omega)$ with $supp(f_{1})\subset\gamma_{1}$ and $supp(f_{2})\subset\gamma_{1}$, can we conclude that $A^{(1)}=A^{(2)}$ and $q^{(1)}=q^{(2)}$ in $\Omega$?
For the bi-harmonic operator, determination of the first order perturbation on a bounded domain $\Omega$ was considered in \cite{KLU3} with partial boundary measurements. The authors showed that, from the Dirichlet-to-Neumann map, one can uniquely determine not only the electric potential $q$, but also the magnetic potential $A$. Again this is different from the situation for the magnetic Schr\"{o}dinger operator where the gauge transformation exists as an obstruction for the recovery of $A$. In this paper, we will improve the uniqueness result in \cite{KLU3} under two different assumptions: in Theorem 1.3, we assume $A^{(1)}=A^{(2)}$ and $q^{(1)}=q^{(2)}$ in a neighborhood of $\partial\Omega$, and show that we still have the uniqueness even when both $\gamma_{1}$ and $\gamma_{2}$ are arbitrarily small; in Theorem 1.4, we assume the inaccessible part of the boundary is contained in a plane, and prove the uniqueness with local data.
\begin{thm}
Let $\Omega\subset\mathbb{R}^{n}(n\geq 3)$ be a bounded domain with $C^{\infty}$ connected boundary. Let $A^{(j)}\in W^{1,\infty}({\Omega};\mathbb{C}^{n})$, $q^{(j)}\in L^{\infty}(\Omega;\mathbb{C}), j=1,2$ be such that the assumption \textbf{(A)} holds for both operators. Assume that $A^{(1)}=A^{(2)}$ and $q^{(1)}=q^{(2)}$ in a neighborhood of the boundary $\partial\Omega$. Let $\gamma_{1},\gamma_{2}\subset\partial\Omega$ be non-empty open subsets of the boundary.
If
$$\Lambda_{A^{(1)},q^{(1)}}(f_{1},f_{2})|_{\gamma_{2}\times\gamma_{2}}=\Lambda_{A^{(2)},q^{(2)}}(f_{1},f_{2})|_{\gamma_{2}\times\gamma_{2}}$$
for all $(f_{1},f_{2})\in (H^{\frac{7}{2}}(\Gamma_{1})\cap\mathcal{E}'(\Gamma_{1}))\times (H^{\frac{3}{2}}(\Gamma_{1})\cap\mathcal{E}'(\Gamma_{1}))$ with $supp(f_{1})\subset\gamma_{1}$ and $supp(f_{2})\subset\gamma_{1}$, then $A^{(1)}=A^{(2)}$ and $q^{(1)}=q^{(2)}$ in $\Omega$.
\end{thm}
In the following theorem, notice that we need the magnetic potential $A$ and electric potential $q$ to be smooth. This is due to the fact that our proof relies on determination of the boundary value of $A$ from the Cauchy data set $\mathcal{C}_{A,q}$. For the bi-harmonic operator, or more generally for poly-harmonic operators, this result was proved only for smooth $A$ and $q$ in \cite{KLU2}.
\begin{thm}
Let $\Omega\subset\{x\in\mathbb{R}^{n}:x_{n}>0\}(n\geq 3)$ be a bounded domain with $C^{\infty}$ connected boundary, and let $\partial\Omega\cap\{x\in\mathbb{R}^{n}:x_{n}=0\}\neq\emptyset$ and $\gamma:=\partial\Omega\backslash\{x\in\mathbb{R}^{n}:x_{n}=0\}$. Let $A^{(j)}\in C^{\infty}(\overline{\Omega};\mathbb{C}^{n})$, $q^{(j)}\in C^{\infty}(\overline{\Omega};\mathbb{C}), j=1,2$ be such that the assumption \textbf{(A)} holds for both operators.
If
$$\Lambda_{A^{(1)},q^{(1)}}(f_{1},f_{2})|_{\gamma\times\gamma}=\Lambda_{A^{(2)},q^{(2)}}(f_{1},f_{2})|_{\gamma\times\gamma}$$
for all $(f_{1},f_{2})\in H^{\frac{7}{2}}(\partial\Omega)\times H^{\frac{3}{2}}(\partial\Omega)$ with $supp(f_{1})\subset\bar{\gamma}$ and $supp(f_{2})\subset\bar{\gamma}$,
then $A^{(1)}=A^{(2)}$ and $q^{(1)}=q^{(2)}$ in $\Omega$.
\end{thm}
This paper is structured as follows: in Section 2 we establish a Carleman type estimate for the bi-harmonic operator and then construct a class of CGO solutions on a bounded domain; in Section 3 we show an integral identity and a Runge type approximation theorem; in Section 4 we construct the CGO solutions we desire in the slab by reflecting the CGO solutions constructed in Section 2; Section 5, 6, 7, 8 are devoted to the proof of Theorem 1.1, 1.2, 1,3 and 1.4 respectively. In Appendix A we prove the solvability of the boundary value problem \eqref{Dirichlet1} and some identities used in the proofs of the main theorems.
\section{Carleman estimate and CGO solutions on a bounded domain}
In this section we construct some CGO solutions on a bounded domain to the equation $\mathcal{L}_{A,q}u=0$. CGO solutions have been intensively utilized in establishing uniqueness result in elliptic inverse boundary value problems. For the construction of various CGO solutions and their application, see \cite{BRUZ,CY,CY2,FKSU,KSU,SU}.
Let $\Omega\subset\mathbb{R}^{n}$, $n\geq 3$, be a bounded domain with $C^{\infty}$ boundary. Consider the equation $\mathcal{L}_{A,q}u=0$ in $\Omega$ with $A\in W^{1,\infty}(\Omega;\mathbb{C}^{n})$ and $q\in L^{\infty}(\Omega;\mathbb{C})$. We will construct CGO solutions of the form
\begin{equation}\label{form}
u(x,\zeta,h)=e^{x\cdot\zeta/h}(a(x,\zeta)+r(x,\zeta,h)).
\end{equation}
based on a Carleman estimate. Here $\zeta\in\mathbb{C}^{n}$ is a complex vector satisfying $\zeta\cdot\zeta=0$, $a$ is a smooth amplitude, $r$ is a correction term, $h>0$ is a small semiclassical parameter. To deal with the perturbation, we extend $A\in W^{1,\infty}(\Omega;\mathbb{C}^{n})$ to a Lipschitz vector field compactly supported in $\mathbb{R}^{n}$, extend $q\in L^{\infty}(\Omega;\mathbb{C})$ as zero to $\mathbb{R}^{n}$. We shall work with $\zeta$ depending slightly on $h$, i.e. $\zeta=\zeta^{(0)}+\zeta^{(1)}$ with $\zeta^{(0)}$ independent of $h$, $\zeta^{(1)}=\mathcal{O}(h)$, and $|\textrm{Re } \zeta^{(0)}|=|\textrm{Im } \zeta^{(0)}|=1$.
Consider the conjugated operator
$$
h^{4}e^{-x\cdot\zeta/h}\mathcal{L}_{A,q}e^{x\cdot\zeta/h}=(h^{2}\Delta+2ih\zeta\cdot\nabla)^{2}+h^{3}A\cdot hD-ih^{3}A\cdot \zeta+h^{4}q
$$
In order to eliminate the lowest order term involving $h$ in this expression, we require
\begin{equation}\label{eq_a}
(\zeta^{(0)}\cdot\nabla)^{2}a=0 \quad\quad \textrm{ in } \Omega.
\end{equation}
As $|\textrm{Re } \zeta^{(0)}|=|\textrm{Im } \zeta^{(0)}|=1$, $\zeta^{(0)}\cdot\nabla$ is a $\bar{\partial}$-operator in appropriate coordinates, so the above equation admits a solution $a=a(x,\zeta^{(0)})\in C^{\infty}(\overline{\Omega})$. To find an appropriate correction term, we need a Carleman type estimate. We will use the semiclassical Sobolev spaces $H^{s}_{\textrm{scl}}(\mathbb{R}^{n})$ ($s\in\mathbb{R}$) with the norm $\|f\|_{H^{s}_{\textrm{scl}}(\mathbb{R}^{n})}=\|\langle hD\rangle^{s}f\|_{L^{2}(\mathbb{R}^{n})}$ where $\langle\xi\rangle=(1+|\xi|^{2})^{\frac{1}{2}}$.
\begin{prop}\label{Carleman}
Suppose $A\in W^{1,\infty}(\Omega;\mathbb{C}^{n})$, $q\in L^{\infty}(\Omega;\mathbb{C})$. Then for $h>0$ sufficiently small, there exists a constant $C>0$ independent of $h$ such that
$$\|u\|_{L^{2}(\mathbb{R}^{n})}\leq Ch^{2}\|e^{x\cdot\zeta/h}\mathcal{L}_{A,q}e^{-x\cdot\zeta/h}u\|_{H^{-1}_{\textrm{scl}}(\mathbb{R}^{n})} \quad\quad u\in C^{\infty}_{c}(\Omega).$$
\end{prop}
\begin{proof}
From \cite[Proposition 4.2]{KS}, we can find a constant $C_{1}>0$ independent of $h$ such that for all $u\in C^{\infty}_{c}(\Omega)$
$$
\|u\|_{L^{2}(\mathbb{R}^{n})}\leq C_{1}h\|e^{x\cdot\zeta/h}\Delta e^{-x\cdot\zeta/h}u\|_{H^{-1}_{\textrm{scl}}(\mathbb{R}^{n})}
$$
Iterate to get
\begin{equation}\label{CE}
\|u\|_{L^{2}(\mathbb{R}^{n})}\leq C^{2}_{1}h^{2}\|e^{x\cdot\zeta/h}\Delta^{2} e^{-x\cdot\zeta/h}u\|_{H^{-1}_{\textrm{scl}}(\mathbb{R}^{n})}
\end{equation}
We can add the zeroth order term $h^{2}q$ to \eqref{CE} since
$$
h^{2}\|qu\|_{H^{-1}_{\textrm{scl}}(\mathbb{R}^{n})}\leq h^{2}\|qu\|_{L^{2}(\mathbb{R}^{n})}\leq h^{2}\|q\|_{L^{\infty}(\mathbb{R}^{n})}\|u\|_{L^{2}(\mathbb{R}^{n})}.
$$
We can add the first order term $h^{2}e^{x\cdot\zeta/h}A\cdot De^{-x\cdot\zeta/h}$ as
$$
h^{2}e^{x\cdot\zeta/h}A\cdot De^{-x\cdot\zeta/h}=h(A\cdot hD+iA\cdot\zeta)
$$
and
$$h\|A\cdot\zeta u\|_{H^{-1}_{scl}(\mathbb{R}^{n})}\leq h\|A\cdot\zeta u\|_{L^{2}(\mathbb{R}^{2})}\leq h\|A\cdot\zeta\|_{L^{\infty}(\mathbb{R}^{n})}\|u\|_{L^{2}(\mathbb{R}^{n})}.$$
$$
\begin{array}{rl}\vspace{1ex}
h\|A\cdot hDu\|_{H^{-1}_{\textrm{scl}}(\mathbb{R}^{n})}\leq & h\displaystyle\sum^{n}_{j=1}\|hD_{j}(A_{j}u)\|_{H^{-1}_{\textrm{scl}}(\mathbb{R}^{n})}+\mathcal{O}(h^{2})\|(\textrm{div} A)u\|_{H^{-1}_{\textrm{scl}}(\mathbb{R}^{n})}\\ \vspace{1ex}
\leq & \mathcal{O}(h)\displaystyle\sum^{n}_{j=1}\|A_{j}u\|_{L^{2}(\mathbb{R}^{n})}+\mathcal{O}(h^{2})\|u\|_{L^{2}(\mathbb{R}^{2})} \\ \vspace{1ex}
\leq & \mathcal{O}(h)\|u\|_{L^{2}(\mathbb{R}^{n})}.
\end{array}
$$
After adding these perturbation terms, we get the desired result.
\end{proof}
Denote $\|f\|^{2}_{H^{1}_{scl}(\Omega)}:=\|f\|^{2}_{L^{2}(\Omega)}+h^{2}\|\nabla f\|^{2}_{L^{2}(\Omega)}$. The following solvability result is an immediate consequence of the above Carleman estimate and the Hahn-Banach Theorem.
\begin{prop}\label{solvability}
Suppose $A\in W^{1,\infty}(\Omega;\mathbb{C}^{n})$, $q\in L^{\infty}(\Omega;\mathbb{C})$. Then for any $f\in L^{2}(\Omega)$, the equation
$$e^{-x\cdot\zeta/h}\mathcal{L}_{A,q}e^{x\cdot\zeta/h}r=f \quad\quad \textrm{ in }\Omega$$
has a solution $r\in H^{1}(\Omega)$ with $\|r\|_{H^{1}_{\textrm{scl}}(\Omega)}\leq\mathcal{O}(h^{2})\|f\|_{L^{2}(\Omega)}$.
\end{prop}
\begin{proof}
We extend $A$ to a compactly supported Lipschitz vector field in $\mathbb{R}^{n}$, extend $q$ and $f$ as zero, and solve the equation in $\mathbb{R}^{n}$. Denote $\mathcal{L}_{\zeta}:=e^{-x\cdot\zeta/h}\mathcal{L}_{A,q}e^{x\cdot\zeta/h}$, the $L^{2}$-adjoint of $\mathcal{L}_{\zeta}$ is given by
$$\mathcal{L}^{\ast}_{\zeta}:=e^{x\cdot\zeta/h}\mathcal{L}^{\ast}_{A,q}e^{-x\cdot\zeta/h}=e^{x\cdot\zeta/h}
\mathcal{L}_{\bar{A},i^{-1}\nabla\cdot\bar{A}+\bar{q}}e^{-x\cdot\zeta/h}.$$
Consider the complex linear functional
$$L:\mathcal{L}^{\ast}_{\zeta}C^{\infty}_{c}(\Omega)\rightarrow\mathbb{C} \quad\quad \mathcal{L}^{\ast}_{\zeta}u\mapsto (u,f)_{L^{2}(\mathbb{R}^{n})}.$$
Applying Proposition \ref{Carleman} with $\mathcal{L}_{A,q}$ replaced by $\mathcal{L}^{\ast}_{A,q}$, we see the map $L$ is well-defined and for any $u\in C^{\infty}_{c}(\Omega)$, we have
$$
|L(\mathcal{L}^{\ast}_{\zeta}u)|=|(u,f)_{L^{2}(\mathbb{R}^{n})}|\leq \|u\|_{L^{2}(\mathbb{R}^{n})}\|f\|_{L^{2}(\mathbb{R}^{n})}\leq Ch^{2}\|\mathcal{L}^{\ast}_{\zeta}u\|_{H^{-1}_{\textrm{scl}}(\mathbb{R}^{n})}\|f\|_{L^{2}(\mathbb{R}^{n})}.
$$
This shows that $L$ is bounded in the $H^{-1}(\mathbb{R}^{n})$-norm. By the Hahn-Banach theorem we can extend $L$ to a bounded linear functional $\tilde{L}$ on $H^{-1}(\mathbb{R}^{n})$ without increasing the norm. Thus, by Riesz representation theorem, there exists $r\in H^{1}(\mathbb{R}^{n})$ such that for all $u\in H^{-1}(\mathbb{R}^{n})$ we have
$$\tilde{L}(u)=(u,r)_{H^{-1}(\mathbb{R}^{n}),H^{1}(\mathbb{R}^{n})} \quad \textrm{ and } \quad \|r\|_{H^{1}_{\textrm{scl}}(\mathbb{R}^{n})}\leq Ch^{2}\|f\|_{L^{2}(\mathbb{R}^{n})}.$$
Here $(u,r)_{H^{-1}(\mathbb{R}^{n}),H^{1}(\mathbb{R}^{n})}$ stands for the $L^{2}$-duality. It follows that $\mathcal{L}_{\zeta}r=f$ in $\mathbb{R}^{n}$, hence also in $\Omega$, and $\|r\|_{H^{1}_{scl}(\Omega)}\leq
\|r\|_{H^{1}_{scl}(\mathbb{R}^{n})}\leq Ch^{2}\|f\|_{L^{2}(\mathbb{R}^{n})}=Ch^{2}\|f\|_{L^{2}(\Omega)}$.
\end{proof}
Now we can complete the construction of the CGO solution in \eqref{form}. Equation \eqref{eq_a} gives $e^{-x\cdot\zeta/h}\mathcal{L}_{A,q}e^{x\cdot\zeta/h}a=\mathcal{O}(h^{-1})$. From Proposition \ref{solvability}, we can find $r\in H^{1}(\Omega)$ with $\|r\|_{H^{1}_{\textrm{scl}}(\Omega)}=\mathcal{O}(h)$ such that
$$
e^{-x\cdot\zeta/h}\mathcal{L}_{A,q}e^{x\cdot\zeta/h}r=-e^{-x\cdot\zeta/h}\mathcal{L}_{A,q}e^{x\cdot\zeta/h}a.
$$
Summing up, we have proved
\begin{prop}\label{existence}
Let $A\in W^{1,\infty}(\Omega;\mathbb{C}^{n})$, $q\in L^{\infty}(\Omega;\mathbb{C})$, and $\zeta\in\mathbb{C}^{n}$ be such that $\zeta\cdot\zeta=0$. Then for all $h>0$ small enough, there exist solutions $u\in H^{1}(\Omega)$ to the equation $\mathcal{L}_{A,q}u=0$ in $\Omega$ of the form
$$u(x,\zeta,h)=e^{x\cdot\zeta/h}(a(x,\zeta^{(0)})+r(x,\zeta,h)),$$
where $a\in C^{\infty}(\overline{\Omega})$ satisfies \eqref{eq_a} and $\|r\|_{H^{1}_{\textrm{scl}}(\Omega)}=\mathcal{O}(h)$.
\end{prop}
\noindent\textbf{Remark:} Sometimes we may need complex geometric optics solutions belonging to $H^{4}(\Omega)$, we can obtain such solutions as follows. Let $\Omega'\supset\supset\Omega$ be a bounded domain with smooth boundary. Extend $A\in W^{1,\infty}(\Omega;\mathbb{C}^{n})$ and $q\in L^{\infty}(\Omega;\mathbb{C})$ to functions in $W^{1,\infty}(\Omega';\mathbb{C}^{n})$ and $L^{\infty}(\Omega';\mathbb{C})$, respectively. By elliptic regularity, the complex geometric optics solutions constructed as above in $\Omega'$ will belong to $H^{4}(\Omega)$.
\section{Integral identity and Runge approximation}
For the bi-harmonic operator, Green's formula gives
\begin{equation} \label{green}
\begin{array}{ll} \vspace{1ex}
&\displaystyle\int_{\Omega}(\mathcal{L}_{A,q}u)\bar{v}\,dx-\int_{\Omega}u\overline{\mathcal{L}^{\ast}_{A,q}v}\,dx=-i\int_{\partial\Omega}\nu(x)\cdot Au\bar{v}\,dS-\int_{\partial\Omega}\partial_{\nu}(-\Delta u)\bar{v}\,dS \\
&+\displaystyle\int_{\partial\Omega}(-\Delta u)\overline{\partial_{\nu}v}\,dS-\int_{\partial\Omega}\partial_{\nu}u\overline{(-\Delta v)}\,dS+\int_{\partial\Omega}u\overline{(\partial_{\nu}(-\Delta v))}\,dS\\
\end{array}
\end{equation}
for all $u,v\in H^{4}(\Omega)$. Here $\mathcal{L}^{\ast}_{A,q}:=\mathcal{L}_{\bar{A},i^{-1}\nabla\cdot\bar{A}+\bar{q}}$ is the adjoint of $\mathcal{L}_{A,q}$, $\nu$ is the unit outer normal vector to the boundary $\partial\Omega$, and $dS$ is the surface measure on $\partial\Omega$.
For $(f_{1},f_{2})\in (H^{\frac{7}{2}}(\Gamma_{1})\cap\mathcal{E}'(\Gamma_{1}))\times (H^{\frac{3}{2}}(\Gamma_{1})\cap\mathcal{E}'(\Gamma_{1}))$ with supp$(f_{1})\subset\gamma_{1}$ and supp$(f_{2})\subset\gamma_{1}$, let $u_{1}\in H^{4}(\Sigma)$ solve
$$\left\{
\begin{array}{rcll}
\mathcal{L}_{A^{(1)},q^{(1)}}u_{1} &=& 0 & \quad\textrm{ in } \Sigma \\
u_{1}=f_{1} \quad\quad \Delta u_{1}&=& f_{2} & \quad\textrm{ on } \Gamma_{1} \\
u_{1}=0 \quad\quad \Delta u_{1}&=& 0 & \quad\textrm{ on } \Gamma_{2} \\
\end{array}
\right.$$
Let $u_{2}\in H^{4}(\Sigma)$ solve
$$\left\{
\begin{array}{rcll}
\mathcal{L}_{A^{(2)},q^{(2)}}u_{2} &=& 0 & \quad\textrm{ in } \Sigma \\
u_{2}=u_{1} \quad\quad \Delta u_{2}&=& \Delta u_{1} & \quad\textrm{ on } \Gamma_{1}\cup\Gamma_{2} \\
\end{array}
\right.$$
Let $w:=u_{2}-u_{1}$, then
\begin{equation} \label{difference}
\mathcal{L}_{A^{(2)},q^{(2)}}w=(A^{(1)}-A^{(2)})\cdot Du_{1}+(q^{(1)}-q^{(2)})u_{1}.
\end{equation}
Suppose $\Lambda_{A^{(1)},q^{(1)}}(f_{1},f_{2})|_{\gamma_{2}}=\Lambda_{A^{(2)},q^{(2)}}(f_{1},f_{2})|_{\gamma_{2}}$, then
\begin{equation}\label{normal}
\partial_{\nu}u_{1}|_{\gamma_{2}}=\partial_{\nu}u_{2}|_{\gamma_{2}}, \quad\quad \partial_{\nu}(\Delta u_{1})|_{\gamma_{2}}=\partial_{\nu}(\Delta u_{2})|_{\gamma_{2}},
\end{equation}
from which we conclude $\partial_{\nu}w=\partial_{\nu}(\Delta w)=0$ on $\gamma_{2}$. We denote
$$l_{1}:=\Gamma_{1}\cap\overline{B}\subset\gamma_{1}, \quad l_{2}:=\Gamma_{2}\cap\overline{B}\subset\gamma_{2}, \quad l_{3}:=\Sigma\cap\partial B.$$
Apparently $\partial(\Sigma\cap B)=l_{1}\cup l_{2}\cup l_{3}$. It follows from \eqref{difference} that $w\in H^{4}(\Sigma)$ is a solution to
$$\Delta^{2}w=0 \quad \textrm{ in } \quad \Sigma\backslash\overline{B}.$$
As $w=\partial_{\nu}w=0$ on $\gamma_{2}\backslash\overline{l}_{2}$, by unique continuation, $w=0$ in $\Sigma\backslash\overline{B}$. Therefore $w=\Delta w=\partial_{\nu}w=\partial_{\nu}(\Delta w)=0$ on $l_{3}$. We record these results here:
\begin{equation} \label{results}
\begin{array}{cl}
w=0 & \textrm{ on } l_{1}\\
w=\partial_{\nu}w=\partial_{\nu}(\Delta w)=0 & \textrm{ on } l_{2} \\
w=\partial_{\nu}w=\Delta w=\partial_{\nu}(\Delta w)=0 & \textrm{ on } l_{3} \\
\end{array}
\end{equation}
If $v$ is a solution of the equation
\begin{equation}\label{eqn_v}
\mathcal{L}^{\ast}_{A^{(2)},q^{(2)}}v=0 \quad\textrm{ in } \Sigma\cap B
\end{equation}
such that
\begin{equation}\label{condition_v}
v=\Delta v=0 \quad \textrm{ on } l_{1}.
\end{equation}
Taking into consideration of $\eqref{difference}$ and $\eqref{eqn_v}$, we apply \eqref{green} to $w$ and $v$ over $\Sigma\cap B$ to get
\begin{equation} \label{identity1}
\begin{array}{ll} \vspace{1ex}
&\displaystyle\int_{\Sigma\cap B}((A^{(1)}-A^{(2)})\cdot Du_{1})\bar{v}\,dx+\int_{\Sigma\cap B}(q^{(1)}-q^{(2)})u_{1}\bar{v}\,dx \\ \vspace{1ex}
=& -i\displaystyle\int_{\partial(\Sigma\cap B)}\nu(x)\cdot A^{(2)}w\bar{v}\,dS-\int_{\partial(\Sigma\cap B)}\partial_{\nu}(-\Delta w)\bar{v}\,dS \\ \vspace{1ex}
&+\displaystyle\int_{\partial(\Sigma\cap B)}(-\Delta w)\overline{\partial_{\nu}v}\,dS-\int_{\partial(\Sigma\cap B)}\partial_{\nu}w\overline{(-\Delta v)}\,dS \\ \vspace{1ex}
&+\displaystyle\int_{\partial(\Sigma\cap B)}w\overline{(\partial_{\nu}(-\Delta v))}\,dS\\
:=&I_{1}+I_{2}+I_{3}+I_{4}+I_{5}
\end{array}
\end{equation}
We analyze each term on the right-hand side and show $I_{j}=0, j=1,\cdots,5$.
$$I_{1}:=-i\displaystyle\int_{\partial(\Sigma\cap B)}\nu(x)\cdot A^{(2)}w\bar{v}\,dS.$$
By \eqref{results}, $w=0$ on $\partial(\Sigma\cap B)$; hence $I_{1}=0$.
$$I_{2}:=-\displaystyle\int_{\partial(\Sigma\cap B)}\partial_{\nu}(-\Delta w)\bar{v}\,dS.$$
By \eqref{condition_v}, $v=0$ on $l_{1}$; by \eqref{results}, $\partial_{\nu}(\Delta w)=0$ on $l_{2}\cup l_{3}$; hence $I_{2}=0$.
$$I_{3}:=\displaystyle\int_{\partial(\Sigma\cap B)}(-\Delta w)\overline{\partial_{\nu}v}\,dS.$$
By definition, $\Delta w=0$ on $l_{1}\cup l_{2}$; by \eqref{results}, $\Delta w=0$ on $l_{3}$; hence $I_{3}=0$.
$$I_{4}:=-\displaystyle\int_{\partial(\Sigma\cap B)}\partial_{\nu}w\overline{(-\Delta v)}\,dS.$$
By \eqref{condition_v}, $\Delta v=0$ on $l_{1}$; by \eqref{results}, $\partial_{\nu}w=0$ on $l_{2}\cup l_{3}$; hence $I_{4}=0$.
$$I_{5}:=\displaystyle\int_{\partial(\Sigma\cap B)}w\overline{(\partial_{\nu}(-\Delta v))}\,dS.$$
By definition, $w=0$ on $l_{1}\cup l_{2}$; by \eqref{results}, $w=0$ on $l_{3}$; hence $I_{5}=0$.
Putting these together, from \eqref{identity1} we obtain
\begin{equation} \label{identity2}
\displaystyle\int_{\Sigma\cap B}((A^{(1)}-A^{(2)})\cdot Du_{1})\bar{v}\,dx+\int_{\Sigma\cap B}(q^{(1)}-q^{(2)})u_{1}\bar{v}\,dx=0.
\end{equation}
for all $u_{1}\in \mathcal{W}(\Sigma)$ and $v\in \mathcal{V}_{l_{1}}(\Sigma\cap B)$. Here for $j=1,2,$ we define some function spaces for later use:
$$
\begin{array}{rl} \vspace{1ex}
\mathcal{W}(\Sigma):=&\{u\in H^{4}(\Sigma):\mathcal{L}_{A^{(1)},q^{(1)}}u=0 \textrm{ in } \Sigma, u|_{\Gamma_{2}}=\Delta u|_{\Gamma_{2}}=0,\\
& \textrm{ supp}(u|_{\Gamma_{1}})\subset\gamma_{1}, \textrm{ supp}(\Delta u|_{\Gamma_{1}})\subset\gamma_{1}\}.\\
\end{array}
$$
$$
\mathcal{V}_{l_{j}}(\Sigma\cap B):=\{v\in H^{4}(\Sigma\cap B):\mathcal{L}^{\ast}_{A^{(2)},q^{(2)}}v=0 \textrm{ in } \Sigma\cap B, v|_{l_{j}}=\Delta v|_{l_{j}}=0\}.
$$
$$
\mathcal{W}_{l_{j}}(\Sigma\cap B):=\{u\in H^{4}(\Sigma\cap B):\mathcal{L}_{A^{(1)},q^{(1)}}u=0 \textrm{ in } \Sigma\cap B, u|_{l_{j}}=\Delta u|_{l_{j}}=0\}.
$$
We would like to replace $u_{1}\in \mathcal{W}(\Sigma)$ in \eqref{identity2} by elements of the space $\mathcal{W}_{l_{2}}(\Sigma\cap B)$. This can be achieved by the following Runge type approximation result.
\begin{prop}\label{runge}
$\mathcal{W}(\Sigma)$ is a dense subspace of $\mathcal{W}_{l_{2}}(\Sigma\cap B)$ in $L^{2}(\Sigma\cap B)$ topology.
\end{prop}
\begin{proof}
It suffices to establish the following fact: for any $g\in L^{2}(\Sigma\cap B)$ such that
$$\displaystyle\int_{\Sigma\cap B} u\overline{g}\,dx=0 \quad\quad \forall u\in \mathcal{W}(\Sigma),$$
we have
$$\displaystyle\int_{\Sigma\cap B} v\overline{g}\,dx=0 \quad\quad \forall v\in \mathcal{W}_{l_{2}}(\Sigma\cap B).$$
To prove this fact, we extend $g$ by zero to $\Sigma\backslash\Sigma\cap B$. Let $U\in H^{4}(\Sigma)$ be the solution of the problem
$$
\begin{array}{rl} \vspace{1ex}
\mathcal{L}^{\ast}_{A^{(1)},q^{(1)}}U=g & \quad \textrm{ in } \Sigma\\ \vspace{1ex}
U=\Delta U=0 & \quad \textrm{ on } \Gamma_{1}\cup\Gamma_{2}.\\
\end{array}
$$
For any $u\in \mathcal{W}(\Sigma)$, Green's formula in the infinite slab $\Sigma$ (see appendix B) gives
$$
0=\displaystyle\int_{\Sigma}u\overline{g}\,dx=\int_{\Sigma}u\overline{\left(\mathcal{L}^{\ast}_{A^{(1)},q^{(1)}}U\right)}\,dx
=\int_{\Gamma_{1}}\overline{\partial_{\nu}U}\Delta u\,dS+\int_{\Gamma_{1}}\overline{\partial_{\nu}\Delta U} u\,dS.
$$
Since $u|_{\Gamma_{1}}$ and $\Delta u|_{\Gamma_{1}}$ can be arbitrary smooth functions supported in $\gamma_{1}$, we conclude that $\partial_{\nu}U|_{\gamma_{1}}=\partial_{\nu}\Delta U|_{\gamma_{1}}=0$. Hence $U$ satisfies $\Delta^{2} U=0$ in $\Sigma\backslash B$, and moreover, $U=\partial_{\nu}U=0$ on $\gamma_{1}\backslash l_{1}$. Thus, by unique continuation, $U=0$ in $\Sigma\backslash B$, and we have $U=\partial_{\nu} U=0$ on $l_{3}$. Similarly $\Delta U$ satisfies $\Delta(\Delta U)=0$ in $\Sigma\backslash B$ and $\Delta U=\partial_{\nu}\Delta U=0$ on $\gamma_{1}\backslash l_{1}$. Again by unique continuation, $\Delta U=0$ in $\Sigma\backslash B$, and we have $\Delta U=\partial_{\nu}\Delta U=0$ on $l_{3}$.
For any $v\in \mathcal{W}_{l_{2}}(\Sigma\cap B)$, using Green's formula on the bounded domain $\Sigma\cap B$ we get
$$
\begin{array}{rl}\vspace{1ex}
\displaystyle\int_{\Sigma\cap B}v\overline{g}\,dx= & \displaystyle\int_{\Sigma\cap B}v\overline{(\mathcal{L}^{\ast}_{A^{(1)},q^{(1)}}U)}\,dx \\ \vspace{1ex}
=& \displaystyle\int_{\Sigma\cap B}\left(\mathcal{L}_{A^{(1)},q^{(1)}}v\right)\overline{U}\,dx+i\int_{\partial(\Sigma\cap B)}\nu(x)\cdot Av \overline{U}\,dS\\ \vspace{1ex}
& +\displaystyle\int_{\partial(\Sigma\cap B)}\partial_{\nu}(-\Delta v)\overline{U}\,dS-\displaystyle\int_{\partial(\Sigma\cap B)}(-\Delta v)\overline{\partial_{\nu}U}\,dS\\ \vspace{1ex}
& +\displaystyle\int_{\partial(\Sigma\cap B)}\partial_{\nu}v\overline{(-\Delta U)}\,dS-\displaystyle\int_{\partial(\Sigma\cap B)}v\overline{(\partial_{\nu}(-\Delta U))}\,dS\\ \vspace{1ex}
=& 0.\\
\end{array}
$$
\end{proof}
Combining \eqref{identity2} with Proposition \ref{runge} we conclude
\begin{prop}
\begin{equation} \label{identity3}
\displaystyle\int_{\Sigma\cap B}((A^{(1)}-A^{(2)})\cdot Du_{1})\bar{v}\,dx+\int_{\Sigma\cap B}(q^{(1)}-q^{(2)})u_{1}\bar{v}\,dx=0.
\end{equation}
for all $u_{1}\in \mathcal{W}_{l_{2}}(\Sigma\cap B)$ and $v\in \mathcal{V}_{l_{1}}(\Sigma\cap B)$.
\end{prop}
\section{Construction of CGO solutions in the infinite slab}
In this section we construct CGO solutions $u_{1}\in \mathcal{W}_{l_{2}}(\Sigma\cap B)$ and $v\in \mathcal{V}_{l_{1}}(\Sigma\cap B)$. Let $\xi,\mu^{(1)},\mu^{(2)}\in\mathbb{R}^{n}$ be such that $|\mu^{(1)}|=|\mu^{(2)}|=1$ and $\mu^{(1)}\cdot\mu^{(2)}=\mu^{(1)}\cdot\xi=\mu^{(2)}\cdot\xi=0$. We set
\begin{equation}
\zeta_{1}:=\displaystyle\frac{ih\xi}{2}+i\sqrt{1-h^{2}\frac{|\xi|^{2}}{4}}\mu^{(1)}+\mu^{(2)}, \quad \zeta_{2}:=-\displaystyle\frac{ih\xi}{2}+i\sqrt{1-h^{2}\frac{|\xi|^{2}}{4}}\mu^{(1)}-\mu^{(2)}.
\end{equation}
Note that $\zeta_{1}\cdot\zeta_{1}=\zeta_{2}\cdot\zeta_{2}=0$, and $(\zeta_{1}+\overline{\zeta_{2}})/h=i\xi$. Here $h>0$ is a small semiclassical parameter. Note also that
\begin{equation}\label{zeta}
\zeta_{1}=i\mu^{(1)}+\mu^{(2)}+\mathcal{O}(h) \textrm{ and } \zeta_{2}=i\mu^{(1)}-\mu^{(2)}+\mathcal{O}(h) \textrm{ as } h\rightarrow 0,
\end{equation}
so $\zeta^{(0)}_{1}=i\mu^{(1)}+\mu^{(2)}$, $\zeta^{(0)}_{2}=i\mu^{(1)}-\mu^{(2)}$.
We first construct $u_{1}\in \mathcal{W}_{l_{2}}(\Sigma\cap B)$. To satisfy the condition $u_{1}|_{l_{2}}=\Delta u_{1}|_{l_{2}}=0$, we reflect $\Sigma\cap B$ with respect to the plane $x_{n}=0$ and denote this reflection by $(\Sigma\cap B)^{\ast}_{0}:=\{(x',-x_{n}):x=(x',x_{n})\in\Sigma\cap B\}$ where $x'=(x_{1},\cdots,x_{n-1})$. We extend the coefficients $A^{(1)}$ and $q^{(1)}$ to $(\Sigma\cap B)^{\ast}_{0}$ as follows: for the components $A^{(1)}_{j},j=1,\cdots,n-1$ and $q^{(1)}$, we extend them as even functions with respect to $x_{n}=0$, for $A^{(1)}_{n}$ we extend it as an odd function with respect to $x_{n}=0$, i.e. we set
$$
\begin{array}{rl} \vspace{1ex}
\tilde{A}^{(1)}_{j}(x)=&
\left\{
\begin{array}{ll} \vspace{1ex}
A^{(1)}_{j}(x',x_{n}) & 0<x_{n}<L \\
A^{(1)}_{j}(x',-x_{n}) & -L<x_{n}<0 \\
\end{array}
\right.
, \quad j=1,\cdots,n-1 \\ \vspace{1ex}
\tilde{A}^{(1)}_{n}(x)=&
\left\{
\begin{array}{ll} \vspace{1ex}
A^{(1)}_{n}(x',x_{n}) & 0<x_{n}<L \\
-A^{(1)}_{n}(x',-x_{n}) & -L<x_{n}<0 \\
\end{array}
\right. \\ \vspace{1ex}
\tilde{q}^{(1)}(x)=&
\left\{
\begin{array}{ll} \vspace{1ex}
q^{(1)}(x',x_{n}) & 0<x_{n}<L \\
q^{(1)}(x',-x_{n}) & -L<x_{n}<0 \\
\end{array}
\right. .\\
\end{array}
$$
For the moment, let us assume $A^{(1)}_{n}|_{x_{n}=0}=0$ so that $\tilde{A}^{(1)}\in W^{1,\infty}((\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{0})$ and $\tilde{q}^{(1)}\in L^{\infty}((\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{0})$. We will come back to the general case after establishing Proposition \ref{curlvanish}.
Proposition \ref{existence} implies that there exist CGO solutions of the form
$$\tilde{u}_{1}(x,\zeta_{1},h)=e^{x\cdot\zeta_{1}/h}(a_{1}(x,\zeta^{(0)}_{1})+r_{1}(x,\zeta_{1},h))\in H^{2}((\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{0})$$
which satisfy the equation $\mathcal{L}_{\tilde{A}^{(1)},\tilde{q}^{(1)}}\tilde{u}_{1}=0$ in the bounded region $(\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{0}$ with
\begin{equation}
((i\mu^{(1)}+\mu^{(2)})\cdot\nabla)^{2}a_{1}=0 \quad \textrm{ in } (\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{0},
\end{equation}
\begin{equation}
\|r_{1}\|_{H^{1}_{\textrm{scl}}((\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{0})}=\mathcal{O}(h).
\end{equation}
Let
\begin{equation} \label{CGO_u}
u_{1}(x):=\tilde{u}_{1}(x',x_{n})-\tilde{u}_{1}(x',-x_{n}) \quad\quad x\in\Sigma\cap B.
\end{equation}
Then it is easy to check that $u_{1}\in \mathcal{W}_{l_{2}}(\Sigma\cap B)$.\\
To construct $v\in \mathcal{V}_{l_{1}}(\Sigma\cap B)$, we notice that $\mathcal{L}^{\ast}_{A,q}=\mathcal{L}_{\overline{A},i^{-1}\nabla\cdot\overline{A}+\overline{q}}$, so $\mathcal{L}^{\ast}_{A^{(2)},q^{(2)}}v=0$ is equivalent to $\mathcal{L}_{A^{(3)},q^{(3)}}v=0$ where $A^{(3)}:=\overline{A^{(2)}}$ and $q^{(3)}:=i^{-1}\nabla\cdot\overline{A^{(2)}}+\overline{q^{(2)}}$. In the following we will construct $v$ such that $\mathcal{L}_{A^{(3)},q^{(3)}}v=0$ with $v|_{l_{1}}=\Delta v|_{l_{1}}=0$. To this end, we reflect $\Sigma\cap B$ with respect to the plane $x_{n}=L$ and denote this reflection by $(\Sigma\cap B)^{\ast}_{L}:=\{(x',-x_{n}+2L):x=(x',x_{n})\in\Sigma\cap B\}$. We extend the coefficients $A^{(3)}$ and $q^{(3)}$ to $(\Sigma\cap B)^{\ast}_{L}$ as follows: for $A^{(3)}_{j},j=1,\cdots,n-1$ and $q^{(3)}$ we extend them as even functions with respect to $x_{n}=L$, for $A^{(3)}_{n}$ we extend it as an odd function with respect to $x_{n}=L$, i.e.
$$
\begin{array}{rl} \vspace{1ex}
\tilde{A}^{(3)}_{j}(x)=&
\left\{
\begin{array}{ll} \vspace{1ex}
A^{(3)}_{j}(x',x_{n}) & 0<x_{n}<L \\
A^{(3)}_{j}(x',-x_{n}+2L) & L<x_{n}<2L \\
\end{array}
\right.
, \quad j=1,\cdots,n-1 \\ \vspace{1ex}
\tilde{A}^{(3)}_{n}(x)=&
\left\{
\begin{array}{ll} \vspace{1ex}
A^{(3)}_{n}(x',x_{n}) & 0<x_{n}<L \\
-A^{(3)}_{n}(x',-x_{n}+2L) & L<x_{n}<2L \\
\end{array}
\right. \\ \vspace{1ex}
\tilde{q}^{(3)}(x)=&
\left\{
\begin{array}{ll} \vspace{1ex}
q^{(3)}(x',x_{n}) & 0<x_{n}<L \\
q^{(3)}(x',-x_{n}+2L) & L<x_{n}<2L \\
\end{array}
\right. .\\
\end{array}
$$
Again, first we assume $A^{(2)}_{n}|_{x_{n}=L}=0$ so that $A^{(3)}_{n}|_{x_{n}=L}=0$, $\tilde{A}^{(3)}\in W^{1,\infty}((\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{L})$ and $\tilde{q}^{(3)}\in L^{\infty}((\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{L})$. The general case will be dealt with below Proposition \ref{curlvanish}.
Proposition \ref{existence} implies that there exist CGO solutions of the form
$$\tilde{v}(x,\zeta_{2},h)=e^{x\cdot\zeta_{2}/h}(a_{2}(x,\zeta^{(0)}_{2})+r_{2}(x,\zeta_{2},h))\in H^{2}((\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{L})$$
which satisfy the equation $\mathcal{L}_{\tilde{A}^{(3)},\tilde{q}^{(3)}}\tilde{v}=0$ in the bounded region $(\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{L}$ with
$$
((i\mu^{(1)}-\mu^{(2)})\cdot\nabla)^{2}a_{2}=0 \quad \textrm{ in } (\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{L},
$$
$$
\|r_{2}\|_{H^{1}_{\textrm{scl}}((\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{L})}=\mathcal{O}(h).
$$
Let
\begin{equation}\label{CGO_v}
v(x):=\tilde{v}(x',x_{n})-\tilde{v}(x',-x_{n}+2L) \quad\quad x\in\Sigma\cap B.
\end{equation}
Then it is easy to check that $v\in \mathcal{V}_{l_{1}}(\Sigma\cap B)$.
We write down the CGO solutions \eqref{CGO_u} and \eqref{CGO_v} explicitly for future references:
\begin{equation}\label{CGO1}
u_{1}(x)=e^{x\cdot\zeta_{1}/h}(a_{1}(x)+r_{1}(x))-e^{(x',-x_{n})\cdot\zeta_{1}/h}(a_{1}(x',-x_{n})+r_{1}(x',-x_{n}))
\end{equation}
\begin{equation}\label{CGO2}
v(x)=e^{x\cdot\zeta_{2}/h}(a_{2}(x)+r_{2}(x))-e^{(x',-x_{n}+2L)\cdot\zeta_{2}/h}(a_{2}(x',-x_{n}+2L)+r_{2}(x',-x_{n}+2L))
\end{equation}
where $a_{1}\in C^{\infty}(\overline{(\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{0}})$ ,$a_{2}\in C^{\infty}(\overline{(\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{L}})$ and
\begin{equation}\label{condition1}
((i\mu^{(1)}+\mu^{(2)})\cdot\nabla)^{2} a_{1}=0 \quad \textrm{ in } (\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{0},
\end{equation}
\begin{equation}\label{condition2}
((i\mu^{(1)}-\mu^{(2)})\cdot\nabla)^{2} a_{2}=0 \quad \textrm{ in } (\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{L},
\end{equation}
\begin{equation}\label{condition3}
\|r_{1}\|_{H^{1}_{\textrm{scl}}((\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{0})}=\mathcal{O}(h),
\end{equation}
\begin{equation}\label{condition4}
\|r_{2}\|_{H^{1}_{\textrm{scl}}((\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{L})}=\mathcal{O}(h).
\end{equation}
\section{Proof of Theorem 1.1}
We are ready to prove our first main theorem. We will substitute the CGO solutions constructed in last section into \eqref{identity3}. To this end we compute
\begin{equation}\label{exp1}
\begin{array}{rl} \vspace{1ex}
e^{x\cdot\zeta_{1}/h}e^{x\cdot\overline{\zeta}_{2}/h}=&e^{ix\cdot\xi} \\ \vspace{1ex}
e^{(x',-x_{n})\cdot\zeta_{1}/h}e^{x\cdot\overline{\zeta}_{2}/h}=&e^{-2\mu^{(2)}_{n}x_{n}/h+ib_{1}} \\ \vspace{1ex}
e^{x\cdot\zeta_{1}/h}e^{(x',-x_{n}+2L)\cdot\overline{\zeta}_{2}/h}=&e^{2\mu^{(2)}_{n}(x_{n}-L)/h+ib_{2}} \\ \vspace{1ex}
e^{(x',-x_{n})\cdot\zeta_{1}/h}e^{(x',-x_{n}+2L)\cdot\overline{\zeta}_{2}/h}=&e^{-2L\mu^{(2)}_{n}/h+ib_{3}} \\
\end{array}
\end{equation}
where $b_{1}, b_{2}, b_{3}\in\mathbb{R}^{n}$ are defined by
$$
\begin{array}{rl} \vspace{1ex}
b_{1}:=&x'\cdot\xi'-\displaystyle\frac{2}{h}\sqrt{1-h^{2}\frac{|\xi|^{2}}{4}}\mu^{(1)}_{n}x_{n}, \\ \vspace{1ex}
b_{2}:=&x'\cdot\xi'+\displaystyle\frac{2}{h}\sqrt{1-h^{2}\frac{|\xi|^{2}}{4}}\mu^{(1)}_{n}(x_{n}-L)+L\xi_{n}, \\ \vspace{1ex}
b_{3}:=&x'\cdot\xi'-\displaystyle\frac{2L}{h}\sqrt{1-h^{2}\frac{|\xi|^{2}}{4}}\mu^{(1)}_{n}-x_{n}\xi_{n}+L\xi_{n}. \\
\end{array}
$$
We further assume that $\mu^{(2)}_{n}>0$, hence for $0<x_{n}<L$ the following pointwise convergence holds as $h\rightarrow 0+$:
\begin{equation}\label{vanish}
\begin{array}{rl} \vspace{1ex}
|e^{(x',-x_{n})\cdot\zeta_{1}/h}e^{x\cdot\overline{\zeta}_{2}/h}|&\rightarrow 0 \textrm{ as } h\rightarrow 0+,\\ \vspace{1ex}
|e^{x\cdot\zeta_{1}/h}e^{(x',-x_{n}+2L)\cdot\overline{\zeta}_{2}/h}|&\rightarrow 0 \textrm{ as } h\rightarrow 0+,\\ \vspace{1ex}
|e^{(x',-x_{n})\cdot\zeta_{1}/h}e^{(x',-x_{n}+2L)\cdot\overline{\zeta}_{2}/h}|&\rightarrow 0 \textrm{ as } h\rightarrow 0+.\\
\end{array}
\end{equation}
Therefore, with the CGO solutions $u_{1}$ and $v$ given by \eqref{CGO1} and \eqref{CGO2}, we conclude from \eqref{condition3} \eqref{condition4} and \eqref{vanish} that
\begin{equation}\label{term2}
h\displaystyle\int_{\Sigma\cap B}(q^{(1)}-q^{(2)})u_{1}\bar{v}\,dx\rightarrow 0 \quad \textrm{ as } h\rightarrow 0+.
\end{equation}
On the other hand, denote $\zeta^{\ast}_{j}=(\zeta'_{j},-(\zeta_{j})_{n})$ for $\zeta_{j}=(\zeta'_{j},(\zeta_{j})_{n})$, $j=1,2$. Using \eqref{CGO1} we compute
\begin{equation}
\begin{array}{rl} \vspace{1ex}
Du_{1}(x)=&-\displaystyle\frac{i\zeta_{1}}{h}e^{x\cdot\zeta_{1}/h}(a_{1}(x)+r_{1}(x))+e^{x\cdot\zeta_{1}/h}(Da_{1}(x)+Dr_{1}(x))\\ \vspace{1ex}
&+\displaystyle\frac{i\zeta^{\ast}_{1}}{h}e^{(x',-x_{n})\cdot\zeta_{1}/h}(a_{1}(x',-x_{n})+r_{1}(x',-x_{n})) \\ \vspace{1ex}
&-e^{(x',-x_{n})\cdot\zeta_{1}/h}(Da_{1}(x',-x_{n})+Dr_{1}(x',-x_{n})).\\
\end{array}
\end{equation}
Therefore, with the CGO solutions $u_{1}$ and $v$ given by \eqref{CGO1} and \eqref{CGO2}, we have from \eqref{zeta} \eqref{condition3} \eqref{condition4} \eqref{vanish} and the dominant convergence theorem that
\begin{equation}\label{term1}
\begin{array}{rl}\vspace{1ex}
&h\displaystyle\int_{\Sigma\cap B}(A^{(1)}-A^{(2)})\cdot Du_{1}\bar{v}\,dx \\ \vspace{1ex}
\rightarrow& (\mu^{(1)}-i\mu^{(2)})\cdot\displaystyle\int_{\Sigma\cap B}(A^{(1)}-A^{(2)}) e^{ix\cdot\xi}a_{1}\overline{a_{2}}\,dx \quad \textrm{ as } h\rightarrow 0+.\\
\end{array}
\end{equation}
Multiplying \eqref{identity3} by $h$ and letting $h\rightarrow 0+$ for the constructed solutions $u_{1}$ and $v$, we obtain from \eqref{term2} and \eqref{term1} that
\begin{equation}\label{identity_A1}
(\mu^{(1)}-i\mu^{(2)})\cdot\displaystyle\int_{\Sigma\cap B}(A^{(1)}-A^{(2)}) e^{ix\cdot\xi}a_{1}\overline{a_{2}}\,dx=0.
\end{equation}
This identity holds for all $a_{1}$ satisfying \eqref{condition1}, $a_{2}$ satisfying \eqref{condition2}, and for all $\mu^{(1)},\mu^{(2)},\xi\in\mathbb{R}^{n}$ such that $\mu^{(1)}\cdot\mu^{(2)}=\mu^{(1)}\cdot\xi=\mu^{(2)}\cdot\xi=0$ and $\mu^{(2)}_{n}>0$. Replace $\mu^{(1)}$ by $-\mu^{(1)}$ and subtract to find
\begin{equation}\label{identity_A2}
\mu^{(1)}\cdot\displaystyle\int_{\Sigma\cap B}(A^{(1)}-A^{(2)}) e^{ix\cdot\xi}a_{1}\overline{a_{2}}\,dx=0.
\end{equation}
\begin{prop} \label{curlvanish}
\begin{equation} \label{curl1}
\partial_{j}(A^{(1)}_{k}-A^{(2)}_{k})-\partial_{k}(A^{(1)}_{j}-A^{(2)}_{j})=0 \textrm{ in } \Sigma\cap B, \quad 1\leq j,k\leq n.
\end{equation}
\end{prop}
\begin{proof}
Obviously $a_{1}=a_{2}=1$ satisfies \eqref{condition1} and \eqref{condition2}. Inserting $a_{1}=a_{2}=1$ in \eqref{identity_A2} we get
\begin{equation}\label{fourier}
\mu^{(1)}\cdot(\widehat{A^{(1)}\chi_{\Sigma\cap B}}(\xi)-\widehat{A^{(2)}\chi_{\Sigma\cap B}}(\xi))=0
\end{equation}
where $\chi_{\Sigma\cap B}$ stands for the characteristic function of the set $\Sigma\cap B$ and $\widehat{A^{(j)}\chi_{\Sigma\cap B}}$ denotes the Fourier transform of $A^{(j)}\chi_{\Sigma\cap B}$.
To show the proposition, it suffices to consider the case when $j\neq k$. Let $e_{1},\cdots,e_{n}$ be the standard orthonormal basis in $\mathbb{R}^{n}$. Let $\xi=(\xi_{1},\cdots,\xi_{n})$ with $\xi_{j}>0, j=1,\cdots,n$. Define
$$\mu^{(1)}=-\xi_{k}e_{j}+\xi_{j}e_{k} \quad\quad 1\leq j,k\leq n, j\neq k.$$
To define $\mu^{(2)}$ we consider two cases: if $j,k$ are such that $1\leq j,k<n$, define
$$\mu^{(2)}=-\xi_{j}\xi_{n}e_{j}-\xi_{k}\xi_{n}e_{k}+(\xi^{2}_{j}+\xi^{2}_{k})e_{n};$$
if $k=n$ and $j$ is such that $1\leq j<n$, define
$$\mu^{(2)}=(-\xi^{2}_{j}-\xi^{2}_{n})e_{l}+\xi_{l}\xi_{j}e_{j}+\xi_{l}\xi_{n}e_{n}$$
with some $l\neq j,n$, which exists since $n\geq 3$. In either case it is easy to check $\mu^{(1)}\cdot\mu^{(2)}=\mu^{(1)}\cdot\xi=\mu^{(2)}\cdot\xi=0$ and $\mu^{(2)}_{n}>0$. For such $\mu^{(1)}$ and $\xi$ we get from \eqref{fourier} that
$$\xi_{j}\cdot(\widehat{A^{(2)}_{k}\chi_{\Sigma\cap B}}(\xi)-\widehat{A^{(1)}_{k}\chi_{\Sigma\cap B}}(\xi))-\xi_{k}\cdot(\widehat{A^{(2)}_{j}\chi_{\Sigma\cap B}}(\xi)-\widehat{A^{(1)}_{j}\chi_{\Sigma\cap B}}(\xi))=0,$$
$1\leq j,k\leq n, j\neq k$ for all $\xi\in\mathbb{R}^{n},\xi_{1}>0,\cdots,\xi_{n}>0$, and thus everywhere by analyticity of the Fourier transform. This completes the proof.
\end{proof}
By Proposition \ref{curlvanish}, we conclude $d A^{(1)}=d A^{(2)}$ in $\Sigma$. As $\Sigma$ is simply connected, there exists a compactly supported $\Phi\in C^{1,1}(\overline{\Sigma})$ such that
$$A^{(1)}-A^{(2)}=\nabla\Phi \quad\quad \textrm{ in } \Sigma.$$
In particular, $\Phi=0$ along $\partial B\cap\Sigma$.
Recall that in the construction of the CGO solutions above, we have assumed that $A^{(1)}_{n}|_{x_{n}=0}=0$ and $A^{(2)}_{n}|_{x_{n}=L}=0$. Now we show why our results are independent of such assumptions. Indeed, for $A^{(1)}$, there exists $\Psi^{(1)}\in C^{1,1}(\overline{\Sigma})$ with compact support such that $\Psi^{(1)}|_{\partial\Sigma}=0$ and $\partial_{\nu}\Psi^{(1)}=-A^{(1)}\cdot\nu$ on $\partial\Sigma$, where as usual $\nu$ is the unit outer normal vector on $\partial\Sigma$. Then $A^{(1)}+\nabla\Psi^{(1)}$ satisfies $(A^{(1)}_{n}+\nabla\Psi^{(1)}_{n})|_{x_{n}=0}=0$. See \cite[Theorem 1.3.3]{H} for the existence of $\Psi^{(1)}$. Similarly, we can find $\Psi^{(2)}\in C^{1,1}(\overline{\Sigma})$ with compact support such that $\Psi^{(2)}|_{\partial\Sigma}=0$ and $\partial_{\nu}\Psi^{(2)}=-A^{(2)}\cdot\nu$ on $\partial\Sigma$. Then $(A^{(2)}_{n}+\nabla\Psi^{(2)}_{n})|_{x_{n}=L}=0$. Therefore, we may replace $A^{(j)}$ by $A^{(j)}+\nabla\Psi^{(j)},j=1,2$ to fulfill the assumption. After the replacement, Proposition \ref{curlvanish} will give $d(A^{(1)}+\nabla\Psi^{(1)})=d(A^{(2)}+\nabla\Psi^{(2)})$ in $\Sigma$. As above we can find a compactly supported function $\Phi'\in C^{1,1}(\overline{\Sigma})$ such that
$$A^{(1)}+\nabla\Psi^{(1)}-A^{(2)}-\nabla\Psi^{(2)}=\nabla\Phi' \quad\quad \textrm{ in } \Sigma.$$
Define $\Phi:=\Phi'-\Psi^{(1)}+\Psi^{(2)}$, then $\Phi\in C^{1,1}(\overline{\Sigma})$ is compactly supported and satisfies $A^{(1)}-A^{(2)}=\nabla\Phi$. In particular $\Phi=0$ on $\partial B\cap\Sigma$. We are back to the same situation.
Next, we establish a proposition which asserts that $\Phi=0$ on $\Gamma_{1}\cup\Gamma_{2}$.
\begin{prop}
$\Phi=0$ along $\partial(\Sigma\cap B)$.
\end{prop}
\begin{proof}
Notice \eqref{condition1} implies that in the expression \eqref{CGO1}, we may replace $a_{1}$ by $g_{1}a_{1}$ if $g_{1}\in C^{\infty}(\overline{(\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{0}\cup(\Sigma\cap B)^{\ast}_{L}})$ satisfies
$$(i\mu^{(1)}+\mu^{(2)})\cdot\nabla g_{1}=0 \quad\quad \textrm{ in } (\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{0}\cup(\Sigma\cap B)^{\ast}_{L}.$$
Thus \eqref{identity_A1} becomes
$$(\mu^{(1)}-i\mu^{(2)})\cdot\displaystyle\int_{\Sigma\cap B}\nabla\Phi g_{1}e^{ix\cdot\xi}a_{1}\overline{a_{2}}\,dx=0.$$
Set $\xi=0$, $a_{1}=a_{2}=1$ and multiply by $i$:
\begin{equation}\label{g}
(i\mu^{(1)}+\mu^{(2)})\cdot\displaystyle\int_{\Sigma\cap B}\nabla\Phi g_{1}\,dx=0
\end{equation}
As $\mu^{(1)}\cdot\mu^{(2)}=0$ and $|\mu^{(1)}|=|\mu^{(1)}|=1$, we can make a change of variable so that $(i\mu^{(1)}+\mu^{(2)})\cdot\nabla$ becomes a $\overline{\partial}$-operator as follows. Complete the set $\{\mu^{(2)}$, $\mu^{(1)}\}$ to an orthonormal basis in $\mathbb{R}^{n}$, say $\{\mu^{(2)},\mu^{(1)},\mu^{(3)},\cdots,\mu^{(n)}\}$; introduce new coordinates $y=(y_{1},\cdots,y_{n})\in\mathbb{R}^{n}$ with respect to this orthonormal basis by defining $y_{1}=x\cdot\mu^{(2)},y_{2}=x\cdot\mu^{(1)},y_{j}=x\cdot\mu^{(j)},j=3,\cdots,n$; in other words, we made an orthogonal transformation $T:\mathbb{R}^{n}\rightarrow\mathbb{R}^{n}$, $T(x)=y$. Denote $z=y_{1}+iy_{2}$ and $\partial_{\overline{z}}=\frac{1}{2}(\partial_{y_{1}}+i\partial_{y_{2}})$. Then $(i\mu^{(1)}+\mu^{(2)})\cdot\nabla=2\partial_{\overline{z}}$, and in the new coordinates \eqref{g} becomes
$$\displaystyle\int_{T(\Sigma\cap B)}g_{1}\partial_{\overline{z}}\Phi\,dy=0$$
for all $g_{1}\in C^{\infty}(\overline{(\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{0}\cup(\Sigma\cap B)^{\ast}_{L}})$ satisfying $\partial_{\overline{z}}g_{1}=0$. Replacing $\mu^{(1)}$ by $-\mu^{(1)}$, in the same way we can show
$$\displaystyle\int_{T(\Sigma\cap B)}g_{2}\partial_{z}\Phi\,dy=0$$
for all $g_{2}\in C^{\infty}(\overline{(\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{L}\cup(\Sigma\cap B)^{\ast}_{0}})$ satisfying $\partial_{z}g_{2}=0$. Taking $g_{j}(y)=g'_{j}(z)\otimes g''_{j}(y''),\; j=1,2, \; y''=(y_{3},\cdots,y_{n})$ and varying $g''_{j}$ yields
$$\displaystyle\int_{T_{y''}}g'_{1}(z)\partial_{\overline{z}}\Phi\,dz\wedge d\overline{z}=0 \quad\quad \displaystyle\int_{T_{y''}}g'_{2}(\overline{z})\partial_{z}\Phi\,dz\wedge d\overline{z}=0$$
Here $T_{y''}$ is the intersection of $T(\Sigma\cap B)$ with the two dimensional plane $\{(y_{1},y_{2},y'')\in\mathbb{R}^{n}:y'' \textrm{ fixed}\}$, $\partial_{\overline{z}}g'_{1}=0$ and $\partial_{z}g'_{2}=0$. Notice that $\partial T_{y''}$ is piecewise smooth. Since
$$d(g'_{1}(z)\Phi\,dz)=g'_{1}(z)\partial_{\overline{z}}\Phi\,d\overline{z}\wedge dz, \quad\quad d(g'_{2}(\overline{z})\Phi\,d\overline{z})=g'_{2}(\overline{z})\partial_{z}\Phi\,dz\wedge d\overline{z},$$
we obtain from Stokes' formula that
$$\displaystyle\int_{\partial T_{y''}}g'_{1}(z)\Phi\,dz=0 \quad\quad \displaystyle\int_{\partial T_{y''}}g'_{2}(\overline{z})\Phi\,d\overline{z}=0.$$
Taking $g'_{2}=\overline{g'_{1}}$ we see that
$$\displaystyle\int_{\partial T_{y''}}g'_{1}(z)\Phi\,dz=0 \quad\quad \displaystyle\int_{\partial T_{y''}}g'_{1}(z)\overline{\Phi}\,dz=0$$
Hence
$$\displaystyle\int_{\partial T_{y''}}g'_{1}(z)\textrm{Re }\Phi\,dz=\displaystyle\int_{\partial T_{y''}}g'_{1}(z)\textrm{Im }\Phi\,dz=0.$$
for all holomorphic functions $g'_{1}\in C^{\infty}(\overline{T_{y''}})$. Arguing as in \cite[Lemma 5.1]{FKSU}, we can find holomorphic functions $F_{1},F_{2}\in C(\overline{T_{y''}})$ such that
$$F_{1}|_{\partial T_{y''}}=\textrm{Re }\Phi|_{\partial T_{y''}} \quad\quad F_{2}|_{\partial T_{y''}}=\textrm{Im }\Phi|_{\partial T_{y''}}.$$
Moreover, $\Delta\textrm{Im }F_{j}=0$ in $T_{y''}$ and $\textrm{Im }F_{j}|_{\partial T_{y''}}=0$. Thus, $F_{j},\; j=1,2,$ are real-valued and thus constant on $T_{y''}$. Therefore, $\Phi$ is constant along $\partial T_{y''}$. In the $x$-coordinate system, we see that the function $\Phi(x)$ is constant on the boundary of the intersection $T^{-1}(\Pi_{y''})\cap (\Sigma\cap B)$ for all $y''\in\mathbb{R}^{n-2}$, where $T^{-1}(\Pi_{y''})$ is defined by
$$T^{-1}(\Pi_{y''}):=\{x=y_{1}\mu^{(2)}+y_{2}\mu^{(1)}+\sum^{n}_{j=3}y_{j}\mu^{(j)}: y_{1},y_{2}\in\mathbb{R},y''=(y_{3},\cdots,y_{n})\}.$$
Setting $\mu^{(1)}=e_{j},\; j=1,\cdots,n-1$ and $\mu^{(2)}=e_{n}$, then varying $y''$ gives that $\Phi$ vanishes on $\partial(\Sigma\cap B)$. This completes the proof.
\end{proof}
To show that $A^{(1)}=A^{(2)}$ consider \eqref{identity_A1} with $a_{2}=1$ and $a_{1}$ satisfying
$$((i\mu^{(1)}+\mu^{(2)})\cdot\nabla) a_{1}=1 \quad \textrm{ in } (\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{0}.$$
This choice is possible thanks to \eqref{condition1}. We have from \eqref{identity_A1} that
$$(i\mu^{(1)}+\mu^{(2)})\cdot\displaystyle\int_{\Sigma\cap B}(\nabla\Phi)e^{ix\cdot\xi}a_{1}\,dx=0$$
Integrating by parts and using the fact that $\Phi=0$ along $\partial(\Sigma\cap B)$ and $\mu^{(1)}\cdot\xi=\mu^{(2)}\cdot\xi=0$ we obtain
$$0=\displaystyle\int_{\Sigma\cap B}\Phi(x)e^{ix\cdot\xi}[(i\mu^{(1)}+\mu^{(2)})\cdot\nabla a_{1}]\,dx=\int_{\Sigma\cap B}\Phi(x)e^{ix\cdot\xi}\,dx.$$
This indicates that Fourier transform of the function $\Phi\chi_{\Sigma\cap B}$ vanishes. Thus $\Phi=0$ in $\Sigma\cap B$, and therefore $A^{(1)}=A^{(2)}$.\vspace{1ex}
Inserting $A^{(1)}=A^{(2)}$ in \eqref{identity3} gives
$$\displaystyle\int_{\Sigma\cap B}(q^{(1)}-q^{(2)})u_{1}\bar{v}\,dx=0.$$
Let $u_{1}$ and $v$ be the CGO solutions given by \eqref{CGO1} and \eqref{CGO2}. Taking the limit $h\rightarrow 0+$, from \eqref{condition3} \eqref{condition4} \eqref{vanish} we get
$$\displaystyle\int_{\Sigma\cap B}(q^{(1)}-q^{(2)})e^{ix\cdot\xi}a_{1}\bar{a}_{2}\,dx=0$$
where $a_{1}$ and $a_{2}$ satisfy \eqref{condition1} and \eqref{condition2} respectively. In particular, for $a_{1}=a_{2}=1$ this identity becomes
\begin{equation}\label{identity_q}
\displaystyle\int_{\Sigma\cap B}(q^{(1)}-q^{(2)})e^{ix\cdot\xi}\,dx=0
\end{equation}
for all $\xi$ such that there exist $\mu^{(1)},\mu^{(2)}\in\mathbb{R}^{n}$ such that
$$\mu^{(1)}\cdot\mu^{(2)}=\xi\cdot\mu^{(1)}=\xi\cdot\mu^{(2)}=0, \quad |\mu^{(1)}|=|\mu^{(2)}|=1,\quad \mu^{(2)}_{n}>0.$$
Write $\xi=(\xi',\xi_{n-1},\xi_{n})$ with $\xi'\in\mathbb{R}^{n-2}$. If $\xi_{n-1}\neq 0$, we can choose
$$\mu^{(2)}=\displaystyle\frac{1}{\sqrt{1+\frac{\xi^{2}_{n}}{\xi^{2}_{n-1}}}}(0_{\mathbb{R}^{n-2}},\frac{-\xi_{n}}{\xi_{n-1}},1),$$
which satisfies $\xi\cdot\mu^{(2)}=0$, $|\mu^{(2)}|=1$ and $\mu^{(2)}_{n}>0$. Since $n\geq 3$, we can find a third unit vector $\mu^{(1)}$ so that $\{\mu^{(1)},\mu^{(2)},\xi\}$ are mutually orthogonal. Thus \eqref{identity_q} indicates that $\widehat{q^{(1)}\chi_{\Sigma\cap B}}(\xi)=\widehat{q^{(2)}\chi_{\Sigma\cap B}}(\xi)$ for $\xi$ with $\xi_{n-1}\neq 0$, and therefore for all $\xi\in\mathbb{R}^{n}$ as both Fourier transforms are continuous functions. This completes the proof of Theorem 1.1.
\vspace{3ex}
\section{Proof of Theorem 1.2}
In this section we show Theorem 1.2. First, arguing as in the proof of Theorem 1.1, we derive identity \eqref{identity3} for all $u_{1}\in \mathcal{W}_{l_{2}}(\Sigma\cap B)$ and $v\in \mathcal{V}_{l_{2}}(\Sigma\cap B)$. Next, we construct CGO solutions to be used in the proof of Theorem 1.2. We have constructed $u_{1}\in \mathcal{W}_{l_{2}}(\Sigma\cap B)$ in \eqref{CGO_u}, now we construct $v\in \mathcal{V}_{l_{2}}(\Sigma\cap B)$. As in the construction of $u_{1}$, we will reflect the coefficients with respect to the plane $x_{n}=0$. Recall that we have introduced $A^{(3)}=\overline{A^{(2)}}$ and $q^{(3)}=i^{-1}\nabla\cdot\overline{A^{(2)}}+\overline{q^{(2)}}$ so that $\mathcal{L}^{\ast}_{A^{(2)},q^{(2)}}=\mathcal{L}_{A^{(3)},q^{(3)}}$. For $A^{(3)}_{j}, j=1,\cdots,n-1$ and $q^{(3)}$, we extend them as even functions with respect to $x_{n}=0$; for $A^{(3)}_{n}$, we extend it as an odd function with respect to $x_{n}=0$, i.e. we set
$$
\begin{array}{rl} \vspace{1ex}
\tilde{A}^{(3)}_{j}(x)=&
\left\{
\begin{array}{ll} \vspace{1ex}
A^{(3)}_{j}(x',x_{n}) & 0<x_{n}<L \\
A^{(3)}_{j}(x',-x_{n}) & -L<x_{n}<0 \\
\end{array}
\right.
, \quad j=1,\cdots,n-1 \\ \vspace{1ex}
\tilde{A}^{(3)}_{n}(x)=&
\left\{
\begin{array}{ll} \vspace{1ex}
A^{(3)}_{n}(x',x_{n}) & 0<x_{n}<L \\
-A^{(3)}_{n}(x',-x_{n}) & -L<x_{n}<0 \\
\end{array}
\right. \\ \vspace{1ex}
\tilde{q}^{(3)}(x)=&
\left\{
\begin{array}{ll} \vspace{1ex}
q^{(3)}(x',x_{n}) & 0<x_{n}<L \\
q^{(3)}(x',-x_{n}) & -L<x_{n}<0 \\
\end{array}
\right. .\\
\end{array}
$$
Without loss of generality we assume $A^{(3)}_{n}|_{x_{n}=0}=0$, as we did before. Then $\tilde{A}^{(3)}\in W^{1,\infty}((\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{0})$ and $\tilde{q}^{(3)}\in L^{\infty}((\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{0})$. Proposition \ref{existence} implies that there exist CGO solutions of the form
$$\tilde{v}(x,\zeta_{2},h)=e^{x\cdot\zeta_{2}/h}(a_{2}(x,\zeta^{(0)}_{2})+r_{2}(x,\zeta_{2},h))\in H^{2}((\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{0})$$
which satisfy the equation $\mathcal{L}_{\tilde{A}^{(3)},\tilde{q}^{(3)}}v=0$ in the bounded region $(\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{0}$ with
\begin{equation}
((i\mu^{(1)}-\mu^{(2)})\cdot\nabla)^{2}a_{2}=0 \quad \textrm{ in } (\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{0},
\end{equation}
\begin{equation}
\|r_{2}\|_{H^{1}_{\textrm{scl}}((\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{0})}=\mathcal{O}(h).
\end{equation}
Let
\begin{equation} \label{CGO_v2}
v(x):=\tilde{v}(x',x_{n})-\tilde{v}(x',-x_{n}) \quad\quad x\in\Sigma\cap B.
\end{equation}
Then it is easy to see that $v\in \mathcal{V}_{l_{2}}(\Sigma\cap B)$.\\
It will be convenient to write down the CGO solutions \eqref{CGO_v2} explicitly for future references:
\begin{equation}\label{CGO3}
v(x)=e^{x\cdot\zeta_{2}/h}(a_{2}(x)+r_{2}(x))-e^{(x',-x_{n})\cdot\zeta_{2}/h}(a_{2}(x',-x_{n})+r_{2}(x',-x_{n}))
\end{equation}
where $a_{2}\in C^{\infty}(\overline{(\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{0}})$ and
\begin{equation}\label{condition5}
((i\mu^{(1)}-\mu^{(2)})\cdot\nabla)^{2} a_{2}=0 \quad \textrm{ in } (\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{0},
\end{equation}
\begin{equation}\label{condition6}
\|r_{2}\|_{H^{1}_{\textrm{scl}}((\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{0})}=\mathcal{O}(h).
\end{equation}
We will substitute the solutions \eqref{CGO1} and \eqref{CGO3} into \eqref{identity3}. To this end we compute
\begin{equation}\label{exp2}
\begin{array}{rl} \vspace{1ex}
e^{x\cdot\zeta_{1}/h}e^{x\cdot\overline{\zeta}_{2}/h}=&e^{ix\cdot\xi} \\ \vspace{1ex}
e^{x\cdot\zeta_{1}/h}e^{(x',-x_{n})\cdot\overline{\zeta}_{2}/h}=&e^{ix\cdot\xi_{+}+2\mu^{(2)}_{n}x_{n}/h}\\ \vspace{1ex}
e^{(x',-x_{n})\cdot\zeta_{1}/h}e^{x\cdot\overline{\zeta}_{2}/h}=&e^{ix\cdot\xi_{-}-2\mu^{(2)}_{n}x_{n}/h} \\ \vspace{1ex}
e^{(x',-x_{n})\cdot\zeta_{1}/h}e^{(x',-x_{n})\cdot\overline{\zeta}_{2}/h}=&e^{i(x',-x_{n})\cdot\xi} \\
\end{array}
\end{equation}
where
$$
\xi_{\pm}=\left(\xi',\pm\frac{2}{h}\sqrt{1-h^{2}\frac{|\xi|^{2}}{4}}\mu^{(1)}_{n}\right).
$$
Moreover, we assume $\mu^{(1)}_{n}\neq 0$ and $\mu^{(2)}_{n}=0$, so $\xi_{\pm}\rightarrow\infty$ as $h\rightarrow 0$. Then we have
\begin{equation}\label{vanish2}
\zeta_{1}\cdot\displaystyle\int_{\Sigma\cap B} (A^{(1)}-A^{(2)})e^{x\cdot\zeta_{1}/h}e^{(x',-x_{n})\cdot\overline{\zeta}_{2}/h}a_{1}a_{2}\,dx\rightarrow 0
\end{equation}
as $h\rightarrow 0$ by the Riemann-Lebesgue lemma. Similarly
\begin{equation}\label{vanish3}
\zeta_{1}\cdot\displaystyle\int_{\Sigma\cap B} (A^{(1)}-A^{(2)})e^{(x',-x_{n})\cdot\overline{\zeta}_{1}/h}e^{x\cdot\overline{\zeta}_{2}/h}a_{1}a_{2}\,dx\rightarrow 0
\end{equation}
as $h\rightarrow 0$. Therefore, multiplying \eqref{identity3} by $h$ and taking the limit $h\rightarrow 0$ gives
$$
\begin{array}{rl}\vspace{1ex}
& \left(\mu^{(1)}-i\mu^{(2)}\right)\cdot\displaystyle\int_{\Sigma\cap B}\left(A^{(1)}-A^{(2)}\right)e^{ix\cdot\xi}a_{1}\bar{a}_{2}\,dx\\ \vspace{1ex}
+&\left((\mu^{(1)})'-i(\mu^{(2)})',-(\mu^{(1)}-i\mu^{(2)})\right)\cdot\displaystyle\int_{\Sigma\cap B}\left(A^{(1)}-A^{(2)}\right)e^{i(x',-x_{n})\cdot\xi}\\ \vspace{1ex}
& a_{1}(x',-x_{n})\overline{a_{2}(x',-x_{n})}\,dx\rightarrow 0.
\end{array}
$$
Set $\tilde{A}^{(2)}=\overline{\tilde{A}^{(3)}}$. After a change of variable, this expression becomes
\begin{equation}\label{mu1}
\left(\mu^{(1)}-i\mu^{(2)}\right)\cdot\displaystyle\int_{(\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{0}}\left(\tilde{A}^{(1)}-\tilde{A}^{(2)}\right)e^{ix\cdot\xi}a_{1}\overline{a}_{2}\,dx=0
\end{equation}
for all $\xi,\mu^{(1)},\mu^{(2)}\in\mathbb{R}^{n}$ with
\begin{equation}\label{perp}
\mu^{(1)}\cdot\mu^{(2)}=\xi\cdot\mu^{(1)}=\xi\cdot\mu^{(2)}=0,\quad |\mu^{(1)}|=|\mu^{(2)}|=1, \quad \mu^{(2)}_{n}=0, \quad \mu^{(1)}_{n}\neq 0.
\end{equation}
Replacing $\mu^{(1)}$ by $-\mu^{(1)}$ to get
\begin{equation}\label{mu2}
\left(\mu^{(1)}+i\mu^{(2)}\right)\cdot\displaystyle\int_{(\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{0}}\left(\tilde{A}^{(1)}-\tilde{A}^{(2)}\right)e^{ix\cdot\xi}a_{1}\overline{a}_{2}\,dx=0.
\end{equation}
Hence, \eqref{mu1} and \eqref{mu2} imply that
\begin{equation}\label{mu3}
\mu\cdot\displaystyle\int_{(\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{0}}\left(\tilde{A}^{(1)}-\tilde{A}^{(2)}\right)e^{ix\cdot\xi}a_{1}\overline{a}_{2}\,dx=0.
\end{equation}
for all $\mu\in\textrm{span}\{\mu^{(1)},\mu^{(2)}\}$ and all $\xi\in\mathbb{R}^{n}$ for which \eqref{perp} holds.\\
Next proposition indicates that $d\tilde{A}^{(1)}=d\tilde{A}^{(2)}$ in $(\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{0}$.
\begin{prop}\label{curlvanish2}
\begin{equation} \label{curl2}
\partial_{j}(\tilde{A}^{(1)}_{k}-\tilde{A}^{(2)}_{k})-\partial_{k}(\tilde{A}^{(1)}_{j}-\tilde{A}^{(2)}_{j})=0 \textrm{ in } (\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{0}, \quad 1\leq j,k\leq n.
\end{equation}
\end{prop}
\begin{proof}
If $n=3$, for any vector $\xi\in\mathbb{R}^{3}$ with $\xi^{2}_{1}+\xi^{2}_{2}>0$, it is easy to see the vectors
$$
\begin{array}{rl}\vspace{1ex}
\mu^{(1)}= & \displaystyle\frac{\tilde{\mu}^{(1)}}{|\tilde{\mu}^{(1)}|}\quad\quad
\tilde{\mu}^{(1)}=(-\xi_{1}\xi_{3},-\xi_{2}\xi_{3},\xi^{2}_{1}+\xi^{2}_{2}), \\ \vspace{1ex}
\mu^{(2)}= & \left(\displaystyle\frac{-\xi_{2}}{\sqrt{\xi^{2}_{1}+\xi^{2}_{2}}},
\displaystyle\frac{\xi_{1}}{\sqrt{\xi^{2}_{1}+\xi^{2}_{2}}},0\right),\\
\end{array}
$$
satisfy \eqref{perp}. Thus, after choosing $a_{1}=a_{2}=1$, \eqref{mu3} gives
\begin{equation}\label{muperp}
\mu\cdot f(\xi)=0 \quad \textrm{ where }f(\xi):=\widehat{\tilde{A}^{(1)}\chi}_{(\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{0}}(\xi)-\widehat{\tilde{A}^{(2)}\chi}_{(\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{0}}(\xi)
\end{equation}
for all $\mu\in\textrm{span}\{\mu^{(1)},\mu^{(2)}\}$. Here $\chi_{(\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{0}}$ stands for the characteristic function of the set $(\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{0}$. We decompose $f(\xi)\in\mathbb{R}^{3}$ as
$$f(\xi)=\alpha(\xi)\xi+f_{\perp}(\xi)$$
where Re$\,\alpha(\xi)$, Im$\,\alpha(\xi)$ are real numbers, and Re$\,f_{\perp}(\xi)$, Im$\,f_{\perp}(\xi)$ are orthogonal to $\xi$. As $n=3$, we conclude that Re$\,f_{\perp}(\xi)$, Im$\,f_{\perp}(\xi)\in\textrm{span}\{\mu^{(1)},\mu^{(2)}\}$. It follows from \eqref{muperp} that $f_{\perp}(\xi)=0$ for all $\xi\in\mathbb{R}^{3}$ with $\xi^{2}_{1}+\xi^{2}_{2}>0$. Hence $f(\xi)=\alpha(\xi)\xi$. Choose $\mu=-\xi_{k}e_{j}+\xi_{j}e_{k},1\leq j,k\leq 3, j\neq k$, where $e_{j}$ is the standard orthonormal basis of $\mathbb{R}^{3}$. This choice of $\mu$ satisfies $\mu\cdot f(\xi)=0.$ Therefore,
$$\xi_{j}\cdot(\widehat{\tilde{A}^{(1)}_{k}\chi(\xi)}-\widehat{\tilde{A}^{(2)}_{k}\chi(\xi)})-
\xi_{k}\cdot(\widehat{\tilde{A}^{(1)}_{j}\chi(\xi)}-\widehat{\tilde{A}^{(2)}_{j}\chi(\xi)})=0$$
for all $\xi\in\mathbb{R}^{3}$ with $\xi^{2}_{1}+\xi^{2}_{2}>0$, and hence everywhere by analyticity of the Fourier transform.
If $n\geq 4$, for any vector $\xi=(\xi_{1},\cdots,\xi_{n})\in\mathbb{R}^{n},\xi_{l}\neq 0,l=1,\cdots,n$, define vectors
$$\mu^{(1)}=(-\xi_{j}\xi_{n})e_{j}+(-\xi_{k}\xi_{n})e_{k}+(\xi^{2}_{j}+\xi^{2}_{k})e_{n}, \quad \mu^{(2)}=-\xi_{k}e_{j}+\xi_{j}e_{k}$$
where $1\leq j,k<n, j\neq k.$ It is easy to check that $\mu^{(1)}\cdot\mu^{(2)}=\mu^{(1)}\cdot\xi=\mu^{(2)}\cdot\xi=0,\mu^{(2)}_{n}=0$ and $\mu^{(1)}_{n}\neq 0$. Thus, after choosing $a_{1}=a_{2}=1$ and $\mu=\mu^{(2)}$, \eqref{mu3} implies
\begin{equation}\label{vanish4}
\xi_{j}\cdot(\widehat{\tilde{A}^{(1)}_{k}\chi(\xi)}-\widehat{\tilde{A}^{(2)}_{k}\chi(\xi)})-
\xi_{k}\cdot(\widehat{\tilde{A}^{(1)}_{j}\chi(\xi)}-\widehat{\tilde{A}^{(2)}_{j}\chi(\xi)})=0 \quad 1\leq j,k<n,j\neq k
\end{equation}
for all $\xi\in\mathbb{R}^{n}$ with $\xi_{l}\neq0,l=1,2,\cdots,n$.
Let $\xi=(\xi_{1},\cdots,\xi_{n})\in\mathbb{R}^{n}$ with $\xi_{l}\neq 0,l=1,2,\cdots,n,$ and let $1\leq j<n.$ Choose indices $k$ and $l$ so that the set $\{j,k,l,n\}$ consists of four distinct numbers. Define
$$\mu^{(1)}=-\xi_{n}e_{j}+\xi_{j}e_{n},\quad\quad \mu^{(2)}=-\xi_{k}e_{l}+\xi_{l}e_{k}.$$
Again one can check that $\mu^{(1)}\cdot\mu^{(2)}=\mu^{(1)}\cdot\xi=\mu^{(2)}\cdot\xi=0,\mu^{(2)}_{n}=0$ and $\mu^{(1)}_{n}\neq 0$. After choosing $a_{1}=a_{2}=1$ and $\mu=\mu^{(1)}$, \eqref{mu3} implies
\begin{equation}\label{vanish5}
\xi_{j}\cdot(\widehat{\tilde{A}^{(1)}_{n}\chi(\xi)}-\widehat{\tilde{A}^{(2)}_{n}\chi(\xi)})-
\xi_{n}\cdot(\widehat{\tilde{A}^{(1)}_{j}\chi(\xi)}-\widehat{\tilde{A}^{(2)}_{j}\chi(\xi)})=0 \quad 1\leq j<n
\end{equation}
for all $\xi\in\mathbb{R}^{n}$ with $\xi_{l}\neq0,l=1,2,\cdots,n$. The result in the case $n\geq 4$ then follows from \eqref{vanish4} and \eqref{vanish5}.
\end{proof}\vspace{1ex}
Arguing as in the proof of Theorem 1.1, we can find compactly supported function $\Phi\in C^{1,1}(\overline{\Sigma\cup\Sigma^{\ast}_{0}})$ such that
$$\tilde{A}^{(1)}-\tilde{A}^{(2)}=\nabla\Phi \quad\quad \textrm{ in } (\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{0}$$
and $\Phi=0$ on $\partial((\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{0})$. In \eqref{mu3}, pick $a_{2}=1$, $a_{1}$ satisfying
$$((i\mu^{(1)}+\mu^{(2)})\cdot\nabla)a_{1}=1 \quad\quad \textrm{ in } (\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{0}.$$
and $\mu=i\mu^{(1)}+\mu^{(2)}$. Integrating by parts we obtain
$$
\begin{array}{rl}\vspace{1ex}
0= & (i\mu^{(1)}+\mu^{(2)})\cdot\displaystyle\int_{(\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{0}}(\nabla\Phi)e^{ix\cdot\xi}a_{1}\,dx\\ \vspace{1ex}
=& \displaystyle\int_{(\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{0}}\Phi(x)e^{ix\cdot\xi}[(i\mu^{(1)}+\mu^{(2)})\cdot\nabla a_{1}]\,dx\\ \vspace{1ex}
=& \displaystyle\int_{(\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{0}}\Phi(x)e^{ix\cdot\xi}\,dx.
\end{array}
$$
This implies that $\Phi=0$ in $(\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{0}$. Hence $\tilde{A}^{(1)}=\tilde{A}^{(2)}$, and therefore, $A^{(1)}=A^{(2)}$ in $\Sigma\cap B$.
As for electric potentials $q^{(1)}$ and $q^{(2)}$, continuing to argue as in the proof of Theorem 1.1, we arrive at
\begin{equation}\label{vanish6}
\displaystyle\int_{(\Sigma\cap B)\cup(\Sigma\cap B)^{\ast}_{0}}(q^{(1)}-q^{(2)})e^{ix\cdot\xi}\,dx=0
\end{equation}
for all $\mu^{(1)},\mu^{(2)},\xi\in\mathbb{R}^{n}$ satisfying \eqref{perp}. For any vector $\xi\in\mathbb{R}^{n}$ with $\xi^{2}_{n-2}+\xi^{2}_{n-1}>0$, the vectors
$$
\begin{array}{rl}\vspace{1ex}
\mu^{(1)}= & \displaystyle\frac{\tilde{\mu}^{(1)}}{|\tilde{\mu}^{(1)}|}\quad\quad
\tilde{\mu}^{(1)}=\left(0_{\mathbb{R}^{n-3}},\displaystyle\frac{-\xi_{n}\xi_{n-2}}{\sqrt{\xi^{2}_{n-2}+\xi^{2}_{n-1}}},
\displaystyle\frac{-\xi_{n}\xi_{n-1}}{\sqrt{\xi^{2}_{n-2}+\xi^{2}_{n-1}}},\displaystyle\sqrt{\xi^{2}_{n-2}+\xi^{2}_{n-1}}
\right),\\ \vspace{1ex}
\mu^{(2)}= & \left(0_{\mathbb{R}^{n-3}},\displaystyle\frac{-\xi_{n-1}}{\sqrt{\xi^{2}_{n-2}+\xi^{2}_{n-1}}},
\displaystyle\frac{\xi_{n-2}}{\sqrt{\xi^{2}_{n-2}+\xi^{2}_{n-1}}},0\right), \\
\end{array}
$$
satisfy \eqref{mu3}. Thus, \eqref{vanish6} holds for all $\xi\in\mathbb{R}^{n}$ with $\xi^{2}_{n-2}+\xi^{2}_{n-1}>0$. We conclude that \eqref{vanish6} also holds for all $\xi\in\mathbb{R}^{n}$ by the analyticity of the Fourier transform. This completes the proof of Theorem 1.2.\vspace{3ex}
\section{Proof of Theorem 1.3}
In this section we prove Theorem 1.3. Let $\Omega_{1}\subset\subset\Omega$ be a bounded sub-domain with $C^{\infty}$ boundary and be such that $\Omega\backslash\bar{\Omega}_{1}$ is connected and $supp(A^{(1)}-A^{(2)})$ and $supp(q^{(1)}-q^{(2)})$ are contained in $\Omega_{1}$.
Let $u_{1}\in H^{4}(\Omega)$ be the solution to the Dirichlet problem
$$\left\{
\begin{array}{rcll}
\mathcal{L}_{A^{(1)},q^{(1)}}u_{1} &=& 0 & \quad\textrm{ in } \Omega \\
u_{1}&=&f_{1} & \quad\textrm{ on } \partial\Omega \\
\Delta u_{1}&=&f_{2} & \quad\textrm{ on } \partial\Omega \\
\end{array}
\right.$$
with $(f_{1},f_{2})\in H^{\frac{7}{2}}(\partial\Omega)\times H^{\frac{3}{2}}(\partial\Omega)$ and $supp(f_{1})\subset\gamma_{1}, supp(f_{2})\subset\gamma_{1}$. Let $u_{2}\in H^{4}(\Omega)$ be the solution to the Dirichlet problem
$$\left\{
\begin{array}{rcll}
\mathcal{L}_{A^{(2)},q^{(2)}}u_{2} &=& 0 & \quad\textrm{ in } \Omega \\
u_{2}&=&f_{1} & \quad\textrm{ on } \partial\Omega \\
\Delta u_{2}&=&f_{2} & \quad\textrm{ on } \partial\Omega. \\
\end{array}
\right.$$
Setting $w=u_{2}-u_{1}$, then
$$\mathcal{L}_{A^{(2)},q^{(2)}}w=(A^{(1)}-A^{(2)})\cdot Du_{1}+(q^{(1)}-q^{(2)})u_{1} \textrm{ in } \Omega.$$
As two Dirichlet-to-Neumann maps agree on $\gamma_{2}$, we have $\partial_{\nu}w=0$ on $\gamma_{2}$. Therefore, $w$ is a solution of
$$\mathcal{L}_{A^{(2)},q^{(2)}}w=0 \quad \textrm{ in } \Omega\backslash\bar{\Omega}_{1}$$
with $w=\partial_{\nu}w=0$ on $\gamma_{2}$. By unique continuation, we obtain that $w=0$ in $\Omega\backslash\bar{\Omega}_{1}$. Thus, $w=\Delta w=\partial_{\nu}w=\partial_{\nu}\Delta w=0$ on $\partial\Omega_{1}$.
Let $v\in H^{4}(\Omega_{1})$ be a solution of
\begin{equation}\label{equation}
\mathcal{L}^{\ast}_{A^{(2)},q^{(2)}}v=0 \quad\textrm{ in } \Omega_{1}
\end{equation}
Using Green's formula \eqref{green} over $\Omega_{1}$, we have
\begin{equation}\label{difference2}
\displaystyle\int_{\Omega_{1}}((A^{(1)}-A^{(2)})\cdot Du_{1})\bar{v}\,dx+\int_{\Omega_{1}}(q^{(1)}-q^{(2)})u_{1}\bar{v}\,dx=0
\end{equation}
for all $v\in H^{4}(\Omega_{1})$ satisfying \eqref{equation} and for all $u_{1}\in \mathcal{W}(\Omega)$, where
$$\mathcal{W}(\Omega):=\{u\in H^{4}(\Omega):\mathcal{L}_{A^{(1)},q^{(1)}}u=0 \textrm{ in }\Omega, supp(u|_{\partial\Omega})\subset\gamma_{1},supp(\Delta u|_{\partial\Omega})\subset\gamma_{1}\}.$$
Let
$$\widetilde{\mathcal{W}}(\Omega_{1}):=\{u\in H^{4}(\Omega_{1}):\mathcal{L}_{A^{(1)},q^{(1)}}u=0 \textrm{ in }\Omega_{1}\}.$$
Again we need a density result to pass from $\mathcal{W}(\Omega)$ to $\widetilde{\mathcal{W}}(\Omega_{1})$.
\begin{prop}
$\mathcal{W}(\Omega)$ is a dense subspace in $\widetilde{\mathcal{W}}(\Omega_{1})$ in $L^{2}(\Omega_{1})$-topology.
\end{prop}
\begin{proof}
It suffices to establish the following fact: for any $g\in L^{2}(\Omega_{1})$ such that
$$\displaystyle\int_{\Omega_{1}} u\overline{g}\,dx=0 \quad\quad \forall u\in \mathcal{W}(\Omega),$$
we have
$$\displaystyle\int_{\Omega_{1}} v\overline{g}\,dx=0 \quad\quad \forall v\in \widetilde{\mathcal{W}}(\Omega).$$
To this end, extend $g$ by zero to $\Omega\backslash\Omega_{1}$. Let $U\in H^{4}(\Omega)$ be the solution of the Dirichlet problem
$$
\begin{array}{rl} \vspace{1ex}
\mathcal{L}^{\ast}_{A^{(1)},q^{(1)}}U=g & \quad \textrm{ in } \Omega\\ \vspace{1ex}
U=\Delta U=0 & \quad \textrm{ on } \partial\Omega.\\
\end{array}
$$
For any $u\in \mathcal{W}(\Omega)$, Green's formula on bounded domain $\Omega$ gives
$$
\begin{array}{rl}\vspace{1ex}
0=&\displaystyle\int_{\Omega}u\bar{g}\,dx=\displaystyle\int_{\Omega}u\overline{(\mathcal{L}^{\ast}_{A^{(1)},q^{(1)}}U)}\,dx \\ \vspace{1ex}
=&-\displaystyle\int_{\partial\Omega}(-\Delta u)\overline{\partial_{\nu}U}\,dS-\displaystyle\int_{\partial\Omega}u\overline{(\partial_{\nu}(-\Delta U))}\,dS\\
\end{array}
$$
where we have used $U=\Delta U=0$ on $\partial\Omega$. Since $u|_{\gamma_{1}}$ and $\Delta u|_{\gamma_{1}}$ can be arbitrary smooth functions supported in $\gamma_{1}$, we conclude that $\partial_{\nu}U|_{\gamma_{1}}=\partial_{\nu}\Delta U|_{\gamma_{1}}=0$. Hence $U$ satisfies $\mathcal{L}^{\ast}_{A^{(1)},q^{(1)}}U=0$ in $\Omega\backslash\Omega_{1}$, and $U=\Delta U=\partial_{\nu}U=\partial_{\nu}(\Delta U)=0$ on $\gamma_{1}$. By unique continuation, $U=0$ in $\Omega\backslash\Omega_{1}$, and therefore, $U=\Delta U=\partial_{\nu} U=\partial_{\nu}(\Delta U)=0$ on $\partial\Omega_{1}$.
For any $v\in \widetilde{\mathcal{W}}(\Omega_{1})$, using Green's formula over $\Omega_{1}$ we get
$$
\begin{array}{rl}\vspace{1ex}
\displaystyle\int_{\Omega_{1}}v\overline{g}\,dx= & \displaystyle\int_{\Omega_{1}}v\overline{(\mathcal{L}^{\ast}_{A^{(1)},q^{(1)}}U)}\,dx \\ \vspace{1ex}
=& \displaystyle\int_{\Omega_{1}}\left(\mathcal{L}_{A^{(1)},q^{(1)}}v\right)\overline{U}\,dx+i\int_{\partial\Omega_{1}}\nu(x)\cdot A\overline{U}v\,dS\\ \vspace{1ex}
& +\displaystyle\int_{\partial\Omega_{1}}\partial_{\nu}(-\Delta v)\overline{U}\,dS-\displaystyle\int_{\partial\Omega_{1}}(-\Delta v)\overline{\partial_{\nu}U}\,dS\\ \vspace{1ex}
& +\displaystyle\int_{\partial\Omega_{1}}\partial_{\nu}v\overline{(-\Delta U)}\,dS-\displaystyle\int_{\partial\Omega_{1}}v\overline{(\partial_{\nu}(-\Delta U))}\,dS\\ \vspace{1ex}
=& 0.\\
\end{array}
$$
\end{proof}
We conclude from this proposition that \eqref{difference2} holds for all $u\in\widetilde{\mathcal{W}}(\Omega_{1})$ and $v\in H^{4}(\Omega_{1})$ satisfying \eqref{equation}.
Let $B\subset\mathbb{R}^{n}$ be an open ball such that $\Omega_{1}\subset B$. The fact that $A^{(1)}=A^{(2)}$ and $q^{(1)}=q^{(2)}$ on $\partial\Omega_{1}$ allows to extend $A^{(j)}$ and $q^{(j)}$ to $B$ in such a way that the extensions, still denoted by $A^{(j)}$ and $q^{(j)}$, coincide on $B\backslash\Omega_{1}$, have compact supports, and satisfy $A^{(j)}\in W^{1,\infty}(B)$, $q^{(j)}\in L^{\infty}(B)$. It follows from \eqref{difference2} that
$$\displaystyle\int_{B}((A^{(1)}-A^{(2)})\cdot Du_{1})\bar{v}\,dx+\int_{B}(q^{(1)}-q^{(2)})u_{1}\bar{v}\,dx=0$$
for all $u_{1},v\in H^{4}(B)$ which are solutions of
$$\mathcal{L}_{A^{(1)},q^{(1)}}u_{1}=0 \textrm{ in } B \quad\quad\quad \mathcal{L}^{\ast}_{A^{(2)},q^{(2)}}v=0 \textrm{ in } B.$$
Now we are in the same situation as in \cite{KLU2} for the bi-harmonic operator, and as in \cite{KLU3} with full boundary measurements. We can construct complex geometric optics solutions as in Proposition \ref{existence}, and proceed as in \cite{KLU2}, \cite{KLU3} and the proof of Theorem 1.1 to show that $A^{(1)}=A^{(2)}$ and $q^{(1)}=q^{(2)}$ in $\Omega$.
\section{Proof of Theorem 1.4}
In this section we prove Theorem 1.4. First, as in the proof of Theorem 1.1 and Theorem 1.2, after applying Green's formula over $\Omega$, we obtain the integral identity
\begin{equation}\label{difference3}
\displaystyle\int_{\Omega}((A^{(1)}-A^{(2)})\cdot Du_{1})\bar{v}\,dx+\int_{\Omega}(q^{(1)}-q^{(2)})u_{1}\bar{v}\,dx=0
\end{equation}
for all $u_{1},v\in H^{4}(\Omega)$ such that
$$\mathcal{L}_{A^{(1)},q^{(1)}}u_{1}=0 \textrm{ in }\Omega,\quad\quad u_{1}|_{x_{n}=0}=(\Delta u_{1})|_{x_{n}=0}=0;$$
$$\mathcal{L}^{\ast}_{A^{(2)},q^{(2)}}v=0 \textrm{ in }\Omega,\quad\quad v|_{x_{n}=0}=(\Delta v)|_{x_{n}=0}=0.$$
Applying the reflection argument as in the proof of Theorem 1.2, we can construct CGO solutions $u_{1}$ and $v$, as in \eqref{CGO1} and \eqref{CGO3}, to the above equations and with the corresponding boundary conditions. Substituting these solutions $u_{1}$ and $v$ into \eqref{difference3} and proceeding as in the proof of Theorem 1.2 we get
\begin{equation}\label{mu4}
\left(\mu^{(1)}-i\mu^{(2)}\right)\cdot\displaystyle\int_{\Omega\cup\Omega^{\ast}_{0}}\left(\tilde{A}^{(1)}-\tilde{A}^{(2)}\right)e^{ix\cdot\xi}a_{1}\bar{a}_{2}\,dx=0
\end{equation}
for all $\xi,\mu^{(1)},\mu^{(2)}\in\mathbb{R}^{n}$ such that
$$\mu^{(1)}\cdot\mu^{(2)}=\xi\cdot\mu^{(1)}=\xi\cdot\mu^{(2)}=0,\quad |\mu^{(1)}|=|\mu^{(2)}|=1, \quad \mu^{(2)}_{n}=0, \quad \mu^{(1)}_{n}\neq 0,$$
where we have introduced the notation $\Omega^{\ast}_{0}:=\{(x',x_{n})\in\mathbb{R}^{n}:(x',-x_{n})\in\Omega\}$.
Applying the boundary reconstruction result \cite[Proposition 4.1]{KLU2} we conclude that $A^{(1)}=A^{(2)}$ on $\bar{\gamma}$, hence $\tilde{A}^{(1)}=\tilde{A}^{(2)}$ on $\partial(\Omega\cup\Omega^{\ast}_{0})$. This allows us to extend $\tilde{A}^{(j)},\; j=1,2,$ to compactly supported vector fields on a large ball $B$ with $\Omega\cup\Omega^{\ast}_{0}\subset\subset B$ and $\tilde{A}^{(1)}=\tilde{A}^{(2)}$ in $B\backslash\Omega\cup\Omega^{\ast}_{0}$. Then \eqref{mu4} leads to
$$\left(\mu^{(1)}-i\mu^{(2)}\right)\cdot\displaystyle\int_{B}\left(\tilde{A}^{(1)}-\tilde{A}^{(2)}\right)e^{ix\cdot\xi}a_{1}\bar{a}_{2}\,dx=0.$$
From Proposition \ref{curlvanish2} we have $d\tilde{A}^{(1)}=d\tilde{A}^{(2)}$ in $B$. Therefore, there exists $\Phi\in C^{1,1}(\overline{B})$ so that
$$\tilde{A}^{(1)}-\tilde{A}^{(2)}=\nabla\Phi \quad\quad\textrm{ in } B.$$
As before we can show that $\Phi=0$ on $\partial(\Omega\cup\Omega^{\ast}_{0})$; in particular, $\Phi=0$ on $\bar{\gamma}$. Now we are facing the same situation as in the proof of Theorem 1.2. Arguing as there we conclude that $A^{(1)}=A^{(2)}$ and $q^{(1)}=q^{(2)}$. This completes the proof of Theorem 1.4.
\begin{appendices}
\section{Solvability of the forward problem in an infinite slab}
In this appendix we provide the proof of the existence of the forward boundary value problem \eqref{Dirichlet1} for the perturbed bi-harmonic operator in an infinite slab. Recall that the perturbed bi-harmonic operator is of the form
$$\mathcal{L}_{A,q}(x,D):=\Delta^{2}+A(x)\cdot D+q(x).$$
The infinite slab is written as $(n\geq 3)$
$$\Sigma=\{x=(x',x_{n})\in\mathbb{R}^{n}:x'=(x_{1},\dots,x_{n-1})\in\mathbb{R}^{n-1}, 0<x_{n}<L\},\quad L>0.$$
whose boundary hyperplanes are
$$\Gamma_{1}=\{x\in\mathbb{R}^{n}:x_{n}=L\} \quad\quad \Gamma_{2}=\{x\in\mathbb{R}^{n}:x_{n}=0\}.$$
We will rewrite the perturbed bi-harmonic equation as a system of equations. For this purpose, let $u=(u_{1},u_{2})$ with $u_{2}=\Delta u_{1}$, define
$$\mathcal{S}u:=
\Delta \left(
\begin{array}{c}
u_{1} \\
u_{2} \\
\end{array}
\right)+
\left(
\begin{array}{cc}
0 & -1 \\
A\cdot D+q & 0 \\
\end{array}
\right)
\left(
\begin{array}{c}
u_{1} \\
u_{2} \\
\end{array}
\right),
$$
then $\mathcal{L}_{A,q}u_{1}=0$ is equivalent to $\mathcal{S}u=0$. We will show the existence of a unique solution to this system with boundary value $(u_{1},u_{2})|_{\partial\Sigma}=(f_{1},f_{2})$.
Poincar\'{e}'s inequality in an infinite slab indicates that the quadratic form
$$u\mapsto \displaystyle\int_{\Sigma}|\nabla u|^{2}\,dx=\displaystyle\int_{\Sigma}(|\nabla u_{1}|^{2}+|\nabla u_{2}|^{2})\,dx$$
is non-negative and densely defined closed on $H^{1}_{0}(\Sigma)\times H^{1}_{0}(\Sigma)$. Associated with this quadratic form, the Laplace operator $-\Delta$ equipped with the domain
$$\mathcal{D}(-\Delta):=\{u\in H^{1}_{0}(\Sigma)\times H^{1}_{0}(\Sigma):\Delta u=(\Delta u_{1},\Delta u_{2})\in L^{2}(\Sigma)\times L^{2}(\Sigma)\}$$
is a non-negative self-adjoint operator on $L^{2}(\Sigma)\times L^{2}(\Sigma)$. Its spectrum is obtained in the following proposition.
\begin{prop}
$\mathcal{D}(-\Delta)=H^{1}_{0}(\Sigma)\cap H^{2}(\Sigma) \times H^{1}_{0}(\Sigma)\cap H^{2}(\Sigma)$. Moreover, the spectrum of $-\Delta$ is purely absolutely continuous and is equal to $[\pi^{2}/L^{2},+\infty)$.
\end{prop}
\begin{proof}
Let $F=(F_{1},F_{2})\in L^{2}(\Sigma)\times L^{2}(\Sigma)$, we will consider
$$-\Delta u=F, \quad\quad u\in\mathcal{D}(-\Delta).$$
Taking the Fourier series with respect to the variable $x_{n}\in [0,L]$ we have
\begin{equation}\label{Fourier}
\begin{array}{lr}
u(x',x_{n})=\displaystyle\sum^{\infty}_{l=1}u_{l}(x')\sin\frac{l\pi x_{n}}{L}, &\quad u_{l}(x')=\displaystyle\frac{2}{L}\displaystyle\int^{L}_{0}u(x)\sin\frac{l\pi x_{n}}{L}\,dx_{n};\\
F(x',x_{n})=\displaystyle\sum^{\infty}_{l=1}F_{l}(x')\sin\frac{l\pi x_{n}}{L}, &\quad F_{l}(x')=\displaystyle\frac{2}{L}\displaystyle\int^{L}_{0}F(x)\sin\frac{l\pi x_{n}}{L}\,dx_{n}.\\
\end{array}
\end{equation}
As usual Parseval's identities hold
$$
\begin{array}{rl}
\|u\|_{L^{2}(\Sigma)\times L^{2}(\Sigma)} & =\displaystyle\frac{L}{2}\displaystyle\sum^{\infty}_{l=1}\|u_{l}\|^{2}_{L^{2}(\mathbb{R}^{n-1})\times L^{2}(\mathbb{R}^{n-1})},\\
\|F\|_{L^{2}(\Sigma)\times L^{2}(\Sigma)} & =\displaystyle\frac{L}{2}\displaystyle\sum^{\infty}_{l=1}\|F_{l}\|^{2}_{L^{2}(\mathbb{R}^{n-1})\times L^{2}(\mathbb{R}^{n-1})}.\\
\end{array}
$$
Comparing the Fourier coefficients $u_{l}$ of $u$ and $F_{l}$ of $F$ we see that they are related by
\begin{equation}\label{Fourieridentity}
\left(-\Delta_{x'}+\displaystyle\frac{l^{2}\pi^{2}}{L^{2}}\right)u_{l}(x')=F_{l}(x'), \quad\quad x'\in\mathbb{R}^{n-1}, l=1,2,\dots.
\end{equation}
The operator $-\Delta_{x'}+\frac{l^{2}\pi^{2}}{L^{2}}$ $(l\geq 1)$, when equipped with the domain $H^{2}(\mathbb{R}^{n-1})$, is self-adjoint on $L^{2}(\mathbb{R}^{n-1})$ with purely absolutely continuous spectrum $[l^{2}\pi^{2}/L^{2},+\infty)$. Hence \eqref{Fourieridentity} has the unique solution
$$u_{l}(x')=\left(-\Delta_{x'}+\displaystyle\frac{l^{2}\pi^{2}}{L^{2}}\right)^{-1}F_{l}(x')\in H^{2}(\mathbb{R}^{n-1}),$$
and moreover, it satisfies the norm estimate
\begin{equation}\label{normestimate1}
\begin{array}{rl}\vspace{1ex}
\|u_{l}\|_{L^{2}(\mathbb{R}^{n-1})\times L^{2}(\mathbb{R}^{n-1})} & \leq\displaystyle\frac{L^{2}}{l^{2}\pi^{2}}\|F_{l}\|_{L^{2}(\mathbb{R}^{n-1})\times L^{2}(\mathbb{R}^{n-1})};\\\vspace{1ex}
\|u_{l}\|_{H^{2}(\mathbb{R}^{n-1})\times H^{2}(\mathbb{R}^{n-1})} & \leq C\|F_{l}\|_{L^{2}(\mathbb{R}^{n-1})\times L^{2}(\mathbb{R}^{n-1})}.\\
\end{array}
\end{equation}
Here and in the following we will name all the constants independent of $l$ as C. By interpolation we obtain
\begin{equation}\label{normestimate2}
\|u_{l}\|_{H^{1}(\mathbb{R}^{n-1})\times H^{1}(\mathbb{R}^{n-1})} \leq \displaystyle\frac{C}{l}\|F_{l}\|_{L^{2}(\mathbb{R}^{n-1})\times L^{2}(\mathbb{R}^{n-1})}.
\end{equation}
Parseval's identities and \eqref{normestimate1} then give
$$
\begin{array}{rl}\vspace{1ex}
\|u\|^{2}_{L^{2}(\Sigma)\times L^{2}(\Sigma)} = & \displaystyle\frac{L}{2}\displaystyle\sum^{\infty}_{l=1}\|u_{l}\|^{2}_{L^{2}(\mathbb{R}^{n-1})\times L^{2}(\mathbb{R}^{n-1})}\\\vspace{1ex}
\leq & C \displaystyle\sum^{\infty}_{l=1}\frac{1}{l^{4}}\|F_{l}\|^{2}_{L^{2}(\mathbb{R}^{n-1})\times L^{2}(\mathbb{R}^{n-1})}\leq C \|F\|^{2}_{L^{2}(\Sigma)\times L^{2}(\Sigma)}.\\
\end{array}
$$
To take care of the first order derivatives, we differentiate with respect to $x_{n}$ to get
$$
\begin{array}{rl}\vspace{1ex}
\|\partial_{x_{n}}u\|^{2}_{L^{2}(\Sigma)\times L^{2}(\Sigma)} = &
\|\displaystyle\sum^{\infty}_{l=1}\displaystyle\frac{l\pi}{L}u_{l}(x')\cos\frac{l\pi x_{n}}{L}\|^{2}_{L^{2}(\Sigma)\times L^{2}(\Sigma)}\\\vspace{1ex}
= & \displaystyle\frac{L}{2}\displaystyle\sum^{\infty}_{l=1}\frac{l^{2}\pi^{2}}{L^{2}}\|u_{l}\|^{2}_{L^{2}(\mathbb{R}^{n-1})\times L^{2}(\mathbb{R}^{n-1})}\leq C \|F\|^{2}_{L^{2}(\Sigma)\times L^{2}(\Sigma)}.\\
\end{array}
$$
Using \eqref{normestimate2} we obtain that for $j=1,2,\cdots,n-1$,
$$
\begin{array}{rl}\vspace{1ex}
\|\partial_{x_{j}}u\|^{2}_{L^{2}(\Sigma)\times L^{2}(\Sigma)} = & \displaystyle\frac{L}{2}\displaystyle\sum^{\infty}_{l=1}\|\partial_{x_{j}}u_{l}\|^{2}_{L^{2}(\mathbb{R}^{n-1})\times L^{2}(\mathbb{R}^{n-1})}\\\vspace{1ex}
\leq & C \displaystyle\sum^{\infty}_{l=1}\frac{1}{l^{2}}\|F_{l}\|^{2}_{L^{2}(\mathbb{R}^{n-1})\times L^{2}(\mathbb{R}^{n-1})}\leq C \|F\|^{2}_{L^{2}(\Sigma)\times L^{2}(\Sigma)}.\\
\end{array}
$$
We proceed to estimate the second order derivatives. For $j,k=1,2,\cdots,n-1$, it follows from \eqref{normestimate1} that
$$
\begin{array}{rl}\vspace{1ex}
\|\partial^{2}_{x_{j}x_{k}}u\|^{2}_{L^{2}(\Sigma)\times L^{2}(\Sigma)} = & \displaystyle\frac{L}{2}\displaystyle\sum^{\infty}_{l=1}\|\partial_{x_{j}x_{k}}u_{l}\|^{2}_{L^{2}(\mathbb{R}^{n-1})\times L^{2}(\mathbb{R}^{n-1})}\leq C \|F\|^{2}_{L^{2}(\Sigma)\times L^{2}(\Sigma)};\\ \vspace{1ex}
\|\partial^{2}_{x_{j}x_{n}}u\|^{2}_{L^{2}(\Sigma)\times L^{2}(\Sigma)} = & \displaystyle\frac{L}{2}\displaystyle\sum^{\infty}_{l=1}\frac{l^{2}\pi^{2}}{L^{2}}\|\partial_{x_{j}}u_{l}\|^{2}_{L^{2}(\mathbb{R}^{n-1})\times L^{2}(\mathbb{R}^{n-1})}\leq C \|F\|^{2}_{L^{2}(\Sigma)\times L^{2}(\Sigma)};\\ \vspace{1ex}
\|\partial^{2}_{x_{n}}u\|^{2}_{L^{2}(\Sigma)\times L^{2}(\Sigma)} = & \displaystyle\frac{L}{2}\displaystyle\sum^{\infty}_{l=1}\frac{l^{4}\pi^{4}}{L^{4}}\|u_{l}\|^{2}_{L^{2}(\mathbb{R}^{n-1})\times L^{2}(\mathbb{R}^{n-1})}\leq C \|F\|^{2}_{L^{2}(\Sigma)\times L^{2}(\Sigma)}.\\
\end{array}
$$
These estimates show that $u\in H^{2}(\Sigma)\times H^{2}(\Sigma)$. The statement concerning the spectrum of $-\Delta$ follows from the fact that
$$-\Delta=\bigoplus^{\infty}_{l=1}\left(-\Delta_{x'}+\displaystyle\frac{l^{2}\pi^{2}}{L^{2}}\right).$$
This completes the proof of the proposition.
\end{proof}
\begin{prop}
Let $A\in W^{1,\infty}({\Sigma};\mathbb{C}^{n})\cap\mathcal{E}'(\bar{\Sigma};\mathbb{C}^{n})$, $q\in L^{\infty}(\Sigma;\mathbb{C})\cap\mathcal{E}'(\bar{\Sigma};\mathbb{C}^{n}).$ Then the operator $\mathcal{S}$, equipped with the domain $H^{1}_{0}(\Sigma)\cap H^{2}(\Sigma) \times H^{1}_{0}(\Sigma)\cap H^{2}(\Sigma)$, is closed and its essential spectrum is equal to $[\pi^{2}/L^{2},+\infty)$.
\end{prop}
\begin{proof}
This follows from the fact that
$$
\left(
\begin{array}{cc}
0 & -1 \\
A\cdot D+q & 0 \\
\end{array}
\right)\Delta^{-1}:L^{2}(\Sigma)\times L^{2}(\Sigma)\rightarrow L^{2}(\Sigma)\times L^{2}(\Sigma)
$$
is a compact operator and that the essential spectrum do not change under relatively compact perturbations.
\end{proof}
This proposition yields the following solvability result. Suppose $A\in W^{1,\infty}({\Sigma};\mathbb{C}^{n})\cap\mathcal{E}'(\bar{\Sigma};\mathbb{C}^{n})$ and $q\in L^{\infty}(\Sigma;\mathbb{C})\cap\mathcal{E}'(\bar{\Sigma};\mathbb{C}^{n})$, then for any $F=(F_{1},F_{2})\in L^{2}(\Sigma)\times L^{2}(\Sigma)$, the boundary value problem
\begin{equation}\label{Dirichlet3}
\left\{
\begin{array}{rll}\vspace{1ex}
\mathcal{S}u= & F &\quad\textrm{ in } \Sigma\times\Sigma \\ \vspace{1ex}
u|_{\partial\Sigma\times\partial\Sigma}= & 0. & \\
\end{array}
\right.
\end{equation}
admits a unique solution $u\in H^{2}(\Sigma)\times H^{2}(\Sigma)$.
Given any $f=(f_{1},f_{2})\in (H^{\frac{7}{2}}(\Gamma_{1})\cap\mathcal{E}'(\Gamma_{1}))\times (H^{\frac{3}{2}}(\Gamma_{1})\cap\mathcal{E}'(\Gamma_{1}))$, we can establish the existence and uniqueness of the solution to the boundary value problem \eqref{Dirichlet1} as follows. \eqref{Dirichlet1} is equivalent to the following boundary value problem for the system
\begin{equation}\label{Dirichlet4}
\left\{
\begin{array}{rll}\vspace{1ex}
\mathcal{S}u= & 0 &\quad\textrm{ in } \Sigma\times\Sigma \\ \vspace{1ex}
u|_{\Gamma_{1}\times\Gamma_{1}}= & f & \\ \vspace{1ex}
u|_{\Gamma_{2}\times\Gamma_{2}}= & 0. & \\
\end{array}
\right.
\end{equation}
Uniqueness of the solution to \eqref{Dirichlet4} follows from the unique solvability of \eqref{Dirichlet3} when $F=0$. To show that \eqref{Dirichlet4} has at least one solution, choose $G\in H^{4}(\Sigma)\cap\mathcal{E}'(\overline{\Sigma})\times H^{2}(\Sigma)\cap\mathcal{E}'(\overline{\Sigma})$ so that $G|_{\Gamma_{1}\times\Gamma_{1}}=f$ and $G|_{\Gamma_{2}\times\Gamma_{2}}=0$; choose $u_{0}$ to be the unique solution of \eqref{Dirichlet3} when $F=-\mathcal{S}G$, then $G+u_{0}$ is a solution for \eqref{Dirichlet4}. This completes the proof that \eqref{Dirichlet1} admits a unique solution in $H^{4}(\Sigma)$ for $f_{1}$ and $f_{2}$.
\section{Green's formula in a slab}
In the proof of Proposition \ref{runge} we used the Green's formula in a slab, in this part we establish this identity.
For $R>0$, define $\Sigma_{R}$ by
$$\Sigma_{R}:=\{x\in\Sigma:|x'|<R\}.$$
We may choose $R>0$ sufficiently large so that $supp(A^{(1)})\subset\overline{\Sigma_{R}}$. Introduce the notations
$$d_{j}(R):=\partial\Sigma_{R}\cap\Gamma_{j}, j=1,2;\quad\quad d_{3}(R)=\partial\Sigma_{R}\cap\Sigma,$$
then $A^{(1)}=0$ on $d_{3}(R)$. Let $u\in W(\Sigma)$ and let $U\in H^{4}(\Sigma)$ be the solution of the problem
$$
\begin{array}{rl} \vspace{1ex}
\mathcal{L}^{\ast}_{A^{(1)},q^{(1)}}U=g & \quad \textrm{ in } \Sigma\\ \vspace{1ex}
U=\Delta U=0 & \quad \textrm{ on } \Gamma_{1}\cup\Gamma_{2}.\\
\end{array}
$$
Apply Green's formula \eqref{green} over the region $\Sigma_{R}$ we obtain
\begin{equation}\label{greenslab}
\begin{array}{rl}\vspace{1ex}
& -\displaystyle\int_{\Sigma_{R}}ug\,dx\\ \vspace{1ex}
=& \displaystyle\int_{\Sigma_{R}}(\mathcal{L}_{A^{(1)},q^{(1)}}u)\overline{U}\,dx-\int_{\Sigma_{R}}u\overline{(\mathcal{L}^{\ast}_{A^{(1)},q^{(1)}}U)}\,dx\\ \vspace{1ex}
=& -\displaystyle\int_{d_{3}(R)}\partial_{\nu}(-\Delta u)\overline{U}\,dS+\int_{d_{1}(R)\cup d_{3}(R)}(-\Delta u)\overline{\partial_{\nu}U}\,dS\\ \vspace{1ex}
& -\displaystyle\int_{d_{3}(R)}\partial_{\nu}u\overline{(-\Delta U)}\,dS+\int_{d_{1}(R)\cup d_{3}(R)}u\overline{\partial_{\nu}(-\Delta U)}\,dS.\\
\end{array}
\end{equation}
We will show that the right hand side converges to
\begin{equation}\label{righthandside}
\int_{\Gamma_{1}}\overline{\partial_{\nu}U}\Delta u\,dS+\int_{\Gamma_{1}}\overline{\partial_{\nu}\Delta U} u\,dS.
\end{equation}
To this end, notice that for $R>0$ sufficiently large,
$$\Delta^{2}u=\Delta^{2}U=0 \quad\quad \textrm{ in } \Sigma\backslash\Sigma_{R},$$
$$\Delta u=\Delta U=0 \quad\quad \textrm{ on } \partial(\Sigma\backslash\Sigma_{R}).$$
According to \cite{MW}, we have
\begin{equation}
\Delta u, \partial_{\nu}(\Delta u), \Delta U, \partial_{\nu}(\Delta U) \textrm{ are of order } \mathcal{O}(|x'|^{-n}) \textrm{ as } |x'|\rightarrow\infty.
\end{equation}
We can estimate the first term on the right hand side of \eqref{greenslab} as follows
$$
\begin{array}{rl}\vspace{1ex}
& \left|-\displaystyle\int_{d_{3}(R)}\partial_{\nu}(-\Delta u)\overline{U}\,dS\right| = \left|\displaystyle\int_{|x'|=R,0<x_{n}<L}\partial_{\nu}(\Delta u)\overline{U}\,dS\right|\\ \vspace{1ex}
\leq & \left(\displaystyle\int_{|x'|=R,0<x_{n}<L}|\partial_{\nu}(\Delta u)|^{2}\,dS\right)^{\frac{1}{2}} \left(\displaystyle\int_{|x'|=R,0<x_{n}<L}|U|^{2}\,dS\right)^{\frac{1}{2}}\\ \vspace{1ex}
\leq & \mathcal{O}(R^{-\frac{n}{2}-1})\left(\displaystyle\int_{|x'|=R,0<x_{n}<L}|U|^{2}\,dS\right)^{\frac{1}{2}}\\ \vspace{1ex}
\leq & \mathcal{O}(R^{-\frac{n}{2}-1})\left(\displaystyle\int_{\partial\Sigma_{R}}|U|\,dS\right)^{\frac{1}{2}}\\ \vspace{1ex}
\leq & C \mathcal{O}(R^{-\frac{n}{2}-1})\|U\|_{H^{2}(\Sigma)} \rightarrow 0 \textrm{ as } R\rightarrow\infty.\\
\end{array}
$$
where in the last step the constant $C$ comes from the trace theorem. Similarly, all the other terms involving $d_{3}(R)$ on the right hand side of \eqref{greenslab} will vanish as $R\rightarrow\infty$. Therefore, after taking the limit $R\rightarrow\infty$ in \eqref{greenslab}, the right hand side will become \eqref{righthandside}, as we have claimed.\\
\end{appendices}
\begin{center}
ACKNOWLEDGEMENT
\end{center}
The author would like to thank Professor Gunther Uhlmann for his constant encouragement and support. The author would also like to thank Dr. Katya Krupchyk for her assistance and helpful discussions. This work is partially supported by the NSF grant DMS 126598.
|
\section{Introduction}
\label{intro}
In conventional classification learning, the aim is to learn a function which assigns
discrete (scalar) labels to unseen objects, given a set of already labeled training set
examples. There exist tasks in computational linguistics or bioinformatics, which are
often described as mappings from input sequences to output sequences. As an example,
in computational linguistics, such tasks include part-of-speech (POS) tagging, named
entity recognition (NER) and shallow parsing~\mycite{lbfgs}
In this work, we focus on discriminative models for sequence learning.
\myshortcite{crf} introduced conditional random fields (CRFs), an undirected graphical
model that models $p({{ \bf {y} }}|{{ \bf {x} }})$ directly, and proposed to use iterative scaling algorithms
for CRF training. Subsequently, ~\myshortcite{lbfgs} demonstrated that preconditioned
conjugate gradient or limited-memory quasi-Newton (L-BFGS) methods offer significant
training speed advantages over iterative scaling. These batch algorithms were found to
be very slow on large sequence labeling problems. On a benchmark data set with about
$35,000$ examples and $18$ million features, L-BFGS method for CRF training required about
\emph{five} days on a reasonably fast machine to design a classifier. Therefore, in this
work, we consider online algorithms where model parameters are updated after
visiting each example.
Note that, though some of the
algorithms discussed here are fundamentally batch algorithms, they have an online feel
and are reasonably fast.
Many real world prediction problems can be posed as structured prediction problems,
where the output is a structured object like a sequence or a tree or a graph.
Large margin methods like support vector machines (SVMs) have shown much
promise for structured output learning~\mycite{tsochan}.
In recent years, several learning algorithms have been proposed to solve the
structured prediction problems involving sequence
labeling. Some prominent methods among them include stochastic gradient descent (SGD)
algorithm for CRFs~\mycite{sgd}, Sequential Minimal Optimization (SMO)~\mycite{m3n},
Cutting Plane method~\mycite{Joa2009},
Sequential Dual Method (SDM)~\mycite{sdm}, exponentiated gradient (EG) methods~\mycite{eg}
for max-margin Markov networks (also called Structural SVMs)
and
structured perceptron~\mycite{spcollins}.
These methods assume exact inference,
which is often computationally expensive.
\myshortcite{Huang}
proposed the variants
of perceptron,
called ``violation-fixing perceptrons",
which use a violation (approximate inference) in each update.
It is important to note that all these algorithms
(except the structured perceptron algorithm) solve either a primal problem or a dual problem.
It will be therefore interesting to compare these algorithms in terms of their speed
and generalization performance achieved by the resulting model.
\myshortcite{nguyen} compared
some prominent algorithms for sequence labeling. However, as pointed out in~\mycite{crfcomp},
this comparison employed different internal feature functions and therefore the comparison was not fair. In the case of sequence labeling, to the best of our knowledge,
there has been no systematic comparison of models obtained by using different algorithms, which solve either a primal
or a dual problem and use the same set of feature functions. We believe that, this evaluation will be useful for practitioners and help
them choose an appropriate method for sequence labeling depending upon the requirement.
{\bf Contributions}:
This work is motivated by the need to compare different sequence labeling
algorithms systematically on real-world data sets and make recommendations about the algorithm selection.
We consider two types of convex loss functions and compare methods which solve the
regularized loss minimization problem (either the primal problem or its equivalent dual).
The loss functions used are: a) A variant of the hinge
loss function used in large-margin classification problems, and b) the negative
log-likelihood function (used in CRFs).
In particular, we compare i) stochastic gradient descent (SGD) method for CRF to solve
the primal problem for regularized loss function (b), ii) Cutting Plane (CP) method
to solve the dual problem of regularized loss function (a), iii) Sequential Dual Method (SDM)
(which can be used to solve the dual problems obtained using either of the loss functions), and
iv) Averaged Structured Perceptron Algorithm (which does not use any objective function).
On a number of experiments carried out on large-scale real-world data sets, we observed that
the sequential dual method for SVMs and stochastic gradient descent methods should be preferred if test set accuracy and likelihood are respectively the evaluation criteria.
Further, the averaged structured perceptron algorithm does achieve comparable
test set accuracy, if not test set likelihood. This is despite the fact that it
does not optimize any objective function!
The paper is organized as follows.
The following section briefly describes
different sequence learning algorithms used in this work.
The details of various experiments
performed on large data sets and their results
are
given in Section~\ref{experiments}.
Our recommendations are presented in Section~\ref{conclusion}.
\section{Sequence Learning Algorithms}
\label{seqla}
In this section, we describe various sequence learning algorithms used in this work.
We assume that a training set $S$ of input-output sequence pairs is available. Let $S = \{ {{ \bf {x} }}_i, {{ \bf {y} }}_i \}^n_{i=1}$ where ${{ \bf {x} }}_i \in \cal X$
and ${{ \bf {y} }}_i \in \cal Y$ for every $i$. The goal is to learn a discriminant
function $g:\cal X \times \cal Y \rightarrow \rr$ over the training set $S$
from which prediction for a input ${{ \bf {x} }}$ is given by
$$
h({{ \bf {x} }}) = \arg\max_{{{ \bf {y} }} \in \cal Y} \;\;g({{ \bf {x} }}, {{ \bf {y} }}).
$$
For sequence labeling problems, the $\arg\max$ computation in the above equation can
be done using dynamic programming like the Viterbi algorithm.
If an input sequence and the corresponding output sequence are of length $L$, and each individual
label of the output sequence can take values from the set $\Sigma$, then the sequence labeling
problem can be considered as a multi-class classification problem with $|\Sigma|^L$ classes. This
demonstrates that the cardinality of $\cal Y$ grows exponentially with the size of ${{ \bf {x} }}$. In this
work, we assume that $g({{ \bf {x} }}, {{ \bf {y} }})$ takes the form of a linear function,
$$
g({{ \bf {x} }}, {{ \bf {y} }}) = {{\bf {w}}}^T f({{ \bf {x} }}, {{ \bf {y} }})
$$
where ${{\bf {w}}}$ is a parameter vector and $f({{ \bf {x} }}, {{ \bf {y} }})$ is a feature vector
relating input ${{ \bf {x} }}$ and ${{ \bf {y} }}$. Note that the feature vectors $f({{ \bf {x} }},{{ \bf {y} }})$
have a crucial effect on the performance of the designed structured classifier~\mycite{crfcomp}.
Using a variant of the hinge loss function, the sequence learning problem can be posed as an
extension of multi-class SVM problem as follows~\mycite{crammer}:
\begin{eqnarray}
\min_{{{\bf {w}}}, {\mbox{\boldmath $\xi$}}} \;\; \frac{\lambda}{2} {\| {{\bf {w}}} \|}^2 + \sum_i \xi_i \nonumber \\
{\rm s.t.} \;\; {{\bf {w}}}^T {\Delta f_i ({\by})} \ge l_i({\by}) - \xi_i \;\; \forall \;i, {{ \bf {y} }}
\label{eq:mm-primal}
\end{eqnarray}
where $\Delta f_i ({\by}) = f({{ \bf {x} }}_i, {{ \bf {y} }}_i) - f({{ \bf {x} }}_i, {{ \bf {y} }})$ and $\lambda > 0$ is a regularization parameter. $l_i({{ \bf {y} }})$ in~(\ref{eq:mm-primal}) is a loss function that quantifies the loss associated with predicting ${{ \bf {y} }}$ instead of the correct output ${{ \bf {y} }}_i$. In sequence learning problems,
a natural choice for the loss function is the Hamming distance,
$$
l_i({{ \bf {y} }}) = \sum^L_{j=1} I(y^j_i \ne y^j)
$$
where $I(\cdot) $ is the indicator function and ${{ \bf {y} }} = (y^1,y^2,\ldots,y^L)$.
By defining the
conditional distribution,
\begin{eqnarray}
p({{ \bf {y} }} | {{ \bf {x} }}; {{\bf {w}}}) = \frac{e^{{{\bf {w}}}^T f({{ \bf {x} }}, {{ \bf {y} }})}}{Z_x}
\label{eq:pd}
\end{eqnarray}
where $Z_x = \sum_{y^\prime} e^{{{\bf {w}}}^T f({{ \bf {x} }}, {{ \bf {y} }}^\prime)}$ is the partition function
and using the negative log-likelihood function, the parameter ${{\bf {w}}}$ can be learned
by solving the following unconstrained optimization problem:
\begin{eqnarray}
\min_{{{\bf {w}}}} \;\; \frac{\lambda}{2} {\| {{\bf {w}}} \|}^2 - \sum_i \log p({{ \bf {y} }}_i | {{ \bf {x} }}_i; {{\bf {w}}})
\label{eq:ll-primal}
\end{eqnarray}
where $\lambda > 0$ is a regularization parameter. CRF training (batch algorithm) involves
solving the problem~(\ref{eq:ll-primal}) in batch mode.
Note that both the problems~(\ref{eq:mm-primal}) and~(\ref{eq:ll-primal}) are convex programming
problems and the optimal solution ${{\bf {w}}}$ can be used for
making the prediction for a specific input ${{ \bf {x} }}$ as
$\arg\max_{{{ \bf {y} }} \in \cal Y} \;\;{{\bf {w}}}^T f({{ \bf {x} }},{{ \bf {y} }})$.
We note that a variety of techniques
have been developed in the literature to solve large scale sequence learning
problems. These include bundle method~\mycite{bundle}, fast Newton-CG method for batch learning of CRFs~\mycite{newtoncgcrf},
SGD and block coordinate methods for
$L_1$ regularized and elastic-net CRFs~\mycite{practicalcrf},
stochastic
meta-descent method~\mycite{crfsgdaccel},
stochastic block coordinate Frank-Wolfe optimization~\mycite{frankwolf},
dual coordinate ascent method~\mycite{gimpelsmith} and
excessive gap reduction technique~\mycite{egspeedup}.
The number of dual variables for the problems~(\ref{eq:mm-primal}) and~(\ref{eq:ll-primal}) is exponentially large and techniques which use
marginal variables (which are polynomial in number and from which dual
variables can be easily derived) were also
proposed in the literature.
These techniques include exponentiated gradient method~\mycite{eg} and
sequential minimal optimization algorithm~\mycite{m3n}.
It will be difficult to
compare every method proposed in the literature. Therefore, in this work, we
restrict ourselves to algorithms which are simple, easy to implement
from a practitioner's viewpoint and solve
the problems~(\ref{eq:mm-primal}) and~(\ref{eq:ll-primal}) or their dual
problems directly.
Recently, adaptive-rate and parameter-free variants
of SGD have also been proposed
for binary classification tasks in
\mycite{adaptivesgd} and \mycite{peskysgd}.
While a simple SGD chooses the learning rate
to be $\frac{1}{t}$, $t$ being the number
of iterations,
adaptive-rate variant of SGD \mycite{adaptivesgd}
updates the learning rate for each component of
the model parameter ${{\bf {w}}}$ by incorporating
gradient-information from the past iterations.
However, tuning is required to find
an initial choice of the
learning rate.
\myshortcite{peskysgd} propose to
automatically choose the learning rate based on
a second-order approximation of the loss function
at a particular component of ${\bf {w}}$,
thus eliminating the need to tune the learning
rate altogether.
However, making the learning rate
automatically-tuned or adaptable,
comes at an added cost of
computing the second order information
or book-keeping gradient information, thus
increasing the iteration complexity of the basic SGD.
Moreover, the results given in \mycite{peskysgd}
show that the generalization performance
achieved by the basic SGD is
comparable and sometimes better than that
achieved by the
adaptive-rate and parameter-free variants
of SGD. Hence, we resort to the basic SGD method with
averaging~\mycite{sgd}, in our experiments.
In the following, we give a brief description of
the algorithms used to solve the problems~(\ref{eq:mm-primal}) and~(\ref{eq:ll-primal}) and compared in this work.
\subsection{Cutting Plane Method ({\bf CP})}
\myshortcite{Joa2009} proposed an equivalent formulation of the problem
~(\ref{eq:mm-primal}), given by~(\ref{eq:Joach})
and presented a cutting-plane algorithm, which is significantly fast on large scale problems.
It was shown
that the dual problem of~(\ref{eq:Joach}) has a sparse solution (that is, the number of non-zero dual variables
is small at optimality).
\begin{align}
\min_{{{\bf {w}}}, \xi} \;\; \frac{\lambda}{2} {\| {{\bf {w}}} \|}^2 + \xi \nonumber \\
{\rm s.t.} \;\; \frac{1}{n} {{\bf {w}}}^T \sum^n_{i=1} {\Delta f_i ({\by})} \ge \frac{1}{n} \sum^n_{i=1} l_i({\by}) - \xi, \nonumber \\
\; \forall \; \{{{ \bf {y} }}_1,\ldots,{{ \bf {y} }}_n\} \in {\cal Y}^n \label{eq:Joach}
\end{align}
Even for large data sets, the size of the quadratic programs that need to be solved
was observed to be very small (as the number of violated constraints was very small) and therefore, the method achieved considerable speed-up.
We note that the bundle method presented in~\mycite{bundle} is similar to
the cutting-plane algorithm and therefore, we do not use it for comparison.
\subsection{Sequential Dual Method for Structural SVMs ({\bf SVM-SDM})}
\myshortcite{sdm} suggested the use of sequential dual method (SDM) to solve the dual problem of~(\ref{eq:mm-primal}).
\begin{align}
\min & \frac{1}{2 \lambda} {\| \sum_{i, {{ \bf {y} }}} {\mbox{\boldmath $\alpha$}_i({\by})} \Delta f_i ({\by}) \|}^2 - \sum_{i, {{ \bf {y} }}} {\mbox{\boldmath $\alpha$}_i({\by})} l_i({\by}) \nonumber \\
{\rm s.t.} & \sum_{{ \bf {y} }} {\mbox{\boldmath $\alpha$}_i({\by})} = 1 \; \forall \; i, \; {\mbox{\boldmath $\alpha$}_i({\by})} \ge 0 \; \forall \; i, {{ \bf {y} }} \label{eq:mm-dual}
\end{align}
The method makes repeated passes over the training data set and optimizes
the dual variables associated with one example at a time, until some
stopping condition is satisfied.
Note that the number of dual variables in~(\ref{eq:mm-dual}) is exponentially large for every example $i$.
However, at optimality, very few of the dual variables are strictly positive.
This fact was used to develop a fast and
efficient algorithm to solve the dual problem~(\ref{eq:mm-dual}).
Some heuristics were also proposed to make the sequential dual method more
efficient.
This method was found to
be an order of magnitude faster than the cutting-plane method on many sequence-learning data sets~\mycite{sdm}.
\subsection{Averaged Stochastic Gradient Descent Method for CRF ({\bf CRF-ASGD})}
Gradient based online methods like stochastic gradient descent (SGD) can be used to solve the CRF primal
problem~(\ref{eq:ll-primal})~\mycite{sgd}. The SGD method is fast and is quite useful when the training data
size is large. It operates by visiting each example and updating the parameter ${{\bf {w}}}$ through a simple update
step. For example, at iteration $t$, using a single training example (say ${{ \bf {x} }}_i$), the following simple
update rule is used:
\begin{align}
{{\bf {w}}}_{t+1} = {{\bf {w}}}_t - \gamma_t (\lambda {{\bf {w}}}_t + f({{ \bf {x} }}_i, {{ \bf {y} }}_i) - E_{p(y|x_i)} f({{ \bf {x} }}_i, {{ \bf {y} }})) \\ \nonumber
\end{align}
where $\gamma_t$ is a learning rate parameter (typically set to $\frac{1}{t+1}$). The expectation term,
$E_{p(y|x_i)} f({{ \bf {x} }}_i, {{ \bf {y} }})$ is calculated with respect to the conditional probability~(\ref{eq:pd}).
This is usually done using a forward-backward algorithm~\mycite{crf}. The SGD method, though simple, requires multiple
passes over the data before it converges to the optimal solution. To overcome this difficulty, ~\myshortcite{asgd}
proposed a method which averages the parameter ${{\bf {w}}}$. An average parameter $\bar{{\bf {w}}}$ is maintained and updated
at every iteration,
\begin{eqnarray}
\bar{{{\bf {w}}}}_{t+1} = \frac{t}{t+1} \bar{{\bf {w}}}_t + \frac{1}{t+1} {{\bf {w}}}_{t+1}.
\end{eqnarray}
This method, called averaged SGD (ASGD), has been demonstrated to make reasonable progress
in the objective function in the initial few iterations. A thorny issue with online
methods like SGD or ASGD methods is the choice of the initial learning rate, $\gamma_0$. With
an improper choice, the methods might become very slow on large data sets. Choosing a suitable
$\gamma_0$ involves taking a random sample of the data set initially and performing SGD or ASGD
updates on this selected sample using different learning rate values and then choosing the best
learning rate as the rate which gives maximum decrease in the objective function. Having determined
the initial learning rate $\gamma_0$, the method then proceeds in the normal fashion on the entire
training set.
\subsection{ Sequential Dual Method for CRF ({\bf CRF-SDM})}
Inspired by the speed-up achieved by SVM-SDM over state-of-the-art methods, we implemented a sequential dual
method to solve the dual of the CRF primal problem~(\ref{eq:ll-primal}).
\myshortcite{memisevic} and \myshortcite{maxent} proposed to
solve the dual of~(\ref{eq:ll-primal}) for multi-class classification problems. These methods
cannot be directly used for sequence learning problems as the number of dual variables is exponentially
large for such problems. Further, the resulting model at optimality is not sparse.
That is, all the possible sequences are used to define the parameter vector,
\begin{align}
{\bf {w}} ({\mbox{\boldmath $\alpha$}})=\frac{1}{\lambda} \sum_{i, {{ \bf {y} }}} {\mbox{\boldmath $\alpha$}_i({\by})} \Delta f_i ({\by}) \label{eq:wdual}
\end{align}
when the following dual problem of~(\ref{eq:ll-primal})
is solved:
\begin{align}
\min \;\; \frac{1}{2 \lambda} {\| {\bf {w}} ({\mbox{\boldmath $\alpha$}}) \|}^2 + \sum_{i, {{ \bf {y} }}} {\mbox{\boldmath $\alpha$}_i({\by})} \log {\mbox{\boldmath $\alpha$}_i({\by})} \nonumber \\
{\rm s.t.} \;\; \sum_{{ \bf {y} }} {\mbox{\boldmath $\alpha$}_i({\by})} = 1 \; \forall \; i, \; {\mbox{\boldmath $\alpha$}_i({\by})} \ge 0 \; \forall \; i, {{ \bf {y} }} \label{crfdual}
\end{align}
One way to alleviate this problem
is to assume that ${\mbox{\boldmath $\alpha$}_i({\by})} \leq \eta$ for many ${ \bf {y} } \in \cal Y$ corresponding to
every example $i$, where $\eta$ is very small (say, $10^{-18}$)
and solve the dual problem with
respect to the sequences in the set, $V_i=\{ { \bf {y} } : {\mbox{\boldmath $\alpha$}_i({\by})} > \eta\}$
using SMO-type algorithm with the
modified Newton method illustrated in \mycite{maxent}.
Note that the CRF-SDM method finds only an \emph{approximate}
solution to~\eqref{crfdual} because of the practical
limitation on the size of the set $V_i$.
Our experiments presented in the next section,
clearly indicate that on many data sets,
such an approximate solution is sufficient to obtain
a comparable generalization performance.
\subsection{Averaged Structured Perceptron ({\bf AvStructPerc})}
Perceptron algorithm for binary classification is simple and does not use any objective function.
\myshortcite{spcollins} proposed the perceptron algorithm for structural learning. After randomly initializing the
weight vector ${{\bf {w}}}$, the algorithm makes repeated passes over the training data set (visiting one
example at a time), until some stopping condition is satisfied. In every pass, after visiting an example
$({{ \bf {x} }}_i, {{ \bf {y} }}_i)$, the following update rule is used if the current weight vector fails to predict
the desired label ${{ \bf {y} }}_i$:
\begin{eqnarray}
{{\bf {w}}} := {{\bf {w}}} + \eta (f({{ \bf {x} }}_i, {{ \bf {y} }}_i) - f({{ \bf {x} }}_i, {{ \bf {y} }}^\prime))
\end{eqnarray}
where ${{ \bf {y} }}^\prime = \arg\max_y \; {{\bf {w}}}^T f({{ \bf {x} }}_i, {{ \bf {y} }})$ and $\eta (> 0)$ is the learning rate parameter.
\myshortcite{spcollins} also proposed a simple refinement to the perceptron algorithm. Defining $w^{t,i}_j$ to be the
value of the $j$-th parameter after the $i$-th training example has been visited in pass $t$ over the
training data, the average parameter ${{\bf {w}}}_{avg}$ is defined as,
\begin{eqnarray}
{{\bf {w}}}_{{avg},j} = \sum^T_{t=1} \sum^n_{i=1} w^{t,i}_j / (nT) \;\; \forall \; j=1,\ldots,d
\end{eqnarray}
where $d$ denotes the dimension of the parameter vector ${{\bf {w}}}$.
In our implementation, we adopted a very efficient way of updating ${{\bf {w}}}_{avg}$.
It was observed that the use of
this refined algorithm gave better generalization performance~\mycite{spcollins}.
In our experiments, we observed
that the value of $\eta$ is crucial to get good generalization performance and therefore, preferred to
choose it using the procedure similar to that used in the SGD method. Since the averaged structured perceptron algorithm
does not optimize any objective function, the learning rate which resulted in the least error on the remaining
training set was chosen.
\begin{table*}[ht]
\caption{\textbf{Summary of data sets}. $n$, $n_{val}$ and $n_{test}$ denote the sizes of the training, validation and test data respectively,
$d$ is the input dimension, $k$ denotes the number of alphabets and $N$ is the feature vector dimension}
\label{dataset_tab}
\begin{center}
\begin{tabular}{|p{20mm}|p{11mm}|p{11mm}|p{11mm}|p{15mm}|p{5mm}|p{15mm}|}
\hline %
Data set & $n$ & $n_{val}$ & $n_{test}$ & $d$ & $k$ & $N$ \\\hline
BioCreative & 6000 & 1500 & 5000 & 102409 & 3 & 307236 \\
BioNLP & 14836 & 3710 & 3856 & 513932 & 11 & 5653373 \\
CoNLL & 14987 & 3684 & 3466 & 651041 & 8 & 5208392\\
dataPOS & 39832 & 2416 & 1700 & 258299 & 45 & 11625480\\
WSJPOS & 28424 & 7107 & 1681 & 446147 & 42 & 18739938 \\\hline
\end{tabular}
\end{center}
\vspace{-.1in}
\end{table*}
We give details
of experimental evaluation in the next section.
\section{Experiments and Discussion}
\label{experiments}
In this section we compare different sequence learning algorithms, mentioned
in Section~\ref{seqla}, on five benchmark data sets:
\begin{itemize}
\item Wall Street Journal POS (WSJPOS)~\mycite{wsjpos}
\item BioNLP~\mycite{bionlp}
\item BioCreative~\mycite{biocre}
\item CoNLL~\mycite{conll2003} and
\item dataPOS~\mycite{datapos}.
\end{itemize}
The characteristics of these data sets
are summarized in Table~\ref{dataset_tab}.
In all our experiments, the feature vector $f({ \bf {x} },{ \bf {y} })$ was
constructed using the combination of first order and token-independent
second order feature functions~\mycite{crfcomp}.
For the first order feature function, we aggregated the $d$ dimensional
feature vector over the nodes for each label. The second order functions
used are independent of the tokens, but capture the dependencies between
the labels of two neighbouring nodes.
In this case, the dimension $N$ of the feature vector is $k^2+d k$.
The value of the hyper-parameter $\lambda$ was set to $10$ and $1$ respectively
for the problems~(\ref{eq:mm-primal}) and~(\ref{eq:ll-primal}).
To compare the test likelihood achieved by batch CRF with that obtained
by the models trained using online algorithms, limited
memory BFGS method (CRF-LBFGS) was used in the batch mode.
To get an idea about the steady state generalization performance of the
resulting models trained using online algorithms, the algorithms were run for a large number of iterations.
Test set accuracy (for large-margin related methods) and likelihood (for CRF related methods) were used as performance measures to
compare different algorithms.
Validation set performance was calculated at the end of every iteration; this
time was not counted for CPU time calculations.
An algorithm could be stopped if there is no significant improvement in
the validation set performance.
This stopping criterion was used for all the methods.
This condition is represented by a black square on each graph in Figures~\ref{biocreativefig}-\ref{wsjposfig}
The computation of $\arg\max_{{ \bf {y} }} {\bf {w}}({\mbox{\boldmath $\alpha$}})^T f({\bx_i},{\by})$ and similar
terms was
performed using the Viterbi algorithm. All experiments were run on a $2.4$GHz Intel Xeon
processor with $16$GB of shared main memory under Linux.
The plots in Figures~\ref{biocreativefig}-\ref{wsjposfig}. depict the performance of different algorithms.
The results are summarized in Table~\ref{largedatacompare}.
{\bf Generalization Performance (Test Set Accuracy)} : The left
panel of Figures~\ref{biocreativefig}-\ref{wsjposfig}. shows the behaviour of test set accuracy of
the resulting models, as a function of CPU time. From these plots, it is
evident that the SVM-SDM method achieves the steady state performance much faster
than other methods. The difference is significant especially
for large data sets like dataPOS and WSJPOS (Figure~\ref{dataposfig} and Figure~\ref{wsjposfig}, left panel). It is worth
noting that the performance of the averaged structured perceptron is comparable with
SVM-SDM on BioCreative and CoNLL data sets(Figure~\ref{biocreativefig} and Figure~\ref{dataconll2003fig}).
On large data sets, however, it showed degrading behaviour
as iterations progressed (Figures~\ref{dataposfig}-\ref{wsjposfig}).
The averaged structured perceptron has this overfitting/overtraining issue
that is bothersome. The methods which solve the
problem~(\ref{eq:ll-primal}) or its dual were not found to be suitable
if test set accuracy is used as a performance measure. The main reason is
that these methods are designed to maximize the likelihood. On the other hand,
methods like CRF-ASGD perform better if test set likelihood is used
as a performance measure. We now discuss this.
\begin{table}[!h]
\caption{\textbf{Comparison of various algorithms at the point of termination based on validation set performance. Acc - Accuracy, ANLL - Average Negative Log-likelihood. The best and the second best results are highlighted in boldface and italic style respectively.}}
\label{largedatacompare}
\begin{center}
\small
\begin{tabular}{|c|c|c|c|c|}
\hline %
Dataset & Algorithm & Time (sec) & \small{Test Acc(\%)} & Test ANLL \\\hline
\small{BioCreative} & \small{AvStructPerc} & {\bf 4.84} & 98.67 & 7800 \\\hline
& \small{SVM-SDM} & {\it 7.21} & {\bf 98.73} & 21820 \\\hline
& \small{CP} & 31.17 & {\it 98.72} & 16420 \\\hline
& \small{CRF-SDM} & 20.95 & 98.6 & {\em 5053} \\\hline
& \small{CRF-ASGD} & 15.16 & 98.38 & 5491 \\\hline
& \small{CRF-LBFGS} & 270.8 & 98.58 & {\bf 4667} \\\hline\hline
BioNLP & \small{AvStructPerc} & {\it 39.23} & {\bf 98.01} & 210600 \\\hline
& \small{SVM-SDM} & {\bf 36.93} & 97.72 & 56170 \\\hline
& \small{CP} & 369.7 & {\it 97.86} & 28730 \\\hline
& \small{CRF-SDM} & 138.5 & 97.75 & 6920 \\\hline
& \small{CRF-ASGD} & 151.7 & 97.77 & {\it 6522} \\\hline
& \small{CRF-LBFGS} & 8850 & 97.76 & {\bf 6439} \\\hline\hline
CoNLL & \small{AvStructPerc} & {\it 34.69} & {\bf 95.8} & 51330 \\\hline
& \small{SVM-SDM} & {\bf 26.85} & {\bf 95.8} & 37060 \\\hline
& \small{CP} & 158.4 & 95 & 34020 \\\hline
& \small{CRF-SDM} & 45.11 & 95.38 & 6297 \\\hline
& \small{CRF-ASGD} & 79.02 & 95.31 & {\it 5877} \\\hline
& \small{CRF-LBFGS} & 1350 & 95.51 & {\bf 5765} \\\hline\hline
dataPOS & \small{AvStructPerc} & {\it 415.3} & 97.14 & 8256 \\\hline
& \small{SVM-SDM} & {\bf 235.2} & {\bf 97.36} & 84900 \\\hline
& \small{CP} & 808.7 & 97.05 & 68310 \\\hline
& \small{CRF-SDM} & 1446 & 97.27 & 6511 \\\hline
& \small{CRF-ASGD} & 8497 & {\it 97.29} & {\it 4261} \\\hline
& \small{CRF-LBFGS} & 609600 & 97.28 & {\bf 4215} \\\hline\hline
WSJPOS & \small{AvStructPerc} & {\it 250.6} & 96.46 & 5576 \\\hline
& \small{SVM-SDM} & {\bf 128.8} & {\bf 96.6} & 627900 \\\hline
& \small{CP} & 870.7 & 96.23 & 511200 \\\hline
& \small{CRF-SDM} & 1620 & 96.56 & 6363\\\hline
& \small{CRF-ASGD} & 4391 & {\it 96.57} & {\it 3954} \\\hline
& \small{CRF-LBFGS} & 463700 & 96.5 & {\bf 3953} \\\hline\hline
\end{tabular}
\end{center}
\end{table}
{\bf Generalization Performance (Test Set Likelihood)} : The behaviour
of test set likelihood as a function of CPU time, is
depicted in the right panel of Figures~\ref{biocreativefig}-\ref{wsjposfig}.
Note that the methods
CRF-ASGD and CRF-LBFGS (batch algorithm) optimize the training set
likelihood directly by solving problem~(\ref{eq:ll-primal}).
On the other hand, the sequential dual method for CRFs (CRF-SDM)
solves the dual problem of~(\ref{eq:ll-primal}).
From the results in Table~\ref{largedatacompare}, we see that
the CRF-LBFGS method, though slow, gave the best test likelihood value
among all
the methods compared.
While the performance of CRF-SDM and CRF-ASGD was comparable on
three data sets,
the CRF-ASGD method clearly outperformed the CRF-SDM method on
large data sets like dataPOS and WSJPOS. This
is mainly due to the large number of alphabets ($k$ in Table 1)
in these data sets. This results in a large number of possible output
sequences for every example.
Since the CRF-SDM method assumes that, only those sequences in the set
$V_i$ have a non-zero value of the dual variable, and the set $V_i$
cannot accommodate all possible sequences associated with example $i$, the
vector ${{\bf {w}}}({\mbox{\boldmath $\alpha$}})$ in~(\ref{eq:wdual}) cannot be accurately determined.
This results in degradation of performance of CRF-SDM, especially on those
data sets where the number of alphabets is large.
Test likelihood performance of CP and SVM-SDM methods was not good
compared to other methods, as these methods were mainly designed to solve the dual
problems obtained using a variant of the hinge loss function.
On the other hand, the averaged structured perceptron, which does not optimize any objective function, performed better than these methods on all the data sets except BioNLP and CoNLL.
The averaged structured perceptron algorithm is simple, easy to implement, fast and gives reasonable test accuracy and likelihood performance.
However, it does begin to overfit eventually, as is evident from Figures~\ref{biocreativefig}-\ref{wsjposfig}.
Some key observations, made in this Section, are summarized in
Table~\ref{perfallmethods}.
{\bf Sensitivity to hyper-parameter selection}: For the Structural SVM and CRF model
design, the effect of hyper-parameter selection on the test set performance was studied by conducting
the following experiment. The hyper-parameter $\lambda$ (in \eqref{eq:mm-primal} or \eqref{eq:ll-primal})
was varied from $10^{-3}$ to $10^{3}$ and the test set accuracy and likelihood values were noted.
The results are reported in Table~\ref{sensitivitytable}.
It is clear from this table that the variation in test accuracy was not large for Structural SVM
over different values of $\lambda$. On the other hand, test likelihood values
varied significantly over a range of $\lambda$ values, when CRF was used.
We however note that, for most of the data sets, $\lambda=10$ was an optimal choice for
Structural SVMs (with accuracy as a measure) and the corresponding choice for CRF was $\lambda=1$ (with likelihood
as a measure). These optimal choices of hyper-parameters were used in all our experiments.
\section{ Conclusion and Recommendations}
\label{conclusion}
In this work, we have done a systematic comparison of different
sequence labeling algorithms in terms of their speed,
ability to reach a good generalization performance (accuracy and
likelihood) fast, and
ability to maintain best generalization performance at the end.
Based on
experimental results on real-world benchmark data sets,
we recommend that a dual method like SVM-SDM which solves the dual
of~(\ref{eq:mm-primal}) is preferred if test accuracy is the evaluation
criterion.
The averaged structured perceptron yields good test accuracy; however it has
an overfitting issue that is bothersome.
On the other
hand, the CRF-ASGD method should be preferred if likelihood is the criterion.
\begin{table}[ht]
\caption{\textbf{Sensitivity of Structural SVM and CRF to} $\lambda$. \textbf{ANLL - Average Negative Log-likelihood. (Not given for Structural SVM as one does not expect any pattern there.) The best results are highlighted in boldface style.} }
\label{sensitivitytable}
\begin{center}
\small
\begin{tabular}{|c|c|c|c|c|}
\hline %
Dataset & $\lambda$ & \multicolumn{2}{c|}{Test Accuracy (\%)} & Test ANLL\\\cline{3-5}
& & Struct SVM & CRF & CRF \\
& & (SDM) & (ASGD) & (ASGD)\\\hline
\small{BioCreative}&$10^{3}$ & 97.27 & 94.17 & 15186 \\\hline
&$10^{2}$ & 98.07 & 97.01 & 9278 \\\hline
&10 & 98.76 & 98.16 & 6070 \\\hline
&1 & {\bf 98.77} & 98.5 & 4894 \\\hline
&$10^{-1}$ & 98.36 & 98.57 & 4725 \\\hline
&$10^{-2}$ & 98.07 & 98.57 & {\bf 4707} \\\hline
&$10^{-3}$ & 97.47 & 98.57 & 4707 \\\hline\hline
\small{BioNLP}&$10^{3}$ & 96.46 & 91.6 & 21105 \\\hline
&$10^{2}$ & 97.83 & 96.07 & 10752 \\\hline
&10 & {\bf 98.15} & 97.45 & 7115 \\\hline
&1 & 98.02 & 97.82 & {\bf 6189} \\\hline
&$10^{-1}$ & 97.56 & 97.88 & 6452 \\\hline
&$10^{-2}$ & 97.29 & 97.88 & 6550 \\\hline
&$10^{-3}$ & 96.5 & 97.88 & 6553 \\\hline\hline
\small{CoNLL}&$10^{3}$ & 91.75 & 89.57 & 12763 \\\hline
&$10^{2}$ & 93.94 & 92.01 & 9076 \\\hline
&10 & {\bf 95.81} & 94.6 & 6680 \\\hline
&1 & 95.6 & 95.54 & {\bf 5630} \\\hline
&$10^{-1}$ & 94.75 & 95.74 & 5659 \\\hline
&$10^{-2}$ & 94.03 & 95.75 & 5715 \\\hline
&$10^{-3}$ & 92.59 & 95.75 & 5716 \\\hline\hline
\small{dataPOS}&$10^{3}$ & 95.1 & 90.7 & 25560 \\\hline
&$10^{2}$ & 96.7 & 95.79 & 9989 \\\hline
&10 & {\bf 97.36} & 97.05 & 5358 \\\hline
&1 & 97.04 & 97.27 & {\bf 4273} \\\hline
&$10^{-1}$ & 96.17 & 97.13 & 5142 \\\hline
&$10^{-2}$ & 95.01 & 96.99 & 6670 \\\hline
&$10^{-3}$ & 94.5 & 96.99 & 6770 \\\hline\hline
\small{WSJPOS}&$10^{3}$ & 93.4 & 87.87 & 21964 \\\hline
&$10^{2}$ & 95.88 & 94.36 & 8933 \\\hline
&10 & {\bf 96.59} & 96.31 & 4769 \\\hline
&1 & 96.25 & 96.53 & {\bf 3939} \\\hline
&$10^{-1}$ & 95.56 & 96.42 & 4544 \\\hline
&$10^{-2}$ & 95.01 & 96.35 & 5159 \\\hline
&$10^{-3}$ & 93.7 & 96.35 & 5188 \\\hline\hline
\end{tabular}
\end{center}
\end{table}
\begin{table*}[h]
\caption{\textbf{Generalization performance characteristics of various algorithms}}
\noindent
\footnotesize
\label{perfallmethods}
\begin{center}
\begin{tabular}{|p{3.3cm}|c|c|c|c|c|c|}
\hline %
& \multicolumn{6}{|c|}{Algorithm}\\
\cline{2-7}
& AvgStrPerc & SVM-SDM & CP & CRF-SDM & CRF-SGD & CRF-LBFGS \\\hline
Ability to reach best test set & \multirow{2}{*}{\checkmark} & \multirow{2}{*}{\checkmark} & \multirow{2}{*}{\xmark} & \multirow{2}{*}{\xmark} & \multirow{2}{*}{\xmark} & \multirow{2}{*}{\xmark} \\
accuracy fast & & & & & & \\\hline
Ability to reach best test set & \multirow{2}{*}{\xmark} & \multirow{2}{*}{\xmark} & \multirow{2}{*}{\xmark} & \multirow{2}{*}{\xmark} & \multirow{2}{*}{\checkmark} & \multirow{2}{*}{\xmark} \\
likelihood fast & & & & & & \\\hline
Ability to maintain & \multirow{2}{*}{\xmark} & \multirow{2}{*}{\checkmark} & \multirow{2}{*}{\checkmark} & \multirow{2}{*}{\checkmark} & \multirow{2}{*}{\checkmark} & \multirow{2}{*}{\checkmark} \\
best test set accuracy & & & & & & \\\hline
Ability to maintain & \multirow{2}{*}{\xmark} & \multirow{2}{*}{\xmark} & \multirow{2}{*}{\xmark} & \multirow{2}{*}{\checkmark} & \multirow{2}{*}{\checkmark} & \multirow{2}{*}{\checkmark} \\
best test set likelihood & & & & & & \\\hline
\end{tabular}
\end{center}
\vspace{-.1in}
\end{table*}
\begin{figure*}[ht]
\begin{minipage}[b]{0.5\linewidth}
\centering
\includegraphics[scale=0.4]{plots/biocreative_testacc.eps}
\end{minipage}
\begin{minipage}[b]{0.5\linewidth}
\centering
\includegraphics[scale=0.4]{plots/biocreative_testlikeli.eps}
\end{minipage}
\begin{minipage}[b]{0.5\linewidth}
\centering
\includegraphics[scale=0.4]{plots/biocreative_testacc_zoominitial.eps}
\end{minipage}
\begin{minipage}[b]{0.5\linewidth}
\centering
\includegraphics[scale=0.4]{plots/biocreative_testlikeli_zoominitial.eps}
\end{minipage}
\begin{minipage}[b]{0.5\linewidth}
\centering
\includegraphics[scale=0.4]{plots/biocreative_testacc_zoomfinal.eps}
\end{minipage}
\begin{minipage}[b]{0.5\linewidth}
\centering
\includegraphics[scale=0.4]{plots/biocreative_testlikeli_zoomfinal.eps}
\end{minipage}
\caption{\textbf{Comparison of Test Accuracy and Test Likelihood for BioCreative dataset.
The plots in rows 2 and 3 are zoomed versions to clearly see certain
behaviour in the initial and final stages respectively.
} }
\label{biocreativefig}
\end{figure*}
\begin{figure}[!h]
\begin{minipage}[b]{0.5\linewidth}
\centering
\includegraphics[scale=0.4]{plots/bionlp_testacc.eps}
\end{minipage}
\begin{minipage}[b]{0.5\linewidth}
\centering
\includegraphics[scale=0.4]{plots/bionlp_testlikeli.eps}
\end{minipage}
\begin{minipage}[b]{0.5\linewidth}
\centering
\includegraphics[scale=0.4]{plots/bionlp_testacc_zoominitial.eps}
\end{minipage}
\begin{minipage}[b]{0.5\linewidth}
\centering
\includegraphics[scale=0.4]{plots/bionlp_testlikeli_zoominitial.eps}
\end{minipage}
\begin{minipage}[b]{0.5\linewidth}
\centering
\includegraphics[scale=0.4]{plots/bionlp_testacc_zoomfinal.eps}
\end{minipage}
\begin{minipage}[b]{0.5\linewidth}
\centering
\includegraphics[scale=0.4]{plots/bionlp_testlikeli_zoomfinal.eps}
\end{minipage}
\caption{\textbf{Comparison of Test Accuracy and Test Likelihood for BioNLP dataset.The plots in rows 2 and 3 are zoomed versions to clearly see certain
behaviour in the initial and final stages respectively.} }
\label{bionlpfig}
\end{figure}
\begin{figure*}[!h]
\begin{minipage}[b]{0.5\linewidth}
\centering
\includegraphics[scale=0.4]{plots/dataconll2003_testacc.eps}
\end{minipage}
\begin{minipage}[b]{0.5\linewidth}
\centering
\includegraphics[scale=0.4]{plots/dataconll2003_testlikeli.eps}
\end{minipage}
\begin{minipage}[b]{0.5\linewidth}
\centering
\includegraphics[scale=0.4]{plots/dataconll2003_testacc_zoominitial.eps}
\end{minipage}
\begin{minipage}[b]{0.5\linewidth}
\centering
\includegraphics[scale=0.4]{plots/dataconll2003_testlikeli_zoominitial.eps}
\end{minipage}
\begin{minipage}[b]{0.5\linewidth}
\centering
\includegraphics[scale=0.4]{plots/dataconll2003_testacc_zoomfinal.eps}
\end{minipage}
\begin{minipage}[b]{0.5\linewidth}
\centering
\includegraphics[scale=0.4]{plots/dataconll2003_testlikeli_zoomfinal.eps}
\end{minipage}
\caption{\textbf{Comparison of Test Accuracy and Test Likelihood for dataCoNLL2003 dataset.
The plots in rows 2 and 3 are zoomed versions to clearly see certain
behaviour in the initial and final stages respectively.
} }
\label{dataconll2003fig}
\end{figure*}
\begin{figure*}[!h]
\begin{minipage}[b]{0.5\linewidth}
\centering
\includegraphics[scale=0.4]{plots/datapos_testacc.eps}
\end{minipage}
\begin{minipage}[b]{0.5\linewidth}
\centering
\includegraphics[scale=0.4]{plots/datapos_testlikeli.eps}
\end{minipage}
\begin{minipage}[b]{0.5\linewidth}
\centering
\includegraphics[scale=0.4]{plots/datapos_testacc_zoominitial.eps}
\end{minipage}
\begin{minipage}[b]{0.5\linewidth}
\centering
\includegraphics[scale=0.4]{plots/datapos_testlikeli_zoominitial.eps}
\end{minipage}
\begin{minipage}[b]{0.5\linewidth}
\centering
\includegraphics[scale=0.4]{plots/datapos_testacc_zoomfinal.eps}
\end{minipage}
\begin{minipage}[b]{0.5\linewidth}
\centering
\includegraphics[scale=0.4]{plots/datapos_testlikeli_zoomfinal.eps}
\end{minipage}
\caption{\textbf{Comparison of Test Accuracy and Test Likelihood for dataPOS dataset.
The plots in rows 2 and 3 are zoomed versions to clearly see certain
behaviour in the initial and final stages respectively.
} }
\label{dataposfig}
\end{figure*}
\begin{figure*}[ht]
\begin{minipage}[b]{0.5\linewidth}
\centering
\includegraphics[scale=0.4]{plots/wsjpos_testacc.eps}
\end{minipage}
\begin{minipage}[b]{0.5\linewidth}
\centering
\includegraphics[scale=0.4]{plots/wsjpos_testlikeli.eps}
\end{minipage}
\begin{minipage}[b]{0.5\linewidth}
\centering
\includegraphics[scale=0.4]{plots/wsjpos_testacc_zoominitial.eps}
\end{minipage}
\begin{minipage}[b]{0.5\linewidth}
\centering
\includegraphics[scale=0.4]{plots/wsjpos_testlikeli_zoominitial.eps}
\end{minipage}
\begin{minipage}[b]{0.5\linewidth}
\centering
\includegraphics[scale=0.4]{plots/wsjpos_testacc_zoomfinal.eps}
\end{minipage}
\begin{minipage}[b]{0.5\linewidth}
\centering
\includegraphics[scale=0.4]{plots/wsjpos_testlikeli_zoomfinal.eps}
\end{minipage}
\caption{\textbf{Comparison of Test Accuracy and Test Likelihood for WSJPOS dataset.
The plots in rows 2 and 3 are zoomed versions to clearly see certain
behaviour in the initial and final stages respectively.
} }
\label{wsjposfig}
\end{figure*}
\clearpage
\bibliographystyle{apalike}
|
\chapter{Data Analysis} \label{chap:analysis}
The ambition of any good experiment is to transform its data into results that agree with the initial investigative expectations. For B-Machine, as per all other CMB experiments, the goal is to produce angular power spectra, sky maps and cosmological parameters. By selecting the best data and using modern computing skills and tools that are currently available to the CMB community, B-Machine's data has been changed into just that.
\section{Interactive Data Language} \label{subsec:IDL}
The majority of the data analysis and reduction is done using the Interactive Data Language (IDL)\footnote{ITT Visual Information Solutions www.ittvis.com}. IDL has been a key developmental tool for CMB analysis and has thousands of routines and functions written by various experts in the field of astronomy, CMB research, and astrophysics. In addition all maps that are generated in IDL use the HEALpix scheme~\citep{gorski99}. At UCSB all significant data analysis uses codes developed throughout the CMB community and also written locally by Peter Meinhold, Rodrigo Leonardi, John Staren, Mike Seifert, and Brian Williams.
\section{Data Selection}
One hurdle to overcome with large data sets is the selection of data with the appropriate noise levels and signal levels. B-Machine has a set of automated calibrations and was meant to be rotating in azimuth once every $\sim70$ s. If these parameters are not being met then the data needs to be stamped as not in standard operating mode. Further, each channel needs to be checked for weather, DC noise level, and DC levels just to name a few. The manner in which this is done was developed by Rodrigo Leonardi when he was processing the WMPOL (\citep{levy08}) data. It is referred to as the compact data set or just CDS for short. The CDS pipeline is a set of IDL routines that divide the data into $\sim70$ s sections and generate statistics for each section; a description of each binned variable is described in Table~\ref{tab:cdsdatafields}. Only channels 1, 2, 3, and 6 were used for statistics on the AC channels and 9, 10, 11, and 14 for the DC channels.
\begin{table}[p]
\begin{center}
\caption[Description of CDS Variables]{Description of CDS Variables \label{tab:cdsdatafields}}
\begin{tabular}{|l|l|}
\hline
Variable name & Description \\
\hline
\hline
AVERAGEAC & Average value for AC channels.\\
\hline
AVERAGEDC & Average value for DC channels for I.\\
\hline
AVERAGEIR & Average Value of the IR sensor.\\
\hline
CENTERCHANTEMP & Temperature of the central horn.\\
\hline
COLDHEADTEMP & Temperature of the cold head.\\
\hline
FILE & File that 70 s segment came from.\\
\hline
FRAMETEMP & Temperature of frame sensor 10 K/V.\\
\hline
GAIN & RF gain of each channel in K/V.\\
\hline
JULIANDAY & Julian day of first sample in data segment.\\
\hline
MAXDC & Maximum DC value for DC channels.\\
\hline
MEANAZ & Average azimuth reading for data Segment.\\
\hline
MEANEL & Average elevation reading for data Segment.\\
\hline
MINDC & Minimum DC value for DC channels.\\
\hline
NOBS & Number of samples for 70 s data segment.\\
\hline
PEAK2PEAK & Difference of min/max values for AC channels.\\
\hline
POLCALPOSITION & Averaged voltage of Polarization Calibrator.\\
\hline
PRIMARYTEMP & Temperature of the primary mirror 10 K/V.\\
\hline
SIGMAAC & Standard deviation of AC channels.\\
\hline
SIGMAAZ & Standard deviation of azimuth.\\
\hline
SIGMACENTCHANTEMP & Standard deviation of temperature for\\
& the central cryogenic amplifier.\\
\hline
SIGMACOLDHEADTEMP & Standard deviation of cryo head temperature.\\
\hline
SIGMAEL & Standard deviation of elevation.\\
\hline
SIGMAFRAMETEMP & Standard deviation of frame temperature.\\
\hline
SIGMAPRIMARYTEMP & Standard deviation of primary mirror\\
& temperature.\\
\hline
SIGMATILTTEMP & Standard deviation of clinometer temperature.\\
\hline
SIGMATILTX & Standard deviation of tilt in the X direction.\\
\hline
SIGMATILTY & Standard deviation of tilt in the Y direction.\\
\hline
STATUS & Average value of status word.\\
\hline
TILTTEMP & Temperature of clinometer $100^{\circ}\mathrm{C}/\mathrm{V}$.\\
\hline
TILTX & Average X tilt $0.407792 ^{\circ} /\mathrm{V}$ at $15.8^{\circ}\mathrm{C}$.\\
\hline
TILTY & Average Y tilt $0.408163 ^{\circ} /\mathrm{V}$ at$15.6^{\circ}\mathrm{C}$.\\
\hline
TIME & The time of the first and last data points.\\
\hline
WHITENOISEAC & Average white noise value for AC channels.\\
\hline
\end{tabular}
\end{center}
\end{table}
The CDS was generated for all data taken at WMRS, including all calibration runs, Jupiter and Tau A scans as well as all specialized sky scanning strategies (NCP scans and stationary data taking). The final data cutting parameters can be seen in Table~\ref{tab:cdsdataselection}. The basic technique for data cutting is straightforward. A histogram for all data is generated for a set of parameters, some sample histograms can be seen in Figure~\ref{fig:datacuthistograms}. The mean value of the gaussian fit is the cut parameter and the standard deviation of the fit is the unit of measure for cutting. A very conservative cut ($2\sigma$)was used to generate a batch of data labeled level 2 data. The list of parameters that yields the best cutting information with the fewest fields is in Table~\ref{tab:cdsdataselection}. These data were uploaded to the Nersc site where the analysis was performed.
\begin{figure}[p]
\begin{tabular}{cc}
\includegraphics[width=7cm,height=6cm]{fig/sigmaAC6.pdf}&
\includegraphics[width=7cm,height=6cm]{fig/AverageDC6.pdf}\\
\includegraphics[width=7cm,height=6cm]{fig/Peak2Peak6.pdf}&
\includegraphics[width=7cm,height=6cm]{fig/WhiteNoise6.pdf}\\
\end{tabular}
\caption[Example data cut histograms with gaussian fits included]{Example data cut histograms with gaussian fits, all data from the central channel 6/14. Top left, is the standard deviation of Q(black) and U(red). Top right, is the average DC value, the odd shape of this curve represents changing sky temperatures over the campaign. Bottom left, is the peak to peak values of T (blue), Q (black), and U (red). Bottom right, is the white noise level of T (blue) , Q (black) and U (red). \label{fig:datacuthistograms}}
\end{figure}
\begin{table}[p]
\begin{center}
\caption[CDS Data Selection Parameters]{CDS Data Selection Parameters \label{tab:cdsdataselection}}
\begin{tabular}{|l|l|}
\hline
Variable name & Description \\
\hline
\hline
$\mathrm{SigmaAC}\_\mathrm{chan}\#$ & Standard Deviation of AC channels \\
\hline
MinDC & Minimum DC value\\
\hline
MaxDC & Maximum DC value\\
\hline
Peak2Peak & Peak to Peak value of AC channels \\
\hline
Nobs & Number of samples in a given bin \\
\hline
MeanEl & Mean elevation \\
\hline
Status & Average value of status word for nominal operation\\
\hline
$\mathrm{WhiteNoiseAC}\_\mathrm{chan}\#$ & White noise level for AC channels\\
\hline
\end{tabular}
\end{center}
\end{table}
\section{Pointing Reconstruction} \label{sec:pointreconstruct}
Currently the largest problem presented in reconstructing maps from the data collected is accuracy of the pointing reconstruction. The majority of the pointing corrections were found using Moon or Tau A crossings, see Chapter~\ref{chap:instrument} Section~\ref{sec:pointing}. A list of days with Moon crossings is in Table~\ref{tab:mooncrossings}, which also includes the angular diameter and the phase of the Moon in a percentage of full. The phase of the Moon was used during beam characterization to account for the thermal profile of the moon depending on its illumination, see Chapter~\ref{chap:telescopechar} Section~\ref{sec:beamcharacterization}.
\begin{table}[p]
\begin{center}
\caption[List of Moon Crossings]{List of Moon Crossings \label{tab:mooncrossings}}
\begin{tabular}{|c|c|c|c|c|}
\hline
Day & Time of & Time of & Ang Diameter & Percentage \\
& Crossing (UT) & Crossing (UT) & (arcseconds) & of Full\\
\hline
\hline
$8/17/2008$ & 7:56-8:45 & & 1870 & $97\%$\\
\hline
$8/18/2008$ & 7:25-8:35 & & 1890 & $96\%$\\
\hline
$8/19/2008$ & 7:20-8:30 & 11:15-12:10 & 1906 & $91\%$\\
\hline
$8/20/2008$ & 7:40-8:20 & 12:40-13:25 & 1920 & $83\%$\\
\hline
$8/21/2008$ & 8:00-8:50 & 13:55-14:40 & 1930 & $74\%$\\
\hline
$8/22/2008$ & 8:40-9:40 & 15:15-15:50 & 1950 & $64\%$\\
\hline
$8/23/2008$ & 9:20-9:55 & & 1950 & $50\%$\\
\hline
$8/26/2008$ & 12:15-12:55 & & 1960 & $20\%$\\
\hline
$8/27/2008$ & 13:25-14:10 & & 1960 & $10\%$\\
\hline
$8/28/2008$ & 14:35-15:20 & & 1955 & $5\%$\\
\hline
$8/28/2008$ & 15:50-16:15 & & 1940 & $1\%$\\
\hline
$9/14/2008$ & 5:35-8:15 & & 1890 & $100\%$\\
\hline
$9/15/2008$ & 5:35-6:45 & 8:50-9:55 & 1915 & $100\%$\\
\hline
$9/16/2008$ & 5:50-6:40 & 10:35-11:10 & 1930 & $97\%$\\
\hline
$9/18/2008$ & 6:40-7:20 & 13:00-13:35 & 1960 & $86\%$\\
\hline
$9/20/2008$ & 8:10-8:50 & 15:20-16:00 & 1960 & $66\%$\\
\hline
$9/21/2008$ & 9:10-9:45 & 16:20-16:45 & 1960 & $58\%$\\
\hline
$9/24/2008$ & 12:25-13:10 & & 1940 & $22\%$\\
\hline
$9/26/2008$ & 14:45-15:25 & & 1910 & $7\%$\\
\hline
$10/14/2008$ & 9:25-10:10 & & 1940 & $100\%$\\
\hline
\end{tabular}
\end{center}
\end{table}
It now appears that a loose Helical Beam Shaft Coupler (Flex Coupler, see Figure~\ref{fig:flexcoupler}) is the culprit of the pointing problems. The Flex Coupler is used to minimize the torque on the Gurley absolute encoder from shaft misalignment. If the stationary shaft that is secured to the underside of B-Machine is slightly misaligned with the absolute encoder that is affixed to the rotating table, small torque variations will cause unusual wearing in the encoder bearings. This can lead to inaccuracies in the position readout that are gradual and very difficult to characterize. A Flex Coupler is placed between the 2 shafts to maintain precise shaft alignment. After thermally cycling for several weeks of observing and testing the Flex Coupler gradually loosened to a point where it would slip slightly ($\sim1-2^{\circ}$) when starting rotation or changing direction. For nominal observing with Moon crossings the data can be salvaged by reconstructing the centroid of the Moon, using this positional offset to adjust all the data for that day. Days with no Moon crossings are more difficult to reconstruct and further efforts to correct were needed. Each of the days that didn't contain Moon crossings were evaluated for Tau A crossings. Of the 15 days with no Moon crossings 9 of these days were salvaged using Tau A. Comparing the Tau A and Moon crossings gave consistent results to within a beam size for 2 days that contained both crossings.
\begin{figure}[p]
\includegraphics[width=13.5cm]{fig/FlexCoupler.pdf}
\caption[Flex coupler for B-Machines azimuth pointing]{Image of the underside of B-Machines rotating table. Three of the support drive cones are visible in the image. Also barely distinguishable is the Flex Coupler (black) between 2 of the support ribs and criss-crossed by green wires. \label{fig:flexcoupler}}
\end{figure}
\begin{figure}[p]
\begin{tabular}{cc}
\includegraphics[width=7cm,height=6cm]{fig/MoonOffsetsCompare16.pdf}&
\includegraphics[width=7cm,height=6cm]{fig/MoonOffsetsCompare36.pdf}\\
\includegraphics[width=7cm,height=6cm]{fig/MoonOffsetsCompare26.pdf}&
\includegraphics[width=7cm,height=6cm]{fig/MoonOffsetsEl.pdf}\\
\end{tabular}
\caption[Pointing offsets for each channel compared to central channel]{Pointing offsets for each channel compared to the central channel. Top left, is channel 1, Top right is channel 2, Bottom left channel 3, and bottom right is the elevation comparison of all channels. The straight lines in the first 3 plots are the expected offset from horns position compared to central horn. For the elevation comparison the offset was removed. The large spike for channel 1 is a problem in the Moon comparison code due to some small clouds in the data and is not a real artifact of the pointing. \label{fig:moonoffsetcompare}}
\end{figure}
Pointing comparisons, see Figure~\ref{fig:moonoffsetcompare}, of days with Moon crossings show the pointing reconstruction to be reliable. Also, in Figure~\ref{fig:moonoffsetcompare} is a comparison of the elevation pointing which is consistent as well. The corrections for the elevation from the Moon pointing data were sub-beam size and reinforces the notion that the only problem with the pointing data is the Flex Coupler and nothing more complicated.
\section{Point Sources} \label{sec:pointsources}
An effort to observe point sources to confirm calibrations, pointing and facilitate comparisons to other experiments was done several times during the observing campaign. Table~\ref{tab:radiosources} gives a list of possible sources; all the sources for this table use Jupiter as an absolute standard to calibrate all of the other sources and bands. All of the reference information is outside of our band-pass making it necessary to extrapolate up to our frequency of 41.5 GHz using the spectral index quoted in \citet{page07} and
\begin{center}
\begin{equation}\label{eqn:specindex}
S\propto \nu^{\beta},
\end{equation}
\end{center}
where $\beta$ is the spectral index and S is the flux. In addition to temperature calibrations, polarization calibrations are also necessary, but a standard candle in polarization presents fewer opportunities to observe and less accurate data for cross calibrating experiments. Only 2 of the sources in Table ~\ref{tab:radiosources} were observable and sufficiently bright to hope to observe on a single drift scan. Tau A, known more commonly as the Crab Nebula, and Jupiter were pursued. Due to the pointing difficulties only Tau A was observed. It turns out that all of the drift scans for Jupiter were slightly off on azimuth and thus no observations were made.
\begin{table}[p]
\begin{center}
\caption[Radio Source Data]{Radio Source Data \citep{Hafez08,page07}}
\begin{tabular}{|c|c|c|c|c|c|}
\hline
Radio & RA & DEC & Flux Density & Spectral & Freq \\
Source & & & (Jy) & Index & (GHz)\\
\hline
\hline
Cas A & $23^h23^m26^s$ & $58^{\circ}48'$ & 182.0 & $-0.69$ & 33\\
\hline
Cyg A & $19^h59^m28^s$ & $40^{\circ}44'2''$ & 36.4 & $-1.21$ & 33\\
\hline
Tau A (M1) & $ 5^h35^m 4^s$ & $22^{\circ}01'1''$ & 299.2 & $-0.23$ & 40.4\\
\hline
NGC7027 & $21^h 7^m 2^s$ & $42^{\circ}14'1''$ & 5.39 & $-0.119$ & 33\\
\hline
Hydra A & $ 9^h18^m 6^s$ & $-12^{\circ}65'4''$ & 0.127 & $0.19$ & 15\\
\hline
Jupiter & & & 146.6 & $0.248$ & 33\\
\hline
Venus & & & 460.3 & $-0.278$ & 33\\
\hline
Saturn & & & 140.5 & $0.00$ & 33 \\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{table}[p]
\begin{center}
\caption[Radio Source Brightness and Flux Extrapolated to 41.5 GHz]{Radio Source Brightness and Flux Extrapolated to 41.5 GHz \citep{Hafez08} \label{tab:radiosources}}
\begin{tabular}{|c|c|c|c|c|c|}
\hline
Radio & Planet & Flux Density & T Brightness & Effective Ta & Angular\\
Source & & (Jy) & (K) & (mK) & Size ($''$)\\
\hline
\hline
& Jupiter & & 157 & 172 & 42-47\\
\hline
& Venus & & 431.61 & 86.6 & 17\\
\hline
& Saturn & & 140.50 & 71.4 & 12\\
\hline
Cas A & & 155.1 & & 82.5 & 5 \\
\hline
Cyg A & & 27.58 & & 14.67 & 0.03\\
\hline
Tau A & & 281.3 & & 112.7 & 6$'$x4$'$\\
\hline
NGC7023 & & 5.25 & & 2.793 & 18\\
\hline
Hydra A & & 0.154 & & 0.0819 & 0.0001 \\
\hline
\end{tabular}
\end{center}
\end{table}
\subsection{Converting Flux Units to Temperature Units}\label{subsect:flux2temp}
Point source measurements are typically quoted in the radio astronomy flux unit Jansky (Jy). This unit is the total flux incident on a telescope for a given source, and conversion of this flux unit to antenna temperature is instrument dependent and requires knowledge of the telescope for conversion. The flux from a radio telescope can be converted to antenna temperature by,
\begin{center}
\begin{equation}\label{eqn:Flux}
Flux(\frac{W}{m^2\cdot Hz})=\frac{k\times Ta}{Ae},
\end{equation}
\end{center}
where Ae is the effective receiving area, Ta is the antenna temperature, and \textit{k} is Boltzmann's constant. We use
\begin{center}
\begin{equation}\label{eqn:effectivearea}
\lambda^2=Solid Angle \times Ae,
\end{equation}
\end{center}
to get the effective receiving area, where $\lambda$ is the center frequency of the band-pass and the solid angle is the integrated gaussian over the sphere for each horn. Using B-Machine's FWHM of $22.2'\pm0.2'$ for the central horn and $24.0'\pm0.2'$ for the off axis horns (see Section~\ref{sec:beamcharacterization}), we obtain for a $1.0$ mK Ta,
\begin{center}
\begin{equation}\label{eqn:conversionconstant}
Flux(\frac{\mathrm{W}}{\mathrm{m}^2\cdot \mathrm{Hz}})=\frac{1.38\cdot10^{-23}\times 1\cdot 10^{-3}}{1.106 \mathrm{m}^2}=1.248 \cdot 10^{-27}\frac{\mathrm{W}}{\mathrm{m}^2 \cdot \mathrm{Hz}}=1.248 \frac{\mathrm{Jy}}{\mathrm{mK}},
\end{equation}
\end{center}
where $1 \mathrm{ Jy} = 10^{-26}~\frac{\mathrm{W}}{\mathrm{m}^2 \cdot \mathrm{Hz}}$. However, standard nomenclature for radio source flux is to quote total flux, which includes both polarizations, while B-Machine measures only one polarization and calibrates with a blackbody. The conversion is doubled, to get $2.496~\frac{\mathrm{Jy}}{\mathrm{mK}}$ for an unpolarized source from the central horn, see Table~\ref{tab:jytomk} for the off axis horns. As a check to test the reliability of the calculation WMAP data was used. The WMAP~\citep{page07} observations of Tau A show a $\sim72$ mK signal and using the published WMAP instrument characteristics and Equations~\ref{eqn:Flux} and~\ref{eqn:effectivearea} gives $\sim74$ mK.
\begin{table}[p]
\begin{center}
\caption[Jansky to Kelvin Conversion Constants]{Conversion Constants for Jansky's to Kelvin \label{tab:jytomk}}
\begin{tabular}{|c|c|c|}
\hline
Constant& $22.2'$ & $24.0'$ \\
\hline
\hline
Solid Angle & $4.725\cdot10^{-5}$ sr & $5.523\cdot10^{-5}$ sr \\
\hline
Ae & $1.106~\mathrm{m}^{2}$ & $0.946~\mathrm{m}^{2}$ \\
\hline
Ta Radio & $1.25\pm.03~\frac{\mathrm{Jy}}{\mathrm{mK}}$ & $1.450\pm.035~\frac{\mathrm{Jy}}{\mathrm{mK}}$ \\
\hline
Ta & $2.50\pm.06~\frac{\mathrm{Jy}}{\mathrm{mK}}$ & $2.90\pm.07~\frac{\mathrm{Jy}}{\mathrm{mK}}$ \\
\hline
\end{tabular}
\end{center}
\end{table}
\subsection{Tau A}
Tau A was used for pointing reconstruction, calibration, and functionality tests. For pointing reconstruction, Tau A was located in each of the daily temperature maps and the day's pointing was adjusted so that Tau A appeared in the appropriate pixels, see Section~\ref{sec:maps}. For calibration and functionality testing, drift scans were used due to the raster scan limitation. Tau A is sufficiently bright that it can be seen over the $1/f$ noise in the temperature channels with minimal data analysis. There were 5 drift scans at different elevations (5 sections) with each of the scans having 7-10 crossings each. Each crossing that had a threshold voltage above a certain level (different for each channel) was used for pointing calibration. The azimuth and elevation for each point was compared to the expected azimuth and elevation of Tau A and the average difference for all crossings of one drift scan is considered the offset for that scan, see Table~\ref{tab:tauaoffsets}. The azimuth and elevation for Tau A was found by using the time stamp in the level 1 data and the known right ascension and declination to convert to azimuth and elevation.
\begin{table}[p]
\begin{center}
\caption[Tau A Offsets for Each of the 5 Drift Scans.]{Tau A Offsets for Each of the 5 Fixed Elevation Drift Scans \label{tab:tauaoffsets}}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
& \multicolumn{2}{|c|}{Expected}&\multicolumn{2}{|c|}{Actual}&\multicolumn{2}{|c|}{Offset}\\
\hline
Pass & Azimuth & Elevation & Azimuth & Elevation & Co-El & Azimuth \\
& (degrees) & (degrees) & (degrees)& (degrees) & (degrees) & (degrees)\\
\hline
\hline
1 & 79.333 & 23.165 & 81.810 & 23.003 & -2.277 & 0.168 \\
\hline
2 & 84.323 & 30.234 & 86.770 & 30.001 & -2.114 & 0.233 \\
\hline
3 & 93.451 & 42.259 & 95.866 & 42.007 & -1.787 & 0.252 \\
\hline
4 & 100.78 & 50.328 & 103.06 & 49.995 & -1.454 & 0.333 \\
\hline
5 & 110.31 & 58.347 & 112.48 & 58.008 & -1.136 & 0.340\\
\hline
\end{tabular}
\end{center}
\end{table}
In Figures~\ref{fig:tauatempmaps},~\ref{fig:tauaqmaps}, and~\ref{fig:tauaumaps} maps of T, Q, and U respectively are presented. The temperature maps are consistent with the expected temperatures from the flux measurements from the WMAP data, with the exception of channel 3. Channel 3 has been consistently problematic, with much higher noise levels and calibration constants that are a factor of 2 higher then the other 3 channels. Channel 3 has a high gate current indicating a leaky gate which contributes significantly to its higher noise temperature. An additional note on the temperature measurement is that the noise levels are dominated by the $\frac{1}{f}$ noise from each channel. B-Machine was designed as a polarimeter and it is surprising that Tau A is visible in the temperature maps. The $Q$ and $U$ maps both consistently show more signal than expected. The full result of the Tau A drift scans are in Table~\ref{tab:tauatemps}. There are 2 sources of error contributing to these figures. First, the pixelization scheme changes the beam size in a non-gaussian fashion, which in turn causes the conversion from flux to temperature units to differ slightly from the derived formula. Secondly, the beam shift as a function of position of the Polarization Rotator is contributing a polarized signal that would not otherwise be present. As the telescope scans across the source the beam is shifting in elevation, and each sector of the rotator sees a slightly different part of the sky. For example, if the horizontal sector is centered on Tau A then the vertical sector won't be, causing an erroneous polarization signal. This effect is only apparent due to the fact that a point source is being observed.
\begin{figure}[p]
\begin{tabular}{cc}
\includegraphics[width=7cm,height=6cm]{fig/chan1Tthesis.pdf}&
\includegraphics[width=7cm,height=6cm]{fig/chan2Tthesis.pdf}\\
\includegraphics[width=7cm,height=6cm]{fig/chan3Tthesis.pdf}&
\includegraphics[width=7cm,height=6cm]{fig/chan6Tthesis.pdf}\\
\end{tabular}
\caption[Tau A temperature maps nside=512]{Tau A temperature maps with nside=512. Top left, channel 1, Top right channel 2, Bottom left channel 3, and bottom right channel 6. Each map is centered on Tau A (RA: $83.6332^{\circ}$ DEC: $22.015^{\circ}$) and represents $28.27$ square degrees. The peak temperatures are $111.0\pm11.0\mathrm{~mK}$, $91.9\pm10.5\mathrm{~mK}$, $134.4\pm15.0\mathrm{~mK}$ and $100.0\pm6.0\mathrm{~mK}$, respectively. \label{fig:tauatempmaps}}
\end{figure}
\begin{figure}[p]
\begin{tabular}{cc}
\includegraphics[width=\textwidth]{fig/SumTThesis.pdf}
\end{tabular}
\caption[Combined Tau A temperature map]{Tau A temperature map with nside=512. All channels were combined using sigma weighting. The map is centered on Tau A (RA: $83.6332^{\circ}$ DEC: $22.015^{\circ}$) and represents $28.27$ square degrees. The centroid temperature of $99.8\pm4.0\mathrm{~mK}$. \label{fig:tauatempsummap}}
\end{figure}
\begin{figure}[p]
\begin{tabular}{cc}
\includegraphics[width=7cm,height=6cm]{fig/chan1Qthesis.pdf}&
\includegraphics[width=7cm,height=6cm]{fig/chan2Qthesis.pdf}\\
\includegraphics[width=7cm,height=6cm]{fig/chan3Qthesis.pdf}&
\includegraphics[width=7cm,height=6cm]{fig/chan6qthesis.pdf}\\
\end{tabular}
\caption[Tau A $Q$ maps]{Tau A $Q$ maps with nside=512. Top left, channel 1, Top right channel 2, Bottom left channel 3, and bottom right channel 6. Each map is centered on Tau A (RA: $83.6332^{\circ}$ DEC: $22.015^{\circ}$) and represents $28.27$ square degrees. The peak temperatures are $-12.6\pm6.0\mathrm{~mK}$, $-8.5\pm4.0\mathrm{~mK}$, $-18.0\pm22.0\mathrm{~mK}$ and $9.8\pm1.2\mathrm{~mK}$, respectively.\label{fig:tauaqmaps}}
\end{figure}
\begin{figure}[p]
\begin{tabular}{cc}
\includegraphics[width=\textwidth]{fig/SumMapQThesis.pdf}
\end{tabular}
\caption[Combined Tau A $Q$ map]{Tau A $Q$ map with nside=512. All channels were combined using sigma weighting. The map is centered on Tau A (RA: $83.6332^{\circ}$ DEC: $22.015^{\circ}$) and represents $28.27$ square degrees, with a centroid $Q$ of $-8.2\pm3.0\mathrm{~mK}$. \label{fig:tauaqsummap}}
\end{figure}
\begin{figure}[p]
\begin{tabular}{cc}
\includegraphics[width=7cm,height=6cm]{fig/chan1Uthesis.pdf}&
\includegraphics[width=7cm,height=6cm]{fig/chan2Uthesis.pdf}\\
\includegraphics[width=7cm,height=6cm]{fig/chan3Uthesis.pdf}&
\includegraphics[width=7cm,height=6cm]{fig/chan6UThesis.pdf}\\
\end{tabular}
\caption[Tau A $U$ maps]{Tau A $U$ maps with nside=512. Top left, channel 1, Top right channel 2, Bottom left channel 3, and bottom right channel 6. Each map is centered on Tau A (RA: $83.6332^{\circ}$ DEC: $22.015^{\circ}$) and represents $28.27$ square degrees. The peak temperatures are $12.4\pm6.0\mathrm{~mK}$, $12.5\pm4.0\mathrm{~mK}$, $15.0\pm22.0\mathrm{~mK}$ and $-8.5\pm1.2\mathrm{~mK}$, respectively.\label{fig:tauaumaps}}
\end{figure}
\begin{figure}[p]
\begin{tabular}{cc}
\includegraphics[width=\textwidth]{fig/SumMapUThesis.pdf}
\end{tabular}
\caption[Combined Tau A $U$ map]{Tau A $U$ map with nside=512. All channels were combined using sigma weighting. The map is centered on Tau A (RA: $83.6332^{\circ}$ DEC: $22.015^{\circ}$) and represents $28.27$ square degrees, with a centroid $U$ of $7.1\pm3.0\mathrm{~mK}$. \label{fig:tauausummap}}
\end{figure}
\begin{figure}[p]
\begin{tabular}{cc}
\includegraphics[width=7cm,height=6cm]{fig/chan1nobsThesis.pdf}&
\includegraphics[width=7cm,height=6cm]{fig/chan2nobsthesis.pdf}\\
\includegraphics[width=7cm,height=6cm]{fig/chan3nobsthesis.pdf}&
\includegraphics[width=7cm,height=6cm]{fig/chan6nobsThesis.pdf}\\
\end{tabular}
\caption[Tau A number of samples per bin maps]{Tau A number of samples per bin maps. Top left, channel 1, top right channel 2, bottom left channel 3, and bottom right channel 6. Each map is centered on Tau A (RA: $83.6332^{\circ}$ DEC: $22.015^{\circ}$) and represents $28.27$ square degrees. The average samples per bin are 62.7 (1.88 s), 61.7 (1.85 s), 61.9 (1.85 s) and 63.0 (1.88 s), respectively. The regions with low counts have been excluded from the averages. \label{fig:tauanobsmap}}
\end{figure}
The pixel size for the maps (nside=512) was selected based on beam shift, number of pixels per beam ($\sim3$), and total number of pixels. This last criterion was necessary due to computational limits on memory usage imposed by IDL.
\begin{table}[p]
\begin{center}
\caption[Tau A Stokes Parameters at 41.5 GHz]{Tau A Stokes Parameters at 41.5 GHz \label{tab:tauatemps}}
\begin{tabular}{|c|c|c|c|}
\hline
Channel & Flux Density & Temperature & Observed Temperature \\
& (Jy) & (mK) & (mK) \\
\hline
\hline
\multicolumn{4}{|c|}{Temperature}\\
\hline
1 & $281\pm6$ & $98\pm6$ & $111.0\pm11$ \\
\hline
2 & $281\pm6$ & $98\pm6$ & $103.4\pm10.5$ \\
\hline
3 & $281\pm6$ & $98\pm6$ & $124.9\pm15.0$ \\
\hline
6 & $281\pm6$ & $110\pm6$ & $95.2\pm6.0$ \\
\hline
Sum & & $101\pm3$ & $99.8\pm4.6$ \\
\hline
\multicolumn{4}{|c|}{Q Stokes parameter}\\
\hline
1 & $-18.5\pm3.0$ & $-6.7\pm1.3$ & $-10.8\pm6.0$ \\
\hline
2 & $-18.5\pm3.0$ & $-6.7\pm1.3$ & $-6.8\pm4.0$ \\
\hline
3 & $-18.5\pm3.0$ & $-6.7\pm1.3$ & $-14.7\pm22.0$ \\
\hline
6 &$ -18.5\pm3.0$ & $-7.4\pm1.3$ & $-9.4\pm1.2$ \\
\hline
Sum & & $-6.9\pm0.7$ & $-8.2\pm3.0$\\
\hline
\multicolumn{4}{|c|}{U Stokes parameter}\\
\hline
1 & $0.5\pm3.0$ & $0.2\pm1.1$ & $16.0\pm6.0$ \\
\hline
2 & $0.5\pm3.0$ & $0.2\pm1.1$ & $15.8\pm4.0$ \\
\hline
3 & $0.5\pm3.0$ & $0.2\pm1.1$ & $11.2\pm22.0$ \\
\hline
6 & $0.5\pm3.0$ & $0.2\pm1.1$ & $8.1\pm1.2$ \\
\hline
Sum & & $0.2\pm0.6$ & $7.1\pm3.0$ \\
\hline
\end{tabular}
\end{center}
\end{table}
\section{Maps} \label{sec:maps}
B-Machine spent 81 days at White Mountain from arrival with the experiment in the truck to final shut down of the facility due to weather and funding constraints. Of the 81 possible days, 40 of these were used for observing. Of the remaining 41 days 8 were used for the initial reconstruction and calibration of B-Machine, 12 of the days B-Machine was down due to mechanical issues, and the final balance was from weather. Of the 40 observing days 4 were used for a dedicated NCP scan which was not usable due to pointing issues and 29 were salvaged from the pointing debacle, see Section~\ref{sec:pointreconstruct}, for map generation. A full observing day consisted of 14 hours, with the time being limited by the Sun's height in the sky. During the final shut down sequence the telescope was allowed to operate during the day to test if Sun crossings would damage the instrument. After inspecting the telescope after a Sun crossing it was determined that operation during day light hours was possible, though thermal cycling would certainly lead to data analysis issues.
Of the 29 days of usable data each channel for each day was cut based on several values generated out of the CDS arrays. In Table~\ref{tab:sparameters} the final selection parameters for each of the channels can be seen.
\begin{table}[p]
\begin{center}
\caption[CDS Data Selection Values]{CDS Data Selection Values, Values are in Volts Unless Otherwise Stated \label{tab:sparameters}}
\begin{tabular}{|c|c|c|c|c|}
\hline
Selection Field & Channel & I & Q & U \\
\hline
\hline
AC Sigma & & & & \\
\hline
& 1 & 0.11 & 0.021 & 0.022 \\
\hline
& 2 & 0.11 & 0.0186 & 0.0203 \\
\hline
& 3 & 0.11 & 0.0156 & 0.0164 \\
\hline
& 6 & 0.11 & 0.00918 & 0.00892 \\
\hline
Peak to Peak & & & & \\
\hline
& 1 & 0.72 & 0.149 & 0.161 \\
\hline
& 2 & 0.78 & 0.142 & 0.155 \\
\hline
& 3 & 0.70 & 0.124 & 0.126 \\
\hline
& 6 & 0.30 & 0.0701 & 0.0680 \\
\hline
White Noise & & & & \\
\hline
& 1 & 0.012 & 0.00336 & 0.00366 \\
\hline
& 2 & 0.013 & 0.00322 & 0.00352 \\
\hline
& 3 & 0.010 & 0.00266 & 0.00281 \\
\hline
& 6 & 0.004 & 0.00154 & 0.00150 \\
\hline
Min DC Value & & & & \\
\hline
& 1 & -0.825 & & \\
\hline
& 2 & -0.922 & & \\
\hline
& 3 & -0.638 & & \\
\hline
& 6 & -0.829 & & \\
\hline
Max DC Value & & & & \\
\hline
& 1 & -0.627 & & \\
\hline
& 2 & -0.703 & & \\
\hline
& 3 & -0.473 & & \\
\hline
& 6 & -0.665 & & \\
\hline
Number of Samples & All & 3000 & counts & \\
\hline
Status & All & 35 & & \\
\hline
Max Elevation & All & 46 & degrees & \\
\hline
\end{tabular}
\end{center}
\end{table}
The final values were determined by using 2 sigma cuts on gaussian fits to the data histograms. Bulk cuts were made based on Healpix position. These cuts can be seen on the maps as rectangular cut outs on the leading and trailing edges of the maps. These cuts correspond to the start and stop of data for each day and are consistent with thermal cycling of optical components. In Figures~\ref{fig:tempmapall} through~\ref{fig:tauazoommap} maps of all 3 Stokes parameters can be seen with all channels and days combined.
\begin{figure}[p]
\begin{center}
\includegraphics[width=9.5cm ]{fig/TempMapAllThesis.pdf}
\caption[Temperature sky map]{Temperature sky map for all days and channels. The three circles are observed point sources Cygnus A, Tau A, and M42 from the top down. The Galaxy can be seen as a blue line on the right side of the figure. The map represents a $53.07\%$ sky coverage, with an average integration time of 20.9 seconds per pixel and is in ecliptic coordinates.\label{fig:tempmapall}}
\end{center}
\end{figure}
\begin{figure}[p]
\begin{center}
\includegraphics[width=10cm ]{fig/QMapAll.pdf}
\caption[Q sky map]{Q sky map. This map is featureless and represents a white noise dominated signal. \label{fig:Qmapall}}
\end{center}
\end{figure}
\begin{figure}[p]
\begin{center}
\includegraphics[width=10cm ]{fig/UMapAll.pdf}
\caption[U sky map]{U sky map. This map is featureless and represents a white noise dominated signal. \label{fig:Umapall}}
\end{center}
\end{figure}
\begin{figure}[p]
\begin{center}
\includegraphics[width=10cm ]{fig/NobsMapAll.pdf}
\caption[Number of observations per bin sky map]{Number of observations per bin map. The features at the top and bottom represent the edges of scans.\label{fig:nobsmapall}}
\end{center}
\end{figure}
\begin{figure}[p]
\begin{tabular}{cc}
\includegraphics[width=7cm,height=6cm]{fig/TauAZoomTemp.pdf}&
\includegraphics[width=7cm,height=6cm]{fig/TauAZoomQ.pdf}\\
\includegraphics[width=7cm,height=6cm]{fig/TauAZoomU.pdf}&
\includegraphics[width=7cm,height=6cm]{fig/TauAZoomNobs.pdf}\\
\end{tabular}
\caption[Zoom in of Tau A region of full map]{Zoom in of Tau A region of full maps for all Stokes parameters and number of sampled bins. Top left, Temperature, top right, $Q$, bottom left, $U$, bottom right, number of binned samples. The small "x" is the position of Tau A.\label{fig:tauazoommap}}
\end{figure}
\section{Angular Power Spectrum}
Preliminary angular power spectra are consistent with noise dominated measurements. For accurate power spectra de-stripping codes, such as Madam 3.5, need to be used to remove the obvious stripping in the current temperature maps, see Figure~\ref{fig:tempmapall}. Included in Figures~\ref{fig:TTPowerSpec} through Figures~\ref{fig:TEPowerSpec} are power spectra for TT, EE, and TE. In addition, power spectra for a random white noise map with standard deviation scaled to that of the data set for the given Stokes parameter are shown. A Jack knife difference was taken for first order instrument noise removal from the maps. The process of Jack knifing the data consisted of splitting the data set into 2 roughly equal length data sets, subtracting the maps, and generating power spectra from the differenced map.
\begin{figure}[p]
\begin{center}
\includegraphics[width=\textwidth ]{fig/ThesisTTPowerSpec.pdf}
\caption[TT Angular Power Spectrum]{TT Angular Power Spectrum with the three main point sources and the galaxy removed before spectrum generation. Black is Jack knifed data, blue is B-Machines entire data set, red is the differenced data, and green is a white noise spectrum.\label{fig:TTPowerSpec}}
\end{center}
\end{figure}
\begin{figure}[p]
\begin{center}
\includegraphics[width=\textwidth ]{fig/ThesisEEPowerSpec.pdf}
\caption[EE Angular Power Spectrum]{EE Angular Power Spectrum in black and a white noise spectrum in blue. The differenced Jack knifed data is in green and consistent with no cosmological signal. \label{fig:EEPowerSpec}}
\end{center}
\end{figure}
\begin{figure}[p]
\begin{center}
\includegraphics[width=\textwidth ]{fig/ThesisTEPowerSpec.pdf}
\caption[TE Angular Power Spectrum]{TE Angular Power Spectrum in black and a white noise spectrum in blue. The differenced Jack knifed data is in green and consistent with no cosmological signal. \label{fig:TEPowerSpec}}
\end{center}
\end{figure}
The TT spectrum differs from a white noise spectrum for 2 main reasons. First, its noise is dominated by $\frac{1}{f}$ noise and secondly the maps are not de-stripped. This accounts for the slope difference of the 2 curves in Figure~\ref{fig:TTPowerSpec}. The polarization map is consistent with a white noise dominated map as expected by our sensitivity given the limited integration time and large sky coverage, see Chapter~\ref{chap:telescopechar} Table~\ref{tab:MaxPhaseTsys}. A first attempt to estimate the power spectrum uncertainties was made using an analytic tool developed by Lloyd Knox \citep{knox96}. Plots of the three interesting power spectra have been generated with error bars in Figures~\ref{fig:TTPowerSpecErr} through ~\ref{fig:TEPowerSpecErr}. Using the errors from the analytical model a goodness of fit test reveals that the TE power spectrum is consistent with the null hypothesis, reduced chi square is $0.65$.
\begin{figure}[p]
\begin{center}
\includegraphics[width=\textwidth ]{fig/ThesisTTPowerSpecerr.pdf}
\caption[TT Angular Power Spectrum with estimate errors]{TT Angular Power Spectrum including estimates of errors. The shape of the power spectrum is clearly indicative of a white noise dominated power spectrum.\label{fig:TTPowerSpecErr}}
\end{center}
\end{figure}
\begin{figure}[p]
\begin{center}
\includegraphics[width=\textwidth]{fig/ThesisEEPowerSpecerr.pdf}
\caption[EE Angular Power Spectrum with estimate errors]{EE Angular Power Spectrum including estimates of error. The shape of the power spectrum is clearly indicative of a white noise dominated power spectrum. \label{fig:EEPowerSpecErr}}
\end{center}
\end{figure}
\begin{figure}[p]
\begin{center}
\includegraphics[width=\textwidth]{fig/ThesisTEPowerSpecerr.pdf}
\caption[TE Angular Power Spectrum with estimate errors]{TE Angular Power Spectrum including estimates of errors. This curve is consistent with no cosmological signal.\label{fig:TEPowerSpecErr}}
\end{center}
\end{figure}
\chapter{Blackbody Temperature to Antenna Temperature} \label{app:blackbody}
Blackbody radiation refers to an idealized object or system which absorbs all radiation incident upon it and re-radiates the energy which is characteristic of the radiating system only, not dependent upon the type of radiation which is incident upon it, see Figure~\ref{fig:Blackbody}. The brightness of radiation from a black body is given by Planck's law,
\begin{equation}
B_{\nu}(T) =\frac{2h\nu^3}{c^2}\frac{1}{\mathrm{e}^{h\nu /kT}-1}
\end{equation}
where $h$ is Planck's constant, $\nu$ is the frequency, $c$ is the speed of light, $k$ is Boltzmann's constant, $T$ is the physical temperature and B$_{\nu}$ is the surface brightness in Watts $\mathrm{m}^{-2}$ $\mathrm{Hz}^{-1}$ $\mathrm{sr}^{-1}$. Typically a transfer standard (gain/calibration) is generated based on the input flux of a telescope versus output voltage level using beam filling objects that radiate as black bodies at different temperatures, see Chapter~\ref{chap:telescopechar} Section~\ref{Sec:Calibration}.
\begin{figure}[p]
\begin{center}
\includegraphics[width = 13.5cm]{fig/Blackbody.pdf}
\caption[Curves of blackbody radiators at different temperatures]{Curves of blackbody radiators at different temperatures retrieved may 28, 2009 from $\mathrm{http://en.wikipedia.org} / \mathrm{wiki} / \mathrm{Black\_body}$. \label{fig:Blackbody}}
\end{center}
\end{figure}
A brightness temperature is what is strived for when observing, but antenna temperature is what is measured. Antenna temperature is the convolution of the brightness temperature with the beam of the instrument and is given by,
\begin{equation}
T_A =\frac{1}{\Omega_A}\int \int T_s(\theta, \phi)P_n(\theta,\phi)\mathrm{d}\Omega,
\end{equation}
where $T_s(\theta,\phi)$ is the source temperature and $P_n(\theta,\phi)$ is the normalized antenna pattern. For very small sources with uniform temperatures this reduces to,
\begin{equation}
T_A =\frac{\Omega_s}{\Omega_A}T_s,
\end{equation}
where $\Omega_s$ and $\Omega_A$ are the source and antenna areas respectively. For large sources that fill the beam (but are small compared to the entire sky) and an antenna pattern that does not have significant side lobe contribution then
\begin{equation}
T_A \simeq T_s.
\end{equation}
\chapter{Conclusion} \label{chap:conclusion}
Though B-Machine didn't yield grand results, it did achieve the goals it set out to meet. B-Machine was fielded and operated in the summer of 2008 at White Mountain Research Station for 2 main reasons: first, to test the polarization rotator chopping strategy and second, to collect data to generate maps and power spectra of the CMB polarization. Jack knife analysis of the the current sky maps show that the map noise integrates down with added integration time and no features become apparent to the tenths of mK level. This demonstrates that no major systematics are polluting the data stream in an odd fashion. Moreover, rebinning of the maps into nside=64 reduces the RMS noise of the map from $1.0\mathrm{~mK}$ to $0.20\mathrm{~mK}$ increasing the sensitivity at low l's at the price of high l sensitivity, yielding yet another avenue for continued analysis. It seems clear that the B-Machine platform was effective as both a test platform for larger experiments (arrays of multiple telescopes for B-mode observations and balloon borne experiments for foreground observations) and a polarimeter for observing the CMB.
Though several problems were encounter when fielding the telescope these problems are easy to solve for the next observing campaign and the next year of data promises to yield a minimum of twice as much data. The addition of point source observations will give a very rich data set for future analysis. While future observations are certainly not guaranteed for B-Machine, it is clear at frequencies below 50 GHz ground based observations can provide excellent results on par with that of satellites, see Table~\ref{tab:compare}. With a very stable $\frac{1}{f}$ knee frequency, B-Machine is a gold mine of data waiting to be taken advantage of. Drawing your attention to row 12 of Table~\ref{tab:compare}, it becomes clear that with a minimal financial and man power investment low frequency observations can be made from the ground and compete with satellites. The reduction in the $f_{\mathrm{knee}}$ frequency in Table~\ref{tab:compare} is done with a linear fit to the data before differencing and is a data analysis technique. In addition, the numbers quoted in Table~\ref{tab:compare} don't take into account that B-Machine style radiometers have a fundamental advantage over both WMAP and Planck-LFI style radiometers in that they are designed as polarimeters and measure $Q$ and $U$ directly through the same RF path.
\begin{table}[p]
\begin{center}
\caption[Comparison of WMAP, Planck, Future B-Machine, and COFE]{Comparison of B-Machine, WMAP \citep{jarosikwmap03}, Planck-LFI \citep{meinhold09}, Future B-Machine, and Cofe \citep{leonardi05} \label{tab:compare}}
\begin{tabular}{|l|c|c|c|c|c|}
\hline
& B-Machine & WMAP & Planck & Future & COFE \\
& & Satellite & LFI & B-machine & \\
\hline
\hline
Center Freq. (GHz) & 41.5 & 41.0 & 44.0 & 40.0 & 20 \\
\hline
T$_{sys}$ (K) & 54 & 59 & 30 & 30 & 10 \\
\hline
T$_{sky}$ (K) & 16 & 2.7 & 2.7 & 16 & 2.7 \\
\hline
$f_{\mathrm{knee}}$ (mHz) & 5 & 4 & 30 & 2.8 & 2.8 \\
\hline
Number of Detectors & 4 & 4 & 8 & 8 & 6\\
\hline
Angular Res. (arcmin) & 22.2 & 30.6 & 30 & 22.2 & 42.0\\
\hline
$\triangle$T (mK $\sqrt{\mathrm{s}}$) & 1.8 & 0.90 & 0.439 & 0.80 & 0.318\\
\hline
$\triangle$P (mK $\sqrt{\mathrm{s}}$) & 1.37 & 0.90 & 0.439 & 0.61 & 0.227\\
\hline
Aggr. $\triangle$P (mK $\sqrt{\mathrm{s}}$) & 0.685 & 0.450 & 0.155 & 0.215 & 0.094\\
\hline
\multicolumn{6}{|c|} {$\sim1\sigma$ sensitivities for 7 months of observations$^{\star}$}\\
\hline
Observing Efficiency & 0.375 & 1.00 & 1.00 & 0.50 & 0.50\\
\hline
Aggr. $\triangle$P /Pixel ($\mu K$)$^{\dag}$& 452.0$^{\ddag}$ & 67.0 & 25.7 & 48.0 & 17.0 \\
\hline
Sky coverage ($\%$) & 53.1 & 100 & 100 & 60 & 60 \\
\hline
\multicolumn{6}{l}{$~~\dag$ Pixel Size $0.5^{\circ} \mathrm{~x~} 0.5^{\circ}$}\\
\multicolumn{6}{l}{$~~\ddag$ From actual B-Machine data set of 40 days}\\
\multicolumn{6}{l}{$~~\star$ For similar observations B-Machine achieves 165 $\mu K$ Aggr. $\triangle$P Per Pixel }\\
\end{tabular}
\end{center}
\end{table}
\section{The Future}
With some minor contributions in time and money B-Machine could get a data set that would rival even that of the Planck Satellite Mission. By changing the current front end LNA's with more modern lower noise LNA's and populate the remaining 4 horns with the same LNA's, the current system temperature would be reduced by a factor of 2, doubling the integration time. B-Machine has the capability to have 8 feed horns and is currently equipped with the hardware and software to run all 8 horns. Drop in replacements for the data acquisition boards would allow for faster DAQ rates reducing the addition of $\frac{1}{f}$ noise from the LNA's and the addition of a 5 point calibration sequence on a monthly basis would improve the performance of the telescope. The sensitivity for all of these upgrades can be seen in the Future B-Machine column of Table~\ref{tab:compare}.
\chapter{Description of the B-Machine Instrument}\label{chap:instrument}
The B-Machine telescope was designed to test a new technique in CMB polarization detection (see Section~\ref{chap:polarizationrotator}) and to measure CMB polarization from a previously established site (White Mountain Research Station, Barcroft, henceforth referred to as WMRS). The construction of the telescope has been an on going process for the last several years. Each of the telescopes subsystems was constructed and tested at UCSB prior to full integration and deployment to WMRS. The majority of the work constructing the telescope was performed by me, with general design and construction help coming from lab personnel including Peter Meinhold, Jared Martinez, Hugh O'Neil, and Andrew Riley. There are also a handful of undergraduates and others that deserve some thanks and a list of them can be found in the acknowledgements section.
\section{Telescope}\label{sec:telescope}
\begin{figure}[p]
\begin{center}
\includegraphics[width = \textwidth]{fig/b_machine_full.pdf}
\end{center}
\caption[B-Machine line drawing]{Line Drawing of B-Machine telescope including Polarization Rotator, dewar, optics, and table. \label{fig:bmachinedrawing}}
\end{figure}
\begin{figure}[p]
\begin{center}
\includegraphics[width = \textwidth]{fig/b_machine_opticsandbeam.pdf}
\end{center}
\caption[Optical design of B-Machine telescope]{Optical design of B-Machine Telescope. The parabolic primary reflector, plane mirror Polarization Rotator, ellipsoidal secondary reflector, and the dewar are shown. Radiation from the sky is focused by the primary onto the common focal point of the primary and secondary where the plane mirror Polarization Rotator is located and then focused by the secondary to the phase center of the central corrugated feed horn.\label{fig:opticaldesign}}
\end{figure}
\subsection{Optical Design}
B-Machine is a modified off-axis Gregorian telescope with a reflecting half wave plate polarization modulator at the confocal point. This design is a slight modification of the BEAST and WMPOL optical design \citep{childers05,figueiredo05, meinhold05, mejia05, odwyer05}. The optics consist of a primary $2.2$ m off-axis parabolic reflector, a $0.9$ m ellipsoidal secondary reflector and a reflecting polarization modulator as seen in Figure~\ref{fig:opticaldesign}. The telescope meets the Dragone-Mizugutch condition for minimal cross-polarization contamination and maximum focal plane area \citep{Dragone78,Mizugutch76}.
\begin{figure}[p]
\includegraphics[width = \textwidth]{fig/skydipchan10.pdf}
\caption[Cross polarization isolation using sky dip]{Sky dip with $\mathrm{T}_{\mathrm{sys}}$ removed, black line is the temperature change as a function of azimuth ($90^{\circ}$ corresponds to zenith), blue and red lines are $Q$ and $U$, respectively. The slope of $Q$ and $U$ is over 20 dB lower than that of $T$ indicating a isolation of $>20$ dB. The Zenith sky temperature calculated from these data is 10 K, consistent with atmospheric models of the site. \label{fig:crosspolarization}}
\end{figure}
The Dragone-Mizugutch condition for an off-axis Gregorian telescope is shown to be
\begin{equation}\tan{\alpha} = \left(\frac{e+1}{1-e}\right)
\tan{\beta},
\end{equation}
by \citet{Dragone78}, where the parameters $\alpha$, $\beta$, and $e$ can be found in Table~\ref{tab:optics}. The Reflecting Polarization Modulator is positioned at the confocal point of the primary and secondary reflectors. The optical system is folded about this point so that the overall beam paths have the same lengths as the Beast \citep{childers05} and WMPOL \citep{levy08} optical design.
\begin{table}[p]
\begin{center}
\caption[Optics Parameters]{Design Parameters of B-Machine Optics \label{tab:optics}}
\begin{tabular}{lc}
\hline\hline\\
Parameter & Value\\\\ \hline \\
Primary focal length (mm) & 1250.0 \\
Primary max. physical dimension (mm) & 2200.0 \\
Primary min. physical dimension (mm) & 1966.1 \\\\
\hline \\
Secondary semimajor axis (mm) & 886.7 \\
Secondary semiminor axis (mm) & 853.4 \\
Secondary focal length (mm) & 240.7 \\
Secondary eccentricity, e & 0.2714 \\\\
\hline\\
Feed angle, $2\alpha$ (degrees) & 58.2 \\
Angle between axes, $2\beta$ (degrees) & 35.4 \\\\
\hline
\end{tabular}
\end{center}
\end{table}
The same mirrors were used on both the Beast telescope and the B-Machine telescope and were stored in a large crate between observing campaigns. When unpacked from storage there was a large blemish in the middle of the primary reflector, as seen in Figure~\ref{fig:PrimaryBlemish}. It is believed that a small amount of water sat on the mirror and degraded the AL surface during storage. The mirrors were installed and used for testing in spite of the damage. While testing, a large polarization offset was observed when viewing the sky. The blemish appeared to be the cause of the offset and the mirrors were sent to Surface Optics Incorporated and resurfaced. The polarization offset was in fact eliminated following the re-coat of the mirrors.
\begin{figure}[p]
\begin{center}
\includegraphics[width = \textwidth]{fig/Primary_blemish_PreRecoat.pdf}
\caption[Blemish on primary mirror from water evaporation]{Large discoloration in the middle of the primary mirror is oxidation of the Aluminum layer from water that was allowed to pool during long term storage. This blemish caused an $\sim5$ K polarized offset. \label{fig:PrimaryBlemish}}
\end{center}
\end{figure}
\subsection{Table}
The mount for B-Machine is a rotating table with a stationary base, the majority of the instrumentation is mounted on the top of the rotating section of the table. The Azimuth drive system for the table consists of a Galil, planetary drive, motor, relative encoder, slip ring, cone bearings, absolute encoder, and control software. The table was retrofitted from a drive cone system \citep{levy08} to a gear reducing direct drive system. A DC motor\footnote{Amtek 40 V} (relative encoder attached to base of motor) was attached via belt to the lower drive cog of the planetary drive gearbox\footnote{Sipco Mechanical Linkage 105}. The upper side of the planetary drive gearbox, which drives the main pulley (19.05") and gives a 127:1 gear reduction between the output of the motor and the table rotation rate. With this gearing the table can rotate anywhere between 3 rotations per minute down to $0.1$ rotations per minute (see Chapter~\ref{chap:telescopechar} Section~\ref{subsec:scanstrategy} for more details). The system is controlled by a servo code originally written for WMPOL and modified, by Marcus Ansman, for use with B-Machine. The code communicates with a multi-axis motion controller\footnote{Galil DMC-2140}, that uses feedback from the relative encoders and the absolute encoders\footnote{Gurley Precision Instruments A25S 16-Bit} to move the table in both azimuth and elevation. The servo computer (the computer that controls the motion, as opposed to the DAQ computer which collects the scientific data) needs constant contact with the Galil for precise motion of the telescope. This communication channel is achieved through a wireless router which enables the moving servo computer to communicate with the stationary Galil (the Galil is mounted to the bottom of the stationary section of the table). The Galil is unable to source sufficient current to power the motors for direct motion control so a Linear Servo Amplifier\footnote{Western Servo Design Inc. LDH-A1-4/15} is used on both axes enabling a small control voltage, $\pm10$ V, to control the high current motors. The elevation drive system is similar in that the same Galil is used, and an absolute and relative encoder are used in tandem to control its motion. The elevation drive uses a linear actuator to drive the experiment up and down. The elevation drive is not used during normal operation; it is fixed (bolted down) to minimize jitter in this axis.
To properly control and power the experiment it is important that the stationary base of the experiment can communicate with the rotating platform. Though wireless communications are the most sensible, they are not suited for higher current or constant voltage applications. Since, it is necessary to route the power for the entire telescope through the connection a slip ring assembly was chosen. The slip ring allows 12 connections between the moving and stationary platforms. Of the 12 possible connections only 10 were used, 5 for the 220/120 V AC power and the remaining 5 for the elevation encoder. These lines consisted of the A incremental encoder phase, 5 V power from Galil to the relative encoder, Galil ground, and $\pm$ Control lines for the elevation linear amplifier.
\subsection{Leveling}
When the experiment was installed into the dome care was taken to level the experiment. Each of the 3 corners of the table rests on 2 6" Aluminum I beams, on top of each of the I beams is multiple thicknesses of shim material (thin pieces of brass) for fine adjustment of the height. A 3 ft long bubble level was used for a rough level of the table and a clinometer\footnote{Applied Geomechanics Inc. Model 904-TH} with a $\pm10^{\circ}$ range was used to level the experiment to operational tolerances, see Figures~\ref{fig:xtilt} and~\ref{fig:ytilt}. The clinometer was read throughout the observing campaign, with each servo file containing X tilt , Y tilt and the temperature of the clinometer. When initially inspecting the clinometer data at UCSB it was found that no small variations in the signal level could be seen due to the input noise level of the data acquisition board. To solve this the signal was run into a times 10.76 amplifying board and then routed into the servo computer.
When inspecting the experiment after reassembly and precursory testing at WMRS it was noticed that 2 of the support/bearing cones between the moving and stationary parts of the table were not always in contact. While adjusting the cones it was found that 3 of the 4 bolts that hold the table top to the drive system were broken, sheared in half, presumably from the 20 miles of unpaved road that the experiment was shipped over. After removing and replacing all of the broken bolts, the level of the experiment was rechecked. The level hadn't changed significantly from previous measurements, but it was re-leveled again as a precaution, see Table~\ref{tab:levelingext}.
\begin{table}[p]
\begin{center}
\caption[Leveling Extremes Before and After Re-leveling]{Leveling Extremes Before and After Re-leveling \label{tab:levelingext}}
\begin{tabular}{|c|c|c|c|}
\hline
Tilt Axis & Min & Max & Total Deviation\\
\hline
Y before & -4.24' & 3.40' & 7.64'\\
\hline
Y After & -2.95' & 3.06' & 6.01'\\
\hline
Y Stationary & -0.07' & 0.06'& 0.13'\\
\hline
X before & -3.56' & 3.57' & 7.13'\\
\hline
X After & -3.23' & 2.76' & 5.99'\\
\hline
X Stationary & -0.07' & 0.06' & 0.13'\\
\hline
\end{tabular}
\end{center}
\end{table}
Leveling monitored through out the campaign shows no significant change in the overall leveling of the instrument from the beginning to the end of observations.
\begin{figure}[p]
\includegraphics[width = \textwidth]{fig/xtiltcrop.pdf}
\caption[X tilt]{Tilt readings from the X axis of the clinometer. Black is the tilt of B-Machine just after fixing of the drive system, red is the tilt after adjusting slightly (current tilt) and blue is tilt measurements with the experiment stationary binned into $360^{\circ}$ bins. Both black and red are averaged over 10 rotations and binned into $1^{\circ}$ sections. The stationary data set is looking at $\sim31^{\circ}$ azimuth and binned into 360 bins for ease of plotting (number of samples per bin similar for all curves).\label{fig:xtilt}}
\end{figure}
\begin{figure}[p]
\includegraphics[width = \textwidth]{fig/ytilt.pdf}
\caption[Y tilt]{Tilt readings from the Y axis of the clinometer. Black is the tilt of B-machine just after fixing of the drive system, red is the tilt after adjusting slightly (current tilt) and blue is tilt measurements with the experiment stationary binned into $360^{\circ}$ bins. Both black and red are averaged over 10 rotations and binned into $1^{\circ}$ sections. The stationary data set is looking at $\sim31^{\circ}$ azimuth and binned into 360 bins for ease of plotting (number of samples per bin similar for all curves).\label{fig:ytilt}}
\end{figure}
\subsection{Pointing} \label{sec:pointing}
Determination of the pointing of the experiment is achieved through the use of 2 16-bit absolute encoders\footnote{Gurely Precision Instruments Model A25S} and precise leveling of the experiment. In addition to the pointing encoders several other pointing checks were used, Moon crossings, CCD images, and point sources. The Moon was observed once or twice a day (depending on the day) for about 50$\%$ of the observing days. Each moon crossing was used to align the beam by fitting all of the data to a moon model and taking topological corrections for the position of the experiment on the surface of the earth. A CCD \footnote{Electrophysics Corp. Model WAT-902} equipped with a motorized zoom lens\footnote{Computer Model V10Z1618} and aligned with the beam, captured images of star patterns sporadically during the data campaign. Upon further inspection the CCD images were inconsistent enough that the pointing corrections found from these images were not used. The final pointing evaluation tool was to find bright point sources and correct using the known position of the source. The only source bright enough to see real time was Tau A (Crab Nebula). Evaluating each day for a Tau A crossing and adjusting pointing gave 9 days of additional data to the Moon data. When comparing days with both Tau A and Moon crossings the pointing was consistent to within a beam size. Due to some mechanical difficulties the Moon and Tau A crossings were used for all of the pointing reconstruction. See Chapter~\ref{chap:analysis} Section~\ref{sec:pointreconstruct} for an in depth discussion of the pointing reconstruction.
\subsection{Data Acquisition}
To keep up with the required data rates B-Machine uses 2 separate data acquisition computers: one computer to collect the scientific data (called the DAQ computer) and another computer for housekeeping data and servo control (called the servo computer). Housekeeping is a generic term used to describe the various pieces of information that are needed to turn the science data into useful information. The servo computer uses a PCI based board\footnote{Measurement Computing Corporation model PCI-DAS6402/16} to read position, tilt, temperatures, time, cryogenic temperatures, and status of gain and calibrator. In tandem with this board the servo computer also incorporates a PCI-DIO24 board to read in a 24 bit synchronization number. The DAQ computer reads in 10 channels of scientific data, synchronization number and time using a USB based board\footnote{IOtech model DaqBoard/3005USB}. Of the remaining 6 science channels 2 are used for the Thermopile and polarization calibrator and 4 are blank though functional.
To recombine the complimentary data sets a synchronization number is generated by using the index pulse from the Polarization Rotator encoder to count each revolution. Each line of the data for both computers has a synchronization number associated with it allowing for recombination of the data sets at a later time. It is essential that the synchronization number is unique on multi day time scales. Given our sample rate of $33.4\mathrm{~Hz}$ and 16 hours of data per day the number will repeat itself every 6 days, giving sufficient time to avoid errors in the recombination process.
\section{Radiometer}\label{sec:radiometer}
\begin{figure}[p]
\includegraphics[width = 13.5cm]{fig/Radiometeroutline.pdf}
\caption[Radiometer outline drawing]{Outline drawing of a single radiometer chain. Radiation enters from the right through a microwave transparent window and is coupled into the first LNA by a corrugated feed horn. Low loss coaxial cable carries the signal from the cryogenic LNA to the room temperature LNA's (front end to back end amplifiers). The signal is then passed through a band-pass filter and rectified using a square law diode. Directly attached to the diode is an IF (Intermediate Frequency) amplifier where it is split off into an AC coupled and DC signal. Each signal is passed through a 1.7 kHz low pass filter before entering the A/D converter. \label{fig:radiometeroutline}}
\end{figure}
Much of the main guts (basic wiring and overall structure) of the radiometer were salvaged and reconstituted from the BEAST experiment \citep{childers05}. Very few of the BEAST RF chains survived the punishment of time and static discharge to be used in B-Machine, but the basic signal path is the same.
The microwave signal enters at a sealed window, a low loss extruded polystyrene material that provides a vacuum seal and a first layer of infrared blocking. Between the window and the $20$ K detector array lies a multilayered IR blocking window. This window consists of several (10-20) thin sheets of low loss extruded polystyrene material, each layer is separated by a small gap and attached to a cooling shield. This allows each layer to run slightly cooler than the one above it enabling the horn array to view an IR source that is significantly colder than ambient temperature. Without IR blocking the thermal radiative load would cause the horns to run significantly warmer than the rest of the array giving rise to thermal load dependent signal which could fluctuate significantly from day to day or hour to hour. The corrugated scalar feed horns and the front end amplifiers are housed in a cryogenically cooled dewar and kept at $\sim20$ K. The RF signal comes out of the vacuum vessel through low loss coaxial cable where it is further amplified, filtered, and rectified by temperature regulated back end RF chains. It then ultimately is saved via analog to digital (A/D) conversion on hard disk for post processing.
\subsection{Feed Horns}
As shown in Figure~\ref{fig:radiometeroutline}, conical corrugated scalar feed horns \citep{villa97} couple the microwave radiation from the sky to the telescope. Figure~\ref{fig:hornretloss} shows the return loss of a horn from data taken on a Vector Network Analyzer (VNA)\footnote{HP8510C Vector Network Analyzer from 45 MHz to 50 GHz}. Designed specifically for CMB experiments the full details of the horn design and testing can be found in \citep{villa98,villa97}.
\begin{figure}[p]
\begin{center}
\includegraphics[width = 13.5cm]{fig/retloss.pdf}
\end{center}
\caption[Measured corrugated feed horn return loss (S11) ]{Measured input return loss (S11) of a Q-band feed horn normalized so that 0 dB is all power reflected. \label{fig:hornretloss}}
\end{figure}
\subsection{Amplifiers}
The detector array on B-Machine is equipped with 5 Low Noise Amplifiers (LNA's) that use FET (field effect transistor) technology, 3 of which are microwave integrated circuits (MIC) based and the 2 remaining are monolithic microwave integrated circuits (MMIC). A MMIC has the majority of the bias network coalesced into one small chip, as opposed to the MIC which requires additional bias electronics and tuning. The 2 different types of circuits can be seen in Figure \ref{fig:MMICVsMIC} in a single package. Each of the LNA stages has approximately 25 dB of gain and has been optimized for the lowest noise temperature in our band-pass (38-45 GHz). Each LNA is followed by a back end module that consists of several room temperature MMIC's, with roughly 60 dB of gain, a band-pass filter and a detector diode. With the exception of the detector diode all of the RF chains were designed, assembled, tuned, and tested at UCSB by either Jeff Childers or myself.
The 2 MMIC based amplifiers were test chips from JPL wafer runs and when testing a stability problem associated with a feedback capacitor on the gate of the first stage was found. The problem is outside of our band pass and presents some minor stability problems for our measurements, but a new wafer run of the chip was too far off to try and implement any new designs. Each of the LNA's is cooled to $\sim20$ K using a CTI Cryogenic Cryodyne refrigerator\footnote{Helic Company Model SC Compressor and Model 350CP Cold Head}. The noise temperature of these LNA's drops approximately linearly with temperature from 300 K to $\sim20$ K, see Figure~\ref{fig:NTVstemp}, greatly reducing the effective integration time necessary to achieve desired error bars. Further cooling yields little to no improvement on the noise temperature and requires significant work implementing.
\begin{figure}[p]
\includegraphics[width = 13.5cm]{fig/45ln1ENtVsTemp.pdf}
\caption[45LN1 noise temperature at ambient and 20 K]{The noise temperature of a 45LN1 MMIC chip at ambient temperature left axis in red and 20 K right axis in blue. Chip fabricated at HRL using a Sandy Weinreb design. Amplifier assembled and tested at UCSB. \label{fig:NTVstemp}}
\end{figure}
The LNA's output is coupled via stainless steel coax (for thermal isolation) and copper coax to a back end amplification block. The coax cable provides both thermal and RF isolation between the front end and back end amplifiers and carries the signal with minimal loss to the room temperature section of RF gain, see Table~\ref{tab:backends}. Each individual component in the back end modules is attached together using a gold plated brass carrier that allows for the replacement and testing of the individual components prior to assembly. Though each back end is much more massive than is necessary, the mass provides thermal stability for the unit. All of the back ends are bolted to a large Aluminum plate that is insulated and thermally regulated. Thermal regulation is achieved through a temperature sensor feed back loop that uses an AD590 temperature sensor and power resistors for heating. The back ends are set to run at 305 K, to stay in thermal regulation during both night and day cycling. Three of the 5 back ends used were from the BEAST experiment with only minor alterations and the additional 2 were the same design but complete rebuilds. The three amp chains from BEAST needed tuning and gain adjustment before use.
\begin{table}[p]
\begin{center}
\caption[Back End Amplifier Blocks]{Components of Back End Amplifier Blocks \label{tab:backends}}
\begin{tabular}{|c|c|c|c|c|c|}
\hline
Channel & Amp 1 & Amp 2 & Amp 3 & Filter (GHz) & Diode\\
\hline
\hline
0 & 44LNA1$\_$80 & ALH192C & HMMC-5040 & 38-45 & 75KC50\\
\hline
1 & 44LNA1$\_$80 & ALH192C & HMMC-5040 & 38-45 & 75KC50\\
\hline
2 & ALH244 & ALH244 & ALH244 & 37-44 & 75KC50\\
\hline
3 & ALH376 & ALH386 & ALH386 & 37-44 & 75KC50\\
\hline
6 & 44LNA1$\_$80 & ALH192C & HMMC-5040 & 38-45 & 75KC50\\
\hline
\end{tabular}
\end{center}
\end{table}
With the exception of the 44LNA1$\_$80 all of the back end chips are commercially available. The 44LNA1$\_$80 was a test chip from JPL and is no longer made due to the availability of mass produced devices. The band-pass for all of the amplifier chains is quite large and it is necessary to define the band-pass with an external filter. The first round of filters (38-45 GHz) from BEAST had a band-pass that was dictated by the minimal in the noise temperature of the front end amplifiers. With the second pass of filters shifted down slightly in frequency to avoid the $45\mathrm{~GHz}$ oxygen line in the atmosphere, see Subsection~\ref{subsect:filters} for complete design details.
\begin{figure}[p]
\includegraphics[width = 13.5cm]{fig/qbe8.pdf}
\caption[Gain profile of typical back end]{Back end gain profile for two back ends used in B-Machine. Neither back end contained a band-pass filter and a 10 dB attenuator was attached to the output of the back end to avoid damaging the test equipment. \label{fig:qbe8}}
\end{figure}
\begin{figure}[p]
\includegraphics[width = 13.5cm]{fig/MMICVsMIC.pdf}
\caption[MIC and MMIC amplifier]{A hybrid test amplifier that contains both MIC amplifiers (left side) and MMIC amplifier (right side). On the MIC side the discrete amplifiers can be seen with the bias and bias protection circuits running in from both sides, the top is the drain side and the bottom is the gate side. The 2 MICs in the circuit give $\sim10$ dB of gain while the MMIC gives $\sim25$ dB of gain. \label{fig:MMICVsMIC}}
\end{figure}
\subsection{Filters}\label{subsect:filters}
Each RF chain has significant gain both above and below the desired band-pass. As a result a coupled line filter was produced to significantly shrink the band-pass of the instrument. The filters reflect all of the out of band power, hence, attenuation is necessary between the output of the final amplification stage and the input of the filter. For every 1 dB of attenuation, 2 dB of isolation is gained. There is typically 6-8 dB of attenuation between the 2 elements and the S22 parameter for the output amps is typically around 10 dB giving an isolation of at least 20 dB. The technique to produce the filters was developed jointly by Jeff Childers, Alan Levy and myself.
The filters are first modeled and optimized for the desired band in Libra/Eesof (a software tool that is the pre-cursor to ADS RF EDA software). Libra supplied the dimensions for the inductive fins and the spacing which was then used to draw a template in AUTOCAD, Figure~\ref{fig:FilterDrawing}. From the template a positive photo mask was produced\footnote{CadArt Services Poway, CA}. Standard substrate material consisting of $1/2$ oz. copper separated from the bottom $2$ oz copper layer by 5 mils of Quflon (Teflon) was etched using standard photo-lithography and Ferric Chloride etching techniques and packaged in a generic housing for use in the back end amplifier blocks.
Early in the fabrication development processing it was found that the bandwidth of the filters shifted slightly up in frequency and narrowed, shown in Figure~\ref{fig:Filterresponse}. This effect was seen uniformly throughout many different filter designs and was compensated for to achieve the desired band-pass for the filters.
\begin{figure}[p]
\includegraphics[width = 13.5cm]{fig/B35_44Adwg.pdf}
\caption[Filter mask outline]{Mask drawing of band-pass filter with simulated response from 35-45 GHz and tested response of 37-44 GHz . The large leads on either side of the filter are meant to couple the line to 50 ohm transmission line or glass bead feedthru. The filter is $0.5$ inches long by $\sim0.1$ inches wide. \label{fig:FilterDrawing}}
\end{figure}
\begin{figure}[p]
\includegraphics[width = 13.5cm]{fig/FilterResponseThesis.pdf}
\caption[Band-pass filter response]{Blue (S11) and red (S21) lines are measured responses from the fabricated filter and the black line is the simulated response from Libra. The discrepancy was known before hand and planned on for final filter results.\label{fig:Filterresponse}}
\end{figure}
\subsection{Data Input}\label{Subsect:datainput}
\begin{figure}[p]
\includegraphics[width = 13.5cm]{fig/DAQIFoutline.pdf}
\caption[DAQ layout drawing]{Simplified layout drawing of the data acquisition system components from the end of the RF gain section (square law detector diode) to the input of the computer (16-bit A/D converter). The pre-amp is attached to the detector diode directly and connected to the remainder of the electronics (located in a shielded Mu Metal box) via coaxial cabling. \label{fig:DAQIFoutline}}
\end{figure}
Following the band-pass filter the signal is rectified using a linear rectifying diode\footnote{Anritsu Company 75K50 Microwave Detector Diode}, see Section~\ref{Sec:Calibration} for response characteristics, that converts the RF power incident on it to a proportional DC voltage. All gain after the diode is below 1 MHz and referred to as audio amplification. First, in line is a 20 times audio amplifier that is connected directly to the diode; this configuration reduces sensitivity to systematic noise from external sources such as ground loops and radiation from wireless devices. The first stage audio gain also provides the appropriate loading for the diode (600 ohms). Following the first stage the signal is routed to the shielded Planck Acquisition box, called this because it was originally used for testing prototype cards for the Planck Satellite Mission. The signal into the Planck box goes to an audio amplifier (x10), from here the signal is split into 2 paths, see Figure~\ref{fig:DAQIFoutline}. The upper path is referred to as the DC signal, since the voltage is proportional to the absolute power that hits the diode. The alternate path is first AC coupled, taking any offset voltages out, and ran through another adjustable amplifier gain stage, and then to an opto-isolator out of the box. Each of the front end LNA's is associated with 2 data channels, a DC channel and an AC coupled channel (primary science data). The AC coupled channel has a switchable gain setting on it that can be switch between x1, x10, x50, and x100. From the output of the Planck Acquisition Box each channel is run through a 1.7 kHz low pass filter and then to the input of the data acquisition board\footnote{Iotech, Inc. DaqBoard/3005USB} on the DAQ computer.
Each of the DC channels is used to calibrate the corresponding AC channel, see Chapter~\ref{chap:telescopechar} Section ~\ref{Sec:Calibration}. A complete understanding of the gain difference (from the adjustable gain amp) between the AC and the DC channels is critical for the proper calibration of the system. To measure the gain of the system each channel had a small (0.00052 V) sine wave input to the x20 pre-amp. Data were taken for approximately one minute for each of the gain settings. The sine wave data for each channel was fit using IDL. The amplitude from the fit data for each sine wave was divided by the input fit for the next gain level. For each of the channels there are 3 possible gain divisions x10$/$x1, x50$/$x10, and x100$/$x50. In the x1 setting there is also the preceding audio gain of $\sim$x200, this is referred to as the front-end gain here. Also, the DC channels were compared to the AC x1 gain channels to confirm no gain difference between AC and DC channels with this setting.
\begin{table}[p]
\begin{center}
\caption[IF Gain Measurements]{IF Gain Measurements~\label{tab:ifgain}}
\begin{tabular}{|c|c|c|c|c|c|}
\hline
AC & Front-end & x10 & x50 & x100 & Total Gain \\
channel & & & & & at x100 \\
\hline
\hline
Channel 1 & 199.112 & 9.983 & 50.247 & 100.827 & 20,075.87 \\
\hline
Channel 2 & 196.588 & 9.957 & 50.334 & 100.782 & 19,812.53 \\
\hline
Channel 3 & 195.244 & 9.958 & 49.901 & 100.312 & 19,585.32 \\
\hline
Channel 6 & 195.825 & 9.991 & 50.413 & 100.648 & 19,709.39 \\
\hline
\end{tabular}
\end{center}
\end{table}
In the standard operating mode the adjustable gain switch is set to x100. The total gain is the gain from the output of the rectifying diode to the input of the IOTech data acquisition board. The x100 column in Table~\ref{tab:ifgain} is the multiplication factor for calibration between the AC and the DC channels in standard observing mode.
\section{Electronics}
The telescope runs on the back of multiple electronic subsystems, ranging from simple power distribution to the digital 24-bit synchronization number.
\subsection{Power Distribution}
\begin{figure}[p]
\includegraphics[width = 13.5cm]{fig/PowerSchematic.pdf}
\caption[Schematic of power distribution]{Layout drawing of the power distribution of B-Machine. Input power is sourced from an independent power grid that gets its power from a battery bank that is charged by either a diesel generator or a bank of solar panels. \label{fig:PowerSchematic}}
\end{figure}
All telescope power is filtered by a Ferro Resonant Uninterrupted Power System\footnote{Ferrups, Eaton Corp. Model FE2.1 kVA}. This system uses 4 12 Volt 35 AH batteries as backup power, allowing sufficient time to shut down all critical systems before a hard shutdown occurs in case of power outages. The input power coming from the station is sourced from either a solar/battery powered inverter or a diesel generator. The power was switched between these two sources depending on the battery status, load, and weather conditions. Management of the station power was a split effort between telescope personnel and WMRS facility employees.
Each day started with the facility staff checking the status of the battery bank first thing in the morning (around 6 am). At this point if the battery status was above $90\%$ the generator would be switched off and power would be sourced from the solar/battery inverters. With the generator off several criteria needed to be evaluated every couple of hours. First, were the solar panels getting direct sunlight, if not what state were the batteries in. Second, how long had the facility been on batteries and under what kind of load. This depended on the number of people that were using the WMRS facilities at any given time. The generator was typically left off until about 6 pm. At this point the generator would be turned on for a couple of hours for night time showering and dinner. Following dinner the generator was turned off and it was the telescope personnel's responsibility to turn the generator on at night before bed. At any point if the charge level of the batteries reached the low $60\%$ level the generator was immediately turned on. The station's power requirements have been greatly reduced by upgrading existing systems to more power friendly devices. Also, the crew will typically conserve power when possible by shutting off overhead lighting and turning off the water heater during solar/battery operation. Even with all of this the power failed due to low battery levels several times. The telescope power is very stable and is buffered by multiple pieces of equipment before getting to the experiment. This arrangement worked well and I must thank the WMRS staff in doing a job that wasn't there responsibility.
Power from the generator or solar/battery arrays is run from the main Pace building out to the telescope buildings via 240 V lines. The power comes into the building and goes directly into the UPS and is split into 1 240 V line and 6 120 V lines. The 240 V AC power that is routed through the slip ring goes directly to the CTI cooler and is not used for anything else. This line has its own ground and the cooler is isolated from the rest of the experiment. One of the 120 V AC lines from the UPS is routed through the slip ring and powers all of the systems on the moving part of the table. There are 7 linear power supplies that convert the 120 V AC power to DC power for the assorted electronics (see sections below). The stationary part of the table uses several of the 120 V lines from the ferro resonant power system for power; this includes the Galil, all azimuth drive systems, and the wireless router that is used for Galil communications.
\subsection{Amplifier Bias}
There are 2 basic amplifier bias schemes used on B-Machine, one for each class of amplifier (MIC and MMIC). Each of the front end LNA's with discrete amplifiers needs separate bias lines for each FET. The circuit that was used was developed at UCSB by Jeff Cook. This circuit allows for 2 different biasing possibilities, constant current or constant voltage. The normal operational state of the bias is the constant current mode. In this mode the drain voltage and gate current is set to the desired points and the gate voltage is servoed until the drain current is within an appropriate tolerance of the set point. In the time ordered data power spectrum a broad bump can be seen around 800 Hz. This bump is the servo rate of the gate control circuit. In addition to biasing, this board also provides some over voltage and over current protection; it limits the maximum and minimum values that the amplifiers can be supplied to, 1.75 V on the drain, 10 mA drain current or $\pm0.4$ V on the gates. These boards are not suited to bias the MMIC's due to the current requirements of the larger chips.
To bias the MMICs, 2 front end chips and all of the back end modules, a constant voltage bias scheme was used. The front end MMICs use a board that is the same bias design as the back end boards with minor modifications for voltage and current readouts. Using the pin outs from the front end MIC bias boards, a front end MMIC bias board was developed that supplied the appropriate readouts to the front panel BNC connectors, see Figure~\ref{fig:MMICBiasSchematic}. Each back end board is directly attached to the back end module, as shown in Figure~\ref{fig:backendmodule}. The back end bias boards provided 3 important elements, bias power, voltage protection, and sequencing. The voltage protection on the gates is provided by diode protection and ensures that any potentially dangerous spikes in the voltages are not conducted to the gates but rather shunted to ground and limits the gate voltages to $\pm0.4$ V. Drain/gate sequencing is required for MMICs in general, otherwise they have the possibility of burning out on power cycling. If the gate voltage is not present when the drain voltage comes on the device can draw excessive amounts of current and burn out the FETs on the chip.
\begin{center}
\begin{figure}[p]
\centering
\includegraphics[width = 19cm, height=5.5cm, angle=90]{fig/BiasSchematic.pdf}
\caption[Schematic of MMIC bias circuit]{Schematic of front end MMIC Bias board. Originally designed for room temperature back end bias and modified to have readout circuitry compatible with MIC bias board pin outs.\label{fig:MMICBiasSchematic}}
\end{figure}
\end{center}
\begin{center}
\begin{figure}[p]
\centering
\includegraphics[width = \textwidth]{fig/backendmodule.pdf}
\caption[Picture of back end module]{Picture of back end module. The RF input is on the left and the diode output is on the right. Stood off above the RF components is the bias board. \label{fig:backendmodule}}
\end{figure}
\end{center}
\subsection{24-bit Synchronization Number}
One of the most important subsystems in B-Machine is the synchronization number (sync number), which allows the data from the 2 different computers, DAQ and servo, to be recombined. Unfortunately, in my haste to get the telescope working there is very little documentation for this board. I will attempt to outline the basic operation of this board and include the fundamental elements (IC chips) that are used.
The essential components of the board are 3 8-bit binary counters with output registers (54LS590\footnote{Texas Instruments}) and a hex inverter chip (74LS04~\footnote{Motorola}) which contains 6 independent inverters. The binary counter chips are cascaded in series such that the previous chip triggers the following chip. The input to the first counter is the index pulse that is generated either from the polarization modulator encoder or the encoder eliminator board, see subsection~\ref{subsec:encodereliminator}. It is necessary to invert the index pulse before it is feed into the first of the 8-bit binary counter chips. The 8-bit chips are divided into 2 4-bit counters, and each 4-bit counter uses the previous binary word to trigger it. From here it is just as one would expect, skipping the gory details of getting the chips to cascade appropriately, with all of the 4-bit words in series being read out by each of the data acquisition boards. One feature of the chip is a clock clear pin; this allows for the entire 24-bit sync number to be reset to zero with the use of a small button built into the board. By pulling the pin high ($+5$ V) the entire word is reset.
\subsection{Encoder Eliminator}\label{subsec:encodereliminator}
On occasion it was necessary to run the experiment without the Polarization Rotator running; this only occurred during testing. For this reason an additional board was installed on the experiment to simulate the encoder output. A Programmable Crystal Oscillator (PXO-600\footnote{Statek Co.}) was used to generate a square wave at either 10 Hz or 30 Hz. The square wave signal was split into 2 waves with one of them getting a 1/4 wave phase shift. The unshifted signal was the A phase and the shifted signal was the B phase of the encoder output. An additional counter chip was used to count 128 pulses of the A phase and send out an index pulse. This setup closely mimics the output of the encoder on the Polarization Rotator encoder.
\subsection{Temperature Sensors}
The temperature is monitored by one of 3 systems. All cryogenic temperature sensors use a biased silicon temperature sensor\footnote{Lake Shore, Inc.} that gets a constant $10$ $\mu$A. The voltage across the diode is temperature dependent and readout via the servo computer. Ambient temperature sensors are mounted to several components on the experiment which include the calibrator, primary and secondary mirrors, polarization rotator, tilt sensor, and frame temperature. These sensors use an AD590\footnote{Analog Devices, Inc.} two-terminal IC temperature transducer that has been calibrated prior to use for temperature readout. The third and final temperature sensors are active and utilize an AD590 for the temperature readout in tandem with a set of power resistors mounted for the heating elements. A control voltage can be set to raise the temperature of an insulated system above ambient.
\begin{figure}[p]
\begin{center}
\includegraphics[width = 13.5cm]{fig/Ambienttemperatures.pdf}
\caption[Ambient temperatures for multiple sensors over one observing day]{Ambient temperatures for multiple sensors over one observing day. Temperatures binned into 1 minute averages for primary reflector (black), secondary reflector (red), frame (blue), calibrator (light blue), and Polarization Rotator (purple). \label{fig:AmbientTemperatures}}
\end{center}
\end{figure}
\chapter{Introduction}
From the beginning of human history man has looked to the sky for answers about our origins. Why are we here? Where did we come from? These questions have for the most part been in the realm of philosophy and religion. With the evolution of Cosmology, science can start to address some of these questions. Plato and his student Aristotle created cosmologies to search for higher meaning, focusing on the Earth/Sun system. Their ideas were exemplified by the Ptolemaic Earth centered system which dominated western thinking for over 2000 years. Not until the sixteenth century did a different Heliocentric Copernican school of thought emerge. At the time this was ground breaking not just changing the position of the Earth and Sun, but also declaring that the Earth was not the center of the Universe. During this time period observational Astronomy was beginning to blossom. Galileo Galilei turned his telescope to the sky and saw multiple satellites orbiting Jupiter, the roughness of the Moons surface and sunspots. These observations directly challenged standard dogma. Tycho Brahe was gathering unprecedented measurements of the heavens, though he still believed in an Earth centered Universe, his data eventually led to Kepler's discovery of elliptical orbits and descriptions of planetary motion. By the 18th century basic foundations of gravity and physics were being laid down by Newton, Euler, and Laplace. This marked the beginning of a truly cosmological type of thinking expanding our Universe to the edges of our Galaxy. For the first time, the Universe was more than just the Earth/Sun system. Herschel presented evidence of a vast network of stars that laid between 2 planes and stretched out a large distance and proposed a method to find our location in this stratum of stars. Man has moved from the center of the Universe to some unspecified position on the edge of our Galaxy amongst many galaxies in a vast sea of space. This is really the start of empirical Astronomy and Cosmology with the advancements in photometry and spectroscopy, chemical properties of celestial objects could be found. Though many of the conclusions were questionable until the early twentieth century, Cosmology truly separated from philosophy into an observational science. Like the Universe our understanding of it had an inflationary epoch from 1915 to 1930. The size of the Universe in human understanding increased exponentially from our galaxy to a possibly infinite space and time Universe. Hubble's discovery that everything was moving away from us in every direction and our new understanding of energy, matter, gravity, space and time, from Einstein, made it possible to realize that in the distant past things were densely packed. The Universe was packed so close that everything was all in one place and was followed by a BIG BANG.
\section{Hot Big Bang}
The Hot Big Bang or standard cosmological model consists of a homogeneous and isotropic Universe whose development is described by the Friedman equations derived from Einstein's field equations of gravitation (General Relativity).
\begin{eqnarray}\label{eqn:FriedmanEquations}
\mathrm{H}^{2}=\left(\frac{\dot{a}}{a}\right)^{2}= \frac{8 \pi \mathrm{G}}{3} \rho -\frac{k c^{2}}{a^{2}} + \frac{\Lambda c^{2}}{3} \nonumber\\
\dot{\mathrm{H}}+ \mathrm{H}^{2}=\frac{\ddot{a}}{a}= - \frac{4\pi \mathrm{G}}{3}\left( \rho + \frac{3p}{c^{2}} \right) + \frac{\Lambda c^{2}}{3},
\end{eqnarray}
where $a$ is the expansion parameter, the dot represents a derivative with respect to proper time $\tau$, G is the gravitation constant, $\rho$ is the mass density, p is the pressure, c is the speed of light in a vacuum, k the curvature parameter and H is the Hubble parameter. The Universes expansion is parameterized by the Hubble constant, $H_0$, where $v = H_0 d$ gives the relationship between the recessional velocity, $v$, of a galaxy and its distance, $d$, from Earth. $H_0$ is the Hubble parameter now and has been measured by the ``Hubble Space Telescope Key Project'' to have a value of $72\pm8$~km~s$^{-1}$~Mpc$^{-1}$ \citep{freedman01}.
\begin{figure}[p]
\begin{center}
\includegraphics[width =\textwidth]{fig/ThermalHistoryUniverse.pdf}
\caption[Radiation temperature and density history of the Universe]{Thermal and density history of the Universe. CMB observations originate from the end of the period when the Universe was a plasma. The top set of lines are the energy densities, matter $\rho_{m}$ and radiation $\rho_{r}$. The lowest line is the CMB temperate and has only one major feature from the slight increase in temperature when the Universe experienced reheating from electron positron annihilation. \label{fig:ThermalHistory}}
\end{center}
\end{figure}
The Big Bang did not occur at a single point in space but rather simultaneously everywhere in the Universe. Our understanding of the Universe starts shortly after the big bang. Before the Planck time, $10^{-43}$ s, General Relativity needs to be modified to take into account quantum corrections which become significant at these scales. From the Heisenberg uncertainty principle in the form
\begin{equation}
\Delta E \Delta t \simeq \hbar,
\end{equation}
it can be seen that at very early times the energy levels would require masses or energies of a black hole. Post Planck time as the Universe cooled, neutrinos decoupled from the primordial plasma, particles froze out, and dark matter began to coalesce. At about 10 s the Universe experienced a brief moment of reheating when the temperature dropped below the threshold energy of the electrons and positrons which began to annihilate releasing energy. As the Universe continued to expand, matter began to dominate the energy density. The cooling continued until nucleosytheis created nuclei up to Lithium and Beryllium, all other heavier elements had to wait millions of years till the formation of the first stars and their eventual death in supernova. When the temperature was sufficiently low the electrons combined with nuclei and created neutral Hydrogen or Helium. With a neutral Universe the mean free path of the typical photon increased to longer than the size of the Universe. The transition from a photon-baryon fluid to a neutral gas marked the time of last scattering of the primordial photons.
\subsection{Surface of Last Scattering}
\begin{figure}[p]
\includegraphics[width=\textwidth]{fig/CobePowerSpectrum.pdf}
\caption[COBE CMB temperature spectrum]{COBE temperature spectrum showing a blackbody spectrum with peak radiation consistent with a 2.72 K blackbody. Errors are smaller than the width of the line. Source: lambda.gsfc.nasa.gov. \label{fig:CobeBlackbody}}
\end{figure}
The surface of last scattering was the last time most of the primordial photons directly scattered off of matter, embedding information about this time into the remanent photon field known as the Cosmic Microwave Background (CMB). The CMB was first discovered by Penzias and Wilson in 1965 \citep{penzias65} and was found to be a uniform blackbody over the entire sky (see Appendix~\ref{app:blackbody} for a brief explanation of blackbody temperature and antenna temperature), Figure~\ref{fig:CobeBlackbody}. The initial radiation field has cooled to the point where the radiation is now in the microwave bands ($\sim$2.72 K). Not until the launch of the COBE satellite \citep{cobe92} where variations found in the background temperature, with the best measurements, to date, of the non-uniformities coming from the WMAP (Wilkinson Microwave Anisotropy Probe \citep{bennett03a}) space mission.
\begin{figure}[p]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=\textwidth]{fig/cobeSkydipole.pdf}\\
\includegraphics[width=\textwidth]{fig/cobeSKy.pdf}\\
\end{tabular}
\caption[COBE sky temperature map]{Top, picture of the dipole as seen by COBE where red is hotter and blue is cooler where the magnitude of the dipole is $~3.353\pm0.024\mathrm{~mK}$. Bottom, picture of the fluctuations in the mean CMB temperature with the dipole and galactic foregrounds removed. Source: lambda.gsfc.nasa.gov \label{fig:CobeWmapSky}}
\end{center}
\end{figure}
\begin{figure}[p]
\includegraphics[width=\textwidth]{fig/WMAPSky.pdf}
\caption[WMAP CMB temperature spectrum]{The CMB fluctuations as seen by the WMAP space mission \citep{bennett03a} which also measured a dipole temperature consistent with COBE. The northern galactic pole is at the top of the map. Source: lambda.gsfc.nasa.gov \label{fig:wmapsky}}
\end{figure}
The temperature fluctuations come from the oscillation of the primordial plasma, caused by quantum fluctuations expanded by inflation. The oscillatory behavior of the perturbed plasma can be described as a forced harmonic oscillator,
\begin{equation}\label{eqn:OscillatorEq}
\ddot{ \Theta } + c_{s}^{2} k^{2} \Theta = - \frac{k^{2}}{3} \Psi -\ddot{ \Phi },
\end{equation}
with $c_{s}$ being the sound speed and $k$ the wave number. Perturbations, where $\Theta$ are perturbations in the metric and $\Psi$ are perturbations in the spatial curvature, manifest themselves as small temperature fluctuations in the background temperature. The wealth of data that we get from the anisotropies comes from the primordial temperature differences with changing $\theta$ and $\phi$ described by,
\begin{equation}
\mathrm{T}\left( \vec{x},\theta,\phi,\eta \right) =\mathrm{T}\left(\eta\right) \left[1 + \Theta \left(\vec{x},\theta,\phi,\eta \right) \right],
\end{equation}
where $\eta$ is the conformal time defined by,
\begin{equation}
\eta=\int\frac{dt}{a(t)}.
\end{equation}
The temperature differences originated from 3 effects described by,
\begin{equation}
\frac{\Delta T}{T}=\frac{\phi}{c^{2}}-\frac{\hat{r} \cdot \vec{v}}{c}+\frac{1}{3}\delta.
\end{equation}
The first term on the right corresponds to the depth of the potential well, second the velocity of the fluid relative to the observer, and the final term the intrinsic temperature of the region. Getting at the information embedded in the measurements of these small, $\frac{\Delta \mathrm{T}}{\mathrm{T}}\approx 10^{-5}$, temperature differences contained in a map, see Figure~\ref{fig:CobeWmapSky}, is done by expanding the temperature anisotropies into their spherical harmonic components,
\begin{equation}
\Theta \left(\vec{x},\theta,\phi,\eta \right)=\sum_{l=1}^{\infty} \sum_{m=-l}^{l} a_{lm}\left(\vec{x},\eta \right) Y_{lm}\left(\theta,\phi\right).
\end{equation}
No predictions of any particular $a_{lm}$ is possible, but information from the distribution from which they are drawn can be made, as long as the fluctuations that generated the parent distribution of the $a_{lm}$'s are described by a Gaussian random process. If so the angular power spectrum is given by,
\begin{equation}
\langle a_{lm}^{\mathrm{T}} a_{l^{'}m^{'}}^{\mathrm{T}*} \rangle = \delta_{l l^{'}} \delta_{m m^{'}}C_{l}^{\mathrm{TT}}.
\end{equation}
The T superscript denotes the cross correlation of the temperature, more on this in Section~\ref{subsec:CMBpol}. It is common to plot $C_{l}$ in a way that removes the monopole, dipole, and corrects for the scale invariance of the power of each $l$ such that,
\begin{equation}
\frac{l(l+1)C_l}{2\pi} = \left(\frac{\Delta T}{T}(\theta)\right)^2.
\end{equation}
Once angular power spectra and maps are generated from a given data set, see Figure~\ref{fig:WMAPPSclTTandTE}, the power spectra and map obtained can be compared to theoretical models to determine which cosmological model has the most likely parameter correspondence.
Many Cosmological parameters have been found with unprecedented accuracy by the WMAP space mission, see Table~\ref{tab:parameters}, and will soon be refined by an order of magnitude by the Planck Space Mission \citep{Planck06}. For Table~\ref{tab:parameters}, BAO is the Baryon Acoustic Oscillations, which searches for the distribution of galaxies in 3 dimensions, and SN is supernovae data. In addition to the temperature anisotropies the CMB is polarized at the microKelvin level and maps of the polarization provide a complimentary data set.
\begin{table}[p]
\caption{Key Cosmological Parameters from WMAP + BAO + SN \label{tab:parameters}}
\begin{center}
\begin{tabular}{l|c|c}
\hline
Parameter & Symbol & Value \\
\hline
\hline
Total density & $\Omega_0$ & $1.0052 \pm 0.0064$ \\
Dark energy density & $\Omega_{\Lambda}$ & $0.721 \pm 0.015$ \\
Matter density & $\Omega_m$ & $0.27 \pm 0.04$\\
Baryon density & $\Omega_B$ & $0.0462 \pm 0.0015$\\
Hubble constant & $H_{0}$ & $70.1\pm 1.3$ km/s/Mpc \\
Age of the Universe & $t_0$ & $13.73 \pm 0.12$ Gyr\\
Age at decoupling & $t_{dec}$ & $375938^{+3148}_{-3115}$ yr \\
Redshift of Reionization & $z_{reion}$ & $10.8 \pm 1.4$ Myr \\
\hline
\end{tabular}
\end{center}
\end{table}
\section{CMB Polarization}\label{subsec:CMBpol}
If a charged particle is illuminated by a quadrupole pattern, such as that in the CMB anisotropies (Figure~\ref{fig:cmbquadpol}), a polarized signal is generated, even if the illuminating radiation is not intrinsically polarized. A polarized electromagnetic wave of the form,
\begin{equation}
\vec{E} = E_x (t) \cos(kz - \omega t + \phi_x)\hat{i} + E_y (t) \cos(kz - \omega t + \phi_y)\hat{j},
\end{equation}
can be completely characterized by its stokes parameters. The parameters are given by
\begin{eqnarray}
I \equiv \langle E^2_x\rangle + \langle E^2_{y}\rangle \\
Q \equiv \langle E^2_x\rangle - \langle E^2_{y}\rangle \\
U \equiv \langle 2 E_x E_y \cos(\phi_y - \phi_x) \rangle \\
V \equiv \langle 2 E_x E_y \sin(\phi_y - \phi_x) \rangle
\end{eqnarray}
where the brackets denote a time average. The Stokes parameter $I$ is the total intensity of the radiation with $I^2 \geq Q^2 + U^2 + V^2$. $Q$ and $U$ describe the linear polarization of the wave and $V$ describes the circular polarization, these are equal to zero for unpolarized radiation. The angle of polarization is defined as,
\begin{equation}
\alpha \equiv\frac{1}{2}\tan^{-1}\left(\frac{U}{Q}\right),
\end{equation}
and the total polarization fraction, $P$, is
\begin{equation}
P \equiv\frac{\sqrt{Q^2 + U^2 + V^2}}{I}.
\end{equation}
$I$ and $V$ are rotationally invariant but $Q$ and $U$ transform under rotation by
\begin{equation}
Q^{\prime} = Q\cos(2\varphi) + U\sin(2\varphi),
\end{equation}
\begin{equation}
U^{\prime} = -Q\sin(2\varphi) + U\cos(2\varphi)
\end{equation}
where $\varphi$ is the rotation angle. However, it is clear that the quantity $Q^2 + U^2$ is rotationally invariant.
\begin{figure}[p]
\begin{center}
\includegraphics[width=\textwidth]{fig/StokesParameters.pdf}
\caption[Stokes parameters in degenerate states]{Examples of Stokes parameters in degenerate states retrieved May 15, 2009 from $\mathrm{en.wikipedia.org} / \mathrm{wiki} / \mathrm{StokesVector}$. \label{fig:stokesparameters}}
\end{center}
\end{figure}
\begin{figure}[p]
\begin{center}
\includegraphics[width = 13.5cm]{fig/CMBQuadpol.pdf}
\caption[Quadrupole illumination of electron]{Geometry of Thomson scattering for generation of polarized signal from unpolarized quadrupole illumination (adapted from \citet{hw97b}, also see website at http://background.uchicago.edu/$\sim$whu/). The incoming unpolarized radiation on the left/right (thick blue lines) is scattered by the free electron either up and down or into and out of the page (depending on the polarization). Similarly the radiation from above/below the free electron is scattered up and down or into and out of the page. The end result when viewing radiation emitted is that the horizontal polarization, comes from a cooler region than the vertical polarization giving rise to a polarized signal.\label{fig:cmbquadpol}}
\end{center}
\end{figure}
To standardize measurements a polarization convention was defined by the International Astronomical Union in 1973 and is summarized by \citet{Hamaker96}. At each point on the celestial sphere a cartesian coordinate system with the x and y axes pointing respectively toward the North and East, and the z axis along the line of sight pointing toward the observer (inwards) for a right-handed system. Though to confuse the issue slightly following the mathematical and CMB literature tradition, HEALPix (the most common pixelization scheme for CMB anisotropy maps) defines a cartesian referential with the x and y axes pointing respectively toward the South and East, and the z axis along the line of sight pointing away from the observer (outwards). This difference introduces a minus sign in U that has to be kept track of for power spectra generation and comparisons.
The recombination of the Universe at the surface of last scattering is not instantaneous but rather takes a finite amount of time. This leaves some fraction of charged particles to interact, through Thomson scattering, with the background anisotropies. This process is expected to give the CMB a polarization signature. Only the quadrupole moments and above generate polarization anisotropies, Figure~\ref{fig:cmbquadpol}, shows how the quadrupole moment causes a polarization from an unpolarized signal. Thomson scattering is only expected to polarize the CMB by $\sim10\%$ and will not generate any circular polarization, hence $V$ is expected to be zero. Though some circular polarization might be generated from gravitational lensing and galactic magnetic fields, this signal will be significantly smaller than that of the linear polarization signature.
Observations of the CMB polarization signature generate maps of the various Stokes parameters. Holding to the Helmholtz's decomposition any sufficiently smooth, rapidly decaying vector field (the Universe was/is finite in extent) can be decomposed into a divergence-free vector field (gradient) and a curl-free (divergence) vector field. The typical nomenclature for CMB, analogous to electromagnetic notation, is an E field (divergence) and a B field (gradient).
\begin{figure}[p]
\includegraphics[width=13.5cm]{fig/EandBpatterns.pdf}
\caption[E and B patterns for different Stokes parameter values]{E and B patterns for different Stokes parameter values (from \citet{zal04}). Note the parity differences in the E and B vector fields.\label{fig:EandBpatterns}}
\end{figure}
Transforming into E-modes and B-modes (E and B from here out) lets us take advantage of the fact that E and B are scalar spin-0 quantities like temperature and the maps can be interpreted similar to that of temperature. Congruent with the temperature expansions E and B can be expanded into spherical harmonics:
\begin{equation}
E(\theta,\phi) = \sum_{l,m} a^E_{lm}Y_{lm}(\theta,\phi)
\end{equation}
\begin{equation}
B(\theta,\phi) = \sum_{l,m} a^B_{lm}Y_{lm}(\theta,\phi).
\end{equation}
giving rise to angular power spectra that are defined by
\begin{equation}
C^{XX^{\prime}}_l \equiv \langle a^{X\ast}_{lm}a^{X^{\prime}}_{lm}\rangle
\end{equation}
where $X$ and $X^{\prime}$ can be $T$, $E$, or $B$ resulting in six possible power spectra $C^{TT}_l$, $C^{TE}_l$, $C^{EE}_l$, $C^{BB}_l$, $C^{TB}_l$, and $C^{EB}_l$. $C^{TT}_l$ denotes the temperature anisotropy angular power spectrum which has been previously discussed , $C^{TE}_l$ is the temperature polarization cross-power spectrum (see Figure~\ref{fig:WMAPPSclTTandTE}), and $C^{EE}_l$ and $C^{BB}_l$ are the $E$-mode and $B$-mode angular power spectra. E and T have an even parity while B has an odd parity, as seen in Figure~\ref{fig:EandBpatterns} and this property reduces the number of power spectra under cross-correlation from $6$ to $4$ since $C^{TB}_l$ and $C^{EB}_l$ are expected to be zero.
\begin{figure}[p]
\includegraphics[width=13.5cm]{fig/WMAPPSclTTandTE.pdf}
\caption[WMAP TT and TE angular power spectra]{WMAP TT and TE power spectra. Solid curves represent the best-fit theory spectrum from $\Lambda$CDM \citep{d09}. Grey area on the left represents the cosmic variance limit and the increase in error bars on the right side are caused by the finite beam size of WMAP. Source: lambda.gsfc.nasa.gov. \label{fig:WMAPPSclTTandTE}}
\end{figure}
\begin{figure}[p]
\includegraphics[width=13.5cm]{fig/PlanckAllPowerSpectra.pdf}
\caption[Planck satellite power spectra estimates]{Temperature and polarization spectra for $\Omega_{tot}=1$, $\Omega_{\Lambda}= 2/3$, $\Omega_{b} h^{2}= 0.02$, $\Omega_{m} h^{2}= 0.16$, $n=1 $, $z_{ri}= 7$, and $E_{i}= 2.2\times 10^{16}$GeV. The dashed lines indicate negative cross correlation and the boxes are the statistical errors of the Planck satellite. Plot courtesy of \citet{Hudodelson02}. \label{fig:PlanckPowerSpectra}}
\end{figure}
The expected EE power spectrum has extremes that correspond to scales where the fluid is in motion, maximizing the quadrupole of the CMB temperature. The motion of the fluid induces a quadrupole moment that correlates to the maximum velocity fields causing the maxima of the EE spectrum to correspond to the minimum of the TT spectrum with maximum correlation in the middle as seen in Figure~\ref{fig:WMAPPSclTTandTE}. Measurement of the polarization power spectra gives an independent confirmation of the temperature results. With the addition of the cross correlated (TE) spectrum the two additional pieces of information can lead to better constrained cosmological parameters and the breaking of degeneracies in different cosmological models. The reionization history of the Universe has a slight effect on the smallest scales of the TT spectrum, but leaves a drastic signature on the EE and TE spectra. The final spectrum as of yet undetected is the BB power spectrum and is one of the very few direct probes of inflation that exists. No direct detection of the BB spectrum has been made up to this point, but efforts to determine the scalar-to-tensor ratio are in the works and more sensitive instruments are constantly being developed. A detection of the BB spectrum would give constraints on the energy scale of inflation and help theorists confine the class of theories for inflation. The currently favored theory for inflation is a single parameter slow role model. The interested reader is encouraged to read through \citet{samtleben07}, \citet{samtleben07}, and \citet{Hudodelson02} for more information on the Cosmic Microwave Background temperature and polarization spectra.
\subsection{Foregrounds}
Unfortunately, observations of both the temperature and polarization signatures of the Big Bang are polluted by material between us and the surface of last scattering, known as foregrounds. The foregrounds have been well characterized for the temperature maps, and template subtractions from these maps have been successful. The difficulty arises when trying to understand the polarization of the foreground sources. The sources include reionization, gravitational lensing, synchrotron radiation, free-free emission, extragalactic point sources, atmosphere, and spinning dust grains. Of the 7 sources only 3 of them present severe problems for polarization observations. Reionization is expected to effect low-$l$ measurements of the EE spectrum, gravitational lensing will effect the B-mode measurements, extragalactic point sources are good for calibration and can be masked out fairly easily and the atmosphere (for ground based telescopes only) is not expected to have any polarization effects, but may contribute other systematic effects (see Section~\ref{subsec:barcroft}). This leaves synchrotron, free-free emission, and spinning dust grains as the primary obstacles. Synchrotron radiation is caused by relativistic charged particles interacting with the Galactic magnetic field and can be highly polarized. While free-free emission is due to electron-ion scattering and is expected to be unpolarized, but through Thomson re-scattering by electrons at the edges of the HII regions will become polarized tangentially to the edges of the clouds up to $\sim10\%$. Spinning dust radiation is not well understood, it is thought that the radiation is generated by electric dipole radiation from small rapidly rotating dust particles and has the potential to be significantly polarized. The signal from spinning dust grains is likely to peak at or around 20 GHz at $100~\mu$K and role off rapidly up to $\sim60$ GHz. Cosmic signals can be distinguished from foregrounds by their frequency dependence and their spatial power spectra. Using polarimeters that cover a wide range of frequencies, large sky coverage, and correlations with other lower frequency observations can generate significant information about polarized foregrounds. A more in depth analysis of foregrounds and their effects can be found in \citep{tegmark00}.
\begin{figure}[p]
\includegraphics[width=\textwidth]{fig/AtmosphereThesis.pdf}
\caption[Atmospheric and foreground emission]{Atmospheric and foreground emissions, generated by ATMOS32. Sky temperature is a sum of $H_{2}O$, $O_{2}$, and $O_{3}$ using a sea level water vapor density of $10 \mathrm{ ~g/m}^{3}$ and foreground spectral indices from \citet{bennett03b}. \label{fig:Atmosphere}}
\end{figure}
An interesting comprehensive treatment of the pertinent foregrounds is treated in excruciating detail in \citet{bennett03b}.
\section{B-Machine at White Mountain Research Station, Barcroft}
With this abridged overview of the CMB it is clear that maps with large sky coverage and widely separated frequency bands will play a role in discovering information about the origins and the current state of our local Universe. I have endeavored to build, field, operate, and analyze data from a telescope that is dedicated to mapping the E-modes and B-modes. B-Machine (named for eventually detecting B-modes not for Brian) has been placed at a high altitude site (see Subsection~\ref{subsec:barcroft}) and has been observing for several months. An in depth description of the instrument and its systems can be found in Chapter~\ref{chap:instrument} and Chapter~\ref{chap:polarizationrotator}. Characterizing the instrument and fielding it has been described in Chapter~\ref{chap:telescopechar} and finally an explanation of the preliminary data set and sky maps is presented in Chapter~\ref{chap:analysis}.
\subsection{Barcroft} \label{subsec:barcroft}
Atmospheric loading plays a significant role in both design and use of a telescope. When testing a telescope at sea level (Santa Barbara, Ca.) the typical sky zenith temperature is around 30 K versus about 10 K at a high altitude site (White Mountain Research Station at Barcroft, Ca.). This is mostly due to colder air temperatures and waters scale height of 2 km. Integrated precipitable water vapor (IPWV) for moderate latitudes at sea level varies from $1-2\mathrm{~cm}$, while at high altitude sites, at appreciable latitudes, only varies from $1-2\mathrm{~mm}$, see \citet{marvil05}. It is critical to field ground based telescopes at high altitude sites because rapid changes of the IPWV can mimic sky signals in either temperature or polarization and noise scales directly with antenna temperature (see Appendix~\ref{app:blackbody}).
We have had a great deal of experience in fielding telescopes at a high altitude site that is a reasonable driving distance from Santa Barbara, California. White Mountain Research Station, Barcroft (referred to as WMRS) has been developed into a reliable site over the past decade. Power and personnel issues have been all but eliminated and with the knowledge and experience gained by the previously 2 fielded telescopes (BEAST \citep{childers05} and WMPol \citep{levy08}) it was an easy decision to place B-Machine at WMRS. A comprehensive site survey was done early on, comparing WMRS to other high altitude sites, and it was found to be akin to others, see \citet{marvil05} for full results. The only major preparation needed at WMRS for the installation of B-Machine was the construction of a building with a fully retractable roof. The vast majority of the work and credit for the successful design and construction of the building goes to Andrew Riley. Construction of a building at a high altitude site is much more difficult than normal construction and Andrew went through some heroics to get the concrete foundation laid and the building made in time for B-Machine to be fielded.
\begin{figure}[p]
\begin{tabular}{cc}
\includegraphics[width=6.75cm,height=6.75cm]{fig/DomeTL.pdf}&
\includegraphics[width=6.75cm,height=6.75cm]{fig/DomeTR.pdf}\\
\includegraphics[width=6.75cm,height=6.75cm]{fig/DomeBL.pdf}&
\includegraphics[width=6.75cm,height=6.75cm]{fig/DomeBR.pdf}\\
\end{tabular}
\caption[B-Machine dome opening sequence]{Opening sequence for B-Machine dome, notice the roof is completely retractable. Andrew Riley is seen in the top left image standing on the 20 ft x 20 ft concrete pad that he laid to construct the dome on. The bottom right picture is from a different angle and has B-Machine in the dome before the baffling had been added. \label{fig:DomeOpening}}
\end{figure}
\chapter{Polarization Rotator}\label{chap:polarizationrotator}
To overcome the effects of $1/f$ noise contributed to the data stream by the HEMT amplifiers, some sort of chopping is required. For temperature experiments a Dicke switch radiometer is typically used, which chops rapidly between 2 temperatures. In B-Machine a new technique, that works in a similar fashion to a half wave plate, to chop between polarization states is being used. The Polarization Rotator consists of a linear polarizing wire grid mounted in front of a plane reflecting mirror (polished Aluminum plate). The wire grid decomposes the input radiation into its 2 polarization components, parallel and perpendicular to the wires. The parallel component is reflected off of the wire grid surface while the perpendicular component passes through the wire grid, where it is reflected off of the plane mirror, passing through the grid again and combining with the parallel component. The spacing between the plane mirror and the grid introduces a phase shift between the two polarization components, effectively rotating the plane of polarization of the input wave. A schematic to illustrate the operation of the polarization modulator is shown in Figure~\ref{fig:WireGridExplained}B. By rotating the grid the incident polarization can be rotated 2 times per revolution, as shown in Figure~\ref{fig:signal}, giving the single polarization sensitive receiver a chop between the 2 polarizations 4 times per revolution.
\begin{figure}[p]
\begin{center}
\includegraphics[width = \textwidth]{fig/PolGridfoamplate.pdf}
\caption[Wire grid, blue foam, and plane mirror picture]{Polarization Rotator assembly with plane mirror on bottom, blue foam spacer (transparent at our frequencies), and wire grid upside down so wires and support ring can be seen.
\label{fig:PolGridFoamPlate}}
\end{center}
\end{figure}
\section{Theory}
Reducing the operation of the Polarization Rotator to a simple example of a polarized signal incident on the polarization modulator is the easiest way to examine its workings. Polarized radiation of the form,
\begin{equation}\label{eqn:polrotIncident}
E_{incident}=E_{0}\cos(\kappa y + \omega t)\hat{j},
\end{equation}
is incident on the top surface of the wire grid with magnitude $E_{0}$, angular frequency $\omega$, and the polarization vector makes an angle of $\theta$ with the wires. Transforming the incident radiation basis into the wire grid basis (assuming wires are in the $\hat{j}'$ direction), see Figure~\ref{fig:WireGridExplained}A, and adding a phase shift, $\delta$, to the polarization that passes through the wire grid and is reflected off of the plane mirror backing plate gives,
\begin{equation}\label{eqn:polrotTransformed}
E_{transformed}=E_{0}\sin(\theta)\cos(\kappa y + \omega t+\delta)\hat{i}'+E_{0}\cos(\theta)\cos(\kappa y + \omega t)\hat{j}'.
\end{equation}
Combining the 2 radiation paths and converting back into the original basis gives,
\begin{eqnarray}\label{eqn:polrotback}
\lefteqn{E_{out}=}\nonumber \\
& E_{0}\sin(\theta)\cos(\theta)[\cos(\kappa y + \omega t+\delta)- \cos(\kappa y + \omega t)]\hat{i}+ \nonumber \\
& E_{0}[\sin^{2}(\theta)\cos(\kappa y + \omega t+\delta) + \cos^{2}(\theta)\cos(\kappa y + \omega t)]\hat{j},
\end{eqnarray}
see Figure~\ref{fig:WireGridExplained}B. Since the detector is sensitive to power it is easier to look at the power in each of the polarization states as a function of angle of the wire grid rather than electric field strength. When averaging over time it is assumed that the detector is at $y=0$ and $P\propto\langle E^{2} \rangle_{t}$, which yields,
\begin{eqnarray}\label{eqn:PowerinE}
P_{\hat{i}}& = & 2E_{0}^{2}\sin^{2}(\theta)\cos^{2}(\theta)\sin^{2}\left(\frac{\delta}{2}\right),\\
P_{\hat{j}}& =& \frac{1}{2}E_{0}^{2}[1-4\sin^{2}(\theta)\cos^{2}(\theta)\sin^{2}\left(\frac{\delta}{2}\right)],
\end{eqnarray}
where $\theta$ is the angle the wires on the wire grid make with the incident polarized signal's polarization angle and $\delta$ is the phase shift. Keeping in mind that the detector is only sensitive to either $P_{\hat{i}}$ or $ P_{\hat{j}}$, a wire angle can be found using the above framework to rotate any arbitrary incident polarization into the single polarization angle that the detector is sensitive too. The reflection and transmission losses have little effect on the outcome but do complicate the calculation significantly. As a result of this, they have been omitted from the calculations here.
\begin{figure}[p]
\begin{center}
\includegraphics[width=\textwidth]{fig/Signal.pdf}
\caption[Model of fractional power as a function of wire angle]{Fraction of the polarized incident power that reaches the detector as a function of Polarization Rotator wire angle . The wire angle is referenced to the horizontal. \label{fig:signal}}
\end{center}
\end{figure}
\subsection{Wire Grid Plane Mirror Spacing}
To make the plane of polarization rotate by $\frac{\pi}{2}$ when the wires are at $45^{\circ}$ the incident radiation that is polarized perpendicular to the wires needs a path length difference from the parallel polarization of $\frac{\lambda}{2}$. From Figure~\ref{fig:WireGridExplained}B the path difference is
\begin{equation}
\begin{array}{l}
\Delta l=AB+BC-AD, \\
AB=BC=\frac{d}{\cos(\theta)} ~~\mbox{and}~~ AD=\frac{2 d \sin^{2}(\theta)}{\cos(\theta)}, \\
\Delta l=\frac{2d}{\cos(\theta)}-\frac{2 d \sin^{2}(\theta)}{\cos(\theta)}=2d \cos(\theta), \\
\end{array}
\label{eqn:WGpathdifference}
\end{equation}
where $\theta$ is the angle of incidence of the radiation. Then for a $\frac{\lambda}{2}$ path difference a wire grid to plane mirror spacing of
\begin{equation}\label{eqn:WGPMspacing}
d=\frac{\lambda}{4\cos(\theta)}
\end{equation}
gives the appropriate phase shift. Using the B-Machine parameters of $\lambda=.7223$ cm and $\theta=.6632$ gives a spacing of $0.229\mbox{ cm} ~\mbox{or}~ 0.09\mbox{ inches}$.
\begin{figure}[p]
\begin{tabular}{cc}
\includegraphics[width=7cm]{fig/axisTransformed.pdf}&
\includegraphics[width=7cm]{fig/WireGridInteraction2.pdf}\\
A) &
B) \\
\end{tabular}
\caption[Basis rotation and wire grid radiation interaction]{$A)$ Incident radiation polarized in the unprimed bases is transformed into the primed bases, a phase shift is added onto the polarization that is perpendicular to the wires and transformed back into the unprimed bases. The wires are aligned to the $\hat{j}'$ axis. $B)$ Side view of wire grid interaction with the incident radiation and the grid plane mirror spacing shown. \label{fig:WireGridExplained}}
\end{figure}
\subsection{Polarization Rotator Response}
With a theoretical construction of the operation of the Polarization Rotator in hand, basic questions to test the viability of the technique can be answered. It is necessary for the rotator to work across the B-Machine band-pass and that the maximum polarization rotation occurs for the centeral frequency of the band-pass. Using the fact that the beam makes an angle of $38^{\circ}$ with the normal to the polarization grid and the expected maximum polarization rotation occurs at $45^{\circ}$, a simple plot (Figure~\ref{fig:powervslambda}) shows that a spacing of 0.09 inches gives the best results. With a bandwidth of $16.87\%$ the efficiency of the rotator is calculated to be $99.41\%$ or has an isolation of $22.28$ dB. It is assumed that at $41.5\mathrm{~GHz}$ the rotator is $100\%$ efficient. There is some ambiguity in the definition of bandwidth, for my purposes here percentage bandwidth is defined as
\begin{equation}\label{eqn:bandwidth}
\beta=\frac{f_{stop}-f_{start}}{f_{bandcenter}},
\end{equation}
where the start and stop frequencies correspond to the 3 dB points of the filter. When the Polarization Rotator has rotated the plane of polarization $90^{\circ}$, the expectation is that the other polarization will have no power in it, but since the rotation is sensitive to frequency, beam size, grid spacing, wire spacing, and angle of incidence some power leaks from one polarization to the other. A measure of this power leakage is the isolation quoted here in decibels (dB). The corrugated feed horns have a FWHM of $18^{\circ}$ which makes the angle of incidence vary from $32.5-50.5^{\circ}$, reducing the isolation a bit more, see Figure~\ref{fig:isovsbeam}.
\begin{figure}[p]
\begin{center}
\includegraphics[width = 13.5cm]{fig/powervslambda.pdf}
\caption[Fraction of input power as function of wavelength]{Fractional input power verses wavelength with the black line corresponding to 41.5 GHz using 0.09 inch wire grid to plane mirror spacing. \label{fig:powervslambda}}
\end{center}
\end{figure}
\begin{figure}[p]
\begin{center}
\includegraphics[width = 13.5cm]{fig/IsoVAngleIncwBandPass.pdf}
\caption[Isolation of Q from U as a function of beam divergence]{Isolation of Polarization Rotator as a function of angle of incidence of the horn. The FWHM of the horn is $18^{\circ}$ and the angle of incidence ranging from $10^{\circ}-90^{\circ}$. \label{fig:isovsbeam}}
\end{center}
\end{figure}
Convolution of beam width and frequency dependence gives an expected isolation of 20 dB. This is in good agreement with experimental values from sky dips as seen in Chapter~\ref{chap:instrument} Figure~\ref{fig:crosspolarization}.
\subsection{Wire Grid}
The wire grid that creates the polarization splitting can be tuned by adjusting the size of the wires and thier spacing, changing the efficiency of the wire grid as a function of wavelength. B-Machine's band-pass is well known and hence tuning of the wire gird is a straight forward exercise in using Equations~\ref{eqn:rp} and \ref{eqn:rn}. These equations were originally derived by Lamb~\citep{Lamb98} in 1898 and derived again and presented independently by ~\citep{Lesurf90}.
\begin{equation}\label{eqn:rp}
|r_{p}|^{2}=\left[1+\left(\frac{2S}{\lambda}\right)^{2}\left(ln\left(\frac{S}{\pi d}\right)\right)^{2}\right]^{-1}
\end{equation}
\begin{equation}\label{eqn:rn}
|r_{p}|^{2}=\frac{(\pi d)^{4}}{(2S\lambda)^{2}[1+(\pi^{2} d^{2})^{2}/(2S\lambda)^{2}]}
\end{equation}
where $d$ is the wire diameter (the calculations assume a wire with a circular cross section) and $S$ is the center to center spacing of wires. The smallest wire width was limited by the fabrication process to be 5 mils ($0.0127$ cm). Seen in Figure~\ref{fig:WGRVsSpacing} the reflectivity of the parallel component is close to $100\%$ at $99.9976\%$, while the reflectivity of the perpendicular component is low at $0.0836\%$.
\begin{figure}[p]
\begin{center}
\includegraphics[width = 13.5cm]{fig/RVsWireSpacing.pdf}
\caption[Wire grid reflectivity as a function of wire spacing]{Reflectivity of normalized power for radiation polarized parallel (red) to the wires and perpendicular (blue) to the wires. The black vertical line is the central frequency of the B-Machine band-pass (41.5 GHz) and 5 mil (0.127 cm) wide wires with a 15 mil (0.0381 cm) center to center spacing. \label{fig:WGRVsSpacing}}
\end{center}
\end{figure}
Several techniques in constructing a durable and robust wire grid were explored before the final usable grid was produced. The first technique was to use a threaded rod (40 pitch) secured in a rigid metal frame and wrap copper wire around end over end. Once wrapped, the wires were secured to the rod with epoxy and one surface of the wires were cut away, leaving the frame with free standing wires. This technique had several problems associated with it. First, it was hard to get the spacing consistent over large frames. Second, the wires were not rigid enough to survive rotating at any reasonable speed. After this technique was abandoned a photo-lithography process to pattern and evaporate wires onto a piece of Polypropylene (Polypro is transparent to microwave radiation) was pursued. Again this technique bore little fruit. The evaporated material was too thin, had poor adhesion to the Polypro and making large patterns with high tolerances with equipment available in the local clean rooms was not possible. Following these failures several manufacturers of flexible circuit boards were contacted and asked to quote on a 12" diameter wire grid, finally settling on All Flex Inc.\footnote{All Flex Inc., Northfield, MN 55057, www.allflexinc.com}
\section{Effects on Telescope Sensitivity}
Receiver sensitivity can be estimated using the radiometer equation,
\begin{equation}\label{eqn:sensitivity}
\sigma_{T}=K\left(\frac{T_{\mathrm{sys}}+T_{\mathrm{sky}}}{\sqrt{\Delta\nu\cdot\tau}}\right),
\end{equation}
where $ \sigma_{T}$ is the root-mean-square noise, $T_{\mathrm{sys}}$ is the system noise temperature,
$T_{\mathrm{sky}}$ is the sky antenna temperature, $\Delta\nu$ is the bandwidth, and $\tau$ is the integration time. $K$ is the sensitivity constant and depends on the type of radiometer being used \citep{daywitt89}. For example, an un-differenced receiver will have $K=1$, while a Dicke switched radiometer will have $K=2$.
\subsection{Demodulation Technique}
When calculating the sensitivity of the instrument no correction factor is added for $\frac{1}{f}$ noise. The sensitivity is considered the white noise limit and it is assumed that the $\frac{1}{f}$ noise is taken into account by the sensitivity constant, $K$, in the radiometer equation. Differencing the signal on time scales much faster than the $\frac{1}{f}$ knee is the typical method used to overcome the associated noise. B-Machine's differencing is done by using a lock-in post-processing software tool. The tool written in IDL (see Chapter~\ref{chap:analysis} Subsection~\ref{subsec:IDL}) multiplies the signal by a square wave which oscillates between $+1$ and $-1$. The phase that aligns the square wave with the appropriate polarization is determined by the orientation of the channels horn to the Polarization Rotators wire grid and is generated using the get max phase procedure; see Chapter~\ref{chap:telescopechar} Subsection~\ref{subsec:GetMaxPhase}.
\subsection{Derivation of Sensitivity Constant}
For a standard total power radiometer the sensitivity constant for the Radiometer Equation~\ref{eqn:sensitivity}, is $1$. All other chopping schemes degrade the sensitivity of a given radiometer. In general, the degradation of the sensitivity falls into one of 2 categories. The first is integration time and the second, error propagation (more error accrues on a given reading when you subtract 2 signals with the same error). For B-Machine the sensitivity constant is calculated using a sine wave chopping technique with a square wave demodulation. This is achieved by starting with a square wave chop and a square wave demodulation (Dicke Switched), then multiplying by the efficiency factor between square and sine wave demodulation to get the final answer.
For a standard Dicke Switched Radiometer half the time is spent looking at each of the loads, thus only $\frac{1}{2}$ of the integration time is spent on a given load which degrades the sensitivity by $\sqrt{2}$. Then differencing the 2 signals and adding their error in quadrature yields another factor of $\sqrt{2}$, and multiplying these gives a sensitivity constant of $K=2$. A sine wave demodulation scheme efficiency factor can be derived by looking at a simple example. If a Dicke Switched Radiometer is chopping between a $0$ K and a $1$ K blackbody with a period of $2\pi$ (for simplicity in the sine wave calculations) a signal of $\pi$ K is observed, see Equation~\ref{eqn:SquareInt}, while a similar radiometer that is sine wave chopped will see a signal of $2$ K, see Equation~\ref{eqn:SinInt}. From this the efficiency factor of $Sine\rightarrow Square$ is $\frac{2}{\pi}$.
\begin{equation}\label{eqn:SquareInt}
\langle T_{Square} \rangle=\int_{0}^{\pi}0dT-\int_{\pi}^{2\pi}1dT=\pi
\end{equation}
\begin{equation}\label{eqn:SinInt}
\langle T_{Sine} \rangle=\int_{0}^{\pi}\frac{\sin(T)}{2}dT-\int_{\pi}^{2\pi}\frac{\sin(T)}{2}dT=2
\end{equation}
Multiplying the efficiency gives a sensitivity constant of $\pi$ for $\Delta T$. Gaining another factor of $1/2$ from the definition of the $Q$ and $U$, Stokes parameters
\begin{equation}
Q=\frac{T_{x}-T_{y}}{2},
\end{equation}
\begin{equation}
Q=\frac{T_{x'}-T_{y'}}{2},
\end{equation}
gives the final sensitivity constant as $\frac{\pi}{2}$ for each of the stokes parameters.
\begin{figure}[p]
\begin{center}
\includegraphics[width = 13.5cm]{fig/SquareVsSine.pdf}
\caption[Sample square wave and sine wave demodulation wave form]{Square wave (blue) and sine wave (red) demodulation wave forms. \label{fig:SquareVsSin}}
\end{center}
\end{figure}
\subsection{$\frac{1}{f}$ Characteristics}
$1/f$ noise is a signal with a power spectral density that is roughly inversely proportional to the frequency. The noise power spectral density $P(f)$ can be described as,
\begin{equation}\label{eqn:PSD}
P(f)\sim\sigma^{2}\left[1+\left(\frac{f}{f_{\mathrm{k}}}\right)^{\alpha}\right],
\end{equation}
where $\sigma$ is the white noise component level, $f_{\mathrm{k}}$ denotes the frequency where the white noise and $1/f$ contribute equally to the total noise and is referred to as the knee frequency, and $\alpha$ characterizes the slope of the power spectrum and is typically $\simeq1$. The noise fluctuates the apparent gain of the system, mimicking a real signal on the sky. By differencing the signal on time scales much shorter than the knee frequency the fluctuations can be minimized. For B-Machine a chop rate of 133.6 Hz between polarizations was used with knee frequencies, see Table~\ref{tab:oof}, of $\sim140 \mathrm{ Hz}$ causing some of the $1/f$ noise to pollute the demodulated data. The chop rate was determined based on maximum sampling rates of data acquisition computers and beam smearing effects.
\begin{table}[p]
\begin{center}
\caption[List of $\frac{1}{f}$ Knees Before and After Demodulation]{List of $\frac{1}{f}$ Knees Before and After Demodulation \label{tab:oof}}
\begin{tabular}{|c|c|c|}
\hline
Channel & Undifferenced & Differenced \\
& Knee (Hz) & Knee (mHz) \\
\hline
\hline
1 & 140.0 & 5.0\\
\hline
2 & 143.0 & 4.4 \\
\hline
3 & 151.0 & 4.4\\
\hline
6 & 135.0 & 5.6\\
\hline
\hline
\end{tabular}
\end{center}
\end{table}
\begin{figure}[p]
\begin{center}
\includegraphics[width = 13.5cm]{fig/PSDThesis.pdf}
\caption[Power spectral distribution before/after demodulation]{Power spectral distribution from the central channel of the cold radiometer viewing sky. Black curve is before demodulation and the blue is after demodulation, solid red lines denote noise floor and red dashed lines mark the knee. \label{fig:PSD}}
\end{center}
\end{figure}
\section{Testing}
The theoretical calculations for the wire grid polarization modulator showed no obvious limitations to stop progress on the development of practical testing. Tests of different reflective materials, including wire grids, were made to determine the reflectivity efficiencies of the different materials compared to a wire grid. As expected Copper and Aluminum plates both had efficiencies over $99\%$. All of the measurements were done by hand, chopping a cold and warm load in reflection off of a plate of the desired material. For the several wire grids on hand all the efficiencies were above $95\%$. More accurate measurements we not possible due to $\frac{1}{f}$ dominating the errors in the measurements. Small imperfections in the efficiencies only contribute small changes in the calibration constants, for the low levels that were measured the efficiency imperfections were small and corrected for in the final calibration constants, see Chapter~\ref{chap:telescopechar} Subsection~\ref{Sec:Calibration}. The final grids were not made as of this point, but were eventually measured on the full integrated telescope.
\begin{figure}[p]
\begin{tabular}{cc}
\includegraphics[width=6.5cm,height=6.5cm]{fig/ThermalSource.pdf}&
\includegraphics[width=6.5cm,height=6.5cm]{fig/2HornOMTtest.pdf}\\
\includegraphics[width=6.5cm,height=6.5cm]{fig/SmallTelescope.pdf}&
\includegraphics[width=6.5cm,height=6.5cm]{fig/BmachineOutside.pdf}\\
\end{tabular}
\caption[The 4 different testing platforms for the Polarization Rotator]{Polarization Rotator platforms for testing. Top left, Polarization Rotator looking at polarized thermal source. Top right, Gunn diode reflecting off of Polarization Rotator and using an Ortho-Mode Transducer to look at both polarizations simultaneously. Bottom left, small telescope setup with beam path in red, beam coming out of page. Bottom right, integrated telescope looking at sky from East side of Broida Hall at UCSB. \label{fig:PolRotTestPlatforms}}
\end{figure}
Several small testing platforms, Figure~\ref{fig:powervslambda}, were constructed to make rudimentary measurements. The first platform was a simple Polarization Rotator using a corrugated feed horn and a polarized thermal source. For this setup the Polarization Rotator used a small wire grid that was manufactured by using a standard lift off lithography technique that evaporated a thin layer of copper onto a Polypropylene film producing wires. The second wire grid was purchased as a calibration standard for WMPOL and affixed to a frame. The thermal source, set to $75^{\circ}$C, placed behind the polarizing wire grid transmitted through the wire grid at one polarization and reflected the ambient load for the other polarization giving a polarized signal of roughly $50^{\circ}$C. Rotating the polarizing grid while taking data caused the waveform to shift as expected, confirming that the system could distinguish between both Q and U signals. The second test setup used a Gunn diode at $41.5$ GHz as a polarized source. In addition to the polarized source an Ortho-Mode Transducer (OMT) was used to observe both polarizations simultaneously, Figure~\ref{fig:OMT}. The first 2 test setups were relatively easy to construct and yielded data that confirmed the calculations in a rough fashion. This gave us the confidence to invest the time to construct and test a small telescope with focusing optics and a 5" Polarization Rotator.
The small telescope was a modified off-axis Gregorian telescope using a 22 inch primary, 8 inch secondary and a 5" Polarization Rotator. The wire grid for the small telescope was a UCSB manufactured small wire grid. These optics and an $18^{\circ}$ FWHM corrugated feed horn gives a $\sim6^{\circ}$ FWHM beam on the sky. The telescope was used over an extended time period to gather information on the functionality of the Polarization Rotator. A crude software pipeline to process the data was setup to view the data in the $Q$ and $U$ Stokes parameters. This presented the first opportunity to verify the $\frac{1}{f}$ characteristics before and after demodulation. A $\frac{1}{f}$ knee of $150$ Hz before demodulation was reduced to $10$ mHz after demodulation. This platform also allowed many different, not well documented, tests, things such as placing Eccosorb and other wire grids in the beam path. This yielded some experience and feel for the modulator before an integrated telescope was made. Secure that the Polarization Rotator met minimal standards for a larger telescope, B-Machine was finally constructed. Full tests of B-Machine with the field rotator and detector chains are explored in Chapter~\ref{chap:telescopechar}.
\begin{figure}[p]
\begin{center}
\includegraphics[width = 13.5cm]{fig/OMT.pdf}
\caption[Comparison of both polarizations using polarized source and OMT]{Comparison of both polarizations using a polarized source at $\sim45^{\circ}$ from horizontal and OMT. The red line is from the horizontal output of the OMT and black lines from the vertical output of the OMT. \label{fig:OMT}}
\end{center}
\end{figure}
\begin{figure}[p]
\begin{center}
\includegraphics[width = 13.5cm]{fig/beamshift.pdf}
\caption[Main lobe response from polarized thermal source]{Main lobe response from polarization modulator while viewing a polarized thermal source. Black is reflected off of plane mirror and red is reflection off of wire grid. \label{fig:BeamShift}}
\end{center}
\end{figure}
\begin{figure}[p]
\begin{center}
\includegraphics[width = 13.5cm]{fig/tquThermalsource.pdf}
\caption[Plot of I, Q and U from rotating polarized thermal source]{Radiometer output from small thermal source mounted on the roof of the Bren building rotated several times to show Q (black) and U (blue) signals. Also added is the quadrature sum in red. \label{fig:TQUThermalsource}}
\end{center}
\end{figure}
\chapter{Telescope Characterization}\label{chap:telescopechar}
The B-Machine telescope was characterized both at UCSB and WMRS Barcroft. During the UCSB characterization phase the focus was mainly on beam shape and noise characteristics of the instrument. While at WMRS efforts to define calibration constants both in temperature and polarization were explored. At the same time the servo system was exercised to determine its operational fringes.
\section{Beam Characterization} \label{sec:beamcharacterization}
A full exploration of the beam size and patterns were made to test possible beam shape problems due to the addition of the Polarization Rotator into the Beast \citep{childers05} and WMPOL \citep{levy08} optics.
\subsection{Gaussian Beam Size}
The beam size was determined using scans of the Moon from August 8th, 2008. The Moon was scanned slowly multiple times giving fine angular resolution and 10-20 crossings in a short time period. The multiple crossings of the Moon made it possible to determine when the beam was centroided. A simulated temperature map of the Moon was generated using modeling software written by Stephen Keihm~\citep{keihm75}. The Moon simulation used phase angle, frequency ($41\mathrm{~Ghz}$), and polarization angle (zero for temperature maps) relative to the Moon equator to generate accurate temperature maps. The phase angle was determined by the percent of the Moon that was illuminated on August 8th, 2008 ($43\%$). Simulated maps convolved with different Gaussian beams where checked for goodness of fit with Moon scans using the chi square technique, Figure~\ref{fig:ChiSquare}. The reduced chi square goodness of fit was generated and the beam size was taken to be where this parameter was minimized. This process was done for the central horn (channel 14) and one of the off axis horns (channel 9), Figure~\ref{fig:BeamSize}, yielding beam FWHM of $22.2'\pm0.2'$ for the central horn and $24.0'\pm0.2'$ for the off axis horn. Beam shapes for the $Q$ stokes parameter were investigated using the Moon scanned data, but inconsistencies in the simulation data (polarization from limbs of Moon not well understood) and issues with saturation made the process untenable. The main source of error in determining the beam size came from the resolution in scanning and the resolution of the simulated Moon maps. Beam sizes from experiments using the same optics have yielded similar results. The Beast Campaign gave FWHM beam size of $23.0'$ and WMPOL $24.0'$ for similar optics. WMPOL used larger corrugated feed horns causing the larger beam size. A direct comparison of a Moon scan done from both B-Machine and Beast telescopes, Figure~\ref{fig:BeastBmachineMoon}, show that a smaller beam from B-Machine is expected. Due to the Polarization Rotator much greater care was taken in aligning the optics for B-Machine than Beast and a slightly smaller (and closer to theoretical size) beam was achieved.
\begin{figure}[p]
\begin{center}
\includegraphics[width = \textwidth]{fig/ChiSquareFit.pdf}
\end{center}
\caption[Reduced chi-square fit of simulated Moon/beam convolution and Moon scan]{Reduced chi-square fit of simulated Moon/beam convolution and Moon scan. Reduced Chi square minimum corresponds to best fit of simulation with data. The red curve is an off axis horn and black is the central horn. Error estimates appear to be to large for the off axis horn. \label{fig:ChiSquare}}
\end{figure}
\begin{figure}[p]
\begin{center}
\includegraphics[height=17cm]{fig/BeamSize.pdf}
\end{center}
\caption[Beam size using Moon vs simulated Moon]{Moon data and simulated Moon data using chi square test to determine best fit of convolved simulated Moon and beam. The black points/line are data taken 08/08/2008 and the red points/line are Moon and beam convolved. Top: central horn with best fit beam FWHM $22.2'\pm0.2'$, Bottom: off axis horn with best fit beam FWHM $24.0'\pm0.2'$. \label{fig:BeamSize}}
\end{figure}
\begin{figure}[p]
\begin{center}
\includegraphics[width = \textwidth]{fig/BeastBmachineoverplot.pdf}
\caption[Comparison of Beast and B-Machine beam shape]{Thick black line is a Moon scan from the Beast telescope using the same optics as B-Machine. The multi color lines are plots of each sector of the B-Machine Moon scan. The dotted red line is the demodulated Moon scan. \label{fig:BeastBmachineMoon}}
\end{center}
\end{figure}
\subsection{Beam Shift}
The central corrugated feed horn observing a polarized aligned source will see maximum signal 4 times per revolution: 2 maxima corresponding to the wires being horizontal (signal reflects off of wires) and 2 corresponding to the wires being vertical (signal reflects off of plane mirror backing plate) relative to the horizon. The additional path length when the wires are vertical causes the beam to shift on the sky. This is a purely geometric effect and for a $0.09\mbox{ inch}$ wire grid to plane mirror spacing, the calculated beam shift should be $6.5'$. Using the $41.5$ GHz source on the roof of the Bren Institute and plotting the signal as a function of elevation gives an observed beam shift of $6.56'\pm0.17'$, see Figure~\ref{fig:MainLobeBeamShift}.
\begin{figure}[p]
\begin{center}
\includegraphics[width = \textwidth]{fig/MainLobeBeamShift.pdf}
\end{center}
\caption[Main lobe beam shift 4 max sectors]{Observed beam shift using polarized source, red and black lines correspond to wires on Polarization Rotator vertical (reflecting off of plane mirror), Green and Red lines wires horizontal (reflecting off of wire grid). Testing done on East side of Broida Hall using the 41.5 GHz source with the polarization aligned with the central horn. \label{fig:MainLobeBeamShift}}
\end{figure}
\subsection{Full Beam Shapes}
On February 26th, 2008 B-Machine was rolled to the top of the loading ramp on the east side of Broida, the frame was aligned using existing markings from previous observing days. The markings are spray painted circles on the ground that are three different colors for each of the three stabilization feet on the experiment. A 41.5 Ghz source was mounted on the roof of the Bren Institute approximately 150 m from the telescope. Though this is still in the near field, $\frac{D^{2}}{L\lambda}=\frac{2.0^{2}}{150\times0.0072}>1$, the source uses a corrugated feed horn and can be treated like a point source. Attached to the source were 2 aerowave attenuators and a Direct Read Attenuator (DRA). The Aerowave attenuators had no attenuation on them and the DRA was set at a different attenuation level for each scan, to increase the dynamic range of the beam measurements.
Each polarization, referenced to the horizon as vertical, horizontal or $45^{\circ}$, of the horn was done in the same fashion. Attenuation was added on the DRA, then an elevation scan was taken followed by an azimuth scan. The DRA was adjusted to the next attenuation level (50 dB, 45 dB, 35 dB, 15 dB, or 0 dB) and then Az/El scans were done again. Using multiple attenuation levels allowed for better signal to noise so that the side lobes could be seen out to the 9th or 10th side lobe. Each scan was pieced together by matching the section of the previous scan with the beginning of the uncompressed sections of the next scan. The scans were trimmed when it was clear that the signal to noise was poor. Due to the hand alignment of the source, the $45^{\circ}$ polarization is close but the horizontal and vertical polarizations are off by a couple of degrees.
The side lobe differences of the beam shapes for horizontal and vertical incident polarization can give rise to a spurious $Q$ signal. If an object, for example the sun at 6000 K, is in the 5th and 6th side lobe (in multiple lobes due to angular size) a $1.3\mbox{ mK}$ $Q$ signal is expected. To confirm the expected value a scan of the sun with multiple crossings was taken. When the data were analyzed the thermal effects overwhelmed the small radiometric effects from the Sun. These effects included heating of baffling, optics, RF chains and surrounding environment. Instead using the source scanned data to get fractional signals by comparing the amplitudes of each side lobe in horizontal and vertical polarizations (see Figure~\ref{fig:TVertHor}) and then looking at the $Q$ signal from the source with polarization at $45^{\circ}$ gives a measure of the beam asymmetries. With the source at $45^{\circ}$ the expected signal is all $U$. Comparing the $Q$ signal (measured) with the scans of the source vertical and horizontal to get the expected signal as a fraction of the height of the side lobe estimates of the telescope side lobe pollution were made. Table~\ref{tab:QoutofTfraction} lists expected and measured numbers for the first 5 side lobes. The 3rd side lobe had some anomalous noise problems and was omitted. The expected numbers represent differences in the side lobe response for the 2 orthogonal polarization states, the measured response includes multiple effects that can't be de-convolved from the side lobe asymmetries. However, the measure of the first side lobe with $0.0\%$ expected conversion (horizontal and vertical side lobe sizes are identical) can give a rough estimate of the other effects such as miss alignment of the source, T to Q conversion from non side lobe effects and U to Q conversion. A source that was unpolarized and provided the needed signal to noise for all of the side lobes would have separated some of the ambiguities better. Table~\ref{tab:QoutofTfraction} provides a good first estimate of side lobe contamination and with the inclusion of all the effects gives information on how close thermal signals can be to the main lobe before they contaminate the survey. For a possible balloon experiment the galactic plane may present some spurious signals that will need to be modeled for template removal.
\begin{table}[p]
\begin{center}
\caption[Fraction of $Q$ Out of $T$ from Side Lobes of Central Horn]{Fraction of $Q$ Out of $T$ from Side Lobes of Central Horn \label{tab:QoutofTfraction}}
\begin{tabular}{|c|c|c|c|c|}
\hline
Side Lobe & Expected & Measured & Error & Fraction of\\
& & & & Main Lobe \\
\hline
\hline
1 & 0.0$\%$ & 6$\%$ & 2$\%$ & $2.0\pm0.4\cdot10^{-4}$\\
\hline
2 & 13$\%$ & 20$\%$ & 2.5$\%$ & $7.5\pm0.2\cdot10^{-5}$\\
\hline
3 & 0.0$\%$ & Noise & N/A & $5.0\pm1.4\cdot10^{-5}$\\
\hline
4 & 11.0$\%$ & 15$\%$ & 3.0$\%$ & $2.9\pm1.0\cdot10^{-5}$\\
\hline
5 & 15.0$\%$ & 20$\%$ & 3.6$\%$ & $2.5\pm0.9\cdot10^{-5}$\\
\hline
6 & 22.0$\%$ & 25$\%$ & 2.7$\%$ & $3.0\pm1.0\cdot10^{-5}$\\
\hline
\end{tabular}
\end{center}
\end{table}
Full beam maps of the central channel were taken for both azimuth and elevation in T, $Q$, $U$ and of all sectors for the source polarized vertically, horizontally, and $45^{\circ}$. Each plot was carefully inspected for any indication of sensitivity to signals well off the bore or major differences in the beam shapes. Any major beam anomalies in $Q$ or $U$ would have presented major problems for this type of chopping technique. All of the beam plots were made and show no unexpected results. A subset of the plots are included here for reference, Figures~\ref{fig:TVertHor} through ~\ref{fig:FullBeamHornCTQ}.
\begin{figure}[p]
\begin{center}
\includegraphics[width = \textwidth]{fig/TVertHorPlot.pdf}
\caption[Full beam pattern for $T$ source vertical and horizontal]{Full beam pattern of the central horn for $T$ with source Vertical (blue) and Horizontal (black). The variations in beam shape can give rise to spurious $Q$ signals. \label{fig:TVertHor}}
\end{center}
\end{figure}
\begin{figure}[p]
\begin{center}
\includegraphics[width = \textwidth]{fig/FullBeam4SectThes45.pdf}
\caption[Full beam pattern for max sectors source polarization at $45^{\circ}$]{Full beam pattern of the central horn for max sectors with source polarization at $45^{\circ}$. \label{fig:FullBeam4sect45}}
\end{center}
\end{figure}
\begin{figure}[p]
\begin{center}
\includegraphics[width = \textwidth]{fig/FullBeam4sectELhorthesb.pdf}
\caption[Full beam pattern for max sectors source polarization horizontal]{Full beam pattern of the central horn for max sectors with source polarization horizontal. \label{fig:FullBeam4secthor}}
\end{center}
\end{figure}
\begin{figure}[p]
\begin{center}
\includegraphics[width = \textwidth]{fig/FullBeamU.pdf}
\caption[Full beam pattern for $Q$ and $U$ source polarization $45^{\circ}$]{Full beam pattern of the central horn for $Q$ (red) and $U$ (black) with source polarization $45^{\circ}$. Ratio of $Q$ to $U$ shows $Q$ to $U$ conversion/isolation and source alignment precision. \label{fig:FullBeamU45}}
\end{center}
\end{figure}
\begin{figure}[p]
\begin{center}
\includegraphics[width = 13.5cm]{fig/FullBeamHornCTQ.pdf}
\caption[Full beam pattern for off axis horn source polarization horizontal]{Full beam pattern for horn C, off axis horn, that is not vertically or horizontally lined up with the central horn. $T$ (black) and $Q$ (blue) with source polarization horizontal. The gap in data where the first side lobe is located is due to lack of on scale data for this region. \label{fig:FullBeamHornCTQ}}
\end{center}
\end{figure}
\section{Calibration}\label{Sec:Calibration}
The phrase "calibration" is used to describe the measurement of the instrumental parameters that characterize the three transfer standards (gains) to convert the data stream which is in Volts to temperature, $Q$, or $U$. Each channel has 2 gains, one for the AC channels and the second for the DC channels, they are scaled by the difference in the IF gain between the AC and DC chains, see Chapter~\ref{chap:instrument} Section~\ref{Subsect:datainput}. A full calibration sequence was done at both the beginning and end of the data taking campaign which includes filling the beam with a known cold load, intermediate load, ambient load and sky load, with daily automated calibration sequences ran for all days. Daily calibrations use an automated ambient load and the sky to calibrate.
\subsection{Temperature}
Typically measurements of the gain are done with 2 thermal loads, Eccosorb (AN72) either at room temperature ($\sim300$ K) or soaked in a LN bath ($\sim73.8$ K at $4$ km). They are put in front of the radiometer, just below the extruded polystyrene window, to make sure they fill the beam. With the voltages that correspond to the given temperatures from each RF chains output and,
\begin{center}
\begin{equation}\label{eqn:gainlin}
G_{0}=\frac{T_{hot}-T_{cold}}{V_{cold}-V_{hot}},
\end{equation}
\end{center}
and
\begin{center}
\begin{equation}\label{eqn:lincal}
V_{out}=G_{0}\left(T_{sky}+T_{sys}\right),
\end{equation}
\end{center}
to get the gain in Kelvin per Volt. Here $T_{hot}$/$V_{hot}$ is the temperature/voltage ratio of the warm load and $T_{cold}$/$V_{cold}$ is the temperature/voltage ratio of the cold load. B-Machine has roughly $60$ dB of gain and uses a diode with a response of $0.5$ mV per $\mu\mathrm{W}$ observing a $15$ K sky with a system temperature of 45 K gives a diode level of $\sim3$ mV. Using the same calculation for a 300 K load gives a diode level of $\sim19$ mV consistent with the measured values. The Anristu 75K50 Microwave Detector Diodes output voltage is proportional to its incident power from 1 mV to 10 mV. When outside of this regime the diodes response is nonlinear. For B-Machine the warm load represents a data point outside of the linear response and hence Equation~\ref{eqn:gainlin} cannot be used. As an alternative the detector response can be modeled using,
\begin{center}
\begin{equation}\label{eqn:nonlincal}
V_{out}=\frac{G_{0}(T_{A}+T_{sys})}{1+bG_{0}(T_{A}+T_{sys})},
\end{equation}
\end{center}
where $V_{out}$ is the output voltage of the radiometer chain, $T_{A}$ is the antenna temperature , $T_{sys}$ is the system temperature, $G_{0}$ is the linear gain response and $b$ is the non-linearity parameter. For the case of a linear receiver $b$ goes to zero and Equation~\ref{eqn:nonlincal} reduces to Equation~\ref{eqn:lincal}.
There were 2 main testing phases for B-Machine in order to get all the necessary parameters so that a fit to Equation~\ref{eqn:nonlincal} could be made. The first phase of the measurements was performed at UCSB. With the detector warm, a test load was used which consisted of a large Aluminum box (no top) insulated with polystyrene sides and a piece of Eccosorb (AN72)\footnote{Emerson $\&$ Cuming Microwave Products} inside. The box was filled with Liquid Nitrogen (LN) so that the top of the Eccosorb was underneath the level of the LN. The test load was positioned directly under the RF window of the dewar and Aluminum walls were erected so that all beam paths ended in the LN. With the detector's beam filled with LN a blue foam load alternated with a room temperature piece of Eccosorb was inserted between the RF window and LN load. In this way the blue foam piece could be well characterized for later calibrations. With the detector warm the gain is low enough that all measurements are well within the linear regime of the diode. All measurements for Table~\ref{tab:BlueFoamChar} are done at the diode (no audio gain involved). Using a similar procedure for getting the blue foam temperature the system temperature was found. The main difference in the procedures is that the detector was cooled and tuned for observing. A compression estimate (based on Beast calibrations) from the diode level readings was used for each channel in getting the system temperature. These estimates turned out to be consistent with the compression numbers given by the model.
\begin{table}[p]
\begin{center}
\caption[Blue Foam Characterization]{Blue Foam Characterization \label{tab:BlueFoamChar}}
\begin{tabular}{|c|c|c|c|c|c|}
\hline
Channel & Offset & Hot & Cold & Blue Foam & Blue Foam\\
& mV & mV & mV & mV & K\\
\hline
\hline
1 / 9 & 1.65 & 6.47 & 4.55 & 4.94 & 44.8 \\
& 1.70 & 6.62 & 4.67 & 5.08 & 45.8 \\
\hline
2 / 10 & 1.23 & 9.44 & 6.48 & 7.08 & 44.2 \\
& 1.24 & 9.44 & 6.49 & 7.08 & 44.06\\
\hline
3 / 11 & 0.89 & 12.01 & 10.46 & 10.82 & 50.63\\
& 1.07 & 12.46 & 10.87 & 11.23 & 49.36\\
\hline
6 / 14 & 0.28 & 2.91 & 2.08 & 2.23 & 39.4 \\
& 0.29 & 2.88 & 2.07 & 2.22 & 40.3 \\
\hline
\end{tabular}
\end{center}
\end{table}
Before B-Machine was set to observe (beginning of season) and just before it was shutdown for the winter (end of season) a calibration sequence was done. The first calibration sequence, see Figure~\ref{fig:gain08072008}, consists of a 73.8 K load (LN filled polystyrene cooler that had several sheets of Eccosorb in it) placed in the beam path just below the dewar. This was followed by several tests of calibrated blue foam, white foam, ambient temperature calibrator and polypropylene in several combinations into the beam. A similar sequence was done at the end of the campaign, see Figure~\ref{fig:gain10142008}, modulo no blue foam. The importance of the blue foam calibration was not discovered until after the observing campaign had been finished and in depth data analysis started.
\begin{figure}[p]
\begin{center}
\includegraphics[width=17.0cm, angle=90]{fig/FullGain1Thesis.pdf}
\caption[Gain calibration sequence 08/07/2008]{Gain calibration sequence (08/07/2008) using LN (73.8 K), Eccosorb (ambient), and Blue Foam (intermediate) loads to calibrate the temperature gain of the system. The telescope was looking at sky when there was no load in the optical path. Each curve is one of the operational channels: red is channel 9, blue is channel 10, green is channel 11, and black is channel 14. \label{fig:gain08072008}}
\end{center}
\end{figure}
\begin{figure}[p]
\begin{center}
\includegraphics[width=17.5cm, angle=90]{fig/FullGain2Thesis.pdf}
\caption[Gain calibration sequence 10/14/2008]{Gain calibration sequence (10/14/2008) using LN (73.8 K) and Eccosorb (ambient)loads to calibrate the temperature gain of the system. The telescope was looking at sky when there was no load in the optical path. Each curve is one of the operational channels: red is channel 9, blue is channel 10, green is channel 11, and black is channel 14. \label{fig:gain10142008}}
\end{center}
\end{figure}
To accurately fit the curve from Equation~\ref{eqn:nonlincal}, voltages and temperatures for each of the thermal loads are found and fit to the model. Ambient temperature and LN temperature can be found from temperature sensors or barometric pressure, respectively and the voltages for these loads from data acquisition readouts. Finding system temperature and added blue foam temperature was determined in previous tests at UCSB. The remaining piece of information to get was the sky temperature which was determined by using sky dips, see Subsubsection~\ref{subsubsec:skytemp}. Fit parameters have been computed for all of the channels, see Table~\ref{tab:fitparameters}.
\subsubsection{Sky Temperature}\label{subsubsec:skytemp}
The sky temperature can be calculated using 2 different methods. First, by using the sky signal between thermal load sources (referred to as the DC method) and second by using a sky dip. The DC method suffers from several critical drawbacks: it is very sensitive to DC voltage drift and $\frac{1}{f}$ noise. By solving Equation~\ref{eqn:zenithtemp} for $T_{Zenith}$ (zenith temperature) and using the approximate gain calculated from thermal loads (estimating compression) gives the zenith sky temperature. A more reliable sky temperature, zenith temperature and rough estimates of the system temperatures can be made using sky dips. A sky dip is generated by slowly driving the elevation up or down giving a decreasing or increasing signal, respectively, from the change in thickness of the atmosphere. In Figure~\ref{fig:skydipfit08072008} the signal can be seen decreasing as a function of elevation. By fitting this signal to a model of the sky temperature, Equation~\ref{eqn:skytemp}, a zenith temperature and system temperature are found. The sky dip method for getting the sky temperature is less prone to systematic error from DC voltage drift and 1/f noise then the DC method. Scans made just after the calibration sequences can be modeled and fit to Equation~\ref{eqn:skytemp}, see Figures~\ref{fig:skydipfit08072008} and~\ref{fig:skydipfit10142008}. This also turns out to give a good fit to the system temperature. A constant 1.5 K is added to the sky temperature from the integration of the x2 signal, see Subsection~\ref{subsect:emissivity}, by the demodulation technique.
\begin{equation}\label{eqn:zenithtemp}
T_{Zenith}=(T_{LN}-((V_{sky}-V_{LN})*Gain))*\cos(elevation)
\end{equation}
where $T_{LN}$ is the temperature of LN, $V_{sky}$ is the voltage from the sky signal and $V_{LN}$ is the voltage from the LN signal.
\begin{equation}\label{eqn:skytemp}
T=T_{sys}+\frac{T_{Zenith}}{\sin(elevation)}
\end{equation}
\begin{table}[p]
\begin{center}
\caption[Fit Parameters for Calibrations]{Fit Parameters for Calibrations \label{tab:fitparameters}}
\begin{tabular}{|c|c|c|c|c|c|c|c|}
\hline
Channel & Max & & & Tsys & Tsys & $\triangle$T & $\triangle Q$ or $\triangle U$\\
AC/DC & Phase & $G_{0}$ & $b$ & DC & Sky Dip & $\mathrm{mK}/\sqrt{\mathrm{s}}$ & $\mathrm{mK}/\sqrt{\mathrm{s}}$\\
& & & & (K) & (K) & & \\
\hline
1/9 & 23 & 0.0108 & 0.00655 & 53.66 & 54.75& 1.74 & 1.93\\
\hline
2/10 & 7 & 0.0119 & 0.00608 & 52.95 & 56.43& 1.72 & 1.91\\
\hline
3/11 & 6 & 0.00723 & 0.0506 & 71.19 & 69.55& 2.10 & 2.33\\
\hline
6/14 & 14 & 0.0136 & 0.0178 & 40.20 & 43.88& 1.44 & 1.60\\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{figure}[p]
\includegraphics[width=\textwidth]{fig/SKydipfits08072008.pdf}
\caption[Sky dip data with fit from 08/07/2008]{Sky dip data for 08/07/2008: the black line is the fit data and the colored lines are observed data. \label{fig:skydipfit08072008}}
\end{figure}
\begin{figure}[p]
\begin{center}
\includegraphics[width=\textwidth]{fig/Skydipfits10142008.pdf}
\caption[Sky dip data with fit from 10/14/2008]{Sky dip data for 10/14/2008: the black line is the fit data and the colored lines are observed data. \label{fig:skydipfit10142008}}
\end{center}
\end{figure}
\subsubsection{Emissivity}\label{subsect:emissivity}
A difference in the signal level when observing sky between the wire grid and the aluminum plate (plane mirror) causes a synchronous signal that is $\frac{1}{2}$ of the periodicity of the polarization signal, see Figure~\ref{fig:x2}. This signal is referred to as the x2 signal. Though the signal level is high the demodulation technique (lock-in amplification) for $Q$ and $U$ is not sensitive to the x2 signal. The x2 signal is the convolution of several effects including loss of the wire grid, emissivity of the wires, emissivity of the blue foam spacer, and emissivity of the Aluminum plate. All of these effects combine to create an overall level difference in the signal when the feed horns are viewing sky off of the wire grid surface or the plane mirror surface.
\begin{figure}
\begin{center}
\includegraphics[width=14.6cm]{fig/x2thesis.pdf}
\caption[Times 2 signal all channels]{Times 2 signal from emissivity difference between wire grid and plane mirror Aluminium plate. Channel 6/14 is black, Channel 1/9 is red, channel 2/10 is blue, channel 3/11 is green, and the thick purple dashed line is a polarized signal on the central channel for periodicity comparison. The demodulation integrates the x2 signal to zero on the AC channels. \label{fig:x2}}
\end{center}
\end{figure}
Data with no wire grid and blue foam should not exhibit the same added signal. Hence a data set that contains both sky signal with and without the wire grid and blue foam spacer on the Polarization Rotator should yield the emissivity of the wire grid. In addition, a calibration must be performed relatively close to the above sequence to determine the sky temperature. Theoretically the gain is not necessary; all of the calculations could be done in voltage but due to diode compression from the data sets and diode levels not being adjusted for nominal scanning mode, calibration is a must.
To calibrate the data set that viewed the sky with and without the wire grid on, a data set made several hours before that contained a sky dip, a LN load and ambient load in the data was used to get the sky temperature. Also, in the calibrating data set was a section of data with the Polarization Rotator in. This was necessary due to extreme diode compression in the data set with and without data set. Assuming a moderate compression of 10$\%$ (consistent with previous diode compression calculations) a gain of $86.9\pm2.0\mathrm{~K/V}$ for the sky dip data and using a model to fit the zenith temperature gives a zenith temperature of $22.30\pm0.52^{\circ}\mathrm{C}$ or a sky temperature of $28.98\pm0.68^{\circ}\mathrm{C}$ at $50.3^{\circ}$ from horizon. Utilizing the sky temperature and the LN target the calibration for the data set with and without the wire grid is $30.34\pm0.76\mathrm{~K/V}$. The signal viewing the sky through the Polarization Rotator with the wire grid is of the form,
\begin{center}
\begin{equation}\label{eqn:Grideff}
T_{out}=T_{in}+T_{amb}(1-\epsilon_{grid}).
\end{equation}
\end{center}
Using the section of the data that is looking at the sky with no wire grid gives the input temperature and the section of data looking at the sky through the wire grid averaged over all sectors gives the output temperature. Solving for $\epsilon$ yields an efficiency of $99.33\pm0.20\%$ using the temperature sensor on the calibrator for the ambient temperature. Also available is an estimate of the efficiency of the polarization calibrator $80.0\pm2.0\%$. This number is not an estimate of the emissivity but rather the reflectivity of the Eccosorb, how well the piecemeal wire grids are put together, and the non-uniformity of the grids used. The compression at the ambient temperature side is dominating the error of these measurements.
\subsection{Polarization}
Converting the signal from voltage to temperature is not the optimal way to run this experiment because of the chopping method. Converting the AC channels into the $Q$ and $U$ Stokes parameters is the preferable mode of operation. The basic idea of getting a polarization calibration is to put a known polarized signal in front of the detector and then demodulate that section of data. The ratio of the demodulated voltage to the known Q or U signal (Q is used for this test) is the calibration constant. Following the temperature calibration sequence with LN on 08/07/2008 and 10/14/2008 the Polarization Calibrator was lowered into the beam and rotated. The Polarization Calibrator is a large round piece of Hexcel covered by several layers of Eccosorb, AN72, covered by a thin layer of Styrofoam. On top of the Styrofoam is a piecemeal wire grid. This wire grid consists of several panels of wire grids that didn't pass the quality control for the Polarization Rotator and were visually aligned and attached together. Finally, a large Styrofoam cover is put over the whole assembly and taped down to the Hexcel for thermal stability and mechanical strength. The Polarization Calibrator is mounted to a frame that rotates in and out of the optical path by a bushing system that lets it rotate about its center, see Figure~\ref{fig:PolCal}. A potentiometer is attached to the end of the bushing system to measure the relative position of the calibrator and the voltage is stored on channel 15 of the DAQ data files. The voltage is calibrated to get the exact position of the wires from the max phase measurement (see Section~\ref{subsec:GetMaxPhase}).
\begin{figure}
\begin{center}
\includegraphics[width = \textwidth]{fig/polcalbeam.pdf}
\caption[Optical design of B-Machine Polarization Calibrator]{Beam paths for the central horn with Polarization Calibrator in. \label{fig:PolCal}}
\end{center}
\end{figure}
The orientation of the wheel dictates the polarization angle and the difference between the sky (zenith) and ambient (Eccosorb) temperatures give the amplitude of the signal. Knowing the rough orientation of the wires on the Polarization Calibrator makes it possible to determine the matching Stokes parameter.
A data set with polarization calibrator in was taken and a section of demodulated data were found that represented a signal that was all $Q$ for the given channels. The revolution that corresponds to the max signal was used to find the minimum (sky zenith) and maximum (Eccosorb or ambient) voltage in the time ordered data (TOD). The temperature difference divided by 2 gives the corresponding $Q$ or $U$ Stokes parameter that the level 1 file for this revolution should give. Taking the ratio of the Stokes parameter and the voltage gives the gain that each channel should be multiplied by to get the corresponding Stokes parameter ($Q$ or $U$).
\begin{equation}\label{eqn:Stokes}
Q=\frac{T_{x}-T_{y}}{2}=\frac{f_{gain}(V_{eccco})-f_{gain}(V_{sky})}{2},
\end{equation}
Where $f_{gain}(\mathrm{~V})$ is the gain model which gives a temperature for the given voltage. See Table~\ref{tab:Polcal08072008} through Table~\ref{tab:Polcalall1014} for the relevant data for the polarization calibrations.
\begin{table}[p]
\caption{Polarization Calibration File Information 08/07/2008 \label{tab:Polcal08072008}}
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
Channel & Rev Num & Max (V) & Min (V) & $\Delta$ T (K) & $\Delta$ V (V)& Pol Gain \\
\hline
\hline
1/9 & 1462640 & -1.029 & -3.400 & 226.49 & 2.371 & 166.22 \\
\hline
2/10 & 1462487 & -1.086 & -3.688 & 224.69 & 2.602 & 148.47 \\
\hline
3/11 & 1461193 & -0.746 & -2.230 & 240.94 & 1.484 & 279.93 \\
\hline
6/14 & 1462484 & -0.939 & -3.750 & 225.804 & 2.911 & 139.284\\
\hline
\end{tabular}
\end{center}
\end{table}
A couple of points need to be made about Table~\ref{tab:polcalall}. First, the temperature calibration model previously derived was used. Secondly, another method of getting the polarization calibration might be to use the voltages to get an absolute temperature for both the min and max voltages and then calculate the Stokes parameter using these temperatures. The reason that this method wasn't used is it is very sensitive to any offset voltages that might have changed since the calibration sequences. The polarization calibration takes into account both the inefficiencies of the polarization calibrator and the wire grid. In addition, a small effect due to the non-symmetric wave form of the polarization signal is buried in the analysis. This effect was explored in some detail and by averaging the demodulation over a full revolution of the Polarization Rotator it can be removed.
\begin{table}[p]
\caption{Calibration Numbers Using All Peaks and Averages for Data Taken on 08/07/2008 \label{tab:polcalall}}
\begin{center}
\begin{tabular}{|c|c|c|c|c|}
\hline
Stokes & Channel 9 & Channel 10 & Channel 11 & Channel 14 \\
\hline
\hline
+Q & 166.574 & 148.385 & 278.396 & 141.609 \\
\hline
-Q & 165.054 & 148.408 & 280.452 & 142.110 \\
\hline
+U & 163.671 & 148.029 & 278.060 & 141.580 \\
\hline
-U & 167.597 & 148.064 & 279.176 & 141.262 \\
\hline
Average & 165.72 & 148.22 & 279.00 & 141.64 \\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{table}
\caption{Calibration Numbers Using All Peaks and Averages for Data Taken on 10/14/2008 \label{tab:Polcalall1014}}
\begin{center}
\begin{tabular}{|c|c|c|c|c|}
\hline
Stokes & Channel 9 & Channel 10 & Channel 11 & Channel 14 \\
\hline
\hline
+Q & 150.000 & 140.843 & 224.678 & 139.915\\
\hline
-Q & 149.191 & 140.244 & 226.366 & 140.473 \\
\hline
+U & 145.991 & 139.341 & 223.443 & 139.888 \\
\hline
-U & 150.260 & 138.980 & 224.293 & 139.718 \\
\hline
Average & 148.86 & 140.10 & 224.70 & 140.00 \\
\hline
\end{tabular}
\end{center}
\end{table}
As a simple test to check that the polarization calibration is consistent, a data set that contained a signal that was all $Q$ from the get max phase data, see Subsection~\ref{subsec:GetMaxPhase}, was normalized and shifted so that a signal that is roughly $0.2$ K peak to peak, was demodulated and a $Q$ signal of $106$ mK was seen. From the definition of $Q$ a polarized signal of $0.2$ K should give a $100$ mK $Q$ signal. The waveform of the Polarization Rotator doesn't yield itself to this type of analysis due to its asymmetries, but it is clear that the calibration for the polarization is at worst $6\%$ off.
\subsubsection{Getting Max Phase}\label{subsec:GetMaxPhase}
Each of the feed horns is coupled via circular to rectangular waveguide transition to the input of a Low Noise Amplifier (LNA). Rectangular waveguides will only carry radiation polarized perpendicular to its long dimension and each of the horns is arbitrarily orientated with respect to the central horn. Some effort was made to align the central horn so that the polarization was parallel to the horizon, but this was a gross alignment and was not used as a standard. To calibrate the horns alignment the telescope required a polarized signal of known polarization with a large signal to noise. When discussing local sources $Q$ is a source polarized vertically or horizontally and $U$ is a source polarized at $45^{\circ}$. Once a signal of say all Q is observed the signal is demodulated with a phase shift for the demodulation square wave of 0 to 31, the shift that yields the maximum signal is the max phase. The max phase corresponds to the number of sectors the Polarization Rotator must rotate from the zero position (the point at which the index pulse of the encoder triggers the sector rest and data sample) before the wires are lined up with the polarization of a given horn. The wire grid is subdivided into 128 sectors or 32 counts per polarization rotation.
To get the max phase for B-Machine, a person was sent to the north ridge of Mount Barcroft with a 41.5 GHz source mounted on a tripod. The source had 2 Aerowave uncalibrated attenuators and one calibrated Direct Read Attenuator (DRA). The source was leveled using a digital level, that is accurate up to $0.01^{\circ}$, to better than $0.1^{\circ}$ and pointed roughly at B-Machine. The source was setup such that when it is leveled on the tripod the emitted polarization is all $Q$. The signal level was adjusted with the use of the DRA so that it was on scale at its maximum for the DC channels. B-Machine then scanned the source taking care that the central horn (channels 6 and 14) saw the source and the signal was on scale. Following the central channel scan the polarization calibrator was placed into the beam and rotated several times for calibration of the other channels. This sequence was done at the beginning and end of the data taking campaign (08/07/2008 and 10/14/2008). Several Polarization Calibration sequences were also done at random intervals in the data taking campaign to confirm relative phase.
It was noticed after taking a look at the max phase data from the first day of calibration that a max phase of 13 or 14 could be achieved depending on the revolution number used for demodulation. This was caused by the source not being on the peak of Mt. Barcroft so the side lobes of the beam were seeing the mountain. Also, better care aligning the source would have made the measurement more accurate. For the second max phase calibration sequence much more care was taken into leveling and placement of the source, placing the source on the peak of Mt. Barcroft. These numbers were not dependent on the revolution for the second max phase calibration. Ultimately, a max phase of 14 for the central channel was used, based on the second calibration sequence and parts of the first.
Once the max phase for the central channel is established a calibration of the Polarization Rotator is easily made. The polarization angle of the Polarization Calibrator as a function of the output voltage on channel 15 can be used to get the max phase of the remaining channels. Data from the rotation of the Polarization Calibrator is demodulated and the revolution number corresponding to the maximum signal for channel 14 (all of the AC channels are not on scale) is the revolution number that corresponds to a signal that is all Q (in the telescope frame). This coincides with the wires being vertical on the Polarization Calibrator. Finding the phase that demodulates to the maximum value for each channel from this revolution of data gives the max phase.
\begin{table}[p]
\begin{center}
\caption[Maximum Phase and System Temperature]{Maximum Phase and System Temperature \label{tab:MaxPhaseTsys}}
\begin{tabular}{|c|c|c|c|c|}
\hline
Chan & Chan & Max & Tsys DC & Tsys \\
AC & DC & Phase & Measurements & sky dips \\
& & & (K) & (K) \\
\hline
\hline
1 & 9 & 23 & 53.66 & 54.75 \\
\hline
2 & 10 & 7 & 52.95 & 56.43 \\
\hline
3 & 11 & 6 & 71.19 & 69.55 \\
\hline
6 & 14 & 14 & 40.20 & 43.88 \\
\hline
\end{tabular}
\end{center}
\end{table}
\subsection{Error}
The errors in calibration will be propagated throughout the experiment to the final answer. It is important to understand the sources of error and thier impact on the calibration constants. The uncertainty based on the measurements can be quantified by looking at the variance of each of the values and combining them in the appropriate fashion, Equation~\ref{eqn:gainerror}.
\begin{center}
\begin{equation}\label{eqn:gainerror}
\Delta G^{2}=G(\frac{\sigma_{Th}^{2}}{T_{h}^{2}}+\frac{\sigma_{Tc}^{2}}{T_{c}^{2}}+\frac{\sigma_{Vh}^{2}}{V_{h}^{2}}+\frac{\sigma_{Vc}^{2}}{V_{c}^{2}})
\end{equation}
\end{center}
The standard deviation for each of the voltage values can be calculated from the data sets. The standard deviations in the temperature are a bit more difficult to quantify. Variations in the thermal conductivity of the Eccosorb loads, radiative loading, and small variations in the surface temperature can all lead to errors between the RF signal and the projected temperature of the loads. For the warm load a separation of statistical and systematic uncertainty is possible, since a temperature sensor is placed directly on the load. Looking at the data from the warm load temperature read out over a $2.2$ minute interval the temperature varied by $0.5$ K and the voltage from the radiometer varied by $0.03$ V ($\sim1.5$ K). This indicates that the surface of the warm load was radiatively cooling faster than the read out was changing. Using the extreme of this measurement of $\pm0.5$ K leads to a change of $0.5\%$ in the gain constant. Figure~\ref{fig:gainvstemp} shows how the calibration constant changes as a function of temperature for the warm and cold loads. The data that were used to get the warm load voltages was much shorter ($\sim$30 s) so this will safely over estimate the error in the warm load temperature. No readout was possible for the cold load temperature, hence an estimation of the deviation will need to be made. Using the entire section of cold load data and looking at the decay in the signal, gives an estimate of the change in cold load temperature as a function of time. The cold load section represents about $52$ seconds of data and shows a change in temperature of roughly $0.5$ K. The data that are used for the calibration sequence is $\frac{1}{4}$ of the cold load data, a small section in the middle of the run. This suggests that the deviation in temperature will be dominated by systematic unknowns such as beam filling, thermalization of Eccosorb, and reflectivity at the surface. A conservative estimate of the deviation from the temperature of LN at altitude ($\sim73.78$ K) is $0.3$ K. The error for the cold load won't be symmetric around the boiling point temperature, because the thermal load is much more likely to be warmer in this case than colder. The temperature of LN at altitude is found using the "Thermophysical Properties of Nitrogen" webpage from NIST. The results from the error calculations are all summarized in Table~\ref{tab:DCgains}.
\begin{table}[p]
\begin{center}
\caption[Voltages, Temperatures and Standard Deviations for Gain Measurements]{Voltages, Temperatures and Standard Deviations for Gain Measurements \label{tab:DCgains}}
\begin{tabular}{|c|c|c|c|c||c|c|c|c|}
\hline
&\multicolumn{4}{c||}{08-07-2008}&\multicolumn{4}{c|}{10-14-2008}\\
\hline
Channel & 9 & 10 & 11 & 14 & 9 & 10 & 11 & 14 \\
\hline
\hline
$V_{hot}$ & 3.91 & 4.47 & 2.58 & 4.66 & 3.88 & 4.23 & 3.01 & 4.40\\
\hline
$\sigma_{Vhot}$ & 0.011 & 0.011 & 0.008 & 0.008 & 0.008 & 0.008 & 0.008 & 0.005 \\
\hline
$V_{cold}$ & 1.35 & 1.49 & 1.03 & 1.49 & 1.41& 1.49 & 1.11 & 1.50\\
\hline
$\sigma_{Vcold}$ & 0.0069 & 0.007 & 0.006 & 0.007 & 0.003 & 0.003 & 0.004 & 0.002 \\
\hline
$T_{hot}$ & 296.8 & 296.8 & 296.8 & 296.8 & 273.46 & 273.46 &273.46 & 273.46\\
\hline
$\sigma_{Thot}^{*}$ & 0.144 & 0.144 & 0.144 & 0.144 & 0.08 & 0.08 & 0.08 & 0.08 \\
\hline
$\sigma_{Thot}$ & 0.5 & 0.5 & 0.5 & 0.5 & 0.5 & 0.5 & 0.5 & 0.5 \\
\hline
$T_{cold}$ & 73.8 & 73.8 & 73.8 & 73.8& 73.8 & 73.8 & 73.8 & 73.8\\
\hline
$\sigma_{Tcold}$ & 0.25 & 0.25 & 0.25 & 0.25& 0.25 & 0.25 & 0.25 & 0.25 \\
\hline
Error$\pm$ & 0.48 & 0.41 & 0.71 & 0.35 & 0.43 & 0.38 & 0.67 & 0.33\\
\hline
Cal & 87.23 & 74.60 & 118.47 & 69.09 & 80.25 & 72.56 & 105.20 & 68.62 \\
\hline
\end{tabular}
\end{center}
\end{table}
All voltages are in Volts, temperatures are in Kelvin and gains are in Kelvin per Volt. The 2 deviations in the warm load temperature have been added in quadrature to get the standard deviation for the final calculation. The first warm load standard deviation is statistical while the second is systematic.
\begin{figure}[p]
\begin{center}
\includegraphics[width=14.6cm]{fig/Gainvstemp.pdf}
\caption[Gain changes from temperature variations in load]{Gain of channel 14 for small variations in load temperatures. The cold load calibrations have been offset by $0.5$ K/V so that the lines are distinguishable from each other. \label{fig:gainvstemp}}
\end{center}
\end{figure}
\section{General Telescope Properties}
In addition to RF and beam characterization other properties of the telescope were quantified for observation.
\subsection{Servo System}
The original controls and programs for controlling the telescope were written for a telescope that scanned between 2 fixed azimuths \citep{levy08}. While B-Machine is set up to continuously scan in azimuth, the necessity for raster ability was not deemed important in the original design. While attempting to raster scan a small source for a full 2-D beam map it was found that the drive system was not finely tuned enough to allow for raster scanning. As a result all point sources both terrestrial and extra solar were observed by fixed elevation scans while slowly scanning between 2 azimuths letting the source drift through (frequently referred to as a drift scan).
\subsection{Scan Strategy} \label{subsec:scanstrategy}
B-Machine needed to balance sample rate limitations, beam smearing effects and beam overlap in the nominal scan strategy for day to day operation. To measure the maximum sample rate, the rate of the Polarization Rotator was slowly increased (rate of rotator proportional to the input voltage) until missed samples in the data were spotted. A final rotation rate of the primary table was determined once a maximal rotation rate of $33.4$ Hz was determined. Using a $70$ s scan rate gives approximately 9 samples per beam in azimuth. In addition the sky will drift approximately $17.5'$ at the worst point per rotation giving 2 samples per beam in elevation. This allows for maximal sky coverage with minimal missing sky area between rotations. Ultimately the rotation rate of B-Machine is held hostage by the upper limit of the data acquisition system.
\subsection{Thermopile}
B-Machine was also equipped with a Thermopile device\footnote{Dexter research, Inc. M5 Thermopile Detector}, which produces a voltage proportional to the difference in temperature of its housing vs. observed temperature. The thermopile has a large field of view and consequently a long Aluminum tube was placed on the end to narrow its field of view from $78^{\circ} \mathrm{~to~} \sim20^{\circ}$. Final analysis of this data will not be addressed in any detail here and will be combined with a data set that is being taken at WMRS with additional Infra-Red sensors.
|
\section{Introduction}
\vspace{-0.1in}
In undirected graphical models or Markov random fields, each node
represents a random variable while the set of edges specifies the
conditional independencies of the underlying distribution. When the
random variables are jointly Gaussian, the models are called \textit{Gaussian
graphical models }(GGMs) or \textit{Gauss Markov random fields}. GGMs,
such as linear state space models, Bayesian linear regression models,
and thin-membrane/thin-plate models, have been widely used in communication,
image processing, medical diagnostics, and gene regulatory networks.
In general, a larger family of graphs represent a larger collection
of distributions and thus can better approximate arbitrary empirical
distributions. However, many graphs lead to computationally expensive
inference and learning algorithms. Hence, it is important to study
the trade-off between modeling capacity and efficiency.
Both inference and learning are efficient for tree-structured graphs
(graphs without cycles): inference can be computed exactly in linear
time (with respect to the size of the graph) using belief propagation
(BP) \citep{pearl1986constraint} while the learning problem can be
solved exactly in quadratic time using the Chow-Liu algorithm \citep{chow1968approximating}.
Since trees have limited modeling capacity, many models beyond trees
have been proposed \citep{choi2009exploiting,comer1999segmentation,bouman1994multiscale,karger2001learning}.
Thin junction trees (graphs with low tree-width) are extensions of
trees, where inference can be solved efficiently using the junction
algorithm \citep{jordan2004graphical}. However, learning junction
trees with tree-width greater than one is NP-complete \citep{karger2001learning}
and tractable learning algorithms (e.g. \citep{abbeel2006learning})
often have constraints on both the tree-width and the maximum degree.
Since graphs with large-degree nodes are important in modeling applications
such as social networks, flight networks, and robotic localization,
we are interested in finding a family of models that allow arbitrarily
large degrees while being tractable for learning.
Beyond thin-junction trees, the family of sparse GGMs is also widely
studied \citep{dobra2004sparse,tipping2001sparse}. These models are
often estimated using methods such as graphical lasso (or $l$-1 regularization)
\citep{friedman2008sparse,ravikumar2008model}. However, a sparse
GGM (e.g. a grid) does not automatically lead to efficient algorithms
for exact inference. Hence, we are interested in finding a family
of models that are not only sparse but also have guaranteed efficient
inference algorithms.
In this paper, we study the family of GGMs with small feedback vertex
sets (FVSs), where an FVS is a set of nodes whose removal breaks all
cycles \citep{vazirani2004approximation}. The authors of \citep{liu2012feedback}
have demonstrated that the computation of exact means and variances
for such a GGM can be accomplished, using message-passing algorithms
with complexity $\calO(k^{2}n)$, where $k$ is the size of the FVS
and $n$ is the total number of nodes. They have also presented results
showing that for models with larger FVSs, approximate inference (obtained
by replacing a full FVS by a pseudo-FVS) can work very well, with
empirical evidence indicating that a pseudo-FVS of size $\calO(\log n)$
gives excellent results. In Appendix \ref{sec:AppenPartitionFunction}
we will provide some additional analysis of inference for such models
(including the computation of the partition function), but the main
focus is maximum likelihood (ML) \textit{learning} of models with
FVSs of modest size, including identifying the nodes to include in
the FVS.
In particular, we investigate two cases. In the first, all of the
variables, including any to be included in the FVS are observed. We
provide an algorithm for exact ML estimation that, regardless of the
maximum degree, has complexity $\calO(kn^{2}+n^{2}\log n)$ if the
FVS nodes are identified in advance and polynomial complexity if the
FVS is to be learned and of bounded size. Moreover, we provide an
approximate and much faster greedy algorithm when the FVS is unknown
\textit{and} large. In the second case, the FVS nodes are taken to
be latent variables. In this case, the structure learning problem
corresponds to the (exact or approximate) decomposition of an inverse
covariance matrix into the sum of a tree-structured matrix and a low-rank
matrix. We propose an algorithm that iterates between two projections,
which can also be interpreted as alternating \textit{low-rank} corrections.
We prove that even though the second projection is onto a highly non-convex
set, it is carried out exactly, thanks to the properties of GGMs of
this family. By carefully incorporating efficient inference into the
learning steps, we can further reduce the complexity to $\calO(kn^{2}+n^{2}\log n)$
per iteration. We also perform experiments using both synthetic
data and real data of flight delays to demonstrate the modeling capacity
with FVSs of various sizes. We show that empirically the family of
GGMs of size $\calO(\log n)$ strikes a good balance between the modeling
capacity and efficiency.
\vspace{-0.15in}
\paragraph*{Related Work}
In the context of classification, the authors of \citep{friedman1997bayesian}
have proposed the tree augmented naive Bayesian model, where the class
label variable itself can be viewed as a size-one observed FVS; however,
this model does not naturally extend to include a larger FVS. In \citep{chandrasekaran2010latent},
a convex optimization framework is proposed to learn GGMs with latent
variables, where conditioned on a small number of latent variables,
the remaining nodes induce a sparse graph. In our setting with latent
FVSs, we further require the sparse subgraph to have tree structure.
\vspace{-0.2in}
\section{Preliminaries}
\vspace{-0.2in}
Each undirected graphical model has an underlying graph $\calG=(\calV,\calE)$,
where $\calV$ denotes the set of vertices (nodes) and $\calE$ the
set of edges. Each node $s\in\calV$ corresponds to a random variable
$x_{s}$. When the random vector $\bx_{\calV}$ is jointly Gaussian,
the model is a GGM with density function given by $p(\mathbf{x})=\frac{1}{Z}\exp\{-\frac{1}{2}\mathbf{x}^{T}J\mathbf{x}+\bh^{T}\mathbf{x}\}$,
where $J$ is the \textit{information matrix} or\textit{ precision
matrix}, $\bh$ is the \textit{potential vector}, and $Z$ is the
\textit{partition function}. The parameters $J$ and $\bh$ are related
to the mean $\boldsymbol{\mu}$ and covariance matrix $\Sigma$ by
$\boldsymbol{\mu}=J^{-1}\bh$ and $\Sigma=J^{-1}$. The structure
of the underlying graph is revealed by the sparsity pattern of $J$:
there is an edge between $i$ and $j$ if and only if $J_{ij}\neq0$.
Given samples $\{\bx^{i}\}_{i=1}^{s}$ independently generated from
an unknown distribution $q$ in the family $\calQ$, the ML estimate
is defined as $q_{\text{ML}}=\arg\min_{q\in\calQ}\sum_{i=1}^{s}\log q(\bx^{i}).$
For Gaussian distributions, the empirical distribution is $\hat{p}(\bx)=\calN(\bx;\hat{\bmu},\hat{\Sigma})$,
where the empirical mean $\hat{\bmu}=\frac{1}{s}\sum_{i=1}^{s}\bx^{i}$
and the empirical covariance matrix $\hat{\Sigma}=\frac{1}{s}\sum_{i=1}^{s}\bx^{i}\left(\bx^{i}\right)^{T}-\hat{\bmu}\hat{\bmu}^{T}$.
The Kullback-Leibler (K-L) divergence between two distributions $p$
and $q$ is defined as $D_{\text{KL}}(p||q)=\int p(\bx)\log\frac{p(\bx)}{q(\bx)}\mathrm{d}\bx$.
Without loss of generality, we assume in this paper the means are
zero.
Tree-structured models are models whose underlying graphs do not have
cycles. The ML estimate of a tree-structured model can be computed
exactly using the Chow-Liu algorithm \citep{chow1968approximating}.
We use $\Sigma_{\text{CL}}=\text{CL}(\hat{\Sigma})$ and $\calE_{\text{CL}}=\text{CL}_{\calE}(\hat{\Sigma})$
to denote respectively the covariance matrix and the set of edges
learned using the Chow-Liu algorithm where the samples have empirical
covariance matrix $\hat{\Sigma}$.
\vspace{-0.15in}
\section{Gaussian Graphical Models with Known FVSs\label{sec:Gaussian-Graphical-Models}}
\vspace{-0.15in}
\noindent \label{subsec:FVS}In this section we briefly discuss some
of the ideas related to GGMs with FVSs of size $k$, where we will
also refer to the nodes in the FVS as \textit{feedback nodes}. An
example of a graph and its FVS is given in Figure \ref{fig:FVS},
where the full graph (Figure \ref{fig:FVSfullgraph}) becomes a cycle-free
graph (Figure \ref{fig:FVStreepart}) if nodes 1 and 2 are removed,
and thus the set $\{1,2\}$ is an FVS. \vspace{-0.2in}
\begin{figure}[H]
\centering{
\begin{minipage}[t]{0.7\columnwidth
\subfloat[]{\includegraphics[width=0.35\columnwidth]{FVSfullgraph2}\label{fig:FVSfullgraph}
}\hfill{}\subfloat[]{\includegraphics[bb=0bp -50bp 360bp 216bp,width=0.35\columnwidth]{FVStreepart2}\label{fig:FVStreepart}
}\vspace{-0.1in}
\caption{A graph with an FVS of size 2. (a) Full graph; (b) Tree-structured
subgraph after removing nodes 1 and 2}
\label{fig:FVS
\end{minipage}
\end{figure}
\vspace{-0.2in}
\noindent Graphs with small FVSs have been studied in various contexts.
The authors of \citep{dinneen2001forbidden} have characterized the
family of graphs with small FVSs and their obstruction sets (sets
of forbidden minors). FVSs are also related to the ``stable sets''
in the study of tournaments \citep{brandt2011minimal}.
Given a GGM with an FVS of size $k$ (where the FVS may or may not
be given), the marginal means and variances $\bmu_{i}=\left(J^{-1}\bh\right)_{i}$
and $\Sigma_{ii}=\left(J^{-1}\right)_{ii}$, for $\forall i\in\calV$
can be computed \textit{exactly} with complexity $\calO(k^{2}n)$
using the feedback message passing (FMP) algorithm proposed in \citep{liu2012feedback},
where standard BP is employed two times on the cycle-free subgraph
among the non-feedback nodes while a special message-passing protocol
is used for the FVS nodes. We provide a new algorithm in Appendix
\ref{sec:acceleratedLatent}, to compute $\det J$, the determinant
of $J$, and hence the partition function of such a model with complexity
$\calO(k^{2}n)$. The algorithm is described and proved in Appendix
\ref{sec:AppenPartitionFunction}.
An important point to note is that the complexity of these algorithms
depends simply on the size $k$ and the number of nodes $n$. There
is no loss in generality in assuming that the size-$k$ FVS $F$ is
fully connected and each of the feedback nodes has edges to every
non-feedback node. In particular, after re-ordering the nodes so that
the elements of $F$ are the first $k$ nodes ($T=V\backslash F$
is the set of non-feedback nodes of size $n-k$), we have that $J=\left[\begin{array}{cc}
J_{F} & J_{M}^{T}\\
J_{M} & J_{T}
\end{array}\right]\succ0,$ where $J_{T}\succ0$ corresponds to a tree-structured subgraph among
the non-feedback nodes, $J_{F}\succ0$ corresponds to a complete graph
among the feedback nodes, and all entries of $J_{M}$ may be non-zero
as long as $J_{T}-J_{M}J_{F}^{-1}J_{M}^{T}\succ0$ (while $\Sigma=\left[\begin{array}{cc}
\Sigma_{F} & \Sigma{}_{M}^{T}\\
\Sigma_{M} & J_{T}
\end{array}\right]=J^{-1}\succ0$ ). We will refer to the family of such models with a given FVS $F$
as $\calQ_{F}$, and the class of models with some FVS of size at
most $k$ as $\calQ_{k}$
\footnote{In general a graph does not have a unique FVS. The family of graphs
with FVSs of size $k$ includes all graphs where there \textit{exists
}an FVS of size $k$
} If we are not explicitly given an FVS, though the problem of finding
an FVS of minimal size is NP-complete, the authors of \citep{bafna1992}
have proposed an efficient algorithm with complexity $\calO(\min\{m\log n,\ n^{2}\})$,
where $m$ is the number of edges, that yields an FVS at most twice
the minimum size (thus the inference complexity is increased only
by a constant factor). However, the main focus of this paper, explored
in the next section, is on \textit{learning }models with small FVSs
(so that when learned, the FVS is \textit{known}). As we will see,
the complexity of such algorithms is manageable. Moreover, as our
experiments will demonstrate, for many problems, quite modestly sized
FVSs suffice. \vspace{-0.15in}
\section{Learning GGMs with Observed or Latent FVS of Size $k$ \vspace{-0.15in}
}
In this section, we study the problem of recovering a GGM from $i.i.d.$
samples, where the feedback nodes are either observed or latent variables.
If all nodes are observed, the empirical distribution $\hat{p}(\bx_{F},\bx_{T})$
is parametrized by the empirical covariance matrix $\hat{\Sigma}=\left[\begin{array}{cc}
\hat{\Sigma}_{F} & \hat{\Sigma}_{M}^{T}\\
\hat{\Sigma}_{M} & \hat{\Sigma}_{T}
\end{array}\right]$. If the feedback nodes are latent variables, the empirical distribution
$\hat{p}(\bx_{T})$ has empirical covariance matrix $\hat{\Sigma}_{T}$.
With a slight abuse of notation, for a set $A\subset\calV$, we use
$q(\bx_{A})$ to denote the marginal distribution of $\bx_{A}$ under
a distribution $q(\bx_{\calV})$. \vspace{-0.1in}
\subsection{When All Nodes Are Observed}
\vspace{-0.1in}
When all nodes are observed, we have two cases: 1) When an FVS of
size $k$ is given, we propose the \textit{conditioned Chow-Liu algorithm,}
which computes the \textit{exact }ML estimate efficiently; 2) When
no FVS is given \textit{a priori}, we propose both an exact algorithm
and a greedy approximate algorithm for computing the ML estimate.\vspace{-0.1in}
\subsubsection{Case 1: An FVS of Size $k$ Is Given.}
\vspace{-0.1in}
\label{sub:knownFVS}
When a size-$k$ FVS $F$ is given, the learning problem becomes solving
\vspace{-0.15in}
\begin{alignat}{1}
p_{\text{ML}}(\bx_{F},\bx_{T}) & =\underset{q(\bx_{F},\bx_{T})\in\calQ_{F}}{\arg\min}D_{\text{KL}}(\hat{p}(\bx_{F},\bx_{T})||q(\bx_{F},\bx_{T})).\label{eq:knownFVS}
\end{alignat}
\vspace{-0.1in}
This optimization problem is defined on a highly non-convex set $\calQ_{F}$
with combinatorial structures: indeed, there are $(n-k)^{n-k-2}$
possible spanning trees among the subgraph induced by the non-feedback
nodes. However, we are able to solve Problem \eqref{eq:knownFVS}
\textit{exactly} using the conditioned Chow-Liu algorithm described
in Algorithm \ref{Algo:givenF}
\footnote{Note that the conditioned Chow-Liu algorithm here is different from
other variations of the Chow-Liu algorithm such as in \citep{kirshner2004conditional}
where the extensions are to enforce the inclusion or exclusion of
a set of edges.
}\textit{ }The intuition behind this algorithm is that even though
the entire graph is not tree, the subgraph induced by the non-feedback
nodes (which corresponds to the distribution of the non-feedback nodes
conditioned on the feedback nodes) has tree structure, and thus we
can find the best tree among the non-feedback nodes using the Chow-Liu
algorithm applied on the conditional distribution. To obtain a concise
expression, we also exploit a property of Gaussian distributions:
the conditional information matrix (the information matrix of the
conditional distribution) is simply a submatrix of the whole information
matrix. In Step 1 of Algorithm \ref{Algo:givenF}, we compute the
conditional covariance matrix using the Schur complement, and then
in Step 2 we use the Chow-Liu algorithm to obtain the best approximate
$\Sigma_{\text{CL}}$ (whose inverse is tree-structured). In Step
3, we match exactly the covariance matrix among the feedback nodes
and the covariance matrix between the feedback nodes and the non-feedback
nodes. For the covariance matrix among the non-feedback nodes, we
add the matrix subtracted in Step 1 back to $\Sigma_{\text{CL}}$.
Proposition \ref{prop:givenF} states the correctness and the complexity
of Algorithm \ref{Algo:givenF}. Its proof included in Appendix \ref{sec:Appen_Prop1}.We
denote the output covariance matrix of this algorithm as $\text{CCL}(\hat{\Sigma})$.\vspace{-0.1in}
\begin{algorithm}[H]
\centering{}\textbf{
\begin{minipage}[t]{0.9\columnwidth
\textbf{Input:} $\hat{\Sigma}\succ0$ and an FVS $F$
\textbf{Output:} $\calE_{\text{ML}}$ and $\Sigma_{\text{ML}}$\textbf{ }
\begin{enumerate}
\item Compute the conditional covariance matrix $\hat{\Sigma}_{T|F}=\hat{\Sigma}_{T}-\hat{\Sigma}_{M}\hat{\Sigma}_{F}^{-1}\hat{\Sigma}_{M}^{T}$
.
\item Let $\Sigma_{\text{CL}}=\text{CL}(\hat{\Sigma}_{T|F})$ and $\calE_{\text{CL}}=\text{CL}_{\calE}(\hat{\Sigma}_{T|F})$.
\item $\calE_{\text{ML}}=\calE_{\text{\text{CL}}}$ and $\Sigma_{\text{ML}}=\left[\begin{array}{cc}
\hat{\Sigma}_{F} & \hat{\Sigma}_{M}^{T}\\
\hat{\Sigma}_{M} & \Sigma_{\text{CL}}+\hat{\Sigma}_{M}\hat{\Sigma}_{F}^{-1}\hat{\Sigma}_{M}^{T}
\end{array}\right]$.
\end{enumerate}
\caption{The conditioned Chow-Liu algorithm}
\label{Algo:givenF
\end{minipage}
\end{algorithm}
\begin{proposition}
Algorithm \ref{Algo:givenF} computes the ML estimate $\Sigma_{\text{ML}}$
and $\calE_{\text{ML}}$, exactly with complexity $\calO(kn^{2}+n^{2}\log n)$.
In addition, all the non-zero entries of $J_{\text{ML}}\stackrel{\Delta}{=}\Sigma_{\text{ML}}^{-1}$
can be computed with extra complexity $\calO(k^{2}n)$.
\label{prop:givenF}
\end{proposition}
\vspace{-0.1in}
\subsubsection{Case 2: The FVS Is to Be Learned\vspace{-0.1in}
}
Structure learning becomes more computationally involved when the
FVS is unknown. In this subsection, we present both exact and approximate
algorithms for learning models with FVS of size no larger than $k$
(i.e., in $\calQ_{k}$). For a fixed empirical distribution $\hat{p}(\bx_{F},\bx_{T})$,
we define $d(F)$, a set function of the FVS $F$ as the minimum value
of \eqref{eq:knownFVS}, i.e.,\vspace{-0.15in}
\begin{equation}
d(F)=\min_{q(\bx_{F},\bx_{T})\in\calQ_{F}}D_{\text{KL}}(\hat{p}(\bx_{F},\bx_{T})||q(\bx_{F},\bx_{T})).\label{eq:setfunction}
\end{equation}
\vspace{-0.1in}
When the FVS is unknown, the ML estimate can be computed exactly by
enumerating all possible $\dbinom{n}{k}$ FVSs of size $k$ to find
the $F$ that minimizes $d(F)$. Hence, the exact solution can be
obtained with complexity $\calO(n^{k+2}k)$, which is polynomial in
$n$ for fixed $k$. However, as our empirical results suggest, choosing
$k=\calO(\log(n))$ works well, leading to quasi-polynomial complexity
even for this exact algorithm. That observation notwithstanding, the
following greedy algorithm (Algorithm \ref{algo:greedynew}), which,
at each iteration, selects the single best node to add to the current
set of feedback nodes, has polynomial complexity for arbitrarily large
FVSs. As we will demonstrate, this greedy algorithm works extremely
well in practice.\vspace{-0.15in}
\begin{algorithm}[H]
\centering{
\begin{minipage}[t]{0.9\columnwidth
\textbf{Initialization: $F_{0}=\emptyset$}
\textbf{For $t=1$ to $k$,}
\begin{alignat*}{1}
k_{t}^{*} & =\underset{k\in V\backslash F_{t-1}}{\arg\min}d(F_{t-1}\cup\{k\}),\ F_{t}=F_{t-1}\cup\{k_{t}^{*}\}
\end{alignat*}
\begin{center}
\caption{Selecting an FVS by a greedy approach}
\vspace{-0.1in}
\label{algo:greedynew}
\par\end{center
\end{minipage}
\end{algorithm}
\vspace{-0.15in}
\subsection{When the FVS Nodes Are Latent Variables}
When the feedback nodes are latent variables, the marginal distribution
of observed variables (the non-feedback nodes in the true model) has
information matrix $\tilde{J}_{T}=\hat{\Sigma}_{T}^{-1}=J_{T}-J_{M}J_{F}^{-1}J_{M}^{T}$.
If the exact $\tilde{J}_{T}$ is known, the learning problem is equivalent
to decomposing a given inverse covariance matrix $\tilde{J}_{T}$
into the sum of a tree-structured matrix $J_{T}$ and a rank-$k$
matrix $-J_{M}J_{F}^{-1}J_{M}^{T}$
\footnote{It is easy to see that different models having the same $J_{M}J_{F}^{-1}J_{M}$
cannot be distinguished using the samples, and thus without loss of
generality we can assume $J_{F}$ is normalized to be the identify
matrix in the final solution
} In general, use the ML criterion\vspace{-0.1in}
\begin{alignat}{1}
q_{\text{ML}}(\bx_{F},\bx_{T}) & =\arg\min_{q(\bx_{F},\bx_{T})\in Q_{F}}D_{\text{KL}}(\hat{p}(\bx_{T})||q(\bx_{T})),\label{eq:latent_objective}
\end{alignat}
\vspace{-0in}
where the optimization is over all nodes (latent and observed) while
the K-L divergence in the objective function is defined on the marginal
distribution of the observed nodes only.
We propose \textit{the latent Chow-Liu algorithm, }an alternating
projection algorithm that is a variation of the EM algorithm and can
be viewed as an instance of the majorization-minimization algorithm.
The general form of the algorithm is as follows:
\begin{enumerate}
\item Project onto the empirical distribution:
\[
\hat{p}^{(t)}(\bx_{F},\bx_{T})=\hat{p}(\bx_{T})q^{(t)}(\bx_{F}|\bx_{T}).
\]
\item Project onto the best fitting structure on all variables:
\[
q^{(t+1)}(\bx_{F},\bx_{T})=\arg\min_{q(\bx_{F},\bx_{T})\in\calQ_{F}}D_{\text{KL}}(\hat{p}^{(t)}(\bx_{F},\bx_{T})||q(\bx_{F},\bx_{T})).
\]
\end{enumerate}
In the first projection, we obtain a distribution (on both observed
and latent variables) whose marginal (on the observed variables) matches
exactly the empirical distribution while maintaining the conditional
distribution (of the latent variables given the observed ones). In
the second projection we compute a distribution (on all variables)
in the family considered that is the closest to the distribution obtained
in the first projection. We found that among various EM type algorithms,
this formulation is the most revealing for our problems because it
clearly relates the second projection to the scenario where an FVS
$F$ is both observed and known (Section \ref{sub:knownFVS}). Therefore,
we are able to compute the second projection \textit{exactly} even
though the graph structure is \textit{unknown} (which allows \textit{any}
tree structure among the observed nodes). Note that when the feedback
nodes are latent, we do not need to select the FVS since it is simply
the set of latent nodes. This is the source of the simplification
when we use latent nodes for the FVS: We have no search of sets of
observed variables to include in the FVS. \vspace{-0.15in}
\begin{algorithm}[H]
\begin{centering}
\textbf{
\begin{minipage}[t]{0.95\columnwidth
\textbf{Input:} the empirical covariance matrix $\hat{\Sigma}_{T}$
\textbf{Output:} information matrix $J=\left[\begin{array}{cc}
J_{F} & J_{M}^{T}\\
J_{M} & J_{T}
\end{array}\right]$, where $J_{T}$ is tree-structured
\begin{enumerate}
\item \noindent Initialization: $J^{(0)}=\left[\begin{array}{cc}
J_{F}^{(0)} & \left(J_{M}^{(0)}\right)^{T}\\
J_{M}^{(0)} & J_{T}^{(0)}
\end{array}\right]$.
\item \noindent Repeat for $t=1,2,3,\ldots$:
\begin{enumerate}
\item \noindent \textbf{P1}: Project to the empirical distribution:\\
$\hat{J}^{(t)}=\left[\begin{array}{cc}
J_{F}^{(t)} & (J_{M}^{(t)})^{T}\\
J_{M}^{(t)} & \left(\hat{\Sigma}_{T}\right)^{-1}+J_{M}^{(t)}(J_{F}^{(t)})^{-1}(J_{M}^{(t)})^{T}
\end{array}\right]$. Define $\hat{\Sigma}^{(t)}=\left(\hat{J}^{(t)}\right)^{-1}$.
\item \noindent \textbf{P2:} Project to the best fitting structure:\\
$\Sigma^{(t+1)}=\left[\begin{array}{cc}
\hat{\Sigma}_{F}^{(t)} & \left(\hat{\Sigma}_{M}^{(t)}\right)^{T}\\
\hat{\Sigma}_{M}^{(t)} & \text{CL}(\hat{\Sigma}_{T|F}^{(t)})+\hat{\Sigma}_{M}^{(t)}\left(\hat{\Sigma}_{F}^{(t)}\right)^{-1}\left(\hat{\Sigma}_{M}^{(t)}\right)^{T}
\end{array}\right]=\text{CCL}(\hat{\Sigma}^{(t)}),$\\
where $\hat{\Sigma}_{T|F}^{(t)}=\hat{\Sigma}_{T}^{(t)}-\hat{\Sigma}_{M}^{(t)}\left(\hat{\Sigma}_{F}^{(t)}\right)^{-1}\left(\hat{\Sigma}_{M}^{(t)}\right)^{T}$.
Define $J^{(t+1)}=\left(\Sigma^{(t+1)}\right)^{-1}.$
\end{enumerate}
\end{enumerate}
\caption{The latent Chow-Liu algorithm}
\label{algo:GaussEM
\end{minipage}
\par\end{centering}
\vspace{-0in}
\end{algorithm}
\vspace{-0.15in}
In Algorithm \ref{algo:GaussEM} we summarize the latent Chow-Liu
algorithm specialized for our family of GGMs, where both projections
have exact closed-form solutions and exhibit complementary structure---one
using the covariance and the other using the information parametrization.
In projection \textbf{P1}, three blocks of the information matrix
remain the same; In projection \textbf{P2,} three blocks of the covariance
matrix remain the same.
The two projections in Algorithm \ref{algo:GaussEM} can also be interpreted
as alternating \textit{low-rank} corrections : indeed,
\begin{eqnarray*}
\text{ }\text{In }\boldsymbol{\text{P1 }} & \hat{J}^{(t)} & =\left[\begin{array}{cc}
\bzero & \bzero\\
\bzero & \left(\hat{\Sigma}_{T}\right)^{-1}
\end{array}\right]+\left[\begin{array}{c}
J_{F}^{(t)}\\
J_{M}^{(t)}
\end{array}\right]\left(J_{F}^{(t)}\right)^{-1}\left[\begin{array}{cc}
J_{F}^{(t)} & \left(J_{M}^{(t)}\right)^{T}\end{array}\right],\\
\text{and in }\boldsymbol{\text{P2 }} & \Sigma^{(t+1)} & =\left[\begin{array}{cc}
\bzero & \bzero\\
\bzero & \text{CL}(\hat{\Sigma}_{T|F})
\end{array}\right]+\left[\begin{array}{c}
\hat{\Sigma}_{F}^{(t)}\\
\hat{\Sigma}_{M}^{(t)}
\end{array}\right]\left(\hat{\Sigma}_{F}^{(t)}\right)^{-1}\left[\begin{array}{cc}
\hat{\Sigma}_{F}^{(t)} & \left(\hat{\Sigma}_{M}^{(t)}\right)^{T}\end{array}\right],
\end{eqnarray*}
where the second terms of both expressions are of low-rank when the
size of the latent FVS is small. This formulation is the most intuitive
and simple, but a naive implementation of Algorithm \ref{algo:GaussEM}
has complexity ${\cal O}(n^{3})$ per iteration, where the bottleneck
is inverting full matrices $\hat{J}^{(t)}$ and $\Sigma^{(t+1)}$.
By carefully incorporating the inference algorithms into the projection
steps, we are able to further exploit the power of the models and
reduce the per-iteration complexity to $\calO(kn^{2}+n^{2}\log n)$,
which is the same as the complexity of the conditioned Chow-Liu algorithm
alone. We have the following proposition.
\begin{proposition}
Using Algorithm \ref{algo:GaussEM}, the objective function of \eqref{eq:latent_objective}
decreases with the number of iterations, i.e., $D_{\text{\text{KL}}}(\calN(0,\hat{\Sigma}_{T})||\calN(0,\Sigma_{T}^{(t+1)}))\leq\calN(0,\hat{\Sigma}_{T})||\calN(0,\Sigma_{T}^{(t)}))$.
Using an accelerated version of Algorithm \ref{algo:GaussEM}, the
complexity per iteration is $\calO$($kn^{2}+n^{2}\log n$).
\label{prop:LatenChowLiu}
\end{proposition}
Due to the page limit, we defer the description of the accelerated
version (\textit{the accelerated latent Chow-Liu algorithm}) and the
proof of Proposition \ref{prop:LatenChowLiu} to Appendix \ref{sec:Appen_Prop2}.
In fact, we never need to explicitly invert the empirical covariance
matrix $\hat{\Sigma}_{T}$ in the accelerated version.
As a rule of thumb, we often use the spanning tree obtained by the
standard Chow-Liu algorithm as an initial tree among the observed
nodes. But note that \textbf{P2} involves solving a combinatorial
problem exactly, so the algorithm is able to jump among different
graph structures which reduces the chance of getting stuck at a bad
local minimum and gives us much more flexibility in initializing graph
structures. In the experiments, we will demonstrate that Algorithm
\ref{algo:GaussEM} is not sensitive to the initial graph structure.\vspace{-0.15in}
\section{Experiments\vspace{-0.15in}
}
In this section, we present experimental results for learning GGMs
with small FVSs, observed or latent, using both synthetic data and
real data of flight delays. \vspace{-0.15in}
\paragraph*{Fractional Brownian Motion: Latent FVS}
We consider a fractional Brownian motion (fBM) with Hurst parameter
$H=0.2$ defined on the time interval $(0,1]$. The covariance function
is $\Sigma(t_{1},t_{2})=\frac{1}{2}(|t_{1}|^{2H}+|t_{2}|^{2H}-|t_{1}-t_{2}|^{2H})$.
Figure \ref{fig:FBMfull} shows the covariance matrices of approximate
models using spanning trees (learned by the Chow-Liu algorithm), latent
trees (learned by the CLRG and NJ algorithms in \citep{choi2011learning})
and our latent FVS model (learned by Algorithm \ref{algo:GaussEM})
using 64 time samples (nodes). We can see that in the spanning tree
the correlation decays quickly (in fact exponentially) with distance,
which models the fBM poorly. The latent trees that are learned exhibit
blocky artifacts and have little or no improvement over the spanning
tree measured in the K-L divergence. In Figure \ref{fig:FBMplot-1},
we plot the K-L divergence (between the true model and the learned
models using Algorithm \ref{algo:GaussEM}) versus the size of the
latent FVSs for models with 32, 64, 128, and 256 time samples respectively.
For these models, we need about 1, 3, 5, and 7 feedback nodes respectively
to reduce the K-L divergence to 25\% of that achieved by the best
spanning tree model. Hence, we speculate that empirically $k=\calO(\log n)$
is a proper choice of the size of the latent FVS. We also study the
sensitivity of Algorithm \ref{algo:GaussEM} to the initial graph
structure. In our experiments, for different initial structures, Algorithm
\ref{algo:GaussEM} converges to the same graph structures (that give
the K-L divergence as shown in Figure \ref{fig:FBMplot-1}) within
three iterations.
\begin{figure}
\includegraphics[width=1\columnwidth]{FBM_all_64nodes}\caption{From left to right: 1) The true model (fBM with 64 time samples);
2) The best spanning tree; 3) The latent tree learned using the CLRG
algorithm in \citep{choi2011learning}; 4) The latent tree learned
using the NJ algorithm in \citep{choi2011learning}; 5) The model
with a size-one latent FVS learned using Algorithm \ref{algo:GaussEM}.
The gray scale is normalized for visual clarity. }
\label{fig:FBMfull}
\end{figure}
\begin{figure}
\vspace{-0.3in}
\begin{minipage}[t]{1\columnwidth
\subfloat[32 nodes]{\includegraphics[width=0.25\columnwidth]{KL_SizeOfFVS_32points_new}}\hfill{}\subfloat[64 nodes]{\includegraphics[width=0.25\columnwidth]{KL_SizeOfFVS_64points_new}}\subfloat[128 nodes]{\includegraphics[width=0.25\columnwidth]{KL_SizeOfFVS_128points_new}}\subfloat[256 nodes]{\includegraphics[width=0.25\columnwidth]{KL_SizeOfFVS_256points_new}}\hfill{}
\caption{The relationship between the K-L divergence and the latent FVS size.
All models are learned using Algorithm \ref{algo:GaussEM} with 40
iterations.}
\label{fig:FBMplot-1}\vspace{-0.1in}
\end{minipage}
\vspace{-0.1in}
\end{figure}
\begin{figure}[t]
\subfloat[True Model]{\includegraphics[width=0.2\columnwidth]{Greedy3FVS}
}\subfloat[KL=12.7651]{\includegraphics[width=0.2\columnwidth]{Greedy0FVS}
}\subfloat[KL=1.3832]{\includegraphics[width=0.2\columnwidth]{Greedy1FVS}
}\subfloat[KL=0.6074]{\includegraphics[width=0.2\columnwidth]{Greedy2FVS}
}\subfloat[KL=0.0048]{\includegraphics[width=0.2\columnwidth]{Greedy3FVS}
\label{fig:3FVS}}
\caption{Learning a GGM using Algorithm \ref{algo:greedynew}. The thicker
blue lines represent the edges among the non-feedback nodes and the
thinner red lines represent other edges. (a) True model; (b) Tree-structured
model (0-FVS) learned from samples; (c) 1-FVS model; (d) 2-FVS model;
(e) 3-FVS model.}
\label{fig:greedy}
\end{figure}
\vspace{-0.15in}
\begin{figure}
\subfloat[Spanning Tree]{\includegraphics[width=0.23\columnwidth]{city200FVS0}
\label{fig:spanningtree}}\hfill{}\subfloat[1-FVS GGM]{\includegraphics[width=0.23\columnwidth]{city200FVS1}
\label{fig:1-fvstree}}\hfill{}\subfloat[3-FVS GGM]{\includegraphics[width=0.23\columnwidth]{city200FVS3}
\label{fig:3-fvstree}}\hfill{}\subfloat[10-FVS GGM]{\includegraphics[width=0.23\columnwidth]{city200FVS10}
\label{fig:10-fvstree}}
\caption{GGMs for modeling flight delays. The red dots denote selected feedback
nodes and the blue lines represent edges among the non-feedback nodes
(other edges involving the feedback nodes are omitted for clarity).}
\end{figure}
\vspace{-0.05in}
\paragraph*{Performance of the Greedy Algorithm: Observed FVS}
In this experiment, we examine the performance of the greedy algorithm
(Algorithm \ref{algo:greedynew}) when the FVS nodes are observed.
For each run, we construct a GGM that has 20 nodes and an FVS of size
three as the true model. We first generate a random spanning tree
among the non-feedback nodes. Then the corresponding information matrix
$J$ is also randomly generated: non-zero entries of $J$ are drawn
\textit{i.i.d. }from the uniform distribution $U[-1,1]$ with a multiple
of the identity matrix added to ensure $J\succ0$. From each generated
GGM, we draw 1000 samples and use Algorithm \ref{algo:greedynew}
to learn the model. For 100 runs that we have performed, we recover
the true graph structures successfully. Figure \ref{fig:greedy} shows
the graphs (and the K-L divergence) obtained using the greedy algorithm
for a typical run. We can see that we have the most divergence reduction
(from 12.7651 to 1.3832) when the first feedback node is selected.
When the size of the FVS increases to three (Figure \ref{fig:3FVS}),
the graph structure is recovered correctly.
\vspace{-0.2in}
\paragraph*{Flight Delay Model: Observed FVS}
In this experiment, we model the relationships among airports for
flight delays. The raw dataset comes from RITA of the Bureau of Transportation
Statistics. It contains flight information in the U.S. from 1987 to
2008 including information such as scheduled departure time, scheduled
arrival time, departure delay, arrival delay, cancellation, and reasons
for cancellation for all domestic flights in the U.S. We want to model
how the flight delays at different airports are related to each other
using GGMs. First, we compute the average departure delay for each
day and each airport (of the top 200 busiest airports) using data
from the year 2008. Note that the average departure delays does not
directly indicate whether an airport is one of the major airports
that has heavy traffic. It is interesting to see whether major airports
(especially those notorious for delays) correspond to feedback nodes
in the learned models. Figure \ref{fig:spanningtree} shows the best
tree-structured graph obtained by the Chow-Liu algorithms (with input
being the covariance matrix of the average delay). Figure \ref{fig:1-fvstree}--\ref{fig:10-fvstree}
show the GGMs learned using Algorithm \ref{algo:greedynew}. It is
interesting that the first node selected is Nashville (BNA), which
is not one of the top \textquotedblleft{}hubs\textquotedblright{}
of the air system. The reason is that much of the statistical relationships
related to those hubs are approximated well enough, when we consider
a 1-FVS approximation, by a spanning tree (excluding BNA) and it is
the breaking of the cycles involving BNA that provide the most reduction
in K-L divergence over a spanning tree. Starting with the next node
selected in our greedy algorithm, we begin to see hubs being chosen.
In particular, the first ten airports selected in order are: BNA,
Chicago, Atlanta, Oakland, Newark, Dallas, San Francisco, Seattle,
Washington DC, Salt Lake City. Several major airports on the coasts
(e.g., Los Angeles and JFK) are not selected, as their influence on
delays at other domestic airports is well-captured with a tree structure.\vspace{-0.15in}
\section{Future Directions\vspace{-0.15in}
}
Our experimental results demonstrate the potential of these algorithms,
and, as in the work \citep{liu2012feedback}, suggests that choosing
FVSs of size $\calO(\log n)$ works well, leading to algorithms which
can be scaled to large problems. Providing theoretical guarantees
for this scaling (e.g., by specifying classes of models for which
such a size FVS provides asymptotically accurate models) is thus a
compelling open problem. In addition, incorporating complexity into
the FVS-order problem (e.g., as in AIC or BIC) is another direction
we are pursuing. Moreover, we are also working towards extending our
results to the non-Gaussian settings. \vspace{-0.15in}
\section*{Acknowledgments\vspace{-0.15in}
}
This research was supported in part by AFOSR under Grant FA9550-12-1-0287.\vspace{-0.15in}
\bibliographystyle{IEEEtran}
|
\section*{Introduction}\label{S:one}
The lambda calculus was originally introduced by Church \cite{Church33,Church41} as a foundation for logic, where functions, instead of
sets, were primitive, and it turned out to be consistent and successful as a tool for formalising all computable functions.
The rise of computers gave a new development to its theoretical studies. The lambda calculus is the kernel of the functional programming
paradigm, because its ordinary parameter-binding mechanism corresponds closely to parameter binding in many functional programming
languages.
At the beginning researchers have investigated lambda calculus by using mainly syntactical methods
and have focused their interest on a limited number of equational extensions of lambda calculus, called \emph{$\lambda$-theories} (see
\cite{Bare}). Lambda theories are congruences on the set of $\lambda$-terms, which contain $\beta$-conversion. They arise by syntactical or
semantical considerations. Indeed a $\lambda$-theory may correspond to a possible operational semantics of the lambda calculus, as well as it
may be induced by a model of lambda calculus through the congruence relation associated with the interpretation function. The set of
$\lambda$-theories is naturally equipped with a structure of complete lattice, whose bottom element is the least $\lambda$-theory $\lambda\beta$, and
whose top element is the inconsistent $\lambda$-theory. The lattice of $\lambda$-theories is a very rich and complex structure of cardinality
$2^{\aleph_0}$ (see, for example, \cite{Bare,LusinS04,ManzonettoS10}). Syntactical techniques are usually difficult to apply in the
study of $\lambda$-theories. Therefore, semantic methods have been extensively investigated.
One of the most important contributions in the area of mathematical programming
semantics was the discovery by Scott \cite{Scott69} in the late 1960s, that complete partial orders, having their own function space as a retract, are models for the untyped lambda calculus.
It took some time, after Scott gave his model construction, for consensus to arise on the general notion of a model of the $\lambda$-calculus. There are mainly two descriptions that one can give: the category-theoretical and the algebraic one. Besides the different languages in which they are formulated, the two approaches are intimately connected (see \cite{Bare}). The categorical notion of model is well-suited for constructing concrete models, while the algebraic one is rather used to understand global properties of models (constructions of new models out of existing ones, closure properties, etc.) and to obtain results about the structure of the lattice of $\lambda$-theories.
The algebraic description of models of $\lambda$-calculus proposes two kinds of structures, viz. the $\lambda$-models based on the notion of combinatory algebra of Curry and Sch\"onfinkel (see \cite{CurryF58,Schonfinkel24}), and the $\lambda$-abstraction algebras of Pigozzi and Salibra (see \cite{PigozziS95,Salibra00,ManzonettoS10}).
Lambda abstraction algebras are intended as an alternative to combinatory algebras since they constitute an equational description of lambda calculus, which keeps the lambda notation and hence all the functional intuitions.
After the construction of the first model by Scott,
a large number of mathematical models for lambda calculus have been introduced in various categories of domains and
were classified into semantics according to the nature of their representable functions \cite{Bare,Berline00,Berline06,Plotkin93}. Scott
continuous semantics \cite{Scott80} is given in the category whose objects are complete partial orders and
morphisms are Scott continuous functions.
Other semantics of lambda calculus were isolated by Berry \cite{Berry78} and Bucciarelli-Ehrhard \cite{BucciarelliE91}:
Berry's stable semantics and Bucciarelli and Ehrhard's strongly stable semantics are refinements of
the continuous semantics introduced to capture the notion of `sequential' Scott continuous function.
All these semantics are structurally and equationally rich \cite{BerlineS06,Kerth95,Kerth01} in the sense that it is possible
to build up $2^{\aleph_0}$-models in each of them, inducing pairwise distinct $\lambda$-theories.
On the other hand, there are results that indicate that Scott's methods, based on a combination of order-theory and topology, may not in general be exhaustive.
The problem of the \emph{theory completeness} of a given semantics asks whether
every consistent $\lambda$-theory arises as the theory of a model in the semantics.
Honsell and Ronchi Della Rocca \cite{HonsellR92} have shown that there exists a $\lambda$-theory that does not arise as the theory of a Scott model.
Analogous results of incompleteness were obtained by Bastonero and Gouy for the stable semantics \cite{BastoneroG99} and by Salibra for all models of lambda calculus that involve monotonicity with respect to some pointed (i.e., with bottom element) partial order \cite{Salibra03}.
This last result removes the belief that pointed
partial orderings are intrinsic to models of the lambda calculus, and that the
incompleteness of a semantics is only due to the richness of the structure of representable functions.
Instead, it is also due to the richness of the structure of the $\lambda$-theories.
The need of more abstract semantics of lambda calculus arises when we recognize the inadequacy of Scott continuous semantics in order to investigate the structure of the lattice of $\lambda$-theories (see \cite[Chapter 4]{Bare} and \cite{Berline00,Berline06}) in itself and in connections with the theory of models.
Since topology refines partial orderings through separation axioms, Salibra \cite{Salibra03} has introduced other topological semantics of lambda calculus, and has shown: (i) the theory completeness of the semantics of lambda calculus given in terms of metrizable topological models; (ii) the theory incompleteness of the semantics of lambda calculus given in terms of topological models, which have no disjoint closures of nonempty open sets.
This last semantics is enough wide to include properly all pointed ordered models.
A natural problem of \emph{equational completeness} then arises for a semantics of $\lambda$-calculus: whether any two $\lambda$-terms
equal in all models of the semantics are $\beta$-convertible. Theory completeness implies equational completeness but the opposite direction is trivially false. The equational
completeness problem for Scott continuous semantics is one of most outstanding open problems of $\lambda$-calculus and it seems to have appeared first in the literature in \cite{HonsellR92}.
There is also an analogous \emph{equational consistency} problem, raised by Honsell and Plotkin in \cite{HonsellP09}: whether
for every finite set $E$ of equations between $\lambda$-terms, consistent with the $\lambda$-calculus, there exists a
Scott model contemporaneously satisfying all equations of $E$. In this paper we answer negatively to this second question in a very wide way. We provide two equations (called the \emph{subtractive equations}) consistent with $\lambda$-calculus, which cannot be contemporaneously satisfied by an arbitrary pointed ordered model of $\lambda$-calculus.
Although many familiar models are constructed by order-theoretic methods, it is also known that there are some models of the lambda calculus that cannot be non-trivially ordered (see \cite{Plotkin96,Salibra03,Selinger03}). In general, we define a combinatory algebra ${\bf A}$ to be \emph{unorderable} if there does not exist a non-trivial partial order on $A$ for which the application operation is monotone. Of course, the existence of unorderable models does not imply that order-theoretic methods are somehow incomplete for constructing models:
an unorderable model can still arise from an order-theoretic construction, for instance as
a subalgebra of some ordered model.
The most interesting result has been obtained by Selinger \cite{Selinger03}, who,
enough surprisingly, has shown that the standard open and closed term models of $\lambda\beta$ and $\lambda\beta\eta$ are unorderable.
As a consequence of this result, it follows that if $\lambda\beta$ or $\lambda\beta\eta$ is the theory of an ordered model, then the denotations of closed terms in that model are pairwise incomparable, i.e. the term denotations form an anti-chain.
This led Selinger \cite{Selinger03} to study the related question of absolute unorderability: a model is absolutely unorderable if it cannot be embedded in an orderable one. Plotkin conjectures in \cite{Plotkin96} that an absolutely unorderable combinatory algebra exists, but the question is still open whether this is so. Selinger has given in \cite{Selinger03} a syntactic
characterisation of the absolutely unorderable algebras in any algebraic variety (equational class) in
terms of the existence of a family of Mal'cev operators. Plotkin's conjecture is thus reduced to the question whether Mal'cev operators are consistent with the lambda calculus or combinatory logic.
The question of absolute unorderability can also be formulated in terms of theories, rather than models.
In this form, Selinger \cite{Selinger03} refers to it as the \emph{order-incompleteness} question: does there exist a $\lambda$-theory which does not arise as the theory of a non-trivially ordered model? Such a problem can be also characterised in terms of connected components of a partial ordering (minimal subsets which are both upward and downward closed): a $\lambda$-theory ${\mathcal T}$ is order-incomplete if, and only if, every ordered model, having ${\mathcal T}$ as equational theory, is partitioned in an infinite number of connected components, each one containing exactly one element. In other words, the partial order is the equality.
Toward an answer to the order-incompleteness problem, we define a consistent $\lambda$-theory ${\mathcal T}$ having
the following properties: (i) ${\mathcal T}$ includes the subtractive equations; (ii) if an ordered model ${\mathcal M}$ has an equational theory that contains ${\mathcal T}$, then ${\mathcal M}$ has an infinite number of connected components among
which that of the looping term $\Omega\equiv (\lambda x.xx)(\lambda x.xx)$ is a singleton set. Moreover, each connected component of ${\mathcal M}$ contains the denotation of at most one $\beta\eta$-normal form.
Compared to absolute unorderability, the above situation still has some missing bits. For example we are not in the position to tell where the denotations
of all unsolvable $\lambda$-terms other than $\Omega$ are placed in the model. Same thing for all the solvable $\lambda$-terms which do not have a $\beta\eta$-normal form.
The inspiration for the subtractive equations comes from the notion of \emph{subtractive variety} of algebras introduced by Ursini in \cite{Ursini}. A variety ${\mathcal V}$ of algebras is subtractive if it satisfies the following identities:
$$s(x,x) = 0;\qquad s(x,0) = x$$
for some binary term $s$ and constant $0$. Subtractive algebras abound in classical algebras. If we interpret the binary operator ``$s$'' as subtraction, and we use the infix
notation, then we can rewrite the above identities as $x-x= 0$ and $x-0 = x$.
In the context of $\lambda$-calculus, the subtractive equations make a certain $\lambda$-term behave like a binary subtraction operator (in curryfied form) whose ``zero'' is the looping $\lambda$-term $\Omega$.
We relativize to an element the notion of absolute unorderability. We say that an algebra ${\bf A}$ is $0$-unorderable if, for every compatible partial order on ${\bf A}$, $0$ is not comparable with any other element of the algebra.
An algebra ${\bf A}$ in a variety ${\mathcal V}$ is absolutely 0-unorderable
if, for any ${\bf B}\in{\mathcal V}$ and embedding $f : {\bf A}\to {\bf B}$, ${\bf B}$ is 0-unorderable.
Generalising subtractivity to $n$-subtractivity ($n\geq 2$),
we give a syntactic characterisation of the absolutely $0$-unorderable algebras with Mal'cev-type conditions. The consistency of the two subtractive equations with $\lambda$-calculus implies the existence of absolutely $\Omega$-unorderable combinatory algebras.
The last result of the paper is a theory incompleteness theorem for topological models of $\lambda$-calculus. This result is a strong
generalisation of an analogous theorem shown in \cite{Salibra03}. The incompleteness theorem is a consequence of a study of conditions
of separability for topological $n$-subtractive algebras.
This article is organised as follows: in Section \ref{prelim} we review the basic definitions of universal algebra, topology and $\lambda$-calculus
which are involved in the rest of the article. Section \ref{plotkin} is devoted to the proof of consistency of the subtractive equations. In Section \ref{sec:Hon-Plo-question}, we negatively answer to the open question of the equational consistency of Scott continuous semantics. The order incompleteness problem is presented in Section \ref{subs:strenghtening}. Section \ref{sec:subtractivity} is devoted to the study of the (absolute) $0$-unorderability in a variety. The main result of Section \ref{sec:subtractivity-topology} is the theory-incompleteness of the semantics of $\lambda$-calculus given in terms of topological models, which are not $T_{2_{1/2}}$-separated in $\Omega$.
\section{Preliminaries}\label{prelim}
\subsection{Partial Orderings}\label{subsec:poset}
Let $(A,\leq)$ be a partially ordered set (poset).
Two elements $a,b$ of $A$ are: (1) {\it comparable}
if either $a \leq b$ or $b\leq a$.
A set $B\subseteq A$ is an
{\it upward} ({\it downward}) closed set
if $b\in B$, $a\in A$ and $b\leq a$ ($a\leq b$) imply $a\in B$.
We denote by $\approx_\leq$ the least equivalence relation on $A$ containing $\leq$.
A \emph{connected component} of $(A,\leq)$ is an equivalence class of $\approx_\leq$.
A connected component can be also characterised as a minimal subset of $A$
which is both upward closed and downward closed.
The poset $(A,\leq)$ is called \emph{connected} if $\approx_\leq$ determines a unique equivalence class.
\subsection{Algebras}\label{algebras}
An \emph{algebraic similarity type} $\tau$ is constituted by a non-empty set of operator symbols
together with a function assigning to each operator $f \in\tau$ a finite \emph{arity}.
Operator symbols of arity 0 are called \emph{nullary operators} or \emph{constants}.
A $\tau$-algebra ${\bf A}$ is a tuple $(A, f^{\bf A})_{f\in \tau}$, where $A$ is a non-empty set and
$f^{\bf A} :A^n\to A$ is an $n$-ary function for every $f\in\tau$ of arity $n$.
Henceforth, the superscript in $f^{\bf A}$ will be dropped whenever the difference
between the operation and the operation symbol is clear from the context, and
a similar policy will be followed in similar cases throughout the paper.
Given two $\tau$-algebras ${\bf A}$ and ${\bf B}$, a {\em homomorphism} from ${\bf A}$ into ${\bf B}$ is a map $g:A\to B$ such that $g(f^{{\bf A}}(a_1,\ldots,a_n)) = f^{{\bf B}}(g(a_1),\ldots,g(a_n))$
for each $n$-ary operator $f\in\tau$ and for all $a_i\in A$.
Given a $\tau$-algebra ${\bf A}$, a binary relation $\phi$ on $A$ is \emph{compatible} if for all $f\in\tau$ of arity $n$,
and for all $a_i,b_i\in A$ we have
$$
a_1\phi b_1,\ldots, a_n \phi b_n \Longrightarrow f(a_1,\ldots,a_n)\phi f(b_1,\ldots,b_n).
$$
A compatible equivalence relation on an algebra ${\bf A}$ is called a \emph{congruence}.
Let ${\mathcal V}$ be a class of $\tau$-algebras and ${\bf A}$ be a $\tau$-algebra.
\begin{defi}
If $X\subseteq A$, then we say that ${\bf A}$ has the {\em universal mapping property for ${\mathcal V}$ over $X$} if, for every ${\bf B}\in {\mathcal V}$ and for every mapping
$g: X\to B$, there is a unique homomorphism $f: {\bf A}\to {\bf B}$ that extends $g$ (i.e., $f(x)=g(x)$ for every $x \in X$).
\end{defi}
\begin{defi}
${\bf A}$ is {\em free in ${\mathcal V}$ over $X$} if ${\bf A}\in {\mathcal V}$, ${\bf A}$ is generated by $X$ under the operations of $\tau$, and ${\bf A}$ has the universal mapping property for
${\mathcal V}$ over $X$.
\end{defi}
In the following we give a concrete characterization of the
free algebra in the class of all algebras of type $\tau$.
Let $X$ be a set of indeterminates. The set $T_\tau[X]$ of {\em $\tau$-terms} over $X$ is defined by induction as follows:
\begin{itemize}
\item $x\in T_\tau[X]$ for every $x\in X$;
\item $c\in T_\tau[X]$ for every constant $c\in\tau$;
\item if $t_1,\ldots, t_n\in T_\tau[X]$, then $f(t_1,\ldots,t_n)\in T_\tau[X]$ for all $f\in\tau$ of arity $n$.
\end{itemize}
A $\tau$-term is \emph{ground} if it does not contain variables.
If $t$ is a term, we write $t\equiv t(x_1,\ldots,x_n)$ if the variables occurring in $t$ are among $x_1,\ldots,x_n$.
If ${\bf A}$ is an algebra of type $\tau$ then every term $t(x_1,\ldots,x_n)$ induces a {\em term operation} $t^{\bf A}:A^n\to A$
defined in the obvious way.
The \emph{free algebra over $X$ in the class of all algebras of type $\tau$} is the algebra
$${{\bf T}_\tau}[X] = (T_\tau[X], f^{{\bf T}_\tau[X]})_{f\in \tau},$$
where
$$f^{{\bf T}_\tau[X]}(t_1,\dots,t_n) = f(t_1,\dots,t_n),\quad \text{for all $f\in \tau$.}$$
An \emph{identity} (or equation) of type $\tau$ is a pair $(t,u)$ of $\tau$-terms, written also $t = u$. A \emph{ground identity} is an equation between ground terms.
An identity $t(x_1,\ldots,x_n) = u(x_1,\ldots,x_n)$ holds in a $\tau$-algebra ${\bf A}$, and we write ${\bf A}\models t = u$, if the $n$-ary term operations $t^{\bf A}$ and $u^{\bf A}$ are equal.
A non-empty class ${\mathcal V}$ of algebras of the same type is:
\begin{enumerate}[label=(\roman*)]
\item a \emph{variety} if it is closed under subalgebras, homomorphic images and direct products;
\item an \emph{equational class} if it is axiomatisable by a set of equations.
\end{enumerate}
\begin{thm} {\rm (Birkhoff \cite{Birkhoff35})} A class of algebras is a variety if, and only if, it is an equational class.
\end{thm}
If ${\mathcal V}$ is a variety of type $\tau$, then
$Eq({\mathcal V})$ is the set of equations satisfied by all algebras in ${\mathcal V}$.
The set $Eq({\mathcal V})$ is a congruence over the free algebra ${\bf T_\tau}[X]$, and the quotient of ${\bf T}_\tau[X]$ by $Eq({\mathcal V})$ is the free algebra ${\bf T}_{\mathcal V}[X]$ in ${\mathcal V}$ over $X$.
\subsection{Topology}\label{top}
By a topological space we shall mean a pair $(X,\tau)$, where $X$ is
a non-empty set and $\tau$ is
a family of subsets of $X$
closed under infinite unions, finite intersections, and including $\emptyset$ and
$X$. Given a point $x$ of a space $X$, we say that $U\subseteq X$ is a neighbourhood of $x$ if there exists an open set $V$ such that
$x\in V\subseteq U$.
A subset $F$ of a space $X$ is closed if $X\setminus F$ is open. The closure of $U\subseteq X$ will be denoted by $\oU$ (as a matter of notation, we write $\ob$ for
$\overline{ \{ b\} }$).
Recall that $a\in \oU$ if $U\cap V \neq \emptyset$ for every open
neighbourhood $V$ of $a$.
A space $X$ is
\begin{itemize}
\item $T_0$ if, for all distinct $a,b\in X$, $a$ has a neighbourhood that does not contain $b$ {\bf or} vice versa.
\item $T_1$ if, for all distinct $a,b\in X$, $a$ has a neighbourhood that does not contain $b$ {\bf and} vice versa.
\item $T_2$ (or Hausdorff) if for all distinct $a,b\in X$ there exist open sets
$U$ and $V$ with $a\in U$, $b\in V$ and $U\cap V = \emptyset$.
\item $T_{2_{1/2}}$ (or completely Hausdorff) if for all distinct $a,b\in X$ there exist open sets
$U$ and $V$ with $a\in U$, $b\in V$ and $\oU\cap \oV = \emptyset$.
\end{itemize}
\noindent The previous axioms of separation can be relativized to
pairs of elements. For example, $a$ and $b$ are
$T_{2_{1/2}}$-separable, if there exist open sets $U$ and $V$ with
$a\in U$, $b\in V$ and $\oU\cap \oV = \emptyset$. $T_2$-, $T_1$-,
$T_0$-separability are similarly defined.
A preorder can be defined on $(X,\tau)$ by
$$a\leq_\tau b\ \text{iff}\ a\in \ob\ \text{iff}\
\forall U\in \tau (a\in U \Rightarrow b\in U).$$
We have
$$\mbox{$(X,\tau)$ is $T_0$ iff $\leq_\tau$ is a partial order.}$$
For any $T_0$-space $(X,\tau)$ the partial order $\leq_\tau$ is called
the \emph{specialization order} of $(X,\tau)$. Note that any continuous map between $T_0$-spaces is
necessarily monotone
and that the order is discrete (i.e. satisfies $a\leq_\tau b$ iff $a = b$)
iff $X$ is a $T_1$-space.
Suppose a space $X$ is not $T_0$. We can obtain a $T_0$-space out of $X$ by the following
well-known construction. Define an equivalence relation $\equiv$ on $X$ as follows: $a\equiv b\ \Leftrightarrow\ a \leq_\tau b\ \text{and}\ b\leq_\tau a$.
Then $X/\equiv$ equipped with the quotient topology is a $T_0$-space.
A space $(X,\tau)$ is \emph{coconnected} if for all non-empty opens $U, V\in \tau$ we have $\oU\cap\oV\neq\emptyset$ (see \cite{Salibra03}).
\subsection{Topological algebras}
A {\it topological algebra} is a pair $({\bf A},\tau)$
where $\sbA$ is an algebra and $\tau$ is a topology on the underlying set $A$
with the property that each basic operation $f^{\bf A}: A^n\to A$ of ${\bf A}$ is continuous,
where the domain of $f$ is endowed with the product topology.
A {\it semitopological algebra} is a pair $({\bf A},\tau)$
where $\sbA$ is an algebra and $\tau$ is a topology on the underlying set $A$
with the property that each unary polynomial of $\sbA$ is
continuous w.r.t. $\tau$.
Every topological algebra is a semitopological algebra.
A map $f: A^n \to A$ is \emph{separated-continuous} if, for every $1 \leq i \leq n$
and every $a_1,\dots,a_{i-1},a_{i+1},\dots,a_n \in A$,
the map $g:A\to A$, defined by
$$g(b) = f(a_1,\dots,a_{i-1},b,a_{i+1},\dots,a_n),\quad\mbox{for all $b\in A$},$$
is continuous.
In a semitopological algebra every term operation is separated-continuous.
\begin{lem} \label{1.3.1bis}
Let $(A,\tau)$ be a topological space. Then every (separated-)continuous map $f: A^n \to A$ is monotone:
$$a_i\leq_\tau b_i\ (1\leq i\leq n)\ \Rightarrow\ f(a_1,\dots,a_n) \leq_\tau f(b_1,\dots,b_n).$$
\end{lem}
\proof
For the sake of simplicity, assume $f$ is binary.
Let $a \leq_\tau b$ and $a' \leq_\tau b'$. We have to show that $f(a,a') \leq_\tau f(b,b')$.
Let $f(a,a')\in U$, where $U$ is open. Then by continuity in first coordinate there exists
an open set $V$ such that $a\in V$ and $f(V,a') \subseteq U$. By $a\leq_\tau b$ it follows that
$b\in V$ so that $f(b,a')\in U$. By using continuity in second coordinate we get that there
exists an open set $W$ such that $a'\in W$ and $f(b,W)\subseteq U$. By $a'\leq_\tau b'$ we get $b'\in W$
and then the conclusion.
\qed
If $(\sbA,\tau)$ is a semitopological algebra then $\leq_\tau$ is a compatible preorder.
\subsection{Lambda calculus}\label{sus:lam-calc}
With regard to the $\lambda$-calculus we follow the notation and terminology of \cite{Bare}. By $\Lambda$ and $\Lambda^o$, respectively, we indicate
the set of $\lambda$-terms and of closed $\lambda$-terms. By convention application associates to the left.
The symbol $\equiv$ denotes syntactical equality. The following are some notable $\lambda$-terms that will be used throughout the paper: $\Omega \equiv (\lambda x.xx)(\lambda x.xx)$; $\ssi \equiv \lambda x.x$; $\ssT \equiv \lambda xy.x$; $\ssF \equiv \lambda xy.y$.
If $M$ is a $\lambda$-term and $\vec{P}\equiv P_1\dots P_n$ is a sequence of $\lambda$-terms, we write $M\vec P$ for the application $MP_1\cdots P_n$.
The $\beta$-reduction will be denoted by $\labelra{\beta}$, while the $\eta$-reduction by $\labelra{\eta}$.
One step of either $\beta$-reduction or $\eta$-reduction will be denoted by $\labelra{\beta\eta}$.
A $\lambda$-term $M$ is \emph{solvable} if it has a \emph{head normal form}, i.e., $M$ is $\beta$-convertible to a term of the
form $\lambda \vec{x}.y\vec{N}$. A $\lambda$-term $M$ is \emph{unsolvable} if it is not solvable.
A {\em $\lambda$-theory} is a congruence on $\Lambda$ (with respect to the operators of abstraction and application) which contains $\alpha\beta$-conversion.
We denote by $\lambda\beta$ the least $\lambda$-theory. The least extensional $\lambda$-theory $\lambda\beta\eta$ is axiomatised over $\lambda\beta$ by the
equation $\lambda x. Mx = M$, where $M\in \Lambda$ and $x$ is not free in $M$.
The $\lambda$-theory generated by a set $X$ of identities is the intersection of all $\lambda$-theories containing $X$.
For a $\lambda$-theory ${\mathcal T}$ we will write ${\mathcal T}\vdash M=N$ (or $M=_{\mathcal T} N$) to indicate the existence of an equational proof of the identity
$M=N$ that uses equations of ${\mathcal T}$.
A $\lambda$-theory is {\em consistent} if it does not equate all $\lambda$-terms, {\em inconsistent} otherwise. It turns out that a $\lambda$-theory
${\mathcal T}$ is inconsistent iff ${\mathcal T}\vdash \ssT = \ssF$. A $\lambda$-theory is \emph{semisensible} if it does not equate solvable and unsolvable
$\lambda$-terms. Semisensible $\lambda$-theories are consistent (see \cite{Bare}). The set of $\lambda$-theories constitutes a complete lattice
w.r.t. inclusion, whose top is the inconsistent $\lambda$-theory and whose bottom is the theory $\lambda\beta$.
An unsolvable is called a \emph{zero term} if it never reduces to an abstraction.
A zero term, that will be used in the rest of the paper, is defined as
follows. Let $B \equiv \lambda x.x(\lambda y.yx)$, $C \equiv \lambda z.zB$ and $\Theta \equiv BC$. By a direct computation we see that the only
possible reduction path starting with $\Theta$ is the following:
$$\Theta \labelra{\beta} C(\lambda y.yC) \labelra{\beta} (\lambda y.yC)B \labelra{\beta} BC \equiv \Theta.$$
The letters $\xi_1,\xi_2,\dots$ denote algebraic variables (holes in Barendregt's terminology \cite{Bare}). Contexts are built up as $\lambda$-terms but also allowing occurrences of algebraic variables. Substitution for algebraic variables is made without $\alpha$-conversion. For example,
$(\lambda x. x\xi)[xy/\xi] = \lambda x.x(xy)$.
We recall the Genericity Lemma of lambda calculus (see \cite[Proposition 14.3.24]{Bare}).
\begin{lem}\label{gen}
Let $M\in \Lambda$ unsolvable and $N\in \Lambda$ having a normal form. Then, for every context $C[\xi]$,
$$C[M] =_{\lambda\beta} N\ \Longrightarrow\ (\forall Q\in \Lambda)\ C[Q] =_{\lambda\beta} N.$$
\end{lem}
Throughout the paper we consider different reductions. If $\labelra{\gamma}$ is a reduction, then we denote by $\mslabelra{\gamma}$ the
reflexive transitive closure of $\labelra{\gamma}$, and we write $=_\gamma$ to denote the reflexive, symmetric and transitive closure of
$\labelra{\gamma}$. Finally we define, as usual, the \emph{$\gamma$-reduction graph} of a term $M$ as the set
${\mathcal G}_\gamma(M) = \{N \in \Lambda : M \mslabelra{\gamma} N\}$.
\subsection{Models of $\lambda$-calculus}
The algebraic description of models of $\lambda$-calculus proposes two kinds of structures, viz. the \emph{$\lambda$-algebras} and the
\emph{$\lambda$-models}, both based on the notion of \emph{combinatory algebra}. We will focus on $\lambda$-models.
A \emph{combinatory algebra} ${\bf A} = (A,\cdot,K, S)$ is a structure with a binary operation called \emph{application} and two distinguished
elements $K$ and $S$ called \emph{basic combinators}. The symbol ``$\cdot$'' is usually omitted from expressions and by convention
application associates to the left, allowing to leave out superfluous parentheses. The class of combinatory algebras is axiomatised by
the equations $K xy = x$ and $S xyz = xz(yz)$. Intuitively, elements on the left-hand side of an application are to be seen as functions
operating on arguments, placed on the right-hand side. Hence it is natural to say that a function $f:A^n \to A$ is \emph{representable}
(\emph{in} ${\bf A}$) if there exists an element $a \in A$ such that $f(b_1,\dots,b_n) = ab_1\dots b_n$ for all $b_1,\dots, b_n \in A$. For example the identity function is
represented by the combinator $I\equiv SKK$ and the projection on the first argument by the combinator $K$.
The axioms of an elementary subclass of combinatory algebras, called \emph{$\lambda$-models}, were expressly chosen to make coherent the definition of interpretation of $\lambda$-terms. In addition to the axioms of combinatory algebra, we have:
\[
\begin{array}{lll}
& \forall xy.(\forall z.\ xz = yz) \Rightarrow \sso x = \sso y & \\
& \sso_2K = K & \\
& \sso_3S = S &
\end{array}
\]
where $\sso_1 \equiv \sso \equiv S(KI)$ and $\sso_{n+1} \equiv S(K\sso)(S(K\sso_n))$.
The combinators $\sso_n$ are made into inner choice operators. Indeed, given any $a \in A$, the element $\sso_n a$ represents the same $n$-ary
function as $a$ and $\sso_n c = \sso_n d$ for all $c,d$ representing the same $n$-ary function.
Let $Env_A$ be the set of $A$-environments, i.e.\ , the functions from the set $\mathrm{Var}$ of $\lambda$-calculus variables to $A$.
For every $x\in \mathrm{Var}$ and $a\in A$ we denote by $\rho[x:=a]$ the environment $\rho'$ which coincides with $\rho$ everywhere except on $x$,
where $\rho'$ takes the value $a$.
When ${\bf A}$ is a $\lambda$-model it is possible to define the following interpretation:
\[
\begin{array}{ll}
{|x|}_\rho^{\bf A} = \rho(x); &\\
{|MN|}_\rho^{\bf A} = {|M|}_\rho^{\bf A} {|N|}_\rho^{\bf A}; &\\
{|\lambda x.M|}_\rho^{\bf A} = \sso a,\ \text{where $a\in A$ is any element representing the function $b\in A \mapsto {|M|}_{\rho[x:=b]}^{\bf A}$.}&
\end{array}
\]
Note that ${|\lambda x.M|}_\rho^{\bf A}$ is well-defined, since each function $b\in A \mapsto {|M|}_{\rho[x:=b]}^{\bf A}$ is representable under the hypothesis that ${\bf A}$ is a $\lambda$-model. This is the kind of interpretation we will refer to.
By the way when $M$ is a closed $\lambda$-term and there is no worry of confusion about the model being considered, we write $|M|$ for ${|M|}_\rho^{\bf A}$.
Each $\lambda$-model ${\bf A}$ induces a $\lambda$-theory, denoted here by $Th({\bf A})$, and called {\em the equational theory of ${\bf A}$}.
Thus, $M = N \in Th({\bf A})$ if, and only if, $M$ and $N$ have the same interpretation in ${\bf A}$.
An \emph{ordered $\lambda$-model} is a pair $({\bf A},\leq)$, where ${\bf A}$ is a $\lambda$-model and $\leq$ is a partial order on $A$ which makes the application operator of ${\bf A}$ monotone in both arguments.
An ordered model $({\bf A},\leq)$ is \emph{non-trivial} if the partial order
is not discrete, i.e., $a<b$ for some $a,b\in A$ (thus $A$ is not a singleton).
A \emph{pointed ordered model} is an ordered model with bottom element.
The {\it term model} ${\mathcal M}_{\mathcal T}$ of a lambda theory ${\mathcal T}$ (see \cite[Def.~5.2.11]{Bare})
consists of the set of the equivalence classes of $\lambda$-terms modulo the lambda theory ${\mathcal T}$
together with the operation of application on the equivalence classes.
By Cor.~5.2.13(ii) in \cite{Bare} ${\mathcal M}_{\mathcal T}$ is a $\lambda$-model such that $Th({\mathcal M}_{\mathcal T})= {\mathcal T}$.
A class ${\mathcal C}$ of $\lambda$-models is
\begin{enumerate
\item \emph{equationally consistent} if, for every finite set $E$ of equations between $\lambda$-terms, consistent with the $\lambda$-calculus, there exists a
model ${\bf A}\in {\mathcal C}$ simultaneously satisfying all equations of $E$.
\item \emph{equationally complete} if any two $\lambda$-terms
equal in all models of ${\mathcal C}$ are $\beta$-convertible.
\item \emph{theory complete} if, for every consistent $\lambda$-theory ${\mathcal T}$, there exists a model ${\bf A}\in {\mathcal C}$ such that $Th({\bf A}) = {\mathcal T}$.
\end{enumerate}
\subsection{The Jacopini\textendash Kuper technique}\label{subs:Jacopini-Kuper}
The Jacopini\textendash Kuper technique, introduced by Jacopini in \cite{Ja} and generalized by Kuper in \cite{Kuper97}, can be used to tackle questions of consistency of equational extensions of lambda calculus.
In this section we review this technique.
Let ${\mathcal T}$ be an arbitrary consistent $\lambda$-theory, $\vec P = \vec Q$ be a set of identities $P_i = Q_i$ ($i=1,\dots, n$) between closed $\lambda$-terms, and ${\mathcal T}'$ be the $\lambda$-theory generated by ${\mathcal T}\cup (\vec P = \vec Q)$.
The idea is to reduce inconsistency of ${\mathcal T}'$ to that of ${\mathcal T}$. If ${\mathcal T}'$ is inconsistent, then there exists a finite equational proof of $\ssT =_{{\mathcal T}'} \ssF$,
and such a proof contains a finite number of applications of equations which are in $\vec P = \vec Q$.
Jacopini\textendash Kuper technique, when applicable, consists in checking two conditions on the sequences $\vec P$ and $\vec Q$, namely
that $\vec P$ is ${\mathcal T}$-\emph{operationally less defined} than $\vec Q$ (see Definition \ref{def:op-def}) and
that $\vec P$ is ${\mathcal T}$-\emph{proof-substitutable} by $\vec Q$ (see Definition \ref{def:proof-subst}).
Under these two conditions, it is possible to remove from the proof of $\ssT =_{{\mathcal T}'} \ssF$ all occurrences of
equations in $\vec P = \vec Q$, thus yielding a proof of $\ssT =_{{\mathcal T}} \ssF$. This is the end of the method, because ${\mathcal T}$ is supposed to be a consistent $\lambda$-theory.
A very useful property for the application of Jacopini-Kuper method, and in particular for proving
${\mathcal T}$-proof-substitutability, is the existence of a Church-Rosser reduction,
whose induced conversion coincides with the equality induced by ${\mathcal T}$ on $\lambda$-terms.
This is not evident from the abstract formulation given in this section, but will be clear in the next one, when we will apply the technique.
\begin{lem}\label{lem:standard-eq-proof}
We have that
${\mathcal T}\cup (\vec P = \vec Q) \vdash M=N$
if, and only if, there exist closed terms $F_1,\dots, F_n$ such that
\[
\begin{array}{llll}
M & =_{{\mathcal T}} & F_1\vec P\vec Q& \\
F_{j}\vec Q\vec P & =_{{\mathcal T}} & F_{j+1}\vec P\vec Q &\text{for $1\leq j\leq n-1$}\\
F_n\vec Q\vec P & =_{{\mathcal T}} & N.&
\end{array}
\]
\end{lem}
\proof By \cite[Theorem 1]{Du}
there exist binary contexts $C_1(\xi_1,\xi_2), \dots, C_n(\xi_1,\xi_2)$ and identities $P_{i_j} = Q_{i_j}$ in $\vec P = \vec Q$ such that
\[
\begin{array}{llll}
M & =_{\mathcal T} & C_1(P_{i_1},Q_{i_1}) & \\
C_j(Q_{i_j}, P_{i_j} ) & =_{\mathcal T} &C_{j+1}(P_{i_{j+1}},Q_{i_{j+1}}) &\text{for $1\leq j\leq n-1$} \\
C_n(Q_{i_n}, P_{i_n}) & =_{\mathcal T} & N. &
\end{array}
\]
It is sufficient to define $F_j \equiv \lambda \vec x\vec y. C_j(x_{i_j},y_{i_j})$,
where $\vec x$ and $\vec y$ are sequences of length $k$ of fresh variables.
\qed
\begin{defi}[Operational definiteness]\label{def:op-def}
We say that $\vec P$ is ${\mathcal T}$-operationally less defined than $\vec Q$
if, for every $\beta\eta$-normal form $N$ and every term $F$, we have that
$$ F\vec P=_{\mathcal T} N \Rightarrow F\vec Q =_{\mathcal T} N. $$
\end{defi}
\begin{defi}[Proof-substitutability]\label{def:proof-subst} We say that $\vec P$ is ${\mathcal T}$-proof-substitutable by $\vec Q$ if
$$\forall F,F'\in\Lambda^o(F \vec P =_{{\mathcal T}} F'\vec P \Rightarrow
\exists G \in\Lambda^o(G \vec P\vec Q =_{{\mathcal T}} F \vec Q \mbox{ and }
G \vec Q\vec P =_{{\mathcal T}} F'\vec Q)).$$
\end{defi}
\begin{thm}\label{thm:reduce-incons}
If $\vec P$ is ${\mathcal T}$-operationally less defined than $\vec Q$ and $\vec P$ is ${\mathcal T}$-proof-substitutable by $\vec Q$, then the $\lambda$-theory ${\mathcal T}'$ generated by ${\mathcal T}\cup (\vec P = \vec Q)$ is consistent.
\end{thm}
\proof
Assume ${\mathcal T}'$ is inconsistent, so that $\ssT =_{{\mathcal T}'} \ssF$.
Then by Lemma \ref{lem:standard-eq-proof} there exists an equational proof of this identity of the form
\[
\begin{array}{llll}
\ssT & =_{{\mathcal T}} & F_1\vec P\vec Q& \\
F_{j}\vec Q\vec P & =_{{\mathcal T}} & F_{j+1}\vec P\vec Q& \text{for $1\leq j\leq n-1$}\\
F_n\vec Q\vec P & =_{{\mathcal T}} & \ssF.&
\end{array}
\]
Now we show how to iteratively transform the above proof of ${\mathcal T}'\vdash \ssT =\ssF$ in a proof of ${\mathcal T}\vdash \ssT=\ssF$.
Suppose $n = 1$, i.e., we have
$\ssT =_{{\mathcal T}} F_1\vec P\vec Q$ and
$F_{1}\vec Q\vec P =_{{\mathcal T}} \ssF$.
Since $\vec P$ is ${\mathcal T}$-operationally less defined than $\vec Q$, from
$\ssT =_{{\mathcal T}} F_1\vec P\vec Q =_{{\mathcal T}} (\lambda \oy. F_1\oy \vec Q) \vec P $ and $F_{1}\vec Q\vec P =_{{\mathcal T}} \ssF$, we get
$\ssT =_{{\mathcal T}} F_1\vec Q\vec Q =_{{\mathcal T}} \ssF$.
Suppose $n > 1$. As before, by the hypothesis we get $\ssT =_{{\mathcal T}} F_1\vec Q\vec Q$
and $F_n\vec Q\vec Q =_{{\mathcal T}} \ssF$.
Let $\vec y$ be a sequence of fresh variables. For each $j = 1,\ldots,n-1$, define
$$H_{j} \equiv \lambda \vec y.F_{j}\vec Q\vec y;\qquad H_{j+1}' \equiv \lambda \vec y.F_{j+1}\vec y\vec Q.$$
By $F_{j}\vec Q\vec P =_{{\mathcal T}} F_{j+1}\vec P\vec Q$ ($j = 1,\ldots,n-1$) we have that
$$H_{j}\vec P =_{{\mathcal T}} H_{j+1}'\vec P.$$
Since $\vec P$ is ${\mathcal T}$-proof-substitutable by $\vec Q$, then
there exist terms $G_{j}$ ($j = 1,\ldots,n-1$) such that
$G_{j}\vec P\vec Q =_{{\mathcal T}} H_{j}\vec Q =_{{\mathcal T}} F_{j}\vec Q\vec Q$ and
$G_{j}\vec Q\vec P =_{{\mathcal T}} H_{j+1}'\vec Q =_{{\mathcal T}} F_{j+1}\vec Q\vec Q$.
Therefore we obtain that
\[
\begin{array}{rclcl}
\ssT & =_{{\mathcal T}} & F_1\vec Q\vec Q & =_{{\mathcal T}} & G_1\vec P\vec Q \\
G_{1}\vec Q\vec P & =_{{\mathcal T}} & F_{2}\vec Q\vec Q & =_{{\mathcal T}} & G_2\vec P\vec Q \\
& \vdots & & \vdots & \\
G_{n-1}\vec Q\vec P & =_{{\mathcal T}} & F_{n}\vec Q\vec Q & =_{{\mathcal T}} & \ssF \\
\end{array}
\]
Therefore one can iterate the argument and get a proof of ${\mathcal T}\vdash \ssT=\ssF$.
\qed
\section{On a question by Honsell and Plotkin}\label{plotkin}
In this section we turn to a question posed in \cite{HonsellP09} by Honsell and Plotkin. The problem is whether or not there exists
a formula $\varphi$ of first-order logic written as a possibly empty list of universal quantifiers followed by a conjunction of
equalities between $\lambda$-terms such that $\varphi$ does not admit pointed ordered models. According to Honsell and Plotkin, this problem falls
under the name of $\Pi_1$-consistency of the class of pointed ordered models.
Since any equation between $\lambda$-terms is equivalent to a suitable equation between closed $\lambda$-terms,
then in the context of $\lambda$-calculus the $\Pi_1$-consistency is equivalent to the equational consistency,
which is the particular case of $\Pi_1$-consistency in which the formula $\varphi$ is quantifier-free.
\subsection{The $\lambda$-theory $\lambda\pi\phi$}\label{sec:theory-lpp}
We introduce two equations between $\lambda$-terms, whose models will be shown to have strong properties with respect to the possible partial orderings they can be endowed with. Of course we have to prove that the $\lambda$-theory $\lambda\pi\phi$ generated by these equations is consistent
and this will be done in Section \ref{subs:Jac-for-lpp}.
The two equations we are going to introduce represent within $\lambda$-calculus the notion of subtractivity, which has been introduced in Universal Algebra
by Ursini \cite{Ursini}.
\begin{defi}\label{sub} An algebra ${\bf A}$ is \emph{subtractive} if there exist a binary term
$s(x,y)$ and a constant $0$ in the algebraic similarity type of ${\bf A}$ such that
$$s(x,x) = 0;\qquad s(x,0)=x.$$
\end{defi}
Subtractive algebras abound in classical
algebras and in algebraic logic since the term $s$ simulates part of subtraction. If
we interpret the binary operator ``$s$'' as subtraction ``$-$'' and we use the infix
notation, then we can rewrite the above identities as $x-x= 0$ and $x-0 = x$.
The following lemma explains why we are interested in subtractive equations.
\begin{lem}\label{lem-ord} Let ${\bf A}$ be a subtractive algebra and $\leq$ be a compatible partial order on $A$. Then, every element $a\neq 0$ of $A$ is incomparable with $0$.
\end{lem}
\proof If $a\leq 0$ then $0= s(a,a) \leq s(a,0)= a$.
If $0\leq a$ then $a = s(a,0) \leq s(a,a) = 0$.
\qed
A general treatment of subtractivity and orderings in Universal Algebra can be found in Section \ref{sec:subtractivity}.
We now define the $\lambda$-theory $\lambda\pi\phi$.
Let $\Theta$ be the term defined in Section \ref{sus:lam-calc}. We define $s(x,y) \equiv \Theta xy$ and $0\equiv \Omega$.
Then the $\lambda$-theory $\lambda\pi\phi$ is defined as the least extensional $\lambda$-theory generated by the following two equations, called the \emph{subtractive equations}:
$$(\pi)\ \ \Theta xx = \Omega;\quad\qquad (\phi)\ \ \Theta x \Omega = x.$$
The intuitive meaning of the equations ($\pi$) and ($\phi$) is that they make the term $\Theta$ behave like a
binary subtraction operator (in curried form) whose ``zero'' is the term $\Omega$.
We have chosen $\Theta$ to represent the binary subtraction operator because the reduction graph of $\Theta$ is as simple as possible among the unsolvables distinct from $\Omega$.
The following theorem illustrates a curious aspect of the equations ($\pi$) and ($\phi$): the choice of $\Omega$ is the right one.
\begin{thm}\label{thm:must-be-Omega} Let $O$ be a $\lambda$-term such that $x\notin FV(O)$, and let ${\mathcal T}$ be any $\lambda$-theory including the identities $\Theta x O = x$ and $\Theta xx = O$. Then ${\mathcal T} \vdash O =Y\ssi$ for every fixpoint combinator $Y$. In particular, if $Y\equiv \lambda f.(\lambda x.f(xx))(\lambda x.f(xx))$ is the Curry fixpoint combinator, then ${\mathcal T} \vdash O = \Omega$.
\end{thm}
\proof
We apply a technique introduced by Gordon Plotkin and Alex Simpson (see \cite{Selinger03}).
Let $Y$ be an arbitrary fixpoint combinator.
Then, for any $\lambda$-term $M$, define $\mu x.M \equiv Y(\lambda x.M)$.
Now let $D \equiv \mu y.\mu x. \Theta xy$. Then we have $D =_\beta \Theta DD =_{\mathcal T} O$ and
$D=_\beta \mu x. \Theta xD =_{\mathcal T}\mu x. \Theta xO =_{\mathcal T} \mu x.x \equiv Y\ssi$. Moreover, $Y\ssi =_\beta \Omega$ if $Y$ is the Curry fixpoint combinator.
\qed
\subsection{The $\lambda$-theory $\lambda\pi$}\label{subs:lp}
The extensional $\lambda$-theory $\lambda\pi$ is axiomatised over $\lambda\beta\eta$ by the equation $(\pi)$. It is consistent because semisensible.
We will show the consistency of $\lambda\pi\phi$ relying on the consistency of $\lambda\pi$.
The following notion of reduction will be useful in the next section to prove the consistency of the $\lambda$-theory $\lambda\pi\phi$
(recall from Section \ref{sus:lam-calc} the definition of $\Theta$ and of its reduction graph ${\mathcal G}_\beta(\Theta)$).
\begin{defi}[$\lambda\pi$-reduction]
We introduce here $\beta\eta\pi$-reduction, notation $\labelra{\beta\eta\pi}$, as the contextual closure of $\labelra{\beta\eta} \cup \labelra{\pi}$, where
$$\Psi M N \labelra{\pi} \Omega\quad\text{if $\Psi\in {\mathcal G}_\beta(\Theta)$ and $\lambda\pi \vdash M = N$}.$$
\end{defi}
Of course the conversion $=_{\beta\eta\pi}$ coincides with the equality induced by ${\lambda\pi}$.
We remark that in \cite{Salibra03} it was introduced a $\lambda$-theory axiomatised by $\Omega xx = \Omega$.
Here we use instead $\Theta xx = \Omega$ for technical reasons. In fact when we apply the Jacopini\textendash Kuper technique in
Section \ref{subs:Jac-for-lpp}, we use the fact that a step of reduction $\Theta x x \labelra{\pi} \Omega$ does not create a new
$\labelra{\pi}$-redex, because $\Omega \not\in {\mathcal G}_\beta(\Theta)$.
Note that if a $\lambda$-term $N$ is a $\beta\eta$-normal form, then it is also a $\beta\eta\pi$-normal form, because every
$\labelra{\pi}$-redex contains also a $\labelra{\beta}$-redex.
\begin{thm}\label{thm:betapil-confluent}\hfill
\begin{enumerate}[label=(\roman*)]
\item The reduction $\labelra{\beta\eta\pi}$ is Church-Rosser;
\item For all terms $M$ and $N$, we have $\Theta MN =_{\beta\eta\pi} \Omega$ iff $\lambda\pi \vdash M = N$.
\end{enumerate}
\end{thm}
\proof\hfill
\begin{enumerate}[label=(\roman*)]
\item First observe that the relation $\mslabelra{\pi}$ is Church-Rosser.
Moreover the relations $\mslabelra{\pi}$ and $\mslabelra{\beta\eta}$ commute, meaning that
whenever $Q \mslabella{\pi} P \mslabelra{\beta\eta} Z$ there exists $P'$ such that $Q \mslabelra{\beta\eta} P' \mslabella{\pi} Z$.
The conclusion follows from the Hindley\textendash Rosen Lemma (see \cite[Prop.~3.3.5]{Bare}), which states that
if two Church-Rosser relations commute, then their union is Church-Rosser too.
\item If $\Theta MN =_{\beta\eta\pi} \Omega$ then by (i) there exists a reduction $\Theta MN \mslabelra{\beta\eta\pi} \Omega$.
But this is possible only if $\Theta MN \mslabelra{\beta\eta\pi} \Psi M'N'$ with $\Theta \mslabelra{\beta} \Psi$,
$M \mslabelra{\beta\eta\pi} M'$, $N \mslabelra{\beta\eta\pi} N'$ and $\lambda\pi \vdash M'= N'$. Therefore
$\lambda\pi \vdash M=N$. \qed
\end{enumerate}
\noindent Another useful result is the forthcoming lemma, which says that all ${\beta\eta\pi}$-reduction paths may be ``simulated''
by a reduction path which contains ${\pi}$-steps only at the end. We indicate by $\to^=_{\beta\eta}$ the reflexive closure of
$\labelra{\beta\eta}$.
\begin{lem}[Factorization]\label{lem:factorization}
If $M \mslabelra{\beta\eta\pi} N$, then there exists $P$ such that $M \mslabelra{\beta\eta} P \mslabelra{\pi} N$.
\end{lem}
\proof
Use iteratively the fact that whenever $M \labelra{\pi} N \labelra{\beta\eta} Q$, then there exists $N'$
such that $M \to^=_{\beta\eta} N' \mslabelra{\pi} Q$.
\qed
\begin{lem}\label{lem:theta-not-omega}
The terms $\Theta$ and $\Omega$ are not ${\beta\eta\pi}$-convertible.
\end{lem}
\proof
By the reduction graph of $\Theta$ and the confluence of $\labelra{\beta\eta\pi}$.
\qed
\subsection{Jacopini\textendash Kuper technique for $\lambda\pi\phi$}\label{subs:Jac-for-lpp}
In this section we apply the Jacopini\textendash Kuper technique explained in Section \ref{subs:Jacopini-Kuper} to prove the consistency
of the theory $\lambda\pi\phi$.
More precisely, the results presented here show that the closure $\lambda x.\Theta x\Omega = \ssi$ of the equation $(\phi)$,
that axiomatizes $\lambda\pi\phi$ over $\lambda\pi$, satisfies the hypotheses of Theorem \ref{thm:reduce-incons}.
\begin{lem}\label{lem:bottom-preorder}
The term $\lambda x.\Theta x \Omega$ is $\lambda\pi$-operationally less defined than $\ssi$.
\end{lem}
\proof
Let $F$ be a $\lambda$-term and $N$ be a $\beta\eta$-normal form, and suppose
\mbox{$F(\lambda x.\Theta x \Omega) =_{\lambda\pi} N$}.
Since ${\lambda\pi}$-reduction is confluent and $N$ is $\beta\eta$-normal, we have that
\mbox{$F(\lambda x.\Theta x\Omega) \mslabelra{\beta\eta\pi} N$}.
By Lemma \ref{lem:factorization} there exists a term $M$ such that
$$F(\lambda x.\Theta x \Omega) \mslabelra{\beta\eta} M \mslabelra{\pi} N.$$
Since $N$ is a $\beta\eta$-normal form and the reduct $\Omega$ of a $\labelra{\pi}$-redex is a $\beta$-redex,
we must have that $M \equiv N$. Therefore we have $\lambda\beta\eta \vdash F(\lambda x.\Theta x \Omega) = N$ and, since
$\lambda x.\Theta x \Omega$ is unsolvable, we can apply the Genericity Lemma of lambda calculus (see Lemma \ref{gen}) to obtain
$\lambda\beta\eta \vdash F\ssi = N$, and hence obviously $\lambda\pi \vdash F\ssi = N$ which is the desired conclusion.
\qed
In Lemma \ref{lem:proof-subst} below we keep track of the residuals of the $\lambda$-term $\lambda x.\Theta x \Omega$ during the reduction of the term
$F(\lambda x.\Theta x \Omega)$. We have three kinds of residuals: $\lambda x.\Psi x\Omega$, $\Psi M\Omega$ and $\Omega$ (with $\Psi\in {\mathcal G}_\beta(\Theta)$) as the following informal example shows:
\[
\begin{array}{llll}
F(\lambda x.\Theta x \Omega) &\mslabelra{\lambda\pi} &\cdots (\lambda x.\Theta x \Omega)\cdots (\lambda x.\Theta x \Omega)\cdots (\lambda x.\Theta x \Omega)\cdots &\\
& \mslabelra{\beta} &\cdots (\lambda x.\Theta x \Omega)\cdots (\lambda x.\Psi x \Omega)\cdots\cdots (\lambda x.\Theta x \Omega)\cdots&\text{($\Theta\mslabelra\beta \Psi$)} \\
& \mslabelra{\lambda\pi} &\cdots (\lambda x.\Theta x \Omega)\cdots (\lambda x.\Psi x \Omega)M\cdots (\lambda x.\Theta x \Omega)\cdots& \\
& \labelra{\beta} &\cdots (\lambda x.\Theta x \Omega)\cdots (\Psi M \Omega)\cdots (\lambda x.\Theta x \Omega)\cdots &\text{($\beta$-reduction)} \\
& \mslabelra{\lambda\pi} &\cdots (\lambda x.\Theta x \Omega)\cdots (\Psi N \Omega)\cdots (\lambda x.\Theta x \Omega)\cdots&\text{($M\mslabelra{\lambda\pi} N$)}\\
& \labelra{\pi} &\cdots (\lambda x.\Theta x \Omega)\cdots (\Omega) \cdots (\lambda x.\Theta x \Omega) \cdots&\text{($N=_{\lambda\pi} \Omega$)}\\
& \mslabelra{\lambda\pi} & \cdots\cdots\cdots&
\end{array}
\]
In order to trace the residuals it is useful to enrich the syntax of $\lambda$-terms with labels as follows:
\[
\begin{array}{lllr}
\quad M,N & ::= & x \mid \lambda x.M \mid MN & \\
& & \mid (\lambda x.\Psi x\Omega)^n \mid (\Psi M\Omega)^n \mid (\Omega)^n & (n\geq 1 \text{ and } \Psi\in {\mathcal G}_\beta(\Theta)) \\
\end{array}
\]
We denote by $\Lambda^{\mathbb{N}}$ the set of labelled terms and we write $\underline{M}$ for the $\lambda$-term, called \emph{erasure} of $M$, obtained by erasing the labels from $M$.
Since $(\Omega)^n$ and $(\lambda x.\Psi x\Omega)^n$ are closed terms, then it is sufficient to extend substitution to labelled terms by setting $(\Psi M\Omega)^n[N/x] = (\Psi M[N/x] \Omega)^n$,
where $\Psi\in {\mathcal G}_\beta(\Theta)$. Then we define a reduction on labeled terms as the smallest contextual reduction $\labelra{\mathrm{lab}}$ (reduction under labels is allowed) satisfying the following clauses, for all labelled terms $M,N$:
\[
\begin{array}{ll}
(\lambda x.M)N \labelra{\mathrm{lab}} M[N/x] & \\
\lambda x.Mx \labelra{\mathrm{lab}} M & \text{ if } x\not\in \mathrm{FV}(M) \\
(\lambda x.\Psi x\Omega)^n M \labelra{\mathrm{lab}} (\Psi M\Omega)^n & \text{ if } \Psi\in {\mathcal G}_\beta(\Theta) \\
\Psi MN \labelra{\mathrm{lab}} \Omega & \text{ if $\Psi\in {\mathcal G}_\beta(\Theta)$ and $\underline{M} =_{\lambda\pi} \underline{N}$.}
\end{array}
\]
Note that (i) $\labelra{\beta\eta\pi} \subseteq \labelra{\mathrm{lab}}$; (ii) if $M\labelra{\mathrm{lab}} N$ then $(\Psi M\Omega)^n \labelra{\mathrm{lab}} (\Psi N \Omega)^n$. If $\sigma$ is a reduction path of labelled terms, then we denote by
$\underline \sigma$ the corresponding reduction path, where all labels are erased.
We will make use of an additional operation on labelled terms. Given terms $M,N \in \Lambda^{\mathbb{N}}$ such that
$\underline{M} \equiv \underline{N}$, we define their \emph{superposition} as the labelled term obtained from the syntax tree of $\underline M$
by adding a possible label $k$ to each subtree $T$ of $\underline M$ according to the following schema:
\begin{itemize}
\item Put $k = m+n$ if $T$ has label $m$ in $M$ and $n$ in $N$;
\item Put $k=m$ if $T$ has label $m$ in $M$ and no label in $N$;
\item Put $k = n$ if $T$ has label $n$ in $N$ and no label in $M$;
\item Put no label otherwise.
\end{itemize}
\begin{lem}\label{lem:proof-subst}
The term $\lambda x.\Theta x \Omega$ is $\lambda\pi$-proof-substitutable by $\ssi$.
\end{lem}
\proof
In this proof $\Psi$ ranges over ${\mathcal G}_\beta(\Theta)$. Let $F_1,F_2$ be closed $\lambda$-terms and suppose $F_1(\lambda x.\Theta x \Omega) =_{\lambda\pi} F_2(\lambda x.\Theta x \Omega)$. Since the reduction
$\labelra{\beta\eta\pi}$ is confluent, then the two sides of the equality are the beginning of two reduction paths $\sigma_1$ and $\sigma_2$ that end in a common term $R$.
Consider now the labelled terms $F_i(\lambda x.\Theta x \Omega)^{i}$ for $i=1,2$. Then there exists a labelled reduction path $\sigma'_i$ starting with $F_i(\lambda x.\Theta x \Omega)^{i}$ such that $\underline{\sigma'_i} \equiv \sigma_i$. Notice that the label $1$ is the unique label occurring in the reduction path $\sigma'_1$, while the label 2 is the unique label occurring in the reduction path $\sigma'_2$.
We denote by $R_i$ the last labelled term in the reduction path $\sigma'_i$. Then we have that $R\equiv \underline R_1\equiv \underline R_2$.
Let $P$ be the term obtained by superposition of $R_1$ and $R_2$. Then the labels of $P$ range over the set ${\mathcal L}=\{ 1,2,3\}$.
We now describe how to extract a witness of $\lambda\pi$-proof-substitutability by suitably modifying $P$. All residuals with label $3$ in $P$ are common to the reduction paths $\sigma'_1$ and $\sigma'_2$. Then, if we mimic the reduction path $\sigma_i$ starting from $F_i\ssi$ ($i=1,2$), we will find in place of the residuals with label $3$
the term $\ssi$ for $(\lambda x.\Psi x \Omega)^{3}$; $M$ for $(\Psi M \Omega)^{3}$ and a term $N$ ($\lambda\pi$-convertible with $\Omega$) for $(\Omega)^{3}$:
\[
\begin{array}{lll}
F_i(\lambda x.\Theta x \Omega) &\mslabelra{\lambda\pi} & \text{($i = 1,2$)} \\
\cdots (\lambda x.\Theta x \Omega)\cdots &\mslabelra{\beta} & \\
\cdots (\lambda x.\Psi x \Omega)\cdots & \mslabelra{\lambda\pi} &\text{($\Theta\mslabelra\beta \Psi$)} \\
\cdots (\lambda x.\Psi x \Omega)M\cdots & \labelra{\beta} & \\
\cdots \Psi M \Omega\cdots & \labelra{\lambda\pi} &\text{($M\mslabelra{\lambda\pi} N$)} \\
\cdots \Psi N \Omega\cdots & \labelra{\lambda\pi} &\text{($N=_{\lambda\pi} \Omega$)}\\
\cdots \Omega \cdots & &\\
\end{array}\qquad\qquad
\begin{array}{lll}
F_i\ssi &\mslabelra{\lambda\pi} & \text{($i = 1,2$)} \\
\cdots \ssi\cdots &\equiv & \\
\cdots \ssi\cdots & \mslabelra{\lambda\pi} & \\
\cdots \ssi M\cdots & \labelra{\beta} & \\
\cdots M\cdots & \labelra{\lambda\pi} &\text{($M\mslabelra{\lambda\pi} N$)} \\
\cdots N\cdots & \equiv &\text{($N=_{\lambda\pi} \Omega$)}\\
\cdots N \cdots & &\\
\end{array}
\]
Then we substitute all occurrences of the label $3$ in the term $P$ as follows:
$$Q \equiv P[\ssi/(\lambda x.\Psi x \Omega)^{3}; M/(\Psi M \Omega)^{3}; \Omega/(\Omega)^{3}].$$
The last substitution $\Omega$ for $(\Omega)^3$ is possible because the term $N$ in the above reduction path (right column) is $\lambda\pi$-convertible with $\Omega$.
We see that, by mimicking the steps in the paths $\sigma_1,\sigma_2$, we have that
$$(\ast)\qquad F_i\ssi =_{\lambda\pi} L_i \text{, where $L_i$ is the erasure of } Q[\ssi/(\lambda x.\Psi x \Omega)^{i}; M/(\Psi M \Omega)^{i}; \Omega/(\Omega)^{i}] \quad (i=1,2) $$
Let $x_1,x_2$ be fresh variables and let $H$ be the term without labels obtained from $Q$ by replacing bottom-up the subterms
$$
\text{ for $i=1,2$ }
\begin{cases}
(\lambda x.\Psi x \Omega)^i & \text{ with $x_i$;} \\
(\Psi M \Omega)^i & \text{ with $x_iM$;} \\
(\Omega)^i & \text{ with $x_i\Omega$.}
\end{cases}
$$
Then the following equivalences hold:
\begin{enumerate}[label=(\alph*)]
\item $L_1 =_{\lambda\pi} H[\ssi/x_1; (\lambda x.\Psi x \Omega)/x_2]$;
\item $L_2 =_{\lambda\pi} H[(\lambda x.\Psi x \Omega)/x_1;\ssi/x_2]$.
\end{enumerate}
Therefore by setting $G\equiv \lambda x_2x_1.H$, we obtain that
\begin{itemize}
\item $G(\lambda x.\Theta x\Omega)\ssi \mslabelra{\beta} H[\ssi/x_1; (\lambda x.\Psi x \Omega)/x_2] =_{\lambda\pi} L_1 = _{\lambda\pi} F_1\ssi$, \qquad by ($\ast$) and (a)
\item $G\ssi(\lambda x.\Theta x\Omega) \mslabelra{\beta} H[(\lambda x.\Psi x \Omega)/x_1;\ssi/x_2] =_{\lambda\pi} L_2 =_{\lambda\pi} F_2\ssi$, \qquad by ($\ast$) and (b)
\end{itemize}
This shows that $G$ is the witness term we were looking for.
\qed
Now we are ready to give the main theorem of the section.
\begin{thm}\label{thm:reduce-incons-lambdaphipi}
The $\lambda$-theory $\lambda\pi\phi$ is consistent.
\end{thm}
\proof
Lemma \ref{lem:proof-subst} and Lemma \ref{lem:bottom-preorder} show that
the hypotheses of Theorem \ref{thm:reduce-incons} are satisfied by the equation that
axiomatizes $\lambda\pi\phi$ over $\lambda\pi$, and therefore $\lambda\pi\phi$ must be consistent.
\qed
\section{On the equational inconsistency of the pointed ordered models}\label{sec:Hon-Plo-question}
In this section we find a counterexample to the equational consistency of the class of pointed ordered models:
we prove that there is no pointed ordered model satisfying the equations ($\pi$) and ($\phi$).
\begin{lem}\label{lem:connected-components}
Let ${\mathcal M}$ be an ordered model such that ${\mathcal M}\models \Theta xx=\Omega \land \Theta x\Omega = x$ (i.e., $Th({\mathcal M}) \supseteq \lambda\pi\phi$). Then for all closed $\lambda$-terms $P$ and $Q$ we have:
\begin{enumerate}[label=(\roman*)]
\item If ${\mathcal M} \not\models\Theta PQ= \Omega$, then the interpretations of $P$ and $Q$ are in distinct connected components of ${\mathcal M}$.
\item The connected component of the interpretation of $\Omega$ is a singleton set.
\end{enumerate}
\end{lem}
\proof (i) Following \cite[Section 4]{Salibra03} we define the subtraction sequence of the pair $(P,Q)$:
$$s_1 \equiv \Theta PQ; \qquad s_{n+1} \equiv \Theta s_n \Omega.$$
By hypothesis ${\mathcal M}\models s_1 \neq\Omega$ and by subtractivity ${\mathcal M}\models s_n = s_1$ for all $n$.
Then the conclusion follows from \cite[Corollary 4.6]{Salibra03}.
(ii) Since ${\mathcal M}$ is a subtractive ordered model then the conclusion follows from Lemma \ref{lem-ord}.
\qed
The situation described by Lemma \ref{lem:connected-components} can be regarded to as a relativized version of absolute unorderability to one fixed element.
In particular the interpretation of $\Omega$ is isolated in every model. This property will be studied in Section \ref{sec:subtractivity}
in the framework of Universal Algebra.
The class of pointed ordered models is not consistent with respect to the set of quantifier-free sentences: the quantifier-free sentence
$\lambda x.\Omega xx = \lambda x.\Omega \land \Omega\neq\Omega\gO(\Omega\ssT\ssi)$ is consistent with the $\lambda$-calculus but no pointed ordered model satisfies
it. This result was shown by Salibra for the $\lambda$-calculus (see the remark after \cite[Corollary 3.6]{Salibra03}) and by
Honsell\textendash Plotkin for the extensional $\lambda$-calculus (see \cite[Theorem 7]{HonsellP09}): Honsell and Plotkin also ask
whether there exist a finite set of equations consistent with the $\lambda$-calculus that are not satisfied by any pointed ordered model.
The following theorem answer their question: the subtractive equations are indeed a counterexample to the equational consistency for the class of pointed ordered models.
\begin{thm}
No pointed ordered model ${\mathcal M}$ simultaneously satisfies the equations $\Theta xx=\Omega\ \land\ \Theta x\Omega = x$.
\end{thm}
\proof
Suppose, by contradiction, that ${\mathcal M}\models \Theta xx=\Omega \land \Theta x\Omega = x$.
Since $|\Omega|^{{\mathcal M}}$ is comparable with $\bot$, then by Lemma \ref{lem:connected-components}(ii) $\Omega$ is interpreted as the bottom
element $\bot$ of ${\mathcal M}$. Since the bottom element is comparable with all other elements of ${\mathcal M}$, from Lemma \ref{lem:connected-components}(ii) it follows that ${\mathcal M}$ is trivial.
\qed
We can also get a stronger result.
\begin{thm}
No connected ordered model simultaneously satisfies the equations $\Theta xx=\Omega\ \land\ \Theta x\Omega = x$.
\end{thm}
\proof
By Lemma \ref{lem:connected-components}(ii).
\qed
\section{On the order-incompleteness of $\lambda$-calculus}\label{subs:strenghtening}
The \emph{order-incompleteness} problem of $\lambda$-calculus, raised by Selinger in \cite{Selinger03}, can be characterised in terms of connected components of a partial ordering:
a $\lambda$-theory ${\mathcal T}$ is order-incomplete if, and only if, every ordered model ${\mathcal M}$ such that $Th({\mathcal M}) = {\mathcal T}$ is partitioned in an infinite number of connected components, each one containing exactly one element. In other words, the partial order is the equality.
So far we have shown that the subtractive equations force the connected component of $\Omega$ in a model to be a singleton set, so that the model cannot be connected as a partial order. However, the
order-incompleteness is something more than disconnected. Toward order-incompleteness, we propose a strengthening ${\mathcal T}$ of the
$\lambda$-theory $\lambda\pi\phi$ having the following property: every ordered model ${\mathcal M}$ such that $Th({\mathcal M}) \supseteq {\mathcal T}$ has an infinite
number of connected components among which that of $\Omega$ is a singleton set. Moreover each connected component contains the
denotation of at most one $\beta\eta$-normal form.
We define a family of unsolvable terms $\Theta_n$ (for $n\geq 0$) obtained as follows:
\begin{itemize}
\item Define inductively $A_0 \equiv x$ and $A_{n+1} \equiv \lambda y.yA_n$, where $y\not\equiv x$. Note that $\mathrm{FV}(A_n) = \{x\}$, for each $n\in{\mathbb{N}}$.
\item Now set $B_n \equiv \lambda x.xA_n$, $C_n \equiv \lambda z.zB_n$ (where $z\not\equiv x$) and $\Theta_n \equiv B_nC_n$.
\end{itemize}
Note that $\Theta_0 \mslabelra{\beta} \Omega$ and
$\Theta_1 \equiv \Theta$ (recall the definition of $\Theta$ from Section \ref{sus:lam-calc}).
\begin{lem}
The terms $\Theta_n$ are closed zero-terms such that
$\lambda\pi\vdash\Theta_n = \Theta_m$ iff $m=n$.
\end{lem}
\proof
It follows from the confluence of the reduction
$\labelra{\beta\eta\pi}$ and from the form of the reduction graphs of the terms in question. Each graph ${\mathcal G}_{\beta\eta\pi}(\Theta_n)$
is a cycle whose edges are only $\labelra{\beta}$ reductions, and ${\mathcal G}_{\beta\eta\pi}(\Theta_n)$ is disjoint from ${\mathcal G}_{\beta\eta\pi}(\Theta_m)$
whenever $n\neq m$.
\qed
We are now going to introduce the above-mentioned strenghtening of $\lambda\pi\phi$. In what follows we let ${\mathcal T}$ be the theory axiomatized over
$\lambda\pi\phi$ by the following equations:
\[
\begin{array}{ll}
\Theta_2\Omega = \ssT; & \\
\Theta_2(\Theta MN) = \ssF, &\quad \text{$M$ and $N$ distinct closed $\beta\eta$-normal forms.}
\end{array}
\]
(recall the definiton of $\ssT$ and $\ssF$ from Section \ref{sus:lam-calc}.)
Next we show that ${\mathcal T}$ is consistent. In order to do that it suffices, by compactness reasons, to prove that any finite subset of the above equations
is eliminable from a proof of ${\mathcal T} \vdash \ssT = \ssF$ via the Jacopini\textendash Kuper technique.
The proof of this fact closely resembles the consistency proof given for $\lambda\pi\phi$ (see Section \ref{subs:Jac-for-lpp}), so
we will just sketch it, only considering the extension of $\lambda\pi$ by three equations $\Theta_2(\Theta MN) = \ssF$, $\Theta_2\Omega = \ssT$ and
$\lambda x.\Theta x\Omega = \ssi$, where $(M,N)$ is an arbitrary but fixed pair of closed distinct $\beta\eta$-normal forms.
Define the two sequences $\vec P = \Theta_2(\Theta MN),\Theta_2\Omega,\lambda x.\Theta x\Omega$ and $\vec Q = \ssF,\ssT,\ssi$.
\begin{lem} $\vec P$ is $\lambda\pi$-operationally less defined than $\vec Q$.
\end{lem}
\proof
As the proof of Lemma \ref{lem:bottom-preorder}.
\qed
\begin{lem}\label{lem:proof-rep-2}
$\vec P$ is $\lambda\pi$-proof-substitutable by $\vec Q$.
\end{lem}
\proof
In this proof $\Psi$ and $\Psi_2$ range, respectively, over ${\mathcal G}_\beta(\Theta)$ and ${\mathcal G}_\beta(\Theta_2)$.
Let $F_1,F_2$ be closed $\lambda$-terms and suppose $F_1\vec P =_{\lambda\pi} F_2\vec P$. Since the reduction
$\labelra{\beta\eta\pi}$ is confluent, then the two sides of the equality are the beginning of two reduction paths $\sigma_1$ and $\sigma_2$ that end in a common term $R$.
Consider now the labelled term
\begin{itemize}
\item $A_1 \equiv F_1(\Theta_2(\Theta MN))^{1}(\Theta_2\Omega)^{4}(\lambda x.\Theta x \Omega)^{10}$
\item $A_2 \equiv F_2(\Theta_2(\Theta MN))^{2}(\Theta_2\Omega)^{5}(\lambda x.\Theta x \Omega)^{11}$
\end{itemize}
Then there exist labelled reduction paths $\sigma'_i$ starting with $A_i$ ($i=1,2$) such that $\underline{\sigma'_i} \equiv \sigma_i$.
We denote by $R_i$ the last labelled term in the reduction path $\sigma'_i$. Then we have $R\equiv \underline R_i$ ($i = 1,2$).
Let $S$ be the term obtained by superposition of $R_1$ and $R_2$. Then the labels of $S$ range over the set ${\mathcal L}=\{1,2,3,4,5,9,10,11\}$. Note that if $S$ has a labelled subterm of the shape $(\Theta_2\Omega)^{l}$, then $l \in \{4,5,9\}$
because the contrary would require $\Theta MN \labelra{\beta\eta\pi} \Omega$ (by Theorem \ref{thm:betapil-confluent}(ii)), which is impossible because it would imply
$\lambda\pi \vdash M = N$, contradicting the consistency of $\lambda\pi$ (as a consequence of B\"{o}hm's Theorem \cite[Thm.~10.4.2]{Bare}).
We now describe how to extract a witness of $\lambda\pi$-proof-substitutability by suitably modifying $S$. All residuals with label $3$ ,$9$, or $21$ in $S$
are common to the reduction paths $\sigma'_1$ and $\sigma'_2$. Then, if we mimic the reduction path $\sigma_i$ starting from $F_i\ssi$ ($i=1,2$), we will find in place
of the residuals with label $21$ the term $\ssi$ for $(\lambda x.\Psi x \Omega)^{21}$; $M$ for $(\Psi M \Omega)^{21}$ and a term $N$ ($\lambda\pi$-convertible with $\Omega$) for $(\Omega)^{21}$.
Similarly those residuals with labels $3$ and $9$ are replaced by $\ssF$ and $\ssT$, respectively. Then we let
$$S' \equiv S[\ssi/(\lambda x.\Psi x \Omega)^{21}; M/(\Psi M \Omega)^{21}; \Omega/(\Omega)^{21};\ssF/(\Theta_2(\Theta MN))^{3};\ssT/(\Theta_2\Omega)^{9}]$$
and we define a term $H$ out of $S'$ by replacing bottom-up some subterms (labeled by $i\in {\mathcal L}$), using fresh variables $x_{1},x_{2},x_3,x_4,x_5,x_{10},x_{11}$ as follows
$$
\text{ for $i=10,11$ }
\begin{cases}
(\lambda x.\Psi x \Omega)^i & \text{with $x_i$;} \\
(\Psi M \Omega)^i & \text{with $x_iM$;} \\
(\Omega)^i & \text{with $x_i\Omega$.}
\end{cases}
\quad
\text{for $i=4,5$ and $j=1,2$}
\begin{cases}
(\Psi_2\Omega)^i & \text{with $x_i$;} \\
(\Psi_2(\Psi MN))^j & \text{with $x_j$.}
\end{cases}
$$
Finally, as in the proof of Lemma \ref{lem:proof-subst}, it is possible to find a term $G$ such that:
\[
\begin{array}{rclcl}
G\vec P \vec Q & \mslabelra{\beta} & H[(\Psi_2(\Psi MN)) / x_1; \ssF / x_{2} ; (\Psi_2\Omega)/ x_{4}; \ssT /x_5; \ssi / x_{10}; (\lambda x.\Psi x \Omega)/ x_{11}]
& =_{\lambda\pi} & F_1\vec Q \\
& & & & \\
G\vec Q \vec P & \mslabelra{\beta} & H[\ssF / x_1;(\Psi_2(\Psi MN)) / x_{2} ;\ssT / x_{4}; (\Psi_2\Omega)/x_5; (\lambda x.\Psi x \Omega)/ x_{10}; \ssi/ x_{11}]
& =_{\lambda\pi} & F_2\vec Q \\
\end{array}
\]
\qed
The following proposition, which relies on Lemma \ref{lem:proof-rep-2}, it is analogous to Theorem \ref{thm:reduce-incons-lambdaphipi}.
\begin{thm}\label{thm:super-cool}
The $\lambda$-theory ${\mathcal T}$ is consistent.
\end{thm}
In \cite{Salibra03} it is shown that the semantics of $\lambda$-calculus given in terms of ordered models with a finite number of connected components is theory incomplete. In Theorem \ref{thm:order-inc} below we improve this result.
\begin{lem}\label{lemma:infinite-components}
Let ${\mathcal M}$ be an ordered model such that $Th({\mathcal M}) \supseteq {\mathcal T}$. Then ${\mathcal M}$ has an infinite number of connected components and it has the following properties:
\begin{enumerate}[label=(\roman*)]
\item The interpretation in ${\mathcal M}$ of two distinct $\beta\eta$-normal forms belongs to different connected components;
\item The connected component of $\Omega$ is a singleton set.
\end{enumerate}
\end{lem}
\proof
Let $M,N$ be two distinct $\beta\eta$-normal forms and suppose, by way of contradiction, that $M$ and $N$ lie in the same connected component
of ${\mathcal M}$. Then ${\mathcal M} \models \Theta MN = \Omega$ by Lemma \ref{lem:connected-components}(i).
From ${\mathcal M} \models \ssF = \Theta_2(\Theta MN)$ and ${\mathcal M} \models \Theta_2\Omega = \ssT$ we derive that ${\mathcal M} \models \ssF = \ssT$,
which contradicts the non-triviality of ${\mathcal M}$. Hence each denotation of a $\beta\eta$-normal form belongs to exactly one connected component.
The second part of the statement follows directly from Lemma \ref{lem:connected-components}(ii).
\qed
\begin{thm}\label{thm:order-inc} The following classes of ordered models are theory-incomplete:
\begin{enumerate}[label=(\roman*)]
\item Models with a finite number of connected components.
\item Models with an infinite number of connected components such that the connected component of $\Omega$ has cardinality $>1$.
\end{enumerate}
\end{thm}
\proof By Lemma \ref{lemma:infinite-components}.
\qed
\section{Subtractivity and orderings}\label{sec:subtractivity}
The inspiration for the subtractive equations comes from a general algebraic framework, developed by Ursini \cite{Ursini}, called \emph{subtractivity}. Salibra in \cite{Salibra03} investigated the weaker notion of \emph{semi-subtractivity}, linking it to properties of ordered models of $\lambda$-calculus.
Here we follow that path illustrating the stronger properties of subtractivity.
We start the section reviewing the connection
established by Selinger in \cite{Selinger03} between the absolute unorderability and the validity of certain Mal'cev-type conditions.
\subsection{Unorderability and absolute unorderability}
Let $\tau$ be an algebraic similarity type and ${\bf A}$ be an algebra of type $\tau$.
We say that ${\bf A}$ is \emph{unorderable} if it admits only equality as a compatible partial order.
The following result is due to Hagemann \cite{Hag73,HM73} (see also Coleman \cite[Theorem 1.6]{Coleman97}).
\begin{thm}\label{thm-unord} Let ${\mathcal V}$ be a variety. Then the following conditions are equivalent:
\begin{enumerate}
\item Every algebra in ${\mathcal V}$ is unorderable;
\item Every compatible preorder on an algebra in ${\mathcal V}$ is symmetric (and thus a congruence).
\item There exist a natural number $n\geq 2$ and ternary terms $p_1,\ldots,p_{n-1}$
in the type of ${\mathcal V}$ such that the following Mal'cev identities hold in ${\mathcal V}$:
\vspace*{-1.6mm}
$$
\begin{array}{rcl}
x & = & p_1(x,y,y); \\
p_i(x,x,y) & = & p_{i+1}(x,y,y) \quad (i=1,\dots,n-2); \\
p_{n-1}(x,x,y) & = & y.
\end{array}
$$
\end{enumerate}
\end{thm}
\proof The equivalence of (2) and (3) is \cite[Theorem 1.6]{Coleman97}. We now prove that (1) implies (2).
Let ${\bf A}\in{\mathcal V}$ and $\leq$ be a compatible preorder on ${\bf A}$.
Let $\approx$ be the congruence on ${\bf A}$ generated by $\leq$, that is,
$$\text{$a \approx b$ iff $a\leq b$ and $b\leq a\quad$ (for all $a,b\in A$).}$$
By hypothesis the partial ordering on the quotient algebra ${\bf A}/\approx$, defined by
$[a]_\approx \sqsubseteq [b]_\approx$ iff $a\leq b$, is trivial.
Then $\leq$ is symmetric.
\qed
\begin{defi}
Let ${\mathcal C}$ be a class of algebras of type $\tau$ and ${\bf A}\in {\mathcal C}$.
We say that ${\bf A}$ is \emph{absolutely unorderable in ${\mathcal C}$} if, for every algebra ${\bf B}\in {\mathcal C}$ and
every embedding $f: {\bf A} \to {\bf B}$ (i.e., injective homomorphism), the algebra ${\bf B}$ is unorderable.
\end{defi}
Let ${\mathcal V}$ be a variety of type $\tau$ and ${\bf A}\in {\mathcal V}$. We denote by $\tau(A)$ the algebraic similarity
type $\tau$ enriched by a constant $\bar a$
for each element $a\in A$. The algebra ${\bf A}$ becomes a $\tau(A)$-algebra by interpreting each constant $\bar a$
with the element $a\in A$.
The {\it equational diagram\/} of ${\bf A}$ is defined as
the set of all \emph{ground identities} $t = u$ of type $\tau(A)$ such that ${\bf A}\models t=u$.
We denote by ${\mathcal V}_{\bf A}$ the variety of type $\tau(A)$ axiomatized by the equational theory $Eq({\mathcal V})$ of ${\mathcal V}$
and the equational diagram of ${\bf A}$.
There is a bijective correspondence between algebras of ${\mathcal V}_{\bf A}$ and pairs $({\bf B},f)$, where
${\bf B}\in{\mathcal V}$ and $f:{\bf A}\to{\bf B}$ is
a $\tau$-homomorphism. Indeed,
the algebra ${\bf B}$ can be turned into a $\tau(A)$-algebra in ${\mathcal V}_{\bf A}$, by interpreting each constant $\bar a$ with the element $f(a)$ of $B$.
If $X$ is a set of indeterminates, then the {\it free extension ${\bf A}[X]$ of ${\bf A}$ by $X$} in the
variety ${\mathcal V}$ is the free $\tau(A)$-algebra over $X$
in the variety ${\mathcal V}_{\bf A}$.
The algebra ${\bf A}[X]$ can be also defined up to isomorphism by the following universal mapping properties: (1) $A\cup X\subseteq A[X]$; (2) ${\bf A}[X]\in {\mathcal V}$; (3) for every ${\bf B}\in {\mathcal V}$, $\tau$-homomorphism $h:{\bf A}\to{\bf B}$ and every function $f:X\to B$, there exists a unique $\tau$-homomorphism $g:{\bf A}[X]\to {\bf B}$ extending $h$ and $f$.
When $X= \{x_1,\dots, x_n\}$ is finite, we write ${\bf A}[x_1,\dots, x_n]$ for ${\bf A}[X]$.
We are now ready to give the main result of this subsection that characterizes those algebras which are absolutely unorderable in a variety. Notice that the equivalence of items (1) and (3) below was shown by Selinger in \cite[Theorem 3.4]{Selinger03}, while the equivalence of (1) and (2) was suggested by a referee.
\begin{thm}\label{ab-thm} Let ${\mathcal V}$ be a variety of type $\tau$ and ${\bf A}\in {\mathcal V}$. Then the following conditions are equivalent:
\begin{enumerate}
\item ${\bf A}$ is absolutely unorderable in ${\mathcal V}$;
\item Every algebra in the variety ${\mathcal V}_{\bf A}$ is unorderable;
\item The algebra ${\bf A}[x,y]$ satisfies the Mal'cev identities of Theorem \ref{thm-unord}.
\end{enumerate}
\end{thm}
\proof (1) $\Rightarrow$ (2) Assume, by contradiction, ${\bf B}$ to be a $\tau(A)$-algebra in ${\mathcal V}_{\bf A}$, which admits a nontrivial compatible partial order $\leq$.
Then there are two distinct elements $a,b\in B$ such that $a\leq b$.
If $f:X\to B$ is an onto map from a set $X$ of indeterminates to $B$, then there are two indeterminates $x,y\in X$ such that $f(x)= a$ and $f(y)=b$. Let $\theta$ be the least compatible preorder on ${\bf A}[X]$ such that
$x\theta y$, and
let $\approx$ be the congruence on ${\bf A}[X]$ generated by $\theta$, that is,
$$\text{$t \approx u$ iff $t\theta u$ and $u\theta t\quad$ (for all $t,u\in {\bf A}[X]$).}$$
Then the relation $\sqsubseteq$, defined by $[t]_{\approx} \sqsubseteq [u]_{\approx}$ iff $t\theta u$, is a compatible partial ordering on ${\bf A}[X]/\!\!\approx$.
By the hypothesis $x\theta y$ we obtain $[x]_{\approx}\sqsubseteq [y]_{\approx}$. Since the map $\iota: {\bf A}\to {\bf A}[X]/\!\!\approx$, defined by $\iota(a) = [a]_{\approx}$, is an embedding and ${\bf A}$ is absolutely unorderable in ${\mathcal V}$, then
we get $[y]_{\approx}= [x]_{\approx}$, so that $y\theta x$ holds.
Since ${\bf B}\in {\mathcal V}_{\bf A}$, then the map $h:A\to B$, defined by $h(a)= {\bar a}^{\bf B}$, is a $\tau$-homomorphism. Consider the
unique $\tau$-homomorphism $g:{\bf A}[X]\to {\bf B}$ extending $h$ and $f$.
Since $\theta \subseteq \{(t,u) : g(t) \leq g(u)\}$, then by $y\theta x$ we get $b=g(y) \leq g(x)=a$, that together with $a\leq b$ implies $a=b$. Contradiction.
(2) $\Leftarrow$ (1) If ${\bf B}\in{\mathcal V}$ and $f: {\bf A}\to{\bf B}$ is an embedding, then ${\bf B}$ can be seen as a $\tau(A)$-algebra in ${\mathcal V}_{\bf A}$, which is unorderable by hypothesis.
(2) $\Leftrightarrow$ (3) follows from Theorem \ref{thm-unord}.
\qed
It is unknown whether there exist absolutely unorderable algebras in the variety of combinatory algebras. Plotkin and Simpson have shown that the Mal'cev
identities are inconsistent with combinatory logic for $n=2$, while Plotkin and Selinger have obtained the same result for $n=3$ (see \cite{Selinger03}). It is unknown whether the Mal'cev identities are consistent with combinatory logic for $n\geq 4$.
In \cite{Salibra03} the second author has shown that there exists a quasi-variety of combinatory algebras which only admits absolutely unorderable algebras.
We recall that a quasi-variety is a class of algebras axiomatized by quasi-identities (i.e., equational implications with a finite number of equational
premises).
\begin{thm} \rm{(\cite{Salibra03})} Let ${\mathcal Q}$ be the quasi-variety of combinatory algebras axiomatized by the identity $\Omega xx=\Omega$ and the quasi-identity $\Omega xy=\Omega \Rightarrow x=y$. Then ${\mathcal Q}$ is nontrivial and every algebra ${\bf A}\in{\mathcal Q}$ is absolutely unorderable in ${\mathcal Q}$.
\end{thm}
\proof
${\mathcal Q}$ is nontrivial because the term model of the consistent $\lambda$-theory axiomatised by $\Omega xx=\Omega$ belongs to ${\mathcal Q}$. Moreover, it is easy to show that every algebra of ${\mathcal Q}$ is unorderable.
\qed
\subsection{Absolute $0$-unorderability}
In the case a variety ${\mathcal V}$ has two constants $0$ and $1$, the Mal'cev identities of Theorem \ref{thm-unord} give:
$$
\begin{array}{rcl}
0 & = & p_1(0,1,1); \\
p_i(0,0,1) & = & p_{i+1}(0,1,1) \quad (i=1,\dots,n-2); \\
p_{n-1}(0,0,1) & = & 1.
\end{array}
$$
If we define the unary term operations $f_i(x)= p_i(0,x,1)$, then the above identities can be written as follows:
\begin{equation}\label{eq}
0= f_1(1);\qquad f_i(0) = f_{i+1}(1) \quad (i=1,\dots,n-2);\qquad f_{n-1}(0) = 1.
\end{equation}
This suggests the following theorem, whose proof is similar to the proof of Theorem \ref{thm-0-unord} below and it is omitted.
\begin{thm}
Let ${\mathcal V}$ be a variety with two constants $0$ and $1$. Then the
constants $0$ and $1$ are incomparable in all ordered algebras in ${\mathcal V}$ if, and only if,
there exist a natural number $n\geq 2$ and unary terms $f_1,\ldots,f_{n-1}$
in the type of ${\mathcal V}$ such that the identities (\ref{eq}) hold in ${\mathcal V}$.
\end{thm}
In the case a variety ${\mathcal V}$ has a constant $0$, then we can relativise the Mal'cev identities of Theorem \ref{thm-unord} as follows:
$$
\begin{array}{rcl}
0 & = & p_1(0,y,y); \\
p_i(0,0,y) & = & p_{i+1}(0,y,y) \quad (i=1,\dots,n-2); \\
p_{n-1}(0,0,y) & = & y.
\end{array}
$$
If we define the binary term operations $s_i(y,x)= p_i(0,x,y)$, then the above identities can be written as follows:
\begin{equation}\label{eq2}
\begin{array}{rcl}
0 & = & s_1(x,x) \\
s_i(x,0) & = & s_{i+1}(x,x) \quad (i=1,\dots,n-2); \\
s_{n-1}(x,0) & = & x.
\end{array}
\end{equation}
This suggests that the absolute unorderability relative to the element $0$ can be expressed by the following identities defining $n$-subtractivity.
\begin{defi}
An algebra ${\bf A}$ is \emph{$n$-subtractive} ($n\geq 2$) if there exist a constant $0$ and $n-1$ binary terms $s_1(x,y),\dots, s_{n-1}(x,y)$ such that ${\bf A}$ satisfies identities (\ref{eq2}).
A variety of algebras is $n$-subtractive if every algebra in ${\mathcal V}$ is $n$-subtractive with respect to the same constant and term operations.
\end{defi}
Then Ursini's subtractivity of Definition \ref{sub} means $2$-subtractivity: this is the strongest notion since it is easy to verify that
an $n$-subtractive algebra is also $m$-subtractive for every $m > n$.
Every model of the two equations $(\pi)$ and $(\phi)$ is subtractive, when we define the binary subtractivity operator $s_1(x,y)$ as the $\lambda$-term $\Theta xy$. As a consequence of the consistency of the $\lambda$-theory $\lambda\pi\phi$, it follows that
there exists a non-trivial subtractive variety of combinatory algebras.
\begin{defi}
An algebra ${\bf A}$ with $0$ is
\begin{enumerate}[label=(\roman*)]
\item \emph{$0$-unorderable} if, for every compatible partial order $\leq$ on $A$ and every $a\in A\setminus\{0\}$, neither $0\leq a$ nor $a\leq 0$.
\item \emph{$0$-symmetric} if, for every compatible preorder $\leq$ on $A$ and every $a\in A\setminus\{0\}$, we have $0\leq a \Leftrightarrow a\leq 0$.
\end{enumerate}
\end{defi}
\begin{prop}\label{subunord} Every $n$-subtractive algebra is $0$-unorderable and $0$-symmetric.
\end{prop}
\proof
Let ${\bf A}$ be $n$-subtractive and $\leq$ be a compatible preorder on $A$. If $a\in A$ and $a\leq 0$, then
$$0= s_{1}(a,a) \leq s_{1}(a,0)= s_{2}(a,a) \leq \dots\leq s_{n-2}(a,0)=s_{n-1}(a,a) \leq s_{n-1}(a,0) = a.$$
If $0\leq a$ a similar reasoning works.
\qed
The following theorem relativizes Theorem \ref{thm-unord} to $0$-unorderability.
\begin{thm}\label{thm-0-unord} Let ${\mathcal V}$ be a variety with a constant $0$. Then the following conditions are equivalent:
\begin{enumerate}
\item Every algebra in ${\mathcal V}$ is $0$-unorderable;
\item Every compatible preorder on an algebra in ${\mathcal V}$ is $0$-symmetric.
\item ${\mathcal V}$ is $n$-subtractive for some $n$.
\end{enumerate}
\end{thm}
\proof (2) $\Rightarrow$ (3):
Define a compatible relation $\prec$ on the free algebra ${\bf T}_{\mathcal V}[x]$ as follows:
$t \prec u$ iff there exists a binary term $p(x,y)$ such that ${\mathcal V}\models t = p(x,x)$ and ${\mathcal V} \models p(x,0) = u$.
The condition $x \prec 0$ is witnessed by the polynomial $p(x,y) \equiv y$. The reflexive and transitive closure $\prec^*$ of $\prec$ is a compatible preorder on ${\bf T}_{\mathcal V}[x]$. Then by hypothesis we derive $0\prec^* x$. This implies the existence of binary terms $s_1,\ldots,s_{n-1}$ which witness $n$-subtractivity.
(3) $\Rightarrow$ (2): By Proposition \ref{subunord}.
(1) $\Leftrightarrow$ (2): as in Theorem \ref{thm-unord}.
\qed
\begin{defi} Let ${\mathcal C}$ be a class of algebras with a constant $0$.
An algebra ${\bf A}\in {\mathcal C}$ is said to be \emph{absolutely $0$-unorderable in ${\mathcal C}$} if, for any algebra ${\bf B}\in {\mathcal C}$ and embedding $f: {\bf A} \to{\bf B}$, ${\bf B}$ is $0$-unorderable.
\end{defi}
\begin{thm}\label{ab2-thm} Let ${\mathcal V}$ be a variety of type $\tau$ and ${\bf A}\in {\mathcal V}$. Then the following conditions are equivalent:
\begin{enumerate}
\item ${\bf A}$ is absolutely $0$-unorderable in ${\mathcal V}$;
\item Every algebra in the variety ${\mathcal V}_{\bf A}$ is $0$-unorderable;
\item The algebra ${\bf A}[x]$ is $n$-subtractive for some $n$.
\end{enumerate}
\end{thm}
\proof (2) $\Leftrightarrow$ (3) follows from Theorem \ref{thm-0-unord}.
(1) $\Rightarrow$ (2): As the corresponding proof of Theorem \ref{ab-thm}, where the role of the indeterminate $y$ is taken by the constant $0$.
(2) $\Rightarrow$ (1): If ${\bf B}\in{\mathcal V}$ and $f: {\bf A}\to{\bf B}$ is an embedding, then ${\bf B}$ can be seen as a $\tau(A)$-algebra in ${\mathcal V}_{\bf A}$, which is $0$-unorderable by hypothesis.
\qed
\begin{cor}\label{mostro}
Every $\lambda$-model satisfying the subtractive identities $\pi$ and $\phi$
is absolutely $\Omega$-unorderable in the variety of combinatory algebras.
\end{cor}
\proof Let ${\bf A}$ be a $\lambda$-model satisfying the subtractive identities $\pi$ and $\phi$.
By \cite[Proposition 20]{Sel-alg} and \cite[Theorem 4.2]{PigozziS98}, for all $\lambda$-terms $M$ and $N$, ${\bf A} \models M=N$ if, and only if, ${\bf A}[x]\models M = N$. Then ${\bf A}[x]$ is also subtractive. Then the conclusion follows from Theorem \ref{ab2-thm}.
\qed
We would like to conclude this section by remarking that
Ursini \cite{Ursini} has shown that subtractive algebras have a good theory of ideals.
We recall that ideals in general algebras generalize normal subgroups, ideals in rings, filters in Boolean or Heyting algebras, ideals in Banach algebra, in $l$-groups, etc. One feature of subtractive varieties is that their ideals are exactly the congruence classes of $0$, but one does not have the usual one-one correspondence ideals-congruences: mapping a congruence $\theta$ to its equivalence class $0/\theta$ only establishes a lattice homomorphism between the congruence lattice and the ideal lattice. This points to another feature: the join of two congruences is a tricky thing to deal with. The join of two ideals in a subtractive algebra behaves nicely: for $I,J$ ideals, we have that $b\in I\lor J$ iff for some $a\in I$, $s(b, a)\in J$. Thanks to the consistency of the subtractive equations with $\lambda$-calculus, the theory of ideals for subtractive varieties can be applied to all $\lambda$-theories extending $\lambda\pi\phi$.
\section{Subtractivity and topology}\label{sec:subtractivity-topology}
In this section we provide a topological incompleteness theorem for the $\lambda$-calculus as a consequence of a study of conditions of separability for $n$-subtractive algebras.
The classification of the models of lambda calculus into
orderable/unorderable models was refined as follows in \cite{Salibra03}.
For every algebra ${\bf A}$, let $T_i^{\bf A}$ ($i = 0, 1, 2, 2_{1/2}$)
be the set of all topologies $\tau$ on $A$
which make $({\bf A},\tau)$ a $T_i$-topological algebra.
It is obvious that in general we have
$$T_0^{\bf A} \supseteq T_1^{\bf A} \supseteq T_2^{\bf A} \supseteq T_{2_{1/2}}^{\bf A}.$$
A topology $\tau$
with a non-trivial specialization order (we have $a <_\tau b$ for some $a,b$)
would be $T_0$ yet not $T_1$, so that
$$\mbox{${\bf A}$ is unorderable iff $T_0^{\bf A} = T_1^{\bf A}$.}$$
\begin{defi}
We say that ${\bf A}$ is of \emph{topological type $i$} ($i = 1, 2, 2_{1/2}$) if $T_0^{{\bf A}} = T_i^{{\bf A}}$.
A variety ${\mathcal V}$ of algebras is of topological type $i$ if every algebra ${\bf A}\in{\mathcal V}$ is of topological type $i$.
\end{defi}
\begin{exa} We recall from \cite{Salibra03} that a lambda theory ${\mathcal T}$ is {\it of topological type $i$} ($i = 1, 2, 2_{1/2}$)
if the term model ${\mathcal M}_{\mathcal T}$ of ${\mathcal T}$ satisfies $T_0^{{\mathcal M}_{\mathcal T}} = T_i^{{\mathcal M}_{\mathcal T}}$.
The lambda theory ${\mathcal B}$, generated by equating two lambda terms
if they have the same B\"ohm tree, is not of type $1$
(see \cite{Bare}).
$\lambda\beta$ and $\lambda\beta\eta$ are of type 1 by Selinger's result \cite{Selinger03},
while the lambda theory $\Pi$, generated by the equation $\Omega xx=\Omega$, was shown of type $2_{1/2}$ in \cite{Salibra03}.
\end{exa}
We now refine the topological axioms of separability as follows.
\begin{defi} Let $(X,\tau)$ be a space and $a\in X$.
We say that $X$ is {\it $T_i$-separated in $a$} ($i =0,1,2, 2_{1/2}$) if, for all $b\in X\setminus \{a\}$,
$a$ and $b$ are $T_i$-separated.
\end{defi}
For every algebra ${\bf A}$ and $a\in A$, let $T_i^{\bf A}(a)$ ($i = 0, 1, 2, 2_{1/2}$)
be the set of all topologies $\tau$ on $A$
which make $({\bf A},\tau)$ a topological algebra which is $T_i$-separated in $a$.
We have
$$\mbox{${\bf A}$ is $a$-unorderable iff $T_0^{\bf A}(a) = T_1^{\bf A}(a)$.}$$
\begin{defi}
We say that ${\bf A}$ is of \emph{topological type $i$ in $a$} ($i = 1, 2, 2_{1/2}$) if $T_0^{{\bf A}}(a) = T_i^{{\bf A}}(a)$.
A variety ${\mathcal V}$ of algebras, whose type contains a constant $a$, is of topological type $i$ in $a$ if every algebra ${\bf A}\in{\mathcal V}$ is of topological type $i$ in $a$.
\end{defi}
We know from Proposition \ref{subunord} that the variety of combinatory algebras generated by the term model of $\lambda\pi\phi$ is of topological type $1$ in $\Omega$.
This result will be improved in the next subsection.
\subsection{The topological incompleteness theorem for $\lambda$-calculus}
The \emph{$i$-diagonal} $Diag_i({\bf A})$ of an $n$-subtractive algebra ${\bf A}$ is the set of elements $a\in A$ such that
$s_i(a,a) =0$.
Notice that (i) $Diag_1({\bf A}) = A$; (ii) $0\in Diag_i({\bf A})$ for all $i$; (iii) $Diag_i({\bf A}) \supseteq Diag_{n-1}({\bf A})$ for every $i$.
\begin{lem}\label{separated} Every $n$-subtractive $T_0$-semitopological algebra $({\bf A},\tau)$ is $T_1$-separated in the element $0$.
\end{lem}
\proof
By Lemma \ref{1.3.1bis} the specialisation order $\leq_\tau$ is compatible. Then the conclusion follows from Proposition \ref{subunord}.
\qed
\begin{cor}\label{closed}
Let $({\bf A},\tau)$ be an $n$-subtractive $T_0$-semitopological algebra. Then we have:
\begin{enumerate}[label=(\roman*)]
\item The singleton set $\{0\}$ is closed;
\item The sets $Diag_i({\bf A})$ ($1\leq i\leq n-1$) are closed.
\end{enumerate}
\end{cor}
\proof (i) An element $a$ belongs to the closure of $\{0\}$ if, and only if, $a\leq_\tau 0$. By Proposition \ref{subunord} ${\bf A}$ is $0$-unorderable. Then the set $\{0\}$ is closed.
(ii) $Diag_i({\bf A})$ is the inverse image of the closed set $0$ with respect to the continuous unary polynomial $s_i(x,x)$.
\qed
\begin{lem}\label{semi} Let $({\bf A},\tau)$ be a semitopological algebra in an arbitrary similarity type, $t(x,y)$ be a binary term operation and $a,b\in A$ be two elements such that
$$t(a,a)=t(b,b);\qquad t(a,b)\neq t(a,a).$$
If $t(a,b)$ and $t(a,a)$ are $T_0$-separated, then $a$ and $b$ are $T_1$-separated.
\end{lem}
\proof Let $c\equiv t(a,a)$ in this proof.
Assume first there exists a neighbourhood $U$ of $c$ with
$t(a,b)\notin U$.
From $t(a,a) = c \in U$ and $t(b,b) = c \in U$ and from the
separated continuity of $t$ it follows that there exist an open neighbourhood
$V$ of $a$ and an open neighbourhood $W$ of $b$ such that
$t(a,V)\subseteq U$ and $t(W,b)\subseteq U$.
The condition $b\in V$ or $a\in W$
contradicts the hypothesis that $t(a,b)\notin U$.
Assume now there exists a neighbourhood $U$ of $t(a,b)$ with $c\notin U$.
From $t(a,b)\in U$ and from the
separated continuity of $s$ it follows that there exist an open neighbourhood
$V$ of $a$ and an open neighbourhood $W$ of $b$ such that
$t(a,W)\subseteq U$ and $t(V,b)\subseteq U$. The condition $b\in V$ or $a\in W$
contradicts the hypothesis that $c\notin U$.
\qed
\begin{cor} Let $(\sbA,\tau)$ be an $n$-subtractive $T_0$-semitopological algebra and let $a,b\in Diag_i({\bf A})$. If $s_i(a,b)\neq 0$
then $a$ and $b$ are $T_1$-separated.
\end{cor}
\proof
By Lemma \ref{separated} $s_i(a,b)$ and $0$ are $T_1$-separated.
Then the conclusion follows from Lemma \ref{semi}.
\qed
The following theorem is a slight generalization of \cite[Theorem 5.2]{Salibra03}.
We are indebted to \cite{Bentz99} for the technique used in the proof.
\begin{thm}\label{sep}
Let $(\sbA,\tau)$ be an $n$-subtractive $T_0$-topological algebra and let $a,b\in Diag_i({\bf A})$.
\begin{enumerate}[label=(\roman*)]
\item If $s_i(a,b)\neq 0$, then $a$ and $b$ are $T_2$-separated in the subspace $Diag_i({\bf A})$.
\item If $s_i(a,b)$ and $0$ are
$T_2$-separated, then $a$ and $b$ are $T_{2_{1/2}}$-separated in the subspace $Diag_i({\bf A})$.
\end{enumerate}
\end{thm}
\proof
(i) By Lemma \ref{separated} there exists an open neighborhood $U$ of $s_i(a,b)$ such that $0\notin U$. From $s_i(a,b) \in U$ and from the
continuity of $s_i$ it follows that there exist an open neighborhood
$V$ of $a$ and an open neighbourhood $W$ of $b$ such that $s_i(V,W)\subseteq U$.
If there exists $d\in V\cap W\cap Diag_i({\bf A})$ then $0 = s_i(d,d) \in U$ contradicting the
hypothesis.
(ii) By hypothesis there exist
an open neighbourhood $V'$ of $s_i(a,b)$ and an open neighbourhood $W'$ of $0$
such that $V' \cap W'= \emptyset$.
Since $s_i$ is continuous and $s_i(a,b) \in V'$, there exist two other open
sets $V$
and $W$ containing
$a$ and $b$, respectively, such that $s_i(V,W) \subseteq V'$. The pre-image of
$\overline{V'}$ under the map $s_i$
is closed. From $s_i(V,W) \subseteq
V'\subseteq \overline{V'}$
the pre-image of $\overline{V'}$, that is closed, contains $V\times W$, so
$s_i(\oV,\oW) \subseteq \overline{V'}$.
We now prove that $\oV \cap \oW\cap Diag_i({\bf A}) = \emptyset$.
Assume, by the way of contradiction, that there is
$e\in \oV\cap\oW\cap Diag_i({\bf A})$.
Since $s_i(\oV,\oW) \subseteq \overline{V'}$ it follows that
$0 = s_i(e,e) \in \overline{V'}$.
But by definition of closure of a set
this is possible only if for every
open neighbourhood $Z$ of $0$, we have that $Z\cap V' \neq \emptyset$.
But this contradicts our initial choice of $V'$ and $W'$
as two open neighbourhoods of $s(a,b)$ and $0$ respectively
with empty intersection.
\qed
\begin{cor}\label{sep2} Let $(\sbA,\tau)$ be a $2$-subtractive $T_0$-topological algebra and let $a,b\in A$.
If $s_1(a,b)\neq 0$, then $a$ and $b$ are $T_{2_{1/2}}$-separated. In particular, for all $a\in A\setminus \{0\}$, $a$ and $0$ are $T_{2_{1/2}}$-separated.
\end{cor}
\proof Recall that $Diag_{1}({\bf A})=A$. By Theorem \ref{sep}(i) and $s_1(s_1(a,b),0)= s_1(a,b)$ we have that $s_1(a,b)$ and $0$ are $T_2$-separated.
Then we apply Theorem \ref{sep}(ii) to get the conclusion.
\qed
We cannot generalise Theorem \ref{sep} and Corollary \ref{sep2} to semitopological algebras as explained by the following counterexample.
\begin{exa}
The Visser topology of $\lambda$-calculus (see \cite{Bare, Visser80}) on the set $\Lambda$ of $\lambda$-terms is the topology generated by the following family of sets: $U\subseteq \Lambda$ is a base open set if it is closed under $\beta$-conversion and it is the complement of an r.e. set. The Visser topology on the term model of the $\lambda$-theory $\lambda\pi\phi$ is the quotient topology of the Visser topology on $\Lambda$. It makes the term model of $\lambda\pi\phi$ a semitopological algebra, but not a topological algebra, because Theorem \ref{sep} and Corollary \ref{sep2} are false for the term model of $\lambda\pi\phi$ with the Visser topology. In fact, the Visser topology on $\Lambda$ was shown hyperconnected by Visser in \cite{Visser80}. We recall that a topology is hyperconnected if the intersection of two arbitrary nonempty open sets is nonempty.
\end{exa}
We conclude the section by applying the above results to $\lambda$-calculus.
\begin{cor}
The variety of combinatory algebras generated by the term model of $\lambda\pi\phi$ is of topological type $2_{1/2}$ in $\Omega$.
\end{cor}
The topological incompleteness theorem of \cite{Salibra03} states that the semantics of $\lambda$-calculus given in terms of coconnected topological models is incomplete (see Section \ref{top} for the definition of coconnected space). Coconnected topological models include all pointed ordered models of $\lambda$-calculus.
In the following theorem we strongly improve the topological incompleteness theorem of \cite{Salibra03}.
Notice that a topological model ${\mathcal M}$ is not $T_{2_{1/2}}$-separated in $\Omega$ if there exists an element $a$ in the model such that for all opens $U,V$ with $a\in U$ and $\Omega\in V$ we have $\oU \cap \oV \neq \emptyset$.
\begin{thm}
The semantics of $\lambda$-calculus given in terms of topological models which are not $T_{2_{1/2}}$-separated in $\Omega$ is theory incomplete.
\end{thm}
\subsection{Separability in $n$-subtractive varieties}
In this last subsection we study conditions of separability for $n$-subtractive varieties of (semi)topological algebras.
After some general results about $n$-subtractive semitopological algebras, we show how $n$-subtracivity
in $T_0$-topological algebras induces a generalized version of Hausdorffness due to J.P. Coleman's. Of course the focus is on
$T_0$-topological algebras because they include the vast majority of partially ordered models of $\lambda$-calculus.
As a particular case of this study we obtain a result stating that any $2$-subtractive $T_0$-topological algebra
is $T_2$-separated in $0$.
\begin{defi}
Let ${\bf A}$ be an $n$-subtractive algebra ($n\geq 2$) and $a\in A\setminus \{0\}$.
The \emph{rank} $\kappa(a)$ of $a$ is
the least natural number $k$ such that $s_{k}(a,0) \neq 0$.
\end{defi}
Since $s_{n-1}(a,0) = a$, then the rank $\kappa(a)$ exists
and we have $1\leq \kappa(a)\leq n-1$.
\begin{lem} $a\in Diag_{\kappa(a)}({\bf A})$.
\end{lem}
\proof If $\kappa(a)=1$ the result follows from the identity $s_1(x,x)=0$. If $\kappa(a)>1$ we have $s_{\kappa(a)}(a,a) = s_{\kappa(a)-1}(a,0) =0$.
\qed
Define, for every $0\leq i \leq n-1$, $R_i = \{ a : \kappa(a) \leq i\}$.
Then $R_0 =\emptyset$, $R_i\subseteq R_{i+1}$ and $R_{n-1} = A/ \{ 0\}$.
\begin{lem}\label{open1} Let $({\bf A},\tau)$ be an $n$-subtractive semitopological algebra.
Then $R_i$ is open for every $i$.
\end{lem}
\proof
If $a\in A$ then $\kappa(a)\leq i$ if, and only if, there is $j\leq i$ such that $s_j(a,0)\neq 0$. Then
$$R_i = \bigcup_{j=1}^i \{a : s_j(a,0)\neq 0\}.$$
Each set $\{a : s_j(a,0)\neq 0\}$ is open because $\{0\}$ is closed, $A\setminus \{0\}$ is open and
$\{a : s_j(a,0)\neq 0\}$ is the inverse image of $A\setminus \{0\}$ by the continuous function $s_j(x,0)$.
\qed
We define for $i\geq 1$
$$\Sigma_{i} = \{ a : \mbox{($\exists U,V\in \tau$) $a\in U$, $0\in V$ and $U\cap V \subseteq R_{i-1}$} \}$$
We have: (i) $\Sigma_1$ is the set of all elements $a$ such that $a$ and
$0$ are $T_2$-separated; (ii) $\Sigma_{n} = A\setminus \{0\}$; $\Sigma_i \subseteq\Sigma_{i+1}$ for every $i$.
\begin{lem}\label{open2} Let $({\bf A},\tau)$ be an $n$-subtractive semitopological algebra.
We have for every $1\leq i\leq n$:
\begin{enumerate}[label=(\roman*)]
\item $\Sigma_{i}$ is open.
\item $R_{i-1}\subseteq \Sigma_{i}$.
\end{enumerate}
\end{lem}
\proo
(i) Let $a\in \Sigma_{i}$. We have to show that there exists an open
neighborhood of $a$ contained within $\Sigma_{i}$. If $a\in \Sigma_{i}$
then the exist $U,V\in \tau$ such that $a\in U$, $0\in V$ and $U\cap
V \subseteq R_{i-1}$. Then $U$ is an open neighborhood of $a$
contained within $\Sigma_{i}$.
(ii) By Lemma \ref{open1} $R_{i-1}$ is open. Moreover, $R_{i-1}$ is
an open neighborhood of each of its elements and we trivially have
$R_{i-1}\cap V \subseteq R_{i-1}$ for every (and then some) open
neighborhood $V$ of $0$.\qed
\noindent In this theorem we improve Lemma \ref{open2}(ii) in the hypothesis that ${\bf A}$ is a $T_0$-topological algebra.
\begin{thm}\label{RS} Let $({\bf A},\tau)$ be an $n$-subtractive
$T_0$-topological algebra. Then we have:
\begin{enumerate}[label=(\roman*)]
\item $a\in \Sigma_{\kappa(a)}$ for every $a\in A\setminus \{0\}$.
\item $R_{i}\subseteq \Sigma_{i}$.
\item $\Sigma_{n-1} = A\setminus \{0\}$.
\end{enumerate}
\end{thm}
\proof
(i) Since $s_{\kappa(a)}(a,0) \neq 0$, then
by Lemma \ref{separated} there exists an open neighbourhood $W$ of $s_{\kappa(a)}(a,0)$ such that $0\notin W$. Then
we have
$$s_{\kappa(a)}(a,0)\in W.$$
By the continuity of $s_{\kappa(a)}$ there exist two open neighbourhoods $U$ and $V$
of $a$ and $0$ respectively such that
$$s_{\kappa(a)}(U,V) \subseteq W.$$
If $\kappa(a) = 1$ and there exists $b\in U\cap V$, then $0= s_1(b,b) \in W$, contradicting the hypothesis on $W$. Then $V\cap U=\emptyset$; thus $a$ and $0$ are $T_2$-separated,
and $a\in \Sigma_1$.
If $\kappa(a) > 1$, then $U\cap V\neq \emptyset$. For every $b\in U\cap V$ we have that
$$s_{\kappa(a)}(b,b) \in W,$$
that implies
$$s_{\kappa(a) -1}(b,0) = s_{\kappa(a)}(b,b) \neq 0.$$
This means that the rank of $b$ is less than the rank of $a$ for every $b\in U\cap V$. Then $U\cap V\subseteq R_{\kappa(a) -1}$, so that $a\in \Sigma_{\kappa(a)}$.
(ii) Trivial by (i).
(iii) follows from (ii).\qed
Let $(A,\tau)$ be a topological space. For every $a\in A$,
define by induction the following
family of subsets of $A$:
\begin{enumerate}
\item $\Gamma_0(a) = \emptyset$;
\item $\Gamma_{i+1}(a) = \{ b : \exists\
\mbox{open $U,V$ with $a\in U$, $b\in V$ and $U\cap V \subseteq \Gamma_i(a)$} \}$.
\end{enumerate}
Coleman \cite{Coleman96,Coleman97} defines $(A,\tau)$ to be \emph{$n$-step Hausdorff} if $\Gamma_n(a) = A\setminus\{ a\}$ for all $a\in A$. $1$-step Hausdorff
is equivalent to $T_2$. A variety of algebras is $n$-step Hausdorff if every topological algebra in the variety is $n$-step Hausdorff.
A variety ${\mathcal V}$ of algebras is $n$-permutable ($n\geq 2$) iff every algebra in ${\mathcal V}$ satisfies the Mal'cev identities of Theorem \ref{thm-unord} (see \cite{BS}).
Every $n$-permutable variety has been shown to be $\lfloor n/2\rfloor$-step Hausdorff by Kearnes and Sequeira \cite{KearnesS02}.
We define a topological algebra $(\sbA,\tau)$ to be \emph{$n$-step Hausdorff in $0$} if $\Gamma_n(0) = A/\{ 0\}$. $1$-step Hausdorff in $0$
is equivalent to $T_2$-separated in $0$.
\begin{prop} Let $(\sbA,\tau)$ be an $n$-subtractive $T_0$-topological algebra. Then $(\sbA,\tau)$ is $n-1$-step Hausdorff in $0$.
\end{prop}
\proof We show by induction that $\Sigma_i \subseteq \Gamma_i(0)$ for all $1\leq i \leq n$. For $i=0$ the result is trivial.
\[
\begin{array}{llll}
\Sigma_{i+1} & = & \{ b : \exists\
\mbox{open $U,V$ with $a\in U$, $b\in V$ and $U\cap V \subseteq R_i$} \}&\text{by definition} \\
& \subseteq & \{ b : \exists\
\mbox{open $U,V$ with $a\in U$, $b\in V$ and $U\cap V \subseteq \Sigma_i$} \} &\text{by Thm. \ref{RS}(ii)} \\
& \subseteq & \{ b : \exists\
\mbox{open $U,V$ with $a\in U$, $b\in V$ and $U\cap V \subseteq \Gamma_i(0)$} \} &\text{by induction hypothesis}\\
& = & \Gamma_{i+1}(0) &\text{by definition}
\end{array}
\]
The conclusion follows because $\Sigma_{n-1}=A\setminus\{0\}$.
\qed
|
\section*{Introduction}
Every compact Riemann surface $X$ is an algebraic curve over $\mathbb C$, and every
meromorphic function on $X$ is an algebraic function. This remarkable fact,
generalized in the GAGA principle, links the analytic with the algebraic in a
fundamental way. A natural problem is then to link this further with
arithmetic; to characterize those Riemann surfaces that can be defined by
equations over $\overline{\mathbb Q}$ and to study the action of the absolute Galois group
$\Gal(\overline{\mathbb Q}/\mathbb Q)$ on these algebraic curves. To this end, Bely\u{\i}\
\cite{Belyi,Belyi2} proved that a Riemann surface $X$ over $\mathbb C$ can be defined
over $\overline{\mathbb Q}$ if and only if $X$ admits a \textsf{Bely\u{\i}\ map}, a map $f:X \to
\mathbb P^1_\mathbb C$ that is unramified away from $\{0,1,\infty\}$. Grothendieck, in his
\emph{Esquisse d'un Programme} \cite{Grothendieck}, called this result ``deep
and disconcerting.''
Part of Grothendieck's fascination with Bely\u{\i} 's theorem was a consequence of
the simple combinatorial and topological characterization that follows from it.
Given a Bely\u{\i}\ map $f:X \to \mathbb P^1_\mathbb C$, the preimage $f^{-1}([0,1])$ of the real
interval $[0,1]$ can be given the structure of a \textsf{dessin} (or \textsf{dessin
d'enfant}, ``child's drawing''): a connected graph with bicolored vertices (so
the two vertices of an edge are colored differently) equipped with a cyclic
ordering of the edges around each vertex. Conversely, a
dessin determines the corresponding Bely\u{\i}\ map uniquely up to isomorphism over
$\mathbb C$ or $\overline{\mathbb Q}$. The idea that one can understand the complicated group
$\Gal(\overline{\mathbb Q}/\mathbb Q)$ by looking at children's pictures casts an alluring spell
indeed. As a consequence, hundreds of papers have been written on the subject,
several books have appeared, and the topic remains an active area of research
with many strands.
In a number of these papers, computation of particular examples plays a key role
in understanding phemonena surrounding Bely\u{\i}\ maps; arguably, part of the
richness of the subject lies in the beauty in these examples. Shabat and
Voevodsky \cite[0.1.1, 0.3]{SV} say on this point:
\begin{quote}
Here we have no general theory and only give a number of examples. The
completeness of our results decrease rapidly with growing genus; we are able
to give some complete lists (of non-trivial experimental material) for genus
$0$, but for genera exceeding $3$ we are able to give only some general
remarks. [...] The main reasons to publish our results in the present state is
our eagerness to invite our colleagues into the world of the divine beauty and
simplicity we have been living in since we have been guided by the Esquisse.
\end{quote}
In spite of this important role, no survey of computational methods for Bely\u{\i}\
maps has yet appeared, and in our own calculations we found many techniques,
shortcuts, and some tricks that others had also (re)discovered. In this
article, we collect these results in one place in the hope that it will be
useful to others working in one of the many subjects that touch the theory of
Bely\u{\i}\ maps. We also give many examples; to our knowledge, the larger examples
are new, unless mentioned otherwise. We assume that the reader has some
familiarity with algebraic curves and with computation, but not necessarily with
the theory of Bely\u{\i}\ maps or dessins; at the same time, we hope that this paper
will also be a useful and comprehensive reference, so we will also make some
remarks for the experts.
We take as input to our methods the simple group theoretic description of
Bely\u{\i}\ maps: there is a bijection between \textsf{permutation triples}
\begin{center}
$\sigma = (\sigma_0, \sigma_1, \sigma_{\infty}) \in S_d^3$ that
satisfy $\sigma_0 \sigma_1 \sigma_{\infty} = 1$
\end{center}
up to simultaneous conjugation in the symmetric group $S_d$, and
\begin{center}
Bely\u{\i}\ maps $f:X \to \mathbb P^1$ of degree $d$
\end{center}
up to isomorphism over $\overline{\mathbb Q}$. In this bijection, the curve $X$ can be
disconnected, such as the trivial cover of degree $d > 1$; the cover $X$ is
connected if and only if (the dessin is connected if and only if) the
corresponding permutation triple $\sigma$ generates a transitive subgroup of
$S_d$, in which case we call $\sigma$ \textsf{transitive}. If $\sigma$
corresponds to $f$ in this bijection, we say that $f$ has \textsf{monodromy
representation $\sigma$}.
Given the description of a Bely\u{\i}\ map $f$ in the compressed form of a
permutation triple, it has proven difficult in general to determine explicitly
an algebraic model for $f$ and the curve $X$. As a result, many authors have
written on this subject of explicit computation of Bely\u{\i}\ maps, usually subject
to certain constraints or within a certain class of examples. That this is a
difficult problem is a common refrain, and the following quote by Magot and
Zvonkin \cite[\S 1]{MagotZvonkin} is typical:
\begin{quote}
An explicit computation of a Belyi function corresponding to a given map is
reduced to a solution of a system of algebraic equations. It may turn out to
be extremely difficult. To give an idea of the level of difficulty, we
mention that our attempts to compute Belyi functions for some maps with only
six edges took us several months, and the result was achieved only after using
some advanced Gr\"obner bases software and numerous consultations given by its
author J.C. Faug\`ere.
\end{quote}
The paper is organized as follows. In Section 1, we collect the basic
background (including a discussion of fields of definition), and mention some
applications and generalizations. In Section 2, we discuss a direct method
using Gr\"obner methods, augmented by the Atkin--Swinnerton-Dyer trick. We then
turn to other, more practical methods. We begin in Section 3 with complex
analytic methods; in Section 4, we consider methods using modular forms; in
Section 5, we consider $p$-adic methods. In Section 6, we briefly discuss
alternative methods for Galois Bely\u{\i}\ maps. In Section 7, we discuss the
delicate subjects of field of moduli and field of definition with an eye to its
implications for computation. In Section 8, we treat simplification and
verification of Bely\u{\i}\ maps, and finally in Section 9 we conclude by
considering some further topics and generalizations. Along the way, we give
explicit examples and pose several questions.
The authors would like to thank Noam Elkies, Ariyan Javanpeykar, Curtis
McMullen, John McKay, David Roberts, Steffen Rohde, Sam Schiavone, Matthias
Sch\"utt, Marco Streng, Bernd Sturmfels, Mark Watkins and Bruce Westbury for their comments on
this work, as well as the referee for his or her many suggestions. The first
author was supported by Marie Curie grant IEF-GA-2011-299887, and the second
author was supported by an NSF CAREER Award (DMS-1151047).
\section{Background and applications}\label{sec:backg}
The subject of explicit characterization and computation of ramified covers of
Riemann surfaces is almost as old as Riemann himself. Klein \cite{Klein} and
Fricke--Klein \cite{FrickeKlein} calculated some explicit Bely\u{\i}\ maps, most
notably the icosahedral Galois Bely\u{\i}\ map $\mathbb P^1 \rightarrow \mathbb P^1$ of degree
$60$ \cite[I, 2, \S 13--14]{Klein}. These appeared when constructing what we
would today call modular functions associated with the triangle groups
$\Delta(2,3,5)$ and $\Delta(2,4,5)$ (see Section \ref{sec:mod}). This in turn
allowed them to find a solution to the quintic equation by using analytic
functions. Around the same time, Hurwitz \cite{Hurwitz} was the first to
consider ramified covers in some generality: besides considering covers of small
degree, he was the first to give the classical combinatorical description of
covers of the projective line minus a finite number of points, which would later
result in Hurwitz spaces being named after him.
Continuing up to the modern day, the existing literature on Bely\u{\i}\ maps with an
explicit flavor is extremely rich: surveys include Birch \cite{Birch},
Jones--Singerman \cite{JonesSingerman,JonesSingerman2}, Schneps \cite{Schneps},
and Wolfart \cite{WolfartDessins}; textbooks include the seminal conference
proceedings \cite{Schnepsbook}, work of Malle--Matzat \cite{MalleMatzat}, Serre
\cite{Serre}, and V\"olklein \cite{Volklein}, mainly with an eye toward
applications to inverse Galois theory, the tome on graphs on surfaces by
Lando--Zvonkin \cite{LandoZvonkin}, and the book by Girondo--Gonzalez-Diaz
\cite{GDG}, which interweaves the subject with an introduction to Riemann
surfaces.
We begin this section by reviewing basic definitions; we conclude by mentioning
applications and generalizations as motivation for further study. (We postpone
some subtle issues concerning fields of moduli and fields of definition until
Section \ref{sec:fomfod}.)
\subsection*{Definitions, and equivalent categories}
Let $K$ be a field with algebraic closure $\overline{K}$. An \textsf{(algebraic)
curve} $X$ over $K$ is a smooth proper separated scheme of finite type over $K$
that is pure of dimension $1$.
We now define precisely the main category of this paper whose objects we wish to
study. A \textsf{Bely\u{\i}\ map over $K$} is a morphism $f:X \to \mathbb P^1$ of curves
over $K$ that is unramified outside $\{0,1,\infty\}$. Given two Bely\u{\i}\ maps
$f_1,f_2:X_1,X_2 \to \mathbb P^1$, a \textsf{morphism} of Bely\u{\i}\ maps from $f_1$ to
$f_2$ is a morphism $g : X_1 \to X_2$ such that $f_1 = f_2 g$. We thereby obtain
a category of Bely\u{\i}\ maps over $K$.
A curve $X$ that admits a Bely\u{\i}\ map is called a \textsf{Bely\u{\i}\ curve}. Bely\u{\i}\
\cite{Belyi,Belyi2} proved that a curve $X$ over $\mathbb C$ can be defined over
$\overline{\mathbb Q}$ if and only if $X$ is a Bely\u{\i}\ curve. Consequently, in what follows,
we may pass freely between Bely\u{\i}\ maps over $\overline{\mathbb Q}$ and over $\mathbb C$: we will
simply refer to both categories as the category of \textsf{Bely\u{\i}\ maps}. The
absolute Galois group $\Gal (\overline{\mathbb Q} /\mathbb Q)$ acts naturally on (the objects and
morphisms in) the category of Bely\u{\i}\ maps (over $\overline{\mathbb Q}$); this action is
faithful, as one can see by considering the $j$-invariant of elliptic curves.
We denote the action by a superscript on the right, so the conjugate of a curve
$X$ over $\overline{\mathbb Q}$ by an automorphism $\tau \in \Gal (\overline{\mathbb Q}/\mathbb Q)$ is denoted by
$X^{\tau}$, and that of a Bely\u{\i}\ map $f$ by $f^{\tau}$.
Let $f:X \to \mathbb P^1$ be a Bely\u{\i}\ map of degree $d$. The ramification of $f$
above $\{0,1,\infty\}$ is recorded in its \textsf{ramification type}, the triple
consisting of the set of ramification multiplicities above $0,1,\infty$,
respectively. Such a ramification type is therefore given by a triple of
partitions of $d$, or alternatively by a triple of conjugacy classes in the
symmetric group $S_d$.
Part of the beauty of subject of Bely\u{\i}\ maps is the ability to pass seamlessly
between combinatorics, group theory, algebraic geometry, topology, and complex
analysis: indeed, one can define categories in these domains that are all
equivalent. In the remainder of this subsection, we make these
categories and equivalences precise; the main result is Proposition
\ref{prop:cateq1} below.
To begin, we record the ramification data, or more precisely the monodromy. A \textsf{permutation triple} of
\textsf{degree} $d$ is a triple $\sigma=(\sigma_0 , \sigma_1 ,
\sigma_{\infty}) \in S_d^3$ such that $\sigma_0 \sigma_1 \sigma_{\infty} = 1$.
Let $\sigma'=(\sigma'_0 , \sigma'_1 , \sigma'_{\infty})$ be another such triple
of degree $d'$. Then a \textsf{morphism} of permutation triples from $\sigma$ to
$\sigma'$ is a map $t : \left\{ 1 , \dots , d \right\} \to \left\{ 1 , \dots ,
d' \right\}$ such that $t(\sigma_0(x))=\sigma_0'(t(x))$ for all $x \in S$ and
the same for $\sigma_1,\sigma_\infty$. In particular, two permutation triples
$\sigma,\sigma'$ are isomorphic, and we write $\sigma \sim \sigma'$ and say they
are \textsf{simultaneously conjugate}, if and only if they have the same degree
$d=d'$ and there exists a $\tau \in S_d$ such that
\begin{align*}
\sigma^\tau = \tau^{-1}(\sigma_0,\sigma_1,\sigma_{\infty})\tau
= (\tau^{-1}\sigma_0\tau,\tau^{-1}\sigma_1\tau,\tau^{-1}\sigma_{\infty}\tau)
= (\sigma'_0,\sigma'_1,\sigma'_{\infty}).
\end{align*}
It is a consequence of the Riemann existence theorem that the category of
Bely\u{\i}\ maps is equivalent to the category of permutation triples. More
precisely, let
\begin{equation} \label{eqn:F2}
F_2 = \langle x,y,z \mid xyz = 1 \rangle
\end{equation}
be the free group on two generators. Given a group $G$, a \textsf{finite
$G$-set} is a homomorphism $\alpha:G \to \Sym(S)$ on a finite set $S$, and a
\textsf{morphism} between finite $G$-sets from $\alpha$ to $\alpha'$ is a map of
sets $t : S \to S'$ such that $\alpha'(g)(t(x))=t(\alpha(g)(x))$ for all $g \in
G$ and $x \in S$. We see that giving a permutation triple is the same as giving
a finite $F_2$-set, by mapping $x,y,z \in F_2$ to $\sigma_0 , \sigma_1 ,
\sigma_{\infty}$, and that two permutation triples are isomorphic if and only if
the corresponding $F_2$-sets are isomorphic.
Returning to covers and topological considerations, we have an isomorphism
\begin{align*}
F_2 \cong \pi_1(\mathbb P^1 \setminus \{0,1,\infty\}) ;
\end{align*}
the generators $x,y,z$ chosen above can be taken to be simple counterclockwise
loops around $0,1,\infty$. We abbreviate $\mathbb P^1_*=\mathbb P^1 \setminus
\{0,1,\infty\}$. The category of finite topological covers of $\mathbb P^1_*$ is
equivalent to the category of finite $\pi_1( \mathbb P^1_*)$-sets; to a cover, we
associate one of its fibers, provided with the structure of $\pi_1(
\mathbb P^1_*)$-set defined by path lifting. Therefore, a Bely\u{\i}\ map gives rise to a
cover of $\mathbb P^1_*$ by restriction, and conversely a finite topological cover of
$\mathbb P^1_*$ can be given the structure of Riemann surface by lifting the complex
analytic structure and thereby yields a map from an algebraic curve to $\mathbb P^1$
unramified away from $\{0,1,\infty\}$.
Let $f$ be a Bely\u{\i}\ map, corresponding to a permutation triple $\sigma$. The
corresponding $F_2$-set $\rho : F_2 \rightarrow S_d$ is called the
\textsf{monodromy representation} of $f$, and its image is called the
\textsf{monodromy group} of $f$. The monodromy group, as a subgroup of $S_d$, is
well-defined up to conjugacy and in particular up to isomorphism, and we denote
it by $G=\Mon (f)$. By the correspondences above, the automorphism group of a
Bely\u{\i}\ map is the centralizer of its monodromy group (as a subgroup of $S_d$).
We consider a final category, introduced by Grothendieck \cite{Grothendieck}. A
\textsf{dessin} $D$ is a triple $(\Gamma,C,O)$ where:
\begin{itemize}
\item[(D1)] $\Gamma$ is a finite graph with vertex set $V$, edge set $E$, and
vertex map $v: E \to V \times V$;
\item[(D2)] $C : V \to \left\{ 0,1 \right\}$ is a bicoloring of the vertices
such that the two vertices of an edge are colored differently, i.e., $C( v
(e)) = \left\{ 0,1 \right\}$ (and not a proper subset) for all edges $e \in
E$; and
\item[(D3)] $O$ is a cyclic orientation of the edges around every vertex.
\end{itemize}
Due to the presence of the bicoloring $C$, the cyclic orientation in (D3) is
specified by two permutations $O_0 , O_1 \in \Sym (E)$ specifying the orderings
around the edges marked with $0$ and $1$ respectively. Note that once the
bicoloring $C$ is given, the possible orientations $O = (O_0 , O_1)$ can be
chosen to be any pair of permutations with the property that two edges $e,e'$
are in the same orbit under $O_0$ (resp.\ $O_1$) if and only if the
corresponding vertices marked $0$ (resp.\ $1$) coincide. A \textsf{morphism} of
dessins is a morphism of graphs $\varphi:\Gamma \to \Gamma'$ such that $\varphi$
takes the bicoloring $C$ to $C'$ (i.e., $C'(\varphi(v))=C(v)$) and similarly the
cyclic orientation $O$ to $O'$.
The category of dessins is also equivalent to that of Bely\u{\i}\ maps. Indeed,
associated to a Bely\u{\i}\ map $f$ is the graph given by $f^{-1}([0,1])$, with the
bicoloring on the vertices given by $f$ and with the cyclic ordering induced by
the orientation on the Riemann surface. Conversely, given a dessin we can
algebraize the topological covering induced by sewing on $2$-cells as specified
by the ordering $O$.
Dessins were introduced by Grothendieck \cite{Grothendieck} to study the action
of $\Gal (\overline{\mathbb Q}/\mathbb Q)$ on Bely\u{\i}\ maps through combinatorics. So far, progress has
been slow, but we mention one charming result \cite{Schneps}; the Galois action
is already faithful on the dessins that are \textsf{trees} (as graphs).
We summarize the equivalences obtained in the following proposition and refer to
Lenstra \cite{Lenstra} for further exposition and references.
\begin{prop}\label{prop:cateq1}
The following categories are equivalent:
\begin{enumroman}
\item Bely\u{\i}\ maps;
\item permutation triples;
\item finite $F_2$-sets; and
\item dessins.
\end{enumroman}
\end{prop}
In particular, the equivalence in Proposition \ref{prop:cateq1} yields the key
bijection considered in this paper:
\begin{equation} \label{eq:corresp}
\begin{gathered}
\left\{ \text{permutation triples $\sigma = (\sigma_0, \sigma_1,
\sigma_{\infty}) \in S_d^3$} \right\} /
\sim
\\ \updownarrow \text{\small{1:1}} \\
\left\{ \text{Bely\u{\i}\ maps $f:X \to \mathbb P^1$ of degree $d$} \right\} /
\cong_{\overline{\mathbb Q}}
\end{gathered}
\end{equation}
where the notions of isomorphism are taken in the appropriate categories.
Concretely, under the correspondence \eqref{eq:corresp}, the cycles of the
permutation $\sigma_0$ (resp.\ $\sigma_1,\sigma_{\infty}$) correspond to the
points of $X$ above $0$ (resp.\ $1,\infty$) and the length of the cycle
corresponds to the ramification index of the corresponding point under the morphism $f$. Note in
particular that because the first set of equivalence classes in
\eqref{eq:corresp} is evidently finite, there are only finitely many
$\overline{\mathbb Q}$-isomorphism classes of curves $X$ with a Bely\u{\i}\ map of given degree.
It is often useful, and certainly more intuitive, to consider the subcategories
in Proposition \ref{prop:cateq1} that correspond to Bely\u{\i}\ maps $f : X \to
\mathbb P^1$ whose source is connected (and accordingly, we say the map is
\textsf{connected}). A Bely\u{\i}\ map is connected if and only if the corresponding
permutation triple $\sigma$ is \textsf{transitive}, i.e., the subgroup $\langle
\sigma_0,\sigma_1,\sigma_{\infty} \rangle$ is a transitive group. Restricting
to transitive permutations gives a further equivalent category of finite index
subgroups of $F_2$: the objects are subgroups $H \leq F_2$ of finite index and
morphisms $H \to H'$ are restrictions of inner automorphisms of $F_2$ that map
$H$ to $H'$. The category of finite index subgroups of $F_2$ is equivalent to
that of finite transitive $F_2$-sets (to a subgroup $H$ of $F_2$, one associates
the $F_2$-set $F_2 / H$). Proposition \ref{prop:cateq1} now becomes the
following.
\begin{prop}\label{prop:cateq2}
The following categories are equivalent:
\begin{enumerate}
\item[(i)] connected Bely\u{\i}\ maps;
\item[(ii)] transitive permutation triples;
\item[(iii)] transitive finite $F_2$-sets;
\item[(iii$'$)] subgroups of $F_2$ of finite index; and
\item[(iv)] dessins whose underlying graph is connected.
\end{enumerate}
\end{prop}
Unless stated otherwise (e.g., Section \ref{sec:fomfod}), in the rest of this
article we will assume without further mention that a Bely\u{\i}\ map is
\emph{connected}; this is no loss of generality, since any disconnected Bely\u{\i}\
map is the disjoint union of its connected components.
\subsection*{Geometric properties and invariants}
Let $f:X \to \mathbb P^1$ be a (connected) Bely\u{\i}\ map over $\overline{\mathbb Q}$. If the cover $f$
is \textsf{Galois}, which is to say that the corresponding extension of function
fields $\overline{\mathbb Q} (X) / \overline{\mathbb Q} (\mathbb P^1)$ is Galois, then we call $f$ a \textsf{Galois
Bely\u{\i}\ map}. More geometrically, this property boils down to the demand that a
subgroup of $\Aut (X)$ act transitively on the sheets of the cover; and
combinatorially, this is nothing but saying that $\Mon (f) \subseteq S_d$ has
cardinality $\#\Mon(f)=d$. Indeed, the monodromy group of a Bely\u{\i}\ map can
also be characterized as the Galois group of its \textsf{Galois closure}, which is
the smallest Galois cover of which it is a quotient.
The genus of $X$ can be calculated by using the Riemann--Hurwitz formula. If we
define the \textsf{excess} $e(\tau)$ of a cycle $\tau \in S_d$ to be its length
minus one, and the excess $e(\sigma)$ of a permutation to be the sum of the
excesses of its constituent disjoint cycles (also known as the \textsf{index} of
the permutation, equal to $n$ minus the number of orbits), then the genus of a
Bely\u{\i}\ map of degree $n$ with monodromy $\sigma$ is
\begin{equation} \label{eq:genus}
g = 1-n + \frac{e(\sigma_0) + e(\sigma_1) + e(\sigma_{\infty})}{2}.
\end{equation}
In particular, we see that the genus of Bely\u{\i}\ map is zero if and only if
$e(\sigma_0) + e(\sigma_1) + e(\sigma_{\infty}) = 2n - 2$.
We employ exponential notation to specify both ramification types and conjugacy
classes in $S_d$. So for example, if $d = 10$, then $3^2 2^1 1^2$ denotes both
the conjugacy class of the permutation $(1\ 2\ 3)(4\ 5)(6\ 7\ 8)$ and the
corresponding ramification type; two points of ramification index $3$, one of
index $2$, and two (unramified) of index $1$.
The \textsf{passport} of a Bely\u{\i}\ map $f:X \to \mathbb P^1$ is the triple $(g,G,C)$
where $g$ is the genus of $X$ and $G \subseteq S_d$ is the monodromy group of
$f$, and $C = (C_0 , C_1, C_{\infty})$ is the triple of conjugacy classes of
$(\sigma_0 , \sigma_1 , \sigma_{\infty})$ in $S_d$, respectively
\cite[Definition 1.1.7]{LandoZvonkin}. Although the genus of the Bely\u{\i}\ map is
determined by the conjugacy classes by equation \eqref{eq:genus}, we still
include it in the passport for clarity and ease. The \textsf{size} of a passport
$(g,G,C)$ is the number of equivalence classes of triples $\sigma =
(\sigma_0,\sigma_1,\sigma_{\infty})$ such that $\langle \sigma \rangle = G$ and
$\sigma_i \in C_i$ for $i=0,1,\infty$.
We will occasionally need slightly altered notions of passport. The
\textsf{ramification passport} of $f$ is the pair $(g,C)$ with conjugacy classes
in $S_d$. Another version of the passport will be considered in Section
\ref{sec:fomfod}. The passport has the following invariance property
\cite{JonesStreit}.
\begin{thm}\label{thm:galinv}
The passport and the ramification passport of a Bely\u{\i}\ map are invariant
under the action of $\Gal (\overline{\mathbb Q}/\mathbb Q)$.
\end{thm}
One can calculate the set of isomorphism classes of permutation triples with
given passport using the following lemma with $G = S_d$.
\begin{lem}\label{lem:doubcos}
Let $G$ be a group and let $C_0,C_1$ be conjugacy classes in $G$ represented
by $\tau_0,\tau_1 \in G$. Then the map
\begin{align*}
C_G (\tau_0) \backslash G / C_G (\tau_1) &\to \{ (\sigma_0,\sigma_1) :
\sigma_0 \in C_0 , \sigma_1 \in C_1 \} /\! \sim_G \\
C_G(\tau_0) g C_G(\tau_1) &\mapsto (\tau_0, g \tau_1 g^{-1}) \nonumber
\end{align*}
is a bijection, where $C_G (\tau)$ denotes the centralizer of $\tau$ in $G$ and $\sim_G$ denotes
simultaneous conjugation in $G$.
\end{lem}
The virtue of this lemma is that double-coset methods in group theory are quite
efficient; by using this bijection and filtering appropriately \cite[Lemma
1.11]{KMSV}, this allows us to enumerate Bely\u{\i}\ maps with a given passport
relatively quickly up to moderate degree $d$. One can also estimate the size of
a passport using character theory; for more on this, see Section
\ref{sec:fomfod}.
\subsection*{Applications}
Having introduced the basic theory, we now mention some applications of the
explicit computation of Bely\u{\i}\ maps.
We began in the introduction with the motivation to uncover the mysterious
nature of the action of $\Gal(\overline{\mathbb Q}/\mathbb Q)$ on dessins following Grothendieck's
\emph{Esquisse}. Dessins of small degree tend to be determined by their passport
in the sense that the set of dessins with given passport forms a full Galois
orbit. However, even refined notions of passport do not suffice to distinguish
Galois orbits of dessins of high degree in general: a first example was
\textsf{Schneps' flower} \cite[\S IV, Example I]{Schneps}. Some further examples
of distinguishing features of non-full Galois orbits have been found by Wood
\cite{Wood} and Zapponi \cite{Zapponi}, but it remains a challenge to determine
the Galois structure for the set of dessins with given passport. Even statistics
in small degree are not known yet; an important project remains to construct
full libraries of dessins. The original ``flipbook'' of dessins, due to
B\'etr\'ema--P\'er\'e--Zvonkin \cite{BPZ}, contained only dessins that were
plane trees but was already quite influential, and consequently systematic
tabulation promises to be just as inspiring.
Further applications of the explicit study of Bely\u{\i}\ maps have been found in
inverse Galois theory, specifically the regular realization of Galois groups
over small number fields: see the tomes of Matzat \cite{Matzat}, Malle--Matzat
\cite{MalleMatzat}, and Jensen--Ledet--Yui \cite{JensenLedetYui}. Upon
specialization, one obtains Galois number fields with small ramification set:
Roberts \cite{Roberts1,Roberts2,Robertsx}, Malle--Roberts \cite{MalleRoberts},
and Jones--Roberts \cite{JonesRoberts} have used the specialization of
three-point covers to exhibit number fields with small ramification set or root
discriminant. The covering curves obtained are often interesting in their own
right, spurring further investigation in the study of low genus curves (e.g.,
the decomposition of their Jacobian \cite{Paulhus}). Finally, a Bely\u{\i}\ map
$f:\mathbb P^1 \to \mathbb P^1$, after precomposing so that $\{0,1,\infty\} \subseteq
f^{-1}(\{0,1,\infty\})$, is an example of a rigid \textsf{post-critically finite
map}, a map of the sphere all of whose critical points have finite orbits.
(Zvonkin calls these maps \textsf{dynamical Bely\u{\i}\ functions} \cite[\S
6]{Zvonkin}.) These maps are objects of central study in complex dynamics
\cite{Bartholdi,Pilgrim}: one may study the associated Fatou and Julia sets.
Bely\u{\i}\ maps also figure in the study of \textsf{Hall polynomials}, (also called
\textsf{Davenport-Stothers triples}) which are those coprime solutions $X(t),Y(t),Z(t) \in
\mathbb C[t]$ of the equations in polynomials
\begin{align*}
X(t)^3 - Y(t)^2 = Z(t)
\end{align*}
with $\deg(X(t)) = 2 m$, $\deg (Y(t)) = 3 m$ and $\deg (Z(t)) = m + 1$. These
solutions are extremal in the degree of $Z$ and are analogues of \textsf{Hall
triples}, i.e.\ integers $x , y \in \mathbb Z$ for which $|x^3 - y^2| =O(\sqrt{|x|})$.
Hall polynomials have been studied by Watkins \cite{Watkins} and by
Beukers--Stewart \cite{BeukersStewart}; Montanus \cite{Montanus} uses the
link with dessins ($X^3 (t) / Y^2 (t)$ is a Bely\u{\i}\ map) to find a formula for
the number of Hall polynomials of given degree. Hall polynomials also lead to
some good families of classical Hall triples \cite{ElkiesWatkins}, as
the following example illustrates.
\begin{exm}
Taking $m = 5$ above, one obtains the following Hall polynomials due to Birch:
\begin{align*}
X(t) & = \frac{1}{9} (t^{10} +6 t^7 +15 t^4 + 12 t) , \\
Y(t) & =
\frac{1}{54} (2 t^{15} + 18 t^{12} + 72 t^9 + 144 t^6 + 135 t^3 + 27) , \\
Z(t) & = -\frac{1}{108} (3 t^6 + 14 t^3 + 27).
\end{align*}
Choosing $t \equiv 3$ mod $6$, we get some decent Hall triples, notably
\begin{align*}
|384242766^3 - 7531969451458^2| &= 14668 \\
|390620082^3 - 7720258643465^2| &= 14857
\end{align*}
for $t=\pm 9$; remarkably, in both cases the constant factor $|x^3 -
y^2| / \sqrt{|x|}$ is approximately equal to the tiny number $3/4$.
\end{exm}
Bely\u{\i}\ maps also give rise to interesting algebraic surfaces. The Bely\u{\i}\ maps
of genus $0$ and degree $12$ (resp.\ $24$) with ramification indices above $0,1$
all equal to $3,2$ correspond to elliptic fibrations of rational (resp.\ K3)
surfaces with only $4$ (resp.\ $6$) singular fibers; given such a fibration, the
associated Bely\u{\i}\ map is given by taking its $j$-invariant. By work of
Beauville \cite{Beauville} (resp.\ Miranda and Persson \cite{MP}), there are $6$
(resp.\ $112$) possible fiber types for these families. This result comes down
to calculating the number of Bely\u{\i}\ maps of given degrees with specified conjugacy
classes with cycle type $(3, \dots , 3)$ and $(2, \dots , 2)$ for $\sigma_0$ and
$\sigma_1$.
Especially in the degree $24$ case, the explicit calculation of these Bely\u{\i}\ maps is
quite a challenge. By developing clever methods specific to this case, this
calculation was accomplished by Beukers--Montanus \cite{BeukersMontanus}. They
find $191$ Bely\u{\i}\ maps, exceeding the $112$ ramification types determined by Miranda
and Persson: this is an instance of the phenomenon mentioned above, that the
passport may contain more than one Bely\u{\i}\ map, so that to a given ramification
triple there may correspond multiple isomorphism classes of Bely\u{\i}\ maps.
One can also specialize Bely\u{\i}\ maps to obtain $abc$ triples: this connection is
discussed by Elkies \cite{ElkiesABC} and Frankenhuysen \cite{Frankenhuysen} to
show that the $abc$ conjecture implies the theorem of Faltings, and it is also
considered by van Hoeij--Vidunas \cite[Appendix D]{vHV2}.
Modular curves and certain Shimura curves possess a natural Bely\u{\i}\ map.
Indeed, Elkies has computed equations for Shimura curves in many cases using
only the extant structure of a Bely\u{\i}\ map \cite{ElkiesSCC,Elkies237}. Another
such computation was made by Hallouin in \cite{HallouinComputation}, where a
more elaborate argument using Hurwitz spaces of four-point covers is used.
Explicit equations are useful in many contexts, ranging from the resolution of
Diophantine equations to cryptography \cite{Schoof}. Reducing these equations
modulo a prime also yields towers of modular curves that are useful in coding
theory. Over finite fields of square cardinality $q$, work of Ihara
\cite{Ihara} and Tsfasman--Vl\u{a}du\c{t}--Zink \cite{TVZ} shows that modular
curves have enough \textsf{supersingular points} that their total number of
rational points is asymptotic with $(\sqrt{q}-1)g$ as their genus grows; this is
asymptotically optimal by work of Drinfeld--Vl\u{a}du\c{t}
\cite{DrinfeldVladut}. By a construction due to Goppa \cite{Goppa}, one obtains
the asympotically best linear error-correcting codes known over square fields.
But to construct and use these codes we need explicit equations for the curves
involved. A few of these modular towers were constructed by Elkies
\cite{ElkiesModularTowers}. There are extensions to other arithmetic triangle
towers, using the theory of Shimura curves, which give other results over prime
power fields of larger exponent \cite{DucetThesis}. For the cocompact triangle
quotients, the modular covers involved are Bely\u{\i}\ maps, and in fact many
congruence towers are unramified (and cyclic) after a certain point, which makes
them particularly pleasant to work with.
There are also applications of explicit Bely\u{\i}\ maps to algebraic solutions of
differential equations \cite{LarussonSadykov}: as we will see in Section
\ref{sec:mod}, subgroups of finite index of triangle groups correspond to
certain Bely\u{\i}\ maps, and the uniformizing differential equations for these
groups (resp.\ their solutions) can be obtained by pulling back suitable
hypergeometric differential equations (resp.\ hypergeometric functions). Kitaev
\cite{Kitaev} and Vidunas--Kitaev \cite{VK} consider branched covers at $4$
points with all ramification but one occuring above three points (``almost
Bely\u{\i}\ coverings'') and apply this to algebraic Painlev\'e VI functions.
Vidunas--Filipuk \cite{VidFil} classify coverings yielding transformations
relating the classical hypergeometric equation to the Heun differential
equation; these were computed by van Hoeij--Vidunas \cite{vHV2,vHV}.
There are applications to areas farther from number theory. Eyral--Oka
\cite{EyralOka} explicitly use dessins (and their generalizations to covers of
$\mathbb P^1$ branched over more than $3$ points) in their classification of the
fundamental groups of the complement in the projective plane of certain
join-type sextic curves of the form $a \prod_i (X - \alpha_i Z) = b \prod_j (X -
\beta_j Z)$. Boston \cite{Boston} showed how three-point branched covers arise
in control theory, specifically with regards to a certain controller design
equation. Finally, dessins appear in physics in the context of brane tilings
\cite{Hananyetal} and there is a moonshine correspondence between genus $0$
congruence subgroups of $\SL_2(\mathbb Z)$, associated with some special dessins,
and certain representations of sporadic
groups, with connections to gauge theory \cite{HeMcKay1,HeMcKay2,HeMcKayRead}.
\section{Gr\"obner techniques}\label{sec:groeb}
We now begin our description of techniques for computing Bely\u{\i}\ maps. We start
with the one that is most straightforward and easy to implement, involving the
solutions to an explicit set of equations over $\mathbb Q$. For Bely\u{\i}\ maps of small
degree, this method works quite well, and considerable technical effort has made
it work in moderate degree. However, for more complicated Bely\u{\i}\ maps, it will be
necessary to seek out other methods, which will be described in the sections that follow.
\subsection*{Direct calculation} \label{sec:directcalc}
The direct method has been used since the first Bely\u{\i}\ maps were written down,
and in small examples (typically with genus $0$), this technique works well
enough. A large number of authors describe this approach, with some variations
relevant to the particular case of interest. Shabat--Voevodsky \cite{SV} and
Atkin--Swinnerton-Dyer \cite{ASD} were among the first. Birch \cite[Section
4.1]{Birch} computes a table for covers of small degree and genus. Schneps
\cite[III]{Schneps} discusses the case of clean dessins of genus $0$ and trees.
Malle \cite{MalleTPR} computed a field of definition for many Bely\u{\i}\ maps of
small degree and genus $0$ using Gr\"obner methods, with an eye toward
understanding the field of definition of regular realizations of Galois groups
and a remark that such fields of definition also give rise to number fields
ramified over only a few very small primes. Malle--Matzat \cite[\S
I.9]{MalleMatzat} use a direct method to compute several Bely\u{\i}\ maps in the
context of the inverse Galois problem, as an application of rigidity.
Granboulan studied the use of Gr\"obner bases for genus $0$ Bely\u{\i}\ maps in
detail in his Ph.D. thesis \cite{GranboulanThesis}. Elkies \cite{ElkiesSCC} used
this technique to compute equations for Shimura curves. Other authors who have
used this method are Hoshino \cite{Hoshino} (and Hoshino--Nakamura
\cite{HoshinoNakamura}), who computed the non-normal inclusions of triangle
groups (related to the Bely\u{\i}{}-extending maps of Wood \cite{Wood}). Couveignes
\cite[\S 2]{CouveignesTools} also gives a few introductory examples.
We explain how the method works by example in the simplest nontrivial case.
\begin{exm} \label{exm:firstex}
Take the transitive permutation triple $\sigma=((1\ 2), (2\ 3), (1\ 3\ 2))$
from $S_3$, with passport $(0,S_3,(2^1 1^1,2^1 1^1,3^1))$. Since these
permutations generate the full symmetric group $S_3$, the monodromy group of
this Bely\u{\i}\ map is $S_3$. The Riemann--Hurwitz formula \eqref{eq:genus} gives the
genus as
\begin{align*}
g = 1 - 3 + \frac{1}{2}(1 + 1 + 2) = 0.
\end{align*}
So the map $f : X\cong\mathbb P^1 \to \mathbb P^1$ is given by a rational function $f(t)
\in K(t)$ where $K \subset \overline{\mathbb Q}$ is a number field. There are two points
above $0$, of multiplicities $2,1$, the same holds for $1$, and there is a
single point above $\infty$ with multiplicity $3$. The point above $\infty$
is a triple pole of $f(t)$; since it is unique, it is fixed by
$\Gal(\overline{K}/K)$; therefore we take this point also to be $\infty$,
which we are free to do up to automorphisms of $\mathbb P_K^1$, and hence $f(t) \in
K[t]$. Similarly, the ramified points above $0$ and $1$ are also unique, so
we may take them to be $0$ and $1$, respectively. Therefore, we have
\begin{align*}
f(t) = ct^2(t+a)
\end{align*}
for some $a,c \in K \setminus \{0\}$ and
\begin{align*}
f(t)-1 = c(t-1)^2(t+b)
\end{align*}
for some $b \in K \setminus \{0,-1\}$. Combining these equations, we get
\begin{align*}
c t^2(t+a) - 1 = c(t^3+at^2) - 1 = c(t-1)^2(t+b) = c(t^3 + (b-2)t^2 +
(1-2b)t + b)
\end{align*}
and so by comparing coefficients we obtain $b=1/2$, $c=-2$, and $a=-3/2$. In
particular, we see that the map is defined over $K=\mathbb Q$ and is unique up to
$\Aut(\mathbb P_\mathbb Q^1) \cong \PGL_2(\mathbb Q)$. Thus
\begin{align*}
f(t) = -t^2(2t-3) = -2t^3+3t^2, \quad f(t)-1 = -(t-1)^2(2t+1).
\end{align*}
If we relax the requirement that the ramification set be $\{0,1,\infty\}$ and
instead allow $\{0,r,\infty\}$ for some $r \neq 0,\infty$, then the form of
$f$ can be made more pleasing. For example, by taking $f(t)=t^2(t+3)$ and
$r=4$ we obtain $f(t)-4=(t-1)^2(t+2)$.
\end{exm}
It is hopefully clear from this example (see Schneps \cite[Definition
8]{Schneps}) how to set up the corresponding system of equations for a Bely\u{\i}\
map on a curve of genus $g=0$: with variable coefficients, we equate the two
factorizations of a rational map with factorization specified by the cycle types
in the permutations triple $\sigma$. We illustrate this further in the
following example; for a large list of examples of this kind, see Lando--Zvonkin
\cite[Example 2.3.1]{LandoZvonkin}.
\begin{exm} \label{exm:largegrob}
To get a small taste of how complicated the equations defining a passport can get, consider the case
$G=\PGL_2(\mathbb F_7)$ with permutation triple $\sigma=(\sigma_0,\sigma_1,\sigma_{\infty})$ given by
\begin{align*}
\sigma_0 = \begin{pmatrix} -1 & 0 \\ 0 & 1 \end{pmatrix} , \quad
\sigma_1 = \begin{pmatrix} -1 & 1 \\ -1 & 0 \end{pmatrix} , \quad
\sigma_{\infty} = \begin{pmatrix} 0 & 1 \\ 1 & 1 \end{pmatrix}.
\end{align*}
The permutation representation of $G$ acting on the set of $8$ elements
$\mathbb P^1(\mathbb F_7)$ is given by the elements
\begin{align*}
(1\ 6)(2\ 5)(3\ 4), \quad (0\ \infty\ 1)(2\ 4\ 6), \quad (0\ 1\ 4\ 3\ 2\ 5\
6\ \infty).
\end{align*}
The corresponding degree $8$ Bely\u{\i}\ map $f:X \to \mathbb P^1$ has passport
\[ (0,
\PGL_2 (\mathbb F_7), ( 2^3 1^2, 3^2 1^2 , 8^1)). \]
After putting the totally ramified
point at $\infty$, the map $f$ is given by a polynomial $f(t) \in \overline{\mathbb Q}[t]$
such that
\begin{equation}\label{eqn:ff1}
f(t) = c a(t)^2 b(t) \quad \text{ and } \quad f(t) - 1 = c d(t)^3 e(t)
\end{equation}
where $c \in \overline{\mathbb Q}^\times$ and $a(t),b(t),d(t),e(t) \in \overline{\mathbb Q}[t]$ are monic
squarefree polynomials with $\deg a(t)=3$ and $\deg b(t)=\deg d(t)=\deg e(t) =
2$. We write $a(t)=t^3+a_2t^2+a_1t+a_0$, etc.
Equating coefficients in (\ref{eqn:ff1}) we obtain the following system of $8$
vanishing polynomials in $10$ variables:
\begin{align*}
& a_0^2 b_0 c - c d_0^3 e_0, \\
& 2 a_1 a_0 b_0 c + a_0^2 b_1 c - 3 c d_1 d_0^2 e_0 - c d_0^3 e_1, \\
& 2 a_2 a_0 b_0 c + a_1^2 b_0 c + 2 a_1 a_0 b_1 c + a_0^2 c - 3 c d_1^2
d_0 e_0 - 3 c d_1 d_0^2 e_1 - c d_0^3 - 3 c d_0^2 e_0, \\
& 2 a_2 a_1 b_0 c + 2 a_2 a_0 b_1 c + a_1^2 b_1 c + 2 a_1 a_0 c + 2 a_0
b_0 c - c d_1^3 e_0 - 3 c d_1^2 d_0 e_1 - 3 c d_1 d_0^2 \\
& \qquad
- 6 c d_1 d_0 e_0 - 3 c d_0^2 e_1, \\
& a_2^2 b_0 c + 2 a_2 a_1 b_1 c + 2 a_2 a_0 c + a_1^2 c + 2 a_1 b_0 c + 2
a_0 b_1 c - c d_1^3 e_1 - 3 c d_1^2 d_0 - 3 c d_1^2 e_0 \\
& \qquad - 6 c d_1 d_0 e_1 - 3 c d_0^2 - 3 c d_0 e_0, \\
& a_2^2 b_1 c + 2 a_2 a_1 c + 2 a_2 b_0 c + 2 a_1 b_1 c + 2 a_0 c - c
d_1^3 - 3 c d_1^2 e_1 - 6 c d_1 d_0 - 3 c d_1 e_0 - 3 c d_0 e_1, \\
& a_2^2 c + 2 a_2 b_1 c + 2 a_1 c + b_0 c - 3 c d_1^2 - 3 c d_1 e_1 - 3 c
d_0 - c e_0, \\
& 2 a_2 c + b_1 c - 3 c d_1 - c e_1.
\end{align*}
Using a change of variables $t \leftarrow t-r$ with $r \in \overline{\mathbb Q}$ we may
assume that $b_1=0$, so $b_0 \neq 0$. Note that if $f(t) \in K[t]$ is defined
over $K$ then we may take $r \in K$, so we do not unnecessarily increase the
field of definition of the map. Similarly, if $d_1 \neq 0$, then with $t
\leftarrow ut$ and $u \in K^\times$ we may assume $d_1=b_0$; similarly if
$e_1 \neq 0$, then we may take $e_1=b_0$. If $d_1=e_1=0$, then $f(t)=g(t^2)$
is a polynomial in $t^2$, whence $a_0=0$ so $a_1 \neq 0$, and thus we may take
$a_1=b_0$. This gives a total of three cases: (i) $d_1=b_0 \neq 0$, (ii)
$d_1=0$ and $e_1=b_0 \neq 0$, and (iii) $d_1=e_1=0$ and $a_1=b_0 \neq 0$. We
make these substitutions into the equations above, adding $c \neq 0$ and
$b_1=0$ in all cases. Note that the equation $c \neq 0$ can be added
algebraically by introducing a new variable $c'$ and adding the equation $c c'
= 1$.
These equations are complicated enough that they cannot be solved by hand, but
not so complicated that they cannot be solved by a Gr\"obner basis. There are
many good references for the theory of Gr\"obner bases
\cite{AdamsLoustaunau,CLO,CLO2,GreuelPfister,KreuzerRobbiano}.
In the degenerate cases (ii) and (iii) we obtain the unit ideal, which does
not yield any solutions. In the first case, we find two conjugate solutions
defined over $\mathbb Q(\sqrt{2})$. After some simplification, the first of the
solutions becomes
\begin{align*}
f(t) = \bigl(2\sqrt{2}t^3-2(2\sqrt{2}+1)t^2+(-4+7\sqrt{2})t+1\bigr)^2
\bigl(14t^2+6(\sqrt{2}+4)t - 8\sqrt{2}+31\bigr)
\end{align*}
with
\begin{align*}
f(t)-432(4\sqrt{2}-5) = \bigl(2t^2-2\sqrt{2}+1 \bigr)^3 \bigl(14t^2
-8(\sqrt{2}+4)t - 14\sqrt{2}+63 \bigr).
\end{align*}
\end{exm}
The direct method does not give an obvious way to discriminate among Bely\u{\i}\
maps by their monodromy groups, let alone to match up which Galois conjugate
corresponds to which monodromy triple: all covers with a given ramification type
are solutions to the above system of equations.
To set up a similar system of equations in larger genus $g \geq 1$, one can for
example write down a general (singular) plane curve of degree equal to $\deg
\varphi$ and ask that have sufficiently many nodal singularities so that it has
geometric genus $g$; the Bely\u{\i}\ map can then be taken as one of the
coordinates, and similar techniques apply, though many non-solutions will
still be obtained in this way by cancellation of numerator and denominator.
\begin{rmk}
Any explicitly given quasiprojective variety $X$ with a surjective map to the
moduli space $\mathcal{M}_g$ of curves of genus $g$ will suffice for this purpose;
so for those genera $g$ where the moduli space $\mathcal{M}_g$ has a simpler
representation (such as $g\leq 3$), one can use this representation instead.
The authors are not aware of any Bely\u{\i}\ map computed in this way with genus
$g \geq 3$.
\end{rmk}
The direct method can be used to compute the curves $X$ with Bely\u{\i}\ maps of
small degree. The curve $\mathbb P^1$ is the only curve with a Bely\u{\i}\ map of degree
$2$ (the squaring map), and the only other curve that occurs in degree $3$ is
the genus $1$ curve with $j$-invariant $0$ and equation $y^2 = x^3 + 1$, for
which the Bely\u{\i}\ map is given by projecting onto the $y$-coordinate. In degree
$4$, there is the elliptic curve of $j$-invariant 1728 with equation $y^2=x^3-x$
with Bely\u{\i}\ map given by $x^2$ and one other given by the the elliptic curve
$y^2 = 4(2x + 9)(x^2 + 2x + 9)$ and regular function $y + x^2 +4 x + 18$. Both
were described by Birch \cite{Birch}.
In the direction of tabulating the simplest dessins in this way, all clean
dessins (i.e.\ those for which all ramification indices above $1$ are equal to
$2$) with at most $8$ edges were computed by Adrianov et al.\
\cite{Adrianovetal}. Magot--Zvonkin \cite{MagotZvonkin} and
Couveignes--Granboulan \cite{CouveignesGranboulan} computed the genus $0$
Bely\u{\i}\ maps corresponding to the Archimedean solids, including the Platonic
solids, using symmetry and Gr\"obner bases. For a very complete discussion of
trees and Shabat polynomials and troves of examples, see Lando--Zvonkin \cite[\S
2.2]{LandoZvonkin}.
In general, we can see that these Gr\"obner basis techniques will present
significant algorithmic challenges. Even moderately-sized examples, including
all but the first few of genus $1$, do not terminate in a reasonable time. (In
the worst case, Gr\"obner basis methods have running time that is doubly
exponential in the input size, though this can be reduced to singly exponential
for zero-dimensional ideals; see the surveys of Ayad \cite{Ayad} and Mayr
\cite{Mayr}.) One further differentiation trick, which we introduce in the next
section, allows us to compute in a larger range. However, even after this
modification, another obstacle remains: the set of solutions can have
positive-dimensional degenerate components. These components correspond to situations where
roots coincide or there is a common factor and are often called \textsf{parasitic
solutions} \cite{Kreines03,Kreines08}. The set of parasitic solutions have been
analyzed in some cases by van Hoeij--Vidunas \cite[\S 2.1]{vHV}, but they remain
a nuisance in general (as can be seen already in Example \ref{exm:largegrob}
above).
\begin{rmk}\label{rmk:naive1}
Formulated more intrinsically, the naive equations considered in this section
determine a scheme in the coefficient variables that is a naive version of the
Hurwitz schemes that will be mentioned in Section \ref{sec:gens}. Besides
containing degenerate components, this naive scheme is usually very
non-reduced. We will revisit this issue in Remark \ref{rmk:naive2}.
When calculating a Bely\u{\i}\ map $f : \mathbb P^1 \to \mathbb P^1$, one usually fixes points
on the source and the target. As we saw most elaborately when working out
equation (\ref{eqn:ff1}), this reduces the problem of calculating a Bely\u{\i}\
map in genus $0$ to finding the points on an affine scheme. The families of
solutions in which numerator and denominator cancel give rise to some of the
degenerate components mentioned in the previous paragraph.
\end{rmk}
\subsection*{The ASD differentiation trick}
There is a trick, due to Atkin--Swinnerton-Dyer \cite[\S 2.4]{ASD} that uses the
derivative of $f$ to eliminate a large number of the indeterminates (``the
number of unknowns $c$ can be cut in half at once by observing that $dj/d\zeta$
has factors $F_3^2 F_2$''). Couveignes \cite{CouveignesTools} implies that this
trick was known to Fricke; it has apparently been rediscovered many times.
Hempel \cite[\S 3]{Hempel} used differentiation by hand to classify subgroups of
$\SL_2(\mathbb Z)$ of genus $0$ with small torsion and many cusps. Couveignes \cite[\S
2,\S 10]{CouveignesCRF} used this to compute examples in genus $0$ of
\textsf{clean} dessins. Schneps \cite[\S III]{Schneps} used this trick to
describe a general approach in genus $0$. Finally, Vidunas \cite{Vidunas1}
applied the trick to differential equations, and Vidunas--Kitaev \cite{VK}
extended this to covers with $4$ branch points.
\begin{exm}
Again we illustrate the method by an example. Take
\begin{align*}
\sigma=((1\ 2), (2\ 4\ 3), (1\ 2\ 3\ 4))
\end{align*}
with passport $(0,S_4,(2^1 1^2, 3^1 1^1 , 4^1))$. Choosing the points $0$ and
$1$ again to be ramified, this time of degrees $2,3$ above $0,1$ respectively,
and choosing $\infty$ to be the ramified point above $\infty$, we can write
\begin{align*}
f(t) = c t^2(t^2+at+b)
\end{align*}
and
\begin{align*}
f(t)-1 = c(t-1)^3(t+d).
\end{align*}
The trick is now to differentiate these relations, which yields
\begin{align*}
f'(t) = ct\left( 2(t^2+at+b) + t(2t+a)\right) &= c(t-1)^2\left( (t-1) +
3(t+d)\right) \\
t(4t^2+3at+2b) &= (t-1)^2\left(4t+(3d-1)\right).
\end{align*}
By unique factorization, we must have $4t^2+3at+2b=4(t-1)^2$ and $4t =
4t+(3d-1)$, so we instantly get $a=-8/3$, $b=2$, and $d=1/3$. Substituting
back we see that $c=3$, and obtain
\begin{align*}
f(t) = t^2(3t^2-8t+6) = (t-1)^3(3t+1) + 1.
\end{align*}
\end{exm}
More generally, the differentiation trick is an observation on divisors that
extends to higher genus, as used by Elkies \cite{Elkies237} in genus $g=1$.
\begin{lem} \label{lem:belyiram}
Let $f: X \to \mathbb P^1$ be a Bely\u{\i}\ map with ramification type $\sigma$. Let
\begin{align*}
\opdiv f = \sum_P e_P P - \sum_R e_R R
\quad \text{and} \quad
\opdiv(f-1) = \sum_Q e_Q Q - \sum_R e_R R
\end{align*}
be the divisors of $f$ and $f-1$. Then the divisor of the differential $df$
is
\begin{align*}
\opdiv df = \sum_P (e_P-1) P + \sum_Q (e_Q-1) Q - \sum_R (e_R+1) R.
\end{align*}
\end{lem}
\begin{proof}
Let
\begin{align*}
D= \sum_P (e_P-1) P + \sum_Q (e_Q-1) Q - \sum_{R} (e_R+1)R.
\end{align*}
Then $\opdiv df \geq D$ by the Leibniz rule. By Riemann--Hurwitz, we have
\begin{align*}
2g-2=-2n+\sum_{P} (e_P-1) + \sum_{Q} (e_Q-1) + \sum_{R} (e_R-1)
\end{align*}
so
\begin{align*}
\deg(D)=2g-2+2n-2\sum_R e_R = 2g-2
\end{align*}
since $\sum_R e_R = n$. Therefore $\opdiv df$ can have no further zeros.
\end{proof}
Combined with unique factorization, this gives the following general algorithm
in genus $0$. Write
\begin{align*}
f(t) = \frac{p(t)}{q(t)} = 1 + \frac{r(t)}{q(t)}
\end{align*}
for polynomials $p(t),q(t),r(t) \in \overline{\mathbb Q} [t]$. Consider the derivatives
$p'(t),q'(t),r'(t)$ with respect to $t$ and let $p_0(t)=\gcd(p(t),p'(t))$ and
similarly $q_0(t),r_0(t)$. Write
\begin{align*}
P(t) = \frac{p(t)}{p_0(t)} \text{ and } \widetilde{P}(t) =
\frac{p'(t)}{p_0(t)}
\end{align*}
and similarly $Q$, etc. Then by unique factorization, and the fact that $P , Q
, R$ have no common divisor, evaluation of the expressions $p(t) - q(t) = r(t)$
and $p'(t) - q'(t) = r'(t)$ yields that $Q(t) \widetilde{R}(t) -
\widetilde{Q}(t) R(t)$ is a multiple of $p_0(t)$, and similarly $P(t)
\widetilde{R}(t) - \widetilde{P}(t)R(t)$ (resp.\ $P(t) \widetilde{Q}(t) -
\widetilde{P}(t) Q(t)$) is a multiple of $q_0(t)$ (resp.\ $r_0(t)$).
These statements generalize to higher genus, where they translate to inclusions
of divisors; but the usefulness of this for concrete calculations is limited and
do not pass to relations of functions, since the coordinate rings of higher
genus curves are usually not UFDs. Essentially, one has to be in an especially
agreeable situation for a statement on functions to fall out, and usually one
only has a relation on the Jacobian (after taking divisors, as in the lemma
above). A concrete and important situation where a relation involving functions
does occur is considered by Elkies \cite{Elkies237}. The methods in his example
generalize to arbitrary situations where the ramification is uniform (all
ramification indices equal) except at one point of the Bely\u{\i}\ curve: Elkies himself treats the Bely\u{\i}\ maps with
passport $(1, \PSL_2 (\mathbb F_{27}), (3^9 1^1, 2^{14}, 7^4))$.
The differentiation trick does not seem to generalize extraordinarily well to
higher derivatives; we can repeat the procedure above and further differentiate
$p'(t),q'(t),r'(t)$, but experimentally this not seem to make the ideal grow
further than in the first step.
\begin{ques}
Is the ideal obtained by adding all higher order derivatives equal to the one
obtained from just adding equations coming from first order derivatives (in
genus $0$)?
\end{ques}
However, Shabat \cite[Theorem 4.4]{Shabat} does derive some further information
by considering second-order differentials; and Dremov \cite{Dremov}
calculates Bely\u{\i}\ maps using the quadratic differential
\begin{align*}
MP(f) = \frac{df^2}{f(1 - f)}
\end{align*}
for a regular function $f$ and considering the equalities following from the
relation
\[ MP(f^{-1}) = -MP(f) / f. \]
It is not immediately clear from these
paper how to use this strategy in general, though.
\begin{ques}
How generally does the method of considering second-order differentials apply?
\end{ques}
The additional equations coming from the differentiation trick not only speed up
the process of calculating Bely\u{\i}\ maps, but they also tend to give rise to a
Jacobian matrix at a solution that is often of larger rank than the direct
system. This is important when trying to Hensel lift a solution obtained over
$\mathbb C$ or over a finite field, where the non-singularity of the Jacobian involved
is essential. (We discuss these methods in sections that follow.)
\begin{rmk}\label{rmk:naive2}
Phrased in the language of the naive moduli space in Remark \ref{rmk:naive1},
the additional ASD relations partially saturate the corresponding equation
ideal, so that the larger set of equations defines the same set of geometric
points, but with smaller multiplicities. (We thank Bernd Sturmfels for this
remark.) Reducing this multiplicity all the way to $1$ is exactly the same as
giving the Jacobian mentioned above full rank.
\end{rmk}
\begin{exm}
The use of this trick for reducing multiplicities is best illustrated by some
small examples.
The first degree $d$ in which the ASD differentiation trick helps to give the
Jacobian matrix full rank is $d=6$; it occurs for the ramification triples
$(2^3,2^3,3^2)$, $(2^2 1^2, 3^2, 4^1 2^1 )$, $(3^2, 3^1 2^1 1, 3^1 2^1
1^1)$, $(3^1 1^3, 4^1 2^1, 4^1 2^1)$, $(4^1 2^1, 4^1 1^2, 3^1 2^1 1^1)$,
and $(4^1 2^1, 3^1 2^1 1^1, 3^1 2^1 1^1)$, where it reduces the
multiplicity of the corresponding solutions from $9,3,3,3,4,3$ respectively to
$1$. Note the tendency of Bely\u{\i}\ maps with many automorphisms to give rise to
highly singular points, as for curves with many automorphisms in the
corresponding moduli spaces.
On the other hand, there are examples where even adding the ASD relations does
not lead to a matrix of full rank. Such a case is first found in degree $7$;
it corresponds to the ramification triples $(4^1 2^1 1^1, 3^1 2^1 1^2 ,
4^1 3^1)$, and throwing in the ASD relations reduces the multiplicity from $8$
to $2$. Unfortunately, iterating the trick does not make the ideal grow
further in this case.
More dramatically, for the ramification triples $(2^4, 3^2 2^1, 3^2
2^1)$ and $(2^3 1^2, 4^2 , 3^2 2^1)$, differentiation reduces some
multiplicities from $64$ to $1$ (resp.\ $64$ to $4$). In the latter case,
these multiplicities are in fact not determined uniquely by the corresponding
ramification type, so that considering these multiplicities gives a way to
split the solutions into disjoint Galois orbits.
\end{exm}
\begin{ques}
How close is the ideal obtained from the differentiation trick (combined with
the direct method) to being radical? Can one give an upper bound for the
multiplicity of isolated points?
\end{ques}
\subsection*{Further extensions}
There can be several reasons why a Gr\"obner basis calculation fails to
terminate. One problem is coefficient blowup while calculating the elimination
ideals. This can be dealt by first reducing modulo a suitable prime $p$,
calculating a Gr\"obner basis for the system modulo $p$, then lifting the good
solutions (or the Gr\"obner basis itself) $p$-adically, recognizing the
coefficients as rational numbers, and then verifying that the basis over $\mathbb Q$ is
correct. This was used by Malle \cite{Malle15,MalleTrinks} to compute covers
with passports $(0 , \Hol(E_8) , (4^1 2^1 1^2,4^1 2^1 1^2,6^1 2^1))$ and $(0 ,
\PGL(\mathbb F_{11}) , (2^5 1^2,4^3,11^1 1^1))$ and similarly Malle--Matzat
\cite{MalleMatzatPSL2Fp} to compute covers for $(0 ,
\PSL_2(\mathbb F_{11}) , (2^4 1^3, 6^1 3^1 2^1, 6^1 3^1 2^1))$ and $(0 ,
\PSL_2(\mathbb F_{13}) , (2^7, 4^3 1^2, 6^2 1^2))$. This idea was also used by
Vidunas--Kitaev \cite[\S 5]{VK}. For further developments on $p$-adic methods
to compute Gr\"obner bases, see Arnold \cite{Arnold} or Winkler \cite{Winkler}.
One can also lift a solution modulo $p$ directly, and sometimes such solutions
can be obtained relatively quickly without also $p$-adically lifting the
Gr\"obner bases: this is the basic idea presented in Section \ref{sec:padic}.
In the work of van Hoeij--Vidunas \cite{vHV2,vHV} mentioned in Section
\ref{sec:backg}, genus $0$ Bely\u{\i}\ functions are computed by using pullbacks of
the hypergeometric differential equation and their solutions. This method works
well when the order of each ramification point is as large as possible, e.g.,
when the permutations $\sigma_0,\sigma_1,\sigma_{\infty}$ contain (almost) solely
cycles of order $n_0,n_1,n_{\infty}$ say, and only a few cycles of smaller
order. For example, this occurs when the cover is Galois, or slightly weaker,
when it is \textsf{regular}, that is to say, when the permutations
$\sigma_0,\sigma_1,\sigma_\infty$ are a product of disjoint cycles of equal
cardinality.
The method of van Hoeij--Vidunas to calculate a Bely\u{\i}\ map $f : X \to \mathbb P^1$ is
to consider the $n$ \textsf{exceptional} ramification points in $X$ of $f$ whose
ramification orders do not equal the usual orders $a,b,c$. One then equips the
base space $\mathbb P^1$ with the hypergeometric equation whose local exponents at
$0,1,\infty$ equal $a,b,c$. Pulling back the hypergeometric equation by $f$, one
obtains a Fuchsian differential equation with singularities exactly in the $n$
exceptional points. The mere fact that this pullback exists implies equations on
the undetermined coefficients of $f$.
For example, when the number of exceptional points is just $n=3$, the
differential equation can be renormalized to a Gaussian hypergeometric
differential equation, which completely determines it. When $n=4$, one obtains a
form of Heun's equation \cite{MovasatiReiter, vHV2}. Heun's equation depends on
the relative position of the fourth ramification point, as well as on an
\textsf{accessory parameter}; still, there are only two parameters remaining in
the computation.
One shows that for fixed $n$ and genus $g$ (taken as $g=0$ later), there are
only finitely many hyperbolic Bely\u{\i}\ functions with $n$ exceptional points.
For small $n$, van Hoeij and Vidunas show that this differential method is
successful in practice, and they compute all (hyperbolic) examples with $n \leq
4$ (the largest degree of such a Bely\u{\i}\ map was $60$).
\begin{ques}
Are there other sources of equations (such as those arising from differential
equations, algebraic manipulation, etc.)\ that further simplify the
scheme obtained from the direct method?
\end{ques}
\section{Complex analytic methods}\label{sec:coan}
In this section we consider complex analytic methods for finding equations for
Bely\u{\i}\ maps. These methods are essentially approximative; a high precision
solution over $\mathbb C$ is determined, from which one reconstructs an exact
solution over $\overline{\mathbb Q}$.
\subsection*{Newton approximation}
We have seen in the previous section how to write down a system of equations
which give rise to the Bely\u{\i}\ map. These equations can be solved numerically
in $\mathbb C$ using multidimensional Newton iteration, given an approximate solution
that is correct to a sufficient degree of precision and a subset of equations of
full rank whose Jacobian has a good condition number (determinant bounded away
from zero). Then, given a complex approximation that is correct to high
precision, one can then use the LLL lattice-reduction algorithm \cite{LLL} (as
well as other methods, such as PSLQ \cite{PSLQ}) to guess algebraic numbers that
represent the exact values. Finally, one can use the results from Section
\ref{sec:veri} to verify that the guessed cover is correct; if not, one can go
back and iterate to refine the solution.
\begin{rmk}\label{rem:algdep}
We may repeat this computation for each representative of the Galois orbit to
find the full set of conjugates for each putative algebraic number and then
recognize the symmetric functions of these conjugates as rational numbers
using continued fractions instead. For example, one can compute each
representative in the passport, possibly including several Galois orbits. The
use of continued fractions has the potential to significantly reduce the
precision required to recognize the Bely\u{\i}\ map exactly.
\end{rmk}
\begin{exm}
Consider the permutation triple
\begin{align*}
\sigma_0 = (1\ 3\ 2)(4\ 6\ 5), \quad
\sigma_1 = (1\ 5\ 2)(3\ 4)(6\ 7), \quad
\sigma_{\infty}=(1\ 3\ 5\ 2\ 6\ 7\ 4).
\end{align*}
From the Riemann--Hurwitz formula, we find that the associated Bely\u{\i}\ curve
$X$ has genus $g=1$. The ramification point of index $7$ on $X$ (over
$\infty$) is unique, so we take it to be the origin of the group law on $X$.
Moreover, since there is a unique unramified point above $0$, we can use a
normal form (due to Tate) of an elliptic curve with a marked point. This is
given by an equation
\begin{align}\label{eq:M11}
y^2 + p_3 y = q(x) = x^3 + p_2 x^2 + p_4 x
\end{align}
with marked point $(0,0)$. The equation \eqref{eq:M11} is unique up to scaling
the coefficients by $u\neq 0$ according to $(p_2 ,
p_3 , p_4 ) \mapsto (u^2 a_2 , u^3 a_3 , u^4 a_4)$, showing that the moduli
spaces $\mathcal{M}_{1,2}$ of genus $1$ curves with two marked points is isomorphic
to the weighted projective space $\mathbb P (2,3,4)$.
Since the origin of the group law of $X$ maps to $\infty$ and $(0,0)$ maps to $0$,
the Bely\u{\i}\ map $f:X \to \mathbb P^1$ of degree $7$ is of the form
\begin{align*}
f(x,y) = (a_3x^3+a_2x^2+a_1x) + (b_2x^2+b_1x+b_0)y = a(x)+b(x)y.
\end{align*}
The ramification above $f=0$ leads to the equation
\begin{align*}
\N_{\mathbb C (x,y) / \mathbb C (x)} ( f(x,y) ) = a(x) (a(x)- p_3 b(x)) - b(x)^2 q(x) =
-b_2^2 x c(x)^3
\end{align*}
where $c(x)$ is the monic polynomial $x^2+c_1x+c_0$. Consideration of the
ramification above $1$ yields
\begin{align*}
\N_{\mathbb C (x,y) / \mathbb C (x)} ( f(x,y) - 1 ) = (a(x) - 1) (a(x) - 1 - p_3 b(x)) -
b(x)^2 q(x) = -b_2^2 d(x)^3 e(x)^2
\end{align*}
where $d(x)=x+d_0$ and $e(x)=x^2+e_1x+e_0$.
This yields $13$ equations in $14$ unknowns. The reason for this is that we are
still free to scale the $p_i$. Here we have to distinguish cases. We first
suppose that the point $(0,0)$ in (\ref{eq:M11}) is not $2$-torsion, or equivalently,
that $p_3 \neq 0$: this is the ``generic'' case. We can then distinguish two further
cases, namely $p_2 \neq 0$ and $p_4 \neq 0$.
Accordingly, we may then ensure $p_2=p_3$ or $p_3=p_4$ by scaling over the ground
field, so that we do not needlessly
enlarge the coefficient of the Bely\u{\i}\ map. In either case, plugging in random choices
for the vector of unknowns $(a,b,c,d,e,p) \in \mathbb C^{14}$
and applying multivariate Newton iteration fails to yield a solution.
To improve the convergence, we now proceed to remove some degenerate cases from
this set of equations. Applying the trick from Example \ref{exm:largegrob}, we
impose that $c_0 d_0 e_0 \neq 0$, as we may since the ramification points are
distinct and $(0,0)$ is a ramification point.
(This in fact assumes that none
of the other ramification points is $(0,-1)$, which leads to a
subcase that turns out not to yield a solution.)
We further insist that $c$ and $e$ do not
have a double root, so $(c_1^2-c_0)(e_1^2-e_0) \neq 0$. This adds $2$
more variables and equations.
Finally, we saturate our equations using the Atkin--Swinnerton-Dyer trick in
Lemma \ref{lem:belyiram}. The
differential $dx/(2 y + p_3)$ is holomorphic and has no zeros or poles, so
denoting derivation with respect to $x$ by $'$, we see that
\begin{align*}
\frac{df}{dx/(2 y + p_3)} = (2 y + p_3) \frac{df}{dx} & = a'(x) (2 y + p_3)
+ b'(x) (2 y + p_3) y + b(x) (2 y + a_3) y' \\
& = (2 b'(x) q (x) + b (x) q'(x) + p_3 a'(x)) + (2 a'(x) - p_3 b'(x)) y
\end{align*}
satisfies
\begin{align*}
\N(((2 y + p_3) (df/dx)) = 49 b_2^2 c(x)^2d(x)^2e(x) .
\end{align*}
This differentation trick thus yields another $8$ equations. But even after
adding these and the nondegeneracy conditions, random choices for an initial
approximation fail to converge to a solution for the new system of $23$
equations in $15$ unknowns.
So we are led to consider the case where $p_3 = 0$, so that the unramified point above
$0$ is $2$-torsion. (Here, there is some extra ambiguity, since the moduli space $X_0(2)=X_1(2)$
is not a fine moduli space.) If we write
$\opdiv(f)=(0,0)+3P_1+3P_2-7\infty$ and $\opdiv(f-1)=2Q_1+2Q_2+3Q_3$, then we
have
\begin{align*}
\opdiv(df)=2P_1+2P_2+Q_1+Q_2+2Q_3-8\infty
\end{align*}
and so we obtain the relations
\begin{align*}
Q_1+Q_2=3Q_3=0, \quad 3Q_3=0, \quad (0,0)+(P_1+P_2)=-Q_3, \quad 2(0,0)=0
\end{align*}
in the group law of $X$. In particular, $P_1+P_2$ is a $6$-torsion point on
$X$. Relations such as these can be used to find extra equations for $X$ and
$f$ by using division polynomials. But again, the new system fails to yield
any solutions; perhaps one can prove non-existence of solutions directly.
Here, we look ahead to the methods of this section and
Section \ref{sec:mod} that allow us to find an approximation
to the solution. It turns out that we only need $3$ decimal places to get the
Newton method converging to a real solution with $p_3 \neq 0$ and $p_2 = p_3$,
approximated by the solution
\begin{align*}
&(a,b,c,d,e,p) \approx \\
&\quad ( 182.7513294, 146.8290694, 29.38993410, -308.3482399, -244.0552479 \\
&\qquad -48.11742858, 0.7992141684, 0.1613326212, 0.1482181605, 0.9764940118, \\
&\qquad 0.2561882114, 1.165925608, 0.4430649844, 163.2364906, 3.003693522 )
\end{align*}
in $\mathbb C^{13}$. The condition number of the system without the additional
Atkin--Swinnerton-Dyer relations is approximately $3.3 \cdot 10^7$; but by adding
some of these relations, this can be decreased to approximately $1.2 \cdot 10^5$.
Using LLL, we recognize this as a putative
solution over $\mathbb Q (\alpha)$ with $\alpha^3 - 3 \alpha + 12 = 0$; then
we verify that the recognized solution is correct using
the methods of Section \ref{sec:veri}. This solution thereby gives rise
to two more complex (conjugate) solutions. Since there are only three permutation triples
with the given ramification passport, we see that we have found all dessins of
the given ramification type, so we need not consider the other cases further.
As mentioned in Remark \ref{rem:algdep}, the standard algorithms to recognize
algebraic dependency work better after symmetrizing over these conjugate
solutions. For the most difficult algebraic number to recognize (which is
$b_2$) using a single solution requires the knowledge of $161$ digits, whereas
recognition as an algebraic number needs only $76$ digits.
If we drop the demand that the unramified point is at $(0,0)$, then we can
simplify the solution somewhat, as in Section \ref{sec:veri}. In Weierstrass
form, we can take $X$ to be given by the curve
\begin{align*}
y^2 = x^3 & + (-541809 \alpha^2 + 898452 \alpha + 2255040) x \\
& + (-2929526838 \alpha^2 + 5759667648 \alpha - 11423888784) .
\end{align*}
and the function $f = a(x) + b (x) y$ by
\begin{align*}
2^{13} 3^{14} 5^5 a(x) & = (1491 \alpha^2 + 6902 \alpha + 10360) x^3 \\
& + (1410885 \alpha^2 + 2033262 \alpha - 4313736) x^2 \\
& + (731506545 \alpha^2 + 15899218650 \alpha + 32119846920) x \\
& - (7127713852353 \alpha^2 + 3819943520226 \alpha + 62260261739784)
\end{align*}
and
\begin{align*}
2^{13} 3^{16} 5^5 b(x) & = (-197 \alpha^2 - 240 \alpha + 528) x^2 \\
& + (906570 \alpha^2 - 546840 \alpha - 8285760) x \\
& - (715988241 \alpha^2 - 2506621464 \alpha - 1458270864) .
\end{align*}
We thank Marco Streng for his help with reducing these solutions.
Applying the methods in Section \ref{sec:mod} already gives
equations that are better than those in the normalized forms (\ref{eq:M11})
considered above; at least experimentally, using the modular method also tends to
give equations of relatively small height.
\end{exm}
As we have seen in the preceding example,
in order for this procedure to work, one needs a good starting approximation to
the solution. In the non-trivial examples that we have computed so far, it
seems that often this approximation must be given to reasonably high precision
(at least 30 digits for moderately-sized examples) in order for the convergence
to kick in. The required precision seems difficult to estimate from above or
below. And indeed the dynamical system arising from Newton's method has quite
delicate fractal-like properties and its study is a subject in itself
\cite{NewtonDynam}.
\begin{ques}
Is there an explicit sequence of Bely\u{\i}\ maps with the property that the
precision required for Newton iteration to converge tends to infinity?
\end{ques}
One way to find a starting approximation to the solution is explained by
Couveignes--Granboulan
\cite{CouveignesCRF,GranboulanThesis,CouveignesGranboulan}. They inductively
use the solution obtained from a simpler map: roughly speaking, they replace a
point of multiplicity $\nu$ with two points of multiplicities $\nu_1,\nu_2$ with
$\nu_1+\nu_2=\nu$. One can use any appropriate base case for the induction,
such as a map having simple ramification. Couveignes \cite{CouveignesCRF} gives
a detailed treatment of the case of \textsf{trees}, corresponding to clean Bely\u{\i}\
polynomials $f(t)$, i.e.\ those with $f(t)-1 = g(t)^2$: geometrically, this
means that the corresponding dessin can be interpreted as a tree with oriented
edges. In this case, after an application of the differentiation trick, one is
led to solve a system of equations where many equations are linear. See
Granboulan \cite[Chapter IV]{GranboulanThesis} for an example with monodromy
group $\Aut(M_{22})$.
\begin{rmk}
There is a misprint in the example of Couveignes \cite[\S 3, pg.\
8]{CouveignesCRF} concerning the discriminant of the field involved, corrected
by Granboulan \cite[p.\ 64]{GranboulanThesis}.
\end{rmk}
So far, it seems that the inductive numerical method has been limited to genus
$0$ Bely\u{\i}\ maps with special features. A similar method was employed by
Matiyasevich \cite{Matiyasevich} for trees: he recursively transforms the
initial polynomial $2t^n-1$ (corresponding to a star tree) into a polynomial
representing the desired planar tree.
\begin{ques}
Can an inductive complex analytic method be employed to compute more
complicated Bely\u{\i}\ maps in practice?
\end{ques}
In particular, the iterative method by Couveignes and Granboulan to find a good
starting value seems to rely on intuition involving visual considerations; can
these be made algorithmically precise?
\subsection*{Circle packing}
Another complex analytic approach is to use \textsf{circle packing methods}. This
technique was extensively developed in work of Bowers--Stephenson
\cite{BowersStephenson}, with a corresponding Java script \texttt{CirclePack}
available for calculations.
Given a dessin (i.e., the topological data underlying a Bely\u{\i}\ map), one
obtains a triangulation of the underlying surface by taking the inverse image of
$\mathbb P^1 (\mathbb R) \subset \mathbb P^1 (\mathbb C)$ together with the corresponding cell
decomposition. Choosing isomorphisms between these triangle and the standard
equilateral triangle in $\mathbb C$ and gluing appropriately, one recovers the Riemann
surface structure and as a result a meromorphic description of the Bely\u{\i}\ map.
However, the Riemann surface structure is difficult to determine explicitly,
starting from the dessin. As an alternative, one can pass to \textsf{discrete
Bely\u{\i}\ maps} instead. To motivate this construction, note that a Riemann
surface structure on a compact surface induces a unique metric of constant
curvature $1,0,-1$ (according as $g=0,1,\geq 2$) so that one can then speak
meaningfully about circles on such a surface. In particular, it makes sense to
ask whether or not there exists a \textsf{circle packing} associated with the
triangulation, a pattern of circles centered at the vertices of this
triangulation satisfying the tangency condition suggested by the triangulation.
Satisfyingly enough, the \textsf{circle packing theorem}, due
Koebe--Andreev--Thurston \cite{Koebe,MardenRodin,Thurston}, states that given a
triangulation of a topological surface, there exists a unique structure of
Riemann surface that leads to a compatible circle packing. This then realizes
the topological map to the Riemann sphere as a smooth function.
In summary, starting with a dessin, one obtains a triangulation and hence a
circle packing. The corresponding discrete Bely\u{\i}\ map will in general
\emph{not} be meromorphic for the Riemann surface structure induced by the
circle packing; but Bowers and Stephenson prove that it does converge to the
correct solution as the triangulation is iteratively hexagonally refined.
The crucial point is now to compute the discrete approximations obtained by
circle packing in an explicit and efficient way. Fortunately, this is indeed
possible; work by Collins--Stephenson \cite{CollinsStephenson} and Mohar
\cite{Mohar} give algorithms for this. The crucial step is to lift the
configuration of circles to the universal cover $H$ (which is either the sphere
$\mathbb P^1 (\mathbb C)$, the plane $\mathbb C$, or the upper half-plane $\mathcal{H}$) and perform the
calculation in $H$. In fact, this means that the circle packing method also
explicitly solves the \textsf{uniformization problem} for the surface involved;
for theoretical aspects, we refer to Beardon--Stephenson
\cite{BeardonStephenson}. Upon passing to $H$ and using the appropriate
geometry, one then \emph{first} calculates the radii of the circles involved
from the combinatorics, before fitting the result into $H$, where it gives rise
to a fundamental domain for the corresponding curve as a quotient of $H$.
An assortment of examples of the circle packing method is given by
Bowers--Stephenson \cite[\S 5]{BowersStephenson}, and numerical approximations
are computed to a few digits of accuracy. This includes genus $0$ examples of
degree up to $18$, genus $1$ examples of degree up to $24$, and genus $2$
examples of degree up to $14$. For determining the conformal structure, this
approach is therefore much more effective indeed than the naive method from
Section \ref{sec:groeb}. Even better, one can proceed inductively from simpler
dessins by using so-called \textsf{dessin moves} \cite[\S 6.1]{BowersStephenson},
which makes this approach quite suitable for calculating large tables of
conformal realizations of dessins.
On the other hand, there are no theoretical results on the number of refinements
needed to obtain given accuracy for the circle packing method \cite[\S
7]{BowersStephenson}. In examples, it is possible for the insertion of a new
vertex to drastically increase the accuracy needed \cite[Figure
25]{BowersStephenson} and thereby the number of discrete refinements needed,
quite radically increasing the complexity of the calculation \cite[\S
8.2]{BowersStephenson}. However, the method is quite effective in practice,
particularly in genus $0$.
More problematically, it seems difficult to recover equations over $\overline{\mathbb Q}$ for
the Bely\u{\i}\ map from the computed fundamental domain if the genus is strictly
positive. One can compute the periods of the associated Riemann surface to some
accuracy, but one still needs to recover the curve $X$ and transfer the Bely\u{\i}\
map $f$ on $X$ accordingly. Moreover, is also not clear that the accuracy
obtained using this method is enough to jump start Newton iteration and thereby
obtain the high accuracy needed to recognize the map over $\overline{\mathbb Q}$. In Section
\ref{sec:mod}, we circumvent this problem by starting straightaway with an
explicit group $\Gamma$ of isometries of $H$ so that $\Gamma \backslash H \cong
X$ and then finding equations for $X$ by numerically computing modular forms
(i.e., differential forms) on $X$.
\begin{exm}
In Figure \ref{fig:m23}, we give an example from an alternate implementation
by Westbury, which is freely available \cite{Westbury} for the case of
genus $0$. In the figure, an outer polygon is inserted instead of a circle to
simplify the calculation of the radii. We show the conformal triangulation
induced by the second barycentric subdivision of the original triangulation
for one of the exactly $2$ covers in Example \ref{ex:mathest} that descend to
$\mathbb R$.
\begin{equation} \label{fig:m23} \notag
\includegraphics[scale=0.8]{mathieu2.pdf}
\end{equation}
\begin{center}
\textbf{Figure \ref{fig:m23}}: A second subdivision for $M_{23}$ \\
\vspace{2ex}
\end{center}
\addtocounter{equation}{1}
Several more subdivisions would be needed to get the solution close enough to
apply Newton's method.
\end{exm}
\subsection*{Puiseux series}
Couveignes--Granboulan \cite[\S 6]{CouveignesGranboulan} proposed an alternative
method using Puiseux series expansions to get a good complex approximation to
the solution so that again multidimensional Newton iteration can kick off.
At every regular point $P$ in the curve $X$, the Bely\u{\i}\ map has an analytic
expansion as a power series in a uniformizer $z$ at $P$ that converges in a
neighborhood of $P$. Similarly, at a ramification point $P$, there is an
expansion for $f$ that is a Puiseux series in the uniformizer $z$; more
specifically, it is a power series in $z^{1/e}=\exp(2\pi i \log(z)/e)$ where $e$
is the ramification index of $P$ and $\log$ is taken to be the principal
logarithm. Now, these series expansions must agree whenever they overlap, and
these relations between the various expansions give conditions on their
coefficients. More precisely, one chooses tangential base points, called
\textsf{standards}, and the implied symbolic relations are then integrated with
respect to a measure with compact support. Collecting the relations, one obtains
a block matrix, the positioning of whose blocks reflects the topology of the
overlaps of the cover used.
Unfortunately, Couveignes and Granboulan do not give an example of this method
in practice, and the most detail they give concerns iterative ad hoc methods
\cite[\S 7]{CouveignesGranboulan}.
\begin{ques}
How effective is the method of Puiseux series in finding a good starting
approximation? Can one prove rigorously that this method gives a correct
answer to a desired precision?
\end{ques}
\subsection*{Homotopy methods}
One idea that has yet to be explored (to the authors' knowledge) is the use of
techniques from numerical algebraic geometry, such as polyhedral homotopy
methods \cite{Bertini,Verschelde}, to compute Bely\u{\i}\ maps. The success of
homotopy methods in solving extremely large systems of equations, including
those with positive-dimensional components, has been dramatic. In broad
stroke, one deforms the solution of an easier system to the desired ones and
carefully analyzes the behavior of the transition matrix (Jacobian) to ensure
convergence of the final solution. Because these methods are similar in spirit
to the ones above, but applied for a more general purpose, it is natural to
wonder if these ideas can be specialized and then combined into a refined
technique tailored for Bely\u{\i}\ maps.
\begin{ques}
Can the techniques of numerical algebraic geometry be used to compute Bely\u{\i}\
maps efficiently?
\end{ques}
A potential place to start in deforming is suggested by the work above and by
Couveignes \cite[\S 6]{CouveignesTools}: begin with a stable curve (separating
the branch points) and degenerate by bringing together the genus 0 components.
The difficulty then becomes understanding the combinatorial geometry of this
stable curve, which is an active area of research.
\subsection*{Zipper method}
Complex analytic techniques can also be brought to bear on Bely\u{\i}\ maps of
extremely large degree, at least for the case of trees, using an extension of
the \textsf{zipper method} due to Marshall--Rohde \cite{Zipper1,Zipper2}. The
zipper method finds a numerical approximation of the conformal map of the unit
disk onto any Jordan region \cite{Marshallweb}. In its extension, this amount
to solving the Dirichlet problem with boundary for the domain of the exterior of
the desired dessin, which can be done quite simply for trees even with thousands
of branches. For example, Marshall and Rohde have computed the dessins
associated to the Bely\u{\i}\ maps $f^n(z)$ where $f(z)=(3z^3-9z-2)/4$, giving a
sequence of Bely\u{\i}\ trees (under the preimage of $[-2,1]$), and by extension one
can obtain complex approximation to Bely\u{\i}\ maps of extremely large degree:
trees with tens of thousands of edges, far beyond the reach of other methods.
\begin{ques}
Does the zipper method extend to higher genus?
\end{ques}
In the latter extension, one would need to consider not only the convergence of
the Bely\u{\i}\ map but also the associated Bely\u{\i}\ curve $X$, so it appears one will
have to do more than simply solve the Dirichlet problem. See also work by
Larusson and Sadykov \cite{LarussonSadykov}, where the connection with the
classical Riemann-Hilbert problem is discussed in the context of trees.
\section{Modular forms}\label{sec:mod}
In this section we continue with the general strategy of using complex analytic
methods but shift our focus in the direction of geometry and consideration of
the uniformization theorem; we work explicitly with quotients of the upper
half-plane by Fuchsian groups and recast Bely\u{\i}\ maps in this language. This
point of view is already suggested by Grothendieck \cite{Grothendieck}:
\begin{quote}
In more erudite terms, could it be true that every projective non-singular
algebraic curve defined over a number field occurs as a possible ``modular
curve'' parametrising elliptic curves equipped with a suitable rigidification?
\ldots [T]he Soviet mathematician Bely\u{\i}\ announced exactly that result.
\end{quote}
As in the last section, the method here uses numerical approximations; however,
the use of modular functions adds considerable more number-theoretic flavor to
the analytic techniques in the previous section.
\subsection*{Classical modular forms}
Let $F_2$ be the free group on two generators as in \eqref{eqn:F2}. Recall that
the map that considers the permutation action of $x,y,z$ on the cosets of a
subgroup yields a bijection
\begin{equation}\label{eq:freecorresp}
\begin{gathered}
\left\{ \text{transitive permutation triples $\sigma = (\sigma_0, \sigma_1,
\sigma_{\infty}) \in S_d^3$} \right\} /\!\sim
\\
\updownarrow \text{\small{1:1}}
\\
\left\{ \text{subgroups of $F_2$ of index $d$} \right\} /\!\sim\,;
\end{gathered}
\end{equation}
here the equivalence relation on triples is again uniform conjugation, and the
equivalence relation on subgroups is conjugation in $F_2$. In particular, by
Proposition \ref{prop:cateq2}, isomorphism classes of (connected) Bely\u{\i}\ maps
are in bijection with the conjugacy classes of subgroups $F_2$ of finite index.
The key observation is now that $F_2$ can be realized as an arithmetic group,
as follows.
The group $\Gamma(1)=\PSL_2(\mathbb Z)=\SL_2(\mathbb Z)/\{\pm 1\}$ acts on the completed upper
half-plane $\mathcal{H}^*=\mathcal{H} \cup \mathbb P^1(\mathbb Q)$ by linear fractional transformations
\begin{align*}
z \mapsto \frac{az+b}{cz+d}, \quad \text{for}\ \pm \begin{pmatrix} a & b \\ c
& d \end{pmatrix} \in \PSL_2(\mathbb Z).
\end{align*}
The quotient $X(1)=\Gamma(1)\backslash\mathcal{H}^*$ can be given the structure of a
Riemann surface of genus $0$ by the uniformizing map $j: X(1) \xrightarrow{\sim}
\mathbb P^1(\mathbb C)$ (often called the \textsf{modular elliptic $j$-function}),
\begin{align*}
j(q)=\frac{1}{q}+744+196884q+21493760q^2+864299970q^3+\dots
\end{align*}
where $q=\exp(2\pi iz)$.
For an integer $N$, we define the normal subgroup $\Gamma(N)$ as the kernel of
the reduction map $\PSL_2(\mathbb Z) \to \PSL_2(\mathbb Z/N\mathbb Z)$. We will be particularly
interested in the subgroup
\begin{align*}
\Gamma(2)= \left\{ \pm \begin{pmatrix} a & b \\ c & d \end{pmatrix} \in
\PSL_2(\mathbb Z) : b \equiv c \equiv 0 \psmod{2} \right\}
\end{align*}
of index $6$, with quotient isomorphic to $\Gamma(1)/\Gamma(2) \cong
\PSL_2(\mathbb F_2)=\GL_2(\mathbb F_2) \cong S_3$. The group $\Gamma(2)$ is in fact
isomorphic to the free group $F_2 \cong \Gamma(2)$: it is freely generated by
$\pm \begin{pmatrix} 1 & 2 \\ 0 & 1 \end{pmatrix},\pm
\begin{pmatrix} 1 & 0 \\ 2 & 1 \end{pmatrix}$, which act on $\mathcal{H}$ by $z
\mapsto z+2$ and $z \mapsto z/(2z+1)$, respectively; the corresponding
action on the upper half plane is free as well.
The quotient $X(2)=\Gamma(2)\backslash\mathcal{H}^*$ is again a Riemann surface of
genus $0$; the action of $\Gamma(2)$ on $\mathbb P^1(\mathbb Q)$
has three orbits, with representatives $0,1,\infty \in
\mathbb P^1 (\mathbb Q)$. We obtain another uniformizing map $\lambda:X(2)
\xrightarrow{\sim} \mathbb P^1(\mathbb C)$ with expansion
\begin{align*}
\lambda(z)=16q^{1/2} - 128q + 704q^{3/2} - 3072q^2 + 11488q^{5/2} - 38400q^3 +
\dots.
\end{align*}
As a uniformizer for a congruence subgroup of $\PSL_2(\mathbb Z)$, the function
$\lambda (z)$ has a modular interpretation: there is a family of elliptic curves
over $X(2)$ equipped with extra structure. Specifically, given $\lambda \in
\mathbb P^1(\mathbb C) \setminus \{0,1,\infty\}$, the corresponding elliptic curve with extra
structure is given by the \textsf{Legendre curve}
\begin{align*}
E : y^2 = x(x-1)(x-\lambda) ,
\end{align*}
equipped with the isomorphism $(\mathbb Z / 2 \mathbb Z)^2 \xrightarrow{\sim} E[2]$ determined
by sending the standard generators to the $2$-torsion points $(0,0)$ and
$(1,0)$.
There is a forgetful map that forgets this additional torsion structure on a
Legendre curve and remembers only isomorphism class; on the algebraic level,
this corresponds to an expression of $j$ in terms of $\lambda$, which is given
by
\begin{equation} \label{eqn:jlambda}
j(\lambda) = 256\frac{(\lambda^2-\lambda+1)^3}{\lambda^2(\lambda-1)^2};
\end{equation}
indeed, the map $X(2) \to X(1)$ given by $j/1728$ is a Galois Bely\u{\i}\ map of
degree $6$ with monodromy group $S_3$, given explicitly by \eqref{eqn:jlambda}.
This map is the Galois closure of the map computed in Example \ref{exm:firstex}.
The cusp $\infty$ plays a special role in the theory of modular forms, and
marking it in our correspondence will allow a suitable rigidification. With this
modification, the correspondence \eqref{eq:freecorresp} becomes a bijection
\begin{equation}\label{eq:G2corresp}
\begin{gathered}
\left\{
\begin{array}{c}
\text{transitive permutation triples $\sigma \in S_d$}
\\
\text{with a marked cycle of $\sigma_{\infty}$}
\end{array}
\right\} /\!\sim
\\
\updownarrow \text{\small{1:1}}
\\
\left\{ \text{subgroups of $F_2 \cong \Gamma(2)$ of index $d$} \right\}
/\!\sim
\end{gathered}
\end{equation}
with equivalence relations as follows: given $\Gamma , \Gamma' \leq \Gamma(2)$,
we have $\Gamma \sim \Gamma'$ if and only if $g \Gamma g^{-1} = \Gamma'$ for
$g$ an element of the subgroups of translations generated by $z \mapsto z + 2$;
and two triples $\sigma,\sigma' \in S_d^3$ with marked cycles $c,c'$ in
$\sigma_{\infty},\sigma'_{\infty}$ are equivalent if and only if they are
simultaneously conjugate by an element $\tau$ with $\tau c \tau^{-1} = c'$.
It is a marvelous consequence of either of the bijections \eqref{eq:freecorresp}
and \eqref{eq:G2corresp}, combined with Bely\u{\i}{}'s theorem, that any curve $X$
defined over a number field is uniformized by a subgroup $\Gamma \leq \Gamma(2)
< \PSL_2(\mathbb Z)$, so that there is a uniformizing map $\Gamma \backslash \mathcal{H}^*
\xrightarrow{\sim} X(\mathbb C)$. This is the meaning of Grothendieck's comment: the
rigidification here corresponds to the subgroup $\Gamma$. In general, the group
$\Gamma$ is \textsf{noncongruence}, meaning that it does not contain a subgroup
$\Gamma(N)$, so membership in the group cannot be determined by congruences on
the coordinate entries of the matrices. This perspective of modular forms is
taken by Atkin--Swinnerton-Dyer \cite{ASD} and Birch \cite[Theorem 1]{Birch} in
their exposition of this subject: they discuss the relationship between modular
forms, the Atkin--Swinnerton-Dyer congruences for noncongruence modular forms,
and Galois representations in the context of Bely\u{\i}\ maps. For more on the
arithmetic aspects of this subject, we refer to the survey by Li--Long--Yang
\cite{LLY} and the references therein.
The description \eqref{eq:G2corresp} means that one can work quite explicitly
with the Riemann surface associated to a permutation triple. Given a triple
$\sigma$, the uniformizing group $\Gamma$ is given as the stabilizer of $1$ in
the permutation representation $\Gamma(2) \to S_d$ given by $x,y,z \mapsto
\sigma_0,\sigma_1,\sigma_\infty$ as in \eqref{eq:G2corresp}. A fundamental
domain for $\Gamma$ is given by \textsf{Farey symbols} \cite{KurthLong}, including
a reduction algorithm to this domain and a presentation for the group $\Gamma$
together with a solution to the word problem in $\Gamma$. These algorithms have
been implemented in the computer algebra systems \texttt{Sage} \cite{Sage} (in a
package for \emph{arithmetic subgroups defined by permutations}, by Kurth,
Loeffler, and Monien) and \texttt{Magma} \cite{Magma} (by Verrill).
Once the group $\Gamma$ has been computed, and the curve $X=\Gamma \backslash
\mathcal{H}^*$ is thereby described, the Bely\u{\i}\ map is then simply given by the
function
\begin{align*}
\lambda: X \to X(2) \cong \mathbb P^1 ,
\end{align*}
so one immediately obtains an analytic description of Bely\u{\i}\ map. In order to
obtain explicit equations, one needs meromorphic functions on $X$, which is to
say, meromorphic functions on $\mathcal{H}$ that are invariant under $\Gamma$.
We are led to the following definition. Let
$\Gamma \leq \PSL_2(\mathbb Z)$ be a subgroup of finite index.
A \textsf{modular form} for $\Gamma \leq \PSL_2(\mathbb Z)$
of \textsf{weight} $k \in 2\mathbb Z$ is a holomorphic function $f: \mathcal{H} \to \mathbb C$ such that
\begin{equation} \label{eqn:automorphy}
f(\gamma z)=(cz+d)^k f(z) \quad \text{ for all $\gamma=\pm \begin{pmatrix} a &
b \\ c & d \end{pmatrix} \in \Gamma$}
\end{equation}
and such that the limit $\lim_{z \to c} f(z)=f(c)$ exists for all \textsf{cusps}
$c \in \mathbb Q \cup \{\infty\}=\mathbb P^1(\mathbb Q)$ (with the further technical condition that
as $z \to \infty$, we take only those paths that remain in a bounded vertical
strip). A \textsf{cusp form} is a modular form where $f(c)=0$ for each cusp $c$.
The space $S_k(\Gamma)$ of cusp forms for $\Gamma$ of weight $k$ is a
finite-dimensional $\mathbb C$-vector space. If $\Gamma$ is torsion-free or $k=2$,
then there is an isomorphism
\begin{equation} \label{eqn:SKOmega}
\begin{aligned}
S_k(\Gamma) &\xrightarrow{\sim} \Omega^{k/2}(X) \\
f(z) &\mapsto f(z)\,(dz)^{\otimes k/2}
\end{aligned}
\end{equation}
where $\Omega^{k/2}(X)$ is the space of holomorphic differential $(k/2)$-forms
on $X$. In any case, evaluation on a basis for $S_k(\Gamma)$ defines a
holomorphic map $\varphi:X \to \mathbb P^{r-1}$ where $r=\dim_\mathbb C S_k(\Gamma)$,
whenever $r \geq 1$. Classical theory of curves yields a complete description
of the map $\varphi$; for example, for generic $X$ of genus $g \geq 3$, taking
$k=2$ (i.e., a basis of holomorphic $1$-forms) gives a \textsf{canonical
embedding} of $X$ as an algebraic curve of degree $2g-2$ in $\mathbb P^{g-1}$, by the
theorem of Max Noether.
Selander--Str\"ombergsson \cite{SelanderStrombergsson} use this analytic method
of modular forms to compute Bely\u{\i}\ maps; this idea was already present in the
original work of Atkin--Swinnerton-Dyer \cite{ASD} and was developed further by
Hejhal \cite{Hejhal} in the context of Maass forms. Starting with the analytic
description of a subgroup $\Gamma \leq \Gamma(2)$, they compute a hyperelliptic
model of a curve of genus $2$ from the knowledge of the space $S_2(\Gamma)$ of
holomorphic cusp forms of weight 2 for $\Gamma$. These cusp forms are
approximated to a given precision by truncated $q$-expansions
\begin{equation} \label{eqn:fzanqn}
f(z)=\sum_{n=0}^{N} a_n q^n,
\end{equation}
one for each equivalence class of cusp $c$ and corresponding local parameter $q$
under the action of $\Gamma$. These expansions \eqref{eqn:fzanqn} have
undetermined coefficients $a_n \in \mathbb C$, and the equation \eqref{eqn:automorphy}
implies an approximate \emph{linear} condition on these coefficients for any
pair of $\Gamma$-equivalent points $z,z'$. These linear equations can then be
solved using the methods of numerical linear algebra. This seems to work well
in practice, and once complex approximations for the cusp forms are known, the
approximate algebraic equations that they satisfy can be computed, so that after
a further Newton iteration and then lattice reduction one obtains an exact
solution. Atkin--Swinnerton-Dyer say of this method \cite[p. 8]{ASD}:
\begin{quote}
From the viewpoint of numerical analysis, these equations are of course very
ill-conditioned. The power series converge so rapidly that one must be
careful not to take too many terms, and the equality conditions at adjacent
points in a subdivision of the sides are nearly equivalent. However, by
judicious choice of the number of terms in the power series and the number of
subdivision points, for which we can give no universal prescription, we have
been able to determine the first 8 or so coefficients [...]\ with 7
significant figures in many cases.
\end{quote}
\begin{ques}
Does this method give rise to an \emph{algorithm} to compute Bely\u{\i}\ maps? In
particular, is there an explicit estimate on the numerical stability of this
method?
\end{ques}
For Bely\u{\i}\ maps such that the corresponding subgroup $\Gamma$ is congruence,
methods of \textsf{modular symbols} \cite{Cremona,Stein} can be used to determine
the $q$-expansions of modular forms using exact methods. The Galois groups of
congruence covers are all subgroups of $\PGL_2 (\mathbb Z / N \mathbb Z)$ for some integer
$N$, though conversely not all such covers arise in this way; as we will see in
the next subsection, since $\PSL_2 (\mathbb Z)$ has elliptic points of order $2$ and
$3$, a compatibility on the orders of the ramification types is required.
Indeed, ``most'' subgroups of finite index in $\PSL_2(\mathbb Z)$ (in a precise sense)
are noncongruence \cite{JonesStAndrews}.
\begin{exm} \label{exm:noncong}
To give a simple example, we consider one of the two (conjugacy classes of)
noncongruence subgroups of index $7$ of $\PSL_2 (\mathbb Z)$, the smallest possible
index for a noncongruence subgroup by Wohlfarht \cite{Wohlfarht}. The cusp
widths of this subgroup are $1$ and $6$. The information on the cusps tells
us that the ramification type of the Bely\u{\i}\ map above $\infty$ is given by
$(6,1)$, whereas the indices above $0$ (resp.\ $1$) have to divide $3$ (resp.\
$2$). This forces the genus of the dessin to equal $0$, with ramification
triple $(6^1 1^1,3^2 1^1, 2^3 1^1)$.
There are exactly two transitive covers with this ramification type, both with
passport $(0,G,(2^3 1^1,3^2 1^1,6^1 1^1))$. Here the monodromy group $G$ is
the Frobenius group of order $42$; the two covers correspond
to two choices of conjugacy classes of order $6$ in $G$.
For one such choice, we obtain the following unique solution up to conjugacy:
\begin{align*}
\sigma_0=(1\ 2)(3\ 4)(6\ 7), \quad \sigma_1=(1\ 2\ 3)(4\ 5\ 6), \quad
\sigma_{\infty}=(1\ 4\ 7\ 6\ 5\ 3).
\end{align*}
A fundamental domain for the action of $\Gamma=\Gamma_7$ is as follows.
\begin{equation} \label{fig:gamma7} \notag
\includegraphics[scale=1.2]{gamma7-pics.pdf}
\end{equation}
\begin{center}
\begin{tabular}{ll}
\toprule
Label & Coset Representative\\
\midrule
$1$ & $\begin{pmatrix} 1 & 0 \\ 0 & 1 \end{pmatrix}$ \\
$2$ & $\begin{pmatrix} 0 & -1 \\ 1 & 0 \end{pmatrix}$ \\
$3$ & $\begin{pmatrix} 1 & 5 \\ 0 & 1 \end{pmatrix}$ \\
$4$ & $\begin{pmatrix} 1 & 1 \\ 0 & 1 \end{pmatrix}$ \\
$5$ & $\begin{pmatrix} 1 & 4 \\ 0 & 1 \end{pmatrix}$ \\
$6$ & $\begin{pmatrix} 1 & 3 \\ 0 & 1 \end{pmatrix}$ \\
$7$ & $\begin{pmatrix} 1 & 2 \\ 0 & 1 \end{pmatrix}$ \\
\bottomrule
\end{tabular}
\hfill
\begin{tabular}{ll}
\toprule
Label & Side Pairing Element\\
\midrule
$s_{1}$ & $\begin{pmatrix} 1 & 6 \\ 0 & 1 \end{pmatrix}$ \\
$s_{2}$ & $\begin{pmatrix} 1 & 0 \\ 1 & 1 \end{pmatrix}$ \\
$s_{3}$ & $\begin{pmatrix} 5 & -6 \\ 1 & -1 \end{pmatrix}$ \\
$s_{4}$ & $\begin{pmatrix} 3 & -7 \\ 1 & -2 \end{pmatrix}$ \\
$s_{5}$ & $\begin{pmatrix} 4 & -17 \\ 1 & -4 \end{pmatrix}$ \\
\bottomrule
\end{tabular}
\end{center}
\begin{center}
\textbf{Figure \ref{fig:gamma7}}: A fundamental domain and side pairing
for $\Gamma_7 \leq \Gamma(1)$ of index $7$ \\
\vspace{2ex}
\end{center}
\addtocounter{equation}{1}
We put the cusp of $\Gamma(1)$ at $t=\infty$ and the elliptic point of order
$3$ (resp.\ $2$) at $t=0$ (resp.z $t=1$). After this normalization, the $q$-expansion for the
Hauptmodul $t$ for $\Gamma$ is given by
\begin{align*}
t(q)=\frac{1}{\zeta} + 0 + \frac{9+\sqrt{-3}}{2^1 3^4} \zeta +
\frac{-3-5\sqrt{-3}}{2^2 3^5} \zeta^2 + \frac{1-3\sqrt{-3}}{2^1
3^7} \zeta^3 + \dots
\end{align*}
where $\zeta=\eta q^{1/6}$ and
\begin{align*}
\eta^6 = \frac{3^{10}}{7^7}(-1494+3526\sqrt{-3}).
\end{align*}
From this, we compute using linear algebra the algebraic relationship between
$t(q)$ and $j(q)$, expressing $j(q)$ as a rational function in $t(q)$ of
degree $7$:
\begin{align*}
j = -\frac{2^6(1+\sqrt{-3})}{(5-\sqrt{-3})^7}\frac{(54\sqrt{-3}t^2 +
18\sqrt{-3}t+(5-3\sqrt{-3}))^3(6\sqrt{-3}t -
(1+3\sqrt{-3}))}{(6\sqrt{-3}t-(1+3\sqrt{-3}))}.
\end{align*}
We will compute this example again using $p$-adic methods in the next section
(Example \ref{ex:noncong-padic}).
\end{exm}
\subsection*{Modular forms on subgroups of triangle groups}
There is related method that works with a \emph{cocompact} discrete group
$\Gamma \leq \PSL_2(\mathbb R)$, reflecting different features of Bely\u{\i}\ maps.
Instead of taking the free group on two generators, corresponding to the
fundamental group of $\mathbb P^1 \setminus \{0,1,\infty\}$, we instead consider
orbifold covers arising from triangle groups, a subject of classical interest
(see e.g.\ Magnus \cite{Magnus}). For an introduction to triangle groups,
including their relationship to Bely\u{\i}\ maps and dessins, see the surveys of
Wolfart \cite{WolfartDessins,WolfartObvious}.
Let $a,b,c \in \mathbb Z_{\geq 2} \cup \{\infty\}$. We define the \textsf{triangle
group}
\begin{align*}
\Delta(a,b,c) = \langle \delta_0,\delta_1,\delta_{\infty} \mid
\delta_0^a=\delta_1^b=\delta_{\infty}^c=\delta_0\delta_1\delta_{\infty}=1 \rangle
\end{align*}
where infinite exponents $a,b,c$ are ignored in the relations. Let
$\chi(a,b,c)=1/a+1/b+1/c-1\in \mathbb Q$. For example, we have $\Delta(2,3,\infty)
\cong \PSL_2(\mathbb Z)$ and $\Delta(\infty,\infty,\infty) \cong F_2 \cong \Gamma(2)$,
so this construction generalizes the previous section. The triangle group
$\Delta(a,b,c)$ is the index $2$ orientation-preserving subgroup of the group
generated by the reflections in the sides of a triangle $T(a,b,c)$ with angles
$\pi/a,\pi/b,\pi/c$ drawn in the geometry $H$, where $H=\mathbb P^1,\mathbb C,\mathcal{H}$
according as $\chi(a,b,c)$ is positive, zero, or negative.
Associated to a transitive permutation triple $\sigma$ from $S_d$ is a
homomorphism
\begin{align*}
\Delta(a,b,c) &\to S_d \\
\delta_0,\delta_1,\delta_{\infty} &\mapsto \sigma_0,\sigma_1,\sigma_\infty
\end{align*}
where $a,b,c \in \mathbb Z_{\geq 2}$ are the orders of
$\sigma_0,\sigma_1,\sigma_\infty$, respectively. (Here we have no index
$\infty$, so $\Delta(a,b,c)$ is cocompact, which is where this method diverges
from that using classical modular forms.) The stabilizer of a point $\Gamma \leq
\Delta(a,b,c)$ has index $d$, and the above homomorphism is recovered by the
action of $\Delta$ on the cosets of $\Gamma$. The quotient map
\begin{align*}
\varphi: X = \Gamma \backslash H \to \Delta \backslash H
\end{align*}
then realizes the Bely\u{\i}\ map with monodromy $\sigma$, so from this description
we have a way of constructing the Bely\u{\i}\ map associated to $\sigma$. More
precisely, as in \eqref{eq:freecorresp}, the bijection \eqref{eq:corresp}
generalizes to
\begin{equation}\label{eq:tricorresp}
\begin{gathered}
\left\{
\begin{array}{c}
\text{permutation triples $\sigma = (\sigma_0, \sigma_1,
\sigma_{\infty}) \in S_d^3$}
\\
\text{such that $a,b,c$ are multiples of the orders of
$\sigma_0,\sigma_1,\sigma_\infty$}
\end{array}
\right\} /\!\sim
\\
\stackrel{1:1}{\longleftrightarrow}
\\
\left\{ \text{subgroups of $\Delta(a,b,c)$ of index $n$} \right\} /\!\sim ,
\end{gathered}
\end{equation}
where the equivalences are as usual: conjugacy in the group $\Delta (a,b,c)$ and
simultaneous conjugacy of triples $(\sigma_0 , \sigma_1 , \sigma_{\infty})$.
(In particular, these triples are not marked, as by contrast they are in
\eqref{eq:G2corresp}, though certainly our construction could be modified in
this way if so desired.)
Explicitly, one obtains the Riemann surfaces corresponding to a subgroup $\Gamma
<\Delta(a,b,c)$ under the bijection \eqref{eq:tricorresp} by
gluing together triangles $T(a,b,c)$ and making
identifications. This gives a conformally correct way to draw dessins and a
method for computing the covers themselves numerically.
This method has been developed in recent work of Klug--Musty--Schiavone--Voight
\cite{KMSV}. Algorithms are provided for working with the corresponding
triangle group $\Delta$, determining explicitly the associated finite index
subgroup $\Gamma$, and then drawing the dessin on $H$ together with the gluing
relations that define the quotient $X=\Gamma \backslash H$. From this explicit
description of the Riemann surface (or more precisely, Riemann $2$-orbifold) $X$
one obtains equations for the Bely\u{\i}\ map $f$ numerically. The main algorithmic
tool for this purpose is a generalization of Hejhal's method replacing
$q$-expansions with power series expansions, due to Voight--Willis
\cite{VoightWillis}. This method works quite well in practice; as an
application, a Bely\u{\i}\ map of degree $50$ of genus $0$ regularly realizing the
group $\PSU_3(\mathbb F_5)$ over $\mathbb Q(\sqrt{-7})$ is computed.
\begin{exm}
Consider the permutation triple $\sigma = (\sigma_0 , \sigma_1 ,
\sigma_{\infty})$, where
\begin{align*}
\sigma_0 &= (1\ 7\ 4\ 2\ 8\ 5\ 9\ 6\ 3) \\
\sigma_1 &=(1\ 4\ 6\ 2\ 5\ 7\ 9\ 3\ 8) \\
\sigma_\infty &= (1\ 9\ 2)(3\ 4\ 5)(6\ 7\ 8) .
\end{align*}
Then $\sigma_0\sigma_1\sigma_\infty=1$ and these permutations generate a
transitive subgroup \[ G \cong \mathbb Z/3\mathbb Z ~\wr~\mathbb Z/3\mathbb Z \leq S_{9} \] of order $81$ and
give rise to a Bely\u{\i}\ map with passport $(0,G,(9^1,9^1,3^3))$. The corresponding
group $\Gamma \leq \Delta(9,9,3)=\Delta$ of index $9$ arising from
\eqref{eq:tricorresp} has signature $(3;-)$, i.e., the quotient $\Gamma
\backslash \mathcal{H}$ is a (compact) Riemann surface of genus $3$. The map
$X(\Gamma) = \Gamma \backslash \mathcal{H} \to X(\Delta) = \Delta \backslash \mathcal{H}
\cong \mathbb P^1$ gives a Bely\u{\i}\ map of degree $9$, which we now compute.
First, we compute a \textsf{coset graph}, the quotient of the Cayley graph for
$\Delta$ on the generators $\delta_0^{\pm}, \delta_1^{\pm}$ by $\Gamma$ with
vertices labelled with coset representatives $\Gamma \alpha_i$ for $\Gamma
\backslash \Delta$. Given a choice of fundamental domain $D_\Delta$ for
$\Delta$ (a fundamental triangle and its mirror, as above), such a coset graph
yields a fundamental domain $D_\Gamma = \bigcup_{i=1}^n \alpha_i D_\Delta$
equipped with a \textsf{side pairing}, indicating how the resulting Riemann
orbifold is to be glued. We consider this setup in the unit disc $\mathcal{D}$,
identifying $\mathcal{H}$ conformally with $\mathcal{D}$ taking a vertex to the center
$w=0$; the result is Figure \ref{fig:trianglefd}. We obtain in this way a
reduction algorithm that takes a point in $z \in \mathcal{H}$ (or $\mathcal{D}$) and
produces a representative $z' \in D_\Gamma$ and $\gamma \in \Gamma$ such that
$z'=\gamma z$.
We consider the space $S_2(\Gamma)$ of cusp forms of weight $2$ for $\Gamma$,
defined as in \eqref{eqn:automorphy} (but note that since no cusps are present
we can omit the corresponding extra conditions). As in \eqref{eqn:SKOmega}, we
have an isomorphism $S_2(\Gamma) \cong \Omega^1(X)$ of $\mathbb C$-vector spaces with
the space of holomorphic $1$-forms on $X$. Since $X$ has genus $3$, we have
$\dim_\mathbb C S_2(\Gamma)=3$. We compute a basis of forms by considering power
series expansions
\begin{align*}
f(w) = (1-w)^2 \sum_{n=0}^{\infty} b_n w^n
\end{align*}
for $f \in S_2(\Gamma)$ around $w=0$ in the unit disc $\mathcal{D}$. (The presence
of the factor $(1-w)^2$ makes for nicer expansions, as below.) We compute
with precision $\epsilon=10^{-30}$, and so $f(w) \approx (1-w)^2
\sum_{n=0}^{N} b_n w^n$ with $N=815$. We use the Cauchy integral formula to
isolate each coefficient $b_n$, integrating around a circle of radius
$\rho=0.918711$ encircling the fundamental domain. This integral is approximated by summing the evaluations at $O(N)$ points on this circle, which can be explictly represented by elements in the fundamental domain $D_\Gamma$ after using the reduction algorithm.
\begin{equation} \label{fig:trianglefd} \notag
\includegraphics[scale=0.88]{dessin-9,9,3.pdf}
\end{equation}
\begin{center}
\begin{tabular}{ll}
\toprule
Label & Coset Representative\\
\midrule
$1$ & 1 \\
$2$ & $\delta_0^{3}$ \\
$3$ & $\delta_0^{-1}$ \\
$4$ & $\delta_0^{2}$ \\
$5$ & $\delta_0^{-4}$ \\
$6$ & $\delta_0^{-2}$ \\
$7$ & $\delta_0^{}$ \\
$8$ & $\delta_0^{4}$ \\
$9$ & $\delta_0^{-3}$ \\
\bottomrule
\end{tabular}
\hfill
\begin{tabular}{ll}
\toprule
Label & Side Pairing Element\\
\midrule
$s_{1}$ & $\delta_1^{}\delta_0^{-2}$ \\
$s_{2}$ & $\delta_1^{-1}\delta_0^{-4}$ \\
$s_{3}$ & $\delta_0^{}\delta_1^{}\delta_0^{3}$ \\
$s_{4}$ & $\delta_0^{}\delta_1^{-1}\delta_0^{4}$ \\
$s_{5}$ & $\delta_0^{-1}\delta_1^{}\delta_0^{-4}$ \\
$s_{6}$ & $\delta_0^{-1}\delta_1^{-1}\delta_0^{3}$ \\
$s_{7}$ & $\delta_0^{2}\delta_1^{}\delta_0^{2}$ \\
$s_{8}$ & $\delta_0^{-2}\delta_1^{}\delta_0^{-3}$ \\
$s_{9}$ & $\delta_0^{3}\delta_1^{}\delta_0^{4}$ \\
\bottomrule
\end{tabular}
\end{center}
\begin{center}
\textbf{Figure \ref{fig:trianglefd}}: A fundamental domain and side pairing
for $\Gamma \leq \Delta(9,9,3)$ of index $9$ \\
\vspace{2ex}
\end{center}
\addtocounter{equation}{1}
We find the \textsf{echelonized} basis
\begin{align*}
x(w) = (1-w)^2 &\left( 1 - \frac{40}{6!}(\Theta w)^6 +
\frac{3080}{9!}(\Theta w)^9 - \frac{1848000}{12!}(\Theta w)^{12} +
O(w^{15})\right) \\
y(w) = (1-w)^2 &\left( (\Theta w) + \frac{4}{4!}(\Theta w)^4 +
\frac{280}{7!}(\Theta w)^{7} - \frac{19880}{10!}(\Theta w)^{10} +
O(w^{13})\right) \\
z(w) = (1-w)^2 &\left( (\Theta w)^3 - \frac{120}{6!}(\Theta w)^6 -
\frac{10080}{9!}(\Theta w)^{9} - \frac{2698080}{12!}(\Theta w)^{12} +
O(w^{15})\right)
\end{align*}
where $\Theta = 1.73179\ldots + 0.6303208\ldots \sqrt{-1}$. The algebraicity
and near integrality of these coefficients are conjectural \cite{KMSV}, so
this expansion is only numerically correct, to the computed precision.
We now compute the image of the canonical map
\begin{align*}
X(\Gamma) = \Gamma \backslash \mathcal{H} &\to \mathbb P^2 \\
w &\mapsto (x(w):y(w):z(w));
\end{align*}
we find a unique quartic relation
\begin{align*}
216 x^3 z - 216 xy^3 + 36 xz^3 + 144y^3 z - 7z^4 = 0
\end{align*}
so at least numerically the curve $X$ is nonhyperelliptic. Evaluating these
power series at the ramification points, we find that the unique point above
$f=0$ is $(1: 0: 0)$, the point above $f=1$ is $(1/6: 0: 1)$, and the three
points above $f=\infty$ are $(0:1:0)$ and $((-1\pm 3\sqrt{-3})/12 : 0 : 1)$.
The uniformizing map $f:X(\Gamma) \to X(\Delta) \cong \mathbb P^1$ is given by the
reversion of an explicit ratio of hypergeometric functions:
\begin{align*}
f(w)= -\frac{1}{8}(\Theta w)^9 - \frac{11}{1280}(\Theta w)^{18} -
\frac{29543}{66150400}(\Theta w)^{27} + O(w^{36}).
\end{align*}
Using linear algebra, we find the expression for $f$ in terms of $x,y,z$:
\begin{align*}
f(w) = \frac{-27z^3}{216x^3 - 108x^2z + 18xz^2 - 28z^3}.
\end{align*}
Having performed this numerical calculation, we then verify on the curve
$X(\Gamma)$ that this rational function defines a three-point cover with the
above ramification points, as in Section \ref{sec:veri}.
\end{exm}
An important feature of methods using modular forms is that it allows a much
more direct algebraic approach to determining the algebraic structure on the
target Riemann surface. There are no ``parasitic'' solutions to discard, just as
when using the more advanced analytic method of Section \ref{sec:coan}.
Moreover, the equation for the source surface are much easier to find than with
the analytic method, where one typically needs to compute period matrices to
high precision.
\begin{ques}
What are the advantages of the \emph{noncocompact} ($q$-expansions)
and \emph{cocompact} (power series expansions) approaches relative to one another? How far (degree,
genus) can these methods be pushed? Can either of these methods be made rigorous?
\end{ques}
\section{$p$-adic methods}\label{sec:padic}
As an alternative to complex analytic methods, we can use $p$-adic methods to
find a solution; in this section we survey this method, and give a rather
elaborate example of how this works in practice. It is simply the $p$-adic
version of the complex analytic method, with the big distinction that finding a
suitable approximation and then Hensel lifting can be much easier; usually
finding a solution over a finite field suffices to guarantee convergence of
Newton approximation.
\subsection*{Basic idea}
The $p$-adic method begins by finding a solution in a finite field of small
cardinality, typically by exhaustive methods, and then lifts this solution using
$p$-adic Newton iteration. Again, lattice methods can be then employed to
recognize the solution over $\overline{\mathbb Q}$. \emph{Turning the `$p$-adic crank'}, as it
is called, has been a popular method, rediscovered many times and employed in a
number of contexts. Malle \cite{MalleM22} used this method to compute
polynomials with Galois groups $M_{22}$, $\Aut(M_{22})$, and $\PSL_3(\mathbb F_4):2$
over $\mathbb Q$. Elkies \cite{Elkies237} computed a degree $28$ cover $f:X \to \mathbb P^1$
with group $G=\PSL_2(\mathbb F_{27})$ via its action on $\mathbb P^1(\mathbb F_{27})$ modulo $29$,
and other work of Elkies \cite{ElkiesM23}, Watkins \cite{Watkins} and
Elkies--Watkins \cite{ElkiesWatkins} have also successfully used $p$-adic
methods to compute Bely\u{\i}\ maps. Elkin--Siksek \cite{ElkinSiksek} used this
method and tabulated Bely\u{\i}\ maps of small degree. Van Hoeij--Vidunas
\cite{vHV2} used this approach to compute a list of examples whose branching is
nearly regular, before extending the direct method \cite{vHV} as explained in
Section \ref{sec:groeb}. More recently, Bartholdi--Buff--von Bothmer--Kr\"oker
\cite{Bartholdi} computed a Bely\u{\i}\ map in genus $0$ that is of degree $13$ and
which arises in a problem of Cui in dynamical systems; they give a relatively
complete description of each of the steps involved.
A foundational result by Beckmann indicates which primes are primes of good
reduction for the Bely\u{\i}\ map; which primes, therefore, can be used in the
procedure above.
\begin{thm}[Beckmann \cite{Beckmann}]\label{thm:Beckmann}
Let $f:X \to \mathbb P^1$ be a Bely\u{\i}\ map and let $G$ be the monodromy group of $f$.
Suppose that $p \nmid \# G$. Then there exists a number field $L$ such that
$p$ is unramified in $L$ and $f$ is defined over $L$ with good reduction at
all primes $\mathfrak{p}$ of $L$ lying over $p$.
\end{thm}
\begin{rmk}
In fact, Beckmann proves as a consequence that under the hypotheses of the
theorem, the prime $p$ is unramified in the field of moduli $K$ of $f$. (For
the definition of the \textsf{field of moduli}, see Section \ref{sec:fomfod}.)
\end{rmk}
If one works with a pointed cover instead, then the statement of Beckmann's
theorem is simpler \cite[Theorem 3]{Birch}. In the notation of this theorem, if
$p$ divides the order of one of the permutations $\sigma$ then $f$ has bad
reduction at $\mathfrak{p}$ \cite[Theorem 4]{Birch}. But for those $p$ that divide $\#
G$ but not any of the ramification indices, it is much harder to find methods
(beyond explicit calculation) to decide whether or not a model of $f$ with good
reduction over $\mathfrak{p}$ exists. Important work in this direction is due to
Raynaud \cite{Raynaud} and Obus \cite{Obus}.
\begin{ques}
Can one perform a similar lifting procedure by determining solutions modulo
primes where $f$ has bad reduction?
\end{ques}
As the matrix of derivatives of the equations used is almost always of full rank
(see Section \ref{sec:groeb}), the most time-consuming part is usually the
search for a solution over a finite field. In order for this method to be
efficient, one must do better than simply running over the potential solutions
over $\mathbb F_q$. Bartholdi--Buff--van Bothmer--Kr\"oker describe \cite[Algorithm
4.7]{Bartholdi} a more careful method for genus $0$, working directly with
univariate polynomials (and rational functions) with coefficients in $\mathbb F_q$. In
the example below, we show an approach that is similar in spirit to theirs and
that works for hyperelliptic curves as well.
When the field of definition is ``generic'' in some sense, then there is often a
split prime of small norm, so this method is often efficient in practice. The
following question still merits closer investigation.
\begin{ques}
How efficiently can a Bely\u{\i}\ map be computed modulo a prime $p$? How far can
one reduce the dimension of the affine space employed in the enumeration?
\end{ques}
In particular, can a ``partial projection'' (partial Gr\"obner basis) be
computed efficiently to reduce the number of looping variables?
\begin{exm} \label{ex:noncong-padic}
We return to the Bely\u{\i}\ maps with ramification type $(6^1 1^1, 3^2 1^1, 2^3
1^1)$ considered in Example \ref{exm:noncong}.
Theorem \ref{thm:Beckmann} suggests to reduce modulo $5$ first. We put the
ramification type $(6,1)$ over $\infty$ and the corresponding points at
$\infty$ and $0$; we can do this without risking an extension of the field of
definition since these points are unique. In the same way, we put the type
$(3^2,1)$ over $0$ and the single point in this fiber at $1$. This defines a
reasonably small system over $\mathbb F_5$ of dimension $7$, which could even be
checked by enumeration. We get the solutions
\begin{align*}
f(t) = \frac{\alpha^{8} (t - 2)^3 (t + \alpha)^3 (t - 1)}{t}
\end{align*}
and its conjugate, where $\alpha$ is a root of the Conway polynomial defining
$\mathbb F_{5^2}$ over $\mathbb F_5$, i.e., $\alpha^2-\alpha+2=0$. At the prime $13$, we
get two solutions defined over $\mathbb F_{13}$:
\begin{align*}
f(t) = \frac{-3 (t^2 + 3t + 8)^3 (t - 1)}{t}, \quad f(t)= \frac{2 (t^2 +
6)^3 (t - 1)}{t}.
\end{align*}
In both cases, the derivative matrices of the equations (with or without ASD)
are non-singular, so we can lift to the corresponding unramified $p$-adic
fields. After a few iterations of the second pair of solutions, we get
the $13$-adic approximations
\begin{align*}
f(t) &= (-3 -5 \cdot 13 - 13^2 + \dots )(t - 1)t^{-1} \\
&\qquad \cdot (t^2 + (3+8\cdot 13 - 2\cdot 13^2+\dots)t + (8 - 3\cdot 13 -
6\cdot 13^2 + \dots))^3 \\
f(t) &= (2 -3 \cdot 13 + 3 \cdot 13^2 + \dots)(t - 1)t^{-1} \\
&\qquad \cdot (t^2 + (-4 \cdot 13 + 6 \cdot 13^2 + \dots) t + (6 - 3 \cdot
13^2 + \dots))^3.
\end{align*}
We continue, with quadratically growing accuracy, in order to use LLL in the
end. This suggests a pair of solutions over $\mathbb Q (\sqrt{-3})$ given by
\begin{align*}
f(t) = \frac{-1+\sqrt{-3}}{4\sqrt{-3}^3 (\sqrt{-3} + 2)^7} \frac{(162 t^2 +
18 (-\sqrt{-3} - 6) t + (\sqrt{-3} + 3))^3 (t - 1)}{t}
\end{align*}
and its conjugate. One verifies as in Section \ref{sec:veri} that this yields
a solution over $\mathbb Q(\sqrt{-3})$ to the given equations and that they are the
requested Bely\u{\i}\ maps. Though we stop here, one could further simplify the
equation even further by suitable scalar multiplications in $t$, or even
better, the general methods described in Section \ref{sec:veri}.
\end{exm}
\begin{exm}
We now illustrate the complexities involved in employing the above method in an
example. It arose during a study of Galois Bely\u{\i}\ maps with monodromy group
$\PSL_2(\mathbb F_q)$ or $\PGL_2(\mathbb F_q)$, undertaken by Clark--Voight
\cite{ClarkVoight}.
Consider the passport with uniform ramification orders $3,5,6$ and monodromy
group $G = \PSL_2(\mathbb F_{11}) \leq S_{11}$. Here the embedding of $G$ in $S_{11}$
results from its conjugation action on the cosets of its exceptional subgroup $A_5$
(and indeed $\# G /\#A_5 = 660/60 = 11$).
Let $f : E \to \mathbb P^1$ be the degree $11$ Bely\u{\i}\ map defined by the above data,
and let $\varphi: X \to \mathbb P^1$ be its Galois closure, with Galois group $G$. We
anticipate \cite{ClarkVoight} that $\varphi$ with its Galois action is defined
over an at most quadratic extension of $\mathbb Q(\sqrt{3},\sqrt{5})$, in which case by
the Galois correspondence the quotient map $f$ will be defined over the same field. We confirm this
by direct computation.
Using the representation of $G$ above, we find that $f$ has
passport
\[ (1,\PSL_2(\mathbb F_{11}),( 3^3 1^2 , 5^2 1^1 , 6^1 3^1 2^1)); \]
in accordance with the construction above, the ramification orders are
divisors of $3,5,6$, and $E$ has genus $1$.
We distinguish the point of ramification degree $6$ above $\infty$ and obtain a
corresponding group law on $E$. We fix two more points by taking the other
points above $\infty$ (with ramification $3$ and $2$, respectively) to be
$(0,1)$ and $(1,y_1)$. We write the equation
\begin{align*}
y^2 = \pi_3 x^3 + \pi_2 x^2 + (y_1^2-\pi_3-\pi_2-1)x+1 = \pi(x)
\end{align*}
for the curve $E$. The Bely\u{\i}\ function $f$ has the form
\begin{align*}
f (x,y) = \frac{q(x)+r(x)y}{(x-1)^2 x^3}
\end{align*}
where $q(x)=q_8 x^8 +\dots+q_0$ and $r(x)=r_6x^6+\dots+r_0$ have degree $8,6$
respectively and the numerator $f{}_{\textup{num}}(x,y)=q(x)+r(x)y$ vanishes to degree $3$
at $(0,-1)$ and $2$ at $(0,-y_1)$.
By the ramification description above $0$, we must have
\begin{equation}\label{phinum}
\begin{gathered}
\begin{aligned}
\N_{\overline{\mathbb Q} (x,y) / \overline{\mathbb Q} (x)} ( f{}_{\textup{num}}(x,y) ) & = q(x)^2 - r(x)^2 \pi(x) \\
& = q_8^2 x^3 (x-1)^2 s(x)^3 t(x)
\end{aligned}
\end{gathered}
\end{equation}
where $s(x)=x^3+s_2x^2+s_1x+s_0$ and $t(x) = x^2 + t_1 x + t_0$, and similarly
above $1$ we should have
\begin{equation}\label{phinum1}
\begin{gathered}
\begin{aligned}
\N_{\overline{\mathbb Q} (x,y) / \overline{\mathbb Q} (x)} ((f(x,y)-1){}_{\textup{num}}) & = (q(x)-(x-1)^2x^3)^2 -
r(x)^2 \pi(x) \\
& = q_8^2 x^3 (x-1)^2 u(x) v(x)
\end{aligned}
\end{gathered}
\end{equation}
where $u(x) = x^2 + u_1 x + u_0$ and $v(x) = x + v_0$.
An approach using Gr\"obner basis techniques utterly fails here, given the
number of variables involved. This calculation is also made more difficult by
the possibility that other Bely\u{\i}\ covers will intervene: the Mathieu group
$M_{11} \hookrightarrow S_{11}$ also has a $(3,5,6)$ triple of genus $1$, and it
is a priori conceivable that $S_{11}$ occurs as well. Discarding these
parasitic solutions is a nontrivial task until one has already computed all of
them along with the correct ones, just as in Section \ref{sec:groeb}.
As explained above, we search for a solution in a finite field $\mathbb F_q$, lift such
a solution using Hensel's lemma (if it applies), and then attempt to recognize
the solution $p$-adically as an algebraic number using the LLL lattice reduction
algorithm. The primes of smallest norm in the field $\mathbb Q(\sqrt{3},\sqrt{5})$
that are relatively prime to $\# \PSL_2(\mathbb F_{11})$ have norm $q=49,59$, so there
is no hope of simply running over all the $\mathbb F_q$-rational values in the affine
space in $y_1 , \pi , q , r , s , t , u , v$, which is $28$-dimensional.
We speed up the search with a few tricks. Subtracting the two equations
(\ref{phinum})--(\ref{phinum1}), we have
\begin{align*}
q_8^2 s(x)^3 t(x) - 2q(x) + (x-1)^2 x^3 = r_8^2 u(x)^5v(x).
\end{align*}
Comparing coefficients on both sides, by degree we see that the coefficients of
$x^9$ and $x^{10}$ of $s(x)^3t(x)$ and $u(x)^5v(x)$ must agree. So we
precompute a table of the possible polynomials of the form $u(x)^5 v(x)$; there
are $O(q^3)$ such, and we sort them for easy table lookup. Then, for each of
the possible polynomials of the form $s(x)^3 t(x)$, of which there are
$O(q^5)$, we match the above coefficients. Typically there are few matches.
Then for each $q_8^2 \in \mathbb F_q^{\times 2}$, we compute $q(x)$ as
\begin{align*}
q(x) = \frac{1}{2} \left(q_8^2 s(x)^3 t(x) - q_8^2 u(x)^5v(x) - (x-1)^2
x^3\right).
\end{align*}
From equation (\ref{phinum}) we have
\begin{align*}
q(x)^2 - q_8^2(x-1)^2x^3 s(x)^3 t(x) = \pi(x)r(x)^2 ,
\end{align*}
so we compute the polynomial on the right and factor it into squarefree parts.
If the corresponding $\pi(x)$ has degree $3$, then we find $r(x)$ as well,
whence also our solution.
Putting this on a cluster at the Vermont Advanced Computing Center (VACC) using
\textsf{Magma} \cite{Magma}, after a few days we have our answer. We find
several solutions in $\mathbb F_{49}$ but only one solution lifts $p$-adically without
additional effort; it turns out the Jacobian of the corresponding system of
equations is not of full rank. After some effort (see also Section
\ref{sec:veri}), we recognize this cover as an $M_{11}$-cover with ramification
$(3,5,6)$, defined over the number field $\mathbb Q(\alpha)$ where
\begin{align*}
\alpha^7 - \alpha^6 - 8\alpha^5 + 21\alpha^4 + 6\alpha^3 - 90\alpha^2 +
60\alpha + 60 = 0.
\end{align*}
We find $62$ solutions in $\mathbb F_{59}$. Note that the $M_{11}$-covers above do not
reappear since there is no prime of norm $59$ in $\mathbb Q(\alpha)$. Only $8$ of
these solutions yield covers with the correct ramification data; our above
conditions are necessary, but not sufficient, as we have only considered the
$x$-coordinates and not the $y$-coordinates. These $8$ covers lift to a single
Galois orbit of curves defined over the field $\mathbb Q(\sqrt{3},\sqrt{5},\sqrt{b})$
where
\begin{align*}
b=4\sqrt{3} + \frac{11+\sqrt{5}}{2};
\end{align*}
with $N(b)=11^2$; more elegantly, the extension of $\mathbb Q(\sqrt{3},\sqrt{5})$ is
given by a root $\beta$ of the equation
\begin{align*}
T^2 - \frac{1+\sqrt{5}}{2}T - (\sqrt{3}+1) = 0.
\end{align*}
The elliptic curve $E$ has minimal model
\begin{align*}
&y^2 + ((\textstyle{\frac{1}{2}}(13\sqrt{5} + 33)\sqrt{3} +
\textstyle{\frac{1}{2}}(25\sqrt{5} + 65))\beta +
(\textstyle{\frac{1}{2}}(15\sqrt{5} + 37)\sqrt{3} + (12\sqrt{5} + 30)))xy \\
&\quad + (((8\sqrt{5} + 15)\sqrt{3} + \textstyle{\frac{1}{2}}(31\sqrt{5} +
59))\beta + (\textstyle{\frac{1}{2}}(13\sqrt{5} + 47)\sqrt{3} +
\textstyle{\frac{1}{2}}(21\sqrt{5} + 77)))y \\
&\quad = x^3 + ((\textstyle{\frac{1}{2}}(5\sqrt{5} + 7)\sqrt{3} +
\textstyle{\frac{1}{2}}(11\sqrt{5} + 19))\beta +
(\textstyle{\frac{1}{2}}(3\sqrt{5} + 17)\sqrt{3} + (2\sqrt{5} + 15)))x^2 \\
&\quad + ((\textstyle{\frac{1}{2}}(20828483\sqrt{5} + 46584927)\sqrt{3} +
\textstyle{\frac{1}{2}}(36075985\sqrt{5} + 80687449))\beta \\
&\qquad +(\textstyle{\frac{1}{2}}(21480319\sqrt{5} + 48017585)\sqrt{3} +
\textstyle{\frac{1}{2}}(37205009\sqrt{5} + 83168909)))x \\
&\quad + (((43904530993\sqrt{5} + 98173054995)\sqrt{3} +
\textstyle{\frac{1}{2}}(152089756713\sqrt{5} + 340081438345))\beta \\
&\qquad +((45275857298\sqrt{5} + 101240533364)\sqrt{3} + (78420085205\sqrt{5}
+ 175353747591))).
\end{align*}
The $j$-invariant of $E$ generates the field $\mathbb Q(\sqrt{3},\sqrt{5},\beta)$, so
this is its minimal field of definition. This confirms that $\varphi:X \to
\mathbb P^1$ as a $G$-cover is defined over an at most quadratic extension of
$\mathbb Q(\sqrt{3},\sqrt{5})$ contained in the ray class field of conductor
$11\infty$, as predicted by the results of Clark--Voight \cite{ClarkVoight}.
\end{exm}
\section{Galois Bely\u{\i}\ maps}\label{sec:galois}
In this short section we sketch some approaches for calculating Galois Bely\u{\i}\
maps, i.e., those Bely\u{\i}\ maps $f:X \to \mathbb P^1$ corresponding to Galois
extensions of function fields. The flavor of these computations is completely
different from those in the other sections, as the representation-theoretic
properties of the Galois group involved are used heavily. In light of the
Galois correspondence, all Bely\u{\i}\ maps are essentially known once the Galois
Bely\u{\i}\ maps are known; however, the growth in degree between the degree of the
Bely\u{\i}\ map and that of its Galois closure makes it very difficult in general to
make this remark a feasible approach to computing general Bely\u{\i}\ maps. We
therefore consider the subject only in itself, and even here we limit ourselves
to the general idea: exploiting representations and finding invariant functions.
The Galois Bely\u{\i}\ maps in genus $0$ correspond to the regular solids, and can
be computed using the direct method (see the end of Section
\ref{sec:directcalc}). The most difficult case, that of the icosahedron, was
calculated first by Klein \cite{Klein}. The Galois Bely\u{\i}\ maps in genus $1$
only occur on curves with CM by either $\mathbb Q(\sqrt{-3})$ or $\mathbb Q(\sqrt{-1})$, and
can therefore be calculated by using explicit formulas for isogenies; see work
of Singerman--Syddall \cite{SingermanSyddall}.
So it remains to consider the case of genus $\geq 2$, where Bely\u{\i}\ maps are
related with hyperbolic triangle groups (see Section \ref{sec:mod}). In genus
$\geq 2$, Wolfart \cite{WolfartObvious} has shown that Galois Bely\u{\i}\ maps can
be identified with quotient maps of curves \textsf{with many automorphisms}, that
is, those curves that do not allow nontrivial deformations that leave the
automorphism group intact and whose automorphism group therefore defines a
zero-dimensional subscheme of the moduli space of curves $\mathcal{M}_g$ of genus $g
\geq 2$. Wolfart \cite{WolfartCM} compares these Bely\u{\i}\ maps with the related
phenomenon of \textsf{Jacobians of CM type}, which define zero-dimensional
subschemes of the moduli space of principally polarized abelian varieties
$\mathcal{A}_g$. In particular, the CM factors of the Jacobians of the Galois
Bely\u{\i}\ curves are essentially known; they come from Fermat curves \cite[\S
4]{WolfartCM}.
A fundamental technique for proving these theorems is to determine the
representation of the automorphism group on the space of differentials, first
considered by Chevalley and Weil \cite{ChevalleyWeil}; this is
elaborated by Berry--Tretkoff \cite{BT} and Streit \cite{StreitHBC}. Once
this is done, one typically recovers the curve by determining the shape of its
canonical embedding, often an intersection of quadrics. (When the canonical
embedding is not injective, the situation is even simpler; since the
hyperelliptic involution is central in the automorphism group, this reduce to
the calculations in genus $0$ mentioned above.) The particular form of the
equations is then determined by being fixed under the action of the automorphism
group, which acts by linear transformations.
\begin{ques}
Can the representation of the automorphism group $G$ on the space of
differentials be used to give a rigorous algorithm for the computation of
$G$-Galois Bely\u{\i}\ maps (with a bound on the running time)?
\end{ques}
Put another way, computing a Galois Bely\u{\i}\ map amounts to determining
$G$-invariant polynomials of a given degree; in some cases, there is a unique
such polynomial with given degree and number of variables, and so it can be
found without any computation.
\begin{exm}\label{exm:gal1}
We illustrate the invariant theory involved by giving an example of a
calculation of a quotient map $X \to X / \Aut(X) \cong \mathbb P^1$ that turns out
to be a Bely\u{\i}\ map; the example was suggested to us by Elkies.
Consider the genus $9$ curve $X$ defined by the following variant of the Bring
equations:
\begin{align*}
v + w + x + y + z &= 0 , \\
v^2 + w^2 + x^2 + y^2 + z^2 &= 0 , \\
v^4 + w^4 + x^4 + y^4 + z^4 &= 0 .
\end{align*}
This curve is known as \textsf{Fricke's octavic curve}, and it was studied
extensively by Edge \cite{Edge}. There is an obvious linear action by $S_5$
on this curve by permutation of coordinates. To find coordinates on the
quotient $X / S_5$ it therefore suffices to look at the symmetric functions in
the variables $v,w,x,y,z$. We see that the power sums with exponents $1,2,4$
vanish on $X$. Since the ring of invariants function for $S_5$ is generated
by the power sums of degree at most $5$, this suggests that we cook up a
function from the power sums $p_3$ and $p_5$ of degree $3$ and $5$. These
functions do not have the same (homogeneous) degree; to get a well-defined
function, we consider their quotient $f = (p_3^5 : p_5^3)$ as a morphism from
$f:X \to \mathbb P^1$.
The intersection of the hyperplanes defined by $p_3 = 0$ and $p_5 = 0$ with
$X$ are finite; indeed, this is obvious since the corresponding functions do
not vanish indentically on $X$. By B\'ezout's theorem, these zero loci are of
degrees $24,40$. But whereas in the former case one indeed obtains
$24$ distinct geometric points in the intersection, one obtains only $20$
geometric points in the latter case. This shows that the ramification indices
over $0$ and $\infty$ of the degree $120$ morphism $f$ are $6$ and $5$.
This is in fact already enough to conclude that there is only one other branch
point for $q$. Indeed, the orbifold $X / \Aut (X)$ is uniformized by the upper
half plane $\mathcal{H}$ since the genus $9$ curve $X$ is, so $X / \Aut (X)$ is a
projective line with at least $3$ branch points for the quotient by the action
of $S_5$. On the other hand, the Riemann--Hurwitz formula shows that adding a
single minimal contribution of $2$ outside the contributions $5$ and $6$
already known from $\infty$ and $0$ already makes the genus grow to $9$, so
additional ramification is impossible. The additional branch point of $f$ can
be found by considering the divisor of $d f$ on $X$; this point turns out to
be $- (15/2)^2$. So the morphism $f : X \to \mathbb P^1$ defined by
\begin{align*}
f (v,w,x,y,z) = \frac{-2^2 p_3^5}{15^2 p_5^3} = -
\left(\frac{2}{15}\right)^2 \frac{(v^3 + w^3 + x^3 + y^3 + z^3)^5}{(v^5 +
w^5 + x^5 + y^5 + z^5)^3}
\end{align*}
realizes the quotient $X \to X / S_5$ as a Bely\u{\i}\ map. Moreover, we see that
the Galois action is defined over $\mathbb Q$, since it is given by permuting the
given coordinate functions on $X$.
In fact we have an isomorphism $\Aut(X) \cong S_5$ since $\Aut (X)$ cannot be
bigger than $S_5$; such a proper inclusion would give rise to a Fuchsian group
properly containing the triangle group $\Delta(2,5,6)$, whereas on the other
hand this group is maximal (by work of Takeuchi \cite{Takeuchi}, or more
generally see Singerman \cite{Singerman} or Greenberg \cite[Theorem
3B]{Greenberg}).
We therefore have found a Galois cover realizing $S_5$ with ramification
indices $2,5,6$. It turns out that this is the only such cover up to
isomorphism. Considering the exceptional isomorphism $\PGL_2(\mathbb F_5) \cong S_5$,
we see that our calculation also yields a Galois cover realizing a projective
linear group.
\end{exm}
\section{Field of moduli and field of definition} \label{sec:fomfod}
Considering Grothendieck's original motivation for studying dessins, it is
important to consider the rather delicate issue of fields of definition of
Bely\u{\i}\ maps. In fact this is not only an engaging question on a theoretic
level, but it is also interesting from a practical point of view. Indeed, as we
have seen in our calculations above, it is often necessary to determine
equations for Bely\u{\i}\ maps by recognizing complex numbers as algebraic numbers.
A bound on the degree $K$ is an important part of the input to the LLL algorithm
that is typically used for this. Moreover, having an estimate for the degree of
$K$ is a good indication of how computable a given cover will be---if the
estimate for the size is enormous, we are very unlikely to succeed in practice!
\subsection*{Field of moduli}
For a curve $X$
defined over $\overline{\mathbb Q}$, the \textsf{field of moduli} $M(X)$ of $X$ is the fixed
field of the group $\{\tau \in \Gal(\overline{\mathbb Q} / \mathbb Q) : X^{\tau} \cong X\}$ on
$\overline{\mathbb Q}$, where as before $X^{\tau}$ is the base change of $X$ by the
automorphism $\tau \in \Gal(\overline{\mathbb Q} / \mathbb Q)$ (obtained by applying the automorphism
$\tau$ to the defining equations of an algebraic model of $X$ over $\overline{\mathbb Q}$). One
similarly defines the field of moduli of a Bely\u{\i}\ map: $M(f)$ is the fixed
field of $\{\tau \in \Gal(\overline{\mathbb Q} / \mathbb Q) : f^{\tau} \cong f\}$ with isomorphisms as
defined in Section 1.
Now let $f:X \to \mathbb P^1$ be a Bely\u{\i}\ map with monodromy representation $\sigma :
F_2 \to S_d$ and monodromy group $G$. By Theorem \ref{thm:galinv}, the monodromy
group $G$ of $f$, considered as a conjugacy class of subgroups of $S_d$, is
invariant under the Galois action. Therefore, given $\tau \in \Gal(\overline{\mathbb Q} / \mathbb Q)$,
the conjugated morphism $f^\tau:X^\tau \to \mathbb P^1$ is a Bely\u{\i}\ map, and its
monodromy representation $\sigma^\tau :F_2 \to S_d$, which is well-defined up to
conjugation, can be taken to have image $G$. Because the Galois action
preserves the monodromy group up to conjugation and the ramification indices
\cite{JonesStreit}, the Bely\u{\i}\ map $f^{\tau}$ has the same passport $P$ as
$f$. We therefore get an action of $\Gal (\overline{\mathbb Q} / \mathbb Q)$ on the set $S$ of Bely\u{\i}\ maps
with passport $P$. Since the stabilizer of an element of $S$ under this action
has index at most $\# S$ in $\Gal (\overline{\mathbb Q} / \mathbb Q)$, we get the following result.
\begin{prop}\label{prop:merebound}
Let $f$ be a Bely\u{\i}\ map with passport $P$ and field of moduli $K$. Then the
degree $[K : \mathbb Q]$ is bounded above by the size of $P$.
\end{prop}
As mentioned at the end of Section \ref{sec:backg}, finding better bounds than
in Proposition \ref{prop:merebound} is far from trivial and a subject of ongoing
research. Experimentally, the bound is often an equality for generic
(non-Galois) Bely\u{\i}\ maps.
\subsection*{Rigidified categories}
Working with Galois Bely\u{\i}\ maps and the additional structure coming from their
automorphism group naturally leads one to consider a new, more rigidified
category \cite{DebesDouai}. A \textsf{$G$-Bely\u{\i}\ map} is a pair $(f,i)$, where $f
: X \to \mathbb P^1$ is a Galois Bely\u{\i}\ map and $i : G \xrightarrow{\sim} \Mon (f)$
is an isomorphism of the monodromy group of $f$ with $G$. A \textsf{morphism} of
$G$-Bely\u{\i}\ maps from $(f,i)$ to $(f',i')$ is an isomorphism of Bely\u{\i}\ maps
$h:X \xrightarrow{\sim} X'$ that identifies $i$ with $i'$, i.e., such that
\begin{equation} \label{eqn:higx}
h( i(g) x) = i'(g) h(x) \text{ for all $g \in G$ and $x \in X$}.
\end{equation}
A \textsf{$G$-permutation triple} is a triple of permutations $(\sigma_0 ,
\sigma_1 , \sigma_{\infty})$ in $G$ such that $\sigma_0 \sigma_1 \sigma_{\infty}
= 1$ and such that $\sigma_0 , \sigma_1 , \sigma_{\infty}$ generate $G$. A
\textsf{morphism} of $G$-permutation triples is an isomorphism of permutation
triples induced by simultaneous conjugation by an element in $G$. The main
equivalence is now as follows.
\begin{prop}\label{prop:cateqG}
The following categories are equivalent:
\begin{enumroman}
\item $G$-Bely\u{\i}\ maps;
\item $G$-permutation triples;
\item surjective homomorphisms $F_2 \to G$.
\end{enumroman}
\end{prop}
We leave it to the reader to similarly rigidify the notion of dessins; it will
not be needed in what follows.
We will need a slight weakening of this notion in the following section. A
\textsf{weak $G$-Bely\u{\i}\ map} is a pair $(f,i)$, where $f : X \to \mathbb P^1$ is a
Galois Bely\u{\i}\ map and $i : H \hookrightarrow \Mon (f)$ is an isomorphism of the
monodromy group of $f$ with a subgroup $H$ of $G$. A \textsf{morphism} of weak
$G$-Bely\u{\i}\ maps from $(f,i)$ to $(f',i')$ is an isomorphism of Bely\u{\i}\ maps $h:
X \xrightarrow{\sim} X'$ such that \eqref{eqn:higx} holds up to conjugation,
i.e., such that there exists a $t \in G$ such that $h(i(g) x)= i' (t g t^{-1})
h(x)$ for all $g \in G$ and $x \in X$.
A \textsf{weak $G$-permutation triple} is a triple of permutations $(\sigma_0 ,
\sigma_1 , \sigma_{\infty})$ in $G$ such that $\sigma_0 \sigma_1 \sigma_{\infty}
= 1$. A \textsf{morphism} of weak $G$-permutation triples is an isomorphism of
permutation triples induced by simultaneous conjugation by an element in $G$.
The equivalence of Proposition \ref{prop:cateqG} now generalizes to the
following result.
\begin{prop}\label{prop:cateqGweak}
The following categories are equivalent:
\begin{enumroman}
\item weak $G$-Bely\u{\i}\ maps;
\item weak $G$-permutation triples;
\item homomorphisms $F_2 \to G$.
\end{enumroman}
\end{prop}
The set of Bely\u{\i}\ maps of degree $d$ can be identified with the set of weak
$S_d$-Bely\u{\i}\ maps. In particular, whereas $G$-Bely\u{\i}\ maps are always
connected, weak $G$-Bely\u{\i}\ maps need not be. The absolute Galois group
$\Gal(\overline{\mathbb Q}/\mathbb Q)$ acts on the set of (weak) $G$-Bely\u{\i}\ maps, so we can again
define the field of moduli of these rigidified Bely\u{\i}\ maps.
Having introduced weak $G$-Bely\u{\i}\ maps, it makes sense to consider passports up
to the action of the monodromy group $G \subset S_d$ instead of the full group
$S_d$. We accordingly define the \textsf{refined passport} of a (not necessarily
Galois) Bely\u{\i}\ map $f:X \to \mathbb P^1$ to be the triple $(g,G,C)$, where $g$ is the
genus of $X$, the group $G$ is the monodromy group of $f$, and $C = (C_0 , C_1 ,
C_{\infty})$ are the conjugacy classes of $\sigma_0,\sigma_1,\sigma_{\infty}$ in
$G$, not the conjugacy classes in $S_d$ included in the (usual) passport.
Fried \cite{Fried} shows how the conjugacy classes $C_i$ change under the Galois
action. Let $\sigma \in \Gal (\overline{\mathbb Q} / \mathbb Q)$, let $n = \#G$, and let $\zeta_n
\in \overline{\mathbb Q}$ be a primitive $n$-th root of unity. Then $\sigma$ sends $\zeta_n$ to
$\zeta_n^a$ for some $a \in (\mathbb Z / n \mathbb Z)^{\times}$. We obtain new
conjugacy classes $C_i^{a}$ by raising a representative of
$C_i$ to the $a$th power. Then for any character $\chi$ of $G$ we have
\begin{align}
\sigma ( \chi (C_i) ) = \chi (C_i^{a}) .
\end{align}
Let $\mathbb Q (\chi (C_i))$ be the field generated by the character values of the
conjugacy classes $C_i$. We have $\mathbb Q (\chi (C_i)) = \mathbb Q$ if and only if all conjugacy
classes of $G$ are rational, as for instance in the case $G = S_d$.
\begin{prop}
Let $(f,i)$ be a weak $G$-Bely\u{\i}\ map with refined passport $R$ and field of
moduli $K$ as a weak $G$-Bely\u{\i}\ map. Then the degree $[K : \mathbb Q (\chi (C_i))]$
is bounded above by the size of $R$.
\end{prop}
Calculating in the category of weak $G$-Bely\u{\i}\ maps can be useful even when
considering Bely\u{\i}\ maps without this additional structure. More precisely, this
is useful when using formulas that approximate the size of a passport. To this
end, let $G$ be a finite group and let $C_0 , C_1 , C_{\infty}$ be conjugacy
classes in $G$. Let $S$ be the set of isomorphism classes of weak $G$-Bely\u{\i}\
maps $\sigma = (\sigma_0 , \sigma_1 , \sigma_{\infty})$ with the property that
$\sigma_i \in C_i$ for $i \in \left\{ 0,1,\infty \right\}$. Then a formula that
goes back to Frobenius (see Serre \cite[Theorem 7.2.1]{Serre}) shows that
\begin{align}\label{eq:serre}
\sum_{(f,i) \in S} \frac{1}{\Aut_G (f,i)} = \frac{\#C_0 \#C_1
\# C_{\infty}}{(\#G)^2} \sum_{\chi} \frac{\chi (C_0) \chi (C_1) \chi
(C_{\infty})}{\chi (1)}.
\end{align}
Here the automorphism group $\Aut_G (f,i)$ is the group of automorphisms of
$(f,i)$ as a weak $G$-Bely\u{\i}\ map. The sum on the left of (\ref{eq:serre}) runs
over all weak Bely\u{\i}\ maps with the aforementioned property; in particular, one
often obtains non-transitive solutions that one does not care about in practice.
When working with mere Bely\u{\i}\ maps (without rigidification as a weak
$G$-Bely\u{\i}\ map), it can still be useful to consider the estimate
(\ref{eq:serre}) when the monodromy group of the Bely\u{\i}\ map is question is
included in $G$. We illustrate this by a few concrete examples.
\begin{exm}\label{ex:serreest}
We start by taking $G$ to be a full symmetric group and give the
above-mentioned estimate for the number of genus $0$ Bely\u{\i}\ maps with
ramification passport
\begin{align*}
(0, (3^2 2^3 , 5^1 4^1 2^1 1^1 , 6^1 4^1 2^1 )),
\end{align*}
Before giving it, we calculate the possible permutation triples up to
conjugacy directly using Lemma \ref{lem:doubcos}. This shows that the number
of solutions is $583$, of which $560$ are transitive. The transitive solutions
all have monodromy group $S_{23}$ and hence trivial automorphism group. On the
other hand, the Serre estimate (\ref{eq:serre}) equals $567 \frac{1}{4}$,
which more precisely decomposes as
\begin{align*}
567 \frac{1}{4} = \frac{560}{1} + \frac{1}{1} + \frac{3}{2} +
\frac{19}{4};
\end{align*}
of the 23 nontransitive solutions, there is only one with trivial automorphism
group, whereas there are $3$ (resp.\ $19$) with automorphism group of
cardinality $2$ (resp.\ $4$). For each of the nontransitive solutions, the
associated Bely\u{\i}\ maps are disjoint unions of curves of genus $1$, such as
those corresponding to the products of the genus $1$ Bely\u{\i}\ maps with
ramification types $(2^3,5^1 1^1,6^1)$ (which always have trivial automorphism
group) with those with ramification types $(3^2,4^1 2^1 ,4^1 2^1 )$ (which
have either $1$ or $2$ automorphisms, depending on the solution).
\end{exm}
\begin{exm}\label{ex:mathest}
Another example is the case $(0, H, (4^4 2^2 1^3 , 4^4 2^2 1^3 , 5^4 1^3))$
with $H \leq M_{23}$. We can identify $M_{23}$-conjugacy classes with
$S_{23}$-conjugacy classes for these groups, as the conjugacy classes of
$S_{23}$ do not split upon passing to $M_{23}$.
The calculations are much more rapid working with $M_{23}$ than for the full
group $S_{23}$. We obtain the estimate $909$, which fortunately enough equals
the exact number of solutions because the corresponding subgroups of $M_{23}$
all have trivial centralizer; this is not the case when they are considered as
subgroups of $S_{23}$. Of these many solutions, it turns out that only $104$
are transitive.
As mentioned before, this estimate only gives the number of weak
$M_{23}$-Bely\u{\i}\ maps; accordingly, permutation triples are only considered
isomorphic if they are conjugated by an element of $M_{23}$ rather than
$S_{23}$. However, since $M_{23}$ coincides with its own normalizer in
$S_{23}$, this coincides with the number of solutions under the usual
equivalence. Directly working with the group $M_{23}$ indeed saves a great
deal of computational overhead in this case.
\end{exm}
An explicit (but complicated) formula, using M\"obius inversion to deal with the
disconnected Bely\u{\i}\ maps, was given by Mednykh \cite{Mednykh} this vein; in
fact, his formula can be used to count covers with specified ramification type
of an arbitrary Riemann surface.
Finally, we mention that in the Galois case, the situation sometimes simplifies:
there are criteria \cite{ConderJonesStreitWolfart, JonesStreitWolfart} for
Galois Bely\u{\i}\ maps to have cyclotomic fields of moduli, in which case the
Galois action is described by a simple powering process known as \textsf{Wilson's
operations}. Additionally, Streit--Wolfart \cite{StreitWolfart} have calculated
the field of moduli of an infinite family of Bely\u{\i}\ maps whose Galois group is
a semidirect product $\mathbb Z_p \rtimes \mathbb Z_q$ of cyclic groups of prime order.
\subsection*{Field of moduli versus field of definition}
We have seen that in all the categories of objects over $\overline{\mathbb Q}$ considered so
far (curves, Bely\u{\i}\ maps, etc.)\ there is a field of moduli for the action of
$\Gal(\overline{\mathbb Q}/\mathbb Q)$. Given an object $Y$ of such a category with field of moduli
$M$, it is reasonable to ask whether $Y$ is \textsf{defined over $M$}, i.e., if
there exists an object $Y_M$ in the appropriate category over $M$ that is
isomorphic with $Y$ over $\overline{\mathbb Q}$, in which case $M$ is said to be a \textsf{field
of definition} of $Y$. For example, if $Y = (X,f)$ is a Bely\u{\i}\ map over
$\overline{\mathbb Q}$, this means that there should exist a curve $X_M$ over $M$ and a Bely\u{\i}\
map $f_M : X_M \rightarrow \mathbb P_M^1$ such that $(X,f)$ can be obtained from
$(X_M,f_M)$ by extending scalars to $\overline{\mathbb Q}$.
We first consider the case of curves. Curves of genus at most $1$ are defined
over their field of moduli. But this ceases to be the case for curves of larger
genus in general, as was already observed by Earle \cite{Earle} and Shimura
\cite{Shimura}. The same is true for Bely\u{\i}\ maps and $G$-Bely\u{\i}\ maps. This
issue is a delicate one, and for more information, we refer to work of
Coombes--Harbater \cite{CoombesHarbater}, D\`{e}bes--Ensalem
\cite{DebesEmsalem}, D\`{e}bes--Douai, \cite{DebesDouai}, and K\"ock
\cite{Kock}.
The obstruction can be characterized as a lack of rigidification. For example, a
curve furnished with an embedding into projective space is trivially defined
over its field of moduli (as a projectively embedded curve). Additionally,
marking a point on the source $X$ of a Bely\u{\i}\ map and passing to the
appropriate category, \cite[Theorem 2]{Birch} states that the field of moduli is
a field of definition for the \textsf{pointed Bely\u{\i}\ curve} \cite[Theorem
2]{Birch}; however, this issue seems quite subtle, and in \cite{Kock} only
auxiliary points with trivial stabilizer in $\Aut (X)$ are used. Note that the
more inclusive version of this rigidification (with possibly non-trivial
stabilizer) was considered in Section \ref{sec:mod} (see e.g.\
\eqref{eq:G2corresp}). As mentioned at the beginning of the previous
subsection, this implication can then be applied to give an upper bound on the
degree of the field of definition of a Bely\u{\i}\ map, an important bit of
information needed when for example applying LLL to recognize coefficients
algebraically.
Note that for a Bely\u{\i}\ map $f : X \rightarrow \mathbb P^1$, the curve $X$ may descend
to its field of moduli in the category of curves while $f$ does not descend to
this same field of moduli in the category of Bely\u{\i}\ maps. Indeed, this can be
seen already for the example $X = \mathbb P^1$, as $\Gal(\overline{\mathbb Q}/\mathbb Q)$ acts faithfully on
the set of genus $0$ dessins. In general, the problem requires careful
consideration of obstructions that lie in certain Galois cohomology groups
\cite{DebesDouai}.
\begin{rmk}
Although in general we will have to contend with arbitrarily delicate
automorphism groups, Couveignes \cite{CouveignesAPropos} proved that every
curve defined over a number field $K$ admits a Bely\u{\i}\ map without
automorphisms defined over $K$. This map will then necessarily not be
isomorphic to any of its proper conjugates.
\end{rmk}
On top of all this, a Bely\u{\i}\ map may descend to its field of moduli \textsf{in
the weak sense}, i.e., as a cover of a possibly non-trivial conic ramified above
a Galois-stable set of three points, rather than \textsf{in the strong sense}, as
a cover of $\mathbb P^1 \backslash \left\{ 0,1,\infty \right\}$ (i.e., in the category
of Bely\u{\i}\ maps over the field of moduli). This distinction also measures the
descent obstruction for hyperelliptic curves, as in work of
Lercier--Ritzenthaler--Sijsling \cite{LRS}. For Bely\u{\i}\ maps, a deep study of
this problem in genus $0$, beyond the general theory, was undertaken by
Couveignes \cite[\S\S 4--7]{CouveignesCRF}: he shows that for the \textsf{clean
trees}, those Bely\u{\i}\ maps with a single point over $\infty$ and only
ramification index $2$ over $1$, on the set of which $\Gal (\overline{\mathbb Q} / \mathbb Q)$ acts
faithfully, the field of moduli is always a field of definition in the strong
sense. Moreover, he shows that in genus $0$, the field of moduli is always a
field of definition in the weak sense as long as the automorphism group of the
Bely\u{\i}\ map is not cyclic of even order, and in the strong sense as long as the
automorphism group is not cyclic.
These considerations have practical value in the context of computations. For
example, Couveignes \cite[\S 10]{CouveignesCRF} first exhibits a genus $0$
Bely\u{\i}\ map that descends explicitly to $\mathbb Q$ in the strong sense. Then, due to
the presence of non-trivial automorphisms of this Bely\u{\i}\ map, one can realize
it as a morphism $f : C \to \mathbb P^1$ for infinitely many mutually non-isomorphic
conics $C$ over $\mathbb Q$. And by choosing $C$ appropriately (not isomorphic to
$\mathbb P^1$ over $\mathbb Q$), Couveignes manages to condense his equations from half a
page to a few lines. Further simplification techniques will be considered in
the next section.
We mention some results on the field of moduli as a field of definition that are
most useful for generic ($G$-)Bely\u{\i}\ maps.
\begin{enumerate}
\item If a curve or ($G$-)Bely\u{\i}\ map has trivial automorphism group, then it
can be defined over its field of moduli, by Weil's criterion for descent
\cite{Weil}.
\item If the center of the monodromy group of a Galois Bely\u{\i}\ map is trivial, then
it can be defined over its field of moduli by the main result in the article
by D\`ebes--Douai \cite{DebesDouai}.
\item A $G$-Bely\u{\i}\ map, when considered in the category of Bely\u{\i}\ maps
(without extra structure) is defined over its field of moduli as a Bely\u{\i}\
map \cite{CoombesHarbater}.
\end{enumerate}
To give an impression of the subtleties involved, we further elaborate on
Example \ref{exm:gal1} from the previous section. Along the way, we will
illustrate some of the subtleties that arise when considering fields of moduli.
As we will see, these subtleties correspond with very natural questions on the
level of computation.
\begin{exm}
Since $\Delta(3,5,5)$ is a subgroup of $\Delta(2,5,6)$ of index $2$, we also obtain
from this example a Bely\u{\i}\ map with indices $3,5,5$ for the group $A_5$ by
taking the corresponding quotient. Indeed, ramification can only occur over
the points of order $2$ and $6$, which means that in fact the cover is a
cyclic degree $2$ map of conics ramifying of order $2$ over these points and
under which the point of order $5$ has two preimages. An equation for this
cover (which is a Bely\u{\i}\ map) can now be found by drawing an appropriate
square root of the function $(s_3^5 / s_5^3) + (15/2)^2$ (which indeed
ramifies of order $6$ over $\infty$ and of order $2$ over $0$) and sending the
resulting preimages $\pm 15/2$ of the point of order $5$ to $0$ and $1$,
respectively.
Alternatively, we can calculate as follows. The full ring of invariant
homogeneous polynomials for $A_5$ (acting linearly by permutation of
coordinates) is generated by the power sums
$p_1 , \dots , p_5$ and the \textsf{Vandermonde polynomial}
\begin{align*}
a = (v - w) (v - x) (v - y) (v - z) (w - x) (w - y) (w - z) (x - y) (x - z)
(y - z) .
\end{align*}
One easily determines the expression for $a^2$ in terms of the $p_i$; setting
$p_1 = p_2 = p_4 = 0$, we get the relation
\begin{align*}
a^2 = \frac{4}{45} s_3^5 s_5 + 5 s_5^3 .
\end{align*}
This suggests that to get a function realizing the quotient $X \to X / A_5$,
we take the map $g : X \rightarrow C$, where $C$ is the conic
\begin{align*}
C : 45y^2 = 4x z + 225 z^2
\end{align*}
and $g$ is given by
\begin{align*}
g(v,w,x,y,z) = (s_3^5 : a s_5 : s_5^3) .
\end{align*}
Note that $Q$ admits the rational point $(1:0:0)$.
This result is not as strong as one would like. As we have seen when
calculating the full quotient $f$, the branch points of $g$ of order $5$ on
$C$ satisfy $(x : z) = (0 : 1)$. But the corresponding points are only defined
over $\mathbb Q (\sqrt{5})$, so this is a descent of a Bely\u{\i}\ map in the weak sense.
We explain at the group-theoretical level what other kinds of descent can be
expected.
There are actually two Galois covers with ramification indices $(3,5,5)$ for
$A_5$ up to isomorphism. The other cover is not found as a subcover of $f$;
when composing with the same quadratic map, we instead get a Galois Bely\u{\i}\
map whose Galois group is the direct product of $A_5$ and $\mathbb Z / 2 \mathbb Z$. The
corresponding curve is given by taking the hyperelliptic cover ramified over
the vertices of an icosahedron, leading to the equation
\begin{align*}
t^2 = s^{20} + 228 s^{15} + 494 s^{10} - 228 s^5 + 1 .
\end{align*}
In particular, this means that the Galois orbit of these covers consists of a
single isomorphism class, as their monodromy groups upon composition differ
\cite{Wood}. As mentioned above, an $A_5$-Bely\u{\i}\ map, considered as a mere
Bely\u{\i}\ map, is defined over its field of moduli as a Bely\u{\i}\ map, so our
equations above can be twisted to a Bely\u{\i}\ map over $\mathbb Q$, that is, with
ramification at three rational points.
However, the Galois cover does \emph{not} descend as an $A_5$-Bely\u{\i}\ map (so
in the strong sense, as a Galois cover unramified outside $\left\{ 0,1,\infty
\right\}$). Indeed, the character table of $A_5$ is only defined over $\mathbb Q
(\sqrt{5})$. Twisting may therefore give a cover defined over $\mathbb Q$, but the
Galois action will then only be defined over $\mathbb Q (\sqrt{5})$ and be
accordingly more complicated. We therefore forgo this calculation and content
ourselves with the symmetric form above.
\end{exm}
For more on the questions considered in this section, see also further work by
Couveignes \cite{CouveignesQuelques}, and in a similar vein, the work of van
Hoeij--Vidunas on covers of conics \cite[\S\S 3.3--3.4]{vHV}, \cite[\S 4]{vHV2}.
We again refer to the fundamental paper of D\`ebes--Douai \cite{DebesDouai}, in
which strong results are given for both Bely\u{\i}\ maps and $G$-Bely\u{\i}\ maps that
suffice in many concrete situations. Admittedly, this subject is a delicate one,
and we hope that computations will help to further clarify these nuances.
\section{Simplification and verification}\label{sec:veri}
Once a potential model for a Bely\u{\i}\ map has been computed, it often remains to
simplify the model as much as possible and to verify its correctness
(independently of the method used to compute it). The former problem is still
open in general; the latter has been solved to a satisfactory extent.
\subsection*{Simplification}
By \textsf{simplifying} a Bely\u{\i}\ map $f : X \to \mathbb P^1$, we mean to reduce the
total (bit) size of the model. Lacking a general method for doing this, we
focus on the following:
\begin{enumerate}
\item If $X$ is of genus $0$, we mean to find a coordinate on $X$ that decreases
the (bit) size of the defining coefficients of $f$.
\item If $X$ is of strictly positive genus, we mean to simplify the defining
equations for $X$. (In practice, this will also lead to simpler coefficients
of the Bely\u{\i}\ map $f$.)
\end{enumerate}
Problem (1) was considered by van Hoeij--Vidunas \cite[\S 4.2]{vHV} under
the hypotheses that one of the ramification points has a minimal polynomial of
degree at most $4$; one tries to find a smaller polynomial defining the
associated number field and changes the coordinate accordingly, which typically
yields one a simpler expression of the Bely\u{\i}\ map.
Problem (1) is directly related with Problem (2) for hyperelliptic curves, since
simplifying the equations for hyperelliptic curves over a field $K$ boils down
to finding a small representative of the $\GL_2 (K)$-orbit of a binary form.
Typically one also requires the defining equation to have integral coefficients.
For the case $K = \mathbb Q$, this leads one to consider the problem of finding simpler
representations for binary forms under the action of the group of integral
matrices $\SL_2 (\mathbb Z)$. This is considered by Cremona--Stoll
\cite{CremonaStoll}, using results from Julia \cite{Julia} to find a binary
quadratic covariant, to which classical reduction algorithms are then applied.
The resulting algorithms substantially reduce the height of the coefficients of
the binary form in practice, typically at least halving the bit size of already
good approximations in the applications \cite{CremonaStoll}. A generalization
to, and implementation for, totally real fields is given in Bouyer--Streng
\cite{BouyerStreng}.
In fact, corresponding results for the simplification of Bely\u{\i}\ maps can be
obtained by taking the binary form to be the product of the numerator and
denominator of the Bely\u{\i}\ map. That the resulting binary form may have double
roots and hence may not correspond to hyperelliptic curves is no problem; see
the discussion by Cremona--Stoll \cite[after Proposition 4.5]{CremonaStoll}.
This problem of reduction is intimately related with the problem of finding a
good model of a Bely\u{\i}\ map or hyperelliptic curve over $\mathbb Z$. Note that even for
the case $K = \mathbb Q$ we have not yet used the full group $\GL_2 (\mathbb Q)$; the
transformations in $\SL_2 (\mathbb Z)$ considered by Cremona and Stoll preserve the
discriminant, but it could be possible that a suitable rational transformation
decreases this quantity while still preserving integrality of the binary form.
An approach to this problem is given by Bouyer--Streng \cite[\S
3.3]{BouyerStreng}.
In general, Problem (2) is much harder, if only because curves of high genus
become more difficult to write down.
\begin{ques}
Are there general methods to simplifying equations of curves defined over a
number field in practice?
\end{ques}
\subsection*{Verification}
Let $f:X \to \mathbb P^1$ be a map defined over a number field $K$ of degree $d$ that
we suspect to be a Bely\u{\i}\ map of monodromy group $G$, or more precisely to
correspond to a given permutation triple $\sigma$ or a given dessin $D$. To show
that this is in fact the case, we have to verify that
\begin{enumroman}
\item $f$ is indeed a Bely\u{\i}\ map;
\item $f$ has monodromy group $G$; and
\item $f$ (or its monodromy representation) corresponds to the permutation
triple $\sigma$; or
\item[(iii)'] the pullback under $f$ of the closed interval $[0,1]$ is
isomorphic (as a dessin) to $D$.
\end{enumroman}
This verification step is necessary for all known methods, and especially when
using the direct method from Section \ref{sec:groeb}; the presence of parasitic
solutions means that not even all solutions of the corresponding system of
equations will be Bely\u{\i}\ maps, let alone Bely\u{\i}\ maps with correct monodromy
group or pullback.
Point (i) can be computationally expensive, but it can be accomplished, by
using the methods of computational algebraic geometry. Not even if $X = \mathbb P^1$ is this
point trivial, since although verifying that a Bely\u{\i}\ map is returned is
easy for dessins of small degree, we need better methods than direct
factorization of the polynomials involved as the degree mounts.
As for point (ii), one simple sanity check is to take a field of definition $K$
for $f$ and then to substitute different $K$-rational values of $t \not\in
\{0,1,\infty\}$. One obtains an algebra that is again an extension of $K$ of
degree $d$ and whose Galois group $H$ must be a subgroup of the monodromy group
$G$ by an elementary specialization argument. So if we are given a finite
number of covers, only one of which has the desired monodromy group $G$, then to
eliminate a cover in the given list it suffices to show that specializing this
cover gives a set of cycle type in $H$ that is not contained in the given
monodromy group $G$ when considered as a subgroup of $S_d$. Such cycle types can
be obtained by factoring the polynomial modulo a small prime of $K$.
There are many methods to compute Galois groups effectively in this way; a
general method is given by Fieker--Kl\"uners \cite{FiekerKlueners}. This method
proceeds by computing the maximal subgroups of $S_d$ and checking if the Galois
group lies in one of these subgroups by evaluating explicit invariants. This
method works well if $G$ has small index in $S_d$. Iterating, this allows on to
compute the monodromy group of a Bely\u{\i}\ map explicitly instead of merely giving
the maximal groups in which is it included. To this end, one may work modulo a
prime $\mathfrak{p}$ of good reduction, and in light of Beckmann's Theorem, we may
still reasonably expect a small prime of the ring of integers of $K$ that is
coprime with the cardinality $\#G$ of the monodromy group to do the job.
Second, one can compute the monodromy by using numerical approximation. This
has been implemented by van Hoeij \cite{vanHoeij}, though one must be very
careful to do this with rigorous error bounds. This idea was used by Granboulan
\cite{Granboulan} in the computation of a cover with Galois group $M_{24}$,
first realized (without explicit equation) by Malle--Matzat
\cite[III.7.5]{MalleMatzat}. In particular, Schneps \cite[\S III.1]{Schneps}
describes a numerical method to draw the dessin itself, from which one can read
off the mondromy. This method is further developed by Bartholdi--Buff--von
Bothmer--Kr\"oker \cite{Bartholdi}, who lift a Delaunay triangulation
numerically and read off the permutations by traversing the sequence of edges
counterclockwise around a basepoint. In particular this solves (iii): if we
express each of the complex solutions obtained by embedding $K \hookrightarrow
\mathbb C$, we may also want to know which cover corresponds to which permutation
triple up to conjugation.
A third and final method is due to Elkies \cite{ElkiesM23}, who uses an
effective version (due to Weil's proof of the Riemann hypothesis for curves over
finite fields) of the Chebotarev density theorem in the function field setting.
This was applied to distinguish whether the Galois group of a given cover was
equal to $M_{23}$ or $A_{23}$. More precisely, one relies on reduction modulo a
prime whose residue field is prime of sufficiently large characteristic (in his
case, $>10^9$) and uses the resulting distribution of cycle structures to deduce
that the cover was actually $M_{23}$. This method has the advantage of using
exact arithmetic and seems particularly well-suited to verify monodromy of large
index in $S_d$.
\section{Further topics and generalizations}\label{sec:gens}
This section discuss some subjects that are generalizations of or otherwise
closely related with Bely\u{\i}\ maps. At the end, we briefly discuss the theoretical
complexity of calculating Bely\u{\i}\ maps.
\subsection*{Generalizations}
Over $\overline{\mathbb F}_p$, one can consider the reduction of Bely\u{\i}\ maps from
characteristic $0$; this is considered in Section \ref{sec:padic} above.
Switching instead to global function fields might be interesting, especially if
one restricts to tame ramification and compares with the situation in
characteristic $0$. As a generalization of Bely\u{\i}'s theorem, over a perfect
field of characteristic $p>0$, every curve $X$ has a map to $\mathbb P^1$ that is
ramified only at $\infty$ by work of Katz \cite{Katz}. But this map is
necessarily wildly ramified at $\infty$ if $g(X) > 0$, so the corresponding
theory will differ essentially from that of Bely\u{\i}\ maps over $\overline{\mathbb Q}$.
If we view Bely\u{\i}'s theorem as the assertion that every curve over a number
field is an \'etale cover of $\mathbb P^1 \setminus \{0,1,\infty\} \cong \mathcal{M}_{0,4}$,
the moduli space of genus $0$ curves with $4$ marked points, then Bely\u{\i}'s
result generalizes to a question by Braungardt \cite{Braungardt}: is every
connected, quasi-projective variety $X$ over $\overline{\mathbb Q}$ birational to a finite
\'etale cover of some moduli space of curves $\mathcal{M}_{g,n}$? Easton and Vakil
also have proven that the absolute Galois group acts faithfully on the
irreducible components of the moduli space of surfaces \cite{EastonVakil}.
Surely some computations in small dimensions and degree will be just as
appealing as in the case of Bely\u{\i}\ maps.
As mentioned on a naive level in Remarks \ref{rmk:naive1} and \ref{rmk:naive2},
another more general way to look at Bely\u{\i}\ maps is through the theory of
\textsf{Hurwitz schemes}, which give a geometric structure to the set
$\mathcal{H}_{n,r}(\overline{\mathbb Q})$ parametrizing degree $n$ morphisms to $\mathbb P^1$ over $\mathbb Q$
that are ramified above $r$ points. The theorem of Bely\u{\i}\ then amounts to
saying that by taking the curve associated to a morphism, one obtains a
surjective map from the union of the $\overline{\mathbb Q}$-rational points of the spaces $\mathcal{H}_{n,3}$ to the union of the
$\overline{\mathbb Q}$-rational points of the moduli spaces of curves $\mathcal{M}_g$ of
genus $g$. We refer to work of Romagny--Wewers \cite{RomagnyWewers} for a more
complete account.
\subsection*{Origamis}
One generalization of Bely\u{\i}\ maps is given by covers called \textsf{origamis}: covers
of elliptic curves that are unramified away from the origin. For a more
complete account on origamis, see Herrlich--Schmith\"usen
\cite{HerrlichSchmithuesen}; Bely\u{\i}\ maps can be obtained from origamis by a
degeneration process \cite[\S 8]{HerrlichSchmithuesen}.
The reasons for considering origamis are many. First, the fundamental group of
an elliptic curve minus a point is analogous to that of the Riemann sphere, in
that it is again free on two generators. The ramification type above the origin
is now given by the image of the commutator of these two generators. The local
information at this single point of ramification reflects less information about
the cover than in the case of Bely\u{\i}\ maps. Additionally, the base curve can be
varied, which makes the subject more subtle, as Teichm\"uller theory makes its
appearance.
An exciting family of special origamis was considered by Anema--Top
\cite{AnemaTop}: they consider the elliptic curve $E : y^2 = x^3 + a x + b$ over
the scheme $B : 4 a^3 + 27 b^2 = 1$ defined by the constant non-vanishing
discriminant $1$ of $E$. Considering the torsion subschemes $E[n]$ over $B$,
one obtains a family of covers over the base elliptic curve $B$ of $j$-invariant
$0$ that is only ramified above the point at infinity and whose Galois groups
are subgroups of special linear groups. It would be very interesting to deform
this family to treat the case of arbitrary base curves, though it is not clear
how to achieve this.
\begin{ques}
How does one explicitly deform special origamis to families with arbitrary
base curves?
\end{ques}
Explicit examples of actual families of origamis were found by
Rubinstein-Salzedo \cite{RS1,RS2}. In particular, by using a deformation
argument starting from a nodal cubic, he obtains a family of hyperelliptic
origamis that are totally ramified at the origin. For the case of degree $3$,
this gives a unique cover of genus $2$. More precisely, starting with an
elliptic curve $E$ with full $2$-torsion in Legendre form
\begin{align*}
y^2 = x (x - 1) (x - \lambda) ,
\end{align*}
the hyperelliptic curve
\begin{align*}
y^2 = \frac{1}{2}\left(-4 x^5 + 7 x^3 - (2 \lambda - 1) x^2 - 3 x + (2 \lambda
- 1) \right)
\end{align*}
admits a morphism to $E$ given by
\begin{align*}
\left( x,y\right) \longmapsto \left( \frac{1}{2} \left( -4 x^3 + 3 x + 1
\right) , \frac{y}{2} \left(-4 x^2 + 1 \right) \right)
\end{align*}
that is only ramified at the points at infinity of these curves.
It is important to note here that the field of moduli of these covers is an
extension of the field of moduli of the base elliptic curve; more precisely, as
suggested by the formulas above, this field of moduli is exactly the field
obtained by adjoint the $2$-torsion of the curve. This is a variation on a
result in Rubinstein-Salzedo \cite{RS1}, where simpler expressions for similar
covers are found in every degree. Amusingly enough, adjoining the full
$2$-torsion of the base curves always suffices to define these covers. This
result is appealing and quite different from the corresponding situation for
Bely\u{\i}\ maps, and therefore we ask the following question.
\begin{ques}
Which extension of the field of moduli is needed to define similar covers
totally ramified above a single point for general curves?
\end{ques}
\subsection*{Specialization}
Covering maps of the projective line with more than 3 ramification points
specialize to Bely\u{\i}\ maps by having the ramification points coincide. In many
cases, the covers in the original spaces are easier to compute, and this
limiting process will then lead to some non-trivial Bely\u{\i}\ maps. This also
works in reverse, and provides another application of computing Bely\u{\i}\ maps.
Hallouin--Riboulet-Deyris \cite{HallouinModuli} explicitly computed
polynomials with Galois group $A_n$ and $S_n$ over $\mathbb Q(t)$ with four branch
points for small values of $n$; starting from a relatively simple ``degenerate''
three-point branched cover, the four-point branched cover is obtained by complex
approximation (using Puiseux expansions). These methods were considerably
augmented by Hallouin in \cite{HallouinStudy} to find another such family with
group $\PSL_2 (\mathbb F_8)$. More recently, K\"onig \cite{Konig} similarly computed
such an extension of $\mathbb Q (t)$ with Galois group $\PSL_5(\mathbb F_2)$, using a $p$-adic
approximation to calculate the initial three-point degeneration. In all
aforementioned cases, the resulting covers can be specialized to find explicit
solution to the inverse Galois problem for the groups involved, and as mentioned
at the end of Section \ref{sec:backg}, the results from \cite{HallouinStudy}
have also found an application in the determination of equations for Shimura
curves \cite{HallouinComputation}.
As mentioned in Section \ref{sec:coan}, Couveignes \cite{CouveignesTools} has
used a patching method to describe more generally the computation of families of
ramified branch covers, using a degeneration to the situation of three-point
covers. More extensive algorithmic methods to deal with this question should
therefore be in reach of the techniques of numerical algebraic geometry.
\subsection*{Complexity}
In this article, we have been primarily concerned with practical methods for
computing Bely\u{\i}\ maps; but we conclude this section by posing a question
concerning the theoretical complexity of this task.
\begin{ques}
Is there an algorithm that takes as input a permutation triple and produces as
output a model for the corresponding Bely\u{\i}\ map over $\overline{\mathbb Q}$ that runs in
time doubly exponential in the degree $n$?
\end{ques}
There is an algorithm (without a bound on the running time) to accomplish this
task, but it is one that no one would ever implement: there are only countably
many Bely\u{\i}\ maps, so one can enumerate them one at a time in some order and use
any one of the methods to check if the cover has the desired monodromy. It
seems feasible that the Gr\"obner method would provide an answer to the above
problem, but this remains an open question. Javanpeykar \cite{Javanpeykar} has
given explicit bounds on the Faltings height of a curve in terms of the degree
of a Bely\u{\i}\ map; in principle, this could be used to compute the needed
precision to recover the equations over a number field.
|
\section{Introduction}
The interaction with the environment often plays an important role in
studies of ultracold atoms, leading to gain or loss of particles. One important
example is the inelastic three-body collision of particles in a BEC which can
be described in mean-field approximation by an imaginary interaction potential,
thus rendering the Hamiltonian non-Hermitian~\cite{Moiseyev2011a, Kagan98a}.
Imaginary potentials also find application in studies of dissipative
optical lattices~\cite{Abdullaev10a, Bludov10a}, and the non-Hermitian GPE has
been derived as the mean-field limit of an open Bose-Hubbard system described
by a master equation in Lindblad form~\cite{Trimborn08a, Witthaut11a}. A gain
of particles is less described in literature, but the feeding of a condensate
from a thermal cloud has been described by a positive imaginary
potential~\cite{Kagan98a}.
We study the dynamics of a BEC in a double-well potential where particles are
coherently removed from one well and injected into the other. The system
is described in mean-field approximation by the GPE. The GPE is known to yield
accurate results for temperatures considerably smaller than the critical
temperature of the condensate but has limitations in the vicinity of dynamic
instabilities~\cite{Anglin01a, Vardi01a}. The removal and injection of
particles is described by an imaginary potential. Both a coherent influx and
outflux have been experimentally realized. A coherent particle loss has, e.g.,
been implemented using a focused electron beam~\cite{Gericke08a} whereas the
influx can be provided from a second condensate exploiting electronic
excitations of the atoms~\cite{Robins08a}.
The gain and loss contributions are chosen in such a way that the resulting
system is $\mathcal{PT}$-symmetric, where $\mathcal{P}$ denotes the parity operator
$\hat{x}\to-\hat{x}$, $\hat{p}\to-\hat{p}$ and $\mathcal{T}$ the time reversal
operator $\hat{p}\to-\hat{p}$, $\mathrm{i}\to-\mathrm{i}$. Even though being non-Hermitian,
such systems can exhibit entirely real eigenvalues in certain parameter
regimes~\cite{Bender98a}. In recent years efforts have been pursued to
establish a quantum theory in which the requirement of Hermiticity is replaced
by the weaker condition of $\mathcal{PT}$ symmetry~\cite{Bender99a, Bender07a} or by the
more general concept of pseudo-Hermiticity~\cite{Mostafazadeh02a,
Mostafazadeh02b, Mostafazadeh02c}. Besides these fundamental approaches to
generalize quantum mechanics a great variety of $\mathcal{PT}$-symmetric systems which
are potentially experimentally accessible has been theoretically
investigated~\cite{Klaiman08a, Schindler11a, Bittner12a, Cartarius12a,
Mayteevarunyoo13a, Graefe12b}. The experimental breakthrough succeeded in
optical waveguide systems~\cite{Klaiman08a, Rueter10a, Guo09a}, now being one
of the main foci in the field of $\mathcal{PT}$ symmetry.
For the system considered $\mathcal{PT}$ symmetry demands that the influx and outflux of
particles into or from the condensate is balanced in such a way that
stationary solutions can be found. Embedding the $\mathcal{PT}$-symmetric double well
into a Hermitian four-well potential is a possible experimental realization of
this system~\cite{Kreibich13a}.
Prior to the experimental realization of a $\mathcal{PT}$-symmetric nonlinear quantum
system it is of utmost importance to have a detailed understanding of the
dynamical behavior resulting from the combination of $\mathcal{PT}$ symmetry and
nonlinearity. In particular it is necessary to study the implications on the
stability properties of the stationary solutions since only stable states are
observable. However, not only the stability of stationary states but also the
dynamics of arbitrary wave packets is relevant because for certain initial wave
packets the number of particles diverges, thus destroying the condensate. We
will address these problems, laying the foundation for future attempts of an
experimental realization of a $\mathcal{PT}$-symmetric Bose-Einstein condensate.
In this article we solve the dimensionless GPE with contact interaction
\begin{equation}
\left[-\Delta+V(\mathbf{r})+8\pi N_0 a|\psi(\mathbf{r},t)|^2\right] \psi(\mathbf{r},t)
=\mathrm{i} \frac{\partial}{\partial t} \psi(\mathbf{r},t),
\label{eq:gpe}
\end{equation}
where $N_0$ is the number of particles and $a$ is the scattering length. The
three-dimensional double-well potential chosen,
\begin{equation}
V(\mathbf{r})= \frac14 x^2 + \frac14 \omega_{y,z}^2(y^2+z^2)
+ v_0 \mathrm{e}^{-\sigma x^2} + \mathrm{i} \gamma x \mathrm{e}^{-\rho x^2},
\label{eq:potential_3d}
\end{equation}
is $\mathcal{PT}$-symmetric since $V(\mathbf{r})=V^*(-\mathbf{r})$ holds. The
potential consists of a three-dimensional harmonic trap with an identical
trapping frequency $\omega_{y,z}$ in $y$ and $z$ direction. The harmonic trap
in $x$ direction is superimposed by a Gaussian barrier with height $v_0$ and
width parameter $\sigma$, thus forming a symmetric double-well potential. The
strength of the antisymmetric imaginary part of the potential is tuned by
$\gamma$, and the parameter $\rho$ is chosen in such a way that the extrema of
the imaginary part coincide with the minima of the double well. The positive
imaginary part can be interpreted as a source of probability density while a
negative part corresponds to a sink. In all following calculations the
values $\omega_{y,z}=2$, $v_0=4$, $\sigma=0.5$, and $\rho\approx0.12$ are fixed
whereas the gain/loss parameter $\gamma$ and the strength of the nonlinearity
$N_0a$ are varied.
Since the system is non-Hermitian the norm of the wave function $\psi$ is not
conserved and due to the nonlinearity of the GPE a different norm changes the
dynamics. The physical interpretation of a change in the norm of the wave
function is a change in the number of particles
\begin{equation}
N = N_0 ||\psi||^2.
\label{eq:particlenumber}
\end{equation}
Without interaction, i.e.\ $Na=0$, the Hamiltonian is separable and is solved
by the product ansatz $\psi(x)\psi_m(y)\psi_m(z)$, where $\psi_m$ is the $m$-th
eigenstate of the one-dimensional harmonic oscillator. Since the energy of the
first excited state of the harmonic oscillator is about one order of magnitude
larger than that of the first excited state of the double well, only the ground
states of the harmonic oscillators in $y$ and $z$ direction are taken into
account. Thus the three-dimensional problem is reduced to one dimension with
the one-dimensional $\mathcal{PT}$-symmetric double-well potential
\begin{equation}
V(x)= \frac14 x^2 + v_0 \mathrm{e}^{-\sigma x^2} + \mathrm{i} \gamma x \mathrm{e}^{-\rho x^2}.
\label{eq:potential_1d}
\end{equation}
This can easily be solved by a numerically exact integration.
The product ansatz does not exactly solve the nonlinear GPE with
contact interaction, however, for small values of $Na$ it is still a good
approximation, and as we will see not only shows qualitatively the same
behavior as the calculations in three dimensions but also quantitatively.
The article is organized as follows. In Sec.~\ref{sec:stationary} the
stationary solutions in three dimensions are presented and a comparison with
the one-dimensional solutions is drawn.
To study the stability of the stationary solutions the Bogoliubov-de Gennes
equations are adapted for non-Hermitian systems and solved numerically in
Sec.~\ref{sec:stability}. In Sec.~\ref{sec:dynamics} the time evolution
of wave packets is investigated using the Bloch sphere formalism. Conclusions
are drawn in Sec.~\ref{sec:conclusion}.
\section{Stationary solutions}
\label{sec:stationary}
The eigenvalue spectrum and eigenstates of a BEC in the $\mathcal{PT}$-symmetric double
well~\eqref{eq:potential_1d} have already been discussed~\cite{Dast12a,
Dast13a}, however, only as a function of the gain/loss parameter $\gamma$. To
study dynamical properties it is more instructive to discuss the eigenvalues as
a function of $Na$, where $N$ is the effective number of particles as defined
in Eq.~\eqref{eq:particlenumber}.
To solve the three-dimensional system we use the \textit{time-dependent
variational principle}~\cite{Mclachlan64a,Rau2010c,Rau10a,Rau10b} whose
application to $\mathcal{PT}$-symmetric nonlinear systems has been discussed
in~\cite{Dast12a}. Our ansatz consists of coupled Gaussian functions
\begin{align}
\psi(\mathbf{r}) = \sum_{\mathrm{k}}
\exp \big[ -A_x^\mathrm{k} (x-q_x^\mathrm{k})^2 + A_{y,z}^\mathrm{k} (y^2+z^2) \notag\\
- \mathrm{i} p_x^\mathrm{k} (x-q_x^\mathrm{k}) + \varphi^\mathrm{k} \big]
\label{eq:tdvp_ansatz}
\end{align}
with $A_x^\mathrm{k}, A_{y,z}^\mathrm{k}, \varphi^\mathrm{k} \in \mathbb{C}$, $q_x^\mathrm{k}, p_x^\mathrm{k}
\in \mathbb{R}$.
The variational principle yields equations of motion for these time-dependent
quantities, and stationary solutions are found as the fixed points of these
equations. The stationary solutions of the three-dimensional
potential~\eqref{eq:potential_3d} presented in this section are gained with
the variational approach whereas the one-dimensional
potential~\eqref{eq:potential_1d} is solved numerically exact. We use up to
two Gaussian functions per well, i.e.\ up to four Gaussians in total, which
yields a small correction to one Gaussian function per well.
\subsection{Spectrum of the real double well}
Figure~\ref{fig:spectrum_3d} shows the eigenvalues of the GPE in the
\begin{figure}%
\centering%
\includegraphics[width=\columnwidth]{spectrum_3d}%
\caption{%
(Color online)
(a) The chemical potential $\mu$ of the eigenvalues without gain/loss,
i.e.\ $\gamma=0$, as a function of the particle number scaled scattering
length. The symmetric (blue solid branch) and antisymmetric (yellow dashed
branch) stationary solutions arise at two almost identical bifurcation
points at $Na\approx-1.4$, and the doubly degenerate symmetry-breaking
eigenstates (green dot-dashed branch) at $Na\approx-0.7$. (b) Introducing
the gain/loss $\gamma=0.02$ hardly changes $\real \mu$ of the
symmetry-breaking ($\mathcal{PT}$-broken) eigenstates but the two $\mathcal{PT}$-symmetric
branches now arise at a tangent bifurcation. (c) The tangent bifurcation of
the two $\mathcal{PT}$-symmetric states in the vicinity of the bifurcation point for
a small value of $\gamma$. (d) The difference between the chemical
potential of the $\mathcal{PT}$-symmetric states and their mean value $\bar{\mu}$.
Different values of $\gamma$ are compared to the case $\gamma=0$ (black
solid line).
}%
\label{fig:spectrum_3d}%
\end{figure}%
three-dimensional $\mathcal{PT}$-symmetric double well~\eqref{eq:potential_3d} with the
lowest total energy. For $\gamma=0$ the spectrum in
Fig.~\ref{fig:spectrum_3d}(a) contains three pairs of states each of which form
a common branch, shown as blue solid, yellow dashed and green dot-dashed lines.
Note that the blue and yellow branches almost lie on top of each other. Every
branch is born in a tangent bifurcation at a critical value of $Na$. At the
three tangent bifurcations not only the eigenvalues but also the eigenstates
coincide.
The two bifurcations for the blue and yellow branches reside at almost
identical values $Na \approx -1.4$. The states arising at these tangent
bifurcations have an equal probability of presence in both wells. At one
bifurcation two states with symmetric (even parity) wave functions arise (blue
solid lines), whereas the two states arising at the other bifurcation (yellow
dashed lines) are antisymmetric (odd parity). At the bifurcation points the
wave functions turn into two peaks strongly confined in the two wells due to
the attractive interaction. For stronger attractive interactions no stationary
solutions exist and the condensate collapses~\cite{Gammal01a, Donley01a}.
The same collapse is observed at approximately half the interaction strength
$Na \approx -0.7$ (green dot-dashed branch) where the wave function is entirely
localized in one well, i.e.\ it corresponds to exactly the same collapse
process of the wave function, but for an asymmetric wave restricted to one side
of the double well. The states arising at this bifurcation are clearly parity
symmetry broken and doubly degenerate because the condensate can be located
either in the right or in the left well. The two uppermost panels of
Fig.~\ref{fig:spectrum} shows the eigenvalues for small moduli of $Na$ in
\begin{figure}%
\centering%
\includegraphics[width=\columnwidth]{spectrum}%
\caption{%
(Color online)
(a) Real and (b) imaginary part of the chemical potential of the
eigenstates for small values of $Na$. From top to bottom the four different
gain/loss contributions $\gamma=0$, $0.02$, $0.04$, and $0.042$ are used.
The eigenvalues of the $\mathcal{PT}$-symmetric and $\mathcal{PT}$-broken solutions in three
dimensions are in excellent agreement with the numerically exact
one-dimensional solutions. Stable branches are highlighted as thick
lines (see Sec.~\ref{sec:stability}).
}%
\label{fig:spectrum}%
\end{figure}%
more detail revealing that the parity symmetry broken states coalesce with the
symmetric solution at $Na\approx-0.0075$ and with the antisymmetric solution at
$Na\approx0.0075$. In the region $|Na|\lesssim0.0075$ the parity symmetry
breaking solutions do not exist. The existence of these solutions is known as
macroscopic quantum self-trapping~\cite{Albiez05a}.
\subsection{Spectrum of the $\mathcal{PT}$-symmetric double well}
Introducing the gain and loss contribution $\gamma=0.02$ changes the eigenvalue
spectrum significantly as can be seen in Fig.~\ref{fig:spectrum_3d}(b). The
even and odd state with lower energies vanish and the two
remaining states only exist for $Na\gtrsim-0.9$. These two states are
$\mathcal{PT}$-symmetric, thus having a symmetric real part and an antisymmetric
imaginary part (see Fig.~\ref{fig:wavefnc}(a),(b)). For the symmetry-breaking
\begin{figure}%
\centering%
\includegraphics[width=\columnwidth]{wavefnc}%
\caption{%
(Color online)
Real (solid lines) and imaginary (dotted lines) part of the wave function
of the $\mathcal{PT}$-symmetric ground (a) and first excited (b) state and the two
$\mathcal{PT}$-broken states with a negative (c) and positive (d) imaginary part of
the chemical potential. For $\gamma=0$ the ground state is parity symmetric
and the first excited state is antisymmetric. For finite values of $\gamma$
the real part of both $\mathcal{PT}$-symmetric wave functions is even and the
imaginary part is odd. The $\mathcal{PT}$-broken states are asymmetric thus breaking
the $\mathcal{PT}$ symmetry of the Hamiltonian.
}%
\label{fig:wavefnc}%
\end{figure}%
states the real part of the chemical potential is approximately identical to
the case $\gamma=0$ but the eigenvalue $\mu$ is now complex. Due to the $\mathcal{PT}$
symmetry of the system, the two imaginary parts of $\mu$ have the same absolute
value but different signs. The wave functions of the $\mathcal{PT}$-broken states are
shown in Fig.~\ref{fig:wavefnc}(c),(d). To understand the behavior of the two
tangent bifurcations of the symmetric and antisymmetric states we compare the
vicinity of the bifurcation points for $\gamma=0$ and $\gamma=10^{-3}$.
Figure~\ref{fig:spectrum_3d}(c) shows that the two bifurcations vanish and a
new tangent bifurcation between the two $\mathcal{PT}$-symmetric states arises.
If we want to compare the eigenvalues of the two $\mathcal{PT}$-symmetric states for
different values of $\gamma$ we have to choose a slightly different
presentation since the difference between the eigenvalues of the two
$\mathcal{PT}$-symmetric states is very small compared to their absolute change.
Figure~\ref{fig:spectrum_3d}(d) shows $\mu-\bar{\mu}$ of the two
$\mathcal{PT}$-symmetric states, where $\bar{\mu}$ denotes the mean value of $\mu$ of
these two states. We immediately see that the tangent bifurcation at which the
two $\mathcal{PT}$-symmetric states coalesce is shifted to greater values $Na$ if the
gain/loss parameter $\gamma$ is increased. For $\gamma\gtrsim0.41$ the
bifurcation point is shifted to positive values of $Na$, i.e.\ the two
$\mathcal{PT}$-symmetric solutions investigated here only exist for repulsive
interaction of the atoms. At $Na\approx-0.8$ the chemical potential of the two
$\mathcal{PT}$-symmetric states becomes equal for all parameters $\gamma$ for which the
$\mathcal{PT}$-symmetric states exist at $Na\approx-0.8$. However, at this point only a
generic degeneracy of the energy eigenvalue $\mu$ occurs, the states
themselves are different.
The relevant $\mathcal{PT}$-symmetric effects of the system, namely the breaking of
$\mathcal{PT}$-symmetry, occur already at small absolute values of $Na$. This regime
is shown in Fig.~\ref{fig:spectrum}. The $\mathcal{PT}$-broken states emerge from the
$\mathcal{PT}$-symmetric state in a pitchfork bifurcation. For an attractive
interaction the $\mathcal{PT}$-broken solutions emerge from the $\mathcal{PT}$-symmetric ground
state whereas in the case of repulsive interaction the $\mathcal{PT}$-broken branches
emerge from the excited $\mathcal{PT}$-symmetric state (cf.\ middle two panels in
Fig.~\ref{fig:spectrum}). The $\mathcal{PT}$-broken solutions arise at approximately the
same absolute value of $Na$ for both attractive and repulsive interaction. As
is known from $\mathcal{PT}$-symmetric systems the eigenvalues of the $\mathcal{PT}$-broken
solutions occur in complex conjugate pairs. For increasing values of the
gain/loss parameter $\gamma$ the tangent bifurcation in which the two
$\mathcal{PT}$-symmetric states vanish is shifted to greater values of $Na$. At the same
time the $\mathcal{PT}$-broken solutions emerge already at smaller absolute values of
$Na$. At strong enough values of $\gamma$ the tangent bifurcation is shifted
to repulsive interactions $Na>0$ and the $\mathcal{PT}$-broken states exist even for
$Na=0$.
Although being solutions of the time-independent GPE the $\mathcal{PT}$-broken
states are no stationary solutions of the time-dependent GPE. They
experience an exponential gain or decay of the norm, thus effectively changing
the nonlinearity parameter $Na$.
In the following we will restrict the discussion to small moduli of $Na$. As
already mentioned in this regime the GPE is in good approximation solved by a
product ansatz, thus reducing the problem to one dimension. The comparison in
Fig.~\ref{fig:spectrum} confirms the excellent agreement between the
calculations in one and three dimensions for the parameter range considered and
justifies the reduction to one dimension used in the following sections.
The solutions for the one-dimensional
system~\eqref{eq:potential_1d} are obtained with numerically exact methods by
integrating the GPE outwards and fulfilling boundary conditions \cite{Dast12a}.
\section{Stability}
\label{sec:stability}
The first step towards understanding the dynamical properties of the system is
the stability analysis of the stationary solutions with respect to small
perturbations. The stability of the $\mathcal{PT}$-symmetric stationary states has
already been discussed rudimentarily in~\cite{Dast12a}. We will shortly review
these results and then focus on the stability in the vicinity of the
bifurcations and the study of the dynamics of $\mathcal{PT}$-broken solutions.
In addition to the stability analysis in one dimension, which is presented in
this section, we investigated the stability in three dimensions by linearizing
the equations of motion of the time-dependent variational principle as
described in~\cite{Dast12a}. In three dimensions excitations in $y$ and $z$
direction may give rise to additional instabilities. However, we found that in
the parameter range considered these instabilities do not occur for
$\omega_{y,z} \gtrsim 3$ and we again observe an excellent agreement between
the calculations in one and three dimensions.
The time-dependent GPE is linearized in the vicinity of the stationary states,
yielding the Bogoliubov-de Gennes equations
\begin{subequations}
\begin{align}
\Delta u=&
\left(V-\mu-\omega-8N_0a\left|\psi_0\right|^2\right)u-4N_0a\psi_0^2v,\\
\Delta v=&
\left(V^*-\mu^*+\omega-8N_0a\left|\psi_0\right|^2\right)v-4N_0a\psi_0^{*2}u.
\end{align}
\label{eq:bdge}
\end{subequations}
A solution of these equations determines the behavior of a perturbation
$\delta\psi(x,t)=u(x)\exp(-\mathrm{i} \omega t)+v^*(x)\exp(\mathrm{i} \omega^* t)$ of a
normalized stationary state at interaction strength $N_0 a$. For real
frequencies $\omega$ the perturbed state performs stable oscillations around
the fixed point. If $\omega$ has a non-vanishing imaginary part it is necessary
to distinguish between two cases. A negative imaginary part describes an
exponentially damped and thus stable perturbation. By contrast, a perturbation
increases exponentially if the imaginary part is positive. If one or more
perturbations have a frequency with positive imaginary part the stationary
solution is unstable otherwise it is stable. Due to the ansatz of the
perturbation for every frequency $\omega$ with amplitudes $(u,v)$ a second
solution with frequency $-\omega^*$ and amplitudes $(v^*, u^*)$ exists.
Therefore all frequencies occur in pairs with positive and negative values of
$\real \omega$. Applying the $\mathcal{PT}$ operator to the Bogoliubov-de Gennes
equations shows that if the stationary state $\psi_0$ has a perturbation
frequency $\omega$ then $\mathcal{PT} \psi_0$ has a perturbation frequency $\omega^*$.
Thus for $\mathcal{PT}$-symmetric eigenstates every frequency is always part of a set of
four solutions with $\pm \real \omega \pm \mathrm{i} \imag \omega$. For $\mathcal{PT}$-broken
eigenstates perturbations occur in pairs $\pm \real \omega + \mathrm{i} \imag \omega$
and since $\mathcal{PT}$-broken solutions are mapped onto each other by application of
the $\mathcal{PT}$ operator their perturbation frequencies are complex conjugate.
\subsection{Stability of the $\mathcal{PT}$-symmetric solutions}
The Bogoliubov-de Gennes equations are solved numerically exact by integrating
the amplitudes $u$ and $v$ outwards and demanding that they vanish at the
boundaries. The first nontrivial Bogoliubov-de Gennes eigenvalue with smallest
absolute real part is shown in Fig.~\ref{fig:stability_pt} for the two
\begin{figure}%
\centering%
\includegraphics[width=\columnwidth]{stability}%
\caption{%
(Color online)
(a),(c) Real and (b),(d) imaginary parts of the eigenvalues
$\omega_\mathrm{g/e}$ of the Bogoliubov-de Gennes equations for (a),(b) the
ground state and (c),(d) the first excited state. Only the first nontrivial
eigenvalue with smallest absolute real part is shown. Due to the symmetries
of the equations the real and the imaginary parts of the eigenvalues both
occur in pairs. Both states are stable for weak interactions but become
unstable in the vicinity of the pitchfork bifurcations.
}%
\label{fig:stability_pt}%
\end{figure}%
$\mathcal{PT}$-symmetric stationary states. Higher excitations are neglected since they
have real eigenvalues thus describing stable perturbations.
For a weak gain/loss contribution $\gamma=0.02$ the stability eigenvalues show
that both $\mathcal{PT}$-symmetric stationary states are stable for small interaction
strengths. The ground state (Fig.~\ref{fig:stability_pt}(a),(b)) becomes
unstable at attractive interactions and the first excited state
(Fig.~\ref{fig:stability_pt}(c),(d)) at repulsive interactions. Both stability
changes occur near the pitchfork bifurcations at which the $\mathcal{PT}$-broken states
emerge. For stronger gain/loss contributions $\gamma=0.04$ the pitchfork
bifurcations at which the two states become unstable are shifted to lower
absolute values of $N_0a$. Additionally we observe that the first excited
state becomes unstable for attractive interactions shortly before it merges
with the ground state at $N_0a\approx-0.065$ in a tangent bifurcation and
vanishes.
For gain/loss contributions $\gamma\gtrsim0.41$ the $\mathcal{PT}$-broken states exist
even for $N_0a=0$ and the $\mathcal{PT}$-symmetric stationary states exist only for
repulsive interactions. In this case the ground state is stable for all
interaction strengths whereas the excited state becomes unstable shortly after
it emerges in the tangent bifurcation at $N_0a\approx0.03$. It is worth noting
that the stability of the $\mathcal{PT}$-symmetric states changes not at the bifurcation
points but only in their vicinity. The reason for this behavior is discussed in
detail at the end of this section.
\subsection{Perturbations of the $\mathcal{PT}$-broken solutions}
Although the $\mathcal{PT}$-broken solutions are no stationary solutions of the
time-dependent GPE it is nevertheless instructive to solve the Bogoliubov-de
Gennes equations for these states. We will see that this is relevant for the
understanding of the dynamics of the condensate as well as the stability of the
stationary $\mathcal{PT}$-symmetric solutions. Figure~\ref{fig:stability_broken} shows
the four Bogoliubov-de
\begin{figure
\centering%
\includegraphics[width=\columnwidth]{stability_broken}%
\caption{%
(Color online)
(a),(c) Real and (b),(d) imaginary part of the Bogoliubov-de Gennes
eigenvalues $\omega_{\mathrm{i}\pm}$ for the $\mathcal{PT}$-broken states with (a),(b)
positive and (c),(d) negative imaginary part of the chemical potential
$\mu$ at $\gamma=0.03$. The four eigenvalues with smallest real parts are
shown. In addition to each eigenvalue $\omega$ a solution with
negative real part $-\omega^*$ exists. All but one eigenvalue of the
$\mathcal{PT}$-broken state with $\imag \mu > 0$ ($\imag \mu < 0$) have a negative
(positive) imaginary part.
}%
\label{fig:stability_broken}%
\end{figure}%
Gennes eigenvalues with smallest absolute real part for the two $\mathcal{PT}$-broken
states using a constant value of the gain/loss parameter $\gamma=0.03$. The
stability eigenvalues of the $\mathcal{PT}$-broken state with $\imag \mu > 0$ in
Fig.~\ref{fig:stability_broken}(a),(b) show that the three eigenvalues with
nonzero real part have a negative imaginary part. In fact there is an infinite
number of further perturbations with negative imaginary part corresponding to
higher excited states of the double well. There is, however, an additional
solution with $\real \omega_{\mathrm{i} +} = 0$ and a positive imaginary part
describing a perturbation which increases exponentially. Since the
$\mathcal{PT}$-broken solutions are not stationary the usual interpretation of the
eigenvalues $\omega$ as stability indicators is invalid. The damped oscillatory
behavior described by the eigenvalues with negative imaginary part and
non-vanishing real part is characteristic for these states as will be seen in
Sec.~\ref{sec:dynamics}.
Since the $\mathcal{PT}$-broken states can be mapped onto each other by application of
the $\mathcal{PT}$ operator the $\mathcal{PT}$-broken state with $\imag \mu < 0$ has the complex
conjugate stability eigenvalues shown in
Fig.~\ref{fig:stability_broken}(c),(d). This is a consequence of the fact that
this state has the same dynamics as the $\mathcal{PT}$-broken state with $\imag \mu > 0$
if evolved in negative time direction.
\subsection{Stability at the bifurcation points}
Both the ground state and the excited state become unstable near a bifurcation
point. Therefore the Bogoliubov-de Gennes eigenvalues in the vicinity of the
bifurcations are now investigated in more detail. Figure~\ref{fig:gap} shows
\begin{figure}%
\centering%
\includegraphics[width=\columnwidth]{gap}%
\caption{%
(Color online)
Imaginary part of the Bogoliubov-de Gennes eigenvalues of all states
involved in (a) the pitchfork bifurcation at $\gamma=0.03$ and (b) the
tangent bifurcation at $\gamma=0.04$. The ground state becomes unstable
before the pitchfork bifurcation at which the $\mathcal{PT}$-broken solutions
arise. For decreasing $N_0a$ the excited state becomes unstable before the
tangent bifurcation at which the two $\mathcal{PT}$-symmetric states coalesce and
vanish. At the bifurcations the stability eigenvalues of the states
involved coalesce since the states themselves become equal. The bifurcation
points are marked by black vertical lines.
}%
\label{fig:gap}%
\end{figure}%
the stability eigenvalues of all states involved in the bifurcations. It can
immediately be seen that the stability properties of the $\mathcal{PT}$-symmetric states
do not change at the bifurcation points but only in their vicinity. The ground
state (Fig.~\ref{fig:gap}(a)) is already unstable for greater values of $N_0a$,
i.e.\ the ground state becomes unstable in a parameter regime where the
$\mathcal{PT}$-broken solutions do not yet exist. This was already found in a two-mode
analysis of a $\mathcal{PT}$-symmetric double well \cite{Rodrigues13a}. A similar
behavior is observed at the tangent bifurcation where the excited state becomes
unstable shortly before the bifurcation point at which the $\mathcal{PT}$-symmetric
solutions vanish (Fig.~\ref{fig:gap}(b)).
This discrepancy is surprising because we know from real nonlinear systems that
the stability of eigenstates changes at bifurcation points. Also in linear
$\mathcal{PT}$-symmetric systems all eigenstates are stable unless a $\mathcal{PT}$-broken
eigenstate with complex eigenvalue exists. In both cases a change in the
stability of a stationary state coincides with a qualitative change in the
spectrum.
As a first step we ensure that the investigation of the lowest-lying
states is sufficient and higher excited states are not responsible for the
discrepancy. Therefore we solved the Bogoliubov-de Gennes equations for the
$\mathcal{PT}$-symmetric double-delta potential studied in~\cite{Cartarius12a,
Cartarius12b}, a system in which only two $\mathcal{PT}$-symmetric and two $\mathcal{PT}$-broken
eigenstates exist. Indeed, also this system shows the discrepancy thus ruling
out higher excited states as its origin \cite{Loehle14a}. It therefore seems
likely that the observed discrepancy is a consequence of the combination that
both a nonlinear and simultaneously $\mathcal{PT}$-symmetric system is investigated. Due
to the non-Hermiticity the norm is not conserved thus the nonlinearity
parameter $Na$ and consequently the spectrum changes with time. As a result the
dynamical properties are not governed by the eigenvalue spectrum at a fixed
interaction strength. Instead it is necessary to consider the whole spectrum as
a function of $Na$, and thus the $\mathcal{PT}$-broken states are already dynamically
accessible in the parameter regime where only $\mathcal{PT}$-symmetric states exist.
To confirm that this property is indeed the reason why the stability does not
change at the bifurcation points we modify the Gross-Pitaevskii nonlinearity
\begin{equation}
|\psi(x,t)|^2 \to \frac{|\psi(x,t)|^2}{\int|\psi(x,t)|^2 \mathrm{d} x}.
\label{eq:norm_independent_nonlin}
\end{equation}
This formulation is equivalent to the mean-field limit of the $\mathcal{PT}$-symmetric
Bose-Hubbard dimer by Graefe et al.\ in which such a discrepancy does not
occur~\cite{Graefe10a}.
Replacing the nonlinearity with~\eqref{eq:norm_independent_nonlin} does not
change the normalized eigenstates of the GPE. It leads, however, to a different
form of the Bogoliubov-de Gennes equations,
\begin{subequations}
\label{eq:bdge_ne}
\begin{align}
\left(-\Delta+V-\omega-\mu-8N_0a\left|\psi_0\right|^2\right)u
-4N_0av\psi_0^2 \notag\\
+4N_0a\left|\psi_0\right|^2\psi_0\int v\psi_0+u\psi_0^*\mathrm{d}^3 r
=0, \\
\left(-\Delta+V^*+\omega-\mu^*-8N_0a\left|\psi_0\right|^2\right)v
-4N_0au\psi_0^{*2} \notag\\
+4N_0a\left|\psi_0\right|^2\psi_0^*\int v\psi_0+u\psi_0^*\mathrm{d}^3r=0.
\end{align}
\end{subequations}
These equations are solved in the vicinity of the bifurcations so that the
results with the adapted nonlinearity~\eqref{eq:norm_independent_nonlin} in
Fig.~\ref{fig:gap_hold} can directly be compared to the stability eigenvalues
\begin{figure}%
\centering%
\includegraphics[width=\columnwidth]{gap_hold}%
\caption{%
(Color online)
Imaginary part of the stability eigenvalues $\omega$ of the Bogoliubov-de
Gennes equations~\eqref{eq:bdge_ne} for the norm-independent
nonlinearity~\eqref{eq:norm_independent_nonlin} in the vicinity of (a) the
pitchfork bifurcation and (b) the tangent bifurcation. A constant
gain/loss parameter (a) $\gamma=0.03$ and (b) $\gamma=0.04$ is used. The
ground state becomes unstable at the pitchfork bifurcation and the excited
state is stable until it vanishes at the tangent bifurcation. The
$\mathcal{PT}$-broken states are now a pure sink or source. The bifurcation points
are marked by black vertical lines.
}%
\label{fig:gap_hold}%
\end{figure}%
obtained with the usual Gross-Pitaevskii nonlinearity in Fig.~\ref{fig:gap}.
Using the adapted nonlinearity~\eqref{eq:norm_independent_nonlin} the stability
of the ground state changes at the bifurcation point, i.e.\ the ground state
becomes unstable as soon as the $\mathcal{PT}$-broken solutions exist. Additionally the
excited state does not show a stability change for attractive interaction but
stays stable until it vanishes. Thus, the behavior observed that the stability
of the $\mathcal{PT}$-symmetric states does not change at the bifurcation points but
only in their vicinity is indeed a result of the norm-dependent nonlinearity of
the GPE.
Also the behavior of the $\mathcal{PT}$-broken states does change by introducing the
adapted nonlinearity. With the nonlinearity~\eqref{eq:norm_independent_nonlin}
the $\mathcal{PT}$-broken states are eigenstates of the time-dependent GPE with
exponentially increasing or decreasing norm proportional to $\exp(2\imag\mu)$.
Furthermore the $\mathcal{PT}$-broken states with positive and negative imaginary part
are now a pure sink or source, respectively.
\section{Wave packet dynamics}
\label{sec:dynamics}
The linear stability analysis in the previous section describes the dynamical
behavior in the vicinity of the eigenstates, and in the case of the
$\mathcal{PT}$-broken solutions for very short time spans after which the spectrum
changes. For finite time spans this is not sufficient. To gain a more
elaborate picture of the dynamics the time evolution of wave packets is
investigated for different values of the gain/loss parameter $\gamma$ and the
nonlinearity parameter $Na$.
We use the \textit{split-operator method} to numerically calculate the time
evolution of wave packets which is known to produce accurate results even for
nonlinear equations as the GPE~~\cite{Feit82a, Fleck76a, Javanainen06a}.
The oscillation of a single wave packet between the wells of the
$\mathcal{PT}$-symmetric double-well potential has already been discussed
in~\cite{Dast12a}. There the square modulus of the wave packets was
investigated to analyze the characteristic phase shift of the oscillations.
However, this representation does not allow for predictions of the behavior of
arbitrary wave packets and the impact of the eigenstates of the system. We will
now choose the Bloch sphere formalism as a different approach to visualize the
time evolution of arbitrary states.
Even though Bloch sphere representations have already been used for
$\mathcal{PT}$-symmetric~\cite{Graefe08b, Graefe10a} as well as
dissipative~\cite{Trimborn08a} two-mode BECs these studies always restricted
the dynamics to the surface of the Bloch sphere. In non-Hermitian nonlinear
systems, in which the norm is not conserved and the associated differential
equation depends explicitly on the norm of the wave function, the dynamics
is in general not restricted to this surface. We will see that the Bloch
sphere provides a significant insight into the dynamical properties
nonetheless.
\subsection{Bloch sphere formalism}
In general the representation as a Bloch sphere is limited to two-level quantum
systems. Since the Hilbert space of the system investigated is not
two-dimensional we use a projection to the space spanned by the $\mathcal{PT}$-symmetric
ground state $\psi_\mathrm{g}$ and excited state $\psi_\mathrm{e}$. In the linear
case $Na=0$ the time evolution of initial wave functions consisting of a linear
superposition of the ground and excited state is restricted to this
two-dimensional space. With interaction $Na\neq0$ this is no longer true,
however, our calculations show that for all time evolutions considered the
projection to two dimensions is still a very good approximation.
We now choose two orthogonal basis vectors of the space spanned by the two
$\mathcal{PT}$-symmetric stationary solutions. The first basis vector is identical to
the normalized ground state $\ket{e_1} = \ket{\psi_\mathrm{g}}$, and the second
basis vector, $\ket{e_2} = \alpha(\ket{\psi_\mathrm{e}} -
\braket{\psi_\mathrm{g}}{\psi_\mathrm{e}} \ket{\psi_\mathrm{g}})$ with normalization
constant $\alpha$, is the component of the excited state $\ket{\psi_\mathrm{e}}$
orthogonal to $\ket{e_1}$ and is selected by application of the Gram-Schmidt
method. Both basis vectors are exactly $\mathcal{PT}$-symmetric. An arbitrary wave
function can be written as
\begin{equation}
\ket{\psi} = c_1 \ket{e_1} + c_2 \ket{e_2} + \ket{\psi_\mathrm{err}}
\label{eq:bloch_projection}
\end{equation}
with $c_1 = \braket{e_1}{\psi}$ and $c_2 = \braket{e_2}{\psi}$. Although all
initial wave packets considered are superpositions of $\ket{e_1}$ and
$\ket{e_2}$ the time evolution will in general leave the space spanned by
$\ket{e_1}$ and $\ket{e_2}$. The norm of $\ket{\psi_\mathrm{err}}$ measures the
error made by the projection to the two-dimensional Hilbert space. In all
calculations presented $\braket{\psi_\mathrm{err}}{\psi_\mathrm{err}} < 0.004$ was
found, justifying the projection.
The basis vectors $\ket{e_1}$ and $\ket{e_2}$ are defined to correspond to the
north and south pole of the Bloch sphere, respectively, by introducing the
spherical coordinates $R\in[0,\infty)$, $\varphi\in(-\pi,\pi]$ and
$\vartheta\in[0,\pi]$ as follows
\begin{subequations}
\label{eq:bloch_coordinates}
\begin{align}
c_1 &= R \, \mathrm{e}^{\mathrm{i}\chi+\mathrm{i}{\varphi}} \cos(\vartheta/2), \\
c_2 &= R \, \mathrm{e}^{\mathrm{i}\chi} \sin(\vartheta/2),
\end{align}
\end{subequations}
with an arbitrary phase $\chi\in\mathbb{R}$. Since
$\braket{\psi_\mathrm{err}}{\psi_\mathrm{err}} \ll 1$ the radius can be identified
with the norm of the wave packet $R \approx ||\psi||$. The orientation of the
basis vectors on the Bloch sphere and the stationary solutions at a fixed value
of $\gamma$ are shown in Fig.~\ref{fig:bloch_orientation}.
\begin{figure}%
\centering%
\includegraphics[width=0.67\columnwidth]{bloch_orientation}%
\caption{%
(Color online)
The Bloch sphere representation using the coordinates defined in
Eq.~\eqref{eq:bloch_coordinates}. The basis vectors $\ket{e_1}$ and
$\ket{e_2}$ correspond to the north and south pole, respectively. On the
front side of the plotted great cycle (black line) the azimuth angle is
$\varphi=0$ and on the back side $\varphi=\pi$. All states residing on the
plane in which this great cycle lies are $\mathcal{PT}$-symmetric and the time
evolution in the system is symmetric with respect to this plane.
Furthermore the great cycle is used as the starting point for trajectories
discussed in Sec.~\ref{sec:bloch_dynamics}. Additionally the four
eigenstates are plotted for $\gamma=0.03$ and the value of $Na$
corresponding to the current location
(cf.\ Sec.~\ref{sec:bloch_eigenstates}).
}%
\label{fig:bloch_orientation}%
\end{figure}%
The $\mathcal{PT}$ symmetry of the system has several implications for the
representation as a Bloch sphere. Since the system considered is non-Hermitian
the norm of a wave packet is not conserved. The norm was identified with the
radius $R$, thus in general the time evolution of an arbitrary wave packet is
not constrained to the surface of the Bloch sphere but it will either dive into
the sphere or leave the surface to larger radii. The two basis vectors are
exactly $\mathcal{PT}$-symmetric thus application of the $\mathcal{PT}$ operator leads to a
complex conjugation of the coefficients $c_1$ and $c_2$ which is equivalent to
the reflection $\varphi \to -\varphi$. Consequently all $\mathcal{PT}$-symmetric states
reside on the plane defined by $\varphi\in\{0,\pi\}$.
A further implication of the $\mathcal{PT}$ symmetry of the system is that if
$\psi(x,t)$ is a solution of the time-dependent GPE then $\psi^*(-x,-t)$ is
also a solution. Since $\mathcal{P}$ reflects the spatial coordinate and
$\mathcal{T}$ applies only a complex conjugation $\psi^*(-x,-t)=\mathcal{PT}\psi(x,-t)$
holds and hence all trajectories are symmetric with respect to the plane
$\varphi\in\{0,\pi\}$, in which the $\mathcal{PT}$ symmetric eigenstates are found.
\subsection{Eigenstates in Bloch sphere representation}
\label{sec:bloch_eigenstates}
Figure~\ref{fig:bloch_nonlin} shows both the eigenstates and time-evolved wave
\begin{figure*}%
\centering%
\includegraphics[width=\textwidth]{bloch_nonlin}%
\caption{%
(Color online)
The dynamics of wave packets projected to the space spanned by the
$\mathcal{PT}$-symmetric ground and excited state illustrated on a Bloch sphere. The
interaction strength $Na=-0.05$ on the surface of the sphere is constant in
all figures whereas the gain/loss parameters are varied, with (a)
$\gamma=0$, (b) $\gamma=0.0025$, (c) $\gamma=0.01$, (d) $\gamma=0.02$, (e)
$\gamma=0.03$, and (f) $\gamma=0.04$. The solutions of the time-independent
GPE are plotted for orientation (pink lines
cf.~Fig.~\ref{fig:bloch_orientation}). The ground state starts in the
center of the sphere and goes through the north pole whereas the
penetration point of the excited state is at the south pole for $\gamma=0$
and wanders on the meridian $\varphi=0$ to the north pole for increasing
values of $\gamma$. The $\mathcal{PT}$-broken solutions emerge from the ground
state. All wave packets shown start on a great cycle through the north
pole, south pole and the excited state. Wave packets starting on the
meridian between the ground state at the north pole and the excited state
on the front side of the sphere evolve to a smaller norm thus diving into
the sphere (closed thick blue lines inside the sphere). Wave packets
starting in the remaining region of the great circle either show
oscillations with a larger norm outside the sphere (closed red lines
outside the sphere) or diverge while encircling the $\mathcal{PT}$-broken solutions
(green lines departing from the left and the right side).
}%
\label{fig:bloch_nonlin}%
\end{figure*}%
packets in the Bloch sphere representation. In this example the Bloch sphere
represents the interaction strength $Na=N_0a=-0.05$. Since a larger radius is
equivalent to a greater amount of particles, $N= N_0||\psi||^2 = N_0 R^2$,
plotting the $\mathcal{PT}$-symmetric and $\mathcal{PT}$-broken solutions of the time-independent
GPE is another way of showing these eigenstates in dependence of $Na$. They are
depicted as thick pink lines in Fig.~\ref{fig:bloch_nonlin} and can be seen
most clearly in Fig.~\ref{fig:bloch_nonlin}(a). Only attractive interactions
are shown thus larger radii relate to more negative values of $Na$.
Both $\mathcal{PT}$-symmetric solutions start at the center of the sphere. The ground
state goes through the north pole and the excited state goes through the south
pole for $\gamma=0$ (Fig.~\ref{fig:bloch_nonlin}(a)). Increasing $\gamma$ the
penetration point of the excited state through the Bloch sphere wanders on the
meridian $\varphi=0$ to the north pole at which the ground state resides
(cf.~Figs~\ref{fig:bloch_nonlin}(b)-(e)). For a critical value of $\gamma$ this
point reaches the north pole, which is almost fulfilled in
Fig.~\ref{fig:bloch_nonlin}(f). For greater values of $\gamma$ the
$\mathcal{PT}$-symmetric solutions even vanish on the surface of the Bloch sphere and
only exist within the sphere. This behavior of the stationary solutions on the
surface of the Bloch sphere can be comprehended by comparison with the
eigenvalue spectrum in Fig.~\ref{fig:spectrum} for different values of $\gamma$
and a fixed value of $Na$.
As already shown in the eigenvalue spectrum the bifurcation at which the
ground and excited state coalesce can also be reached by tuning $Na$. This can
be seen in Fig.~\ref{fig:bloch_nonlin}(e) where the two $\mathcal{PT}$-symmetric states
coalesce at a critical radius outside the Bloch sphere resulting in the
closed circle of the thick pink lines. For smaller values of $\gamma$ shown in
Fig.~\ref{fig:bloch_nonlin}(a)-(d) the bifurcation point lies at a larger
radius outside the figure.
Since attractive interactions are shown the two $\mathcal{PT}$-broken solutions emerge
from the ground state. For the chosen interaction strength on the Bloch sphere
$Na=-0.05$ the $\mathcal{PT}$-broken solutions emerge outside the sphere for
$\gamma\gtrsim0.022$ and inside the sphere for $\gamma\lesssim0.022$. The
$\mathcal{PT}$-broken solution with $\imag\mu>0$ ($\imag\mu<0$) lies on the left (right)
side of the symmetry plane.
\subsection{Dynamics on the Bloch sphere}
\label{sec:bloch_dynamics}
After this short discussion of the eigenstates we now address the time
evolution. For all calculations shown in Fig.~\ref{fig:bloch_nonlin} the
initial wave packets are normalized and a linear superposition of $\ket{e_1}$
and $\ket{e_2}$, thus $R=1$ and $\braket{\psi_\mathrm{err}}{\psi_\mathrm{err}}=0$.
The azimuth angle is either $\varphi=0$ or $\pi$, therefore all initial wave
packets start on a great circle of the Bloch sphere through the north pole,
south pole, and the excited stationary state. They are integrated in positive
and negative time direction.
For the case $\gamma=0$ shown in Fig.~\ref{fig:bloch_nonlin}(a)
the norm and therefore the radius are conserved quantities. All trajectories
stay on the surface of the Bloch sphere. Both the stationary ground and excited
state are elliptic fixed points and therefore stable.
This behavior changes drastically if a small gain/loss parameter
$\gamma=0.0025$ is introduced as done in Fig.~\ref{fig:bloch_nonlin}(b). Due to
the gain and loss, the norm of the wave packets are no longer conserved and the
trajectories no longer run on the surface of the Bloch sphere. We identify two
different regions on the great circle for initial wave packets which are
delimited by the ground state, viz.\ the north pole, and the excited state. All
wave packet starting in the region on the front side between the two fixed
points (thick blue lines) evolve to a smaller norm inside the Bloch sphere
whereas wave packets starting on the second region on the back side (red lines)
evolve to a higher norm outside the sphere. The two $\mathcal{PT}$-symmetric stationary
states are again elliptic fixed points, thus being stable. The sum of all
oscillating trajectories of the wave packets define closed surfaces which
cannot be penetrated by other trajectories.
Increasing the gain/loss parameter to $\gamma=0.01$ leads to the situation
shown in Fig.~\ref{fig:bloch_nonlin}(c). Again the wave packets starting on the
front region between the two fixed points oscillate to a smaller norm and
define a closed surface inside the Bloch sphere. However, the wave packets
starting on the back region and oscillating outside the Bloch sphere do no
longer define a closed surface. Instead an additional type of trajectories with
diverging norm encircling the $\mathcal{PT}$-broken eigenstates appears. For
$\gamma=0.02$ the amount of diverging trajectories increases as can be seen in
Fig.~\ref{fig:bloch_nonlin}(d). These trajectories arrive from the vicinity of
the $\mathcal{PT}$-broken state with $\imag\mu<0$ on the right side, touch the sphere,
and leave the region of the sphere to larger radii $R$ encircling the path of
the $\mathcal{PT}$-broken state with $\imag\mu>0$ on the left side. This illustrates the
role of the $\mathcal{PT}$-broken solutions for the dynamics of the condensate as sink
and source.
Another qualitative change in the dynamical behavior is found for $\gamma=0.03$
in Fig.~\ref{fig:bloch_nonlin}(e). In agreement with the investigation of the
linear stability in Fig.~\ref{fig:stability_pt} the ground state is unstable
for $Na=-0.05$, viz.\ exactly on the surface of the Bloch sphere. For lower
values of $\gamma$ there is a region around the ground state in which only
stable oscillations originate. For $\gamma=0.03$ the ground state is unstable
and all wave packets starting on the $\varphi=\pi$ meridian behind the ground
state are diverging. Wave packets starting before the ground state still show
stable oscillations evolving inside the sphere.
Finally for $\gamma=0.04$ shown in Fig.~\ref{fig:bloch_nonlin}(f) the two
$\mathcal{PT}$-symmetric stationary solutions are almost identical and most wave packets
starting on the great circle on the Bloch sphere diverge. Only wave packets
starting in a small region around the stable excited state and in the region
between the two stationary states still show stable oscillations. For even
greater values of $\gamma$ the stationary solutions do no longer exist on the
surface of the Bloch sphere and no stable oscillations starting on the great
circle can be observed.
\section{Conclusion and outlook}
\label{sec:conclusion}
We studied the implications of $\mathcal{PT}$ symmetry on the dynamical behavior and
stability of a BEC with contact interaction in a
double-well potential.
Solving the Bogoliubov-de Gennes equations for non-Hermitian systems showed
that the two $\mathcal{PT}$-symmetric stationary solutions with smallest chemical
potential are stable as long as the $\mathcal{PT}$-broken states do not exist.
However, the ground state becomes dynamically unstable in the vicinity of the
pitchfork bifurcation at which the $\mathcal{PT}$-broken states emerge from the ground
state at an attractive interaction strength. Analogously the excited state
becomes unstable at repulsive interactions in the vicinity of the pitchfork
bifurcation. The discrepancy between the bifurcations and the points at which
the stability changes could be traced back to the norm dependency of the
Gross-Pitaevskii nonlinearity.
Due to the non-Hermiticity of the system the dynamics is not governed by
isolated fixed points but an infinite number of eigenstates which solve
the time-independent GPE for the varying number of particles. The dynamics of
the condensate was visualized using the Bloch sphere formalism although the
dynamics is not constrained to the surface of the sphere due to the
nonlinearity and non-Hermiticity of the system. Applying a small gain and loss
of particles leaves the typical dynamics in a real double-well potential mostly
intact but the trajectories now run slightly above or below the surface of the
sphere describing a condensate with more or less particles,
respectively. For stronger gain/loss contributions an additional type of
trajectories arises which describes a condensate localized in one well with
a diverging number of particles. These diverging trajectories encircle the
$\mathcal{PT}$-broken eigenstates of the time-independent GPE. However it is still
possible to choose initial wave packets that show stable oscillations. In fact
we observed that all $\mathcal{PT}$-symmetric wave functions which initially dive into
the sphere always show stable oscillations. If the gain and loss is further
increased most wave packets diverge and stable oscillations are only found in a
small region in the vicinity of the excited state.
Understanding the dynamics of a BEC in a $\mathcal{PT}$-symmetric double well is the
first step towards an experimental realization of a $\mathcal{PT}$-symmetric quantum
system and the starting point for studies in more complex potentials and with
additional interaction types like the dipolar interaction. Additionally it
would be highly desirable to obtain a microscopic description of the in- and
outcoupling process represented by an imaginary potential in the mean-field
limit. Analyzing the bifurcation scenario at strong attractive interaction
strengths switching from two tangent bifurcations to one in the presence of
gain and loss as visible in Fig.~\ref{fig:spectrum_3d} is an interesting task
for future work from a more theoretical point of view and can probably be
achieved using the analytic continuation described in~\cite{Dast13a}.
|
\section{Introduction}
In the intermediate energy region of a few GeV, the cross sections for charged lepton production
are dominated by quasielastic and various inelastic processes involving pion production induced by charged
and neutral currents in neutrino and antineutrino reactions from nuclear targets.
The experimental observations of strange particles through weak interaction induced $\Delta S =0 $ and $|\Delta S| = 1 $ processes are quite limited.
These are limited both by statistics as well as by the large systematic errors. Also theoretically there are only a few works available in literature.
However, the availability of high intensity neutrino and antineutrino beams in present generation neutrino
experiments has opened up the possibility of experimentally studying these processes with better statistics.
In this work, we summarize our results of the total scattering cross sections for single kaon/antikaon, eta, associated particle production
for neutrino/antineutrino induced processes as well as antineutrino induced single hyperon production. The details of the model are given in
Refs.~\cite{RafiAlam:2010kf,Alam:2012zz,Alam:2013,Alam:2013woa,Singh:2006xp}.
The calculations are performed using a microscopical model based on the
chiral perturbation theory($\chi PT$) at the level of baryon and meson octet~\cite{RafiAlam:2010kf} for the background terms and for $K^-/{\bar K^0}$ production resonant
mechanism is introduced by the inclusion of the
decuplet baryons~\cite{Alam:2012zz}.
For $|\Delta S| =1 $ hyperon production cross section induced by antineutrinos we followed the prescription
of Ref.~\cite{Singh:2006xp}.
\section{Results and Discussions}
\begin{figure}[t]
\begin{center}
\includegraphics[height=5.5cm, width=7.5cm]{xsec_nu_weak.eps}
\includegraphics[height=5.5cm, width=7.5cm]{xsec_nubar_weak.eps}
\caption{$\sigma$ vs $E_{\nu_\mu/ \bar \nu_\mu}$ for $|\Delta S|=1$, K(left panel) and $\bar K$(right panel) production.}
\label{fig:total_single_xsec}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[height=5.5cm, width=7.5cm]{associated_nu_v2.eps}
\includegraphics[height=5.5cm, width=7.5cm]{associated_nubar_v2.eps}
\caption{$\sigma$ vs $E_{\nu_\mu/ \bar \nu_\mu}$ for $\Delta S=0$ process induced by $\nu$(left panel) and
$\bar\nu$(right panel).}
\label{fig:total_ass_xsec}
\end{center}
\end{figure}
Here we are discussing some of the results for $\Delta S =0 $ and $|\Delta S| = 1$ processes,
which would be quite useful in the analysis of neutrino oscillation physics, in estimation of the background in proton decay searches besides their own intrinsic importance.
For single kaon production ($|\Delta S| = 1$), we have considered the following reactions:
\begin{small}
\begin{eqnarray}\label{kaon}
\nu_\mu + p \rightarrow \mu^- + K^+ + p ~~~~~~& \bar \nu_\mu + p \rightarrow \mu^+ + K^- + p \nonumber \\
\nu_\mu + n \rightarrow \mu^- + K^0 + p ~~~~~~& \bar \nu_\mu + p \rightarrow \mu^+ +\bar K^0 + n \nonumber \\
\nu_\mu + n \rightarrow \mu^- + K^+ + n ~~~~~~& \bar \nu_\mu + n \rightarrow \mu^+ + K^- + n \; .
\end{eqnarray}
\end{small}
In Fig.~\ref{fig:total_single_xsec}, we have presented the results for $\nu(\bar\nu)$ induced K($\bar K$) production cross sections.
The kaon production gets contribution from contact term, kaon pole term, u-channel diagram and pion/eta in flight term~\cite{RafiAlam:2010kf}.
For the antikaon production besides the background terms, we have also taken the contribution from lowest lying $\Sigma^*(1385)$ resonance~\cite{Alam:2012zz}.
We find that the contact term is dominant, followed by the u-channel diagram with a $\Lambda$ intermediate state and the $\pi$ exchange term.
The kaon pole contributions are negligible. We have used a global dipole form factor with a mass of 1 GeV and multiplied it with the hadronic current.
We observe similar features in the results of antikaon production cross section as in the case of kaon production,
and the contribution from $\Sigma^*(1385)$ resonance is very small.
In Fig.~\ref{fig:total_ass_xsec}, we have presented the results for neutrino and antineutrino induced associated kaon production cross section.
We have considered the following reactions for $\nu(\bar\nu)$ charged current induced associated
particle production process:
\begin{small}
\begin{eqnarray}\label{associated}
\nu_\mu + n \rightarrow \mu^- + K^+ + \Lambda~~~~~~&\bar \nu_\mu + p \rightarrow \mu^+ + K^0 +
\Lambda~~ \nonumber\\
\nu_\mu + p \rightarrow \mu^- + K^+ + \Sigma^+ ~~~~~~& \bar \nu_\mu + p \rightarrow \mu^+ + K^0 +
\Sigma^0~ \nonumber\\
\nu_\mu + n \rightarrow \mu^- + K^+ + \Sigma^0~ ~~~~~~& \bar \nu_\mu + p \rightarrow \mu^+ +
K^+ + \Sigma^- \nonumber\\
\nu_\mu + n \rightarrow \mu^- + K^0 + \Sigma^+ ~~~~~~& \bar \nu_\mu + n \rightarrow \mu^+ + K^0
+ \Sigma^-
\end{eqnarray}
\end{small}
The various form factors that appear in the vector and axial vector currents
have been determined using the prescription given in
Refs.~\cite{Singh:2006xp,Cabibbo:2003cu}.
In this case also we find that the contact term has the largest contribution to
the cross section followed by s-channel and kaon pole terms. We find that the
cross section for the reaction channel with a $\Lambda$ in
the final state are in general larger than that for the reactions where $\Sigma$
is in the
final state. This can be understood by looking at the relative strength of the
coupling, for example, the ratio of couplings squared for the vertices
$NK^0\Lambda$ to $NK^0\Sigma^0$ is $\frac{g^2_{NK\Lambda}}{g^2_{NK\Sigma}}
\simeq 14$. Furthermore, the cross section for the $\Lambda$ production is
favored by
the available phase space due to its small mass relative to $\Sigma$ baryons.
However, for the antineutrino induced process,
the mechanism where $\Lambda$ is in the final state is not dominating. We
find that the neutrino induced reactions
are the dominant source of $K^+$ production whereas antineutrino induced reactions
favor $K^0$ production.
In Fig.~\ref{fig:eta_xsec}, we have presented the results for eta production cross section and $Q^2$ distribution, for the following processes:
\begin{small}
\begin{eqnarray}\label{eq:eta_weak_process}
\nu_\mu + n \rightarrow \mu^- + \eta + p ~~~~~~~~~~~~\bar \nu_\mu + p \rightarrow \mu^+ + \eta + n
\end{eqnarray}
\end{small}
In these processes from symmetry only s- and u-channel nucleon pole terms contribute. Besides Born-terms, we have also considered
$S_{11}$(1535) and $S_{11}$(1650) resonances. For the resonant mechanism we parameterized the vector part of the from factors using the
helicity amplitudes~\cite{Alam:Prep}. We derived Goldberger-Treiman relation for the axial couplings and assumed a dipole
form for $Q^2$ dependence for the axial form factors.
We find the dominance of $S_{11}$(1535) resonance followed by nucleon pole terms. $Q^2$ distribution in the case of eta production is almost
flat in nature.
\begin{figure}[tbh]
\begin{center}
\includegraphics[height=5.5cm, width=7.5cm]{eta_xsec.eps}
\includegraphics[height=5.5cm, width=7.5cm]{q2_eta.eps}
\caption{$\sigma$(left panel) and the $Q^2$ distribution(right panel) for $\nu_\mu / \bar\nu_\mu $ induced $\eta$ production off the nucleon.}
\label{fig:eta_xsec}
\end{center}
\end{figure}
At low energies, quasi-elastic production
of hyperons induced by antineutrinos is possibile
\begin{small}
\begin{eqnarray}\label{reaction1}
\bar \nu_l + p \rightarrow l^+ + \Lambda ~~~~~~~~\bar \nu_l + p \rightarrow l^+ + \Sigma^0 ~~~~~~~~ \bar \nu_l + n \rightarrow l^+ + \Sigma^-
\end{eqnarray}
\end{small}
\begin{figure}
\begin{center}
\includegraphics[height=6.5cm,width=15cm]{avg_dsdQ2_mb_SK_pi0.eps}
\caption{$Q^2$ distributions for $\bar\nu_\mu$ induced reaction in $^{12}C$ averaged over the MiniBooNE flux,
and for $\bar\nu_{e,\mu}$ induced reaction in $^{16}O$ averaged over the atmospheric antineutrino flux at Super-Kamiokande
are shown with nuclear medium and FSI effects for $\pi^0$ production. The $\pi^0$ production
from hyperon excitations have been scaled by a factor of 1.33.}
\label{fig:hyp_xsec}
\end{center}
\end{figure}
At the energies of MiniBooNE~\cite{AguilarArevalo:2008qa} and
atmospheric~\cite{Ashie:2005ik} $\bar\nu$ experiments, mainly pion production is
considered through the
$\Delta S=0$ resonant mechanism with the dominance of $\Delta$ in the
intermediate state. In the case of antineutrino reactions in
nucleon and nuclei there is an additional contribution to the pion production
from the quasielastic
$|\Delta S|=1$ processes mentioned in Eq.~\ref{reaction1}, in which hyperons
like $\Lambda, \Sigma^{-,0}$ can be produced and decay subsequently to pions.
These processes are generally Cabibbo suppressed as compared to $\Delta$
production process
but could be important in the low energy region of antineutrinos. We find that
at low energies
due to threshold effects and phase space considerations these processes give a
significant contribution to the $\pi$ production~\cite{Alam:2013}.
In Fig.\ref{fig:hyp_xsec}, we present the results of $Q^2$ distribution averaged
over the MiniBooNE~\cite{AguilarArevalo:2008qa} and
atmospheric~\cite{Honda:2006qj}
antineutrino spectra, where $\pi^0$ production is obtained in the $\Delta$
dominance model following Ref.~\cite{Athar:2007wd} as well as $\pi^0$'s
contribution
from hyperons following Ref.~\cite{Singh:2006xp}. These results are obtained in
$^{12}C$ for MiniBooNE and in $^{16}O$ for atmospheric antineutrinos. The effect of nuclear medium like Fermi motion
and Pauli blocking is negligible. Furthermore, we find that when the contribution for the pions coming from all the hyperons
mentioned in Eq.\ref{reaction1}
is taken together there is no net change in the overall pion production due to final state interaction(FSI) effects,
as it increases in the $\Lambda$ production and decreases in the $\Sigma$ production in the nuclear medium~\cite{Alam:2013}.
We observe that in the peak region of $Q^2$ distribution, the contribution of
$\pi^0$ from the hyperon excitations is almost 75$\%$ to the
contribution of $\pi^0$ from the $\Delta$ excitation. Similar is the
observation for $\pi^-$ production~\cite{Alam:2013}. Thus we find that for
antineutrino experiments
in the energy region of 0.6-1.0 GeV, the pion contribution from hyperons are
significant.
One of the authors(MSA) is thankful to PURSE program of D.S.T., Govt. of India
and the Aligarh Muslim University for the financial support. This research was
supported by the Spanish Ministerio de Economía y Competitividad and
European FEDER funds under Contracts FIS2011-28853-C02-01, by Generalitat
Valenciana under Contract No.
PROMETEO/20090090 and by the EU HadronPhysics3 project, Grant Agreement No.
283286.
\section*{References}
|
\section{Introduction}
Cosmological measurements suggest that `dark energy' is the dominant energy component of the Universe, accounting for approximately 70 per cent of the energy density at the present day
\citep[e.g.][]{Blake2011cos,Blake2011distz,Conley2011,Anderson2012,Hinshaw2012,Padmanabhan2012,Planck2013CosmoParams}.
We have no theory that simultaneously explains both its existence and magnitude, which suggests that the standard model of particle physics, quantum physics, or our theory of gravity are incomplete. The simplest model of dark energy corresponds to Einstein's cosmological constant --- a constant energy density, $\rho$, with negative pressure, $p$, such that $p=-\rho$,
but it could also take a more exotic form such as a dynamical fluid with negative pressure, a scalar potential field, or can be accounted for by a modified theory of gravity such as $f(R)$ \citep{Nojiri2007}. In all models, dark energy can be characterized by its equation of state $w\equiv p/\rho$. Measuring the present value of $w$ and any time variation provides us with crucial information about the underlying physics of dark energy.
The properties of dark energy can be probed by studying its influence on the expansion of the Universe.
Standard candles and rulers are tools for mapping this expansion. Standard candles have had central roles in major cosmological discoveries, from the use of Cepheid variable stars in the discovery of the expanding Universe by \citet{Hubble1929} to the more recent use of Type Ia Supernovae (SNe) in the discovery of the accelerated expansion and `dark energy' \citep{Riess1998,Perlmutter1999}.
To investigate the Universe and its expansion, it is important to use data sets from various different probes to break degeneracies between cosmological parameters, and therefore derive tighter constraints \citep{Bahcall1999, Huterer2001,Levine2002, Melchiorri2003,Wang2004, Mantz2010},
Using multiple probes simultaneously is also important for consistency checks between independent methods and to understand and mitigate systematic errors.
Presently, several cosmological probes are being used to study the Universe, including: SNe, baryon acoustic oscillations (BAO), the cosmic microwave background (CMB), weak lensing (WL), and galaxy clustering (CL).
These probes provide us with important and complementary ways to investigate the properties of dark energy, but at present, with the exception of BAO, all of these methods for making cosmological measurements are restricted to relatively low redshifts. Proposed SN measurements may only be observed out to a redshift of $z<2.5$ \citep{Grogin2011,Koekemoer2011}, though the majority of measurements will remain at $z<1.5$, due to their relative faintness at these redshifts and observational magnitude limits, as well as a decrease in SN rates at high redshifts \citep{Albrecht2006, Hook2013}. Galaxy surveys [e.g. \textit{Sloan Digital Sky Survey} (SDSS) \citep{Percival2010}, \textit{WiggleZ} \citep{Blake2011cos}, and \textit{Baryon Oscillation Spectroscopic Survey} (BOSS) \citep{Anderson2012}], from which WL, CL, and BAO measurements are made, are also currently restricted to low redshifts ($z\lesssim 1$). Some BAO measurements have been obtained in a higher redshift regime by probing distant galaxies through Lyman$-\alpha$ absorption in quasar spectra \citep{Busca2012,Slosar2013}. Future galaxy surveys are predicted to extend the redshift range probed by WL and CL to $z\sim3$ but such surveys will not be completed in the next 5-10 years.
With information over a larger redshift range, we can more easily identify time-evolving behaviour in dark energy, if it is present. Fig. \ref{fig:distmods} shows that a large range of models are all reasonably consistent with the low-redshift data points and become more easily distinguishable with the inclusion of high-redshift measurements. The models of dark energy we have shown in this figure are described by the Chevallier--Polarski--Linder, CPL, parametrization, $w(z) = w_0+w_zz/(1+z)$.
Active galactic nuclei (AGN) have been proposed as high-redshift cosmological probes by \citet{Watson2011} using a technique called reverberation mapping. AGN display a tight empirical radius-luminosity ($r_{\rm BLR}-L(5100{\rm \AA})$) relation \citep{Koratkar1991,Wandel1999,Kaspi2000,Bentz2006, Bentz2009,Bentz2013}, which makes them suitable as standardizable candles. Here, $r_{\rm BLR}$ is the distance between the central accretion disc and the broad-emission-line region (BLR),
where nebular clouds reprocess the accretion disc continuum radiation of luminosity, $L$, into emission line photons.
The value of $r_{\rm BLR}$ is measured from the observed time lag between the nuclear continuum and broad-emission-line light-curve variations, taken to be the light travel time between the central source and the BLR \citep[see, e.g.,][]{Peterson2001}.
Since AGN are numerous, highly luminous, and persistent sources of light that are present over a broad range of epochs, they are good candidates for distance measurements. Currently the $r_{\rm BLR}-L$ relationship spans five orders of magnitude in the optical continuum luminosity at 5100\AA~ from $10^{41}$ to $10^{46}$ erg/s with an observed scatter in the relationship as low as 0.13 dex \citep[equivalent to 0.33 mag in the distance modulus; ][]{Bentz2013}, with a clear potential for further reduction in the scatter \citep{Watson2011} making the relationship a reasonable tool for dark energy investigations.
The $r_{\rm BLR}-L$ relationship is anchored in well-understood photoionization physics \citep{Peterson1997,Osterbrock2006}.
AGN broad emission lines are emitted when photoionization equilibrium is attained in the BLR. For systems with the same ionization parameter, gas densities, and ionizing spectral energy distributions (SEDs), this equilibrium occurs at a specific radius. In AGN, at least to the first order, this condition holds, and as a consequence the simple relationship, $r_{\rm BLR}\propto L^{1/2}$, is expected. This $r_{\rm BLR}-L$ relationship can be translated to $\tau/\sqrt{F}\propto D_L$, where $\tau$ is the measured time delay ($\tau=r_{\rm BLR}/c$), $F$ is the measured flux of the object, and $D_L$ is the luminosity distance. Thus a Hubble diagram can be constructed - see fig. 2 of \citet{Watson2011}.
The constancy of the ionization parameter, gas densities, and SED between AGN is supported by the agreement between the predicted and observed $r_{\rm BLR}-L$ relationship \citep{Bentz2009,Bentz2013} and the uniformity in AGN spectra \citep{VandenBerk2001, VandenBerk2004, Dietrich2002}. Despite this, the potential for intrinsic variation in this $r_{\rm BLR}-L$ relationship with black hole characteristics or metallicity, for example, may need to be tested further.
Besides AGN, gamma-ray burst \citep[GRB][]{Ghirlanda2006,Speirits2007, Liang2008, Diaferio2011,Wei2013}, Type II SNe \citep[SNe II][]{Poznanski2010}, and the supernovae associated with gamma ray bursts (GRB-SNe, Li \& Hjorth, in prep.) may also have potential to be standardisable candles, but at this stage, there is no strong evidence to support the use of these probes. New variability surveys make GRBs a highly sought after high-redshift standard candle (HzSC) candidate; however, the physics is still not well known. In contrast, AGN physics is better understood and much of the measurement scatter for AGN can be attributed to known correctable systematics \citep[][Kilerci-Eser et al., in prep.]{Watson2011}. Accordingly, AGN are likely to be our best candidate at this time.
In our analysis, we determine the requirements for an HzSC to be a competitive cosmological probe regardless of the type of standard candle. We investigate how this general standard candle can complement existing and future constraints on dark energy properties from Type Ia SNe, BAO, and the CMB. To investigate the properties of dark energy, we consider both a linear, time-evolving dark energy equation of state and a parametrization-independent piecewise equation-of-state model. We also consider how well a general standard candle can measure the Hubble parameter in independent redshift bins, and make an estimate of the dark energy density function. Similar work has been done by \citet{Goliath2001,Huterer2001, Frieman2003, Linder2003,Salzano2013} but only with standard candle measurements with the redshift capabilities of SNe and only in conjunction with CMB measurements. We extend the redshift range of the standard candle probe in this case and also consider the inclusion of BAO constraints. Other authors have looked at how future surveys will constrain dark energy but do not consider the possibility of an HzSC \cite[e.g.,][]{Albrecht2006, Sarkar2008,Eisenstein2011}.
The aim of this paper is to predict the power of standard candle measurements for constraining dark energy properties and determine the optimal redshift distribution required to set the tightest constraints. From this analysis we can make a judgement about how useful HzSCs are as cosmological probes and define an optimal survey strategy. This paper is organized as follows: Section~\ref{sec:prediction} describes the analysis methods we have implemented, Section~\ref{sec:probes} outlines the observables and data sets we use, while Section 4 details the parameters we fit. We present the results of our analysis in Section~\ref{sec:effect}. Discussion and conclusions are presented in Section~\ref{sec:discuss}.
\begin{figure}
\includegraphics[width=0.5\textwidth]{snecontoursw0wa}
\caption{Hubble diagram showing a distance modulus, $\mu$, normalized to that expected for a $\Lambda$CDM universe (dotted black curve). Blue and red curves show possible $w(z)$ models. The models of dark energy we show in this figure are described by the Chevallier-Polarski-Linder, CPL, parametrization, $w(z) = w_0+w_zz/(1+z)$. $\Lambda$CDM is therefore given by $w_0=1$; $w_z=0$. The Union2 SN data [grey for individual measurements and black for redshift binned results (Weighted arithmetic mean)] are also shown to demonstrate the range of redshifts currently probed by standard candle measurements. All models shown are hard to distinguish with only these data points. The lines with the same colours represent models that are hard to distinguish with only high-redshift measurements (due to uncertainty in the absolute magnitude) but are easy to distinguish with both high- and low-redshift measurements. The purple, brown, and green shaded regions shown at the bottom of the plot represent the predicted redshift limits of future SN, BAO, and proposed HzSC measurements, respectively. }
\label{fig:distmods}
\end{figure}
\section{Analysis Methods}\label{sec:prediction}
We predict the cosmological information that can be extracted from measurements of a reliable HzSC. In particular, we concentrate on how well an HzSC, with an extended redshift range, can further constrain the properties of dark energy over and above existing and predicted future SN, BAO, and CMB measurements. In order to make these predictions, we employ two different methods:
\begin{enumerate}
\item Parameter fitting ($\chi^2$), and
\item Fisher matrix analysis.
\end{enumerate}
The parameter fitting method uses a $\chi^2$ analysis on mock standard candle data and tests the likelihood of the data given the model. It is the more accurate of the two methods, but can be computationally time consuming. The Fisher matrix method is a second order Gaussian likelihood estimation that is very popular in predicting the constraints on various cosmological parameters due to its simplicity and speed \citep[e.g.][]{Albrecht2006,Bassett2011,More2012}; however, it fails when the errors in the parameter space are non-Gaussian, which is common for individual cosmological probes. Nonetheless, Fisher predictions are reasonably reliable for large numbers of standard candle measurements and for combinations of probes, because as the constraints become tighter, the uncertainties become more Gaussian.
\subsection{Parameter fitting ($\chi^2$)}
Our first method of analysis is likelihood testing using real data from SN, BAO, and CMB measurements and mock catalogues of standard candle data and future SN and BAO data.
The likelihood that the data are consistent with the model is $\mathcal{L} \propto \exp[-\chi^2/2 ]$.
We explore the parameter space using either a grid approach, for simple models (such as the linear dark energy equation-of-state parametrization), or Markov Chain Monte Carlo (MCMC) analysis, for models with many parameters (such as the piecewise parametrization). We restrict our parameter space to $50.0\leq H_0\leq100.0$ and $0.15\leq\Omega_m\leq0.4$, where $H_0$ is the present day Hubble constant, and $\Omega_m$ is the present matter density fraction. {Independent measurements of $\Omega_m$ \citep[e.g.][]{Samushia2013} and $H_0$ \citep[e.g.][]{Sandage2006,Riess2011} are consistent with these parameter ranges. For each trial cosmology in our parameter space, picked by MCMC or grid method, we calculate the $\chi^2$ value,
\begin{equation}
\chi^2(\mathcal{P}_{\rm mod}) = \sum_{ij} \left[x_i^m(\mathcal{P}_{\rm mod})-x_i^d\right] C_{ij}^{-1}\left[ x_j^m(\mathcal{P}_{\rm mod})-x_j^d\right],
\label{eq:chi2}
\end{equation}
where $x_i^m(\mathcal{P}_{\rm mod})$ is the predicted observable given the model parameters, $x_i^d$ is the observed value, and $C_{ij}^{-1}$ is the inverse covariance matrix of the observable $x^d$. If the measurements of $x^d$ are independent then the covariance matrix is diagonal, such that $C_{ii} = \sigma_i^2$ where $\sigma_i$ is the uncertainty in the $x_i^d$ measurement.
If the observable has the form, $x = f(\mathcal{P})+K$, where $K$ is a constant, and if no prior knowledge of $K$ is assumed at all, we can analytically marginalize over this constant nuisance parameter ($K\in [-\infty,\infty]$). The revised $\chi^2$ equation used for this purpose is given in Appendix \ref{sec:fishermarg}.
\subsection{Fisher Matrices}
The Fisher matrix method is a method of predicting constraints on your parameter space without real or simulated data. It is a second order approximation of the likelihood. It assumes Gaussian uncertainty on the parameters being fit, as opposed to only Gaussian uncertainties on the measured quantities as in the previous analysis. Using a fiducial model and expected measurement uncertainties, the likelihood in the nearby parameter space is predicted.
The Fisher matrix is defined such that its inverse is the covariance matrix
\begin{equation}
[\mathcal{F}]^{-1} = [{C}] = \left[ \begin{array}{cc}
\sigma_{\alpha}^2 & \sigma_{\alpha\beta} \\
\sigma_{\alpha\beta} & \sigma_{\beta}^2 \end{array} \right],
\label{eq:covmatrix}
\end{equation}
where $\sigma_{\alpha}$ is the uncertainty associated with an arbitrary parameter $\lambda_{\alpha}$, and $\sigma_{\alpha\beta} = \rho\sigma_{\alpha}\sigma_{\beta}$ where $\rho$ is known as the correlation coefficient, which varies from 0 (independent) to 1 (completely correlated). Once the Fisher matrix is known, the inverse gives the best possible constraints that we can derive for the parameters given the observed data. According to the Cramer-Rao inequality, the Fisher matrix gives a lower bound on the parameter uncertainty $\sigma$ in parameter $\lambda_{\alpha}$,
\begin{equation}
\sigma_{{\alpha}} \geq \sqrt{\left(F^{-1}\right)_{\alpha\alpha}}.
\end{equation}
For $N$ model parameters $\lambda_{\alpha},\lambda_{\beta},\ldots, \lambda_N$, the Fisher matrix, $F$, is an $N\times N$ symmetric matrix. Consider $b$ observables, $f_1,f_2,\ldots,f_b$ (such as $\mu$), with which you attempt to measure cosmological parameters, $\lambda_n$ (such as $\Omega_m, \Omega_{\Lambda}, w$), where each observable is related to the model parameters by some function, e.g. $\mu = \mu(\Omega_m,\Omega_{\Lambda}, \ldots)$. Then the elements of the Fisher matrix are given by,
\begin{equation}
F_{\alpha\beta} = \sum_b \frac{1}{\sigma_{b}^2} \frac{\partial f_b}{\partial \lambda_{\alpha}}\frac{\partial f_b}{\partial \lambda_{\beta}},
\label{eq:genfisher}
\end{equation}
where each element is summed over the observables. The derivatives of the parameters required for the Fisher matrix analysis are given in Appendix~\ref{sec:fishercalc}.
To marginalize over any parameter, the Fisher matrix is inverted, then the associated rows and columns for that parameter are removed from the matrix, and the inverted Fisher matrix is once again inverted to give the revised Fisher matrix.
Useful formulae for performing this marginalization stably are described in the appendix of \citet{Matsubara2004}.
Computationally Fisher matrices are much simpler than performing the full likelihood analysis and is therefore common practice when forecasting the precision of a future survey \citep{Albrecht2006, Eisenstein2011}.
\subsection{Quality Measures of Constraining Power}
To quantify the improvement achieved with the addition of a new HzSC, we calculate the predicted change in the constraints of various dark energy parameters. These include the Hubble parameter in several redshift bins $H(z_i)$, a piecewise fit to the dark energy equation-of-state $w(z_i)$, and a linear parametrization of a time-varying dark energy equation-of-state, $w(z)=w_0+w_zz/(1+z)$ .
In the latter case, we also consider the figure of merit (FoM) suggested by the Dark Energy Task Force \citep[DETF; ][]{Albrecht2006}, given by the inverse of the area within the 95 per cent confidence level ($2\sigma$) contours of the parameters $w_0$ and $w_z$,
\begin{equation}
{\rm FoM} = \frac{1}{\Delta\chi^2\pi\sqrt{\det {\rm Cov}(w_0,w_z)}},
\label{eq:FoMproper}
\end{equation}
where $\Delta\chi^2 = 6.17$ for two parameters.\footnote{In general, if the likelihood surfaces for all the parameters are Gaussian, any $N$-dimensional volume is proportional to the square root of the determinant of the covariance matrix of $\{\lambda_i\}$, $\sqrt{\det {\rm Cov}(\lambda_1,\lambda_2,\ldots)}$. For $N=2$, the 1$\sigma$ or 2$\sigma$ confidence level contours of the parameters $\lambda_1$ and $\lambda_2$ are ellipses with the enclosed area given by $\Delta\chi^2\pi\sqrt{\det {\rm Cov}(\lambda_1,\lambda_2)}$, where $\Delta\chi^2$ given by 2.30 or 6.17, respectively \citep{Wang2008}. } Despite equation \ref{eq:FoMproper} being the definition of the FoM that appears in the DETF report, the more recognizable form of the DETF FoM is given by the expression $[\sigma~w_p\times\sigma~w_a]^{-1}$, which is equivalent to
\begin{equation}
[\sigma~w_p\times\sigma~w_a]^{-1}= \frac{1}{\sqrt{\det {\rm Cov}(w_0,w_z)}} \approx 19.38\times {\rm FoM},
\end{equation}
where $w_p=w(z_p)$ is the dark energy equation-of-state value at the pivot redshift, and the pivot redshift is the redshift at which $w(z)$ has the smallest uncertainty. The transformation between ($w_0,~w_z$) coordinates to ($w_p,~w_a$) coordinates is linear and the Fisher matrix in the ($w_p,~w_a$) variables is $F'=M^TFM$, with $\det M=1$. It follows that the error ellipse in the $w_p- w_a$ plane has the same area as the equivalent ellipse in the $w_0- w_z$ plane.
We show both values of FoM in our results for simplicity. Throughout our analysis, we assume a flat universe. As a consequence, our results are not directly comparable with those from the DETF, who allowed for curvature. We only consider a flat universe due to the strength of the current constraints on the curvature given by CMB measurements \citep{Planck2013}.
\section{Observables and Data}\label{sec:probes}
Standard candle measurements provide us with the luminosity distance, $D_L$, through the relationship between measured flux, $F$, and intrinsic luminosity. Because $F$ is measured without an absolute luminosity calibration, what is actually calculated from the observables is the distance modulus,
\begin{equation}
\mu = m-M = 5 \log[D_L(\mathcal{P},z)] + \mathcal{M}
\label{eq:distmod}
\end{equation}
where $\mathcal{M} = \log_{10} (c/H_0) +25$, is a constant over which we marginalize (equation \ref{eq:revisechi2} absorbs the uncertainty in the absolute magnitude, $M$, as well as the uncertainty in $H_0$), and $\mathcal{P}$ are the parameters that describe the universe and influence the expansion. The luminosity distance, $D_L$, is given by,
\begin{equation}
D_L = (1+z)D_M
\end{equation}
where $z$ is the redshift and $D_M$ is the comoving tangential distance, defined as
\begin{equation}
D_M = \frac{c}{H_0} \frac{1}{\sqrt{\Omega_k}} \sinh(\sqrt{\Omega_k}\chi) = R_0 \sinh(\sqrt{\Omega_k}\chi),
\end{equation}
where $R_0 = c/(H_0 \sqrt{\Omega_k})$ is the present day scale factor, and $\Omega_k$ is the equivalent energy density fraction of the curvature.
The dimensionless comoving distance is,
\begin{equation}
\chi(z) = \int_0^z \frac{H_0}{H(z)} dz
\end{equation}
The general form of the Hubble parameter, $H(z)$, for a Friedman-Robertson-Walker metric is given by
\begin{eqnarray}
\frac{H(z)^2}{H_0^2} = \Omega_r(1+z)^4 + {\Omega_m}(1+z)^3 + {\Omega_k}(1+z)^2 \nonumber\\ + {\Omega_x}\exp\left\{3\int_0^{1/(1+z)} \frac{dz}{1+z}\left[1+w(z)\right]\right\} ,
\label{eq:Hubblegeneral}
\end{eqnarray}
where $\Omega_r$ is the current normalized radiation energy density (including the relativistic neutrino density), and $\Omega_x$ is the current normalized dark energy density. The energy density is normalized with the critical density, such that $\Omega=\rho/\rho_c$, where $\rho_c$ is the critical energy density of the universe for which the spatial geometry is flat (or Euclidean).
\subsection{Data}\label{sec:data}
\subsubsection{Existing Data}
\paragraph*{Type Ia supernova:}
For our analysis we use the Supernova data from the \textit{SuperNova Legacy Survey} \citep[SNLS;][]{Conley2011} which is a compilation of the first three-year results from the SNLS survey with other supernova surveys (Contains: 123 low-$z$, 93 SDSS, 242 SNLS, and 14 \textit{Hubble Space Telescope} SN measurements). The details of our fitting procedure, including stretch, and colour corrections, are discussed in Appendix~\ref{sec:Snestuff}.
\paragraph*{Baryon acoustic oscillations:}
Large-scale structure measurements and consequently BAO measurements can be distilled into simple parameters that can be used to constrain cosmology. Two common parameters are used to express the cosmological information from the BAO measurement: $A(z)$ and $d_z$. The acoustic parameter, $A(z)$, was introduced by \citet{Eisenstein2005} and is given by
\begin{equation}
A(z) = \frac{D_V(z)\sqrt{\Omega_m H_0^2}}{cz},
\end{equation}
where the $D_V$ is the `dilation scale' distance,
\begin{equation}
D_V(z) = \left(D_M^2\frac{cz}{H(z)}\right)^{1/3}.
\end{equation}
The ratio of the sound horizon scale to the dilation scale was given the symbol $d_z$ by \citet{Percival2010},
\begin{equation}
d_z = \frac{r_s(z_d)}{D_V(z)},
\end{equation}
where $z_d$ is the redshift at the `baryon-drag epoch' and $r_s(z_d)$ is the comoving sound horizon at the baryon-drag epoch. The general expression for $r_s(z)$ is given by \citep{Komatsu2011}
\begin{equation}
r_s(z) = \frac{c}{\sqrt{3}}\int_0^{1/(1+z_d)} \frac{da}{a^2H(a)\sqrt{1+(3\Omega_b/4\Omega_{\gamma})a}}
\end{equation}
where $\Omega_{\gamma} = 2.469\times10^{-5}(T_{\rm eff}/2.725)^4$ is the normalized pure radiation density, $a$ is the normalized scale factor and is related to the redshift by $a^{-1}=1+z$, and $T_{\rm eff}$ is the effective temperature of the CMB. We use the fitting formula for $z_d$ defined by \citet{EisensteinHu1998}.
The values of these two parameters from \textit{Six-degree-field Galaxy Survey} (6dFGS) \citep{Beutler2011}, SDSS \citep{Percival2010}, WiggleZ \citep{Blake2011distz}, and BOSS \citep{Anderson2012} are shown in Table~\ref{tab:acousticparameter}.
\begin{table}
\caption{The BAO distance data set from the 6dFGS, SDSS, WiggleZ and BOSS surveys. }
\begin{tabular}{cccc}
\hline
Survey & $z$ & $d_z$ &$ A(z)$ \\
\hline
6dFGS & 0.106 & $\mathbf{0.336 \pm 0.015}$ & $0.526 \pm 0.028$ \\
SDSS & 0.2 & $\mathbf{0.1905 \pm 0.0061}$ & $0.488 \pm 0.016$ \\
SDSS & 0.35 & $\mathbf{0.1097 \pm 0.0036}$ & $0.484 \pm 0.016$ \\
WiggleZ & 0.44 & $0.0916 \pm 0.0071$ & $\mathbf{0.474 \pm 0.034}$ \\
WiggleZ & 0.6 & $0.0726 \pm 0.0034$ & $\mathbf{0.442 \pm 0.020}$ \\
WiggleZ & 0.73 & $0.0592 \pm 0.0032$ & $\mathbf{0.424 \pm 0.021}$ \\
BOSS & 0.57& $\mathbf{0.0731\pm0.0018}$ &-\\
\hline \\
\end{tabular}%
\medskip \\
Notes: {Measurements of the distilled parameters $d_z$ and $A(z)$ are quoted. The values in \textbf{bold} are the values we have used in our analysis.}
\label{tab:acousticparameter}%
\end{table}%
We use the $d_z$ parameter for our analysis of the 6dfGS, SDSS and BOSS data and the $A(z)$ parameter when we are considering the WiggleZ data, corresponding to the officially released parameters in the associated papers. These parameters provide the best depiction of the BAO data in each survey. The distilled parameter, $A(z)$, is the most appropriate choice for quantifying the WiggleZ data as it is uncorrelated with $\Omega_mh^2$, but because of the shape of the clustering pattern marginalized over for the 2dfGS, SDSS, and BOSS data, the $d_z$ parameter is the best representation.
Therefore, for the $\chi^2$ analysis we define $x^d= \left[d_{0.106} [{\rm 6dfGS}]\right.$, $d_{0.2}[{\rm SDSS}]$, $d_{0.35} [{\rm SDSS}]$, $A(0.44) [{\rm WiggleZ}]$, $A(0.6) [{\rm WiggleZ}]$, $A(0.73) [{\rm WiggleZ}]$, $d_{0.57} [{\rm BOSS}]\left.\right]$ and $C_{ij}^{-1}$ is a $7\times7$ matrix made up from a combination of the individual errors from the individual measurements from 6dfGS and BOSS and the defined covariance matrices from the WiggleZ \citep{Blake2011distz} and SDSS \citep{Percival2010} data. We have not included any covariance terms between the surveys, despite the fact that the WiggleZ and SDSS surveys share a sky overlap of 500 square degree for redshift range $z < 0.5$. Given that the SDSS measurement is derived across an 8000 square degree sky area and the uncertainties in both measurements contain a significant shot noise component, the resulting covariance is negligible \citep{Blake2011distz}. We have assumed Gaussian errors in the BAO distances. Non-Gaussian tails may be non-negligible \citep{Percival2007, Percival2010, Bassett2010}, but studying their effect is beyond the scope of this paper.
The parameters we consider in the cosmological fitting are $\lambda_{\alpha}=[\Omega_m,~w,~H_0,~\Omega_b]$. We marginalize over $\Omega_b$ and $H_0$ values after the various probes are combined as the cosmological parameters do not have independent probability distributions.
\paragraph*{Cosmic microwave background (CMB):}
We included the CMB data in our cosmological fits using the \textit{Planck}
\citep{Planck2013CosmoParams} results. We use the CMB `distance priors': the shift parameter, $\mathcal{R}$, the acoustic parameter, $\ell_A$, and the redshift at the decoupling epoch, $z_{*}$. The shift parameter is given by
\begin{equation}
\mathcal{R} = \frac{\sqrt{\Omega_m H_0^2}}{c}D_M(z_{*}).
\end{equation}
The acoustic parameter is given by the expression
\begin{equation}
\ell_A = \frac{\pi D_M(z_{*})}{r_s(z_{*})}
\end{equation}
where $r_s(z_{*})$ is the sound horizon at recombination. The redshift at photon decoupling is $z_{*}$ and we implement the \citet{HuSugiyama1996} fitting formula.
The measured \textit{Planck} `distance priors' (D. Parkinson 2013, private communication) from the \citet{Planck2013CosmoParams} are $[\mathcal{R},\ell_A ,z_{*}]= [1.7440\pm0.011, 301.62\pm0.19,1090.02\pm0.42]$. These parameters capture most of the constraining power of the CMB data for the dark energy properties \citep{Komatsu2009}. \citet{WangWang2013} published values for $\mathcal{R}$ and $\ell_A$ from the Planck data but did not include a value of $z_{*}$. Their findings are consistent with the values stated here.
We again consider the parameters $\lambda_{\alpha}=[\Omega_m,~w,~H_0,~\Omega_b]$,
marginalizing over $\Omega_b$ and $H_0$ values after the various probes are combined.
\subsubsection{Mock data}
Mock catalogues were constructed to simulate future
HzSC measurements (AGN, GRB, SN II) and future SN and BAO measurements. The future SN and BAO predictions are taken from the Stage III and IV predictions from \citet{Albrecht2006}. Stage III are intermediate-scale, near-future projects, and Stage IV are large-scale, longer-term future projects. Table \ref{tab:stage3and4surveys} shows proposed Stage III and IV surveys.
Our fiducial model is set as a flat $\Lambda$ cold dark matter ($\Lambda$CDM) universe with the maximum likelihood parameters we determined from the joint SNLS3, SDSS, 6dF, WiggleZ, BOSS, and \textit{Planck} constraints, using an MCMC chain with $w=-1$:
\begin{description}
\item [$\Omega_m = 0.30$] (matter energy density),
\item [$H_0 = 69.45$] (the Hubble constant),
\item [$N_{\rm eff} = 3.04$] (effective number of neutrino-like relativistic degrees of freedom),
\item [$T_{\rm eff} = 2.7255$] (effective temperature).
\end{description}
\begin{table*}
\caption{Proposed future Stage III and IV cosmological surveys \citep{Yoo2012}. Dark energy projects are classified into four stages:
Stage I-completed projects that have already released data, Stage II-on-going projects, Stage
III-intermediate-scale, near-future projects, and Stage IV-large-scale, longer-term future
projects. }
\begin{tabular}{rp{3cm}p{3cm}p{3cm}p{3cm}}
\hline
Probes & SN Ia & CMB & BAO & WL \\
\hline
Stage III & DES, Pan-STARRS4, ALPACA, ODI & ALPACA, CCAT & DES, HETDEX, BigBOSS, ALPACA & DES, Pan-STARRS4, ALPACA, ODI \\
Stage IV & LSST, WFIRST, SNAP, JWST & EPIC, LiteBIRD, B-Pol & LSST, SKA, WFIRST, Euclid, JWST & LSST, SKA, WFIRST, Euclid \\
\hline
\end{tabular}\\
\raggedright
\underline{References:}\\
ALPACA \citep{Corasaniti2006}, BigBOSS \citep{BigBOSS}, B-Pol \citep{DeBernardis2009}, CCAT \citep{Radford2007}, DES \citep{Lin2006}, EPIC \citep{Bock2009}, Euclid \citep{Laureijs2011}, HETDEX \citep{Hill2008}, JWST \citep{JWST2009}, LiteBIRD \citep{Hazumi2011}, LSST \citep{Ivezic2008}, ODI \citep{ODI2002}, Pan-STARRS4 \citep{PanSTARRS2004}, SKA \citep{TorresRodriguez2007}, SNAP \citep{SNAP2007}, WFIRST \citep{Spergel2013}.
\label{tab:stage3and4surveys}%
\end{table*}%
\paragraph*{Mock high-$z$ standard candle measurements}
a standard candle mock catalogue is constructed by generating perfect distance modulus data according to our fiducial model, and adding random Gaussian error of the order of the predicted scatter.
We generate various mock catalogues for a number of mock surveys varying the redshift range and distribution. We assume, unless specifically mentioned, that the uncertainty in the distance modulus measurement for the standard candle is 0.2 mag, chosen following the expected achievable scatter discussed by \citet{Watson2011} for AGN measurements. We generally consider a large mock standard candle catalogue with 2000 distance measurements. This number was chosen as it is directly comparable to Stage III SN numbers for the individual predicted spectroscopic or photometric surveys. We extend this study to consider AGN distributions and realistic observational restrictions in an upcoming paper.
We have assumed independence between individual standard candle measurements. Correlations could be induced by shared peculiar velocities if close enough to be influenced by the same overdensity (e.g. galaxy cluster) or by lensing magnification if close to the same line of sight. However our HzSC candidates are typically at a high enough redshift that peculiar velocity effects are negligible and widely spread enough over the sky that nearby lines of sight are rare, therefore assuming the individual measurements are not correlated is reasonable. As a consequence, the covariance matrix $C_{ij}$ is a diagonal matrix where $C_{ii} = \sigma_{\mu_i}^2$.
\paragraph*{Future SN and BAO constraints}
the mock future SN and BAO measurements are also constructed according to our fiducial model with Gaussian scatter. The future BAO and SN predictions are taken from the Stage III and IV
predictions from \citet{Albrecht2006}.
The predicted SN measurements from DETF \citep{Albrecht2006}, are limited to $z<1.7$, but future surveys on the \textit{Hubble Space Telescope} and \textit{James Webb Space Telescope} (JWST) are now expected to observe SNe out to $z<2.5$ \citep{Grogin2011,Koekemoer2011} and possibly $z<3.5$ \citep{Hook2013}. For consistency with the DETF predictions we do not include these objects in our future SN predictions. The small number of objects they will find should be considered as part of our predictions for HzSCs. \citet{Salzano2013} have investigated how the high-$z$ SN measurements improve existing SN constraints.
The predicted measurements given by DETF \citep{Albrecht2006} break the BAO measurement into the perpendicular and transverse components rather than the previously described angle-averaged measurements ($d_z$ and $A(z)$).
The predicted DETF BAO constraints are consequently given in terms of $\log(D_M(z))$, and $\log(H(z))$. We follow this prescription in our future BAO predictions. We have used both the ground and space based predictions for the BAO and SN predictions. The specifications of the predicted Stage III and IV surveys are described in Table \ref{tab:stage3and4specs}.
We have chosen an intermediate value for the systematic error values for both the SN and BAO measurements [$\sigma_F=0.03$ (associated with photo-$z$ uncertainty),~$\sigma_{L/Q}=0.02$ (associated to the linear and quadratic components of $z$ dependent SN evolution), see \citet{Albrecht2006} for full description] such that the resulting predictions lie directly between the optimistic and pessimistic cases.
\begin{table*}
\caption{The specifications of the predicted Stage III and IV SN and BAO survey measurements.}
\begin{center}
\begin{tabular}{cccccccc}
\hline
SNe & & & & & & & \\
\hline
Stage & Type & N & Redshift Range & ${\sigma_D}^{a}$ & ${\sigma_F}^{b}$ & ${\sigma_{L/Q}}^{c}$& Expected year of Completion\\
III & Spectroscopic & 2001 & $0.01<z<1.0$ & 0.15 & 0.00 & 0.02/$\sqrt{2}$& 2017 (HETDEX) \\
III & Photometric & 2001 &$0.01<z<1.0$ & 0.12 & 0.03 & 0.02/$\sqrt{2}$& 2017 (DES)\\
IV & Spectroscopic & 2498 & $0.01<z<1.7$ & 0.10 & 0.00 & 0.02/$\sqrt{2}$& $>2020$ (SNAP, WFIRST), 2023 (JWST)\\
IV & Photometric & 191679 & $0.01<z<1.2$ & 0.10 & 0.03&0.02/$\sqrt{2}$& 2032 (LSST)\\
\hline
BAO & & & & & & \\
\hline
Stage & Type &Sky Area (deg$^2$) & Redshift Range & ${\sigma_F}^{b}$&& & Expected year of Completion \\
III & Spectroscopic & 2000 &$0.5<z<1.3$ & 0.00& & & 2014 (BOSS), 2017 (HETDEX) \\
III & Spectroscopic & 300 & $2.3<z<3.3$ & 0.00 && & 2014 (BOSS) \\
III & Photometric & 4000 & $0.5<z<1.4$ & 0.03&& & 2017 (DES) \\
IV & Spectroscopic & 20000 & $0.01<z<1.5$ & 0.00& & & 2021 (BigBOSS), $> 2024$ (SKA) \\
IV & Spectroscopic & 10000 & $0.5<z<2.0$ & 0.00 & & &2021 (JDEM), 2023 (JWST) \\
IV & Photometric & 20000 & $0.2<z<3.5$ & 0.03 && & 2032 (LSST) \\
\hline
\end{tabular}
\end{center}
\raggedright
$^{a}$ The uncertainty of the corrected apparent magnitudes due solely to variations in the properties of SNe.\\
$^{b}$ The uncertainty associated with photometrically determined redshifts.\\
$^{c}$ The uncertainty associated with any redshift dependence in the SN population. The $L$ and $Q$ stand for the linear and quadratic components of evolution.\\
\underline{References:}\\
HETDEX \citep{HETDEXdate}, DES \citep{DESdate}, SNAP/JDEM \citep{JDEM2009}, WFIRST \citep{DecadalPlan2010}\, LSST \citep{LSSTdate}, BOSS \citep{BOSSdate}, BigBOSS \citep{BigBOSSdate}, JWST \citep{JWSTdate}.
\label{tab:stage3and4specs}
\end{table*
We did not consider future CMB constraints at this point as no survey is predicted to supersede Planck, nor will we consider the constraints from WL and clustering measurements, both from redshift surveys and X-ray identification, as it is beyond the scope of this study.
\section{Fitting Parameters}
\subsection{Hubble Parameter determination}\label{sec:Hubble}
For a flat universe, by transforming the distance modulus into a comoving distance we can extract an estimate of the Hubble parameter at $z$ through numerical differentiation as
\begin{equation}
H(z) = \frac{1}{c} \left[\frac{{\rm d}D_M(z)}{{\rm d}z}\right]^{-1}.
\label{eq:piecewisehubble}
\end{equation}
This technique was proposed by \citet{Wang2005} and allows an independent determination of the Hubble parameter at different redshifts. The Hubble parameter can also be measured through various other techniques, such as BAO measurements and age--redshift relationships. Table~\ref{tab:hubble} in Appendix \ref{sec:HubbleParamMeas} shows existing measurements of the Hubble parameter and their associated techniques. For a generic dark energy density evolutions, $\rho_x(z)$, the general Hubble parameter formulation given in Equation \ref{eq:Hubblegeneral} can be simplified to take the form,
\begin{eqnarray}
\frac{H(z)^2}{H_0^2} = \Omega_r(1+z)^4 + {\Omega_m}(1+z)^3 + {\Omega_k}(1+z)^2 + {\Omega_x}\frac{\rho_x(z)}{\rho_x(0)}.\nonumber \\
{}
\end{eqnarray}
Given precise measurements of the current matter density fraction $\Omega_m$, and assuming a flat universe with a relatively negligible current radiation density fraction $\Omega_r$, the dark energy density function, $\rho_x(z)/\rho_x(0)$, can trivially be determined from $H(z)$ at low redshifts. Here, $\rho_x(0)$ is the current dark energy density fraction. We set $\Omega_m=0.261\pm0.037$ as determined from anisotropic clustering of galaxies in the CMASS DR9 \citep{Samushia2013} in combination with $H_0$ measurements \citep{Riess2011}, and the full \textit{Wilkinson Microwave Anisotropy Probe 9} (WMAP9) likelihood \citep{Hinshaw2012} for a $w_z$CDM model. For our analysis we broke the SN and HzSC measurements into evenly spaced redshift bins, to match the convention of the Stage III and IV predictions.
\subsection{Dark energy equation-of-state: $w(z)$}
All models of dark energy can be characterized by their equation-of-state $w$, which may evolve with time. Therefore, crucial information about the underlying physics of dark energy can be obtained by measuring the present value of $w$ and any time variation. The two main strategies for investigating the evolution of the dark energy equation-of-state are to (i) assume a $w(z)$ parametrization and fit to existing data or (ii) determine the value of $w(z)$ in different redshift bins, independent of a dark energy parametrization. The first method can more precisely determine the $w(z)$ behaviour if the parametrization represents the true dark energy evolution. If the dark energy behaves differently than the predicted parametrization, this approach is possibly misleading. The second approach is statistically noisy as it depends on the first and second derivatives of the distance with respect to redshift. A consequence is that the uncertainties on the measurements of $w(z)$ can become substantial. On the other hand, it does not require any a priori assumptions about the properties of the dark energy and, as such, can more easily identify exotic behaviour. We will consider both approaches in our analysis.
\subsubsection{Linear $w(z)$ parametrization}\label{sec:parameter}
We initially consider the linear parametrization of the dark energy equation of state given by the expression,
\[w(z)=w_0+w_z z/(1+z).\]
This parametrization was first proposed by \citet{Chevallier2001} and \citet{LinderCPL2003} and is commonly used throughout the literature\footnote{This parametrization is equivalent to the common $w(a)= w_0+(1-a)w_a$, where $w_z=w_a$ but we use the notation $w_z$ as we primarily refer only to redshift in our analysis.}. For a cosmological constant ($\Lambda$CDM) model, the dark energy equation of state is characterized by $w_0=-1$ and $w_z=0$. The dimensionless Hubble parameter for this parametrization is given by,
\begin{eqnarray}
\frac{H(z)^2}{H_0^2} &=& \Omega_r(1+z)^4 + {\Omega_m}(1+z)^3 + {\Omega_k}(1+z)^2 \nonumber\\
&& + \;{\Omega_x}(1+z)^{3(1+w_0+w_z)} e^{-3w_z z/(1+z)}.
\end{eqnarray}
\subsubsection{Redshift-binned piecewise $w(z)$}\label{sec:wzpiece}
Next we consider the value of $w(z)$ in different redshift bins, independent of a dark energy parametrization. Under the assumption of general relativity, the dark energy equation of state can be expressed as
\begin{equation}
w(z) = -\frac{(2/3)(1+z)(\frac{{\rm d}\chi}{{\rm d}z})^{-1}\frac{{\rm d}^2\chi}{{\rm d}z^2}-1}{1-(\frac{{\rm d}\chi}{{\rm d}z})^{2}\Omega_m (1+z)^3},
\label{eq:wzdirect}
\end{equation}
where $\chi$ is the dimensionless comoving distance defined earlier \citep{Daly2004}. Determining this directly in independent redshift bins, through numerical differentiation, as was done for the Hubble parameter in Equation \ref{eq:piecewisehubble}, is very noisy due to data limitations and the discreteness of the individual measurements. Instead, we consider $w(z)$ as a piecewise function, with a constant equation-of-state parameter within each redshift bin, and fit the parameters $w_1,w_2,\ldots, w_i$ using the Monte Carlo analysis, where $w_i$ is the dark energy equation of state corresponding to the $i$th redshift bin, $z_i$. No priors are put on the value of $w_i$.
By choosing $w(z)$ to be a piecewise constant function (or step function), rather than calculating $w(z)$ directly in each redshift bin as in Equation~\ref{eq:wzdirect}, we can fit all the data at once using MCMC and easily incorporate existing SN, BAO, and CMB data, and mock catalogues of future data. This maximizes the information that can be gleaned by the finite number of distance measurements in our samples. However, this process creates correlations in $w(z)$ between the bins, as the distance, $D_M(z)$, requires an integration over 0 to $z$. The correlations are captured by the covariance matrix, given by $[C] = \left\langle {w} {w^T} \right\rangle -\left\langle {w}\right\rangle\left\langle {w^T}\right\rangle $. To decorrelate the equation-of-state parameters, we follow the prescription of \citet{Huterer2005} and transform the $w$ chains through an orthogonal matrix rotation that diagonalizes the inverse covariance matrix. This is equivalent to applying a weighting function to the correlated $w_i$ values. The new, uncorrelated $w_i$ are given as a linear combination of the correlated $w_i$ described by the weight function.
The dimensionless Hubble parameter in this case is given by the expression
\begin{eqnarray}
\frac{H(z_{N-1}< z \leq z_N)^2}{H_0^2} = \Omega_r(1+z)^4 + {\Omega_m}(1+z)^3 + {\Omega_k}\nonumber\\ \times
(1+z)^2 + {\Omega_x}\left({1+z}\right)^{3(1+w_N)}\prod_{i=0}^{N-1} \left[{1+{\rm max}(z_i)}\right]^{3(w_i-w_{i+1})}
\end{eqnarray}
where $N$ is the redshift bin where $z$ resides and max$(z_i)$ is the maximum redshift in the $i$th redshift bin.
We divide the redshift range into the following five bins: $0.0\leq z_1<0.3$, $0.3\leq z_2<0.8$, $0.8\leq z_3<1.2$, $1.2\leq z_4<4.0$, $4.0\leq z_5$. The first four bins are constrained by SN, BAO, and HzSC measurements (roughly a low redshift bin, two medium redshift bins, and a high redshift bin), and the highest redshift bin is constrained entirely by CMB measurements, and therefore is largely uncorrelated with the lower redshift bins (i.e., there is no contribution from preceding bins in the weighting function of bin 5). We assume a flat universe and evaluate $w_i$ in each redshift bin $i$ by fitting the data.
\section{Results: Constraints from High-Redshift Standard Candles}\label{sec:effect}
In this section, we quantify the power of standard candle measurements for constraining dark energy properties using the analysis methods described above. Our primary concern is the optimal redshift range for future standard candle measurements.
Defining the optimal redshift range is difficult, as it will depend on the model of $w$ being tested. The redshift range that best constrains a constant $w$ will not be the same as that which best constrains a variable $w$. In this section we investigate this multidimensional question and discuss several aspects of the result.
Realistically, the number density of standard candles and the observing capabilities and strategy of a survey will set the number, the measurement uncertainties, and the redshift distribution of standard candle measurements. We consider the effect of observational restrictions on our cosmological predictions, for the specific case of AGN as our standard candle in a forthcoming paper. For the time being, we consider uniform and non-uniform distributions of a general standard candle, spanning a range of redshifts.
For most of our analysis, we keep the number and scatter in our mock HzSC measurements fixed and alter only the redshift distribution. Greater numbers and higher precision will both give monotonic advantages in constraining dark energy. This is obvious from the role of $\sigma$ in Equations (\ref{eq:chi2}) and (\ref{eq:genfisher}). If we simply assume the main sources of scatter in the standard candle are observational (statistical), rather than intrinsic (systematic), we can consider the improvement in parameter constraints as a trade-off between the number of measurements and the precision of the measurements. The resulting constraints follow the general relationship $\sigma^2_{\alpha} \propto \sigma_\mu^2/N$, where ${\alpha}$ can represent ($\Omega_m,~\Omega_x,~w,~\ldots$) and $\sigma_\mu$ is the uncertainty in the distance modulus.
We illustrate this in Fig.~\ref{fig:sigmanumplot} for a flat $w_z$CDM parametrization and an optimal double Gaussian distribution of standard candle measurements, which we discuss in the next subsection.
\begin{figure*}
\includegraphics[width=1.0\textwidth]{sigmanumplotlandscape.eps}
\caption{A representation of the trade-off between number of standard candle measurements and the uncertainty in the measurements that dictate the overall dark energy equation-of-state constraints. Each contour represents a constant absolute uncertainty in $w_0$ (left) or $w_z$ (middle), or the FoM (right).
These constraints are calculated with the Fisher matrix methods using the current SN, CMB, and BAO measurements plus additional $N$ standard candles with a double Gaussian distribution with $(\bar{z}_1,\bar{z}_2,\Sigma_1,\Sigma_2)=(0.0,2.0,0.25,0.25)$.
The constraints follow the general relationship $\sigma^2_{w_0,w_z} [1/{\rm FoM}] \propto \sigma_\mu^2/N$.}
\label{fig:sigmanumplot}
\end{figure*}
\subsection{Parameterized Models}\label{sec:linearparamsection}
We consider a CPL dark energy parametrisation, as discussed in Section \ref{sec:parameter}, and initially investigate how the addition of an HzSC affects the likelihood constraints on a flat $w_z$CDM universe.
We combine the constraints from 2000 mock standard candle measurements, uniformly distributed over a large redshift range $0.01\leq z\leq 4.0$, with existing and predicted future SN, BAO, and CMB constraints. The resulting $\Omega_m-w_0$ and $w_0-w_z$ confidence contours are shown in Fig~\ref{fig:contours}. With respect to the current constraints, the introduction of the HzSC measurements makes a marked impact on the precision with which we can determine the matter density and equation-of-state parameters. This improvement is mainly attributed to the large number of measurements (despite the lower precision compared to SNe); however, it is also influenced by the large redshift range of the mock standard candle measurements. Having a larger redshift range predominantly reduces the uncertainty in the $w_z$ parameter. However, when the HzSCs are combined with the predicted future constraints, we find a smaller effect. This is because of the large number and higher precision of predicted future SN data (compared to our mock HzSCs) and the precision and wide redshift range of the predicted future BAO data.
In the $\Omega_m-w_0$ plane, the tilt of the contours is affected by the inclusion of the HzSCs, due to some orthogonality between the probes. As a result, the improvement is more distinguishable than in $w_0-w_z$, where orthogonality is weak. Consequently, it is difficult to break degeneracies in the equation-of-state parameter using distance probes such as standard candles and standard rulers alone. This degeneracy occurs because $w_0$ does not appear independently in the expression for Hubble parameter. We therefore find that 2000 additional standard candle measurements at high redshift will not improve constraints on this model compared to Stage IV probes, which is not surprising given the relative number of SN and BAO measurements in Stage IV. However, we will see in Section \ref{sec:piecewise} that improvements \textit{are} gained when considering more flexible models of dark energy.
\begin{figure*}
\subfloat{ \includegraphics[width=0.45\textwidth]{agn_omw0all}}
\subfloat{ \includegraphics[width=0.45\textwidth]{agn_omw0waall}}
\caption{The 1$\sigma$ and 2$\sigma$ level confidence contours of the cosmological parameters in the $\Omega_m-w_0$ plane (left) and $w_0-w_z$ plane (right), calculated using Fisher matrix methods. The black contours show the current only SN, BAO, and CMB constraints, the blue/purple contours show the predicted future constraints from Stage III (middle plot) and Stage IV (bottom plot) SN, BAO, and CMB data, and the yellow/gold contours in each panel are the combination of those constraints with HzSCs. HzSCs provide significant improvement over current constraints, and are competitive with Stage III probes.}
\label{fig:contours}
\end{figure*}
This initial investigation is not realistic in the sense that it would be infeasible to carry out a HzSC survey with uniform redshift sampling and range for this large number of objects, simply because of the realistic number density and redshift distribution of AGNs, SNe II, or GRBs. It also gives us very little information about which redshifts are important for dark energy investigations. In order to inform future HzSC surveys of a more optimal, and possibly more realistic, survey strategy, we investigate the constraints on dark energy parameters gained by considering various distributions of HzSCs. To do this, we consider: (1) a uniform redshift distribution and alter its (a) maximum and (b) minimum redshift cut-offs and (2) a redshift distribution described by a Gaussian function, where the mean and range of redshifts probed is varied by altering the mean and width of the Gaussian function.
\subsubsection{Maximum redshift cut-off (Case 1a)}\label{sec:zmax}
We set the minimum redshift of our 2000 HzSC measurements to $z_{\rm min}=0.01$ and varied the maximum redshift, $z_{\rm max}$, within the range $0.1\leq z_{\rm max}\leq4.0$. The HzSC measurements were uniformly distributed in redshift from $z_{\rm min}$ to $z_{\rm max}$.
For each redshift configuration, we calculate the individual $w_0$ and $w_z$ constraints as well as the FoM. The resulting constraints are shown by the red solid curves in Fig.~\ref{fig:zFoMcombined}.
In the `Current' case, where the HzSC measurements are combined with the current data, the constraints initially become stronger as the redshift range increases, but there is a maximum in the FoM at $z_{\rm max}\sim 2$, and beyond this point the constraints weaken. The constraints initially strengthen with redshift as time-evolving behaviour in $w(z)$ becomes easier to identify, and $\Omega_m$ constraints tighten with an extended redshift range. The turnover is due to a combination of two effects: (1) a uniform HzSC distribution leads to a relative decrease in the number density of low-$z$ measurements (where dark energy is dominant) as the redshift range is extended; (2) by $z\sim2$ the energy density of dark energy in the fiducial model is an order of magnitude smaller than the matter energy density, so its influence on the expansion (and measured luminosity distances) is minimal compared to that of the matter density. By increasing the $z_{\rm max}$ beyond this point, high sensitivity is required to obtain additional information about the dark energy parameters. Overall, the improvement gained with the addition of the HzSCs to Current cosmological probes is primarily due to the large increase in the total number of distance measurements, but the extended redshift range of the HzSCs also reduces degeneracies between $w_0$ and $w_z$.
The resulting constraints are therefore very dependent on the redshift distribution of HzSCs. It should be noted here that the CPL dark energy parametrization we investigate was expressly designed as a probe of low-$z$ dynamics, so our results are to some extent a consequence of the parametrization choice.
Once Stage III observations have been completed, SNe and BAO will be competitive with the HzSCs we have simulated here. At that point, the orientation of the constraints, in the $w_0$-$w_z$ parameter space, start to play a larger role. Different redshift ranges rotate the degeneracy direction in the $w_0$--$w_z$ plane.
In this case, the FoM no longer experiences a turn over and continues to improve with higher $z_{\rm max}$ values. This is a consequence of the other distance probes (SNe and BAO) contributing mostly only at relatively lower redshifts and supports the need for HzSCs, which now complement their lower $z$ counterparts by adding information about the behaviour of dark energy at high $z$.
By Stage IV, the constraints are already so strong that adding HzSCs gives negligible improvement in the $w_0$ and $w_z$ constraints. Nonetheless, Fig. \ref{fig:zFoMcombined} shows that the FoM is still improved with the addition of HzSCs at Stage IV because of an increased correlation between the two equation-of-state parameters, thus decreasing the area of the $w_0$--$w_z$ ellipse without significantly reducing its extent in either parameter.
Despite the fact that 2000 HzSC measurements, with our prescribed level of measurement uncertainty, only provide a slight improvement on the combined Stage III and Stage IV measurements in the $w_z$ model, we find that 2000 HzSC measurements are overall competitive as individual probes compared to the individual predicted future SN and BAO measurements.
Fig.~\ref{fig:zFoMcombinedcurrentfuture} shows the relative predicted improvement over the current FoM with the individual addition of HzSC measurements, Stage III- and Stage IV- SN and BAO constraints. The HzSC constraints are roughly equivalent to or greater than the predicted Stage III constraints, but to be able to compete with or surpass the Stage IV measurements a large number ($n>2000$) or more precise measurements ($\sigma_{\mu}< 0.2$ mag we assumed here) are still required.
\begin{figure*}
\includegraphics[width=1.\textwidth]{combined3x3}
\caption{The improvement 2000 standard candles would contribute relative to the dark energy equation-of-state parameter baseline constraints of the combined SN, BAO, and CMB measurements (black curves) based on current data (left-hand plot) and future Stage III (middle plot) and Stage IV (right-hand plot) data, in a flat $w_z$CDM cosmology. The predicted improvement is shown for the three different HzSC redshift distributions discussed in Section \ref{sec:linearparamsection}. The red (solid) [purple (dashed)] curves show results for Case 1a (1b) where the maximum (minimum) redshift is varied for a uniform distribution of standard candles with a fixed minimum [maximum] redshift at $z=0.01$ ($z=4.0$). The green (dot-dashed) curves show results for Case 2, where the mean redshift is varied for a Gaussian distribution of standard candles with width $\Sigma_z = 1.0$. Higher FoM values indicate stronger constraints.}
\label{fig:zFoMcombined}
\end{figure*}
We note also that while we have only shown how HzSCs might improve constraints on $w_0$ and $w_z$ for a flat $w_z$CDM universe model, we also forecasted the effect on other cosmological parameters. The magnitude of the predicted constraints and the behaviour as a function of $z_{\rm max}$ depends on the cosmological parameter of interest and the parametrization tested. For example, the density parameters ($\Omega_m$, $\Omega_x$) always prefer a long redshift range, which is contrary to what is observed in Fig. \ref{fig:zFoMcombined} for the $w_0$ and $w_z$ parameters. The complexity of the universe model assumed (e.g., $\Lambda$CDM, $w$CDM, $w_z$CDM, or different flavours of $w_z$CDM) also affects the predicted constraints and changes the optimal redshift distribution. Allowing curvature to vary weakens the strength of the dark energy equations-of-state constraints slightly and tends to shift the optimal $z_{\rm max}$ value for $w_0$ and $w_z$ to a lower redshift. The degradation of the constraints with curvature is expected as there is a well-known degeneracy between dark energy and curvature for purely geometric probes like standard candles \citep{Linder2005,Knox2006,Huang2007,Hlozek2008}.
\begin{figure}
\includegraphics{CombinedFoMcurrentfuture}
\caption{The predicted improvement of the FoM in the $w_0-w_z$ plane in a flat $w_z$CDM model compared to the current constraints. The non-dotted lines represent the predicted improvement from two uniform distributions with variable maximum (red solid) and minimum (purple dashed) cut-off redshift and one Gaussian distribution with a variable mean redshift (green dot-dashed). These representations are consistent with Fig. \ref{fig:zFoMcombined}. The other lines represent the predicted improvement on the current constraints gained with the different stage future experiments. A 2000 large HzSC survey can marginally compete with the Stage IIIs SNe and BAO but cannot compete with either Stage IV SN or BAO predicted constraints.}
\label{fig:zFoMcombinedcurrentfuture}
\end{figure}
\subsubsection{Minimum redshift cut-off (Case 1b)}\label{sec:zmin}
Next we consider the effect of changing the minimum redshift. We set the maximum redshift to $z_{\rm max}=4.0$ and vary the minimum redshift within the range $0.01\leq z_{\rm min}\leq3.9$.
The purple dashed curves in Fig.~\ref{fig:zFoMcombined} show that all constraints are maximized when the redshift distribution of HzSCs extends to $z=0$. The loss of cosmological information as low-redshift data are removed arises
not only because dark energy is dominant at low redshifts but also due to uncertainties in the Hubble constant and in the absence of absolute luminosity calibration. In this absence, cosmological information is gained from the shape of the observed Hubble diagram (${\rm d}\mu(z)/{\rm d}z$), rather than the absolute value of $\mu(z)$. Therefore, long redshift ranges are preferred, and high- and low-redshift standard candle information must be on the same absolute magnitude scale in order to robustly probe the evolution of the expansion and minimize systematic uncertainties as a function of $z$.
As a consequence, in the fiducial cosmology, low-redshift standard candle measurements are just as, or more, important as their high-redshift counterparts.
We further illustrate the dependence of the constraints on the redshift in Fig.~\ref{fig:agnomw0waagnfisherchi} by showing how the standard candle-only likelihood contours, in the $w_0$--$w_z$ plane, change with redshift range. We consider three redshift distributions:
\begin{enumerate}
\item{$0.01\leq z\leq 1.0$, }
\item{$0.3\leq z\leq 4.0$, }
\item{$0.01\leq z\leq 4.0$.}
\end{enumerate}
When only low redshifts are probed (case i), the $w_0$ parameter is well constrained, but the $w_z$ parameter is only weakly constrained. At low redshifts, the dynamics of dark energy and consequently the expansion, are predominantly controlled by the value of $w_0$. On the other hand, probing only high redshift information (case ii) may be more sensitive to $w_z$, but without low redshift constraints, a degeneracy arises between the two equation-of-state parameters and two very different $w(z)$ models may be hard to distinguish. A visual example of this situation is illustrated in Fig.~\ref{fig:distmods} where different models with curves of the same colour appear almost identical at high redshift.
Instead, it is optimal to probe a broad redshift range (case iii) to gain tight constraints on $w_0$ while restricting the possible values of $w_z$.
Fig.~\ref{fig:agnomw0waagnfisherchi} was created using the $\chi^2$ analysis instead of the Fisher matrix analysis, as the Fisher matrix analysis is generally not a good representation of the likelihood of individual probes. See Appendix \ref{sec:fishercomp} for more discussion on this point. The presence of the bend or kink in the contours is primarily due to uncertainties in the matter density \citep{Goliath2001,Wolz2012}, which cause degeneracies in the $w_0$ and $w_z$ plane. When standard candle measurements are combined with current CMB and BAO data this uncertainty diminishes significantly, and the bend disappears.
\begin{figure}
\includegraphics{agn_omw0waagnfisherchi.eps}
\caption{ The 1$\sigma$, 2$\sigma$ and 3$\sigma$ level confidence contours for three different standard candle survey regimes: (a) $0.01\leq z\leq 1.0$ (dot-dash green), (b) $0.3\leq z\leq 4.0$ (dash red), and (c) $0.01\leq z\leq 4.0$ (solid black). Current SN and BAO measurements are restricted to $z\lesssim1$. In this regime, $w_0$ is well constrained, but to constrain $w_z$, we clearly require both high- and low- redshift measurements. The likelihoods were calculated using $\chi^2$ analysis as the HzSC-only constraints are not well approximated as Gaussian, unlike the constraints obtained from a combination of all the probes.}
\label{fig:agnomw0waagnfisherchi}
\end{figure}
\subsubsection{Single Gaussian distribution (Case 2)} \label{sec:gauss}
The two previous cases, while instructive, are over-simplified in their assumption of a uniform redshift distribution of standard candles. Indeed, some redshift ranges may have more power in terms of constraining cosmological parameters (see Fig.~\ref{fig:distmods}). Also, uniform redshift distributions are difficult to achieve in practice due to the small volumes present at low redshift and survey flux limits at larger redshifts even if survey design can attempt to mitigate these effects to some degree.
Here we consider a Gaussian redshift distribution for our mock HzSC sample.
By varying the mean redshift and width of this Gaussian we can directly probe the relative importance of different redshift regimes in constraining cosmology. We invoke a simple Gaussian distribution with a (variable) mean redshift ($\bar{z}$) and redshift spread ($\Sigma_z$), with the number density of standard candle measurements is given by:
\begin{equation}
\mathcal{N}(z)= \frac{1}{2\pi\Sigma_z^2}e^{-(z-\bar{z})^2/{2\Sigma_z^2}}.
\end{equation}
We initially investigate a constant redshift span by holding the redshift spread $\Sigma_z$ fixed at $\Sigma_z=1.0$. This is approximately consistent with a relatively deep flux-limited survey (e.g., 2SLAQ; \citet{Croom2009}). The green dot-dashed curves in Fig. \ref{fig:zFoMcombined} show the parameter constraints we compute as a function of $\bar{z}$. The distributions are truncated at zero to avoid unrealistic (negative) redshifts, but the total number of HzSC measurements always remains constant.
In the Current case, the constraints from a Gaussian redshift distribution survey with $\bar{z} =0$, are optimal and the resulting constraints are directly comparable to those found in the uniformly distributed case with redshift within the range $0.01<z<2.0$. This once again solidifies the importance of low-redshift measurements.
In the Stage IV case, the FoM constraints gained from a Gaussian distribution of $\Sigma_z=1$ are no longer competitive with the strongest uniform distribution configurations, and while there is still a preference for a low mean redshift it is not as pronounced as it was in the Current and Stage III cases. The low redshift regime is well constrained by the future SN and BAO measurements, in this case, and broadening the redshift range is now the more efficient approach to constrain dark energy.
When the range of redshifts probed, $\Sigma_z$, is also allowed to vary, the constraints do not simply tighten for a decreasing $\bar{z}$ and increasing $\Sigma_z$ as one may naively expect. Fig.~\ref{fig:gaussandwidth} shows that the optimal value of $\bar{z}$ is marginally dependent on our choice of $\Sigma_z$. For a narrow survey (i.e. small $\Sigma_z$) a small but non-zero $\bar{z}$ is optimal because there are two opposing influences at play: the first, is the preference for low redshift measurements, and the second, is the preference for a larger redshift range. The relative power of these two preferences depends on the priors imposed. For the current case, where cosmological constraints are relatively weak, only an HzSC survey narrower than $\Sigma_z<0.5$ will prefer a non-zero $\bar{z}$, while in the Stage III case, the upper width threshold is closer to $\Sigma_z<0.7$ for a non-zero mean. As our constraints improve with the anticipated Stage III and Stage IV measurements, the preference for a wide redshift range dominates over the need for low redshift measurements.
\paragraph{Double Gaussian:}
while the global maximum in the FoM shown in Fig.~\ref{fig:gaussandwidth} is at low redshifts it is interesting to note that there is also a small local maximum or plateau in the Current constraints (shown in the inset of Fig.~\ref{fig:gaussandwidth}).
Therefore, information about the dark energy equation-of-state parameter (in the linear parametrization and fiducial model) can be gained from high-redshift data. Due to this, we consider a double Gaussian redshift distribution, $\mathcal{N}(\bar{z}_1,\bar{z}_2,\Sigma_{z(1)},\Sigma_{z(2)})$, which probes both the high- and low-redshift regimes. As low redshift data has been found to be of importance, we set one Gaussian distribution to be stationary at a mean redshift of $\bar{z}_1=0.0$ with a spread in redshift of $\Sigma_{z(1)}=0.25$.
We then allow the $\bar{z}_2$ and $\Sigma_{z(2)}$ of the second Gaussian distribution to vary. Each distribution contains 1000 measurements. In Fig.~\ref{fig:doublegauss} the resultant dark energy constraints are shown: in all three cases FoM is maximized when $\Sigma_{z(2)}$ is large. The predicted Current and Stage III case FoM values are both maximized for $1\leq\bar{z}\leq2$, while the predicted Stage IV FoM value increases in a roughly monotonic fashion with $\bar{z}_2$.
We find that the constraints predicted based on this redshift distribution are marginally superior to the previous distributions tested (Sections \ref{sec:zmax} and \ref{sec:zmin}). This indicates that the optimal redshift regime for an HzSC for a flat wCDM model is something like a double Gaussian distribution.
It should be noted that having a double Gaussian distribution of HzSC measurements is different from simply using SN data at low redshift and a different standard candle at high redshift, the low- and high-redshift measurements must be calibrated to the same relative distance scale.
\subsubsection{Conclusion}
In summary, with only our current SN, BAO, and CMB constraints the most efficient way of extracting information about dark energy is to focus our observations on the low-$z$ regime but once the Stage III and IV measurements are completed, and constraints tighten, then it is more informative for HzSCs to probe over a longer redshift range.
\begin{figure}
\includegraphics[width=.45\textwidth]{gauss1x3_1}
\caption{The predicted improvement in the dark energy FoM parameter for 2000 Gaussian distributed HzSCs over the SNe+BAO+CMB baseline (black dotted curves) for Current (top), Stage III (middle), and Stage IV (bottom) constraints, in a $w_z$CDM cosmology. Coloured curves show the predictions as a function of $\bar{z}$ for a single Gaussian distribution of HzSC measurements with different values of $\Sigma_z$. The inset in the top panel shows a zoomed in section of the FoM values.}
\label{fig:gaussandwidth}
\end{figure}
\begin{figure}
\includegraphics[width=.45\textwidth]{doublegauss1x3_1}
\caption{Same as Fig.~\ref{fig:gaussandwidth} except the 2000 HzSC measurements are equally split between a stationary Gaussian distribution with $\bar{z}_1=0.0$ and $\Sigma_{z(1)}=0.25$, and a Gaussian distribution with variable redshift mean ($\bar{z}_2$) and spread ($\Sigma_{z(2)}$). Coloured curves represent the same $\Sigma_z$ magnitudes as Fig.~\ref{fig:gaussandwidth} but refer only to changes in $\Sigma_{z(2)}$.}
\label{fig:doublegauss}
\end{figure}
\subsection{Piecewise models}\label{sec:piecewise}
\begin{figure*}
\includegraphics[width=1.0\textwidth]{hubblecomboz.eps}
\caption{The relative expansion history (Top) and the dark energy density function (Bottom) found in uncorrelated redshift bins from current (blue) and future (green: Stage III, purple: Stage IV) SN (filled) and BAO (open) data, 2000 uniformly distributed simulated HzSC measurement with two redshift ranges $0.01\leq z\leq 4.0$ (gold-filled) and $2.5\leq z\leq 4.0$ (orange-filled), and other age-$z$ measurements as listed in Table~\ref{tab:hubble} (black-circle). A fiducial $\Lambda$CDM cosmology $(\Omega_{\rm m},\Omega_{\Lambda}, w)=(0.26,0.74)$ is shown by the black curve. The current data are consistent with our fiducial model. Note that the boxes for the current BAO measurements enclose the full range of redshifts included in each measurement, but the weighted mean is generally offset from the centre of the bin (e.g. the highest redshift current BAO measurement has an effective redshift of 2.3). High redshift measurements will be able to give constraints in the currently unconstrained redshift regime. }
\label{fig:hubbleconstantsne}
\end{figure*}
\subsubsection{Hubble Parameter and Dark Energy Density}
We applied the numerical derivation technique, described in Section~\ref{sec:Hubble}, to the SNLS SN data\footnote{In order to be able to extract the Hubble parameter from the SNLS analysis we set the stretch and colour parameters as constant, with values $\alpha=1.45$ and $\beta=3.16$. These values correspond to the best-fitting values found by \citet{Conley2011} for a flat $w$CDM model when only considering statistical error.},
{future mock SN measurements}, and mock HzSC measurements. The mock HzSC catalogue consists of a uniform distribution of 2000 measurements ranging over $0.01\leq z\leq4.0$.
The results are shown in the top panel of Fig.~\ref{fig:hubbleconstantsne} with Hubble parameter measurements from existing BAO data \citep{Blake2012,Chuang2012,Busca2012}, existing differential galaxy age measurements \citep{Simon2005,Stern2010,Moresco2012,Zhang2012}, and predicted future BAO data. It should be noted that when we refer to the BAO measurement here, we combine the BAO scale with the Alcock \& Paczynski effect \citep{Alcock1979}. This effect measures redshift space distortions in the shape of the correlation function or power spectrum. It is the combination of these two measurements that gives such precise $H(z)$ measurements. For more details on the existing Hubble parameter measurements see Appendix \ref{sec:HubbleParamMeas}. AGN or another HzSC and future BAO measurements will probe the high redshift regime. The fiducial 2000 HzSC measurements examined here will not be able to compete with the precision of future high-redshift BAO measurements of $H(z)$, per bin, but they can reach somewhat higher redshifts.
As described in Section~\ref{sec:Hubble}, we can also derive the redshift evolution of the dark energy density, $\rho_x(z)/\rho_x(0)$, assuming a flat universe and given a precise measurement of the matter density fraction. The bottom panel of Fig.~\ref{fig:hubbleconstantsne} shows the resulting estimates for the dark energy density function. {Current} measurements are consistent with $\rho_x(z)/\rho_x(0)=1.0$ (equivalently, $w=-1$) though the constraints are much weaker in the high redshift bins.
At present, the uncertainty in $\Omega_m$ is a dominant source of uncertainty in the dark energy density function estimation, and due to this, an increase in the precision of the Hubble parameter will only provide a relatively small improvement in the overall uncertainty levels of the dark energy density function.
Future independent measurements of $\Omega_m$ from CL and lensing will help diminish this restriction.
As the dark energy density function measurements are independent of adjacent redshift bins, and depend only on the local $\partial D_m/\partial z$ derivative, we do not require an absolute luminosity calibration or low redshift measurements to obtain high-redshift information. As a consequence we can concentrate our standard candle measurements in the high-redshift regime. For a $2.5\leq z \leq4.0$ uniform regime, we can increase the precision on the high redshift values of $\rho_x(z)/\rho_x(0)$ measurement by about 25 per cent compared to a $0.01\leq z\leq 4.0$ redshift distribution (orange compared to gold, in Fig.~\ref{fig:hubbleconstantsne}). This is simply due to an increase in the number of measurements in this redshift range.
The error bars of the dark energy density function depend on both the number of measurements in each redshift bin, as seen above, and the precision of the measurements. Once again, if we assume the main sources of scatter in the standard candle are statistical, rather than systematic, we can consider the improvement in parameter constraints as a trade-off between the number of measurements and the precision of the measurements. It follows a general $\sigma_{H(z)}^2 \propto \sigma_{\mu}^2/N$ relationship, as observed in the previous section. Therefore, to compete with the predicted future high-redshift Stage III BAO measurements for the dark energy density function, we require either $\sim20000$ HzSC measurements, with $\sigma_{\mu}=0.2$ mag or, 2000 HzSC measurements, with $\sigma_{\mu}\sim0.06$ mag.
\subsubsection{Direct determination of $w(z)$}
In Section \ref{sec:parameter} we investigated a simple parametrization of the dark energy equation-of-state and found that depending on the parameter and the complexity of the parametrization the optimal redshift distribution was variable. Presently, we do not have a strong theory about the true form of dark energy and to properly investigate the potential of HzSC data, we need to consider a more general form of the dark energy. We adopt a general piecewise step function of the equation-of-state as described in Section \ref{sec:wzpiece} for this purpose, and test how the addition of 2000 uniformly distributed HzSC measurements in the redshift range $0.01\leq z \leq 4.0$ affects the constraints on the dark energy equation of state in combination with existing SN, BAO, and CMB constraints, compared to the predicted constraints from future surveys.
\begin{figure*}
\includegraphics[width=1.0 \textwidth]{comboplotcurrent}
\caption{Uncorrelated estimates of the derived piecewise dark energy equation-of-state $w(z_i)$ from current SN, BAO, and CMB data (a). The solid black lines correspond to the median value of $w(z_i)$, the dark and light grey shaded regions correspond to the 68 and 95 percent confidence levels and the thin red line corresponds to $w(z)=-1$ ($\Lambda$CDM). This will form the base line in which we add our mock HzSC measurements, Stage III and Stage IV constraints. The coloured histograms in panel (b) show the corresponding normalized likelihood histograms for the five redshift bins we consider. Panel $(c)$ shows the weighting function which transforms the correlated $w(z_i)$ values into uncorrelated $w(z_i)$ values. The new uncorrelated $w_i$ are given as a linear combination of the correlated $w_i$ described by the weight function. There are no prior constraints on the $w_i$ values.}
\label{fig:westcurr}
\end{figure*}
\begin{figure*}
\includegraphics[width=1.0 \textwidth]{likelihoodall}
\caption{The normalized likelihoods of the uncorrelated $w(z_i)$ values for the Current (orange), $0.01<z<4$ HzSC (black), Stage III (purple), and Stage IV (green) constraints, for the four lowest redshift bins [far left: $z_i<0.3$, middle left: $0.3<z_i<0.8$, middle right: $0.8<z_i<1.2$, far right: $1.2<z_i<4.0$]. The thin red line represents $w(z)=-1$. The inset of the far right plot shows a zoomed in section of the likelihood for $w(1.2<z_i<4.0$). We do not show the $z_i>4.0$ bin because it is poorly constrained.}
\label{fig:likelihoodall}
\end{figure*}
The derived $w(z_i)$ values for the current SN, BAO and CMB measurements are shown in Fig. \ref{fig:westcurr}. This shows the current status of the field and will act as the base line to which we add our mock HzSC measurements, Stage III, and Stage IV constraints.
The lower three redshift bins ($z<1.2$) are well constrained by the current data, but beyond $z=1.2$ the constraints become much weaker. At present, only the CMB measurements and 5 SN measurements contribute to constraining the two highest redshift bins.
The current data are consistent with a cosmological constant ($w(z)=-1$).
The value of $w(1.2<z_i<4.0)$ is also consistent with $w=-1$ and its maximum likelihood values coincide with $w=-1$ (see Panel (b) of Fig. \ref{fig:westcurr} and far right plot Fig. \ref{fig:likelihoodall}), but it remains largely unconstrained for values of $w(z)<-1$. The likelihood displays an almost flat distribution tail. The non-negligible tail in the likelihood curve causes the median value for $w(1.2<z_i<4.0)$ to be significantly offset from the maximum likelihood value. To reduce the extent of this tail and make strong constraints on the value of $w(1.2<z<4.0)$ we require additional information, for example, HzSC constraints, and/or Stage III and IV constraints. We take a special interest in this redshift range for this reason. Also, despite having the largest redshift range, the last bin is not well constrained, because other parameters, such that $\Omega_m$, $H_0$ and $\Omega_bh^2$ have a more dominant effect than $w(z>4.0)$ on the observed CMB parameters. This remains true for all the cases we consider.
The weight functions for each redshift bin are shown in Panel (c) of Fig. \ref{fig:westcurr}, where the composition of each redshift bin is distinguished by a different colour (black: $0.0\leq z_1<0.3$, red: $0.3\leq z_2<0.8$, green: $0.8\leq z_3<1.2$, orange: $1.2\leq z_4<4.0$, purple: $4.0\leq z_5$). As an example, the uncorrelated value of $w_i$ for the lowest redshift bin (black) is a linear combination of $w(z<0.3)$ ($\sim85\%$ contribution), $w(0.3<z<0.8)$ ($\sim14\%$ contribution), and $w(0.8<z<1.2)$ ($\sim1\%$ contribution). The $ z_5\geq 4.0$ redshift bin (purple) is predominately constrained by CMB measurements (that is, $z>4.0$) and, as a consequence, largely uncorrelated with the lower redshift bins. This is apparent in its weighting function, which has no or little contribution from the lower redshift bins. In general our low redshift constraints agree with those found by previous authors \citep{Said2013,WangDai2013}, but beyond a redshift $z>0.8$ our constraints were found to be weaker than those found in either \citet{Said2013,WangDai2013}. In both of those studies independent $H(z)$ measurements and a prior on $H_0$ were included in their constraints, and the highest redshift bin was held constant at $w_i=-1$. These factors may explain the discrepancy with our results.
Stage III constraints show a marked improvement in the lower three redshift bins over the current constraints.
The $w(1.2<z_i<4.0)$ is also markedly improved, with the introduction of 10 high redshift BAO measurements, however, a long distribution tail is still present. This is evident in the inset of Fig. \ref{fig:likelihoodall} (Purple). As mentioned, the last redshift bin is again not well constrained, and consequently not shown in the histogram, but with the introduction of the Stage III data an upper limit on the value of $w(z_i>4.0)$ becomes apparent. The upper limit appears to be approximately $w(z_i>4.0)<0.7$ (99.99\% upper bound).
This upper limit arises because a high value of $w$ corresponds to the dark energy behaving more like matter (when $w>-1/3$ dark energy is attractive), and since we have tight constraints on the matter density and Hubble constant, the data can not accommodate more matter at early epochs. Therefore, this upper limit represents the value of $w$ for which the dark energy will have a detectable effect on the CMB measurements. However, no lower bound is expected as the lower the value of $w$, the more negligible the dark energy density is at early times (recall $\Omega(z) \propto [1 + \Delta z]^{3(1+w)}$)\footnote{This equation only holds true over a redshift range where $w$ remains constant.}, and the more negligible is the dark energy's effect on the expansion.
\begin{figure*}
\includegraphics[width=1.0 \textwidth]{westsolid}
\caption{Same as Fig. \ref{fig:westcurr} Panel (a) and (c) but for $0.01<z<4.0$ HzSC, Stage III and Stage IV constraints. All these constraints use the current SN, BAO, and CMB measurements as a baseline.}
\label{fig:westsolid}
\end{figure*}
The addition of HzSC measurements to the current constraints considerably strengthens the constraints on $w(1.2<z_i<4.0)$ (Fig. \ref{fig:westsolid}, left). A slight tail in the likelihood distribution is still present (black curve, Fig. \ref{fig:likelihoodall}), but it has a steeper decline relative to the current and Stage III constraints. As in the Stage III case, an upper limit on $w(z_i>4)$ is observed of approximately $w(z_i>4)\lesssim -0.4$ (99.99\% upper bound). The presence of this predicted upper limit suggests that both Stage III and HzSC constraints may be able to rule out early time dark energy.
With the introduction of Stage IV constraints, we see significant improvement in all bins (Fig. \ref{fig:westsolid}, right). The Stage IV measurements introduce 530 SN and 27 BAO measurements in the $1.2<z_i<4.0$ bin, distributed according to \citet{Albrecht2006}, and as a consequence can tightly constrain the value of $w(z)$ in this redshift range. It also has the additional advantage of stronger $\Omega_m,~ H_0$, and $\Omega_b$ constraints, which allow the signature of $w(z)$ to be more identifiable.
In the previous section, where we consider a piecewise Hubble parameter and dark energy density function, we introduced a sample of HzSC measurements that only occupied the high redshift regime and found a marked improvement in the high redshift constraints. We have used the same technique here with the piecewise $w(z)$ case, with uniformly distributed HzSC measurements over the redshift range $1.2<z<4$. The resulting $w(z)$ constraints are shown in Fig. \ref{fig:comboplothigh}. Naively, we would expect to see an improvement in the $w(z_i>4.0)$ constraints as we saw in the Hubble parameter, but due to the reduced redshift range, the constraints on $\Omega_m$ weaken, and correspondingly weaken the $w(z_i>1.2)$ constraints. Therefore, exclusively obtaining high redshift measurements is not beneficial for investigating dark energy, when combined with the current constraints, and a full redshift range is optimal. This is analogous to what we find in the linear $w(z)$ parametrization analysis.
This did not occur in the dark energy density parameter ($\rho_x(z)/\rho_x(0)$) estimates as the matter density was measured independently.
\begin{figure*}
\includegraphics[width=1.0 \textwidth]{comboplothigh}
\caption{Same as Fig. \ref{fig:westcurr} for the combination of current SN, BAO and CMB data with 2000 $z>1.2$ HzSC measurements. In panel (a) we have overlaid the 68 and 95 percentile constrains for the $0.01<z<4.0$ HzSC constraints (dot-dashed line). }
\label{fig:comboplothigh}
\end{figure*}
Finally, we also considered the addition of 2000 uniformly distributed $0.01<z<4.0$ HzSC measurements in combination with Stage IV constraints. The resulting $w(z_i)$ values and likelihood curves are shown in Fig. \ref{fig:comboplotfuture4agn}. With the addition of the HzSC measurements, the constraints are improved by $\sim30$ per cent. This added precision may give new insight into the nature of dark energy and help to cut down the number of allowable dark energy theories.
As we mentioned earlier, in none of the cases was the last bin well constrained. This bin is constrained solely by the CMB data.
In some previous studies this last bin was held constant at $w=-1$ \citep{Sullivan2007, Sarkar2008,Said2013,WangDai2013}, but we did not want to impose this restriction on our general $w(z)$ expression, as we did not want our data to restrict our model and hinder the revelation of surprises in $w(z)$ if they exist. Despite that, if the low redshift bins are consistent with $w(z)=-1$, then the dark energy density becomes negligible at high redshifts. As a consequence, determining $w(z)$ beyond this point will require a colossal number of precise distance measurements, and as we saw from our predictions, this may only allow us to determine an upper limit. This prediction is highly dependent on our choice of fiducial model. If we have an underlying varying $w(z)$, it may have a large effect on the expansion of the Universe in the high redshift regime and be more easily detectable, but current data do not support this hypothesis. In either case, HzSCs are valuable tools for probing these high-redshift regimes, especially in the presence of exotic forms of dark energy.
\begin{figure*}
\includegraphics[width=1.0 \textwidth]{comboplotfuture4agn}
\caption{Same as Fig. \ref{fig:westcurr} for Stage IV and $0.01<z<4.0$ HzSC constraints. In panel (a) we have overlaid the 68 and 95 percentile constrains for the Stage IV only constraints (dot-dashed line).}
\label{fig:comboplotfuture4agn}
\end{figure*}
\subsection{Large Scale HzSC Survey}
So far, we have only considered 2000 HzSC measurements. However, if given the same considerations as Stage IV measurements (e.g., time-scale, cost, and researcher hours), including a dedicated telescope and well planned observation strategy, it is not unrealistic to consider of the order of 50,000 HzSCs with $\sigma_{\mu}=0.2$ or equivalently 12,500 HzSCs with $\sigma_{\mu}=0.1$. To investigate the potential of this number of HzSC measurements, we once again consider a flat $w_z$CDM parametrization of the dark energy equation-of-state and uniformly distribute the HzSC measurements over the redshift range $0.01<z<4.0$. The resulting $w_0$--$w_z$ probability contours are shown in Fig. \ref{fig:1e5hzsc}. The constraints calculated for 50,000 HzSCs measurements (or 12,500 with $\sigma_{\mu}=0.1$) are predicted to be comfortably stronger than the predicted combined Stage IV constraints.
The likelihood of this number is strongly dependent on the observational requirements of the proposed HzSC candidate, and while these predictions do not include any systematic errors which can seriously limit the resulting constraints, HzSCs can have a huge potential for exploring the properties of dark energy given sufficient resources.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{agn_omw0wastage4}
\caption{The 1$\sigma$ and 2$\sigma$ level confidence contours in the $w_0$--$w_z$ plane, calculated using the Fisher matrix methods. The black contours show the current SN, BAO and CMB constraints only, the blue/purple contours show the predicted future constraints from Stage IV SN, BAO and CMB data, and the yellow/gold contours show the constraints from 50,000 uniformly distributed HzSCs with $\sigma_{\mu}=0.2$ or equivalently 12,500 HzSCs with $\sigma_{\mu}=0.1$, combined with the current constraints. The predicted constraints from 50,000 (12,500) HzSC measurements exceed the predicted constraints from the joint Stage IV SN and BAO measurements.}
\label{fig:1e5hzsc}
\end{figure}
\section{Discussion and Conclusions}\label{sec:discuss}
Our analysis concentrates on the cosmological constraints that can be obtained from any HzSC.
We (1) constructed mock standard candle catalogues with a range of distributions, and (2) using Fisher matrix and $\chi^2$ likelihood analyses, predicted how well HzSC measurements could further constrain the dark energy properties when combined with existing and predicted future SN, BAO, and CMB measurements, and (3) assessed the optimal distribution of HzSC measurements for this type of investigation. Determining whether HzSC measurements could constrain time-evolution in the dark energy equation of state was of primary concern. We approached time-evolution in the equation of state by considering two dark energy models: (1) a linearly parametrized form of the equation-of-state and (2) a piecewise equation-of-state.
The HzSCs show their real strength when constraining a general piecewise $w(z)$ parametrization, especially in the so far unconstrained redshift range, $1.2<z<4.0$. Even with 2000 HzSC measurements, the constraints in that range from an HzSC uniformly distributed with $0.01<z<4.0$ surpass the predicted Stage III constraints, and complement the Stage IV constraints well.
For the linear parametrization case, we generally found that measurement uncertainty, or equivalently the number, and redshift distribution of the HzSCs both have a large effect on the constraints. For a uniform distribution of HzSCs, we observed a general strengthening in the $w_0$ and $w_z$ constraints when the maximum redshift was increased, especially when combined with the predicted future SN and BAO constraints. \citet{Linder2003} found analogous results despite using an alternative dark energy model (i.e. $w =w_0 + w'z$). They also argue that having a long redshift baseline decreases the effects of measurement systematics on the $w(z)$ constraints, which we did not consider. Both ours and Linder's results are dependent on the low redshift regime being well constrained. The influence of dark energy on the expansion is greatest at low redshifts, so in the absence of low-redshift measurements, the constraining power of a standard candle is critically diminished.
We determined that a double Gaussian-like distribution was optimal for this type of investigation. This agrees with the results of \citet{Frieman2003}, who find the optimal distribution for SN measurements (when combined with CMB measurements) for constraining a linear dark energy equation-of-state is bimodal, with one population at low redshift and the other above a redshift of 1.0 (although they also use a different parametrization of the equation-of-state parameter). \citet{Frieman2003} restrict their investigation to the low-redshift regime, and did not include BAO measurement prediction. Therefore our work is able to test the optimal regime for an HzSC in the current state of affairs more robustly.
SNe have been predicted to be observable out to $z<2.5$ \citep{Grogin2011,Koekemoer2011}, but the expected number density of SN measurements beyond $z>1.5$ is low \citep{Albrecht2006, Hook2013}, and the observed scatter is expected to increase with redshift \citep{Albrecht2006,Conley2011}. On the other hand, a large number of AGN (87822 quasars over 3275 deg$^2$) have been measured between a redshift range $0.058<z<5.855$ \citep[SDSS data release 9][]{Paris2012}, making either AGN or an equivalent HzSC, potentially the superior distance probes for achieving the optimal distributions in either the uniform or double Gaussian case, and, therefore constraining $w_z$CDM.
Note that it is not sufficient to combine two different standard candles (for example, strictly HzSC measurements and low-redshift supernovae), unless the two standard candles can be placed on the same relative distance scale. The advantage of a long lever-arm in redshift is negated if the high- and low-redshift populations are distinct, because two different marginalizations over absolute magnitude are required. We gain cosmological information solely from the overall shape of the Hubble diagram and any uncertainty between the scaling of the high- and low-redshift populations affects the accuracy with which the shape can be reconstructed.
However, if we only consider a single Gaussian distribution, a low-redshift mean was preferable and in this redshift regime SNe are likely to be the superior probes of $w$ due to the small observational scatter and minimal observational requirements. In saying that, the strongest $w(z)$ constraints gained from a single Gaussian distribution were found to be weaker than the strongest constraints found for either a uniform or double Gaussian distribution.
The technological requirements and observational resources required to use AGN as cosmological distance indicators are already in place. Given access to the necessary resources, it is a real possibility for the community to obtain the prescribed number of AGN observations contemporaneously with Stage III and well before Stage IV is fully completed.
When combined with the Stage III and Stage IV measurements, 2000 HzSC measurements, with our prescribed level of measurement uncertainty, only provide a slight improvement on the $w_z$CDM constraints. However, we saw in Fig.~\ref{fig:zFoMcombinedcurrentfuture} that 2000 HzSC constraints are competitive with the individual predicted Stage III SN and BAO constraints. Although 2000 HzSC measurements can not compete with the individual Stage IV constraints, a 50,000 large HzSC survey with $\sigma_{\mu}=0.2$ (or equivalently 12,500 large HzSC survey with $\sigma_{\mu}=0.1$) can obtain significantly superior $w_0$--$w_z$ constraints than the combined SN and BAO Stage IV constraints (Fig \ref{fig:1e5hzsc}). This number is highly optimistic, but depending on the observational requirements of the proposed HzSC, it may be possible to accomplish within the Stage IV timeline. Additionally, a measurement uncertainty of $\sigma_{\mu}=0.1$ is feasible to achieve for AGN \citep[][ Kilerci-Eser et al. 2014, in preparation]{Watson2011}.
Regardless of the ability for the HzSC to independently constrain dark energy, it will nonetheless act as an independent probe with respect to all other methods, thereby providing a means to more effectively intercalibrate and evaluate systematic uncertainties across the different methods. This is not only a desired but a critical aspect of cosmological distance measurements if we are to constrain the dark energy equation of state.
Interestingly, the amplitude of the measurement uncertainties not only directly affect the constraints that an HzSC can place on all the cosmological parameters, where smaller uncertainties provide stronger constraints. However, they also influence the choice of `optimal redshift distribution' of the HzSC that can place the best constraints on the dark energy equation-of-state parameters. For example, when considering a double Gaussian distribution, once again anchoring one low-redshift Gaussian with $\bar{z}_1=0.0$ and $\Sigma_{z(1)}=0.25$, and allowing the mean redshift of the second Gaussian to vary, we find that as the scatter is reduced the constraints tighten, as expected, but we also find that the optimal mean redshift tends towards a lower value.
This suggests that we could tailor our observation strategy to the quality of the probe.
\paragraph*{Caveats:}
generally, we only consider a flat universe, and though flatness is well supported by observations \citep{Planck2013}, this assumption may influence the resulting constraints. We briefly considered the effect of loosening this assumption, and the effects varied depending on the model and parameter of interest. The behaviour of the equation-of-state parameter constraints were not hugely affected by the choice of flatness, though the optimal redshift range was altered slightly.
Throughout our investigations we assumed a $\Lambda$CDM model as our fiducial model to construct our mock catalogues of future measurements. As a consequence, the constraints we derive are affected by this choice. Also we do not consider any source of systematic errors in our predictions, but systematic errors can become a dominant limitation in dark energy investigations. We have not included systematic effects in this investigation, as presently we have insufficient knowledge of the candidate HzSCs to predict the type and magnitude of the possible systematics that may arise. As a consequence, the results of this study are limited by this omission, and represent the most optimistic case. It is crucial for future studies, in which possibly AGN, GRBs, or core collapse SNe are measured, that the systematics are fully investigated.
Gravitational lensing magnification, due to intervening structure along the line of sight, introduces scatter in our luminosity distance measurements and will have a degrading effect on the associated constraints \citep{Holz2005}.
However, rather than just being a source of noise, that adds scatter to the magnitudes, the scatter actually has a specific signature, and measuring the lensing signal in standard candles is an interesting new way to test theories of dark energy, because it measures the effect of the distribution of mass along the line of sight on the paths of the photons \citep{Smith2014}. Therefore some of the information lost due to the increase in scatter may be gained from studying the lensing signal.
It should be noted that we do not consider modified gravity models in our investigations, so whether high redshift candles are useful tools in modified gravity models depends on how far the models' predicted expansion history deviates from a flat-$\Lambda$CDM model. Other types of measurements, such as growth of structure and cosmic shear, will likely be very useful in distinguishing between such models and standard $\Lambda$CDM \citep{Cardone2012}. Coincidently, AGNs can also provide extra insight into gravity theory, as reverberation mapping can directly measure the mass of the central supermassive black hole \citep{Peterson2004, Bentz2009, Denney2010, Grier2012,Barth2013}.
\paragraph*{Conclusion:}
HzSCs can be useful in constraining cosmological models, particularly those with a temporally varying dark energy. The number and accuracy needed for standard candle measurements to be competitive with future high-redshift SN and BAO measurements will require significant, long-term observing programs. Nonetheless, seeking to obtain HzSC measurements is a worthwhile enterprise considering (1) the nature of dark energy remains unknown, so gaining additional understanding of it is a significant priority, and (2) there are added benefits of obtaining independent and complementary cosmic distance measurements as a means to further inter-calibrate and cross-check current methods. Also, by using AGN as our HzSC, we will help shed light on their nature and on galaxy -- black hole co-evolution, since using reverberation mapping methods doubles the use of these measurements, allowing for the mass of the quasar black hole to be measured as well.
In a forthcoming paper, we investigate the requirements of a realistic, competitive AGN survey more closely.
\section*{Acknowledgements}
AK would like to
acknowledge the support provided by the University of Queensland and Australian Commonwealth Government via the Australian Postgraduate Award.
TMD acknowledges the support of the Australian Research Council through a Future Fellowship, FT100100595, and the ARC Centre of Excellence for All Sky Astrophysics, CE110001020.
MV acknowledges support from a
FREJA Fellowship granted by the Dean of the Faculty of Natural Sciences at the
University of Copenhagen and a Marie Curie International Incoming Fellowship.
MV thanks the Kavli Institute for Theoretical Physics at University of California,
Santa Barbara for their hospitality while finalizing this work.
The research leading to these results has received funding from the People
Programme (Marie Curie Actions) of the European Union's Seventh Framework
Programme FP7/2007-2013/ under REA grant agreement No. 300553 (MV and KD).
KD would like to acknowledge support from the National Science Foundation through grant no. AST-1302093.
The Dark Cosmology Centre is funded by the Danish National Research Foundation. This research was supported in part by the National Science Foundation
through Grant No. NSF PHY11-25915 to the Kavli Institute for Theoretical Physics (MV).
\bibliographystyle{mn2efix}
|
\section{Brandes' algorithm}
For completeness we present the Brandes algorithm here.
\label{sec:Brandes-steps}
\begin{algorithm}
\begin{algorithmic}[1]
\STATE {\bf for} every $v \in V$ {\bf do} $\bc(v) \leftarrow 0$ \label{brandes-init}
\FOR {every $s \in V$}
\STATE run Dijkstra's SSSP from $s$ and compute $\sigma_{st}$ and $P_s(t),
\forall \ t \in V \setminus \{s\}$ \label{brandes-dijkstras}
\STATE store the explored nodes in a stack $S$ in non-increasing distance from $s$
\STATE accumulate dependency of $s$ on all $t \in V \setminus {s}$ using Algorithm~\ref{algo:accumulate} \label{brandes-accum}
\ENDFOR
\end{algorithmic}
\caption{Betweenness-centrality($G=(V,E)$) (from \cite{Brandes01})}
\label{algo:brandes}
\end{algorithm}
\REM{
\section{Missing details from Section~\ref{sec:vert-update}}
\label{app:vert-update}
The following algorithm computes the new distances and number of shortest paths after a
vertex update. This algorithm was briefly described in Section~\ref{sec:vert-update}.
\begin{algorithm}
\begin{algorithmic}[1]
\REQUIRE $E_i(v)$ with updated weights $\weight'$ \\
\hspace{0.2in} $d(s,t)$ and $\sigma_{st}$, $\forall \ s, t \in V$
\ENSURE $d'(s,v), \sigma'_{sv}, \newpaths$
\STATE $\hat{\sigma}'_{sv} \leftarrow 0$, $\sigma'_{sv} \leftarrow \sigma_{sv}$, $currdist \leftarrow d(s,v)$
\FOR {each edge $(u_i,v) \in E_i(v)$}
\IF {$currdist = d(s,u_i) + \weight'(u_i,v)$}
\STATE $\sigma'_{sv} \leftarrow \sigma'_{sv} + \sigma_{su_i}$
\STATE $\hat{\sigma}'_{sv} \leftarrow \hat{\sigma}'_{sv} + \sigma_{su_i}$
\ELSIF {$ currdist > d(s,u_i) + \weight'(u_i,v)$}
\STATE $currdist \leftarrow d(s,u_i) + \weight'(u_i,v)$
\STATE $\sigma'_{sv} \leftarrow \sigma_{su_i}$
\ENDIF
\ENDFOR
\STATE $d'(s,v) \leftarrow currdist$
\end{algorithmic}
\caption{Compute-Dist-to-v $(s, E_i(v))$}
\label{algo:compute-dist-vert}
\end{algorithm}
\section{Proof of Lemma~\ref{lem:update-rdag}}
Suppose $(a,b) \in \DAG'_R(s)$. This implies that there is at least one shortest from $s$ to $b$ in $G_R'$ that uses
the edge $(a,b)$. And similarly a shortest path from $b$ to $s$ that uses the edge $(b,a)$ in $G'$. To show that $(a,b) \in X$ we consider the possible values for $flag(b,s)$:
\begin{enumerate}
\item If $flag(b,s) = \mbox{UN-changed}$ then the set of shortest paths from $b$ to $s$ in $G$ and $G'$ are
the same, so there is a shortest path from $b$ to $s$ that uses the edge $(b,a)$ in $G$. Further, since $\DAG_R(s)$ is correctly kept, the edge $(a,b)$ is added to $X$ from Step~\ref{rdag:step1}.
\item If $flag(b,s) = \mbox{WT-changed}$ then every shortest path from $b$ to $s$ in $G'$ goes through the updated vertex $v$. This implies that any shortest path $\pi'_{bs} \in G'$ is of the form $\pi'_{bs} = b \leadsto v \leadsto s$. Similarly every shortest path from $s$ to $b$ in $G_R'$ goes through $v$ and, since $(a,b) \in \DAG'_R(s)$, at least one shortest path in $G_R'$ is of the form $s \leadsto v \leadsto a \rightarrow b$. Therefore a shortest path of the form $b \rightarrow a \leadsto v \leadsto s$ is in $G'$, and the edge $(b,a)$ must be one of the outgoing edges of $b$ in $\DAG'(b)$ that lies on a shortest path from $b$ to $v$ in $G'$. So $(a,b) \in R_b$. Thus the edge $(a,b)$ is added to $X$ by Step~\ref{rdag:step2}.
\item If $flag(b,s) = \mbox{NUM-changed}$ then, there exist at least one path from $b$ to $s$ in $G'$ that goes through $v$, and additionally there can be shortest paths from $b$ to $s$ in $G'$ that do not go through $v$. Since, $(a, b) \in \DAG'_R(s)$, the edge $(b, a)$ lies on one or both types of paths. Suppose $(b,a) \in \pi_{bs}$, which does not use $v$. Then the path $\pi_{bs}$ is a shortest
path in $G$ and hence $(a, b) \in \DAG_R(s)$. In this case, $(a,b)$ is added to $X$ by Step~\ref{rdag:step1}.
If $(a, b) \in \pi'_{bs}$ which contains $v$, then since $\pi'_{bs}$ uses the vertex $v$,
we know that $\pi'_{bs} = b \leadsto v \leadsto s$.
Hence, similar to the case (2) above, we conclude that $(a,b) \in R_b$.
Therefore the edge $(a, b)$ is added to $X$ by Step~\ref{rdag:step2}.
\end{enumerate}
Suppose edge $(a,b) \in X$. To show that $(a, b) \in \DAG'_R(s)$
we consider the different steps in Algorithm~\ref{algo:update-rdag} where $(a,b)$ can be added
to $X$:
\begin{enumerate}
\item The edge $(a,b)$ is added to $X$ by Step~\ref{rdag:step1}. This implies that the edge ${(a,b) \in \DAG_R(s)}$. Thus, there exists a shortest path in $G_R$ of the form $s \leadsto a \rightarrow b$. Therefore, there exists a shortest path in $G$ from $b$ to $s$, say $\pi_{bs} = b \rightarrow a \leadsto s$. Note that we execute Step~\ref{rdag:step1} when $flag(b,s) = \mbox{UN-changed}$ or $flag(b, s) = \mbox{NUM-changed}$.
For either value of the flag every shortest path from $b$ to $s$ in $G$ is also shortest path in $G'$.
Therefore, the path ${\pi'_{sb} = s \leadsto a \rightarrow b}$ is a shortest path in $G_R'$ and hence the edge $(a,b) \in \DAG'_R(s)$.
\item The edge $(a,b)$ is added to $X$ by Step~\ref{rdag:step2}. Thus, $(a,b) \in R_b$. This implies that the edge $(b,a)$ is on a shortest path in $G'$ from $b$ to $v$. Moreover, we add $(a,b)$ in Step~\ref{rdag:step2} when $flag(b, s) = \mbox{NUM-changed}$ or $flag(b,s) = \mbox{WT-changed}$.
Therefore, there exists at least one shortest path from $b$ to $s$ in $G'$ that goes through $v$. Since
the edge $(b,a)$ is on a shortest path from $b$ to $v$ in $G'$, the edge $(b, a)$ lies on at least one shortest path from $b$ to $s$ in $G'$.
Therefore, $(a,b)$ lies on at least one shortest path from $s$ to $b$ in $G_R'$, and this establishes that $(a,b) \in \DAG'_R(s)$.
\end{enumerate}
This completes the proof of the lemma.
\end{appendix}
\section{Cache-oblivious Implementation}
\label{sec:cache-obv}
Our incremental algorithms have efficient cache-oblivious implementaions.
We describe here an efficient cache-oblivious implementation of
Algorithm \ref{algo:update-bc} (Incremental-BC).
As noted in Section~\ref{sec:intro},
$scan(r) = r/B$ and $sort(r) = (r/B) \log_M r$ are desirable cache-efficient
bounds for a computation of size $r$.
Let $V=\{1, \ldots, n\}$.
Further, assume that for an $s, t$ pair, we have the values
$d(s,t), \sigma(s,t), flag(s,t)$ stored in
an array $\A[1..n^2]$ ordered by the first component
$s$ and then by the second component $t$ (i.e., stored as an $n\times n$
row major
matrix).
To perform Step 2 of Algorithm \ref{algo:update-bc} using a scan, we extract
the subarray $\A_{v \bullet}$
containing the entries with $v$ as the first component, and
the subarray $\A_{\bullet u}$ containing the entries with $u$ as the second component.
We scan the three arrays $\A, \A_{v \bullet}$, and $\A_{\bullet u}$
to compute the updated values $d', \sigma'$,
and $flag$ for each pair $s,t$ in the order they appear in $\A$.
These are stored in $\A'[1..n^2]$.
Overall, Step 2 incurs $O(scan(n^2))$ cache misses.
We execute the {\bf for} loop in Step~3 in increasing order of $s\in V$,
and hence we will access successive segments of $\A'$.
For Step 4, we sort the edges $(a,b)$ in $\DAG(v)$ and in
$\DAG(s)$ in nondecreasing order
of $b$. Across all $s\in V$ this can be performed in
$n \cdot sort(\mVmax)$ cache misses.
Then, Algorithm \ref{algo:update-dag} can be executed in $scan(n^2) + n \cdot scan(\mStarV)$ cache misses
across all sources since we only need $flag(s,b)$ when we examine
edge $(a,b)$ in $\DAG(s)$.
Instead of the
reverse topologically sorted order for the stack $S$ in Step 5
we use nonincreasing order of $d'(s,t)$, $t\in V$ (as in the Brandes
static algorithm). This is computed in $sort(m_s^*)$ cache misses for
source $s$ and $n \cdot sort(\mVmax)$ across all sources.
In the final step, Step~6, we execute Algorithm \ref{algo:accumulate}.
As input to this algorithm, for a given $s$, we need to generate for
every $w \in V$, the predecessor list $P_s(w)$. For this,
for every $v \in P_s(w)$, we store
the value $d'(s,v)$. This computation is done by sorting the edges $(a, b) \in DAG(s)$
by the first component $s$. This sorted list will be a subsequence of the
row for $s$ in $\A'[1..n^2]$,
and $d'(s,v)$ can be copied to each edge $(v, w)$ in $scan(m_s^{*})$ cache misses.
The $P_s(v)$ lists are then generated with another sort, and each entry
in the predecessor list will contain the associated $d'$ value.
Over all sources $s$, this computation takes $O(scan(n^2) + n \cdot sort(\mVmax))$
cache misses.
Having generated the predecessor lists, we need to execute
Algorithm \ref{algo:accumulate} for each source $s$.
The
cache-oblivious implementation of Algorithm \ref{algo:accumulate}
is somewhat different from the earlier pseudocode. We use
an optimal cache-oblivious max-priority queue $Z$ \cite{ArgeBDHM07}, that is
initially empty.
Each element in $Z$ has an ordered pair $(d'(s,v), v)$ as its key value,
and also has auxiliary data as described below.
Consider the execution of Step~4 in Algorithm \ref{algo:accumulate} for vertices
$v\in P_s(w)$.
Instead of computing the contribution of $w$ to $\delta_{s\bullet}(v)$ for each
$v\in P_s(w)$ when $w$ is processed, we insert an element into $Z$ with
key value $(d'(s,v),v)$ and auxiliary data
$(w,\sigma_{sw}, \delta_{s\bullet}(w))$.
With this scheme, entries will be extracted from $Z$ in nonincreasing values of
$d'(s,v)$, and all entries for a given $v$ will be extracted consecutively.
We compute $\delta_{s\bullet}(v)$ as these extractions occur from $Z$. Note
that the stack $S$ is needed only
to identify the sinks in $\DAG(s)$ (which will have no entry in $Z$).
So, we can dispense with $S$, and instead initially insert an
element with key value $(d'(s,t),t)$ and NIL auxiliary data for each
sink $t$ in $\DAG(s)$.
Using~\cite{ArgeBDHM07}, Step 6 takes $sort(m_s^*)$ cache misses
for source $s$, hence over all sources,
Step 6 takes $O(n \cdot sort (\mVmax))$.
Thus, the overall
cache-oblivious incremental algorithm for betweenness centrality incurs
$O(scan(n^2) + n \cdot sort(m'))$ cache misses, where $m'= \mVmax + m^*_v$.
This is a significant
improvement over any algorithm that computes shortest paths, as is the case
with the static Brandes algorithm, since all of
the known algorithms for shortest paths
face the bottleneck of unstructured accesses to adjacency lists
(see, e.g., \cite{MehlhornM02}).
\section{Incremental edge update}
\label{sec:increment}
In this section we present our algorithm to recompute BC
of all vertices in a graph $G = (V, E)$ after an incremental edge update (i.e., adding a new edge $(u, v)$ or decreasing the edge weight of an existing edge $(u,v)$). We extend this to a
vertex update in the next section. We first consider directed graphs.
Let $G' = (V, E')$ denote the graph obtained after an edge update to ${G=(V,E)}$. Let $d(s,t), \sigma_{st}$, and $\delta_{s \bullet}(t)$ denote the distance from $s$ to $t$ in $G$,
the number of shortest paths from $s$ to $t$ in $G$, and the dependency of $s$ on $t$ in $G$ respectively, and let $d'(s,t), \sigma'_{st}$, and $\delta'_{s \bullet}(t)$ denote these parameters in the graph $G'$.
Our incremental algorithm relies on maintaining the SSSP DAG rooted at every $s \in V$ after an edge update.
Let $\DAG(s)$, $\DAG'(s)$ denote the SSSP DAG rooted at each $s$ in $G$ and in $G'$ respectively.
We show how to efficiently maintain these DAGs after an update. The updated DAGs give us the
updated $P'_s(t)$ for every $s, t \in V$. We also show how to maintain for every
$s, t \in V$, $d'(s,t)$ and $\sigma'_{st}$. Then, using Algorithm~\ref{algo:accumulate},
we obtain the updated BC scores for all vertices.
We begin by making some useful observations.
\begin{lemma}
\label{lem:sets-equal}
Let $(u, v)$ denote the edge on which the weight is decreased.
Then, for any vertex $x \in V$, the set of shortest paths from $x$ to $u$
is the same in $G$ and $G'$, and we have
\[
d'(x, u) = d(x, u) \mbox { and } d'(v, x) = d(v, x) ; ~~~ \sigma'_{xu} = \sigma_{xu} \mbox{ and } \sigma'_{vx} = \sigma_{vx}
\]
\end{lemma}
\begin{proof}
Since edge weights are positive, the edge $(u,v)$ cannot lie on a shortest path to $u$ or from $v$. The lemma follows.
\qed
\end{proof}
By Lemma \ref{lem:sets-equal}, $\DAG(v) = \DAG'(v)$ after weight of $(u,v)$ is decreased.
The next lemma shows that after the weight of $(u,v)$ is decreased we
can efficiently obtain the updated values $d'(s,t)$ and $\sigma'_{st}$ for any $s,t \in V$.
\begin{lemma}
\label{lem:update-st}
Let the weight of edge $(u,v)$ be decreased to $\weight'(u, v)$, and for a
any given pair of vertices $s,t$, let
$D(s,t) = d(s, u) + \weight'(u, v) + d(v, t)$. Then,
\begin{enumerate}
\item If $d(s, t) < D(s,t)$, then
$d'(s, t) = d(s,t)$ and $\sigma'_{st} = \sigma_{st}$.
\item If $d(s, t) = D(s, t)$, then $d'(s, t) = d(s,t)$ and
$\sigma'_{st} = \sigma_{st} + (\sigma_{su} \cdot \sigma_{vt})$.
\item If $d(s, t) > D(s,t)$, then
$d'(s, t) = D(s, t)$ and $\sigma'_{st} = \sigma_{su} \cdot \sigma_{vt}$.
\end{enumerate}
\end{lemma}
\begin{proof}
Case 1 holds because the shortest path distance
from $s$ to $t$ remains unchanged and no
new shortest path is created in this case.
In case 2,
the shortest path distance from $s$ to $t$ remains unchanged, but there
are $\sigma_{su} \cdot \sigma_{vt}$ new shortest paths from $s$ to $t$
created via edge $(u, v)$.
In case 3,
the shortest path distance from $s$ to $t$ decreases and all new
shortest paths pass through $(u,v)$.
\qed
\end{proof}
By Lemma \ref{lem:update-st}, the updated values $d'(s,t)$ and
$\sigma'_{st}$ can be computed in constant time for each pair $s,t$.
Once we have the updated $d'(\cdot)$ and $\sigma'_{(\cdot)}$ values, we
need the updated predecessors $P'_{s}(t)$ for every $s, t$ pair for
the betweenness centrality algorithm. In order to obtain these updated
predecessors efficiently,
we maintain the SSSP DAG rooted at every source $s\in V$.
The next section gives a very simple algorithm to maintain an
SSSP DAG after an incremental edge update.
\subsection{Updating an SSSP DAG}
Let $\DAG(s)$, $\DAG'(s)$ denote the single source shortest path DAG
rooted at $s$ in $G$ and $G'$ respectively.
Recall that since the update is a decrease in the edge weight of $(u, v)$,
we know that $\DAG'(v) = \DAG(v)$.
For an $s, t$ pair we define a $flag(s,t)$ to capture the change in distance/number of shortest
paths from $s$ to $t$ after the decrease in the weight of the edge $(u,v)$
as follows.
\begin{equation}
flag(s,t) =
\begin{cases}
\mbox{NUM-changed} \textrm{ \hspace{0.2in} if $d'(s,t) = d(s,t)$ and $\sigma'_{st} > \sigma_{st}$ } \\
\mbox{WT-changed} \textrm{ \hspace{0.3in} if $d'(s,t) < d(s,t)$} \\
\mbox{UN-changed} \textrm{ \hspace{0.32in} if $d'(s,t) = d(s,t)$ and $\sigma'_{st} = \sigma_{st}$ } \\
\end{cases}
\end{equation}
By Lemma \ref{lem:update-st}, $flag(s,t)$ can be computed in constant time
for each pair $s,t$.
On input $s$ and the updated edge $(u,v)$,
Algorithm~\ref{algo:update-dag} (Update-DAG) constructs a set of edges
$H$ using these $flag$ values,
together with $\DAG(s)$ and $\DAG(v)$.
We will show that $H$ contains exactly the edges in $\DAG'(s)$.
We begin by initializing $H$ to empty (Step~\ref{dag-init}). We then consider each edge $(a, b)$ in $\DAG(s)$ (in Steps~2--4) and in $\DAG(v)$ (in Steps~5--7) and
depending on the value of $flag(s,b)$ decide whether to include it in the set $H$. Finally, we check if the updated edge $(u,v)$ will be inserted in $H$ (in Steps~8--9).
It is clear from the algorithm that the time taken is linear in the size of $\DAG(s)$ and $\DAG(v)$
which is bounded by $O(m^{\ast}_s + m^{\ast}_v)$.
\begin{algorithm}
\begin{algorithmic}[1]
\REQUIRE $\textrm{DAG}(s)$, $\textrm{DAG}(v)$, and $flag(s,t), \forall t\in V$
\ENSURE An edge set $H$ after weight of edge $(u,v)$ has been decreased
\STATE $H \leftarrow \emptyset$ \label{dag-init}
\FOR {each edge $(a,b) \in \DAG(s)$ and $(a,b) \neq (u,v)$}
\IF {$flag(s,b) = \mbox{UN-changed}$ or $flag(s,b) = \mbox{NUM-changed}$}
\STATE $H \leftarrow H \cup \{(a,b)\}$
\ENDIF
\ENDFOR
\FOR {each edge $(a,b) \in \textrm{DAG}(v)$}
\IF {$flag(s,b) = \mbox{NUM-changed}$ or $flag(s,b) = \mbox{WT-changed}$}
\STATE $H \leftarrow H \cup \{(a,b)\}$
\ENDIF
\ENDFOR
\IF {$flag(s,v) = \mbox{NUM-changed}$ or $flag(s,v) = \mbox{WT-changed}$}
\STATE $H \leftarrow H \cup \{(u,v)\}$
\ENDIF
\end{algorithmic}
\caption{Update-DAG($s,(u,v)$)}
\label{algo:update-dag}
\end{algorithm}
\begin{lemma} \label{lem:dag-correct}
Let $H$ be the set of edges output by Algorithm~\ref{algo:update-dag}.
An edge $(a,b) \in H$ if and only if $(a,b) \in \DAG'(s)$.
\end{lemma}
\begin{proof}
Since the update is an incremental update on edge $(u,v)$, we note that
for any $b$, a shortest path $\pi'_{sb}$ from $s$ to $b$ in $G'$ can be of two types: \\
(i) $\pi'_{sb}$ is a shortest path in $G$.
Therefore every edge on such a path is present in $\DAG(s)$ and
each such edge is added to $H$ in Steps~2--4 of Algorithm~\ref{algo:update-dag}.\\
(ii) $\pi'_{sb}$ is not a shortest path in $G$. However, since
$\pi'_{sb}$ is a shortest path in $G'$,
therefore $\pi'_{sb}$ is of the form $s \leadsto u \rightarrow v \leadsto b$.
Since shortest paths from $s$ to $u$ in $G$ and $G'$ are unchanged
(by Lemma \ref{lem:sets-equal}), the edges in the sub-path
$s \leadsto u$ are
present in $\DAG(s)$ and they are added to $H$ in
Steps~2--4 of Algorithm~\ref{algo:update-dag}. Finally, since shortest paths from $v$ to any $b$
in $G$ and $G'$ remain unchanged, the edges in the sub-path $v \leadsto b$ are present in $\DAG(v)$ and therefore
are added to $H$ in Steps~5--7 of Algorithm~\ref{algo:update-dag}.
\qed
\end{proof}
\REM{
\begin{proof}
Suppose $(a,b) \in \DAG'(s)$.
This implies that
there is at least one shortest from $s$ to $b$ in $G'$ that uses
the edge $(a,b)$. If $(a,b)$ is not the updated edge $(u,v)$, to
show that $(a,b) \in H$ we consider the possible values for $flag(s,b)$:
\begin{enumerate}
\item If $flag(s,b) = \mbox{UN-changed}$ then the set of shortest paths from $s$ to $b$ in $G$ and $G'$ are
the same. Further, as $(a, b) \in \DAG'(s)$, it implies that
${(a,b) \in \DAG(s)}$, hence $(a,b)$ is added to $H$ by Step $4$ of the Algorithm~\ref{algo:update-dag}.
\item If $flag(s,b) = \mbox{WT-changed}$ then every shortest path from $s$ to $b$ in $G'$ uses the updated edge $(u,v)$. This implies that
any shortest path $\pi'_{sb} \in G'$ is of the form $\pi'_{sb} = s \leadsto u \rightarrow v \leadsto b$. Moreover, since $(a, b) \in \DAG'(s)$,
there exists at least one shortest path from $s$ to $b$ in $G'$ of the form ${s \leadsto u \rightarrow v \leadsto a \rightarrow b}$. Thus, the edge $(a,b) \in \DAG'(v)$.
However, we know that $\DAG'(v) = \DAG(v)$.
Hence the edge $(a,b)$ is added to $H$ by Step~7 of the Algorithm~\ref{algo:update-dag}.
\item If $flag(s,b) = \mbox{NUM-changed}$ then there is at least one shortest path
from $s$ to $b$ in $G'$ that
does not use the edge $(u,v)$ and at least one shortest path
from $s$ to $b$ in $G'$ that uses the edge $(u,v)$.
Since, $(a, b) \in \DAG'(s)$, the edge $(a, b)$ lies on
one or both types of paths.
Suppose $(a,b) \in \pi_{sb}$, which does not use $(u,v)$.
Then the path $\pi_{sb}$ is a shortest
path in $G$ and hence $(a, b) \in \DAG(s)$. In this case, $(a,b)$ is added to $H$ by Step~4 of Algorithm~\ref{algo:update-dag}.
If $(a, b) \in \pi'_{sb}$ which contains $(u,v)$, then since $\pi'_{sb}$
uses the edge $(u, v)$,
we know that $\pi'_{sb} = s \leadsto u \rightarrow v \leadsto b$.
Hence, similar to the case (2) above, we conclude that $(a, b) \in \DAG'(v)$.
This implies that $(a, b) \in \DAG(v)$. Therefore the edge $(a, b)$ is added to $H$ by
Step~7 of Algorithm~\ref{algo:update-dag}.
\end{enumerate}
Finally, if $(a,b)$ is the updated edge $(u,v)$ and $(a,b)\in \DAG'(s)$, then either $\sigma'(s,b) > \sigma(s,b)$ or
$d'(s,b) < d(s,b)$,
hence $(a,b)$ will be inserted in $H$ by Step~9.\\ \\
Suppose edge $(a,b) \in H$. To show that $(a, b) \in \DAG'(s)$
we consider the different steps in Algorithm~\ref{algo:update-dag} where $(a,b)$ can be added
to $H$:
\begin{enumerate}
\item The edge $(a,b)$ is added to $H$ by Step~4. This implies that the edge ${(a,b) \in \DAG(s)}$.
Thus, there exists a shortest path in $G$ from $s$ to $b$, say $\pi_{sb} = s \leadsto a \rightarrow b$.
Note that we execute Step~4 when $flag(s,b) = \mbox{UN-changed}$ or $flag(s, b) = \mbox{NUM-changed}$.
For either value of the flag every shortest path from $s$ to $b$ in $G$ is
also shortest path in $G'$.
Therefore, the path $\pi_{sb}$ is a shortest path in $G'$ and hence the edge $(a,b) \in \DAG'(s)$.
\item The edge $(a,b)$ is added to $H$ by Step~7. This implies that the edge ${(a,b) \in \DAG(v)}$.
Thus, there exists a shortest path in $G$ from $v$ to $b$, say $\pi_{vb} = v \leadsto a \rightarrow b$ that contains $(a,b)$.
Since decreasing the weight of the edge $(u,v)$ does not change
shortest paths from $v$ to any other vertex,
$\pi_{vb}$ continues to be a shortest path from $v$ to $b$ in $G'$.
We execute Step~7 when $flag(s, b) = \mbox{NUM-changed}$ or
$flag(s,b) = \mbox{WT-changed}$.
Therefore, there exists at least one shortest path from $s$ to $b$ in $G'$ that uses the updated edge $(u,v)$.
Hence the path
$\pi'_{sb} = \pi'_{su} \cdot (u,v) \cdot \pi_{vb} $, where $\pi'_{su}$
is a shortest path in $G'$, must be a shortest path in $G'$, and this
establishes that $(a,b) \in \DAG'(s)$.
\item The edge $(a,b)$ is added to $H$ by Step~9. This implies that $(a,b)$ is the updated edge in $G$ and there is at least a new shortest path from $s$ to $b$ going through it. Hence $(a,b) \in \DAG'(s)$.
\end{enumerate}
This completes the proof of the lemma.
\qed
\end{proof}
}
\subsection{Updating betweenness centrality scores}
In this section we present
Algorithm~\ref{algo:update-bc} (Incremental-BC), which
updates the BC scores
for all vertices after an incremental edge update.
The algorithm takes as input the graph $G = (V,E)$ and
the updated edge $(u,v)$, together with the current
values of $d(s,t)$ and $\sigma_{st}$ for all $s,t \in V$,
and the current shortest path dag $\DAG(s)$, for every $s \in V$.
We begin by initializing the BC score to $0$ for every vertex (Step~\ref{update-bc-init}).
For every pair $s,t$, we compute the updated $d'(s,t)$ and $\sigma'_{st}$ using Lemma~\ref{lem:update-st} (Step~\ref{update-bc-dist}).
Next, in Step~\ref{update-bc-dag} we compute $\DAG'(s)$
for every source $s \in V$, using Algorithm~\ref{algo:update-dag}.
Note that
this gives us the updated predecessor values $P'_s(t)$, for
every $s, t \in V$. We perform a
topological sort of $\DAG'(s)$ to obtain an ordering of $v \in V$ and
store the ordered vertices in a stack $S$.
Finally, using $S$, the $\sigma'_{st}$, and the $P'_s(t)$, we
run Brandes' accumulation of dependencies (Algorithm~\ref{algo:accumulate})
to compute the updated BC scores.
\begin{algorithm}
\begin{algorithmic}[1]
\REQUIRE updated edge $(u, v)$ with new weight $\weight'(u,v)$, \\
\hspace{0.2in} $d(s,t)$ and $\sigma_{st}$, $\forall \ s, t \in V$; $~~~ \DAG(s), \forall \ s \in V$
\ENSURE $\bc'(v)$, $\forall v\in V$ \\
\hspace{0.3in} $d'(s,t)$ and $\sigma'_{st}$ $\forall \ s, t \in V $; $~~~ \DAG'(s), \forall \ s \in V$
\STATE {\bf for} {every $v \in V$} {\bf do} $\bc'(v) \leftarrow 0$ \label{update-bc-init}
\STATE {\bf for} {every $s, t \in V$} {\bf do} compute $d'(s,t), \sigma'_{st}$, $flag(s,t)$ \hspace{0.2in} // use Lemma \ref{lem:update-st} \label{update-bc-dist}
\FOR {every $s \in V$}
\STATE Update-DAG$(s,(u,v))$ \hspace{1.7in} // use Algorithm~\ref{algo:update-dag} \label{update-bc-dag}
\STATE stack $S \leftarrow$ vertices in $V$ in a reverse topological order in $\DAG'(s)$ \label{update-bc-topo}
\STATE Accumulate-dependency$(s, S)$ \hspace{1.25in} // use Algorithm~\ref{algo:accumulate} \label{update-bc-accum}
\ENDFOR
\end{algorithmic}
\caption{Incremental-BC($G=(V,E)$)}
\label{algo:update-bc}
\end{algorithm}
The correctness of our algorithm follows from the correctness of maintaining $d(\cdot), \sigma_{(\cdot)}$, and the updated DAGs. We now bound
the time complexity of our algorithm. In
Step~\ref{update-bc-dist} of Algorithm~\ref{algo:update-bc} we spend
constant time to compute the updated $d'(s,t)$ and $\sigma'_{st}$ values
for an $s,t$ pair, hence $O(n^2)$ time for all pairs.
In Step~\ref{update-bc-dag}, we obtain $\DAG'(s)$ for each $s\in V$ using Algorithm~\ref{algo:update-dag}, which takes time $O(m^{\ast}_s + m^{\ast}_v)$ for source $s$.
This amounts to a total of
$O(m^{\ast}_v n + \sum_{x \in V} m^{\ast}_x) = O( (\mVmax + m^{\ast}_v) n)$.
Finally, Steps~\ref{update-bc-topo} and \ref{update-bc-accum} take time
linear in the size of the updated DAG. Thus the time complexity of our
Incremental-BC algorithm for an edge update is bounded by
$O( (\mVmax + m^{\ast}_v) n + n^2)$. This establishes Theorem
\ref{thm:main1-intro} for directed graphs for an update on a single edge.
We remark that both Update-DAG and Increment-BC use very
simple data structures: arrays, lists, and stacks.
Thus our algorithms are very simple to implement, and should have small
constant factors in their running time.
\paragraph{\bf Undirected graphs:}
Consider an undirected positive edge weighted graph $G = (V, E)$.
To obtain an incremental algorithm
for $G$, we first construct the corresponding directed graph
$G_D = (V, E_D)$
from $G$ by replacing every edge $\{a, b\} \in E$ by two
directed edges $(a, b)$ and $(b,a)$, both with
the weight of the undirected edge $\{a,b\}$.
Since edge weights are positive, no shortest path can contain a cycle,
hence the SSSP sub-graph rooted at $s$ in $G$ is the same as the SSSP DAG
rooted at $s$ in $G_D$.
Our incremental algorithm for $G$ for an edge update on $\{u, v\}$
is now simple: we apply two incremental edge
updates on $G_D$, one on $(u, v)$ and another on $(v, u)$.
Since the number of edges in $G_D$ is only twice
the number of edges in $G$, all other parameters also increase by only
a constant factor. This establishes Theorem \ref{thm:main1-intro} for edge updates in
undirected graphs.
\section{Introduction}\label{sec:intro}
Betweenness centrality (BC) is a widely-used measure
in the analysis of large complex networks. The BC of a node $v$
in a network is the fraction of all shortest paths in the network that go through $v$,
and this measure is often used as an index that determines the relative importance of
$v$ in the network.
Some applications of
BC
include analyzing social interaction networks \cite{KA12},
identifying lethality in biological networks \cite{PCW05},
and identifying key actors in terrorist networks \cite{Krebs02,Coffman}.
Given the changing nature of the networks under consideration,
it is desirable to have algorithms that compute
BC
faster than computing it from scratch after every change. Our main contribution
is the first incremental algorithm for computing
BC
after an incremental update on an edge or on a vertex
that is provably faster on sparse graphs than the widely used
static algorithm by Brandes~\cite{Brandes01}.
By an {\it incremental update} on an edge $(u,v)$ we mean a decrease in
the weight of an existing edge $(u,v)$, or the addition of a new edge
$(u,v)$ with finite weight if $(u,v)$ is not present in the graph;
in an incremental vertex update, updates can occur on any subset of edges
incident to $v$, including the addition of new edges.
Let $G = (V, E)$ be a
graph with positive real edge weights. Let
$n=|V|$ and $m=|E|$.
To state our result we need the following definitions.
For a vertex $x \in V$,
let $\mStarX$ denote the number
of edges that lie on shortest paths through $x$.
Let $\mVmax$ denote the average over all $\mStarX$, i.e.,
$\mVmax = \frac{1}{n} \sum_{x \in V}{\mStarX}$.
Finally, let $\mStar$ denote the total number of edges that lie on shortest paths
in $G$. For our incremental bound, we consider the maximum of each of these terms in
the two graphs before and after the update.
Here is our main result.
\begin{theorem}
\label{thm:main1-intro}
After an incremental update on a vertex $v$
in a directed or undirected
graph with positive edge weights, the
betweenness centrality of all vertices
can be recomputed in $O(m' \cdot n + n^2)$ time, where
$m'= \mVmax + m_v^{\ast}$.
\end{theorem}
Our method is to efficiently maintain the single source shortest paths
(SSSP) directed acyclic graph (DAG) rooted at each source $s\in V$, and
therefore ensure that after every change we only examine the edges
that lie on shortest path DAGs.
Since $m_v^{\ast} \leq m^*$ for all $v\in V$, and
$\mStar \le m$, the worst case time for our incremental
algorithm is bounded by $O(mn + n^2)$, which
is a $\log{n}$ factor improvement over Brandes' algorithm on
sparse graphs. Moreover, our bound in terms of $\mVmax$ and $m_v^*$ results
in much better bounds for many dense graphs.
For example, as noted in \cite{KKP93}, it
is known \cite{FG85,HZ85,LRP89} that $\mStar = O(n \log{n})$
with high probability in a complete graph where edge weights are chosen from a
large class of probability distributions, including the uniform distribution on the range $\{1, \ldots, n^2 \}$, and
our algorithm is much faster than~\cite{Brandes01}
on these graphs.
Our algorithm is very simple, especially for the edge update case,
and we present an efficient
cache-oblivious version that incurs $O(scan(n^2) + n \cdot sort(m'))$
cache misses.
Here, for a cache of size $M$ that can hold $B$ blocks, $scan(r)= r/B$ and
$sort(r)= (r/B) \cdot \log_M r$ with a tall cache ($M \geq B^2$). Both $scan$ and
$sort$ are measures of good caching performance (even though $sort(r)$
performs $r\log r$ operations, the base of $M$ in the $\log$ makes $sort(r)$
preferable to, say $r$ cache misses). In contrast, the Brandes algorithm
calls Dijkstra's algorithm, which is affected by unstructured accesses
to adjacency lists that lead to large caching costs (see, e.g.,
\cite{MehlhornM02}).
In our incremental algorithm, we assume that the BC score
for all vertices has been computed and the SSSP DAG rooted at every vertex
is available. This can be achieved
by running Brandes' algorithm.
We present an alternate algorithm based
on the Hidden Paths algorithm in Karger~et~al.~\cite{KKP93}, which computes
APSP in
$O(m^* n + n^2 \log {n})$ time.
This leads to
a static BC algorithm with the same time bound as the Hidden Paths algorithm,
which is faster than Brandes' when $\mStar = o(m)$
and $m= \omega(n \log n)$. We also note that substituting Pettie's
$O(mn + n^2 \log\log n)$~\cite{Pettie04} all-pairs shortest paths (APSP) algorithm for directed graphs
or the $O(mn \cdot \log \alpha(m,n))$~\cite{PR05}
APSP algorithm for undirected graphs (where $\alpha$ is an
inverse-Ackermann function) in place of Dijkstra's algorithm leads to
static BC algorithms with the same time complexity as the corresponding
APSP algorithm, and both improve on Brandes.
Our incremental algorithm has better bounds than any of these APSP algorithms
on sparse graphs, and is no worse (and likely better) on dense graphs.
\subsection{Related work}
The notion of betweenness centrality
was formalized by
Freeman~\cite{Freeman77} who also defined
other measures such as closeness centrality.
Approximation algorithms and parallel algorithms for
BC have been considered in \cite{BaderKMM07,GeisbergerSS08},
\cite{MadduriEJBC09} respectively.
Lee~et~al.~\cite{Lee12}
present a framework called QUBE (Quick Update of BEtweenness centrality)
which allows edges to be inserted
and deleted from the graph, and
recently, Singh~et~al.~\cite{SinghGIS13} build on the
work of Lee~et~al.~\cite{Lee12} to allow
nodes to be added and deleted.
Another recent work by Kas~et~al.\cite{KasWCC13}
deals with the problem of incremental BC and their algorithm is based on the
dynamic all pairs shortest paths algorithm by Ramalingam and Reps~\cite{RR96}.
All the above mentioned papers give encouraging experimental results,
but there are no performance guarantees.
Closest to our work is the incremental algorithm by
Green~et~al.~\cite{GreenMB12}
for
a single edge insertion in
unweighted graphs, which maintains a breadth-first search (BFS)
tree rooted at every source $s$ and identifies vertices for which
the distance (BFS level)
from $s$ has changed. However, unlike our algorithm, they do not maintain
the BFS DAG and hence
need to run a BFS for all vertices
whose distances have changed.
Thus, in the worst case, their algorithm takes
$\Theta (mn + n^2)$ time for compute the
BC of all vertices in an {\em unweighted} graph.
In contrast,
our algorithm, which takes time $O(m' \cdot n + n^2)$
even in the weighted case and even for a vertex update,
improves on the algorithm in \cite{GreenMB12} for unweighted graphs
when $m' = o(m)$.
\subsubsection{ Organization:}
Section~\ref{sec:prelims} presents
the notation we use,
and discusses Brandes' algorithm. Section~\ref{sec:increment} presents our
basic
BC algorithm for an incremental {\it edge} update, and this algorithm is
extended to a vertex update in
Section 4. Finally, in Section~\ref{sec:cache-obv} and Section~\ref{sec:static}
we give brief sketches of our cache-efficient and static algorithms.
\section{Preliminaries and background}
\label{sec:prelims}
We first consider directed graphs.
Let $G = (V, E)$ denote
a directed graph with positive real edge weights, given by
$\weight: E \rightarrow \mathcal{R}^{+}$.
Let $\pi_{st}$ denote a path from $s$ to
$t$ in $G$. Define $\weight(\pi_{st}) = \sum_{e \in \pi_{st}} \weight(e)$ as the weight
of the path $\pi_{st}$.
We use $d(s,t)$ to denote the weight of a shortest path from $s$
to $t$ in $G$, also called its \emph{distance}.
The following notation was developed by Brandes~\cite{Brandes01}.
For a source $s$ and a vertex $v$, let $P_s(v)$ denote the predecessors of $v$ on shortest
paths from $s$, i.e.,
\begin{eqnarray}
\label{eqn:pred-defn}
P_s(v) = \{ u \in V: (u, v) \in E \mbox{ and } d(s, v) = d(s, u) + \weight(u, v) \}
\end{eqnarray}
Further, let $\sigma_{st}$ denote the number of shortest paths from
$s$ to $t$ in $G$ (with $\sigma_{ss}=1$).
Finally, let $\sigma_{st} (v)$ denote the number of shortest paths from $s$ to $t$ in $G$ that pass through $v$.
It follows from the definition that,
\begin{eqnarray}
\label{eq:pair-dep}
\sigma_{st}(v) =
\begin{cases}
0 \ \ \ \ \ \ \ \ \ \ \ \ \mbox { if $d(s,t) < d(s, v) + d(v, t)$} \\
\sigma_{sv} \cdot \sigma_{vt} \ \ \ \ \mbox{otherwise}
\end{cases}
\end{eqnarray}
The dependency of the pair $s, t$ on an intermediate vertex $v$ is
defined in~\cite{Brandes01} as the {\em pair dependency}
$\delta_{st}(v) = \frac{\sigma_{st}(v)} {\sigma_{st}}$.
\vspace{.05in}
For $v \in V$, the {\it betweenness centrality} $\bc(v)$ is defined by Freeman~\cite{Freeman77} as:
\begin{eqnarray}
\label{eqn:bc-defn}
\bc(v) = \sum_{s \neq v, t \neq v} \frac{\sigma_{st}(v)}{\sigma_{st}} = \sum_{s \neq v, t \neq v} \delta_{st}(v)
\end{eqnarray}
The following two-step procedure computes BC for all $v \in V$:
\begin{enumerate}%
\item For every pair $s, t \in V$, compute $\sigma_{st}$.
\item For every vertex $v \in V$, and for every $s, t$ pair, compute $\sigma_{st}(v)$ (Equation~\ref{eq:pair-dep}), and then compute $\bc(v)$ (Equation~\ref{eqn:bc-defn}).
\end{enumerate}
Step~1 above can be achieved by $n$ executions of Dijkstra's single source shortest paths
algorithm.
Therefore Step~1 takes time $O(mn + n^2 \log{n})$ time,
if we use a
priority queue
with $O(1)$ amortized cost for the decrease-key operation.
For every vertex $v$, Step~2
takes $O(n^2)$ time, since there are $O(n^2)$ pair dependencies.
This gives a $\Theta(n^3)$ time algorithm to compute BC for all vertices.
Thus, the bottleneck of the above algorithm is the second step which explicitly sums up the pair
dependencies for every vertex.
To obtain a faster algorithm for sparse graphs, Brandes~\cite{Brandes01} defined the dependency of a vertex $s$ on a vertex $v$ as:
$\delta_{s \bullet} (v) = \sum_{t \in V \setminus {\{v, s\}}} \delta_{st}(v)$.
Brandes~\cite{Brandes01} also made the useful observation that the partial sums satisfy a recursive
relation. In particular, the dependency of a source $s$ on a vertex $v \in V$ can be written as:
\begin{eqnarray}
\label{eqn:rec-depend}
\delta_{s \bullet} (v) = \sum_{w: v \in P_s(w)} \frac{\sigma_{sv}}{\sigma_{sw}} \cdot \left ( 1 + \delta_{s \bullet}(w) \right)
\end{eqnarray}
(See \cite{Brandes01} for a proof of Equation~\ref{eqn:rec-depend}.) The above equation gives
an efficient algorithm for computing BC described in the next section.
\subsection{Brandes' algorithm}
We present a high level overview of Brandes' algorithm (this high-level
algorithm is given in Appendix~\ref{sec:Brandes-steps}).
The algorithm begins by
initializing the BC score for every vertex to $0$. Next,
for every $s \in V$, it executes Dijkstra's SSSP algorithm. During this step,
for every $t \in V$, it computes $\sigma_{st}$, the number of shortest paths from $s$ to $t$, and
$P_s(t)$, the set of predecessors of $t$ on shortest paths from
$s$. Additionally, the algorithm stores the vertices $v \in V$ in a stack $S$ in order of non-increasing value
of $d(s, v)$. Finally, to compute the BC score,
the algorithm accumulates the dependency of $s$.
We now elaborate on this final step, which is given in
Algorithm~\ref{algo:accumulate} (Accumulate-dependency).
This algorithm
takes
as its input a source $s$ for which Dijkstra's SSSP algorithm has
been executed and the stack $S$ containing
vertices ordered by distance from $s$. The algorithm repeatedly extracts a vertex from $S$ and
accumulates the dependency
using Equation \ref{eqn:rec-depend}.
The time taken by Algorithm~\ref{algo:accumulate} is linear in the size of the $\DAG$
rooted at $s$, i.e., it is $O(m^{\ast}_{s})$.
Note that in Brandes' algorithm, the set $S$ contains vertices $v \in V$
ordered in non-increasing value of $d(s,v)$. However, for the dependency accumulation it suffices
that
$S$ contains vertices $v \in V$ ordered in the reverse topological order of the $\DAG(s)$.
Such an ordering ensures that the dependency of a vertex $w$ is {\em accumulated} to any of its predecessor $v$,
only after all the successors of $w$ in $\DAG(s)$ have been processed.
This observation is useful for our
incremental algorithm, since topological sort can be performed in linear time.
\begin{algorithm}
\begin{algorithmic}[1]
\REQUIRE
For every $t \in V$: $\sigma_{st}, P_s(t)$ \\
\hspace{0.25in}A stack $S$ containing $v \in V$ in a suitable order (non-increasing $d(s,v)$ in \cite{Brandes01})
\STATE {\bf for} every $v \in V$ {\bf do} $\delta_{s\bullet}(v) \leftarrow 0$
\WHILE {$S \neq \emptyset$}
\STATE $w \leftarrow$ pop$(S)$
\STATE {\bf for} $v \in P_s(w)$ {\bf do} $\delta_{s \bullet}(v) \leftarrow \delta_{s \bullet}(v) + \frac{\sigma_{sv}}{\sigma_{sw}} \cdot \left( 1 + \delta_{s \bullet}(w) \right)$
\STATE {\bf if} $w \neq s$ {\bf then} $\bc(w) \leftarrow \bc(w) + \delta_{s \bullet}(w)$
\ENDWHILE
\end{algorithmic}
\caption{Accumulate-dependency($s,S$) (from \cite{Brandes01})}
\label{algo:accumulate}
\end{algorithm}
\section{Static betweenness centrality}
\label{sec:static}
In this section we present static algorithms that compute betweenness
centrality faster than the Brandes algorithm.
We first consider an algorithm that is based on the Hidden Paths algorithm by
Karger et al.~\cite{KKP93} together with Brandes' accumulation technique.
The Hidden Path algorithm runs Dijkstra's
SSSP in parallel from each vertex. It identifies all pairs shortest paths while
only examining the edges in $E^*$, the set of edges that actually lie
on some shortest path.
A similar algorithm with the same running time of
$O(\mStar n+n^2\log n)$ was developed independently by McGeoch~\cite{McGeoch95}; here $m^* = |E^*|$.
Our Static-BC algorithm is presented as Algorithm~\ref{algo:static}.
In Step 1 we run the Hidden Paths algorithm to compute $E^*$ as well as the
shortest path distances for every pair of vertices.
This is the step with the dominant cost,
while in Steps 2--8 the complexity is strictly related to the size of $E^*$. In Steps 2--4, we identify the edges in each shortest path DAG and in each $P_s(v)$: for every edge $(u,v) \in E^{\ast}$,
if ${d(s,u)+\weight(u,v) = d(s,v)}$, then we add the edge $(u,v)$
to $\DAG(s)$ and the vertex $u$ to $P_s(v)$.
The overall time spent for constructing the DAGs and the predecessor lists is bounded by $O( \mStar n )$.
Step~\ref{static-comp-dist} counts the number of shortest paths from $s$ to $v$ for all $v \in V$,
by traversing $\DAG(s)$ according to the topological order of its vertices, maintained in the double-ended queue $Q$ (created in Step 6) used as a queue. We accumulate the path counts for a vertex $v$ according to the formula $\sigma_{sv} = \sum_{(u,v) \in \DAG(s)}\sigma_{su}$.
This takes time linear in the size of $\DAG(s)$. Therefore, across all sources, we spend time which is bounded by
$O(n^2 + \sum_{s \in V}{m^{\ast}_s}) = O(\mVmax n + n^2)$.
Finally in
Step 8, using $Q$ as a stack (reverse topological order), we call Accumulate-dependency($s,Q$) (Algorithm~\ref{algo:accumulate}) to
accumulate dependencies. Thus the overall running time of this static BC
algorithm is $O(m^*n + n^2 \log n)$.
The correctness of the algorithm follows from the correctness of
the Hidden Paths algorithm and the Brandes' accumulation technique.
\begin{algorithm}
\begin{algorithmic}[1]
\STATE Using the Hidden Paths algorithm, compute $E^{\ast}$, and $d(s,t)$ for every $s,t \in V$
\FOR {each node $s \in V$}
\FOR {each $(u,v) \in E^*$}
\STATE {\bf if} $d(s,u)+\weight(u,v)=d(s,v)$ {\bf then} add $(u,v)$ to $\DAG(s)$ and $u$ to $P_s(v)$
\ENDFOR
\ENDFOR
\FOR {each DAG$(s)$}
\STATE compute a dequeue $Q$ containing the nodes of $\DAG(s)$ in topological order
\STATE for all $v\in V$, compute $\sigma_{sv} = \sum_{(u,v) \in \DAG(s)}\sigma_{su}$ by accumulating path counts on vertices extracted from $Q$ in queue order (topological order)\label{static-comp-dist}
\STATE Accumulate-dependency($s,Q$), using $Q$ in stack order \hspace{.10in} //use Algorithm~\ref{algo:accumulate}
\ENDFOR
\end{algorithmic}
\caption{Static-BC($G=(V,E)$)}
\label{algo:static}
\end{algorithm}
Algorithm \ref{algo:static} can be expected to run faster than
Brandes' on many graphs since $m^*$ is often much smaller than $m$.
However, its worst-case running time is asymptotically the same as Brandes'.
We observe that if we replace the Hidden Paths algorithm in Step 1 of
Algorithm \ref{algo:static} with any other APSP algorithm that identifies
a set of edges $E'\supseteq E^*$, we can use $E'$ in place of $E$ in Step 3,
and obtain a correct static BC algorithm that runs in time
${O(m'n+n^2 + T')}$, where $m' = |E'|$ and $T'$ is the running time of
the APSP algorithm used in Step 1.
In particular, if we use
one of the faster APSP
algorithms for positive real-weighted graphs
(Pettie~\cite{Pettie04} for directed graphs or \cite{PR05} for
undirected graphs) in Step 1, we can obtain asymptotically
faster BC algorithms than Brandes' by using $E'=E$.
With Pettie's algorithm~\cite{Pettie04}, we obtain an $O(mn + n^2 \log\log n)$
time algorithm for static betweenness centrality in directed graphs, and
with the algorithm of Pettie and Ramachandran~\cite{PR05}, we obtain a
static betweenness centrality algorithm for undirected graphs that runs in
$O(mn \cdot \log \alpha(m,n))$, where $\alpha$ is an inverse-Ackermann
function.
\section{Incremental vertex update}
\label{sec:vert-update}
Here we show how to extend the incremental edge update algorithm to handle
an incremental update to a vertex $v$ in $G = (V, E)$, where we allow
an incremental edge update on any subset of edges incoming to and outgoing from the vertex $v$.
Our algorithm
is a natural extension of the algorithm for single edge update.
As in the algorithm for edge update, for every $s, t \in V$, we maintain
$d(s, t), \sigma_{st}$, and for every $s \in V$, $\DAG(s)$, the SSSP DAG rooted at $s$ in $G$.
Once we have the updated DAGs, we use a topological sort and accumulate dependencies in the reverse topological
order to get updated BC scores for all vertices.
However, instead of working only with the graph $G$,
here we also work with
the graph $G_R = (V, E_R)$, which is obtained by reversing every edge in $G$. That is $(a, b) \in E_R \textrm { iff } (b, a) \in E$.
Thus, for every $s \in V$, we also maintain $\DAG_R(s)$, the SSSP DAG rooted at $s$ in $G_R$.
Broadly, our vertex update on $v$ is processed as follows.
Let $E_i(v)$ and $E_o(v)$ denote the set of updated edges incoming to $v$ and outgoing from $v$ respectively.
We process
$E_i(v)$ and $E_o(v)$ in two steps.
Let $G'$ denote the graph
obtained by applying to $G$ the
updates on edges in $E_i(v)$. Let $G''$ denote the
graph obtained by applying to $G'$ updates on edges in $E_o(v)$.
For $s, t \in V$, let $d(s,t), d'(s,t)$ and $d''(s,t)$ denote the distance from $s$ to $t$ in $G, G'$, and $G''$ respectively.
We use a similar notation for other terms, such as $\sigma_{st}$ and $\DAG(s)$.
The main observation that we use is the following:
Since $E_i(v)$ contains updated edges incoming to $v$, the SSSP DAGs rooted at $v$ in $G$ and
in $G'$ are the same, that is, $\DAG(v) = \DAG'(v)$. We show that
Algorithm~\ref{algo:update-dag} is readily adapted to obtain $\DAG'(s)$, for all $s\in V$.
At this point, our goal is to apply the updates in $E_o(v)$ to $G'$ and obtain
$\DAG''(s)$ for every $s$. To achieve this efficiently, we first
obtain $\DAG'_R(s)$ for every $s$.
The reason to work with the reverse graph $G_R'$ and DAGs in the reverse graph is that
the edges in $E_o(v)$ are in fact incoming edges to $v$ in $G'_R$. Hence our
method to maintain DAGs when incoming edges are updated works as it is on $G'_R$
and we obtain $\DAG''_R(s)$, for every $s$. Finally, using $\DAG''_R(s)$ for every $s$
we efficiently build $\DAG''(s)$.
We now give details of each step of our algorithm starting with the graph $G$ till
we obtain the $\DAG'_R(s)$ for every $s$.
\begin{enumerate}
\item[(A)] {\bf Compute $d'(s, v)$ and $\sigma'_{sv}$ for any $s$:}
In this step we show how to compute in $G'$ the distance and number of shortest paths to $v$ from any $s$. We make
the following definitions:
For $(u_j, v) \in E_i(v)$, let
$D_j(s, v) = d(s, u_j) + \weight'(u_j, v)$.
Since the updates on edges in $E_i(v)$ are incremental, it follows that
\begin{eqnarray}
d'(s, v) = \min \{ d(s, v), \min_{j: (u_j, v) \in E_i(v)} \{D_j(s, v) \} \}
\end{eqnarray}
Further, if $d'(s, v) = d(s, v)$, we define
\begin{eqnarray}
\newpaths = | \{\pi'_{sv} : \mbox {$\pi'_{sv}$ is a shortest path in $G'$ and $\pi'_{sv}$ uses $e \in E_i(v)$} \}|
\end{eqnarray}
We also need to compute $\sigma'_{sv}$, the number of shortest paths from $s$ to $v$ in $G'$.
It is straightforward to compute $d'(s,v)$, $\sigma'_{sv}$, and $\newpaths$ in $O(|E_i(v)|)$ time.
Algorithm~\ref{algo:compute-dist-vert} gives the details of this step.
\begin{figure}
\begin{minipage}{0.5\linewidth}
\begin{algorithm} [H]
\begin{algorithmic}[1]
\REQUIRE $E_i(v)$ with updated weights $\weight'$ \\
\hspace{0.2in} $d(s,t)$ and $\sigma_{st}$, $\forall \ s, t \in V$
\ENSURE $d'(s,v), \sigma'_{sv}, \newpaths$
\STATE $\hat{\sigma}'_{sv} \leftarrow 0$, $\sigma'_{sv} \leftarrow \sigma_{sv}$, $currdist \leftarrow d(s,v)$
\FOR {each edge $(u_i,v) \in E_i(v)$}
\IF {$currdist = d(s,u_i) + \weight'(u_i,v)$}
\STATE $\sigma'_{sv} \leftarrow \sigma'_{sv} + \sigma_{su_i}$
\STATE $\hat{\sigma}'_{sv} \leftarrow \hat{\sigma}'_{sv} + \sigma_{su_i}$ \label{alg4:updnewpath}
\ELSIF {$ currdist > d(s,u_i) + \weight'(u_i,v)$}
\STATE $currdist \leftarrow d(s,u_i) + \weight'(u_i,v)$
\STATE $\sigma'_{sv} \leftarrow \sigma_{su_i}$
\ENDIF
\ENDFOR
\STATE $d'(s,v) \leftarrow currdist$
\end{algorithmic}
\caption{Compute-Dist-to-v $(s, E_i(v))$}
\label{algo:compute-dist-vert}
\end{algorithm}
\end{minipage}
\begin{minipage}{0.5\linewidth}
\vspace{-0.18in}
\begin{algorithm}[H]
\begin{algorithmic}[1]
\REQUIRE $\DAG_R(s)$, and $R_t, flag(s,t), \forall t\in V$
\ENSURE An edge set $X$ after update on edges in $E_i(v)$
\STATE $X \leftarrow \emptyset$
\FOR {each edge $(a,b) \in \DAG_R(s)$} \label{rdag:for1start}
\IF {$flag(b,s) = \mbox{UN-changed}$ or $flag(b,s) = \mbox{NUM-changed}$}
\STATE $X \leftarrow X \cup (a,b)$ \label{rdag:step1}
\ENDIF
\ENDFOR
\FOR {each $b \in V \setminus \{s\}$} \label{rdag:for2start}
\IF {$flag(b,s) = \mbox{NUM-changed}$ or $flag(b,s) = \mbox{WT-changed}$}
\STATE $X \leftarrow X \cup R_b$ \label{rdag:step2}
\ENDIF
\ENDFOR
\end{algorithmic}
\caption{Update-Reverse-DAG($s$, $E_i(v)$)}
\label{algo:update-rdag}
\end{algorithm}
\end{minipage}
\end{figure}
\item[(B)] {\bf Compute $d'(s,t)$ and $\sigma'(s, t)$ for all $s, t$:} Assuming that we have computed $d'(s, v), \sigma'_{sv}$
and $\hat{\sigma}'_{sv}$, we show that the values $d'(s, t)$ and $\sigma'(s, t)$ can be computed efficiently. We state
Lemma~\ref{lem:update-st-vert} which captures this computation. This lemma is similar to Lemma~\ref{lem:update-st} in the edge update case.
\begin{lemma}
\label{lem:update-st-vert}
Let $E_i(v)$ denote the set of edges incoming to $v$ which have been updated.
Let $G'$ denote the graph obtained by applying update in $E_i(v)$ to $G$.
For any $s \in V$ and $t \in V \setminus \{v\}$, let $D(s,t) = d'(s,v) + d(v,t)$.
\begin{enumerate}
\item If $d(s, t) < D(s,t)$, then
$d'(s, t) = d(s,t)$ and $\sigma'_{st} = \sigma_{st}$.
\item If $d(s, t) = D(s, t)$ and $d(s,v) = d'(s,v)$, then $d'(s, t) = d(s,t)$ and
${\sigma'_{st} = \sigma_{st} + \newpaths \cdot \sigma_{vt}}$.
\item If $d(s, t) = D(s, t)$ and $d(s,v) > d'(s,v)$, then $d'(s, t) = d(s,t)$ and
${\sigma'_{st} = \sigma_{st} + \sigma'_{sv} \cdot \sigma_{vt}}$.
\item If $d(s, t) > D(s,t)$, then
$d'(s, t) = D(s, t)$ and $\sigma'_{st} = \sigma'_{sv} \cdot \sigma_{vt}$.
\end{enumerate}
\end{lemma}
\item[(C)] {\bf Compute $\DAG'(s)$ for every $s$:} Assuming that we have computed $d'(s,t)$ and $\sigma'(s,t)$ for
all $s, t \in V$, we now show how to obtain the updated DAGs. Note that we can readily compute
the value $flag(s,t)$ for every $s, t$, using the updated distances and number of shortest paths.
The algorithm to compute $\DAG'(s)$ for any $s \in V$ is very similar to Algorithm~\ref{algo:update-dag} in
the edge update case. The only modification to Algorithm~\ref{algo:update-dag} we need is in
Steps 8--9 where we consider a single edge $(u, v)$ -- instead in this case
we consider every edge in $E_i(v)$.
\item[(D)]{\bf Compute $\DAG'_R(s)$ for every $s$:}
We now need to update $\DAG_R(s)$, for every $s$, for which we use Algorithm~\ref{algo:update-rdag}.
This algorithm requires for every $t \in V$, a set $R_t$ which is defined as follow:
\begin{eqnarray}
{R_t = \{(a,t) \ | \ (t,a) \in \DAG'(t) \textrm{ and } \weight'(t,a) + d'(a,v) = d'(t,v) \}}
\end{eqnarray}
The set $R_t$ is the set of (reversed) outgoing edges from $t$ in $\DAG'(t)$ that lie on a shortest path from $t$ to $v$ in $G'$.
Consider an
edge $e=(a,b)$ in the updated $\DAG'_R(s)$. If $e$ is in $\DAG_R(s)$, it is added to $\DAG'_R(s)$ by Steps \ref{rdag:for1start}--\ref{rdag:step1}. If $e$ lies on a new shortest path present only in $G_R'$,
its reverse must also lie on a shortest path that goes through $v$ in $G'$, and it will be added to $\DAG'_R(s)$ by the $R_b$ during Steps \ref{rdag:for2start}--\ref{rdag:step2}. ($R_b$ could also contain edges on old shortest paths through $v$ already processed in Steps \ref{rdag:for1start}--\ref{rdag:step1}, but even in that case each edge
is added to $\DAG'_R(s)$ at most twice by Algorithm \ref{algo:update-rdag}.)
Note that we do not need to process edges $(u_j,v)$
in $E_i$ separately (as with edge $(u,v)$ in Algorithm 2),
because these edges will be present in the relevant $R_{u_j}$.
The correctness of Algorithm~\ref{algo:update-rdag} follows from
Lemma~\ref{lem:update-rdag}. This lemma similar to Lemma~3 and its proof
is in the Appendix.
\begin{lemma}
\label{lem:update-rdag}
An edge $(a,b) \in X$ if and only if $(a,b) \in \DAG'_R(s)$ after the incremental update of the set $E_i(v)$.
\end{lemma}
\REM{
\begin{proof}
Suppose $(a,b) \in \DAG'_R(s)$. This implies that there is at least one shortest from $s$ to $b$ in $G_R'$ that uses
the edge $(a,b)$. And similarly a shortest path from $b$ to $s$ that uses the edge $(b,a)$ in $G'$. To show that $(a,b) \in X$ we consider the possible values for $flag(b,s)$:
\begin{enumerate}
\item If $flag(b,s) = \mbox{UN-changed}$ then the set of shortest paths from $b$ to $s$ in $G$ and $G'$ are
the same, so there is a shortest path from $b$ to $s$ that uses the edge $(b,a)$ in $G$. Further, since $\DAG_R(s)$ is correctly kept, the edge $(a,b)$ is added to $X$ from Step 5.
\item If $flag(b,s) = \mbox{WT-changed}$ then every shortest path from $b$ to $s$ in $G'$ goes through the updated vertex $v$. This implies that any shortest path $\pi'_{bs} \in G'$ is of the form $\pi'_{bs} = b \leadsto v \leadsto s$. Similarly every shortest path from $s$ to $b$ in $G_R'$ goes through $v$ and, since $(a,b) \in \DAG'_R(s)$, at least one shortest path in $G_R'$ is of the form $s \leadsto v \leadsto a \rightarrow b$. Therefore a shortest path of the form $b \rightarrow a \leadsto v \leadsto s$ is in $G'$, and the edge $(b,a)$ must be one of the outgoing edges of $b$ in $\DAG'(b)$ that lies on a shortest path from $b$ to $v$ in $G'$. So $(a,b) \in R_b$. Thus the edge $(a,b)$ is added to $X_s$ by Step 8.
\item If $flag(b,s) = \mbox{NUM-changed}$ then, there exist at least one path from $b$ to $s$ in $G'$ that goes through $v$, and additionally there can be shortest paths from $b$ to $s$ in $G'$ that don't go through $v$. Since, $(a, b) \in \DAG'_R(s)$, the edge $(b, a)$ lies on one or both types of paths. Suppose $(b,a) \in \pi_{bs}$, which does not use $v$. Then the path $\pi_{bs}$ is a shortest
path in $G$ and hence $(a, b) \in \DAG_R(s)$. In this case, $(a,b)$ is added to $X$ by Step~5.
If $(a, b) \in \pi'_{bs}$ which contains $v$, then since $\pi'_{bs}$ uses the vertex $v$,
we know that $\pi'_{bs} = b \leadsto v \leadsto s$.
Hence, similar to the case (2) above, we conclude that $(a,b) \in R_b$.
Therefore the edge $(a, b)$ is added to $X$ by Step~8.
\end{enumerate}
Suppose edge $(a,b) \in X$. To show that $(a, b) \in \DAG'_R(s)$
we consider the different steps in Algorithm~\ref{algo:update-rdag} where $(a,b)$ can be added
to $X$:
\begin{enumerate}
\item The edge $(a,b)$ is added to $X$ by Step~5. This implies that the edge ${(a,b) \in \DAG_R(s)}$. Thus, there exists a shortest path in $G_R$ of the form $s \leadsto a \rightarrow b$. Therefore, there exists a shortest path in $G$ from $b$ to $s$, say $\pi_{bs} = b \rightarrow a \leadsto s$. Note that we execute Step~4 when $flag(b,s) = \mbox{UN-changed}$ or $flag(b, s) = \mbox{NUM-changed}$.
For either value of the flag every shortest path from $b$ to $s$ in $G$ is also shortest path in $G'$.
Therefore, the path ${\pi'_{sb} = s \leadsto a \rightarrow b}$ is a shortest path in $G_R'$ and hence the edge $(a,b) \in \DAG'_R(s)$.
\item The edge $(a,b)$ is added to $X$ by Step~8. Thus, $(a,b) \in R_b$. This implies that the edge $(b,a)$ is on a shortest path in $G'$ from $b$ to $v$. Moreover, we add $(a,b)$ in Step~8 when $flag(b, s) = \mbox{NUM-changed}$ or $flag(b,s) = \mbox{WT-changed}$.
Therefore, there exists at least one shortest path from $b$ to $s$ in $G'$ that goes through $v$. Thus, since the edge $(b,a)$ is on a shortest path from $b$ to $v$ in $G'$, then it is also on at least a shortest path from $b$ to $s$ in $G'$. Therefore, $(a,b)$ is in at least one shortest path from $s$ to $b$ in $G_R'$, and this establishes that $(a,b) \in \DAG'_r(s)$.
\end{enumerate}
This completes the proof of the lemma.
\qed
\end{proof}
}
\end{enumerate}
At this point, after having Steps~(A)--(D) above executed, we have processed the updates in $E_i(v)$ and obtained the modified distances $d'(\cdot)$, modified counts $\sigma'_{(\cdot)}$ and $\DAG'(s)$
for every $s$. In addition, we have obtained the modified reverse DAG for every $s$. To process the updates in $E_o(v)$,
we work with $G_R'$. Since we are processing incoming edges in $G_R'$, our earlier steps apply
unchanged, and we
obtain modified values for $d''(\cdot)$, $\sigma''_{(\cdot)}$, and $\DAG''_R(s)$ for every $s$. Finally,
using Algorithm~\ref{algo:update-rdag} we obtain the $\DAG''(s)$ for every $s$.
This completes our vertex update and to compute the updated BC values, we apply Brandes' accumulation technique (Algorithm~\ref{algo:accumulate}).
\paragraph{Complexity:}
Computing $d'(s,v),\sigma'_{sv}$ and $\hat{\sigma}'_{sv}$ requires time ${O(|E_i(v)|) = O(n)}$ for each $s$, and hence $O(n^2)$ time
for all sources.
Applying Lemma \ref{lem:update-st-vert}
to all pairs of vertices takes
time $O(n^2)$.
For any $s$, the complexity of modified Algorithm~\ref{algo:update-dag} becomes $O(m_s^{*}+ \mStar_v +n)$ which
leads to a total of $O(m' \cdot n + n^2)$.
Creating a set $R_t$
requires at most
${O(E^*\cap \{\textrm{outgoing edges of }t\})}$,
so the overall complexity for all
the sets is $O(m^*)$. Finally, we bound the complexity
of Algorithm \ref{algo:update-rdag}:
the algorithm adds $(a,b)$ in a
reverse DAG edge set $X$ at most twice.
Since $\sum_{s \in V}{|E(\DAG'(s))|} = \sum_{s \in V}{|E(\DAG'_R(s))|}$,
at most ${O(\mVmax n)}$ edges can be inserted into all the sets $X$
when Algorithm~\ref{algo:update-rdag} is executed
over all sources. Finally, since applying the updates
in $E_o(v)$ requires a symmetric
procedure starting from the reverse DAGs, the final complexity bound of $O(m'~\cdot~n~+~n^2)$ follows.
|
\section{Introduction}
During the last three decades there have been dramatic advances
in both theoretical understanding of the requirements for control of
quantum dynamics and the technology that is needed for the execution of
proposed control paradigms\cite{ric2,Shap}.
Experimental verifications of the theory for systems as diverse as
control of population transfer in Bose-Einstein condensates (BECs) and
in chemical reactions have been reported
\cite{Baum1,Gaub,Baum2,Ditt,Half,Scha,Scha2,Bas}.
A particularly useful subgroup of the proposals for control of
quantum dynamics of a system rely on adiabatic transfer via the slow
variation of an external field that is applied to the system. However,
experimental exploitation of such control schemes can be rendered difficult
by the occurrence of unwanted internal decoherence processes and by external
noise; both of these difficulties can be reduced or avoided if the adiabatic
transfer process can be speeded up sufficiently to permit population transfer
to compete successfully with the time-dependence of the perturbations.
Indeed, with this goal in
mind, several methods for the acceleration of quantum dynamics, including
adiabatic dynamics, have been proposed. These methods include the
counter-diabatic protocol \cite{ric}, frictionless quantum driving \cite{ber3},
invariant-based inverse engineering \cite{mug1},
and fast forward scaling \cite{mas1,mas2,mas3,mas4,torr2},
which is also used for protection of quantum
states from potential fluctuations \cite{mas5}.
Lattice models are widely used to describe quantum systems, examples of
which are a BEC in an optical lattice, a network of nonlinear waveguides
and optical fibers, and a superconducting ladder of Josephson junctions.
For example, motivated by the potential applicability to quantum computation,
and by the opportunity to simulate aspects of complex electronic behavior in
crystalline matter, many remarkable features of BECs in optical lattices have
been studied \cite{Mor2}.
The existing studies clearly reveal the value of the ability to manipulate
BECs in optical lattices for the purpose of preparing well-defined quantum
states.
We have been stimulated by
this observation to
to extend
the theory of accelerated adiabatic transfer to lattice systems so as to
determine the potential that drives
specified state-to-state population transfer
without excitation of unwanted quantum states.
In this paper we provide a
derivation of that driving potential, and we apply the theory to site-to-site
population transfer of a BEC in a quasi-one-dimensional
optical lattice.
We show that modulation of the lattice potential can
transfer the population of the BEC between sites of the
lattice with high fidelity and without unwanted excitations.
The theory
developed is applicable to any lattice in which the on-site potential is
tunable. We also demonstrate the robustness of the accelerated population
transfer to variation (approximation) of the driving potential.
In Sec. \ref{model}
we present the framework of the theory of accelerated quantum
adiabatic dynamics in a lattice system and discuss its relationship with
the corresponding theory for a continuous system.
In Sec. \ref{Control of BEC in optical lattice potential} we study
accelerated population transfer in a Bose-Einstein condensate in a
one-dimensional optical lattice potential. The robustness of the method
with respect to approximation of the driving potential is studied
in Sec. \ref{Comments}.
An Appendix provides a brief description of the basic theory of acceleration
of non-adiabatic quantum dynamics.
\section{Fast-forward transformation in discrete systems}
\label{model}
We consider a lattice system in which the dynamics is governed by a
discrete time-dependent Schr$\ddot{\mbox{o}}$inger equation
\begin{eqnarray}
i\frac{d\Psi(m,t)}{dt} &=& \sum_l \tau_{m,l} \Psi(l,t) \nonumber\\
&&+ \frac{V_0(m,R(t))}{\hbar}\Psi(m,t),
\label{se0}
\end{eqnarray}
where $l,m$ denote sites and $t$ time, respectively, and
$\tau_{m,l} = \tau_{l,m}^\ast$ is the rate of hopping
between sites $m$ and $l$.
The potential $V_0$ is modulated by a parameter $R$, which is a function of $t$.
If the parameter $R$ changes slowly enough from $R_i$ to $R_f$,
and if the initial state is the $n$th energy eigenstate
of the Hamiltonian with potential $V_0(R_i)$,
the wave function of the state on site $m$
changes from $\phi_n(m,R_i)$ to $\phi_n(m,R_f)$ modulo the
dynamical and adiabatic phases of the states.
The wave function $\phi_n(m,R)$ is a solution of the time-independent
Schr$\ddot{\mbox{o}}$inger equation
\begin{eqnarray}
&&\sum_l \hbar\tau_{m,l} \phi_n(l,R) + V_0(m,R)\phi_n(m,R) \nonumber\\
&&= E_n(R) \phi_n(m,R).
\label{se1}
\end{eqnarray}
On the other hand, when the parameter $R$ changes at a non-zero rate,
transitions occur to other levels. Our purpose is to derive a potential
that drives the state from $\phi_n(m,R_i)$ to $\phi_n(m,R_f)$
in some short time $T_F$ without unwanted
excitations to other states. For that purpose we consider an intermediate
state whose wave function is represented as
\begin{eqnarray}
\Psi_{FF}(m,t) &=& \phi_n(m,R(t))\exp\big{[}if(m,t)\big{]} \nonumber\\
&&\times\exp\Big{[}-\frac{i}{\hbar}\int_0^tE_n(R(t'))dt'\Big{]}.\ \ \ \
\label{psiFF}
\end{eqnarray}
Note that Eq. (\ref{psiFF}) contains the additional phase $f(m, t)$,
and that the intermediate state connects the initial state $\phi_n(m,R_i)$ and
the target state $\phi_n(m,R_f)\exp\Big{[}
-\frac{i}{\hbar}\int_0^{T_F}E_n(R(t'))dt'\Big{]}$ in time $T_F$.
We require that this additional phase
vanishes at $t=0$ and at $t=T_F$, and
we assume that the intermediate state satisfies the time-dependent
Schr$\ddot{\mbox{o}}$inger equation
\begin{eqnarray}
&&i\frac{d\Psi_{FF}(m,t)}{dt} = \sum_l \tau_{m,l} \Psi_{FF}(l,t) \nonumber\\
&&+ \frac{V_{FF}(m,t)}{\hbar}\Psi_{FF}(m,t),
\label{se2}
\end{eqnarray}
in which $V_{FF}(m,t)$ is the driving potential.
We seek the driving potential that generates
$\phi_n(m,R_f)\exp\Big{[}-\frac{i}{\hbar}\int_0^{T_F}E_n(R(t'))dt'\Big{]}$
from $\phi_n(m,R_i)$.
Although we do not aim to generate the adiabatic phase, that uniform
phase can be tuned by a uniform potential if necessary.
To find the forms of the driving potential and the additional phase $f(m,t)$
we substitute Eq. (\ref{psiFF}) into the Schr$\ddot{\mbox{o}}$inger
equation (\ref{se2}) and we use Eq. (\ref{se1}) to rearrange the resulting
equation. The imaginary part of the resultant equation leads to
\begin{eqnarray}
&& \dot{R} \mbox{Re}\big\{ \phi_n^\ast(m,R)\pa_R\phi_n(m,R) \big\} \nonumber\\
&=& \sum_l \mbox{Im} \Big{(}
\tau_{m,l}\phi_n^\ast(m,R)\phi_n(l,R) \nonumber\\
&&\times \big\{ \exp\big{[} i\big{(}f(l,t) - f(m,t)\big{)} \big{]} - 1\big\}
\Big{)}.\nonumber\\
\label{eq_f}
\end{eqnarray}
The solution of Eq. (\ref{eq_f}) yields the additional
phase $f(m,t)$, and the real part gives the driving
potential as a functional of $f$, $V_0$, $R$ and $\phi_n$:
\begin{eqnarray}
&&V_{FF}(m,t) = V_0(m,R(t)) \nonumber\\
&&+ \sum_l \mbox{Re} \Big\{ \hbar\tau_{m,l}
\frac{\phi_n(l,R(t))}{\phi_n(m,R(t))} \nonumber\\
&& \times \big{(}1-\exp\big{[}i\{f(l,t)-f(m,t)\}\big{]}\big{)} \Big\}
\nonumber\\
&&-\hbar\dot{f}(m,t)
- \hbar\dot{R}
\mbox{Im}\Big{[}\frac{\pa_R\phi_n(m,R(t))}{\phi_n(m,R(t))}\Big{]}.
\label{eq_VFF}
\end{eqnarray}
It is necessary that
$R$ satisfies the conditions
\begin{eqnarray}
R(0) &=& R_i\nonumber\\
R(T_F) &=& R_f.
\end{eqnarray}
If we take the boundary conditions to be
\begin{eqnarray}
\dot{R}(0) = \dot{R}(T_F) = 0,
\end{eqnarray}
$f(m,t)$ vanishes at $t=0$ and at $t=T_F$
(see Eq. (\ref{eq_f}) ),
and the intermediate state coincides with the target state at $T_F$.
The driving potential is obtained by substituting the additional
phase into Eq. (\ref{eq_VFF}).
With the boundary conditions
\begin{eqnarray}
\ddot{R}(0) = \ddot{R}(T_F) = 0
\end{eqnarray}
the driving potential coincides with $V_0$ at $t=0$ and at $t=T_F$.
The time-dependence of R is arbitrary except for the requirement
imposed by the above boundary conditions. The driving potential
depends on the time-dependence of $R$.
In the case that the hopping rate and the wave function are real,
Eq. (\ref{eq_f}) and Eq. (\ref{eq_VFF}) simplify to
\begin{eqnarray}
&&\ \dot{R} \pa_R\phi_n(m,R) \nonumber\\
&=& \sum_l\tau_{m,l}\phi_n(l,R) \sin[ f(l,t) - f(m,t) ],\nonumber\\
\label{eq_f2}
\end{eqnarray}
and
\begin{eqnarray}
&&V_{FF}(m,t) = V_0(m,R(t)) \nonumber\\
&&+ \sum_l \hbar\tau_{m,l}
\frac{\phi_n(l,R(t))}{\phi_n(m,R(t))} \nonumber\\
&& \times \{1-\cos[f(l,t)-f(m,t)]\} - \hbar\dot{f}(m,t).\nonumber\\
\label{eq_VFF2}
\end{eqnarray}
We note that Eq. (\ref{eq_f}) implies that for $\dot{R}$
sufficiently large there is no solution for $f(m,t)$.
That is, there is a lower limit to the control time $T_F$.
This property is not seen in the fast-forward theory for continuous
systems \cite{mas2}.
Eqs. (\ref{eq_f}) and (\ref{eq_VFF}), for $f$ and for $V_{FF}$,
reduce to the corresponding equations for continuous systems
shown in Ref. \cite{mas2} in the limit that the differences between
adjacent sites of $f$ and of $\phi_n$ are small.
The theory of acceleration of non-adiabatic quantum dynamics in a
continuous system is described in Ref. \cite{mas1}.
Following the same analysis as in Ref. \cite{mas1},
the key elements of the theory of accelerated non-adiabatic
quantum dynamics in a lattice system are exhibited in the Appendix.
The analysis described above can be
straightforwardly extended to the case when a nonlinear
Schr$\ddot{\mbox{o}}$inger equation is the basic descriptor of the
system dynamics.
Consider
\begin{eqnarray}
&&i\frac{d\Psi(m,t)}{dt} = \sum_l \tau_{m,l} \Psi(l,t)
\nonumber\\
&&+ \frac{V_0(m,R(t))}{\hbar}\Psi(m,t) + \frac{c}{\hbar}|\Psi(m,t)|^2\Psi(m,t),
\nonumber\\
\label{se0_non}
\end{eqnarray}
where $c$ is a constant.
We assume the same form of the wave function of the intermediate state
$\Psi_{FF}$ as in Eq. (\ref{psiFF}).
Then $\phi_n$ is a solution of the time-independent nonlinear
Schr$\ddot{\mbox{o}}$inger equation
\begin{eqnarray}
&&\sum_l \hbar\tau_{m,l} \phi_n(l,R) + V_0(m,R)\phi_n(m,R) \nonumber\\
&&+ c|\phi_n(m,R)|^2\phi_n(m,R)= E_n(R) \phi_n(m,R).\nonumber\\
\label{se1_non}
\end{eqnarray}
We assume that the intermediate state wave function
is defined by the nonlinear Schr$\ddot{\mbox{o}}$inger equation
\begin{eqnarray}
&&i\frac{d\Psi_{FF}(m,t)}{dt} = \sum_l \tau_{m,l} \Psi_{FF}(l,t)
\nonumber\\
&&\hspace{2.3cm} + \frac{V_{FF}}{\hbar}(m,t)\Psi_{FF}(m,t) \nonumber\\
&&\hspace{2.3cm} + \frac{c}{\hbar}|\Psi_{FF}(m,t)|^2\Psi_{FF}(m,t).\ \ \ \ \
\label{se2_non}
\end{eqnarray}
We can derive the equations for the additional phase and the
driving potential in the same manner as for the
linear Schr$\ddot{\mbox{o}}$inger equation.
The resultant equations are the same as
Eqs. (\ref{eq_f}) and (\ref{eq_VFF}), respectively.
The nonlinear term influences the driving potential through $\phi_n$
in Eq. (\ref{se1_non}).
\section{Site-to-site population transfer of a BEC in an optical lattice}
\label{Control of BEC in optical lattice potential}
As an example, we now consider site-to-site population transfer of
a BEC in an optical lattice. The lattice is defined by an external potential
that is the sum of a spatially linear potential, which is tunable,
and a stationary periodic potential
\begin{eqnarray}
V_{ext}(\larger,t) = \xi(t) z + U_L(x,y) \sin^2 (2\pi z/\lambda),\
\label{Vext1}
\end{eqnarray}
where $\lambda/2$ is the wavelength (period) of the potential.
We consider the case that the mean field condensate interaction is negligible.
A discrete model of the BEC in a tilted trap was introduced in Ref. \cite{Tro},
using the tight binding approximation.
In the tight binding approximation the condensate order parameter is written as
\begin{eqnarray}
\Phi(\larger,t) = \sqrt{N_T}\sum_m\Psi(m,t)\varphi(\larger-\larger_m),
\label{eq_Phi}
\end{eqnarray}
where $N_T$ is the total number of atoms and
$\varphi(m,\larger)=\varphi(\larger-\larger_m)$
is the condensate wave function localized in the $m$th trap with location
$\larger_m$.
We assume that
$\int \varphi(m,\larger)\varphi(m+1,\larger)d\larger = 0$
and $\int \varphi^2(m,\larger) d\larger = 1$.
Using Eq. (\ref{eq_Phi}),
the Gross-Pitaevskii equation can be rewritten to read \cite{Tro}
\begin{eqnarray}
i\hbar\frac{\pa}{\pa t}\Psi(m,t) &=&
- K\big{[}\Psi(m-1,t)+\Psi(m+1,t)\big{]}\nonumber\\
&&+ \frac{\xi(t)\lambda m}{2} \Psi(m,t),
\label{GP1}
\end{eqnarray}
where
\begin{eqnarray}
K &\simeq& -\int d\larger\Big{[}
\frac{\hbar^2}{2m_0}\nab\varphi(m,\larger)\cdot\nab \varphi(m+1,\larger)
\nonumber\\
&&+ \varphi(m,\larger) V_{ext}(\larger) \varphi(m+1,\larger)\Big{]},
\end{eqnarray}
with $m_0$ the mass of an atom.
$K$ is independent of $m$ because of the orthogonality
$\int \varphi(m,\larger)\varphi(m+1,\larger)d\larger = 0$.
Equation (\ref{GP1}) then can be rewritten as
\begin{eqnarray}
i\frac{\pa}{\pa t}\Psi(m,t) &=&
\tau\big{[}\Psi(m-1,t)+\Psi(m+1,t)\big{]}\nonumber\\
&&+ \frac{V(m,t)}{\hbar} \Psi(m,t),
\label{GP2}
\end{eqnarray}
with
\begin{eqnarray}
\tau &=& -K/\hbar,
\end{eqnarray}
and
\begin{eqnarray}
V(m,t) &=& \frac{1}{2}\xi(t)\lambda m.
\end{eqnarray}
We demonstrate the acceleration of population transfer
for a BEC in a lattice with this model.
Our goal is the transfer of population to the ground state of
the linear potential with $\xi=\xi_f$ from the ground state of the
linear potential with $\xi=\xi_i$.
We take $\xi_i = -\xi_f$
so that the population is transferred from one side of the lattice
to the opposite side of the lattice.
\subsection{A three-site model}
We consider first a three-site model with site potential
\begin{eqnarray}
V_0(m,R(t)) = \hbar\omega R(t) m.
\label{V01}
\end{eqnarray}
In Eq. (\ref{V01}), the constant frequency $\omega$ is defined by
\begin{eqnarray}
\omega = -\frac{\xi_i\lambda}{2\hbar} = \frac{\xi_f\lambda}{2\hbar},
\end{eqnarray}
and the time-dependence of $R(t)$ is chosen to be
\begin{eqnarray}
R(t) = R_0 + \frac{2}{T_F}\Big{[}
t-\frac{T_F}{2\pi}\sin\Big{(}\frac{2\pi}{T_F}t\Big{)} \Big{]}.
\end{eqnarray}
We take $R_0=-1$,
so that $V_0(m,R(t))$ changes from $\xi_i\lambda m /2$
to $\xi_f\lambda m /2$ in time $T_F$, and take
the hopping rate in Eq. (\ref{se0}) to be
\begin{eqnarray}
\tau_{m,l} = \tau (\delta_{m,l-1} + \delta_{m,l+1}).
\end{eqnarray}
We calculated the additional phase and driving potential for this
model system using Eqs. (\ref{eq_f2}) and (\ref{eq_VFF2}), respectively,
with the parameter set $T_F=4.2$ ms, $\omega=2.14$ /ms,
$\hbar/2K=0.35$ ms and $\lambda=850$ nm \cite{Tro}.
The time-dependence of the additional phase is shown in Fig. \ref{f_stf3_3site},
where we choose $f(1,t)=0$.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=7.5cm]{figure1.eps
\end{center}
\caption{\label{fig:epsart}
Time-dependence of the additional phase.}
\label{f_stf3_3site}
\end{figure}
The driving potential $V_{FF}(m,t)$, shown in Fig. \ref{VFF_stf3},
differs from $V_0(m,R(t))$ for $0<t<T_F$,
and is equal to $V_0$ at $t=0$ and $t=T_F$.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=7.5cm]{figure2.eps
\end{center}
\caption{\label{fig:epsart} (Color online)
Time-dependence of $V_{FF}(m,t)/\hbar$.
The unit of time is 1 ms.
The inset shows the time-dependence of the fidelity, defined
by $F(t) = |<\phi_0|\Psi>|$.}
\label{VFF_stf3}
\end{figure}
We have simulated the evolution of the model system driven by $V_{FF}(m,t)$
from the ground state corresponding to $V_{0}(m,R(0))$.
That evolution is monitored by the fidelity
\begin{eqnarray}
F(t) = |<\phi_0|\Psi>|,
\end{eqnarray}
where $|\phi_0>$ is the ground state of the instantaneous
Hamiltonian $H_0(R(t))$ and $|\Psi>$ is
the state driven by the potential $V_{FF}(m,t)$.
The time-dependence of the fidelity is shown in the inset to
Fig. \ref{VFF_stf3};
it is equal to unity at $T_F$.
A comparison of the population evolution under $V_0(m,R(t))$ and
under $V_{FF}(m,t)$ is shown
in Fig. \ref{dyn_H0_stf3}.
We note that the non-adiabatic transfer generates unwanted
excitations, with the population of each site deviating from that
evolving under the instantaneous Hamiltonian
(dotted lines in Fig. \ref{dyn_H0_stf3}).
The fidelity of the population evolution driven by $V_0(m,R(t))$
is 0.938 at $T_F$.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=7.5cm]{figure3.eps
\end{center}
\caption{\label{fig:epsart} (Color online)
Time-dependence of the population evolution under $V_0(m,R(t))$
(dashed and solid lines)
and $V_{FF}(m,t)$ (dotted lines).
The evolution under the instantaneous Hamiltonian is also shown
with dotted lines.
The notation is $\Psi_m=\Psi(m,t)$. }
\label{dyn_H0_stf3}
\end{figure}
\subsection{A four-site model}
We have also examined accelerated population transfer of a BEC in a
four-site model.
The parameters used for these calculations are the same
as for the three-site model except that $\omega = 0.714$ /ms.
The population of the ground state of the instantaneous Hamiltonian
for each site is shown in Fig. \ref{p_ex_4site}.
The initial state is located mainly at sites 3 and 4,
while the target state is located mainly at sites 1 and 2.
The time-dependence of the driving potential is shown in Fig. \ref{VFF_4site}.
The time-dependence of the fidelity are compared
in the inset to Fig. \ref{VFF_4site}.
The solid curve and the broken curve correspond to the
dynamics with $V_{FF}$ and $V_0$, respectively.
We note that the fidelity
decreases and does not recover at $T_F$ in the $V_0$ generated dynamics
because of unwanted excitations whereas for the $V_{FF}$ generated dynamics
the fidelity becomes unity at $t = T_F$.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=7.5cm]{figure4.eps
\end{center}
\caption{\label{fig:epsart} (Color online)
Population of the ground state of the instantaneous Hamiltonian
in the four-site model system.}
\label{p_ex_4site}
\end{figure}
\begin{figure}[h!]
\begin{center}
\includegraphics[width=7.5cm]{figure5.eps
\end{center}
\caption{\label{fig:epsart} (Color online)
Time-dependence of $V_{FF}(i,t)/\hbar$.
The unit of time is 1 ms.
The inset shows the time-dependence of the fidelity.}
\label{VFF_4site}
\end{figure}
\section{Comments}
\label{Comments}
It is one matter to calculate the exact driving potential required
to transfer the BEC population between sites with perfect fidelity,
but it is another matter to generate that potential in a real experiment.
It is usually the case that in real experiments we cannot generate a
perfect rendition of a specified potential.
Then, the robustness of the proposed population transfer method to
variation of the driving potential is important.
We can test the efficiency of our proposed transfer process
to approximation of the driving potential by considering population
transfer under a driving potential that is proportional to the site number:
\begin{eqnarray}
V_{app}(j,t) = \hanaV(t) j.
\label{eq_Vapp}
\end{eqnarray}
In Eq. (\ref{eq_Vapp}), $\hanaV(t)$
is a function designed so that $V_{app}$ approximates
the exact driving potential.
For the three-site model, for transfers between ground states,
$V_{app}$ coincides with $V_{FF}$ because
\begin{eqnarray}
&&\phi_n(1,R) \big{[}2\phi_n^2(3,R)-\phi_n^2(2,R)\big{]} \nonumber\\
&& \ = \phi_n(3,R) \big{[}2\phi_n^2(1,R)-\phi_n^2(2,R)\big{]},\ \ \
\end{eqnarray}
for any $R$.
This property also holds for second and third eigenstates of the
instantaneous Hamiltonian, although the driving potential
depends on the level $n$.
Thus the simple potential defined in Eq. (\ref{eq_Vapp}) can
transfer population in the three-site model without unwanted excitation.
The approximation $V_{app}(j,t) = \hanaV(t)j$
is not exact for the four-site model,
but it is a good approximation to $V_{FF}$ for that model. We show the
difference between $V_{app}$ and $V_{FF}$ for the four-site model in
Fig. \ref{VFF_com_4site}.
In general, $V_{FF}$ is well approximated by $V_{app}$ with a larger deviation
near $t = T_F /2$ than in other time domains (Fig. \ref{V0_VFF_com_4site}).
The fidelity of
the population transfer in the four-site system driven by $V_{app}$ is
0.997 at $T_F$ whilst the fidelity of the population transfer
driven by $V_0$ is 0.916.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=7.5cm]{figure6.eps
\end{center}
\caption{\label{fig:epsart} (Color online)
Comparison of $V_{app}/\hbar$ and $V_{FF}/\hbar$.
The unit of the vertical axis is 1/ms.
The inset shows the time-dependence of $V/\hbar$ for $T_F/3\le t\le
2T_F/3$.}
\label{VFF_com_4site}
\end{figure}
\begin{figure}[h!]
\begin{center}
\includegraphics[width=7.5cm]{figure7.eps
\end{center}
\caption{\label{fig:epsart} (Color online)
Comparison of $\hanaV (t)/\hbar$ (red solid curve) and $\omega R(t)$ (black
broken curve).
The unit of the vertical axis is 1/ms.}
\label{V0_VFF_com_4site}
\end{figure}
Our derivation of the driving potential that accelerates adiabatic
population transfer in a lattice reveals a striking difference between
a lattice system and a continuous system.
Specifically, in the lattice system there is lower limit to $T_F$.
This limit derives from the condition for the additional phase in
Eq. (\ref{eq_f}),
which gives the lower limit for $\dot{R}$ for each $R$
depending on $\phi_n(R)$, that is,
trajectory of the evolution of the system.
We believe that the accelerated population transfer scheme described
in this paper can be used for the coherent control of many quantum
systems which are described by chain or lattice models.
|
\section{Introduction}
\label{intro}
A field theory of imaginary-time and space based on the assumption that vacuum is a thermodynamical system is constructed. It is figured to be filled with off-shell particles, and may communicate with physical ones as being the so-called virtual particles. The macroscopic observable of the system, the temperature, is introduced through the imaginary-time, and the theory is built according to Matsubara's formalism \cite{matsubara} for either a fermionic or a bosonic many-particle system. The quantizations of the field operators are achieved in terms of the Matsubara frequency, instead of the energy, and 3-momentum. The propagators are derived for the real-time and the imaginary-time from the same partition functions through the path integral approach \cite{shankar93};
the loop corrections of QED according to the corresponding Feynman rules are calculated respectively for the real-time and the imaginary-time. The propagators of the real-time are obtained by summing over all of the Matsubara frequencies, $\omega_n$, then the analytical continuation is performed from the imaginary-time to its real axis. They become the traditional propagators in the limit of zero temperature, which is parametrized by the variable $\beta$ $(=\frac{1}{k_{\rm B}T})$, as it gets close to infinity. This is because that the additional factors in the two-point correlation functions besides the conventional components are the density functions of fermions and bosons, $1\mp n_{\rm F,B}(\xi_{\bf p})$, where $n_{\rm F,B}(\xi_{\bf p})=\frac{1}{e^{\beta\xi_{\bf p}}\pm 1}$ and $\xi_{\bf p}$ is the energy carried by the particle.
If the temperature of the hypothetical vacuum is comparable to that of the microwave background radiations (CMB) \cite{ryden}, which is around a few Kelvins, the factor $\beta$ is about $10^4 \,\, ( {\rm eV}^{-1} )$, or equivalently $10^{16} \,\,( {\rm TeV}^{-1} )$ so that the exponential function is vastly large in the denominator and the density function becomes unity. As particles are created in the modern accelerators, this condition could be easily fulfilled. Although it appears that the density functions violate the Lorentz invariance, for a very tiny temperature it would, on the contrary, explains why the Lorentz invariance holds in the many-particle field theory. For a typical loop integral in this formalism, the calculations compose of a traditional Feynman integral containing the density functions and the residues of the poles from the density functions.
The loop corrections of the real-time, for the real part of the physical observables, such as self-energies, etc., give negligible contribution, therefore they are free of the ultraviolet divergences (UV). Beyond the threshold of the ingoing momenta, the same imaginary part as in the field theory is generated from the branch cut.\par
The loop calculations for the imaginary-time, $\tau$, are similar to the traditional ways; only the zeroth-component of the four-momentum for a particle is replaced by $i\omega_n$, where $\omega_n$ is equal to $\frac{2n\pi}{\beta}$ or $\frac{(2n+1)\pi}{\beta}$ for a boson or a fermion. Instead of integrating $\tau$ from 0 to $\beta$, the lower bound of the imaginary-time is replaced by an infinitesimal $\beta_0$ in the imaginary-time evolution operator for the perturbation theory. The loop integrals at $\beta_0$ provide a reference point for radiative corrections, and the UV divergences for the imaginary-time could be removed automatically
without introducing any counter term to absorb them. The radiative corrections derived from QED, such as the anomalous magnetic dipole moment, $g-2$, of the electron are consistent with the results in this formalism. Moreover, the renormalization group equations can be also derived as functions of $\beta$, which plays the role of the renormalization scale $\mu$ in field theory. They are consistent with the results of those in the $\overline{MS}$ scheme of renormalization \cite{grozin}. One thing related to this idea is the Tolman-Ehrenfest relation \cite{tolman30}, and from the studies in the general relativity it implies that the physical time is proportional to the proper time by the formula $t=\beta\tau_p$, where the ratio $\beta$ is often regarded as the "speed of time", and $t$ is then called the thermal time \cite{rovelli93}. Both of them imply that the temperature of the vacuum could determine the physical scale. It will be shown that they are related to the sacle transformations applied in the imaginary-time theory. On the other hand, there have been many efforts in studying the many-particle theory of the relativistic quantum fields \cite{landsman87}. Some of the approaches toward a QED or a QCD plasma \cite{bellac96} are similar to those presented here; instead of studying an on-shell many-particle system, some of the differences from their works are to treat the vacuum as a macroscopic system, which only off-shell particles are filling in, and to seek regenerating the known results in loop computations in the conventional field theories. In the work of \cite{huang13b} that follows this one, the relations between the imaginary-time hamiltonian with various vacuum effects are discussed, such as the Casimir effect \cite{casimir} and the van der Waals forces \cite{vanderwaals}. For the two effects, the thermal theory of vacuum not only generates consistent results with the conventional calculations but also yields the cutoff functions to automatically regulate the divergences, while in the precedent approaches the regularization functions are added intentionally. Other effects, like Unruh effect \cite{unruh} and the Hawking radiation \cite{hawking}, also exhibit many agreements. In short, the proposed thermal vacuum provides a solid thermal bath for the uniformly accelerated observer in the Unruh's thought experiment to observe the black-body radiation. Meanwhile, in the general relativity, one theoretical source of the black-body radiation is the black hole, which establishes a unique environment, the event horizon, for the virtual photons to radiate and make the black hole evaporate in a very slow pace, and agreements between the imaginary-time formalism and the viewpoint from Hawking's approach are explained in the same paper. From the last two effects, they correspond an acceleration or a surface gravity, $g$, to an effective temperature, $T=\frac{\hbar g}{2\pi c k_B}$. To estimate the temperature of the vacuum, it is about $4\times 10^{-20}K$ for the surface gravity on earth or $\sim6\times10^{-8}K$ for a black hole of a sun's size. This supports the assumption that the propagators of the particles deduced from the imaginary-time formalism in an infinitesimal temperature would become the conventional ones, and thus secure the Lorentz invariance. To extend and show the usefulness of the imaginary-time field theory, the cosmological constant can be derived through the approaches of the DeWitt-Schwinger representation \cite{dewitt75} and the Casimir effect in ref. \cite{huang13c}. One of the unique features for the cosmological constant, the ratio $w=-1$, in the equation of state, $p=w \rho$, can be obtained without any trouble from the divergence. In a recent article \cite{huang13d}, an application on the quantization of the weak gravitational field is discussed.
\par
In the next section, the free and the interaction Lagrangian of electrons and photons in the imaginary-time and space are discussed, as well as their relations to the scale invariance. The interaction range of the imaginary-time in the $S$-matrix is given a nonzero lower bound, which leads to the cancellations of the UV divergences. The Green functions of both kinds of particles are derived for the real-time and the imaginary-time in Section \ref{greenfunction}. In the following section, the radiative corrections, such as the self-energy, are performed and the comparisons with the results from the field theory will be checked. Then the renormalization group equations with respective to the variation of the vacuum temperature are presented in Section \ref{rgeq}. In the end, a conclusion will be given. In the appendices, some details of the calculations are provided for reader's convenience.
\section{Lagrangian in imaginary-time}
\label{lagimag}
\subsection{Fermion Lagrangian}\label{fermionL}
Here we may start with the path integral approach for fermions, and what in the following is basically generalized from the derivations in ref. \cite{shankar93}. Let's consider a partition function over an imaginary-time variable, ${{\tau}}$,
\begin{eqnarray*}
Z&\equiv&{\rm Tr}\,e^{-\beta \mathcal{K}}=\prod_{\{{\tau}\}}{\rm Tr}\, e^{-\beta({\tau}) \mathcal{\mathcal{K}}({\bf \tau})\Delta {\tau }},
\end{eqnarray*}
where $\mathcal{\mathcal{K}}({\tau})$ $(=\hat{H}-\mu_{\rm ch} \hat{N})$ is a normal ordered operator, $\mathcal{K}(\psi^\dagger({\tau}),\psi({\tau}))$. The operator $\hat{H}$ and $\hat{N}$ are the Hamiltonian and the number of the particle, and $\mu_{\rm ch}$ is the chemical potential. A notation, which is used throughout the paper, is the boldface that indicates a 3-dimensional vector, such as the position vector, ${\bf x}$, and 3-momentum, ${\bf p}$. The fermion fields, which are generalized to four dimensions of the imaginary-time and space, $\psi(\tau, \bf x)$ and $\psi^\dagger (\tau, \bf x)$, are the so-called grassmann numbers in the path integral formalism. For the first step, the exponential function is divided into products of infinitesimal changes with respect to the variation of the imaginary time, $\tau$. After summing over all of the functional changes of the fields, we may obtain from the Hamiltonian density of Dirac particles, $ \mathcal{K}(\tau, {\bf x})=\psi^\dagger(\tau, {\bf x})(-i\gamma^0\vec{\gamma}\cdot \vec{\nabla}+m_{\rm f}\gamma^0-\mu_{\rm f})\psi(\tau, {\bf x})$:
\begin{eqnarray}
Z&=&\int e^{\int^\beta_0 d\tau\int d^3{\bf x} \psi^\dagger(\tau,{\bf x})\left(-\frac{\partial}{\partial \tau}+i\gamma^0\vec{\gamma}\cdot\vec{\nabla}-m_{\rm f}\gamma^0+\mu_{\rm f} \right)\psi(\tau,\bf{x})}[d\psi^\dagger(\tau,{\bf x}) d\psi(\tau,\bf{x})],\label{partitionZ2}
\end{eqnarray}
where $\mu_{\rm f}$ is the chemical potential of fermions. The notation for a vector, $\vec{v}$, means a 3 dimensional vector. In Appendix \ref{appA}, the details of the derivation are provided. Here introduce a rescaling factor $e^{\mu_{\rm f}\tau}$ for a transformation of the field operator, $\psi_r(\tau,\vec{\bf x})=e^{\mu_{\rm f}\tau}\psi(\tau,\vec{\bf x})$. With the inclusion of this factor, the reference point of the energy of a fermion is shifted to the Fermi surface, since the term of the chemical potential $\mu_{\rm f}$ is removed from the new Lagrangian.
From above, the rescaled imaginary-time, space and mass are
\begin{eqnarray*}
\tau_r=\frac{3}{2\mu_{\rm f}}\left(e^{\frac{2}{3}\mu_{\rm f}\tau}-1\right),\hspace{.3cm}
\vec{\bf x}_r= e^{\frac{2}{3}\mu_{\rm f}\tau}\vec{\bf x}, \hspace{.3cm}
m_r= e^{-\frac{2}{3}\mu_{\rm f}\tau}m\hspace{.3cm}{\rm and}\hspace{.3cm}
\beta_r=\frac{3}{2\mu_{\rm f}}\left(e^{\frac{2}{3}\mu_{\rm f}\beta}-1\right),
\end{eqnarray*}
so that the new partition function becomes
\begin{eqnarray}
Z&=&\int e^{\int^{\beta}_0 d\tau_r \int d^3{\bf x}_ r\psi^\dagger_r(\tau,{\bf x})\left(-\frac{\partial}{\partial \tau_r}+i\gamma^0\vec{\gamma}\cdot\vec{\nabla}_r-m_r\gamma^0\right)\psi_r(\tau,\bf{x})}[d\psi^\dagger_r(\tau,{\bf x}) d\psi_r(\tau,\bf{x})].\label{partitionF}
\end{eqnarray}
The subscript, $r$, will be dropped hereafter without causing any ambiguity. In the following, besides fermion's propagator of the imaginary-time is derived from this partition function, that of the real-time can also be obtained from it through the summation of the Matsubara frequency and the analytic continuation from the imaginary-time, $\tau$, to the real-time, $t$.
\subsection{Interaction Lagrangian in QED} \label{IntQED}
The QED Lagrangian in the imaginary-time and space, $(\tau, \vec{\bf x})$, can be identified from the partition function, eq. (\ref{partitionF}). After replaced with the covariant derivative, it becomes
\begin{eqnarray*}
\mathcal{L}_{\rm Dirac}+\mathcal{L}_{\rm int}&=&
\left.\bar{\psi} (i \slashed{D}_\tau-m )\psi
,\right.
\end{eqnarray*}
where the covariant derivative is ${D}^\mu_\tau= \left(\frac{\partial }{i\partial \tau},{\vec{\nabla}}\right)+ie {A}^\mu(\tau,\vec{\bf x}).$ The Lagrangian of the imaginary-time is invariant under the following gauge transformations
\begin{eqnarray*}
&&\hspace{1.7cm}\psi(\tau, \vec{\bf x})\rightarrow e^{i\Lambda(\tau, \vec{\bf x})} \psi(\tau, \vec{\bf x}),\\
&&A_0\rightarrow A_0+\frac{i}{e}\partial_{\tau}\Lambda(\tau,\vec{\bf x}),\,\,\,
\vec{A}\rightarrow \vec{A}-\frac{1}{e}\vec{\nabla}\Lambda(\tau,\vec{\bf x}),
\end{eqnarray*}
where $A_0$ is the time component of the vector potential and $\vec{A}$ is for the spatial dimensions.
So the action of the interaction Lagrangian, $\mathcal{S}_{\rm int}$, in the generating function is
\begin{eqnarray}
i\mathcal{S}_{\rm int}&=&i\int d(i\tau) d^3{\bf x}\,\mathcal{L}_{\rm int}(\tau, \vec{\bf x})=e\int d\tau d^3{\bf x}\,
\bar{\psi} \slashed{{A}}\psi. \label{Linteraction}
\end{eqnarray}
The only difference from the usual interaction action is an extra imaginary number $i$, therefore as
we apply the corresponding Feynman rules, the corresponding factor for each vertex is $e\gamma^\mu$, instead of $-ie\gamma^\mu$.
\subsection{Scale invariance and thermal time}\label{scaleinv}
The scale invariance \cite{difrancesco97} is one of the important features in diverse fields of science. In statistical mechanics, it is used to
study phase transitions, and is found that near the critical point the fluctuations happen at all length scales \cite{goldenfeld92}. In the study under the imaginary-time and space, the same feature can also been found. The Lagrangian $\mathcal{L}_{\rm Dirac}$ in eq. (\ref{partitionF}) for fermions is
\begin{eqnarray}
\mathcal{L}_{\rm Dirac}(\psi,\psi^\dagger,\tau,{\bf x})&=&\psi^\dagger(\tau,{\bf x})\left(-\frac{\partial}{\partial \tau}+i\gamma^0\vec{\gamma}\cdot\vec{\nabla}-m_{\rm f}\gamma^0\right)\psi(\tau,\bf{x}).\label{Lfermion}
\end{eqnarray}
As for the photons the corresponding Lagrangian, $\mathcal{L}_{\rm Maxwell}(A_\mu, \tau,\vec{\bf x})$, including the term of the chemical potential, $\mu_\gamma$, could be obtained by replacing the time, $t$, with the imaginary-time, $-i\tau$; combined with the gauge fixing term, $\mathcal{L}_{\rm fix}$, they are:
\begin{eqnarray}
&&\hspace{-0.5cm}\mathcal{L}_{\rm Maxwell}+\mathcal{L}_{\rm fix}
=-\frac{1}{4}F_{\mu\nu}F^{\mu\nu}+\mu_\gamma A_\mu A^\mu-\frac{\zeta}{2} (\partial_\mu A^\mu)^2\nonumber\\ \label{Lphoton}
\hspace{-1cm}&=&\frac{1}{2} A_\nu\left(-\frac{\partial^2}{\partial \tau^2}+\partial_i\partial^i \right)A^\nu+\frac{\zeta-1}{2}\left(-A_0 \frac{\partial^2}{\partial\tau^2} A^0+A_i \partial^i\partial_k A^k\right) +\mu_\gamma A_\mu A^\mu,\nonumber\\
\end{eqnarray}
where the indices, 0, and $i\,\,(=1,2,3)$ refer to the components of the imaginary-time variable, $\tau$, and 3-dimensional space, ${\bf x}$. The factor $\zeta$ is the gauge parameter. The integration by parts has been used to derive from the first to the second line in eq. (\ref{Lphoton}).
eq. (\ref{Lfermion}) and (\ref{Lphoton}) are invariant under the transformations
\begin{eqnarray}
\psi(\tau,\bf{x})&\rightarrow&
s^{3/2}\psi_{0}(\tau,{\bf x}),\hspace{.5cm} \tau\rightarro
s^{-1}\tau_0,\hspace{.5cm} {\bf x}\rightarrow
s^{-1}{\bf x}_0 \nonumber\\
A^\mu(\tau,\bf{x})&\rightarrow
s A^\mu_{0}(\tau,{\bf x}),\hspace{.5cm}\mu_\gamma\rightarrow s^2\, \mu_{\gamma,0} \hspace{.5cm} {\rm and}\,\,\,\, m\rightarrow s\, m_0, \label{scaletrans}
\end{eqnarray}
where $s$ is a scale factor. In the representation of the momentum space, $(\omega_n,{\bf p})$, in expansions like eq. (\ref{psix}) and (\ref{photonex}), the transformation are
\begin{eqnarray*}
\psi(\omega_n,\bf{p})&\rightarrow
s^{{5}/{2}}\psi_{0}(\omega_n,{\bf p}),\hspace{.5cm}\omega_n\rightarro
s\,\omega_{0,n},\\
A^\mu(\tau,\bf{x})&\rightarrow
s^3\hspace{.2cm}A^\mu_{0}(\tau_r,{\bf x}),\hspace{.9cm}
{\bf p}\rightarro
s\,{\bf p}_0.
\end{eqnarray*}
Here a temperature dependent cutoff may be introduced, the maximal number of the Matsubara frequency, $N_{\rm max}$, is related to it by:
\begin{eqnarray}
\omega_{\rm cutoff}=\frac{2\pi N_{\rm max}}{\beta},\label{cutoff1}
\end{eqnarray}
The number $N_{\rm max}$ is a constant and is only constrained by the total number of particles in the system. Obviously, it is meaningless that the maximal number of mode is larger than the total number of the particles in a many-particle system. This assumption will be useful as the UV divergences are considered in Section \ref{oneloop}.
The cutoff frequencies between two different temperatures implies a scale factor,
\begin{eqnarray}
s=\frac{\beta_0}{\beta},\label{sfactor}
\end{eqnarray}
with the assumption that $N_{\rm max}$ is the same regardless of different temperatures. As for the loop integrals in Section \ref{oneloop}, the 3-momentum integration and frequency summation are performed separately. In the field theory, no matter what regularization of divergence is used, the cutoffs for each of four dimensions of the momentum are equal quantities. The same fashion is adopted in the following calculations, so, in a similar way, we may define a cutoff for the 3-momentum phase space
\begin{eqnarray}
\Lambda&=&\frac{2\pi N_{\rm max}}{\beta}.\label{cutoff2}
\end{eqnarray}
The change of the scale leads to the theory of the renormalization group; this is known from the efforts made by Kadanoff \cite{kadanoff66} in 1966, Wilson \cite{wilson75} et al. in 1975\,. An example proposed is the spins in a solid; the renormalization group describes the couplings' variation as observed in different sizes of blocks. The effective Lagrangian may be obtained after the scale transformation and integrating out the momentum phase space between different cutoffs, $\Lambda$ and $s\Lambda$. The coupling constants and masses are re-defined in the new scale. Another example is the Hawking radiation \cite{hawking}, for the field theory in the curved space-time, the ground states are defined separately near the event horizon and in the distance, the match of the conformally invariant wave functions for incident rays and outgoing rays leads to the discovery of the Hawking temperature. In the formalism presented in this paper, the Lagrangian densities possesses the same features as those in their works; in addition, the cutoffs, in eq. (\ref{cutoff1}) and (\ref{cutoff2}), obey the transformation law, $\Lambda=s \Lambda_0$, without the need of integration between two scales. As a result, the action in the partition function is scale invariant. Another important concept developed over decades is the thermal time \cite{rovelli93}, especially in the discussions of the possibilities for a quantum gravity theory. The connection between the thermodynamics and the general relativity has been discussed with intense literatures. It may be started from the Tomita-Takesaki theorem \cite{takesaki70} to derive a time flow from a generic thermal physical state, and the Unruh effect and the Hawking radiation are shown to relate to this idea. The thermal time, $t$, is related to the geometrical time, $\tau_p$, or say the proper time, by a simple formula: $t=\beta\tau_p$. This agrees with the scale transformations for the imaginary-time in eq. (\ref{scaletrans}) for different thermal times. As early as the 1930s, it was figured in a stationary spacetime with a time-like Killing vector field $\xi$, a temperature of the vacuum satisfies the Tolman-Ehrenfest relation \cite{tolman30}
\begin{eqnarray}
T||\xi||=const.\,,
\end{eqnarray}
where $||\xi||=\sqrt{g_{ab}\xi^a\xi^b}$ is the norm of $\xi$. A thermal equilibrium point of view was adopted regarding the gravity, and in the Newtonian limit, the gravitational field was related to the gravitational field by
\begin{eqnarray}
\frac{\nabla T}{T}=\frac{\vec{g}}{c^2},
\end{eqnarray}
where $\vec{g}$ is the strength of the gravitational field. As for the spatial dimensions, in Section 5 the temperature is shown to play the same role to vary the scale of the dimensions as the renormalization scale $\mu$ in renormalization. The Tolman-Ehrenfest relation, the Hawking radiation and the result of the renormalization group from the imaginary-time theory all imply that the temperature of vacuum determines the physical scales. In conclusion, the transformations from eq. (\ref{scaletrans}) and (\ref{sfactor}) can be regarded as those between coordinates of flat space-time to hold the Lagrangian invariant under scale changes.
More discussions are given in Section \ref{oneloop}.
\begin{figure}[t]
\begin{center}
\subfigure[]{\includegraphics[width=5cm]{cyclictau1d.eps}}
\hspace*{0.07\textwidth}
\subfigure[]{\includegraphics[width=5cm]{cyclictau2d.eps}}
\hspace*{0.0\textwidth}
\caption{\small From the KMS condition, the Green function is cyclic with respect to the imaginary-time $\tau$. Therefore, for $n=1$, in eq. (\ref{perturbative}) the intergation domain of $\tau$ is an open-closed interval $(0,\beta]$ in order to avoid the overlap of the periodic domains at points $-\beta$, 0, $\beta$, $2\beta$...\, as shown in (a). Similarly, for a specific time order $\tau_1>\tau_2$ in the case of $n=2$,
the periodic domains in the plane of $\tau_1$-$\tau_2$, the origin has to be taken out from the gray triangular domain A to prevent the confliction. The same reason and result can also apply on and be obtained for the cases of the other time order $\tau_2>\tau_1$ and the higher dimensions of $\tau$, $n>2$.
}
\label{Fig:Cyctau}
\end{center}
\end{figure}
\subsection{Perturbation theory for imaginary-time}\label{perturbation}
Similar to the perturbation for the real-time, the $S$-matrix is the imaginary-time evolution operator from 0 to $\beta$:
\begin{eqnarray*}
\langle {\bf p}_1,\cdots,{\bf p}_n | \,S\, | {\bf k}_1,\cdots,{\bf k}_m \rangle=\langle {\bf p}_1,\cdots,{\bf p}_n | \,T_\tau\,e^{-\int^\beta_0 d\tau H_{int}(\tau)}\, | {\bf k}_1,\cdots,{\bf k}_m \rangle,
\end{eqnarray*}
where $H_{int}(\tau)=\int d^3{\bf x} \,\mathcal{H}_{int}(x)$ and $T_\tau$ is the operator of the imaginary-time ordering. According to the KMS condition \cite{kms}, the Green functions of imaginary-time are cyclic in the interval from zero to $\beta$. As illustrated in Figure \ref{Fig:Cyctau} (a), the domain has to be an open-closed interval. Therefore we have to slightly modify the above definition by replacing the lower bound of the imaginary-time, zero, with $\beta_0$:
\begin{eqnarray*}
\lim_{\beta_0\rightarrow 0^+}\hspace{.3cm}\langle {\bf p}_1,\cdots,{\bf p}_n | \,T_\tau \,e^{-\int^\beta_{\beta_0} d\tau H_{int}(\tau)}\, | {\bf k}_1,\cdots,{\bf k}_m \rangle,
\end{eqnarray*}
\begin{figure}[t]
\begin{center}
{\includegraphics[height=5cm,width=5cm]{taudomain4.eps}}
\hspace*{0.0\textwidth}
\caption{\small After combined with the contributions from two different time orders, $\tau_1>\tau_2$ and $\tau_2>\tau_1$, as shown in Figure \ref{Fig:Cyctau} (b), the integration domain for two imaginary-time variables, $\tau_1$ and $\tau_2$, is illustrated as above. The gray area is the new integration domain with the infinitesimal white square removed from the multi-dimensional integration of $\tau$. }
\label{Fig:taudomain}
\end{center}
\end{figure}
where $\beta_0$ is an infinitesimal positive number. It may not appear to make a difference from the above expression. As we will find out later, the infinitesimal shift of the lower integration bound of $\tau$ gives a new reference point for the radiative corrections, and it automatically removes the UV divergences without the need to introduce any counter term. The expansion of the imaginary-time evolution operator
can be written as
\begin{eqnarray}\label{newIT}
&&\hspace{-1cm}\lim_{\beta_0\rightarrow 0^+} e^{-\int^\beta_{\beta_0} d\tau H_{int}(\tau)}= \lim_{\beta_0\rightarrow 0^+} \sum_{n=0}^\infty \frac{(-1)^n}{n!} \int^\beta_{\beta_0} d\tau_1\dots d\tau_n T_\tau\left\{ H_{int} (\tau_1)\dots H_{int} (\tau_n)\right\}\nonumber\\
&&\hspace{-1.0cm}= \sum_{n=0}^\infty \frac{(-1)^n}{n!} \left(\int^\beta_{0} d\tau_1\dots d\tau_n T_\tau\left\{ H_{int} (\tau_1)\dots H_{int} (\tau_n)\right\}\right.\nonumber\\
&&
\left.\hspace{2.5cm}-\lim_{\beta_0\rightarrow 0^+} \int^{\beta_0}_0 d\tau_1\dots d\tau_n T_\tau\left\{ H_{int} (\tau_1)\dots H_{int} (\tau_n)\right\}\right).\label{perturbative}
\end{eqnarray}
As for the multi-dimensional integration of the imaginary-time in the above expression, a 2-dimensional case is illustrated in Figure \ref{Fig:Cyctau} (b), a small neighborhood around the origin in the $\tau_1$-$\tau_2$ plane has to be removed from the integration domain. From the new definition, the radiative corrections computed from this imaginary-time evolution operator have to subtract the contributions of $\tau$ from 0 to $\beta_0$. In the two dimensional case, after combined with all possible time orders, as shown in Figure \ref{Fig:taudomain}, a small white square, which is the area from 0 to $\beta_0$ for $\tau_1$ and $\tau_2$, is excluded from the integration domain. Even though $\beta_0$ is infinitely close to zero, the contributions are divergent if the integral of $\beta$ also has UV divergence. In this definition, the UV divergences are canceled spontaneously along with the consideration of the consistency in scale, which will be depicted in Section \ref{phi3}. In Section \ref{oneloop}, all of the radiative corrections that are computed for the imaginary-time have to be subtracted from the contributions from 0 to $\beta_0$, as shown as the second term in the parenthesis of eq. (\ref{newIT}). The integral of $\beta_0$ plays a similar role to a counter term as in the renormalization, and the computed radiative corrections that are shifted to the reference point at $\beta_0$ will be called renormalized radiative corrections throughout the paper for convenience and also to emphasize their correspondences in the field theory.
\subsubsection{Examples from $\phi^3$- and $\phi^4$-theory}
\label{phi3}
Here to provide some examples to manifestly explain how the UV divergences are canceled by a closed-open domain of the imaginary-time $\tau$, $(0,\beta]$, where the lower bound will be denoted as $\beta_0(=0^+)$. The action of the interaction hamiltonian of the $\phi^3$-theory is
\begin{eqnarray*}
S_I=\int^\beta_0d\tau \int d^3{\bf x}\,H_I(\tau, {\bf x}),\,\,{\rm where}\,\,H_I(\tau, {\bf x})=\frac{\lambda}{3!}\, \phi^3(\tau,{\bf x}),
\end{eqnarray*}
where $\lambda$ is the coupling constant. The coupling $\lambda$ has dimension [mass], and it is a function of $\beta$ because of the scale invariance of the action, namely $\lambda(\beta)=\frac{\beta_1}{\beta}\lambda_1$, where $\lambda_1$ is the coupling constant at another temperature $\beta_1$. They are assumed to be massless particles here. Consider the correlation function
\begin{eqnarray*}
\langle 0|T_\tau\{ \phi(x)\phi(y)e^{ \int^\beta_0d\tau \int d^3{\bf x}\,H_I(\tau, {\bf x})}\}|0\rangle
\end{eqnarray*}
to its one-loop level
\begin{eqnarray*}
\langle 0|T_\tau\left\{ \phi(x)\phi(y){ \int^\beta_{\beta_0}d\tau_{z_1} \int d^3{\bf z}_1\,H_I(\tau_{z_1}, {\bf z}_1)}
{ \int^\beta_{\beta_0}d\tau_{z_2} \int d^3{\bf z}_2\,H_I(\tau_{z_2}, {\bf z}_2)}\right\}|0\rangle.
\end{eqnarray*}
According to the difinition given in eq. (\ref{newIT}), the above expression can be separated into two counterparts
\begin{eqnarray}
&&\hspace{-0.8cm}=\langle 0|T_\tau\left\{ \phi(x)\phi(y){ \int^\beta_{0}d\tau_{z_1} \int d^3{\bf z}_1\,H_I(\tau_{z_1}, {\bf z}_1)}
{ \int^\beta_{0}d\tau_{z_2} \int d^3{\bf z}_2\,H_I(\tau_{z_2}, {\bf z}_2)}\right\}|0\rangle\nonumber\\
&&\hspace{9cm}-\{\beta\rightarrow\beta_0\}.\hspace{.8cm}
\label{ham1}
\end{eqnarray}
The first term is well known, and the second one is supposed to play the role of counter term. As $\beta_0\rightarrow 0^+$, it would be interesting to see what happens to the second one, especially two of the Green functions are expressed in different scales, or say different temperatures $\beta$ and $\beta_0$. For example, the coordinates $x$ and $z_{0,1}$ in $\acontraction[0.5ex]{}{\phi}{(x)}{\phi} \phi(x)\phi_0(z_{0,1})$ are of different scales, where the subscript-0 corresponds to the temperature parameter $\beta_0$ for the the coordinate variables or fields. In order to compare the difference of the first and the second term in eq. (\ref{ham1}),
we may perform the scale transformation on the coordinate variables and the field operators by
\begin{eqnarray*}
\tau_{z_{0,i}}=\frac{\beta_0}{\beta}\tau_{z_i},\,\, {\bf z}_{0,i}=\frac{\beta_0}{\beta} {\bf z}_i,\,\,{\rm and}\,\,\phi_0 (z_{0,i})=\frac{\beta}{\beta_0}\phi (z_i),\,\,{\rm where}\,\, i=1,2.
\end{eqnarray*}
We may consider the self-energy diagram, $\raisebox{-3.mm}{\psfig{figure=selfephi3a.eps, width=2cm,height=0.8cm}}$, for the second term in eq. (\ref{ham1}) as below. Since we don't know what is the Green function of different scales, like $\acontraction[0.5ex]{}{\phi}{(x)}{\phi}
\phi(x)\phi_0(z_{0,i})$, scale transformations for the external field operators from $\phi(x)\,\,\left(=\frac{\beta_0}{\beta}\phi_0(x_0)\right)$ and $\phi(y)\,\,\left(=\frac{\beta_0}{\beta}\phi_0(y_0)\right)$ and needed. Thus,
\begin{eqnarray*}
&&\hspace{-2cm}\lambda_0^2\int^{\beta_0}_{0}d\tau_{z_{0,1}} d^3{\bf z}_{0,1}\, \int^{\beta_0}_{0}d\tau_{z_{0,2}} d^3{\bf z}_{0,2}\,\langle 0|\left\{
\acontraction[1ex]{}{\phi}{(x)}{\phi}
\phi(x)\phi_0(z_{0,1})
\acontraction[1ex]{}{\phi}{_0(z_{0,1})}{\phi}
\phi_0(z_{0,1})\phi_0(z_{0,2})
\acontraction[1ex]{}{\phi}{_0(z_{0,2})}{\phi}
\phi_0(z_{0,2})\phi_0(z_{0,1})
\acontraction[1ex]{}{\phi}{_0(z_{0,2})}{\phi}
\phi_0(z_{0,2})\phi(y)
\right\}|0\rangle,\\
&&\hspace{-1cm}=\lambda_0^2 \left(\frac{\beta_0}{\beta}\right)^2\frac{1}{\beta_0}\sum_n\int d^3{\bf p}_0\frac{1}{p_{0,n}^2}\left(\frac{1}{\beta_0}\sum_m\int d^3{\bf k}_0 \frac{1}{k^2_{0,m}}\frac{1}{(k_{0,m}+p_{0,n})^2}\right)\frac{1}{p_{0,n}^2}e^{-ip_{0,n}\cdot (x_0-y_0)},\\
&&\hspace{-1cm}=\frac{1}{\beta}\sum_n\int d^3{\bf p}\frac{1}{p_{n}^2}\left(\frac{\lambda^2}{\beta_0}\sum_m\int d^3{\bf k}_0 \frac{1}{k^2_{0,m}}\frac{1}{(k_{0,m}+p_{0,n})^2}\right)\frac{1}{p_{n}^2}e^{-ip_{n}\cdot (x-y)},
\end{eqnarray*}
where $p_n=(i\omega_n,{\bf p})$ and $k_{0,n}=(i\omega_{0,m},{\bf k}_0)$.
The factor $\left(\frac{\beta_0}{\beta}\right)^2$ in the second line is from the scale transformations of $\phi (x)$ and $\phi (y)$, and in the third line the scale transformations on $p_{0,n}$ and $\lambda_0=\frac{\beta}{\beta_0}\lambda$ are performed. The loop integral now is in the parenthesis of the third line. We may explain the idea of cancellation with some examples. Consider the imaginary-time Feynman integral without coupling constants,
\begin{eqnarray*}
I_1(\Delta)&=&
\frac{1}{\beta}\sum_{n=1}^{N_{\rm max}}\frac{1}{\sqrt{\big(\frac{ n}{\beta}\big)^2+\Delta}}\,\,\left(\simeq \int^\Lambda_0 dx \frac{1}{\sqrt{x^2+\Delta}},\,\,{\rm for}\,\,\beta\gg 1 \right),\\
I_2(\Delta)&\equiv&\lim_{\beta_0\rightarrow0}\frac{1}{\beta_0}\sum_{n=1}^{N_{\rm max}}\frac{1}{\sqrt{\big(\frac{n}{\beta_0}\big)^2+\Delta_0}},
\end{eqnarray*}
where the continuous variable $x\,(=n/\beta)$ has a dimension of [mass], $\Delta$ is the function of the external momenta or masses and $\Lambda=\frac{N_{\rm max}}{\beta}$. To simplify the expressions, the factor of $(2\pi)$ in the Matsubara frequency is ignored. The above sums are from the loop integral, $\frac{1}{\beta}\sum_n^{N_{\rm max}}\int\frac {d^3{\bf k}}{(k^2_n+\Delta)}$, after integrating over the 3-momentum. As $\beta\gg 1$, $I_1$ becomes an integral over $x$. The term $\Delta_0$ in $I_2$ is negligible after taking the limit of $\beta_0\rightarrow 0$. In this case, there are $\beta_0$ coming from outside the loop integral and they happen to be canceled out, but we should pay attention that this definition just include $\beta_0$ from inside the loop integral. We obtain
\begin{eqnarray*}
\lambda^2 I_1(\Delta)&=&\lambda^2\log\left(\mathsmaller{\frac{\Lambda+\sqrt{\Lambda^2+\Delta^2}}{\sqrt{\Delta}}}\right)\simeq
\lambda^2\log\Lambda+...\,,\\
\lambda^2 I_2(\Delta_0)&=&\lambda^2\sum_{n=1}^{N_{\rm max}}\frac{1}{n}=\lambda^2(\log N_{\rm max}+\gamma_{\rm E}).
\end{eqnarray*}
After substituting $\Lambda=\frac{N_{\rm max}}{\beta}$ into $I_1$, it is obviously that the terms with $N_{\rm max}$ are the same in $I_1$ and $I_2$. Thus a clean cancellation, $\lambda^2 I_1(\Delta)-\lambda^2 I_2(\Delta)$, can be fulfilled. As a double check for the integrals without UV divergence, consider an example in the vertex diagram, $\raisebox{-3.mm}{\psfig{figure=vertexphi3.eps, width=2cm,height=1cm}}$, similar to eq. (\ref{vertexqed}):
\begin{eqnarray*}
I_3(\Delta)&=
\frac{1}{\beta}\sum_{n=1}^{N_{\rm max}}\frac{1}{\left(\big(\frac{n}{\beta}\big)^2+\Delta^2\right)^{\frac{3}{2}}}\,\,\left(\simeq \int^\Lambda_0 dx \frac{1}{({x^2+\Delta^2})^{\frac{3}{2}}} ,\,\,{\rm for}\,\,\beta\gg 1 \right),\\
I_4(\Delta_0)&=&\lim_{\beta_0\rightarrow 0}\frac{1}{\beta_0}\sum_{n=1}^{N_{\rm max}}\frac{1}{\left(\big(\frac{n}{\beta_0}\big)^2+\Delta_0^2\right)^{\frac{3}{2}}}=0.
\end{eqnarray*}
For $\beta_0$ inside the loop and with its limit to zero, we will have $I_4$ as a reference point for $I_3$, and $I_4(\Delta_0)= \lim_{\beta_0\rightarrow 0} \sum\frac{\beta_0^2}{n^3}= 0$. The triangular Feynman integral is then
$\lambda^2 I_3(\Delta)$, as the same as the traditional integral. In the summations of the Matsubara frequency in $I_2$ and $I_4$, the limiting process of $\beta_0$ inside the loop is treated as a definition. As we may remember from the conventional calculations of the Casimir effect \cite{casimir}, the difference of the continuous and discrete potential functions is calculated. The discrete potential is due to the discrete mode number of the electromagnetic standing waves between the two plates, as the momentum in the normal direction of the plates is $\frac{2\pi n}{L}$, where $L$ is a small distance, similar to the case of $\beta_0\rightarrow 0$ in the Matsubara frequency $\frac{2\pi n}{\beta_0}$. This could be an analogy for such an assumption and condition. Therefore, it could be tricky to give those reference integrals certain kinds of conditions, but it is also important to see if it is universal to all perturbative theories.\par
Let's also see an example from $\phi^4$-theory. The corresponding interaction Lagrangian is $\mathcal{L}_I=\frac{\lambda}{4!}\phi^4$, where the coupling constant $\lambda$ is dimensionless. For a self-energy diagram, $\raisebox{-3.mm}{\psfig{figure=selfephi4.eps, width=1.6cm,height=1cm}}$, similarly perform the scale transformation on $\phi(x)$ and $\phi(y)$ first, we have
\begin{eqnarray*}
&&\hspace{-0cm}\lambda\int^{\beta_0}_{0}d\tau_{z_{0}} d^3{\bf z}_{0}\,\,\langle 0|\left\{
\acontraction[1ex]{}{\phi}{(x)}{\phi}
\phi(x)\phi_0(z_{0})
\acontraction[1ex]{}{\phi}{_0(z_{0})}{\phi}
\phi_0(z_{0})\phi_0(z_{0})
\acontraction[1ex]{}{\phi}{_0(z_{0})}{\phi}
\phi_0(z_{0})\phi(y)
\right\}|0\rangle,\\
&&\hspace{-0cm}=\lambda \left(\frac{\beta_0}{\beta}\right)^2\frac{1}{\beta_0}\sum_n\int d^3{\bf p}_0\frac{1}{p_{0,n}^2}\left(\frac{1}{\beta_0}\sum_m\int d^3{\bf k}_0 \frac{1}{k^2_{0,m}}\right)\frac{1}{p_{0,n}^2}e^{-ip_{0,n}\cdot (x_0-y_0)},\\
&&\hspace{-0cm}=\left(\frac{\beta_0}{\beta}\right)^2\frac{1}{\beta}\sum_n\int d^3{\bf p}\frac{1}{p_{n}^2}\left\{\frac{\lambda}{\beta_0}\sum_m\int d^3{\bf k}_0 \frac{1}{k^2_{0,m}}\right\}\frac{1}{p_{n}^2}e^{-ip_{n}\cdot (x-y)}.
\end{eqnarray*}
The part in the curly bracket is the imaginary-time Feynman integral for $\beta_0$. We may notice that there is an extra factor
$\big(\frac{\beta_0}{\beta}\big)^2$ from outside the loop, and it has to be taken into account in the result.
From the above two examples. we may generalize the above the derivation by considering
\begin{eqnarray*}
\langle 0|\phi(x_1)...\phi(x_E)\prod_{i=1}^V \int d^dz_iH_I(z_i)|0\rangle,
\end{eqnarray*}
where $E$ is the number of the external legs and $V$ is the number of the vertices. The extra factors of $\big(\frac{\beta}{\beta_0}\big)$ are given by
\begin{eqnarray*}
\left(\frac{\beta}{\beta_0}\right)^{V\cdot \lambda_d} \cdot \left(\frac{\beta_0}{\beta}\right)^{E} \cdot
\left(\frac{\beta}{\beta_0}\right)^{4E-4} \cdot \left(\frac{\beta}{\beta_0}\right)^{-2E},
\end{eqnarray*}
where $\lambda_d$ is the mass dimension of the coupling constant.
The first is from all of the coupling constants, and the second is from the scale transformation of $E$ external field operators $\phi(x)=\frac{\beta_0}{\beta}\phi_0(x_{0})$. The third is due to the integrations of external momentum space and $-4$ in the exponent is from the $\delta$-function to ensure momentum and frequency conservations; the external propagators give the last ratio factor. To conclude, the extra factor that has to be taken into account for $\beta_0$ is
\begin{eqnarray*}
\left(\frac{\beta}{\beta_0}\right)^{V\cdot \lambda_d+E-4}.
\end{eqnarray*}
Thus, for the diagram of $\phi^3$-theory $\,\raisebox{-3.mm}{\psfig{figure=selfephi3a.eps, width=2cm,height=0.8cm}}$, we have $V=2$, $\lambda_d=1$ and $E=2$, the factor is one. For the one $\raisebox{-3.mm}{\psfig{figure=vertexphi3.eps, width=2cm,height=1cm}}$, $V=3$, $\lambda_d=1$ and $E=3$, and the factor is $\big(\frac{\beta}{\beta_0}\big)^2$. As for the diagram of $\phi^4$-theory, $\raisebox{-3.mm}{\psfig{figure=selfephi4.eps, width=1.6cm,height=1cm}}$, $V=1$, $\lambda_d=0$ and $E=2$, the factor is $\big(\frac{\beta_0}{\beta}\big)^2$.
\section{Green's function}
\label{greenfunction}
\subsection{Propagator of fermions}
\subsubsection{Real-time}
In terms of the real-time, it is obtained according to an analytic continuation from the imaginary-time $\tau$, after the Matsubara frequencies are summed. As for the quantization of the fields, the creation and annihilation operators $a^s_{\omega_n,{{\bf p}}}$, $b^s_{\omega_n,{{\bf p}}}$ are quantized for the respective Matsubara frequency, $\omega_n$, and 3-momentum ${\bf p}$, and they will be treated as the grassmann numbers latter in the functional formalism.
The field operators are expanded by the Fourier expansion and transform for the respective discrete and continuous phase spaces as
\begin{eqnarray}
\psi(\tau,x)=\frac{1}{\beta}\sum_{n}\int\frac{d^3{\bf p}}{(2\pi)^3}\,\psi(\omega_n,{\bf \vec{p}})e^{-i\omega_n\tau+i\vec{\bf p}\cdot \vec{\bf x}},\nonumber\\
{\rm and }\hspace{.5cm}\psi^\dagger(\tau,x)=\frac{1}{\beta}\sum_{n}\int\frac{d^3{\bf p}}{(2\pi)^3}\,\psi^\dagger(\omega_n,{\bf \vec{p}})e^{i\omega_n\tau-i\vec{\bf p}\cdot \vec{\bf x}},\label{psix}
\end{eqnarray}
where in the momentum phase space:
\begin{eqnarray}
\psi(\omega_n,{\bf p})=\frac{1}{\sqrt{2\xi_{\bf p}}}\sum_s\left(a^s_{\omega_n,{\bf{p}}}u^s({\bf p})+{b^s}^\dagger_{-\omega_n,\bf -p}{{v}^s}({\bf -p})\right),\nonumber\\
{\rm and }\hspace{.5cm}\psi^\dagger(\omega_n,{\bf p})=\frac{1}{\sqrt{2\xi_{\bf p}}}\sum_s\left(b^s_{-\omega_n,\bf -p}{{v}^s}^\dagger({\bf -p})+{a^s}^\dagger_{\omega_n,{\bf{p}}}{u^s}^\dagger({\bf p})\right)
\label{psip}
\end{eqnarray}
The superscript $s$ indicates the spin state of the spinor $u^s({\bf p})$ or $v^s(-{\bf p})$. The spinors satisfy the Dirac equation, such as $(\slashed{p}-m_{\rm f})u({\bf p})=0$ and so on.
In the traditional way, the factor $1/\sqrt{2\xi_{\bf p}}$ is inserted in the expansions, eq. (\ref{psix}), to ensure the Lorentz Invariance. In fact, even without this factor, it can be shown that the Lorentz invariance is still hold for the propagators in this Matsubara frequency expansion and the resultant 2-point correlation function in eq. (\ref{diracprop}), will not be modified. The action, $\mathcal{A}$, from the partition function $Z$ is
\begin{eqnarray*}
\mathcal{A}=\int^{\beta}_0 d\tau\int d^3 \vec{\bf x}\,\,\mathcal{L}_{\rm Dirac}(\tau, \vec{\bf x_r})
= \frac{1}{\beta}\sum_{\omega_n}\int d^3 \vec{\bf p}\,\,\mathcal{L}_{\rm Dirac}(\omega_n, \vec{\bf p}),
\end{eqnarray*}
where the Lagrangians in the respective representations are
\begin{eqnarray}
\mathcal{L}_{\rm Dirac}(\tau, \vec{\bf x})&=&\psi^\dagger(\tau,{\bf x})\left(-\frac{\partial}{\partial \tau}+i\gamma^0\vec{\gamma}\cdot\vec{\nabla}-m_{\rm f}\gamma^0\right)\psi(\tau,\bf{x}),\nonumber\\
\mathcal{L}_{\rm Dirac}(\omega_n, \vec{\bf p})&=&\psi^\dagger(\omega_n,{\bf p})\left(i\omega_n-\gamma^0\vec{\gamma}\cdot\vec{\bf p}-m_{\rm f}\gamma^0\right)\psi(\omega_n,\bf{p}),\label{Lagrangian}
\\
\hspace{-2.cm}&=&\hspace{-.3cm}\sum_{s}\left.\left(i\omega_n-\xi_{\bf p}\right){a^s}^\dagger_{\omega_n,{\bf{p}}}a^s_{\omega_n,{\bf{p}}}\right.
-\left.\sum_{s}\left(i\omega_n+\xi_{\bf p}\right)b^s_{\omega_n,\bf p}{b^s}^\dagger_{\omega_n,\bf p}\right. .\label{Lagrangian2}
\end{eqnarray}
The lines from eq. (\ref{Lagrangian}) to (\ref{Lagrangian2}) are derived by using the Dirac equation, and the factor, $2\xi_{\bf p}$, coming from the spinor products, ${u^r}^\dagger({\bf p})u^s({\bf p})$ and ${v^s}^\dagger({\bf p})v^s({\bf p})$, are canceled by those in eq. (\ref{psip}).
The correlation function for two different imaginary-time and space points, $(\tau_x,{\bf x})$ and $(\tau_y,{\bf y})$, is related to the one for the Matsubara frequency and 3-momentum as follows
\begin{eqnarray*}
&&\langle\bar{\psi}(\tau_x,\vec{\bf x}){\psi}(\tau_y,\vec{\bf y})\rangl
=\frac{1}{\beta^2}\sum_{n,m}\int\frac{d^3{\bf p}}{(2\pi)^3}\frac{d^3{\bf k}}{(2\pi)^3}
\langle \bar{\psi}(\omega_n,{\bf p})\psi(\omega_m,{\bf k})\rangle
e^{-i\omega_n\tau_x+i\omega_m\tau_y+i{\bf p}\cdot {\bf x}-i{\bf k}\cdot {\bf y}}.
\end{eqnarray*}
As the creation and annihilation operators are regarded as grassmann numbers, the partition function is treated as a functional of them; the correlation function for the Matsubara frequency and 3-momentum can be obtained by inserting eq. (\ref{psip}) into:
\begin{eqnarray*}
&&\langle {\psi}(\omega_n,{\bf p})\bar{\psi}(\omega_m,{\bf k})\rangle \\
&=&
-\sum_{s}u^s({\bf p})\bar{u}^s({\bf p})\frac{\beta\delta_{mn}(2\pi)^3\delta^3({\bf p-k})}{2\xi_{\bf p}(i\omega_n-\xi_{\bf p})}-\sum_{s}v^s({\bf -p})\bar{v}^s({\bf -p})\frac{\beta\delta_{mn}(2\pi)^3\delta^3({\bf p-k})}{2\xi_{\bf p}(i\omega_n+\xi_{\bf p})}.
\end{eqnarray*}
As for the case of $\tau_x>\tau_y$, we may obtain the retarded propagator by summing over the Matsubara frequency with the help of eq. (\ref{sumfermi}), and choose the semicircle below the real axis of $p_0$ for the contour. This is achievable as the analytic continuation of the variable $\tau= it$ is applied. The retarded fermion propagator is derived from
\begin{eqnarray}
&&\hspace{-1.5cm}\langle{\psi}(\tau_x,\vec{\bf x})\bar{\psi}(\tau_y,\vec{\bf y})\rangle_{\rm Ret}
=\int_{\otimes}\frac{d^4 p}{(2\pi)^4}\frac{i}{\slashed{p}-m_{\rm f}}e^{-p_0(\tau_x-\tau_y)+i{\bf p}\cdot ({\bf x-y})}(1-n_{\rm F}(p_0)),\label{diracpropR}
\end{eqnarray}
where we have used the relations
\begin{eqnarray*}
\sum_s u^s({\bf p})\bar{u}^s({\bf p})&=&\xi_{\bf p}\gamma_0-\vec{\bf p}\cdot \vec{\gamma}+m_{\rm f} =\slashed{p}+m_{\rm f},\\
\sum_s v^s({\bf p})\bar{v}^s({\bf p})&=&\xi_{\bf p}\gamma_0-\vec{\bf p}\cdot \vec{\gamma}-m_{\rm f} =\slashed{p}-m_{\rm f},
\end{eqnarray*}
and the formulas in Appendix \ref{appD} for the sum of Matsubara frequencies for fermions. The notation, $\int_\otimes$, indicates that the contributions from the poles of the density function have to be excluded by either adjusting the contour off them, such as illustrated in fig. \ref{Fig:threshold} (b), or removing the contributions from the enclosed residues. For fermions, there are poles at $p_0=\pm\frac{\pi}{\beta}$, $\pm\frac{3\pi}{\beta}$... . For the other imaginary-time ordering of the field operators, $\tau_y>\tau_x$,
\begin{eqnarray*}
&&\langle \bar{\psi}_{\alpha}(\omega_m,{\bf k}){\psi}_{\beta}(\omega_n,{\bf p})\rangle \\
&=&
\sum_{s}u^s_\beta({\bf p})\bar{u}^s_\alpha({\bf p})\frac{\beta\delta_{mn}(2\pi)^3\delta^3({\bf p-k})}{2\xi_{\bf p}(i\omega_n-\xi_{\bf p})}+\sum_{s}v^s_\beta({\bf -p})\bar{v}^s_\alpha({\bf -p})\frac{\beta\delta_{mn}(2\pi)^3\delta^3({\bf p-k})}{2\xi_{\bf p}(i\omega_n+\xi_{\bf p})}.
\end{eqnarray*}
The chosen contour is the upper semicircle of the $p_0$-complex plane.
\begin{eqnarray}
&&\hspace{-1cm}\langle\bar{\psi}_{\alpha}(\tau_y,\vec{\bf y}){\psi}_{\beta}(\tau_x,\vec{\bf x})\rangle_{\rm Adv}
=\int_{\otimes}\frac{d^4 p}{(2\pi)^4}\frac{i}{\slashed{p}-m_{\rm f}}e^{p_0(\tau_x-\tau_y)+i{\bf p}\cdot ({\bf x-y})}n_{\rm F}(p_0).\label{diracpropA}
\end{eqnarray}
eq. (\ref{diracpropR}) and (\ref{diracpropA}) have shown the retarded and advanced propagators for fermions.
By choosing the Feynman boundary conditions as in the field theory \cite{peskin95}, we may define the corresponding Feynman propagator as follows
\begin{eqnarray*}
S_F(\tau_x-\tau_y, {\bf x- y})& \equiv &\Theta (\tau_x-\tau_y)\langle0|\psi(\tau_x,{\bf x})\bar{\psi}(\tau_y,{\bf y})|0\rangle-\Theta (\tau_y-\tau_x)\langle0|\bar{\psi}(\tau_y,{\bf y})\psi(\tau_x,{\bf x})|0\rangle\\
&=&\int_{\otimes} \frac{d^4{ p}}{(2\pi)^4}e^{-p_0(\tau_x-\tau_y)+i{\bf p}\cdot ({\bf x}-{\bf y})}\frac{i}{\slashed{p}-m_{\rm f}+i\varepsilon}\left(1-n_F(p_0)\right).
\end{eqnarray*}
The analytic continuation of the imaginary-time $\tau$ to the real-time $t$ by making $\tau=it$ may be applied as soon as the Matsubara frequencies are summed, there is no confusion that we do the replacement now
\begin{eqnarray}
S_F(t_x-t_y, {\bf x- y})
&=&\int_\otimes \frac{d^4{ p}}{(2\pi)^4}e^{-i p\cdot (x-y)}\frac{i}{\slashed{p}-m_{\rm f}+i\varepsilon}\left(1-n_F(p_0)\right).\label{diracprop}
\end{eqnarray}
In the limit of $\beta \rightarrow \infty$, the above Feynman propagator becomes the one that we are familiar with. The density function becomes irrelevant as to suppress the violation of the Lorentz invariance. In the tree-level, the four momentum integral $\int_\otimes \frac{d^4{ p}}{(2\pi)^4} \rightarrow \int \frac{d^4{ p}}{(2\pi)^4}$ is a usual one, since the poles do not lie in the real axis of $p^0$.
\begin{figure}[t]
\begin{center}
\subfigure[]{\includegraphics[width=5cm]{feynD1rev.eps}}
\hspace*{0.07\textwidth}
\subfigure[]{\includegraphics[width=5cm]{feynD2rev.eps}}
\hspace*{0.0\textwidth}
\caption{\small The Feynman diagrams are drawn for a fermion being produced and propagating through space between two space-time points, $(t_x, {\bf x} )$ and $(t_y, {\bf y} )$. (a) for $t_x>t_y$, the chance for a fermion with a momentum $p$ to be created at $y$ and annihilated at $x$ is proportional to $1-n_F(p_0)$, since there is filled with $n_F(p_0)$ of fermions. (b) for $t_x<t_y$, as a fermion is annihilated at $y$ and created at $x$, the probability is proportional to $n_F(p_0)$, because it cannot be annihilated without the existence in the first place. }
\label{Fig:feynDF}
\end{center}
\end{figure}
\subsubsection{Imaginary-time}
From the functional approach, the field operators are treated as grassmann numbers without the use of the spinors. According to the Lagrangian in eq. (\ref{Lagrangian}), the correlation function for the imaginary-time can also be obtained directly:
\begin{eqnarray*}
&&\hspace{-.5cm}\langle \bar{\psi}(\omega_n,{\bf p})\psi(\omega_m,{\bf k})\rangle
=\beta \delta_{nm} (2\pi)^3\delta^{(3)}({\bf p-k})\frac{1}{(i\omega_{n}\gamma_0 -\vec{\gamma}\cdot\vec{{\bf p}}-m_{\rm f})}.
\end{eqnarray*}
Followed by the same approaches in the field theory, the Feynman propagator between the two points, $(\tau_x, {\bf x})$ and $(\tau_y, {\bf y})$, in the imaginary-time and space is
\begin{eqnarray}
S_{\rm F}(\tau_x-\tau_y, {\bf x- y})
=\frac{1}{\beta}\sum_{n={\rm odd}}\int \frac{d^3{ \bf p}}{(2\pi)^3}e^{-i\omega_n(\tau_x-\tau_y)+i{\bf p}\cdot ({\bf x}-{\bf y})}\frac{1}{(i\omega_{n}\gamma_0 -\vec{\gamma}\cdot\vec{{\bf p}}-m_{\rm f})}.\hspace{.5cm}\label{Imtpropfermi}
\end{eqnarray}
\subsection{Propagator of photons}
From the Lagrangian densities in eq. (\ref{Lphoton}), the partition function of the photon field is known as
\begin{eqnarray}
Z&=&\int e^{\int^{\beta}_0 d\tau \int d^3{\bf x}\left(-\frac{1}{4}F^{\mu\nu}F_{\mu\nu}-\frac{\zeta}{2}\partial_\mu A^\mu\partial_\nu A^\nu+\mu_\gamma A_\mu A^\mu\right)}[dA_\sigma(\tau,{\bf x}) dA^\sigma(\tau,\bf{x})],\label{PhotonZ}
\end{eqnarray}
where $\zeta$ is the gauge parameter. The choice of the gauge will be $\zeta=1$ in the following calculations. The role that the chemical potential $\mu_\gamma$ plays is similar to the mass squared, $m^2_\gamma=\frac{\mu_\gamma}{2}$. This term will be ignored in the derivation of photon's propagators and will be considered after the chemical potential is computed in Section \ref{chempot}.
\subsubsection{Real-time}
\begin{figure}[t]
\begin{center}
\subfigure[]{\includegraphics[width=5cm]{feynD3.eps}}
\hspace*{0.07\textwidth}
\subfigure[]{\includegraphics[width=5cm]{feynD4.eps}}
\hspace*{0.0\textwidth}
\caption{\small Feynman diagrams are illustrated for a boson being produced and propagating through space between two space-time points, $(t_x, {\bf x} )$ and $(t_y, {\bf y} )$. (a) for $t_x>t_y$, the chance for a boson with a momentum $q$ to be annihilated at a latter time, $t_x$, is proportional to $1+n_B(q_0)$, since there are $1+n_B(q_0)$ of bosons in the vacuum. (b) for $t_x<t_y$, as a boson to be annihilated at $t_x$ ahead of the event at $t_y$, the chance is proportional to $n_B(q_0)$, because it can only annihilated $n_B(q_0)$ of bosons before $t_y$. }
\label{Fig:feynDF}
\end{center}
\end{figure}
One of the differences in quantizing the photon fields is that it is expanded by the Matsubara frequency, $\omega_n\,\,(=\frac{2\pi n}{\beta})$, instead of the energy.
The other is that the expansion is the Fourier expansion without intentionally adding the factor $\frac{1}{\sqrt{2E}}$ to ensure the Lorentz Invariance. It can be shown later that the Lorentz invariance is secured as $\beta\rightarrow \infty$, like in the case of fermion's. The field operator of photons is expanded with respective to the Matsubara frequency and the 3-momentum as follows
\begin{eqnarray}
A_\mu(\tau,{\bf x}
&\equiv&\frac{1}{\beta}\sum_n \int \frac{d^3 {\bf q}}{(2\pi)^3}A_\mu(\omega_n,{\bf q})e^{-i\omega_n\tau+i{\bf q\cdot x}},\label{photonex}
\end{eqnarray}
where the field operator in the momentum representation is
\begin{eqnarray*}
A_\mu(\omega_n,{\bf q})&=&\frac{1}{\sqrt{2|{\bf q}|}}
\sum^3_{\lambda=0}
\left(a^\lambda_{\omega_n,{\bf q}}\epsilon^\lambda_\mu({\bf q})+
a^{\lambda\dagger}_{-\omega_n,{\bf -q}}\epsilon^{\lambda*}_\mu({\bf -q})\right).
\end{eqnarray*}
The action, $i\mathcal{A}$, of the photon fields, is
\begin{eqnarray*}
i\mathcal{A}=\int^{\beta}_0 d\tau \int d^3 \vec{\bf x}\,\,\mathcal{L}_0(A_\mu,\tau, \vec{\bf x})
= \frac{1}{\beta}\sum_{\omega_n}\int d^3 \vec{\bf q}\,\,\mathcal{L}_0(a^\lambda_{\bf q}, a^{\lambda\dagger}_{\bf q},\omega_n, \vec{\bf q}),
\end{eqnarray*}
where the Lagrangian of free photons, denoted as $\mathcal{L}_0$, includes only the first two terms in eq. (\ref{Lphoton}). The term with the chemical potential density $\mu_\gamma$ is dropped temporarily and will be taken into consideration later in the computation of the photon's self-energy. We choose the gauge parameter to be $\zeta=1$, the Lagrangian in the momentum space is
\begin{eqnarray*}
\mathcal{L}_0(a^\lambda_{\bf q}, a^{\lambda\dagger}_{\bf q}, \omega_n,\vec{\bf q})
&=&\frac{1}{4|{\bf q}|}(\omega_n^2+|{\bf q}|^2)\sum_\lambda (-g_{\lambda\lambda})
\left(a^\lambda_{\omega_n,{\bf q}} a^{\lambda\dagger}_{\omega_n,{\bf q}} + a^{\lambda\dagger}_{\omega_n,{\bf q}} a^\lambda_{\omega_n,{\bf q}}\right),
\end{eqnarray*}
where the matrix $g_{\lambda\lambda'}={\rm diag}(1,-1,-1,-1)$. The propagator of photons in two representations can be related by
\begin{eqnarray*}
\hspace{-0.8cm}\left\langle A_\mu(\tau_x,\vec{\bf x}) A_\nu(\tau_y,\vec{\bf y})\right\rangle
=\frac{1}{\beta^2}\sum_{n,m}\int \frac{d^3 {\bf q}}{(2\pi)^3} \frac{d^3 {\bf k}}{(2\pi)^3}
\left\langle A_\mu(\omega_n,{\bf q}) A_\nu(\omega_m,{\bf k})\right\rangle e^{-i\omega_n\tau_x-i\omega_m\tau_y+i{\bf q\cdot x}+i{\bf k\cdot y}}.
\end{eqnarray*}
The result of the summation for the polarization vectors is $\sum_{\lambda=0}^{3}g_{\lambda\lambda}\epsilon^\lambda_\mu \epsilon^\lambda_\nu=g_{\mu\nu}$.
The two-point correlation function for the momentum representation can be derived as follows:
\begin{eqnarray}
\hspace{-1cm}\left\langle A_\mu(\omega_n,{\bf q}) A_\nu(\omega_m,{\bf k})\right\rangl
&=&\beta\delta_{-n,m}(2\pi)^3\delta^3({\bf q+k })\frac{-g_{\mu\nu}}{2|{\bf q}|}\left\{\frac{1}{i\omega_n+|{\bf q}|}-
\frac{1}{i\omega_n-|{\bf q}|} \right\}.\nonumber\\
\label{PhotonProp1}
\end{eqnarray}
The sum of the polarization vectors is replaced by a negative metric tensor. When computing the propagator in the imaginary-time and space representation, assume $\tau_x>\tau_y$ and sum over the Matsubara frequency, then we may obtain a two-point correlation function of a 4-momentum integral by introducing a complex integral for a variable $q_0$ with a semi-circle contour in the lower $q_0$-complex plane. We use the formulas in eq. (\ref{sumbose}) in Appendix \ref{appD} to sum over the Matsubara frequencies for bosons.
The retarded propagator is
\begin{eqnarray*}
\left\langle A_\mu(\tau_x,\vec{\bf x}) A_\nu(\tau_y,\vec{\bf y})\right\rangle_{\rm Ret}
&=&\int_\otimes \frac{d^4 { q}}{(2\pi)^4} \frac{-ig_{\mu\nu}}{q^2}
\left(1+\frac{1}{e^{\beta q_0}-1}\right)
e^{-q_0(\tau_x-\tau_y)+i{\bf q\cdot( x-y)}}.
\end{eqnarray*}
The contour is chosen to enclose the two residues of the poles at $q_0=-|{\bf q}|$ and $|{\bf q}|$. The resultant expression is similar to the one in the field theory except the factor $1+n_B(q_0)$. In the limit of large value of $\beta$, the density function becomes unity as in the case of fermions. The integral sign, $\int_\otimes$, remind us of that the poles of the density function have not to be enclosed by the contour or their residues inside the contour have to be excluded; this will be taken into account when the radiative corrections are computed in the following sections. Analytically continuation from the imaginary-time to the real-time is made by letting $\tau=i t$; the corresponding Feynman propagator can be obtained in a similar manner to the retarded one:
\begin{eqnarray*}
D_{{\rm F }}^{\mu\nu}(t_x-t_y, {\bf x- y})
&=&\int_\otimes \frac{d^4 { q}}{(2\pi)^4} \frac{-ig^{\mu\nu}}{q^2+i\varepsilon}
\left(1+n_B(p_0)\right)
e^{-i{ q\cdot( x-y)}}.
\end{eqnarray*}
\subsubsection{Imaginary-time}
Like fermion's propagator of the imaginary-time in eq. (\ref{Imtpropfermi}), the photon's imaginary-time propagator is derived from the Lagrangian of photons in eq. (\ref{Lphoton}):
\begin{eqnarray}
D_{{\rm }}^{\mu\nu}(\tau_x-\tau_y, {\bf x- y})
=\frac{1}{\beta}\sum_n\int \frac{d^3{ \bf q}}{(2\pi)^3}e^{-i\omega_n(\tau_x-\tau_y)+i{\bf q}\cdot ({\bf x}-{\bf y})}\frac{-1}{q^2_n}\left(g^{\mu\nu}-(\zeta-1)\frac{q_n^\mu q_n^\nu}{q^2_n}\right),\hspace{.6cm}\label{}
\end{eqnarray}
where $q_n=(i\omega_n, {\bf q})$. The corresponding Matsubara frequency, $\omega_n=\frac{2\pi n}{\beta}$, and $n$ is an integer. As mentioned, the chemical potential has not been included in the above, and will be included in the computations of radiative corrections.
\section{One-loop radiative corrections}
\label{oneloop}
\subsection{Self-energy of photons}
\subsubsection{Real-time}\label{photonreal}
\begin{figure}[t]
\begin{center}
\subfigure[]{\includegraphics[width=5cm]{sephoton1b.eps}}
\hspace*{0.07\textwidth}
\subfigure[]{\includegraphics[width=5cm]{sephoton2b.eps}}
\hspace*{0.0\textwidth}
\caption{\small One-loop self-energy Feynman diagrams of a photon for different time ordering of the interaction vertices: (a) A photon annihilates first and creates an electron and a positron, later the two annihilate and a photon is created, the probability for this process is proportional to the product of $1-n_{\rm F}(k_0+p_0)$ and $n_{\rm F}(k_0)$. (b) A photon disappears due to the annihilation of an electron and a positron in the background, and its momentum has been carried away by another photon that is produced earlier. The probability is proportional to the product of $1-n_{\rm F}(-k_0)$ and $n_{\rm F}(-k_0-p_0)$. }
\label{Fig:feynSEphoton}
\end{center}
\end{figure}
The radiative corrections in QED with the propagators that are formulated in the previous section will be applied on loop calculations. Feynman rules are similar except the extra density functions that are attached to each propagating particle and the redundant residues from the same density functions that have to be carefully dealt with. As a photon carries a momentum, $p^\mu$, an electron and a positron are created and annihilates with momenta $k^\mu+p^\mu$ and $k^\mu$, as shown in fig. \ref{Fig:feynSEphoton}. The 4-velocity of the whole statistical system measured by an observer is denoted as $u^\mu$. In the theory of relativity, the energy of a particle, with a 4-momentum $p^\mu$, measured by the system is written as a scalar product, $(p\cdot u)$. As the observer is at rest with respective to the system, $u^\mu=(1,0,0,0)$, the energy of the particle then happens to be $p^0$. In this section, the calculations for the self-energy of a photon may be proceeded in a similar way from the those in ref. \cite{peskin95},
with $\Delta=m^2_{\rm f}-x(1-x)p^2 $, after the Feynman parametrization is applied and the loop momentum is shifted from $q^\mu$ to $l^\mu$:
\begin{eqnarray}
i\Pi_2^{\mu\nu}(p)=-4e^2\int^1_0 dx\int_\otimes\frac{d^4 l}{(2\pi)^4}\frac{2l^\mu l^\nu -g^{\mu\nu}l^2-2x(1-x)p^\mu p^\nu+g^{\mu\nu}(m^2_{\rm f}+x(1-x)p^2)}{(l^2-\Delta)^2}\nonumber\\
\hspace{0cm}\times
\frac{1}{e^{-\beta l\cdot u-\log b}+1}\frac{1}{e^{\beta l\cdot u-\log a}+1},\hspace{.5cm}
\label{SEofphoton}
\end{eqnarray}
where $l^\mu=k^\mu+x p^\mu$, $a=e^{\beta x p\cdot u}$, $b=e^{\beta(1- x)p\cdot u}$ and $x$ is the Feynman parameter. In general, the integrals with different powers of the denominator after the Feynman parameterization are in the form:
\begin{eqnarray}
&&\hspace{-1.5cm}\int_\otimes \frac{d^4 l}{(l^2-\Delta)^\lambda}\left\{1-n_{\rm F}((k+p)\cdot u)\right\}\left\{1-n_{\rm F}(-k\cdot u)\right\}\nonumber\\
&&\hspace{-1.5cm}
=\int_\otimes \frac{d^4 l}{(l^2_0-{\bf l}^2-\Delta)^\lambda}\frac{1}{e^{-\beta l\cdot u-\log b}+1}\frac{1}{e^{\beta l\cdot u-\log a}+1}.\label{photonInt}
\end{eqnarray}
After applying the Wick rotation and taking into account the shifting of the poles from the density functions, the integral becomes
\begin{eqnarray*}
&&\hspace{-1.cm}i\oint_{\rm RC,LC} \frac{(-1)^\lambda d l_{\rm E}d^3 {\bf l}}{(l^2_{\rm E}+{\bf l}^2+\Delta)^\lambda}\frac{1}{e^{i\beta l_{\rm E}-\log a}+1}\frac{1}{e^{-i\beta l_{\rm E}-\log b}+1}-({\rm residues\,\,of\,\,\,2\,\,n_F }),\\
&&\hspace{-1.5cm}=
\begin{cases}
i\oint_{\rm RC} d l_{\rm E}\int d^3 {\bf l} \frac{(-1)^\lambda}{(l^2_{\rm E}+{\bf l}^2+\Delta)^\lambda}\frac{1}{e^{i\beta l_{\rm E}-\log a}+1}\frac{1}{e^{-i\beta l_{\rm E}-\log b}+1}+2\pi i\frac{ab}{(1-ab)\beta}\sum_{n\in {\rm I}}\int d^3{\bf l}\frac{(-1)^\lambda}{\left[\Delta+{\rm l}^2+(\frac{(2n+1)\pi}{\beta}-i\frac{\log a}{\beta})^2\right]^\lambda}\\
i\oint_{\rm LC} d l_{\rm E}\int d^3 {\bf l} \frac{(-1)^\lambda}{(l^2_{\rm E}+{\bf l}^2+\Delta)^\lambda}\frac{1}{e^{i\beta l_{\rm E}-\log a}+1}\frac{1}{e^{-i\beta l_{\rm E}-\log b}+1}-2\pi i\frac{ab}{(ab-1)\beta}\sum_{n\in {\rm I}}\int d^3{\bf l}\frac{(-1)^\lambda}{\left[\Delta+{\rm l}^2+(\frac{(2n+1)\pi}{\beta}-i\frac{\log b}{\beta})^2\right]^\lambda},
\end{cases}
\end{eqnarray*}
\begin{figure}[t]
\begin{center}
\subfigure[]{\includegraphics[width=5.5cm]{contourpole0.eps}}
\hspace*{0.0\textwidth}
\subfigure[]{\includegraphics[width=5.5cm]{contourpole1.eps}}
\hspace*{0.0\textwidth}
\caption{\small (a) The traditional contour are drawn after the Wick rotation is applied. The contour in blue, including (A) and (B), is running clockwise and those in red for the poles from one of the density function are counterclockwise. (b) The closed contour is represented by the blue lines with four segments (A), (B), (C) and (D). The integral in eq. (\ref{Eq:integralL}) is for the contours (A) and (B). The thick line of (E) is the branch cut of the integrand in the complex plane of $l_0$ as $\Delta<0$. The value $\Lambda_t$ is the cutoff for the energy. }
\label{Fig:threshold}
\end{center}
\end{figure}
where we have changed the variable by making $l_0=il_{\rm E}$. Before the Wick rotation being applied, the residues of the poles from the density functions are either above or below the real axis of the complex variable $l_0$, as illustrated in fig. \ref{Fig:threshold}. After the rotation, the poles that have to be taken into account are either in the right- or left-hand side of the imaginary axis, depending on which contour is adopted. As $\beta \gg 1$, for the integration along the imaginary axis, the density functions in the first integral of both contours become unity, it changes to a traditional Feynman integral used in field theory. In practical calculations, the formulas below eq. (\ref{photonInt}) include an integration over a closed contour and a series of residues, and the contour can be chosen as the outlines of the rectangular box, (A), (B), (C) and (D), as shown in fig. \ref{Fig:threshold}, where the contour (B) runs down the complex plane and gets around the poles. The integral for the contour (B) comprises two parts, one is for the vertical lines between two poles, (B.1), and the other is the semicircles to avoid the poles, (B.2), whose resides happen to be half values of the full poles. For the first, take the right contour as an example, the parametrization of the contour is $l_0=il_{E}+(\log a/\beta)$, where $l_E$ is from $(2n-1)\pi/\beta$ to the next pole $(2n+1)\pi/\beta$. The factor in the integrand of eq. (\ref{photonInt}), $1/(e^{\beta l_0-\log a}+1)$, following the contour is
\begin{eqnarray*}
\frac{1}{e^{i\beta l_E}+1}=\frac{1}{2}-i\frac{\sin\beta l_E}{2+2\cos\beta l_E},
\end{eqnarray*}
while other density function along the contour (B.1), $1/(e^{i\beta l_E-\log a-\log b}+1)$ becomes unity for large value of $\beta$.
The imaginary part are canceled by the upper and the lower halves of the contour (B), and can be ignored from now. Therefore, the sum of the integral ng all of the vertical lines, (B.1), is half size of the value for that of the contour (A) due to the factor $\frac{1}{2}$ in the real part. For the semicircles, the density function gives a factor $-2\pi/\beta$ to the residue for each pole, and it can be regarded as $-\Delta l_E$ as $\beta\gg 1$. The contribution from the half residues happens to be same as that from the first part, therefore the combined result for the integral along the contour (B) can be expressed similar to the that along the imaginary axis as it runs through the contour, $l_0=il_E+\log a/\beta$. As we choose the new contour, the pole of the factor $1/(l_0^2-{\bf l}^2-\Delta)$, as indicated by (F) in the above figure, could be outside the rectangular box, and its residue is proportional to
\begin{eqnarray}
\frac{1}{2\sqrt{{\bf l}^2+\Delta}}\frac{1}{e^{-\beta \sqrt{{\bf l}^2+\Delta}-\log b}+1}\frac{1}{e^{\beta \sqrt{{\bf l}^2+\Delta}-\log b}+1}.\label{pole1}
\end{eqnarray}
The residue is vanishing due to the last factor in eq. (\ref{pole1}) for large loop momentum ${\bf l}^2$ and large $\beta$; the same can also be applied for the LC. Based on the above discussions, the result of eq. (\ref{photonInt}) for the contours (A) and (B) can be written as
\begin{eqnarray}\label{Eq:integralL}
&&\hspace{-1cm}(A)+(B)=
i\int^{\Lambda_t}_{-\Lambda_t} d l_{\rm E}\int d^3 {\bf l} \frac{(-1)^\lambda}{(l^2_{\rm E}+{\bf l}^2+\Delta)^\lambda}+i\int^{\frac{2\pi N_m}{\beta}}_{-\frac{2\pi N_m}{\beta}}d l_{\rm E}\int d^3{\bf l}\frac{(-1)^{\lambda+1}}{\left[\Delta+{\bf l}^2+\left(l_{\rm E}-i\frac{\log a}{\beta} \right)^2\right]^\lambda}.\nonumber
\end{eqnarray}
The integrals along the contours, (B.1) and (B.2), are combined into the second integral in the above.
The integrations of the loop momentum that we will apply are separated into two steps, the first is to integrate the 3-momentum, then the energy, instead of being treated equivalently in the field theory.
For the case of $\lambda=1$, and use the cutoffs, $\Lambda_t=\Lambda=\frac{2\pi N_{\rm max}}{\beta}$, which are introduced in Section \ref{scaleinv}, the two integrals of (A) and (B) become
\begin{eqnarray}
&&\hspace{-1cm}(A)+(B)=
-i\int^{\Lambda_t}_{-\Lambda_t} d l_{\rm E}\int d^3 {\bf l} \frac{1}{(l^2_{\rm E}+{\bf l}^2+\Delta)}+i\int^{\frac{2\pi N_m}{\beta}}_{-\frac{2\pi N_m}{\beta}}d l_{\rm E}\int d^3{\bf l}\frac{1}{\left[\Delta+{\bf l}^2+\left(l_{\rm E}-i\frac{\log a}{\beta} \right)^2\right]}\nonumber\\
\label{Eq:IntL1}
&&\hspace{-.0cm}=
-4\pi i\left[\left(\frac{\Lambda \Lambda_t }{(\Delta+\Lambda^2+\Lambda_t^2)}-\frac{\Lambda_t}{\sqrt{\Delta +\Lambda_t^2}}\tan^{-1}\left(\frac{\Lambda}{\sqrt{\Delta+\Lambda^2_t}}\right)
\right)\frac{\log^2 a}{\beta^2}+\cdots\right].
\end{eqnarray}
As the cutoff $\Lambda_t$ and $\Lambda$ are of the same size, from power countings, the coefficient of the order $\frac{\log^4 a}{\beta^4}$ is proportional to $\frac{1}{\Lambda^2}$, so it is negligible, as well as those of higher orders. The correction to the order of $p^2_0\,(\propto\frac{\log^2 a}{\beta^2})$ gives a non-negligible contribution; the value in the parenthesis of order $p_0^2$ is $-0.285398$ when $\Delta$ is neglected. The mass of the photon is expected to be $<10^{-18}{\rm eV}$ \cite{tu05}, so this non-trivial corrections needs carefully examinations to see if it can be removed. As we discussed the contour in fig. \ref{Fig:threshold}, the segmants of (C) and (D) are not included in the discussion so far, and it can be proved that the contributions from these two are of the same size as (A)+(B) up to a minus sign. For the contours of (C) and (D), the integrals are
\begin{eqnarray*}
(C)+(D)&=&-4\pi\int^{-\frac{\log a}{\beta}}_0dx\sqrt{\Delta+(\Lambda_t+ix)^2}\tan^{-1}\left(\frac{\Lambda}{\sqrt{\Delta+(\Lambda_t+i x)^2}}\right)\\
&&\hspace{0.5cm}-4\pi\int^0_{-\frac{\log a}{\beta}}dx\sqrt{\Delta+(-\Lambda_t+ix)^2}\tan^{-1}\left(\frac{\Lambda}{\sqrt{\Delta+(-\Lambda_t+i x)^2}}\right)\\
&&\hspace{-2cm}=\,\,4\pi i\left[ \left(\frac{\Lambda_t\Lambda}{\Delta+\Lambda^2+\Lambda_t^2}-
\frac{\Lambda_t}{\sqrt{\Delta +\Lambda_t^2}}\tan^{-1}\left(\frac{\Lambda}{\sqrt{\Delta+\Lambda^2_t}}\right)
\right)\frac{\log^2 a}{\beta^2
+O\left(\frac{\log^4 a}{\beta^4}\right)\right],
\end{eqnarray*}
which happens to cancel the correction in eq. (\ref{Eq:IntL1}). As it applies to the self-energy of the photon, the combined results contribute zero to the next-leading order. As for $\lambda=2$, the corresponding integrals for the contours (A) and (B), are
\begin{eqnarray*}
&&\hspace{-1cm}(A)+(B)=
i\int^{\Lambda_t}_{-\Lambda_t} d l_{\rm E}\int d^3 {\bf l} \frac{1}{(l^2_{\rm E}+{\bf l}^2+\Delta)^2}-i\int^{\frac{2\pi N_m}{\beta}}_{-\frac{2\pi N_{\rm max}}{\beta}}d l_{\rm E}\int d^3{\bf l}\frac{1}{\left[\Delta+{\bf l}^2+\left(l_{\rm E}-i\frac{\log a}{\beta} \right)^2\right]^2}\\
&&\hspace{-1cm}
=2\pi i\left[\left(-\frac{1}{\Lambda\Lambda_t}+\frac{2\Lambda\Lambda_t}{(\Lambda^2+\Lambda^2_t)^2}+\frac{\Lambda_t}{\Lambda(\Lambda^2+\Lambda^2_t)}-\frac{1}{\Lambda_t^2}\tan^{-1}\left(\frac{\Lambda}{\Lambda_t}\right)\right)\frac{\log^2a}{\beta^2
+O\left(\frac{\log^4 a}{\beta^4}\right)\right].
\end{eqnarray*}
The result of the above approaches zero once the cutoffs are taken to infinity, and that of (C) and (D) can be proved to be vanished in the same manner. Therefore, the whole integral for different $\lambda$ contributes nothing to the self-energy as they are used in eq. (\ref{SEofphoton}). The only nonzero contribution to the self-energy happens when $\Delta < 0$ and the branch cut appears along the contour (A).
Since the integral along contour (A) is identical to the traditional loop integral as $\beta\gg 1$, it gives the same imaginary part over the threshold as that from the field theory.
\subsubsection{Imaginary-time}\label{photonimg}
The one-loop radiative corrections to the self-energy of photon are computed according to the Lagrangian in eq. (\ref{Lphoton}) with the gauge $\zeta=1$. From the corresponding Feynman rules, the self-energy to the next-leading order, $\Omega_{2,\beta}^{\mu\nu}$, is
\begin{eqnarray}
\Omega_{2,\beta}^{\mu\nu}(p_n{\bf })=(-1)e^2\frac{1}{\beta}\sum_{m}\int\frac{d^3{\bf q}}{(2\pi)^3}
{\rm Tr}\left[\gamma^\mu\frac{1}{\slashed{q}_m-m_{\rm f}}\gamma^\nu\frac{1}{(\slashed{p}_n-\slashed{q}_m)-m_{\rm f}}\right],\label{omega}
\nonumber
\end{eqnarray}
where $p^0_m=i\frac{\omega_m}{\beta}$ and $\omega_m=\frac{2\pi}{\beta}(m+\frac{1}{2})$. The minus sign is added for a closed fermion loop. The self-energy of the photon is taking a general form
\begin{eqnarray}
&&\hspace{-0.7cm}\Omega_{2,\beta}^{\mu\nu}(p_n)\equiv\left(g^{\mu\nu}-\frac{p^\mu_n p^\nu_n}{p^2_n}\right)\left(p^2_n\Omega_T(p^2_n)-M(p^2_n)\right)+\frac{p^\mu_np^\nu_n}{p^2_n}\left(p^2_n\Omega_L(p^2_n)-M(p^2_n)\right),
\hspace{1cm}\label{SEphoton}
\end{eqnarray}
We make a shorthand for the factor
$\Delta(p_n^2)=m^2_{\rm f}-x(1-x){p}_n^2$.
The contribution to the correction mass, $M(p^2_n)$, is from the term of the metric tensor, $g^{\mu\nu}$, therefore it is the same for the longitudinal and the transverse parts.
The integration over 3-momentum
are taken first while the cutoff $\Lambda$ is kept finite until the whole integration is finished.
One the other hand, in the limit of $\beta_0\rightarrow 0$, according to Section \ref{phi3}, an extra scaling factor $(\frac{\beta_0}{\beta})^2$ has to be added. The dimensionless quantities, like $\Omega_{0T}
$ and $\Omega_{0L}$, are free from the extra factor, only $M_0$ has to be corrected. It is noticed that there is in fact no quadratic divergence for the photon mass in QED due to gauge- and Lorentz-invariance. However the Lorentz invariance is only secured for $\beta\gg 1$ in the imaginary-time formalism, there would be no wonder that the quadratic divergence will appear in the following calculation.
The arguments of the self-energy functions can be dropped, since they do not depend on them, and we obtain
\begin{eqnarray}
\Omega_{0T}
=
-\left.\frac{2\alpha}{3\pi}(\log N_{\rm max}+C_E)+\frac{\alpha}{3\pi^2}\left(G+\frac{\pi}{4}\right),
\right. \hspace{1.3cm}&
\hspace{-1cm}\Omega_{0L}=\frac{\alpha}{3\pi^2}\left(G+\frac{\pi}{4}\right),\nonumber\\
M_0
=\,\,\frac{2\alpha}{\pi}\left(\frac{\beta_0}{\beta}\right)^2
\left\{ \frac{4\pi N_{\rm max}^2}{\beta_0^2}+\left(G+\frac{\pi}{4}\right)\frac{m^2_{\rm f }}{\pi}-\frac{\pi^2}{12\beta_0^2}\right\},
\label{betaInftyTps}
\end{eqnarray}
where $C_E=-0.036489...$ for $\sum_{m=1}^{N_{\rm max}}\frac{1}{m-\frac{1}{2}}=\log N_{\rm max}+C_E$. Besides, we have used $\sum_{m=1}^{N_{\rm max}}\left({m-\frac{1}{2}}\right)= \frac{N^2_{\rm max}}{2}$, $\sum_{m=1}^{N_{\rm max}}$\hspace{.5cm}$\left({m-\frac{1}{2}}\right)^2=\frac{N_{\rm max}}{3}\left( \mathsmaller{N^2_{\rm max}}-\frac{1}{4}\right)$, etc.\,.
The factor $G\,\,(=0.915966...)$ is Catalan's constant. It is obtained from the sum, $\sum_{n=0}^\infty \frac{(-1)^n n}{(2n-1)^2}=\frac{G}{2}+\frac{\pi}{8}$, as the integral is expanded in powers of $1/\Lambda$. \par
As for the other limiting case, for $\beta \gg 1$, the sum of the fermionic frequency becomes an integral over the variable $\omega$, that is $\frac{2\pi}{\beta}\sum\rightarrow\int d\omega$.
When integration over $d\omega$, as shown in Appendix \ref{appB}: eq. (\ref{omegaT2}) and (\ref{omegaL2}), the following treatment for the inverse tangent function is adopted by making
\begin{eqnarray*}
\tan^{-1}\left(\frac{\Lambda}{\sqrt{\omega^2+\Delta}}\right)=\frac{\pi}{2}-
\tan^{-1}\left(\frac{\sqrt{\omega^2+\Delta}}{\Lambda}\right).
\end{eqnarray*}
This will make a good expansion of $1/\Lambda$ when the cutoff is taken to infinity.
The corrections to the photon mass is independent from the incoming momentum. We will drop its arguments hereafter. We may set $\Lambda_t=\Lambda=\frac{2\pi N_{\rm max}}{\beta}$ and obtain
\begin{eqnarray}
\Omega_T(p^2_n)&=&-\frac{2\alpha}{\pi}\left(\frac{G}{6\pi}+\frac{1}{3}\log N_{\rm max}-\frac{1}{3}\log \beta+\frac{1}{3}\log (4\pi)-\int^1_0dxx(1-x)\log\Delta(p^2_n)\right),\nonumber \\
\Omega_L(p^2_n)&=&\frac{G\alpha}{3\pi^2},\hspace{.5cm}{\rm and}\hspace{.5cm}M=\frac{2\alpha}{\pi}\left( {\frac{4\pi}{\beta^2}N_{\rm max}^2}+\frac{Gm^2_{\rm f}}{\pi} \right).\label{beta0ps}
\end{eqnarray}
In Section \ref{perturbation}, the radiative corrections, similar to the renormalized conditions, are defined as the differences between $\beta$ and $\beta_0\,\, (=0^+)$ in eq. (\ref{newIT}); they can be written for the respective quantities such as follows
\begin{eqnarray*}
&&\hat{\Omega}_T(p^2_n)\equiv{\Omega}_T(p^2_n)-{\Omega}_{0T},\hspace{.3cm}
\hat{M}\equiv{M}-{M}_{0},\,\
{\rm and}\hspace{.5cm} \hat{\Omega}_L(p^2_n)\equiv{\Omega}_L(p^2_n)-{\Omega}_{0L}.
\end{eqnarray*}
The renormalized results for transverse and longitudinal parts are
\begin{eqnarray}
\hat{\Omega}_T(p^2_n)&=&-\frac{2\alpha}{\pi}\left(-\frac{1}{3}C_E+\frac{\alpha}{24}+\frac{1}{3}\log 4\pi -\frac{1}{3}\log \beta-\int^1_0dxx(1-x)\log\Delta(p_n^2)\right),\nonumber\\
\hat{\Omega}_L(p^2_n)&=&-\frac{\alpha}{12\pi}, \hspace{.2cm}{\rm and}\hspace{.2cm}
\hat{M}
=\,\,\frac{2\alpha}{\pi}
\left\
\frac{Gm^2_{\rm f }}{\pi}+\frac{\pi^2}{12\beta^2}\right\}.\hspace{.3cm}
\label{pitpil}
\end{eqnarray}
For the corrected photon mass squared, $\hat{M}$, looks non-trivial, we will leave it temporarily and be back on this later. It can be compared with the self-energy of the gluon in a QCD plasma \cite{landsman87}; its Debye mass is
\begin{eqnarray}
m_D&=&g^2\left(\frac{N_f T^2}{6}+\frac{N_c T^2}{3}+\frac{N_f \mu_q^2}{2\pi^2}\right),
\end{eqnarray}
where $\mu_q$ is the quark chemical potential and $g$ is the QCD coupling constant. We know that in a sense the $\mu_q$ is equivalent to a mass term in the Lagrangian, therefore the correction to the Debye mass of a gluon in QCD plasma is similar to that of a photon in the imaginary-time formalism of vacuum. They all depend on the temperature and fermion masses or chemical potentials in a similar form.
We now combine the tree-level and the one-loop contribution for the photon propagator, as we assume it carries a mass $m_\gamma$:
\begin{eqnarray}
&&\frac{-g^{\mu\nu}}{(i\omega_n)^2-|{\bf q}|^2-m^2_{\gamma}}+\frac{-g^{\mu\alpha}}{(i\omega_n)^2-|{\bf q}|^2-m^2_{\gamma}}
(\hat{\Omega}_{\alpha\beta})\frac{-g^{\beta\nu}}{(i\omega_n)^2-|{\bf q}|^2-m^2_{\gamma}}+\dots\nonumber
\\
\hspace{-3.8cm}&=&
-\frac{g^{\mu\nu}(1-\hat{\Omega}_T)}{(i\omega_n)^2-{\bf q'}^2 }+\frac{p^\mu_ n p^\nu_n(\hat{\Omega}_L-\hat{\Omega}_T)}{(p^2_n-m^2_{\gamma})^2}+\dots,
\label{propcorr}
\end{eqnarray}
where ${\bf q'}^2={\bf q}^2+(m^2_{\gamma}+\hat{M})(1-\hat{\Omega}_L)$. The role of a photon's intrinsic mass is played by the chemical potential $\mu_{\gamma}$ in the Lagrangian, eq. (\ref{Lphoton}), so the effective mass of the photon to the leading order can be defined from above as
\begin{eqnarray}
m^2_{\rm eff,\gamma}=\frac{\mu_{\gamma}}{2}+\hat{M}.\label{meffphoton}
\end{eqnarray}
Unlike the effective mass of the fermion which will be defined in eq. (\ref{effmassF}), the correction from the imaginary-time context will shift the pole of the photon propagator as seen in eq. (\ref{PhotonProp1}).
To see if the effective mass of the photon is nearly zero or not, the calculation of the chemical potential, which is ignored in the derivation of the photon propagator, is necessary and will be performed below. On the other hand, the renormalized amplitude of the photon is defined as
\begin{eqnarray}
Z_3\equiv 1-\hat{\Omega}_T.
\end{eqnarray}
Its dependence on $\beta$ results in one of the renormalization group equation, and will be discussed in detail in Section \ref{rgeq}.
\subsubsection{Running coupling and Lamb shift}
The radiative corrections to the self-energy of photons give rise to the slight change of the electric potential $V({\bf x})$ in QED. In this section, we will see if the same results are concluded for the imaginary-time formalism.
The correction to the amplitude can approximated by setting $\omega_n=0$, since $\hat{\Omega}_T$ depends weakly on it, especially as $\beta \gg 1$. To Ignore the constants in the expression, the quantity can be redefined in the ${\bf q}^2$ dependence of the effective charge
\begin{eqnarray*}
{\Omega}_2(0,{\bf q})&\equiv& \hat{\Omega}_T(0,{\bf q})-\hat{\Omega}_T(0,{\bf 0}
=-\frac{2\alpha}{\pi}\int^1_0dx (1-x)\log\left(\frac{m^2_{\rm f}}{m^2_{\rm f}+x(1-x){\bf p}^2}\right).
\end{eqnarray*}
The expression is the same as what is obtained in the field theory \cite{peskin95} in the non-relativistic limit.
In the non-relativistic limit, the potential $V({\bf x})$ is obtained from the formula
\begin{eqnarray*}
V({\bf x})&=&-e^2\int \frac{d^3 {\bf p}}{(2\pi)^3}\frac{e^{i{\bf p\cdot x}}}{{\bf p}^2(1-\Omega_2(0,{\bf p}))}.
\end{eqnarray*}
From above, we may see that it also gives the same results as in QED for the effective potential.
\subsubsection{Ward identity}
One of the important requirements in the field theory is to check the gauge invariance of the theory, as it is related to the conservation of the momentum. As to the self-energy of the photon, it states
\begin{eqnarray}
\Pi^{\mu\nu}(p) p_\nu=0,\label{wardI}
\end{eqnarray}
and generates zero mass for the photon.
Based on the discussions in Section \ref{photonreal}, the real part of the self-energy of the photon, $\Pi^{\mu\nu}$ is highly suppressed for $\beta\rightarrow \infty$ along the contour as shown in fig. \ref{Fig:threshold}.
This means eq. (\ref{wardI}) is automatically satisfied for the real part. As to the imaginary part from the branch cut, examine how the dimensional regularization does to ensure the Ward identity.
The reason is that it provides an equality between the $d$-dimensional integrals:
\begin{eqnarray}
\int \frac{d^d l_E}{(2\pi)^{{d}}} \frac{\left(-\frac{2}{d}+1\right)l^2_E}{(l^2_E+\Delta)^2}=
-\int \frac{d^d l_E}{(2\pi)^{d}} \frac{\Delta}{(l^2_E+\Delta)^2}.\label{wi}
\end{eqnarray}
Regardless of the real parts of the both integrals, for $d=4$, both of the imaginary parts are coming from the term, $\frac{\Delta}{(4\pi)^2}\log \Delta$, in the outcomes of above integrals, therefore eq. (\ref{wi}) is contented for the imaginary part. Thus the self-energy tensor of the photon from the real-time formalism satisfies eq. (\ref{wardI}). As for the imaginary-time propagators, we may know from the above one-loop calculations, it does contribute a small radiative correction to photon's mass. On the other hand, we know from the traditional QED, the Lorentz- and gauge-invariance result in zero correction to the mass for photons. In the imaginary-time formalism, the Lorentz invariance does not hold in the imaginary-time Lagrangian densities as shown in Section \ref{lagimag}. Thus there is no wonder there are an appearance of quadratic divergence in photo's self-energy and a nonzero mass correction. Fortunately, the quadratic divergence can be removed by its counterpart at the high energy limit, $\beta_0$, and the nonzero mass correction is canceled by its chemical potential, which will be explained right below. Therefore after taking into account all of the factors, the self-energy correction to photon's propagator can be expressed in the form of $\Omega^{\mu\nu}(q)=\Omega(q) (g^{\mu\nu}-\frac{q^\mu q^\nu}{q^2})$ for both real-time and imaginary-time propagators,
and the Ward identity can be secured.
\subsubsection{Chemical potential}\label{chempot}
The calculation presented here is generalized from Appendix \ref{appC}. In the 4-momentum space $(i\omega_n,{\bf p})$, the number operator of the vector field is proportional to the product of the photon field operators, such as
\begin{eqnarray*}
\hat{N}(i\omega_n,{\bf p})\equiv :A^\mu(i\omega_n,{\bf p}) A_\mu(i\omega_n,{\bf p}):&=&\frac{1}{2}\sum^{}_{\lambda=1,2}:\left(a^\lambda_{\omega_n,{\bf p}}a^{\lambda\dagger}_{\omega_n,{\bf p}}+
a^{\lambda\dagger}_{\omega_n,{\bf p}}a^\lambda_{\omega_n,{\bf p}}\right):,\,\,
\end{eqnarray*}
where $\lambda$ denotes the polarization states. The representation in space and imaginary-time: $\hat{N}(\tau,{\bf x})=A^\mu (\tau,{\bf x}) A_\mu (\tau,{\bf x})$. They are related by
\begin{eqnarray*}
\frac{1}{\beta}\sum_n\int \frac{d^3{\bf p}}{(2\pi)^3}\,\hat{N}(\omega_n,{\bf p})=\int^\beta_0 d\tau\int d^3{\bf x}\,\hat{N}(\tau,{\bf x}).
\end{eqnarray*}
The interaction Hamiltonian in the imaginary-time and space, which is proposed in Section \ref{IntQED}, is $\hat{V}=e\int d^3{\bf x}\bar{\psi}(x)\gamma^\mu\psi(x)A_\mu=\int d^3{\bf x}H_{int}$. We may rewrite eq. (\ref{chemicalp}) in Appendix \ref{appC} in terms of the chemical potential density, $\mu_\gamma$, as
\begin{eqnarray}
\beta \,\mu\rightarrow\int^\beta_0d\tau\int d^3{\bf x}\,\mu_\gamma(\tau,{\bf x})&=&-\sum_{n=1}\frac{1}{n!}\langle \frac{\delta (\int^\beta_0d\tau\int d^3{\bf x}H_{int})^n}{\delta (A^\mu A_\mu)}\rangle.
\label{chemicalp1}
\end{eqnarray}
The first non-vanishing term is for $n=2$, the variation, $\delta$, is placed inside the integral on $H_{int}^2$. The leading order term of the right-hand-side becomes
\begin{eqnarray}
\int^\beta_0d\tau\int d^3{\bf x}\,\mu_\gamma(\tau,{\bf x})&=&-\frac{e^2}{2!}\langle \left(\int^\beta_0d\tau\int d^3{\bf x}\bar{\psi}\gamma^\mu \psi\right)^2\rangle.
\label{chemicalp1}
\end{eqnarray}
The Lorentz index, $\mu$, is contracted with the other one, so that the right-hand-side is kept as a scalar. It
then can be computed
\begin{eqnarray*}
\hspace*{-1.cm}
{\rm RHS}&=&-\frac{e^2}{2!}\int^\beta_0d\tau\int d^3{\bf x}\int^\beta_0d\tau'\int d^3{\bf x'}{\rm Tr} \,S_F(x-x')\gamma^\mu S_F(x'-x)\gamma_\mu,\\
&=&-\frac{e^2}{2!}\int^\beta_0d\tau\int d^3{\bf x}\frac{1}{\beta}\sum_{n}\int \frac{d^3{\bf q}}{(2\pi)^3}{\rm Tr} \,\frac{1}{\slashed{q}_n-m_{\rm f}}\gamma^\mu \frac{1}{\slashed{q}_n-m_{\rm f}}\gamma_\mu .
\end{eqnarray*}
The dependence of $x$ in the first line of the above
disappears after integrating out the variable $x'$.
As compared with eq. (\ref{chemicalp1}), we may obtain the chemical potential density:
\begin{eqnarray*}
\mu_\gamma&=&-\frac{e^2}{2\beta}\sum_{n}\int \frac{d^3{\bf q}}{(2\pi)^3}{\rm Tr} \,\frac{1}{\slashed{q}_n-m_{\rm f}}\gamma^\mu \frac{1}{\slashed{q}_n-m_{\rm f}}\gamma_\mu .
\end{eqnarray*}
The above can be derived directly from eq. (\ref{omega}) by setting the incoming momentum, $p_n=0$ and contracting the two Lorentz indices, so in a similar way to define the renormalization condition for the chemical potential density we may obtain
\begin{eqnarray*}
\hat{\mu}_\gamma&=&-2{\hat{M}} .
\end{eqnarray*}
The subtraction from the value at $\beta_0=0^+$ have been applied in the above. The corrected mass of photon from the self-energy is canceled by the chemical potential density, so that the effective mass, which is defined in eq. (\ref{meffphoton}), diminishes to the order of $O(\alpha)$.
\subsection{Self-energy of fermions}
\subsubsection{Real-time}
In fig. \ref{Fig:feynSEelectron}, it shows that two situations of a fermion emitting and absorbing a photon during its propagation. The self-energy of the fermion can be expressed in the form of
\begin{eqnarray*}
\Sigma(p)=\slashed{p}\Sigma_V(p^2)+m_{\rm f}\Sigma_S(p^2).
\end{eqnarray*}
The scalar quantity, $\Sigma_V(p^2)$, can be obtained by taking $\frac{1}{4}{\rm Tr}\slashed{p}$ on both sides,
\begin{eqnarray*}
-i\Sigma_(p)&=&-e^2\int_\otimes\frac{d^4k}{(2\pi)^4}\gamma^\mu \frac{(\slashed{p}+\slashed{k})+m_{\rm f}}{[(k+p)^2-m_{\rm f}^2]k^2}\gamma^\nu (1-n_F((k+p)\cdot u))(1+n_B(-k\cdot u)).
\end{eqnarray*}
After projecting onto $\slashed{p}$, their coefficient functions are
\begin{eqnarray*}
\Sigma_V(p^2)
&=&2ie^2\int d\alpha_1 d\alpha_2\delta(\alpha_1+\alpha_2-1)(1-\alpha_1)\int_\otimes \frac{d^4 K}{(2\pi)^4}\frac{1}{[K^2-\Delta(p^2)]^2}\\
&&\hspace{5cm}\times
\frac{1}{e^{-\beta(K+(1-\alpha_1)p)\cdot u}+1}\frac{-1}{e^{\beta(K-\alpha_1p )\cdot u}-1},
\end{eqnarray*}
where $\Delta(p^2)=-\alpha_1\alpha_2p^2+\alpha_1m^2_{\rm f}$
and $a=e^{\beta(1-\alpha_1)p\cdot u}$ and $b=e^{\beta\alpha_1p\cdot u}$. As for the scalar part of self-energy function,
\begin{eqnarray*}
\Sigma_S(p^2
&=&-i4e^2\int_\otimes \frac{d^4 k}{(2\pi)^4}\frac{1}{[(k+p)^2-m^2_{\rm f}]k^2}
(1-n_F((k+p)\cdot u))(1+n_B(-k\cdot u)).
\end{eqnarray*}
In general, the 4-momentum integral, $I _{\rm SE, \lambda}$, where $\lambda$ is the power of the denominator, with $n_F$ and $n_B$ can be written in the form of
\begin{eqnarray*}
&&\hspace{-1.5cm}I_{\rm SE, \lambda}=\int_\otimes \frac{d^4 l}{(l^2-\Delta)^\lambda}\left\{1-n_{\rm F}((k+p)\cdot u)\right\}\left\{1+n_{\rm B}(-k\cdot u)\right\}\\
&&\hspace{-0.6cm}
=\int_\otimes \frac{d^4 l}{(l^2_0-{\bf l}^2-\Delta)^\lambda}\frac{1}{e^{-\beta l\cdot u-\log b}+1}\frac{-1}{e^{\beta l\cdot u-\log a}-1},
\end{eqnarray*}
where $l=k+x p$, $a=e^{\beta xp\cdot u}$ and $b=e^{\beta(1- x)p\cdot u}$. It can explicitly be written for the left and right contours of the imaginary axis as follows
\begin{eqnarray*}
&&\hspace{-0.7cm} I_{\rm SE, \lambda}=\oint_{\rm RC,LC} \frac{d l_{\rm 0}d^3 {\bf l}}{(l^2_{\rm 0}-{\bf l}^2-\Delta)^\lambda}\frac{-1}{e^{\beta l_{\rm 0}-\log a}-1}\frac{1}{e^{-\beta l_{\rm 0}-\log b}+1}\pm({\rm residues\,\,of\,\,\,n_B\,\,and\,\,n_F }),\\
&&\hspace{-1cm}=
\begin{cases}
i\oint_{\rm RC} d l_{\rm E}\int d^3 {\bf l} \frac{(-1)^\lambda}{(l^2_{\rm E}+{\bf l}^2+\Delta)^\lambda}\frac{-1}{e^{i\beta l_{\rm E}-\log a}-1}\frac{1}{e^{-i\beta l_{\rm E}-\log b}+1}+2\pi i\frac{ab}{(1+ab)\beta}\sum_{n\in {\rm I}}\int d^3{\bf l}\frac{(-1)^{\lambda+1}}{\left[\Delta+{\rm l}^2+(\frac{2n\pi}{\beta}-i\frac{\log a}{\beta})^2\right]^\lambda}\\
i\oint_{\rm LC} d l_{\rm E}\int d^3 {\bf l} \frac{(-1)^\lambda}{(l^2_{\rm E}+{\bf l}^2+\Delta)^\lambda}\frac{-1}{e^{i\beta l_{\rm E}-\log a}-1}\frac{1}{e^{-i\beta l_{\rm E}-\log b}+1}-2\pi i\frac{(-1)ab}{(ab+1)\beta}\sum_{n\in {\rm I}}\int d^3{\bf l}\frac{(-1)^{\lambda+1}}{\left[\Delta+{\rm l}^2+(\frac{(2n+1)\pi}{\beta}-i\frac{\log b}{\beta})^2\right]^\lambda},
\end{cases}
\end{eqnarray*}
where $il_{\rm E}=l_0$. Take the right contour for example, as $\beta\gg1$, the factor $\frac{ab}{ab+1}$ becomes unity; the integral and the summation can be expressed as
\begin{figure}[t]
\begin{center}
\subfigure[]{\includegraphics[width=5cm]{seelectron1.eps}}
\hspace*{0.07\textwidth}
\subfigure[]{\includegraphics[width=5cm]{seelectron2.eps}}
\hspace*{0.0\textwidth}
\caption{\small One-loop self-energy Feynman diagrams of an electron for different time ordering of the interaction vertices: (a) An electron annihilates first and creates another electron and a photon, later the two annihilate and an electron is created, the probability for this process is proportional to the product of $1-n_{\rm F}(k_0+q_0)$ and $n_{\rm B}(k_0)$. (b) An electron disappears due to the annihilation of a positron and a photon in the background, and its momentum has been carried away by another electron that is produced earlier. The probability is proportional to the product of $1+n_{\rm B}(-k_0)$ and $n_{\rm F}(-k_0-q_0)$. }
\label{Fig:feynSEelectron}
\end{center}
\end{figure}
\begin{eqnarray*}
&&\hspace{0cm}=
i\int^{\Lambda_t}_{-\Lambda_t} d l_{\rm E}\int d^3 {\bf l} \frac{(-1)^\lambda}{(l^2_{\rm E}+{\bf l}^2+\Delta)^\lambda}+i\int^{\frac{2\pi N_{\rm max}}{\beta}}_{-\frac{2\pi N_{\rm max}}{\beta}}d l_{\rm E}\int d^3{\bf l}\frac{(-1)^{\lambda+1}}{\left[\Delta+{\rm l}^2+(l_{\rm E}-i\frac{\log a}{\beta})^2\right]^\lambda}.
\end{eqnarray*}
As discussed before, the cutoff is chosen as $\Lambda_t=\frac{2\pi N_{\rm max}}{\beta}$, which is the same as the cutoff of 3-momentum, $\Lambda$.
As for $\lambda=2$, the corresponding integral of the segments (A) and (B), similar to the contour in fig. \ref{Fig:threshold}, is
\begin{eqnarray*}
&&\hspace{-1cm}
=2\pi i\left[\left(-\frac{1}{\Lambda\Lambda_t}+\frac{2\Lambda\Lambda_t}{(\Lambda^2+\Lambda^2_t)^2}+\frac{\Lambda_t}{\Lambda(\Lambda^2+\Lambda^2_t)}-\frac{1}{\Lambda_t^2}\tan^{-1}\left(\frac{\Lambda}{\Lambda_t}\right)\right)\frac{\log^2a}{\beta^2
+O\left(\frac{\log^4 a}{\beta^4}\right)\right]
\end{eqnarray*}
The result is suppressed by the cutoffs of the energy and the momentum, as well as (C) and (D), and approaches zero while taking their values to infinities; except the imaginary part if there is a branch cut.
\subsubsection{Imaginary-time}
The self-energy of an electron is calculated according to the Lagrangian of fermions in eq. (\ref{Lfermion}) and the interaction Lagrangian in eq. (\ref{Linteraction}).
As in photon's self-energy, the fermion is assigned with a 4-momentum, $p^\mu_n\,\left(=(i\omega_n,{\bf p})\right)$, where $\omega_n=\frac{n\pi}{\beta}$ and $n$ is an odd integer.
Its self-energy takes the form:
\begin{eqnarray*}
\Xi \,(p^2_n)=\slashed{p}_n\Xi_V(p^2_n)+m_{\rm f}\,\Xi_S(p^2_n).
\end{eqnarray*}
Similar to the procedures that are taken for the real-time case, as for the gauge, $\zeta=1$,
\begin{eqnarray*}
\Xi (p^2_n)=\frac{1}{\beta}\sum_m\int\frac{d^3{\bf k}}{(2\pi)^3}(e\gamma^\mu) \frac{1}{(\slashed{p}_n+\slashed{k}_m)-m_{\rm f}}(e\gamma^\nu)\frac{-1}{k^2_m}\left.g_{\mu\nu}\right..
\end{eqnarray*}
After contracting with the vector, $\slashed{p}_n$, and use the relation:
\begin{eqnarray*}
\Xi_V(p^2_n)=\frac{1}{4p^2_n}{\rm Tr}[\slashed{p}_
\Xi] &{\rm and}&
\Xi_S(p^2_n)=\frac{1}{4m_{\rm f}}{\rm Tr}
\Xi].
\end{eqnarray*}
The loop momentum is $k_m\,(=(i\omega_m,{\bf k}))$.
The function of the self-energy, $\Xi_V(p^2_n)$, can be obtained after taking the projection onto 4 momentum, $\slashed{p}_n$, and applying the Feynman parametrization:
\begin{eqnarray}
\Xi_V(p^2_n)
&&\hspace{-0.0cm}=\frac{\alpha}{2\pi}\int^1_0 d\alpha_1 (1-\alpha_1)\frac{2\pi}{\beta}\sum_{m}\frac{1}{[(\frac{2\pi m}{\beta}+i\alpha_1p_n^0)^2+\Delta(p^2_n)]^{1/2}},
\end{eqnarray}
where the shifted loop momentum, $K^\mu_n=k^\mu_n+\alpha_1 p^\mu_n$, and $\alpha_i\,(i=1,2,3)$ are the Feynman parameters.
As the factor $\beta$ gets large, the summation $\frac{2\pi}{\beta}\sum$ is approximated by the integral $\int d\omega$ from $-\Lambda_t$ to $\Lambda_t$. It becomes
\begin{eqnarray}\label{sigmaV}
&&\hspace{-1cm}\Xi_V(p^2_n)
\label{xiv}
=\frac{\alpha}{4\pi}\left(2\log \Lambda_t+2\log 2- \log \Delta(p^2_n) \right).
\end{eqnarray}
The self-energy at an infinite temperature can be expressed as
\begin{eqnarray}\label{BsigmaV}
\Xi_{0V}(p^2_n)&=&
\frac{\alpha}{2\pi}\int d\alpha_1 d\alpha_2\delta(\alpha_1+\alpha_2-1)(1-\alpha_1)\sum_{m}\frac{1}{|m|}\nonumbe
=\frac{\alpha}{4\pi}\left(2\log N_{\rm max}+2\gamma_{\rm E}\right),
\end{eqnarray}
where the integer, $m$, goes from $\pm 1$ to $\pm N_{\rm max}$ and $\gamma_{\rm E}\,(=0.577216...)$ is the Euler-Mascheroni constant. As for the other function of the self-energy,
\begin{eqnarray}
\Xi_{0S}(p_n^2
\label{xis}
&=&-\frac{\alpha}{\pi}\int^1_0 d\alpha_1\frac{2\pi}{\beta_0}\sum_{m}\frac{1}{[(\frac{2\pi m}{\beta_0}+i\alpha_1p^0_n)^2+\Delta(p^2_n)]^{1/2}}.
\end{eqnarray}
Similarly, the scalar part of the self-energy at infinite temperature can be expressed as
\begin{eqnarray*}
\Xi_{0S}(p^2_n)&=&-
\frac{\alpha}{\pi}\int^1_0 d\alpha_1 \sum_{m}\frac{1}{|m|}
=-\frac{\alpha}{\pi}\left(2\log N_{\rm max}+2\gamma_{\rm E}\right),
\end{eqnarray*}
when $N_{\rm max}$ is extremely large. From eq. (\ref{newIT}), we could have the renormalized self-energy functions for the imaginary-time as follows
\begin{eqnarray*}
\hat{\Xi}_V\equiv\Xi_V(\beta, p^2_n)-\Xi_{0V},\,\,{\rm and}\hspace{.4cm}
\hat{\Xi}_S\equiv\Xi_S\,(\beta, p^2_n)-\Xi_{0S}.
\end{eqnarray*}
from eq. (\ref{sigmaV}) and (\ref{BsigmaV}), let $\Lambda_t=\frac{2\pi N_{\rm max}}{\beta}$,
\begin{eqnarray}
\hat{\Xi}_V(\beta, p^2_n)&=&\frac{\alpha}{4\pi}\left(2\log \frac{2\pi}{\beta}+2\log 2-2\gamma- \log \Delta(p^2_n) \right),\nonumber\\
\hat{\Xi}_S(\beta, p^2_n)&=&-\frac{\alpha}{\pi}\left(2\log \frac{2\pi}{\beta}+2\log 2-2\gamma- \log \Delta(p^2_n) \right).\label{xivxis}
\end{eqnarray}
The dependence of $\beta$ in the above results is related to the renormalization group equations, and will be further explained in Section \ref{rgeq}. We combine the tree-level and the one-loop contribution
\begin{eqnarray*}
&&\hspace{-0.8cm}\frac{1}{i\omega_n\gamma^0-\slashed{\bf p}-m_{\rm f}}+\frac{1}{i\omega_n\gamma^0-\slashed{\bf p}-m_{\rm f}}
(\hat{\Xi})\frac{1}{i\omega_n\gamma^0-\slashed{\bf p}-m_{\rm f}}+\dots
=
\frac{1}{i\omega_n\gamma^0-\slashed{\bf p}-m_{\rm f}-\hat{\Xi}}.
\end{eqnarray*}
We may approximate the self-energy as $\hat{\Xi} (\omega_n, |{\bf p}|) \approx \left({i\omega_n}\gamma_0-\slashed{\bf p}\right)\hat{\Xi}_V(0,|{\bf p}|)+m_{\rm f}\hat{\Xi}_S(0,|{\bf p}|) $, since the self-energy functions $\hat{\Xi}_V(\omega_n,|{\bf p}|)$ and $\hat{\Xi}_S(\omega_n,|{\bf p}|)$ in eq. (\ref{xivxis}) have only weak dependence on the Matsubara frequency. The propagator for the imaginary-time may be derived to
\begin{eqnarray}
\approx\frac{1}{i\omega_n\gamma^0-\slashed{\bf p}-m_{\rm f}-(\left({i\omega_n}\gamma^0-\slashed{\bf p}\right)\hat{\Xi}_V+m_{\rm f}\hat{\Xi}_S)}
=\frac{1+\hat{\Xi}_V(0,|{\bf p}|)}{(i\omega_n\gamma^0-\slashed{\bf p})-m_{\rm f}(1+\hat{\Xi}_V+\hat{\Xi}_S)}
\label{effmassF}
\end{eqnarray}
where the effective mass of the fermion is defined as $m_{\rm f,\rm eff}=m_{\rm f}\left(1+(\hat{\Xi}_V+\hat{\Xi}_S)(0,|{\bf p}|)\right)$. One thing needs to emphasize is that this effective mass does not come into the play of the traditional propagator $1/(\slashed{p}-m_{\rm f})$ since in the summation of $1/(i\omega_n-\xi^{\rm eff}_{\bf p})$ over the Matsubara frequencies of eq. (\ref{diracprop}) the effective mass, $m_{\rm eff}$, would just add corrections to the density function. The pole of the traditional propagator is not shifted from the correction of the imaginary-time. From the previous section discussing the corrections for the real-time, no real part is generated, therefore a fermion mass defined by the pole of its propagator will always be a fixed value. As we have seen from the approach of partition function in eq. (\ref{Lagrangian}) and eq. (\ref{Lagrangian2}), such a replacement could be made, $\frac{1}{i\omega_n\gamma^0-\slashed{\bf p}-m_{\rm f}}\rightarrow \frac{1}{(i\omega_n-\xi_{\bf p})\gamma^0}$. The one-loop corrected propagator for the imaginary-time can be replaced by
\begin{eqnarray*}
\hspace{0cm}&&\frac{1+\hat{\Xi}_V}{i\omega_n\gamma^0-\slashed{\bf p}-m_{\rm f,eff}}\rightarrow\frac{1+\hat{\Xi}_V}{(i\omega_n-\xi^{\rm eff}_{\bf p})\gamma^0}.
\end{eqnarray*}
The renormalized amplitude to a fermion field, $Z_2$, can be defined as
\begin{eqnarray}
Z_2\equiv1+\hat{\Xi}_V(0,|{\bf p}|).\label{Z2}
\end{eqnarray}
The amplitude will give rise to the renormalization group equation with respective to the variation of the variable, $\beta$, and will be discussed in Section \ref{rgeq}.
\subsection{Vertex correction}
\subsubsection{Real-time}
The computation of the vertex corrections for the real-time is similar to those in the previous sections.
The nonzero contribution is the imaginary part, while the real part is zero along the rectangular contour similar to that in fig. \ref{Fig:threshold} (b). Here the derivations of the loop integrals and the residues of the poles from the three density functions are shown. The corresponding Feynman diagram showing the incoming, outgoing and loop momenta are illustrated in fig. \ref{Fig:vertexC} (a). The correction to the vertex, $\delta \Gamma^\mu(p',p)$, is derived from
\begin{eqnarray*}
\hspace{-0.8cm}\delta \Gamma^\mu(p',p)
&=&\int_\otimes\frac{d^4k}{(2\pi)^4}\frac{-ig_{\nu\rho}}{(k-p)^2+i\varepsilon}\bar{u}(p')(-ie\gamma^\nu)
\frac{i(\slashed{k'}+m)}{k'^2-m^2_{\rm f}+i\varepsilon}\gamma^\nu \frac{i(\slashed{k}+m)}{k^2-m^2_{\rm f}+i\varepsilon}
(-ie\gamma^\nu)u(p) \\
&&\hspace{3.8cm}\times n_{\rm B}\left((k-p)\cdot u\right)\left\{1-n_{\rm F}((k+q)\cdot u)\right\}\left\{1-n_{\rm F}(k\cdot u)\right\},\\
&=&4ie^2\int_\otimes\frac{d^4l}{(2\pi)^4}\int^1_0dxdydz\delta(x+y+z-1)\\
&&\times\bar{u}(p')\left[\gamma^\mu\cdot\left(-\frac{1}{2D^2}-\frac{\Delta}{2D^3}+(1-x)(1-y)\frac{q^2}{D^3}+(1-4z+z^2)\frac{m^2_{\rm f}}{D^3}\right)\right.\nonumber\\
&&\left.\hspace{.5cm}+\frac{i\sigma^{\mu\nu }q_\nu}{2m_{\rm f}D^3}(2m^2_{\rm f}z(1-z))\right]u(p)\frac{-1}{e^{\beta (l\cdot u)-\log c}-1}\frac{1}{e^{-\beta l\cdot u-\log a}+1}\frac{1}{e^{-\beta l\cdot u+\log b}+1}.
\end{eqnarray*}
where $D=l^2-\Delta+i\varepsilon$, $\Delta=-xyq^2+(1-z)^2m^2_{\rm f}$ and $a=e^{\beta (1-\alpha_2)q\cdot u +\beta\alpha_3 (p\cdot u)}$, $b=e^{\beta \alpha_2 (q\cdot u )-\beta\alpha_3 (p\cdot u)}$ and $c=e^{\beta \alpha_2(q\cdot u) +\beta(1-\alpha_3 )p\cdot u}$. For a convenient purpose, we may write down their products: $ac=e^{\beta (p+q)\cdot u}$, $c/b=e^{\beta p\cdot u}$ and $ab=e^{\beta q\cdot u}$, which are constants without the dependence of the loop momentum $l$.
As $\beta \gg 1$, $ac$, $c/b$ and $ab\gg 1$.
A general form of the loop integrals for the vertex corrections, $I_{\rm VE,\lambda}$, of t\begin{figure}[t]
\begin{center}
\subfigure[]{\includegraphics[width=5cm]{vertexFD1a.eps}}
\hspace*{0.0\textwidth}
\subfigure[]{\includegraphics[width=5cm]{contourpole2.eps}}
\caption{\small (a) The Feynman diagram for the vertex correction is illustrated, and the momentum of each particle is shown for either the imaginary-time or the real-time. (b)
The contour for the vertex correction is represented by the blue lines with two segments (A) and (B). The branch cut, for $\Delta<0$, is denoted as (E). On the both side of the imaginary axis, the dots indicate the poles from three density functions. }
\label{Fig:vertexC}
\end{center}
\end{figure}he loop momentum, $k^\mu$, for the power, $\lambda$, of the denominator is written as
\begin{eqnarray*}
&&\hspace{-0cm}I_{\rm VE, \lambda}=\int_\otimes \frac{d^4 l}{(l^2-\Delta)^\lambda}n_{\rm B}\left((k-p)\cdot u\right)\left\{1-n_{\rm F}((k+q)\cdot u)\right\}\left\{1-n_{\rm F}(k\cdot u)\right\}\\
&&\hspace{-3cm}=
\begin{cases}
i\oint_{\rm RC} d l_{\rm E}\int d^3 {\bf l} \frac{(-1)^\lambda}{(l^2_{\rm E}+{\bf l}^2+\Delta)^\lambda}\frac{1}{e^{i\beta l_{\rm E}-\log c}-1}\frac{1}{e^{-i\beta l_{\rm E}-\log a}+1}\frac{1}{e^{-i\beta l_{\rm E}+\log b}+1}+2\pi i\frac{ac^2}{(b+c)(1+ac)\beta}\sum_{n\in {\rm I}}\int d^3{\bf l}\frac{(-1)^\lambda}{\left[\Delta+{\rm l}^2+(\frac{2n\pi}{\beta}-i\frac{\log c}{\beta})^2\right]^\lambda}\\
i\oint_{\rm LC} d l_{\rm E}\int d^3 {\bf l} \frac{(-1)^\lambda}{(l^2_{\rm E}+{\bf l}^2+\Delta)^\lambda}\frac{1}{e^{i\beta l_{\rm E}-\log c}-1}\frac{1}{e^{-i\beta l_{\rm E}-\log a}+1}\frac{1}{e^{-i\beta l_{\rm E}+\log b}+1}-2\pi i\frac{abc}{(b+c)(1-ab)\beta}\sum_{n\in {\rm I}}\int d^3{\bf l}\frac{(-1)^\lambda}{\left[\Delta+{\rm l}^2+(\frac{(2n+1)\pi}{\beta}-i\frac{\log b}{\beta})^2\right]^\lambda}\\
\hspace{8.65cm}-2\pi i\frac{ac}{(ac+1)(ab-1)\beta}\sum_{n\in {\rm I}}\int d^3{\bf l}\frac{1}{\left[\Delta+{\bf l}^2+(\frac{(2n+1)\pi}{\beta}-i\frac{\log a}{\beta})^2\right]^\lambda}.
\end{cases}
\end{eqnarray*}
Even though on the left-hand side there are twice number of the poles more than those on the right, the coefficient of the second set of residues, $\frac{ac}{(ac+1)(ab-1)}$, diminishes as $\beta\gg1$. In fact, the sum of the two coefficients for the residues on the left is equal to the that on the right,
\begin{eqnarray*}
\frac{ac^2}{(b+c)(1+ac)}=\frac{abc}{(b+c)(1-ab)}+\frac{ac}{(ac+1)(ab-1)},
\end{eqnarray*}
so the two sets of residues on one side are comparable to the one on the other. As for the results of the above calculation, the same logic applies as those for the self-energy integrals, the only nonzero contribution is the imaginary part due to the branch cut, (E) in fig. \ref{Fig:vertexC}, under the condition $\Delta<0$.
\subsubsection{Imaginary-time}\label{vertexIm}
The Feynman diagram for the vertex loop correction is shown in fig. \ref{Fig:vertexC} (a), including the related external and internal momenta. In the calculations of the vertex correction, there is one thing needed to be clarified. As the Gordon decomposition is applied, the on-shell propertyties of the external momenta $p$ and $p'$ have to be used through $(\slashed{p}-m_{\rm f})u(p)=(\slashed{p'}-m_{\rm f})u(p')=0$. However, in the imaginary-time formalism, the external energy-momentum $p_n$ and $p'_{n'}$ are in the forms of ${p}^\mu_{n}=(i\omega_{n},{\bf p})$ and ${p'}^\mu_{n'}=(i\omega_{n'},{\bf p}')$ when they are in the loop. Thus, as the operator ${p}_n$ acts on the physical states $u(p)$, the analytic continuation has to be applied $p_n\rightarrow p$, as well as $p'_{n'}\rightarrow p'$, outside the loop. That means that the energy $p_0$ has to rotate from the imaginary-axis to the real axis after the loop integral is completed. It is similar to the Wick rotation. When the loop momentum is rotated from the real axis to the imaginary one in the conventional way, the external momenta are in fact carried by the shifted loop momentum along the Wick rotation implicitly.
In addition, for a convenient reason in the calculations, the denominator, which is obtained after the Feynman parametrization, is denoted as $D=(p_m+yq_{n''}-xp_n)^2-{\bf l}^2-\Delta(q^2_{n''})$, and the factor $\Delta(q^2_{n''})=-xy q^2_{n''}+(1-z)^2m^2_{\rm f}+zm_\gamma^2$, where $m_\gamma$ is the mass of photon. The variables $x$, $y$ and $z$ are the Feynman parameters. The nonzero photon
mass $m_\gamma$ is given to avoid the infrared divergences at this stage.
The correction to the vertex with a fermionic frequency $\omega_m=\frac{2\pi}{\beta}(m+\frac{1}{2})$ is
\begin{eqnarray}
&&\hspace{-.8cm}\delta \Gamma^\mu(\{p'_{n'}{\bf }\},\{p_n{\bf }\})\nonumber\\
&&\hspace{-1.5cm}=\frac{1}{\beta}\sum_{m}\int\frac{d^3{\bf k}}{(2\pi)^3}\frac{g_{\nu\rho}}{(k_m-p_n)^2}\bar{u}(p'_{n'})(e\gamma^\nu)
\frac{(\slashed{k}_m+\slashed{q}_{n''}+m)}{(k_m+q_{n''})^2-m^2_{\rm f}}\gamma^\mu \frac{(\slashed{k}_m+m)}{k^2_m-m^2_{\rm f}}
(e\gamma^\rho)u(p_n),\nonumber
\\%\end{eqnarray*}
&&\hspace{-1.5cm}=\frac{4e^2}{\beta}\sum_{m}\int^1_0dxdydz\delta(x+y+z-1)\bar{u}(p'_n)\left[\gamma^\mu\cdot\left(-\frac{1}{16\pi\left[(\omega_m-iyq_{0''}+ixp_0)^2+\Delta(q^2_{n''}{\bf })\right]^{1/2}}\right.\right.\nonumber\\
&&\hspace{-1.0cm}\left.
-\left(\frac{\Delta(q_{n''}{\bf })}{2}-(1-x)(1-y)q^2_{n''}-(1-4z+z^2)m^2_{\rm f}\right)\frac{1}{32\pi \left[(\omega_m-iyq_{0''}+ix p_0)^2+\Delta(q^2_{n''}{\bf })\right]^{3/2}}\right)\nonumber\\
&&\hspace{-1.cm}\left.+\frac{i\sigma^{\mu\nu }q_{{n''}\nu}}{2m_{\rm f}}(2m^2_{\rm f}z(1-z)) \frac{1}{32\pi \left[(\omega_m-iyq_{0''}+ixp_0)^2+\Delta(q^2_{n''}{\bf })\right]^{3/2}}\right]u(p_n).\label{vertexqed}
\end{eqnarray}
Under the condition of $\beta\gg 1$, the summation over, $\frac{2\pi}{\beta}\sum_m$, is replaced by an integration, $\int d\omega$:\vspace{-.2cm}
\begin{eqnarray}
\delta \Gamma^\mu(\{p_{n'}\},\{p_n\})
&=&\frac{\alpha}{\pi}\int^1_0dxdydz\delta(x+y+z-1)\bar{u}(p'_n)\left[\gamma^\mu\cdot\left(-\log 2\Lambda+\frac{1}{2}\log \Delta(q^2_{n''})\right.\right.\nonumber\\
&&\left.-\left(\frac{\Delta(q^2_{n''})}{2}-(1-x)(1-y)q^2_{n''}-(1-4z+z^2)m^2_{\rm f}\right)\frac{1}{2\Delta(q^2_{n''})}\right)\nonumber\\
&&\hspace{0cm}\left.+\frac{i\sigma^{\mu\nu }q_{{n''}\nu}}{2m_{\rm f}}(2m^2_{\rm f}z(1-z)) \frac{1}{2\Delta(q^2_{n''})}\right]u(p_n).\label{vertex1}
\end{eqnarray}
Remember that we analytically continue $u(p_n)\rightarrow u(p)$ and $u(p'_{n'})\rightarrow u(p')$, while the Gordon decomposition is used in the above. We may retrieve the form factors from above by defining
\begin{eqnarray*}
F_1(q_{n''})&=&-\frac{\alpha}{2\pi}\left(\log N_{\rm max}+\log 4\pi-\log \beta+\frac{5}{4}\right)\\&&+\frac{\alpha}{\pi}\int^1_0dz\int^{1-z}_0dx\left(\frac{1}{2}\log \Delta(q^2_{n''})\right.\left.
\hspace{0cm}+\frac{zq^2_{n''}+2(1-3z+z^2)m^2_{\rm f}}{\Delta(q^2_{n''})}\right),\\
F_2(q_{n''})&=&\frac{\alpha}{\pi}\int^1_0 dz \int^{1-z}_0 dx \frac{m^2_{\rm f}z(1-z)}{\Delta(q^2_{n''})}.
\end{eqnarray*}
The corresponding form factors $F_{1}(p^2_{n''})$ and $F_{2}(p^2_{n''})$ in the limit of high temperature are obtained by ignoring the terms that are not related to $1/\beta_0$ since they are negligible as compared to those with $1/\beta_0$, when $\beta_0 \rightarrow 0$. In addition, the Matsubara frequencies from the imaginary part of external 4-momenta, $\omega_n$ and $\omega_{n''}$, can also be ignored as the sum of loop frequency $\omega_m$ is going up to a very large number $N_{\rm max}$; it will be explained more by eq. (\ref{shiftingeffect}). The vertex correction at $\beta_0$ is
\begin{eqnarray}
&&\hspace{-2cm}\delta \Gamma_0^\mu(\{p_{n'}\},\{p_n\})=\frac{\alpha(2\pi)}{\pi\beta_0}\sum_{m}\int^1_0dxdydz\delta (x+y+z-1)\bar{u}(p'_n)\left[\gamma^\mu\cdot\left(-\frac{1}{2\left|\omega_m+y\omega_{n''}-x\omega_n\right|}\right.\right.\nonumber\\
&&\hspace{1cm}\left.-\left(\frac{\Delta(p^2_{n''})}{2}-(1-x)(1-y)q^2_{n''}-(1-4z+z^2)m^2_{\rm f}\right)\frac{1}{4 \left|\omega_m+y\omega_{n''}-x\omega_n\right]^{3}}\right)\nonumber\\
&&\hspace{1cm}\left.+\frac{i\sigma^{\mu\nu }q_{{n''}\nu}}{2m_{\rm f}}(2m^2_{\rm f}z(1-z)) \frac{1}{4 \left|\omega_m+y\omega_{n''}-x\omega_n\right|^{3}}\right]u(p_n),\nonumber\\
&&\hspace{.7cm}=-\frac{\alpha}{2\pi}\bar{u}(p'_n)\left.\gamma^\mu\cdot\left(\log N_{\rm max}+C_E\right)\right.
u(p_n).\label{vertex2}
\end{eqnarray}
It is obtained from the sum, $\sum_{m=1}^{N_{\rm max}}\frac{1}{m+\frac{1}{2}}\left(=\log N_{\rm max}+C_E\right)$.
The terms with the denominator, $\left|\omega_m+y\omega_{n''}-x\omega_n\right|^3$, diminish due to the fact that, besides the constants, $\alpha$, $\pi$ and so on, it contains the factor, $\beta^2_0$, and the sum over the integer $m$, $\sum_{m=1}^{\infty}\frac{1}{(m+\frac{1}{2})^3}=0.414398...$, is a finite number. Thus, it becomes zero after taking the limit of $\beta_0\rightarrow 0$. This is similar to the case of $I_4$ explained in Section \ref{phi3}.
During the above derivation, the shift of the frequency, $\omega_m\rightarrow\omega_m+y\omega_{n''}-x\omega_{n'}$, is applied, then the summation range changes from the symmetric one, $\sum^{N_{\rm max}}_{-N_{\rm max}}$ to an asymmetric one, $ \sum^{N_{1}}_{-N_{2}}$, where $N_{1,2}=N_{\rm max}\pm\Delta N$, where the difference of the two cutoffs is $\Delta N=y\omega_{n''}-x\omega_{n'}$. The terms in eq. (\ref{vertex2}) involving the cutoffs is
\begin{eqnarray}
-\frac{\alpha}{4\pi}\left(\log N_{1}+\log N_{2}\right)=-\frac{\alpha}{2\pi}\left\{\log N_{\rm max}+\frac{1}{2}\log \left(1+\frac{\Delta N}{N_{\rm max}}\right)+\frac{1}{2}\log \left(1-\frac{\Delta N}{N_{\rm max}}\right)\right\}.\hspace{.5cm}\label{shiftingeffect}
\end{eqnarray}
As long as $N_{\rm max}\gg\Delta N$, the shifting effect can be ignored. Therefore the form factors in the limit of small $\beta_0$ are
\begin{eqnarray*}
F_{0,1}(q^2_{n''})=-\frac{\alpha}{2\pi}\left(\log N_{\rm max}+ C_E\right), \,\, {\rm and}\hspace{.5cm}F_{0,2}(q^2_{n''})=0.
\end{eqnarray*}
The renormalized form factors, which can be defined from eq. (\ref{newIT}), are given as below with a new reference points at $\beta_0=0^+$:
\begin{eqnarray*}
\hat{F}_{1}(q^2_{n''})&\equiv&F_{1}(q^2_{n''})-F_{0,1}(q^2_{n''}),\\
&=&-\frac{\alpha}{2\pi}(C_E+\log 4\pi-\log \beta+\frac{5}{4})+\frac{\alpha}{\pi}\int^1_0dz\int^{1-z}_0dx\left(\frac{1}{2}\log \Delta(q^2_{n''})\right.\\
&&\hspace{6cm}\left.
+\frac{zq^2_{n''}+2(1-3z+z^2)m^2_{\rm f}}{\Delta(q^2_{n''})}\right),\\
\hat{F}_{2}(q^2_{n''})&\equiv&F_{2}(q^2_{n''})-F_{0,2}(q^2_{n''})=\frac{\alpha}{\pi}\int^1_0 dz \int^{1-z}_0 dx \frac{z}{1-z}=\frac{\alpha}{2\pi}.
\end{eqnarray*}
The form factor $\hat{F}_{1}(q^2_{n''})$ also possesses the infrared divergence. From the result of $\hat{F}_{2}(q^2_{n''})$, the ratio $\frac{g-2}{2}=\frac{\alpha}{2\pi}$, which is consistent with what QED predicts for the one-loop level. It is known that there is an infrared divergence in the function $\hat{F}_{1}(q^2_{n''})$, and a tiny photon mass $m_\gamma$ is given to avoid the catastrophe temporarily. In a subsequent work \cite{huang14b} related to this issue, the infrared divergence can be regularized by giving a non-zero minimal Matsubara frequency $\big(\frac{2\pi}{\beta}\big)^2$ in the loop momentum. The added frequency is to emphasize the discreteness of the Matsubara frequency, since it is treated as a continuous variable in our calculations as $\beta\gg 1$. And in the case of QED presented here for the vertex correction, the loop frequency is not allowed to be zero for any fermionic propagator.
\section{Renormalization group equations}
\label{rgeq}
\subsection{Renormalized amplitude $Z_2$ and electron's effective mass $m_{\rm eff}$}
From one-loop corrections for the imaginary-time in the previous section, the corrections to the amplitudes of the fields are obtained.
The renormalized constant of a fermion is defined from eq. (\ref{Z2}) before being taken into the limit of $\beta\gg1$:
\begin{eqnarray*}
Z_2&=&1+\hat{\Xi}_V|_{\omega_0=0},\\
&=&1+\frac{\alpha}{2\pi}\int^1_0 d\alpha_1 (1-\alpha_1)\frac{2\pi}{\beta}\sum_{m}\frac{1}{[(\frac{2\pi m}{\beta})^2+\Delta(0,{\bf p})]^{1/2}}-\frac{\alpha}{4\pi}\sum_m\frac{1}{|m|}.
\end{eqnarray*}
We may be interested in the change of the renormalized amplitude with respect to the temperature. Take derivative of the above formula:
\begin{eqnarray*}
\frac{\partial \log Z_2}{\partial \log \beta
&=&-\frac{\alpha}{2\pi}\int^1_0 d\alpha_1 (1-\alpha_1)\left.\frac{2\pi}{\beta}\right.\sum_{m}\frac{\Delta(0,{\bf p})}{[(\frac{2\pi m}{\beta})^2+\Delta(0,{\bf p})]^{3/2}}.
\end{eqnarray*}
While $\beta\gg1$, the summation becomes an integration $\frac{2\pi}{\beta}\sum_m\rightarrow \int d\omega$:
\begin{eqnarray*}
\left.\frac{\partial \log Z_2}{\partial \log \beta}\right|_{\beta\gg 1}
&=&-\frac{\alpha}{2\pi}\int^1_0d\alpha_1(1-\alpha_1)\int^\infty_{-\infty} d\omega\frac{\Delta(0,{\bf p})}{[\omega^2+\Delta(0,{\bf p})]^{3/2}}=-\frac{\alpha}{2\pi}.
\end{eqnarray*}
We may change the variable $\beta$ to the temperature, $T\,(=\frac{1}{\beta})$, so that
\begin{eqnarray*}
\left.\frac{\partial \log Z_2}{\partial \log T}\right|_{T\simeq 0}
&=&\frac{\alpha}{2\pi}.
\end{eqnarray*}
This is the same as the renormalization group equation that is obtained from $\overline{MS}$ scheme \cite{grozin} for the renormalized amplitude of a fermion with the renormalization scale, usually denoted as $\mu$, is replaced by the temperature. As we shall see from below, all of the renormalization group equations to the next-to-leading order in the field theory are the same as those derived from the imaginary-time formalism. in the other limit, $\beta\rightarrow 0$, the term $\Delta(0,{\bf p})$ in the denominator can be ignored before the summation is done, so that
\begin{eqnarray*}
\left.\frac{\partial \log Z_2}{\partial \log \beta}\right|_{\beta\rightarrow 0}
&=&\lim_{\beta\rightarrow 0}-\frac{\alpha \zeta(3)}{2\pi}\frac{\beta^2\Delta}{(2\pi)^2}=0,
\end{eqnarray*}
and change the variable $\beta$ to temperature $T$
\begin{eqnarray*}
\hspace{0.0cm} \left.\frac{\partial \log Z_2}{\partial \log T}\right|_{T\rightarrow \infty}
=0.
\end{eqnarray*}
In fact the above result for the limit of infinite value of $\beta$ can be obtained directly in eq. (\ref{xivxis}).
Now consider the effective mass of a fermion, which is defined as $m_{\rm f, eff}=m\left(1+(\hat{\Xi}_V+\hat{\Xi}_S)(0,|{\bf p}|)\right)$, from eq. (\ref{xivxis}) and (\ref{effmassF}), take the derivative of the effective mass
\begin{eqnarray*}
\frac{\partial \log m_{\rm eff}}{\partial \log \beta}&=&\frac{\alpha}{2\pi}\int^1_0 d\alpha_1 (1+\alpha_1)\left.\frac{2\pi}{\beta}\right.\sum_{m}\frac{\Delta(0,{\bf p})}{[(\frac{2\pi m}{\beta})^2+\Delta(0,{\bf p})]^{3/2}},
\end{eqnarray*}
so as $\beta \gg 1 $, transform the summation into an integral, we have
\begin{eqnarray*}
\left.\frac{\partial \log m_{\rm eff}}{\partial \log \beta}\right|_{\beta\gg1
&=&\frac{\alpha}{2\pi}\int^1_0 d\alpha_1 (1+\alpha_1)\int^\infty_{-\infty} d\omega\frac{\Delta(0,{\bf p})}{[\omega^2+\Delta(0,{\bf p})]^{3/2}}=6\frac{\alpha}{4\pi}.
\end{eqnarray*}
In the limit of $\beta\rightarrow 0$, the value is zero in a similar way to that of $Z_2$. We may change it into a function of temperature $T$,
\begin{eqnarray*}
\left.\frac{\partial \log m_{\rm eff}}{\partial \log T}\right|_{T\simeq 0}=-6\frac{\alpha}{4\pi},&{\rm and}\hspace{.5cm}&
\left.\frac{\partial \log m_{\rm eff}}{\partial \log T}\right|_{T\rightarrow \infty}=0.
\end{eqnarray*}
The result for the zero temperature is consistent as the renormalization equation, $\frac{\partial \log m_{\rm eff}}{\partial \log \mu}=-6\frac{\alpha}{4\pi}$, in field theory. As discussed in the previous section, this correction to the fermion's mass only goes to the corrected fermion density function and does not shift the pole of the propagator.
\subsection{Renormalized amplitude $Z_3$ and charge renormalization constant $Z_\alpha$ }
In the limit of infinite $\beta$, we may use the results from the previous sections for the discussions here.
For the photon's renormalized amplitude, it is defined from the one loop correction to the propagator in eq. (\ref{propcorr}),
\begin{eqnarray*}
Z_3&\equiv& 1-\hat{\Omega}_T(\beta,0,{\bf p}).
\end{eqnarray*}
The corresponding renormalization group equations in the limit of infinite value of $\beta$, which can be easily obtained from eq. (\ref{pitpil}), as well as in the other limit of $\beta\rightarrow 0$, are
\begin{eqnarray*}
\hspace{-0cm}\left.\frac{\partial \log Z_3}{\partial \log \beta}\right|_{\beta\gg 1}=-\frac{8}{3}\frac{\alpha}{4\pi}, &{\rm and}\hspace{.5cm}& \hspace{-0cm}\left.\frac{\partial \log Z_3}{\partial \log \beta}\right|_{\beta\rightarrow 0}=0.
\end{eqnarray*}
Like the traditional renormalization constants defined in the field theory, we may also define the factor $Z_1=1+\hat{F}_{1}$ to describe the vertex correction. In a similar way, the correction to electron's charge
can be defined as $e_0=Z^{\mathsmaller{\frac{1}{2}}}_\alpha e$, and it is related to the other renormalized constants by $Z_\alpha=Z_1^2Z_2^{-2}Z_3^{-1}$. It is obvious that, in the framework of the imaginary-time, the Ward identity still holds in the same fashion, so that $Z_1=Z_2$ and $Z_\alpha=Z^{-1}_3$. We may obtain
\begin{eqnarray*}
\hspace{-0cm}\left.\frac{\partial \log Z_\alpha}{\partial \log T}\right|_{T\simeq 0}
=-\frac{8}{3}\frac{\alpha}{4\pi},&{\rm and}\hspace{.5cm}&\hspace{-0cm}\left.\frac{\partial \log Z_\alpha}{\partial \log T}\right|_{T\rightarrow \infty}
=0.
\end{eqnarray*}
This is the same as the renormalization group equation, $\frac{\partial \log Z_\alpha}{\partial \log \mu}=-\frac{8}{3}\frac{\alpha}{4\pi}$, from the $\overline{MS}$ scheme of renormalization. This may lead to a discussion of Landau pole \cite{gellmann,landau55}, which states that coupling constants become infinite in a finite momentum scale.
In the theory presented here, the renormalization group equations are identical with those predicted in the field theory, as the temperature of the vacuum is regarded as a parameter of scale, only in the low temperature limit.
The variation of couplings, with respect to a high $T$, vanishes, as shown in above. Therefore the electromagnetic coupling, $\alpha$, does not diverge over all range of temperature within this picture.
\section{Conclusion}\vspace*{-0cm}
In this paper, the vacuum is treated as a thermodynamical system, which is described by microscopic observables, and is applied on the calculations of radiative corrections as those in the field theory. The theory has been
constructed and compared with the conventional formalism of QED. It shows that, in a temperature close to zero, the mathematical formalisms coincide with those in the traditional theory, such as the propagators of the electron and the photon, and so on. The imaginary-time, $\tau$, is introduced with a finite range $(\beta_0,\beta)$, where $\beta_0(=0^+)$ is infinitesimal, in the construction of the partition functions; the field operators are expanded by the Matsubara frequencies and 3-momenta. The formalism can be separated into two parts for the real-time and the imaginary-time from the same partition function. The real-time evolution of the field is achieved by the analytic continuation of the variable $\tau$, by setting $\tau=it$, after Matsubara frequencies are summed. The one-loop radiative corrections have been calculated for the self-energy of the electron, the vacuum polarization of the photon, and the vertex correction. The results from the imaginary-time are always being real values, such as the self-energy; their UV divergences, instead of introducing counter terms, could be removed by subtracting the counterpart quantities, like corrections to the field amplitudes, masses, etc., at the reference point, $\beta_0= 0^+$. These values at a very high temperature lose the dependence of the external momenta, and are similar to the roles of the counter terms in the field theory. Those "counter terms" are automatically equipped in the theory from the lower integration bound of the imaginary-time, $\beta_0$, so the theory is intrinsically free of UV divergences.
The correction to the self-energy of a fermion contribute to the propagator $1/(\slashed{p}_n-m)$, which only affects the density function and the field amplitude, and does not shift the pole of the propagator for the real-time. As for the self-energy of the photon, the imaginary-time corrections make a shift to photon's mass, which is defined as the pole of the propagator of the real-time. The one-loop mass correction to the photon is non-negligible and is canceled by its chemical potential, so that agreeable comparisons to the experimental results can be secured. The loop corrections from the real-time is found to be always finite and imaginary, because they comprise the conventional loop integrals and the residues from the density functions, which would cancel each other.
The integrals along closed contours contribute nothing except the branch cuts, when the incoming momenta excesses the threshold. The resultant imaginary part is the same as those in the traditional Feynman integrals. The loop corrections from the real-time and the imaginary-time happen to account for the imaginary and the real parts of the calculations that are known in the field theory. As a result, it can be seen that this imaginary-time formalism in the low temperature limit is similar to the mathematical structure of the field theory. The consistency of the whole theory still needs more efforts to check, but if there is any consistency problem in the theory it should not be like those in other imaginary-time theory for plasma because of their mathematical similarities.
Another important conclusion of this paper is the renormalization group effects. The Lagrangian densities are invariant under the scale transformation. From the Tolman-Ehrenfest relation that was discovered in the general relativity, the temperature determines the speed of time. Similar discussions for all of four dimensions could also found in Wilson's approach in discussing the renormalization group in condensed matter physics and the Hawking radiation in the curved space-time field theory. In the calculations provided here, the corrections to the renormalized amplitudes of the fields and fermion mass are given from the imaginary-time corrections; the renormalization group equations can be derived from them with respective to the variation of the temperature. The amazing point of these computation is that it shows the the same renormalization coefficients at a temperature close to zero as those obtained in the field theory. Besides, this theory for electrodynamic forces does not generate infinite values of couplings, the so-called Landau pole, as the variation of the coupling is found to vanish at a very high temperature. In addition to these derived results, the field theory in this imaginary-time formalism incorporates various concepts in the traditional field theory with those in the thermodynamics and the general relativity. The Einstein equation has been known before that it can be perceived in a thermodynamical perspective, such as the Tolman-Ehrenfest relation. In the studies of the black hole, it is shown that the area of the event horizon is proportional to the entropy of the black hole. Besides, the field theory in the curved space-time is usually studied in different vacuum states in which the space-time is under conformal transformations. They are combined in the construction of the imaginary-time field theory and make deep connections with the knowledge obtained in the particle physics. To test its applicability on other physics of the vacuum, various effects, such as the Casimir effect, have been discussed in the following paper and many agreements can be proved. On the other hand, in the cosmology one important role that could be played by the vacuum is the dark energy. In another work, the cosmological constant can be computed without the interference of the divergence in the imaginary-time field theory through the DeWitt-Schwinger representation and the traditional calculation of the Casimir force. Both results give the same ratio for the equation of state, $w=-1$.
|
\section{Introduction}
Along this work, $\kappa$ and $\lambda$ denote infinite cardinals. By $C_p(X)$ we mean the space of the continuous real functions on $X$ with the topology of the pointwise convergence, while $C_k(X)$ denotes the space of the continuous real functions on $X$ with the compact-open topology.
Recall that for a topological space $X$ and a point $x\in X$, the {\bf tightness of $X$ at $x$}, denoted by $t(x,X)$, is the least cardinal $\kappa$ with the property that if $x\in \overline{A}$ for any $A\subset X$, then there exists a $B\in[A]^{\leq\kappa}$ such that $x\in\overline{B}$. The {\bf tightness of the space $X$} is the supremum of all cardinals $t(x,X)$ for $x\in X$, that we denote by $t(X)$.
Following Arhangel'skii in \cite{Arhan1}, we say that a topological space $X$ is {\bf productively $\kappa$-tight at $x\in X$} if, for any space $Y$ with $t(Y)\leq \kappa$, one has $t((x,y),X\times Y)\leq \kappa$ for any $y\in Y$, and we denote this by $\kappa\in Sp(x,X)$. Naturally, the space $X$ is {\bf productively $\kappa$-tight} if $\kappa\in Sp(x,X)$ for all $x$ in $X$, and we write $\kappa\in Sp(X)$ to denote that.
In \cite{Arhan1}, Arhangel'skii gave an internal characterization to the productively $\kappa$-tightness property of a Tychonoff space $X$ by using the concept of $\kappa$-singular families. A family $\mP$ of collections of subsets of $X$ is {\bf $\kappa$-singular} at $x\in X$ if the following holds:
\begin{enumerate}[(a)]
\item for each $\xi\in \mP$ there exists $A\in\xi$ such that $|A|\leq \kappa$;
\item each $\xi\in\mP$ is centered (has the finite intersection property);
\item for any $O\subset X$ open with $x\in O$ there is some $\xi \in \mP$ and $A\in \xi$ such that $A\subset O$;
\item for any $\mE\subset \mP$ with $|\mE|\leq \kappa$ it is possible to choose $A(\xi)\in \xi$ for each $\xi\in \mE$ such that $x\not\in \overline{\bigcup_{\xi\in\mE}A(\xi)}$.
\end{enumerate}
Note that in the presence of conditions $(b),(c)$ and $(d)$, $(a)$ is equivalent to ask condition $(a)'$: every member of any $\xi\in\mP$ is a subset of $X$ with cardinality at most $\kappa$. Then, we say that $X$ is $\kappa$-singular at $x\in X$ if there exists a family $\mP$ $\kappa$-singular at $x$. With this terminology we may state Arhangel'skii characterization for $Sp(x,X)$:
\begin{theorem}[Arhangel'skii, \cite{Arhan1}, Theorem 3.5]\label{ArhanSp}
For a Tychonoff space $X$ and $x\in X$, $\kappa\in Sp(x,X)$ if, and only if, $X$ is not $\kappa$-singular at $x$.
\end{theorem}
Actually, Arhangel'skii has obtained a third equivalence in the above result, related with the tightness of $x$ in any Hausdorff compactification of $X$. Using this other characterization, Uspenskii showed the following:
\begin{theorem}[Uspenskii, \cite{Uspenskii}, Theorem 1]\label{Uspenskii} For a Tychonoff space $X$, $\kappa\in Sp(C_p(X))$ if, and only if, every open covering $\mU$ for $X_\kappa$ has a subcovering $\mU'$ with cardinality less than or equal to $\kappa$, where $X_{\kappa}$ is the topological space obtained by declaring open the $G_{\kappa}$-sets of $X$.
\end{theorem}
The main result of this work is a generalization of the last result, by using the concepts of bornologies and avoiding the compactification argument of Uspenskii.
\section{Remarks about bornologies}
A {\bf bornology} $\mB$ on a topological space $X$ is an ideal of subsets of $X$ that covers the space. By a (compact) base $\mB'$ for a bornology $\mB$ on $X$, we mean a subset of $\mB$ that is cofinal with respect to inclusion (such that all its elements are compact).
For a topological space $X$ and a bornology $\mB$ on $X$ with compact base, we call the {\bf topology of uniform convergence on} $\mB$, denoted by $\mathcal{T}_{\mB}$, as the topology on $C(X)$ having as a neighborhood base at each $f\in C(X)$ the sets of the form $$\langle B,\varepsilon\rangle[f]:=\{g\in C(X):\forall x\in B(|f(x)-g(x)|<\varepsilon)\},$$
for $B\in\mB$ and $\varepsilon>0$. By $C_{\mB}(X)$ we mean the space $(C(X),\mathcal{T}_{\mB})$.
\begin{remark}
In fact, since $\mT_{\mB}$ is obtained from a separating uniformity over $C(X)$, it follows that $C_{\mB}(X)$ is a Tychonoff space (see McCoy and Ntantu \cite{McCoy}).
\end{remark}
We say that $\mC$ is a $\mB$-covering for $X$ if for every $B\in\mB$ there is a $C\in\mC$ such that $B\subset C$. Following the notation of Caserta \emph{et al.} (\cite{Caserta2}), we denote by $\mathcal{O}_{\mB}$ the collection of all open $\mB$-coverings for $X$. When $\mU\in\mO_{\mB}$ is such that $X\not\in\mU$, we say that $\mathcal{U}$ is nontrivial; note that if $\mU\in\mO_{\mB}$ is nontrivial, then $\mathcal{U}\setminus F$ is an open $\mB$-covering for $X$ for any $F\in[\mathcal{U}]^{<\omega}$.
Also, we denote by $l_{\mB}(X)$ the {\bf $\mB$-Lindel\"of degree of $X$}, that is the smallest transfinite cardinal $\kappa$ such that for every open $\mB$-covering for $X$ there exists a $\mB$-subcovering $\mU'\subset \mU$ with $|\mU'|\leq \kappa$.
\begin{example}
The main examples of bornologies with compact base on a topological space $X$ are the bornologies $\mathcal{F}=[X]^{<\omega}$ and $\frak{K}=\{A\subset X:\exists K\subset X$ compact and $A\subset K\}$ $-$ if $X$ is a Hausdorff space, then $\frak{K}=\{A\subset X:\overline{A}$ is compact$\}$. For $\mB=\mathcal{F}$, one has $C_{\mF}(X)=C_p(X)$ and the $\mF$-coverings are usually called by $\omega$-coverings of $X$; we denote by $\Omega$ the collection of all open $\omega$-coverings. Also, if $X$ is Hausdorff, then for $\mB=\frak{K}$ it follows that $C_{\frak{K}}(X)=C_{k}(X)$ and the $\frak{K}$-coverings for $X$ are the so called $K$-coverings for $X$; the family of all $K$-coverings will be denoted by $\mK$.
\end{example}
\section{Some results with bornologies}
Let $\alpha\geq \omega$ be an ordinal. Recall that the game $\gone^{\alpha}(\mA,\mC)$ denotes the two players game played as follows: for every inning $\gamma<\alpha$, player I chooses an element $A_\gamma\in\mA$, and then player II picks an $a_\gamma\in A_\gamma$; player II wins if $\{a_\gamma:\gamma<\alpha\}\in \mC$; we denote by $\gone(\mA,\mC)$ when $\alpha=\omega$. In the following, we should use the families $\mO_{\mB}$ as well as the family $\Omega_x:=\{A\subset X\setminus\{x\}:x\in \overline{A}\}$. We denote by $\0$ the zero function and the open interval $\left(-\frac{1}{n+1},\frac{1}{n+1}\right)$ by $I_n$, for all $n\in\omega$.
The next Lemma enable us to translate some closure properties of $C_{\mB}(X)$ as ``$\mB$-covering'' properties of $X$, and \emph{vice-versa}.
\begin{lemma}
Let $X$ be a Tychonoff space and let $\mB$ be a bornology with compact base on $X$.
\begin{enumerate}[(a)]
\item If $\mU\in\mO_{\mB}$ is nontrivial, then $A=\{f\in C_{\mB}(X):\exists U\in\mU(f\upharpoonright X\setminus U\equiv 1)\}\in\Omega_{\0}$.
\item Let $A\subset C_{\mB}(X)$, $n\in\omega$ and let $\mU=\{f^{-1}[I_n]:f\in A\}$. If $\0\in\overline{A}$, then $\mU\in\mO_{\mB}$ $($and possibly $X\in\mU$$)$.
\end{enumerate}
\end{lemma}
\begin{proof}
Essentially the same proof of Lemma 2.2 of Caserta \emph{et al.} \cite{Caserta2}.
\end{proof}
The following Theorem is based on a result from Scheepers (\cite{Scheepers1}, Theorem 23):
\begin{theorem}\label{Scheepers1}
Let $X$ be a Tychonoff space and let $\mB$ be a bornology with compact base on $X$. Player II has a winning strategy
in $\gone(\mathcal{O}_{\mB},\mathcal{O}_{\mB})$ played on $X$ if, and only if, player
II has a winning strategy in
$\gone(\Omega_{\0},\Omega_{\0})$ played on $C_\mB(X)$.
\end{theorem}
\begin{proof}
Let $\Phi$ be a winning strategy for player II in $\gone(\mathcal{O}_{\mB},\mathcal{O}_{\mB})$ on $X$. For each $A\in\Omega_{\0}$ and $n\in\omega$, let $\mathcal{U}_n(A)=\{ f^{-1}[I_n]:f\in A\}$, which is an open $\mB$-covering for $X$ by the above Lemma. Before we define a winning strategy for player II in $\gone(\Omega_{\0},\Omega_{\0})$, note that without loss of generality we may assume that for every inning $n\in\omega$, player I does not play $A_n\in\Omega_{\0}$ such that $X\in \mU_n(A_n)$. In fact, if player I chooses any $A_n\in \Omega_{\0}$ such that $X\in\mU_n(A_n)$, then player II picks an $f_n\in A_n$ such that $f_n^{-1}[I_n]=X$; if $X\in\mU_n(A_n)$ for infinitely many $n$, then it is easy to see that $\{f_n:n\in\omega\}\in\Omega_{\0}$. So, we may assume that $X\not\in\mU_n(A_n)$ for all $n$.
We define a winning strategy $\eta$ for player II in the game $\gone(\Omega_{\0},\Omega_{\0})$ on $C_\mB(X)$ as follows: for every inning $n\in\omega$, let $\eta((A_0,\dots,A_n))=f_n\in A_n$, where $f_n^{-1}[I_n]=\Phi((\mU_0(A_0),\dots,\mU_n(A_n)))$ is an open set that belongs to the open $\mB$-covering $\mU_n(A_n)$. We will show that $\eta$ is a winning strategy.
Let $(A_0,f_0, A_1,f_1,\dots,A_n, f_n,\dots)$ be a whole play in the game $\gone(\Omega_{\0},\Omega_{\0})$, where for each $n$, $A_n$ is the move of player I and $f_n=\eta((A_0,\dots,A_n))$ is the answer of player II at the inning $n$. Since we have $X\not\in\mU_n(A_n)$ for all $n\in\omega$, it follows that $(\mU_n(A_n))_{n\in\omega}$ is a sequence of nontrivial open $\mB$-coverings for $X$, hence it is a valid sequence of moves for the player I in $\gone(\mO_{\mB},\mO_{\mB})$. Now, let $C_n=\Phi((\mU_m(A_m):m\leq n))$ for all $n$ and note that $f_n^{-1}[I_n]=C_n$. Since $\Phi$ is a winning strategy for player II, the play $(\mU_0(A_0),C_0,\dots,\mU_n(A_n),C_n,\dots)$ in the game $\gone(\mO_{\mB},\mO_{\mB})$ is won by player II. This yields that $\{C_n:n\in\omega\}$ is a nontrivial open $\mB$-covering for $X$, thus $\{C_n:n\geq j\}\in\mO_{\mB}$ for any $j\in\omega$, since we are excluding only finitely many $C_n$'s. Then, we have $\{f_n:n\in\omega\}\in\Omega_{\0}$, as desired.
Conversely, let $\psi$ be a winning strategy for player II in $\gone(\Omega_{\0},\Omega_{\0})$ on $C_\mB(X)$. For each nontrivial $\mU\in\mO_{\mB}$, let $A(\mU)=\{f\in C_{\mB}(X):\exists U\in\mU(f\upharpoonright X\setminus U\equiv 1)\}$, which is an element of $\mO_{\mB}$ by the above Lemma. Clearly we may suppose that the player I does not play trivial open $\mB$-coverings.
We define a winning strategy $\rho$ for player II in $\gone(\mathcal{O}_{\mB},\mathcal{O}_{\mB})$ on $X$ as follows: if $(\mU_0,\dots,\mU_n)$ is a sequence of nontrivial open $\mB$-coverings played by player I, let $\rho((\mU_0,\dots,\mU_n))=U_n\in\mU_n$, where $U_n$ is such that $\psi((A(\mU_0),\dots,A(\mU_n)))\upharpoonright X\setminus U_n\equiv 1$. We now show that $\rho$ is a winning strategy for player II in $\gone(\mO_{\mB},\mO_{\mB})$.
Let $(\mU_0,U_0,\dots,\mU_n,U_n,\dots)$ be a play in the game $\gone(\mO_{\mB},\mO_{\mB})$, where $\mU_n$ and $U_n=\rho((\mU_0,\dots,\mU_n))$ are the moves of player I and player II at the inning $n$, respectively. Calling $f_n=\psi((A(\mU_0),\dots,A(\mU_n)))$ for each $n$, we have that $$(A(\mU_0),f_0,\dots,A(\mU_n),f_n,\dots)$$ is a valid play in $\gone(\Omega_{\0},\Omega_{\0})$, which is won by player II since $\psi$ is a winning strategy. So, $\{f_n:n\in\omega\}\in\Omega_{\0}$, from which it follows that $\{U_n:n\in\omega\}\in\mO_{\mB}$.
\end{proof}
Actually, with similar adaptations of the arguments and definitions given by Scheepers in \cite{Scheepers1}, one can prove the following:
\begin{theorem}
Let $X$ be a Tychonoff space, $\mB$ be a bornology with compact base on $X$ and $\alpha\geq\omega$ be a countable ordinal. Player II has a winning strategy in $\gone^{\alpha}(\mO_{\mB},\mO_{\mB})$ played on $X$ if, and only if, player II has a winning strategy in $\gone^{\alpha}(\Omega_{\0},\Omega_{\0})$ played on $C_{\mB}(X)$.
\end{theorem}
By puting $\mB=\mF$ on the above theorem we obtain the original Scheepers' result. If $\mB=\frak{K}$ and $\alpha=\omega$, then we have the following:
\begin{corollary}\label{res1}
Let $X$ be a Tychonoff space. Player II has a winning strategy
in $\gone(\mK,\mK)$ played on $X$ if, and only if, player
II has a winning strategy in
$\gone(\Omega_{\0},\Omega_{\0})$ played on $C_k(X)$.
\end{corollary}
The next theorem lead us to the generalization of Theorem \ref{Uspenskii} that we are looking for. First, observe that since a bornology $\mB$ is a structure on the set $X$, it makes sense to use $l_{\mB}(X_\kappa)$ to denote the $\mB$-Lindel\"of degree of the $\kappa$-modification of $X$, i.e., the least transfinite cardinal $\lambda$ such that for every $\mB$-covering $\mU$ for $X$ made by $G_{\kappa}$-sets there exists a $\mB$-subcovering $\mU'\subset \mU$ with $|\mU'|\leq\lambda$. We would like to thank Angelo Bella for his suggestions that improved one of the implications of our original statement.
\begin{theorem}\label{Uspenskii1}
For a Tychonoff space $X$ and a bornology $\mB$ with compact base on $X$, $\kappa\in Sp(C_{\mB}(X))$ if, and only if, $l_{\mB}(X_\kappa)\leq \kappa$.
\end{theorem}
\begin{proof}
Let $\mathcal{G}$ be a nontrivial $\mB$-covering for $X$ made by $G_{\kappa}$-sets which does not contain any $\mB$-subcovering of cardinality less than or equal to $\kappa$. We will show that $C_{\mB}(X)$ is not a productively $\kappa$-tight space. Since $C_{\mB}(X)$ is a homogenous space, by Theorem \ref{ArhanSp} it follows that it is equivalent to show that $C_{\mB}(X)$ is $\kappa$-singular at $\0$.
For each $B\in\mB$, there exists $G\in\mathcal{G}$ such that $\overline{B}\subset G$ and $G=\bigcap\mU(G)$, where $\mU(G)$ is a collection of open subsets of $X$ and $|\mU(G)|\leq\kappa$. Without loss of generality, we may assume that $\mU(G)$ is closed under finite intersections. Since $X$ is a Tychonoff space, for any $U\in\mU(G)$ we may fix a function $f_{B,U}\in C(X)$ satisfying $f_{B,U}\upharpoonright \overline{B}\equiv 0$ and $f_{B,U}\upharpoonright (X\setminus U)\equiv 1$. Now, define $A_{B,G,U}=\{f_{B,V}:V\in\mU(G),V\subset U\}$ and let $\mA_{B,G}=\{A_{B,G,U}:U\in\mU(G)\}$. We claim that $\mathbb{P}=\{\mA_{B,G}:B\in\mB,G\in\mathcal{G}, \overline{B}\subset G\}$ is $\kappa$-singular at $\0$.
By construction, each member of $\mA_{B,G}$ has cardinality at most $\kappa$. Furthermore, given $U_0,\dots,U_{n}\in\mU(G)$ for some $G\in\mathcal{G}$ with $\overline{B}\subset G$, we have that $U=\bigcap_{i<n+1}U_i\in\mU(G)$ and then $f_{B,U}\in A_{B,G,U_0}\cap\dots\cap A_{B,G,U_{n}}$, from which it follows that $\mathcal{A}_{B,G}$ is a centered family. Moreover, for an arbitrary neighborhood $\langle B,\varepsilon\rangle[\0]$ of $\0$, there exists $G\in\mathcal{G}$ such that $\overline{B}\subset G$, thus $A_{B,G,U}\subset\langle B,\varepsilon\rangle[\0]$ for every $U\in\mU(G)$.
Finally, let $\{\mA_{B_\alpha,G_\alpha}:\alpha<\kappa\}\subset\mathbb{P}$. By our assumption about $\mathcal{G}$, it follows that $\{G_\alpha:\alpha<\kappa\}$ is not a $\mB$-covering for $X$ and so there is some $B\in\mB$ satisfying $B\not\subset G_\alpha$ for all $\alpha<\kappa$. Since each $G_\alpha=\bigcap\mU(G_\alpha)$, there exists $U_{\alpha}\in \mU(G_\alpha)$ such that $B\setminus U_\alpha\ne\emptyset$. Note that by the way we defined the functions in $A_{B_{\alpha},G_{\alpha},U_{\alpha}}$, it follows that each element of $A_{B_{\alpha},G_{\alpha},U_{\alpha}}$ takes value 1 in some point of $B$. Thus, the neighborhood $\langle B,\frac{1}{2}\rangle[\0]$ does not intersect $A_{B_{\alpha},G_{\alpha},U_{\alpha}}$ for all $\alpha<\kappa$, and so $\0\not\in\overline{\bigcup\{A_{B_{\alpha},G_{\alpha},U_{\alpha}}:\alpha<\kappa\}}$. Therefore, the family $\mathbb{P}$ is $\kappa$-singular at $\0$, as desired.
Conversely, suppose that every $\mB$-covering for $X$ made by $G_\kappa$-sets has a $\mB$-subcovering of cardinality less than or equal to $\kappa$. We show that $C_{\mB}(X)$ is not $\kappa$-singular at $\0$. Let $\mP=\{\mathcal{A}_{\alpha}\}_{\alpha\in I}$ be a collection satisfying conditions $(a)',(b)$ and $(c)$ in the definition of $\kappa$-singular family. We will show that condition $(d)$ does not hold.
For each $\alpha \in I$, $A\in \mathcal{A}_{\alpha}$ and $n\in\omega$, consider the set $\mathcal{U}_{A,n}=\left\{f^{-1}\left[I_n\right]:f\in A\right\},$ and let $\frak{U}_n=\{\bigcap\mathcal{U}_{A,n}:\alpha \in I$ and $A\in\mathcal{A}_{\alpha}\}$. We claim that $\frak{U}_n$ is a $\mB$-covering for $X$ made by $G_\kappa$-sets. Indeed, given $B\in\mB$, we have $\langle B,\frac{1}{n+1}\rangle[\0]$ a neighborhood of $\0$ and, by condition $(c)$, it follows that there are $\alpha \in I$ and $A\in\mathcal{A}_{\alpha}$ ($|A|\leq \kappa$ by condition $(a)'$) such that $A\subset \langle B,\frac{1}{n+1}\rangle[\0]$, so $\mU_{A,n}$ has cardinality at most $\kappa$ and $B\subset \bigcap\mathcal{U}_{A,n}$.
By the hypothesis, each $\frak{U}_n$ has a $\mB$-subcovering $\frak{U}'_n=\{U_{\lambda,n}:\lambda<\kappa\}$. Note that for each $\lambda<\kappa$ and $n\in\omega$ we may choose $\alpha_{\lambda,n}\in I$ and $C_{\lambda,n}\in \mathcal{A}_{\alpha_{\lambda,n}}$ with $U_{\lambda,n}=\bigcap\mathcal{U}_{C_{\lambda,n},n}$. We claim that $\{\mathcal{A}_{\alpha_{\lambda,n}}:\lambda<\kappa,n\in\omega\}$ witnesses that $(d)$ does not hold.
In fact, for each $\lambda<\kappa$ and $n\in\omega$ choose $A_{\lambda,n}\in\mA_{\alpha_{\lambda,n}}$, and let $\langle B,\varepsilon\rangle[\0]$ be a neighborhood of $\0$. Fix $n\in\omega$ with $\frac{1}{n+1}<\varepsilon$. Since $\frak{U}'_{n}$ is a $\mB$-subcovering, there exists an $U_{\lambda,n}$ with $B\subset U_{\lambda,n}$. So, we have $C_{\lambda,n}\subset\langle B,\varepsilon\rangle[\0]$, from which it follows that $A_{\lambda,n}\cap\langle B,\varepsilon\rangle[\0]\ne\emptyset$, since $A_{\lambda,n}\cap C_{\lambda,n}\ne \emptyset$ by condition $(b)$. This ends the proof.
\end{proof}
By making $\kappa=\aleph_0$ on the above theorem, we obtain the following:
\begin{corollary}
Let $X$ be a Tychonoff space and let $\mB$ be a bornology with compact base on $X$. $C_{\mB}(X)$ is productively countably tight if, and only if, every $\mB$-covering for $X$ made by $G_{\delta}$-sets has a countable $\mB$-subcovering.
\end{corollary}
\section{Further applications}
Recall that a space $X$ is said to be an {\bf Alster space} if for every $K$-covering for $X$ made by $G_\delta$'s there is a countable subcovering. In \cite{Alster}, it was showed that if $X$ is an Alster space, then $X\times Y$ is a Lindel\"of space for any Lindel\"of space $Y$, i.e., $X$ is a {\bf productively Lindel\"of space}, and under {\bf CH}, every productively Lindel\"of space with a base of cardinality at most $\omega_1$ is an Alster space.
So, it is natural to say that a topological space $X$ is {\bf strongly Alster} if every $K$-covering for $X$ made by $G_{\delta}$-sets has a countable $K$-subcovering. Since strongly Alster condition trivially implies Alster condition, which in turns implies productively Lindel\"ofness, we have obtained the following
\begin{corollary}\label{res2}
For a Tychonoff space $X$, $C_{k}(X)$ is productively countably tight if, and only if, $X$ is a strongly Alster space.
\end{corollary}
\begin{corollary}\label{ckpl}Let $X$ be a Tychonoff space. If $C_k(X)$ is productively countably tight, then $X$ is productively Lindel\"of.
\end{corollary}
Indeed, the strongly Alster condition is stronger than the Alster one. To see this, note that in a space in which every compact set is a $G_\delta$-set, the Alster condition is equivalent to $\sigma$-compactness, while strongly Alster condition is equivalent to hemicompactness. Thus, $C_k(\QQ)$ is not productively countably tight, since $\QQ$ is not hemicompact, while $C_p(\QQ)$ is productively countably tight, since $C_p(\QQ)$ is first countable. On the other hand, since the space $P$ of the irrational numbers is not an Alster space, it follows that $C_{\mB}(P)$ is not productively countably tight for every bornology $\mB$ on $P$ with compact base (cf. Corollary \ref{lastcor}).
We denote by $\sone(\mA,\mC)$ the following selection principle: for each sequence $(A_n)_{n\in\omega}$ of elements of $\mA$ there exists a sequence $(a_n)_{n\in\omega}$ with $a_n\in A_n$ for all $n\in\omega$ such that $\{a_n:n\in\omega\}\in\mC$. In \cite{Kocinac}, Ko\v cinac proved\footnote{A slightly adaptation of Ko\v cinac's proof shows that for any bornology $\mB$ with compact base on $X$, $C_{\mB}(X)$ satisfies $\sone(\Omega_\0,\Omega_\0)$ iff $X$ satisfies $\sone(\mO_{\mB},\mO_\mB)$.} the next result.
\begin{theorem}[Ko\v cinac, \cite{Kocinac}, Theorem 2.2]\label{Kocinac} For a Tychonoff space $X$, $C_k(X)$ satisfies $\sone(\Omega_\0,\Omega_\0)$ if, and only if, $X$ satisfies $\sone(\mK,\mK)$.
\end{theorem}
Also, in \cite{Aurichi}, Aurichi and Bella had obtained the following:
\begin{theorem}[Aurichi and Bella, \cite{Aurichi}, Corollary 2.4 and Theorem 2.5]\label{Aurichi} Let $X$ be a Tychonoff space. If player II has a winning strategy in $\gone(\Omega_{x},\Omega_{x})$ played on $X$, then $\aleph_0\in Sp(x,X)$. Also, if $\aleph_0\in Sp(x,X)$, then $X$ satisfies $\sone(\Omega_{x},\Omega_{x})$.
\end{theorem}
By changing $X$ for $C_k(X)$ on the above theorem, we obtain some interesting implications by using the results proved so far, that we summarize on the next diagram. Note that each arrow has the number of the results from which the implication follows, also, we write II $\uparrow \gone(\mA,\mC)(Y)$ to mean that player II has a winning strategy in $\gone(\mA,\mC)$ played on the space $Y$.
\small{
\begin{center}
\begin{tikzpicture}\label{diagram}
\matrix(m)[matrix of nodes, row sep=15mm, column sep=10mm,
jump/.style={text width=20mm,anchor=center},
txt/.style={anchor=center},
]
{
II $\uparrow\gone(\Omega_{\0}, \Omega_{\0})$ ($C_k(X)$) &
\begin{tabular}{c}
$C_k(X)$ is productively\\
countably tight
\end{tabular}
&
$\sone(\Omega_\0,\Omega_\0)(C_k(X))$\\
II $\uparrow\gone(\mK,\mK)(X)$ &
\begin{tabular}{c}
$X$ is strongly Alster
\end{tabular} &
$\sone(\mK,\mK)(X)$\\
&\begin{tabular}{c}
$X$ is Alster
\end{tabular}&\\
&\begin{tabular}{c}
$X$ is productively Lindel\"of
\end{tabular} &\\
};
\draw[<->] (m-2-2) to node[auto]{\ref{res2}} (m-1-2);
\draw[<->] (m-2-1) to node[auto]{\ref{res1}} (m-1-1);
\draw[<->] (m-2-3) to node[auto]{\ref{Kocinac}} (m-1-3);
\draw[->] (m-1-1) to node[auto]{\ref{Aurichi}} (m-1-2);
\draw[->] (m-1-2) to node[auto]{\ref{Aurichi}} (m-1-3);
\draw[->] (m-2-2) to (m-3-2);
\draw[->] (m-3-2) to node[auto]{\cite{Alster}} (m-4-2);
\draw[->] (m-2-1) to node[auto]{\ref{res1}+\ref{res2}+\ref{Aurichi}} (m-2-2);
\draw[->] (m-2-2) to node[auto]{\ref{res2}+\ref{Kocinac}+\ref{Aurichi}} (m-2-3);
\end{tikzpicture}
\end{center}}
\normalsize
Since any bornology $\mB$ with compact base is a subset of $\frak{K}$, it follows that any $K$-covering for $X$ is also a $\mB$-covering. Thus, Corollary \ref{ckpl} is a particular case of the following
\begin{corollary}\label{lastcor}
For a Tychonoff space $X$, if there exists a bornology $\mB$ in $X$ with a compact base such that $C_\mB(X)$ is productively countably tight, then $X$ is an Alster space, and hence $X$ is productively Lindel\"of.
\end{corollary}
|
\section{Introduction}
Electronic diodes are one of the key elements in modern electronic devices, and even play an important role in our daily lives. The most common function of a diode is to allow an electric current to pass in one direction, while blocking current in the opposite direction, thus enabling the unidirectional current flux and the rectification of an electrical signal. To date, many contributions have been made for the rectification of different types of energy flows. For example, thermal diode for thermal flow [See Ref.~\cite{thermal} and references therein], acoustic diode for sonic wave~\cite{acoustic,asym} and electromagnetic diode for electromagnetic wave~\cite{optics1,sfm,tunable_nature,left_hand,lin,e_diode,on-chip,hu,diode_passive} have been demonstrated both theoretically and experimentally. As a particular example, violation of the Lorentz reciprocity in optical system prevents the light from retracing its directional transmission and, thus, facilitates certain potential applications, such as optical diodes and isolators~\cite{comment_fan}. There are several mechanisms which allow to break the reciprocity, including magneto-optical effect ~\cite{time,DJ,on-chip,davoyan}, nonlinear material~\cite{optics1,sfm,left_hand,lin,e_diode,hu,diode_passive} and structure with time-dependent refractive index ~\cite{fan_theory,fan_experiment}. High-speed and on-chip optical circuit would require all-optical modulation, in which using nonlinear optical effect of the system could be one of the solutions.~\cite{diode_passive}.
\begin{figure}[htb]
\includegraphics[width=\columnwidth]{fig1}
\caption{\label{fig:fig1}
(Color online) Schematic view of the reversible nonreciprocal response based on the nonlinear Fano resonance. The arrows indicate the direction of the forward and backward transmission. The same system can support waves propagation predominantly either in one (a) or opposite (b) direction for different excitation signals. The color scale stands for the intensity of the field in each cite which are two particular solutions of Eq. 1. Nonlinear Fano defect is shown by a dash ring while others are linear.
}
\end{figure}
Reconfigurable light-driven isolator attracts much attention due to its capacity of switching on/off the interband transition, which manifests itself as the transition from nonreciprocal to reciprocal response, by using a control light beam~\cite{Russell}. Recently, an interesting phenomenon named {\em reversible optical nonreciprocity} was proposed where a nonreciprocal response can be flipped from transmitting a signal predominantly in one direction to opposite one~\cite{aem_non}. However, the contrast of such system is low due to the limited scattering channel. And the possibility of dynamical manipulation has not been addressed yet. By offering such possibilities for dynamically tuning of the nonreciprocity, one can direct the energy flow in real-time and thus realize more advanced control of the wave propagating systems. For example, diodes with tunable forward and backward transmission rates were proven to be significant in manipulating the properties of the wave rectification~\cite{tunable_nature}. Furthermore, if the properties of nonreciprocity can be controlled instantaneously, a lot of basic wave propagation devices, such as switch, router and rectifier etc., could be realized by employing this effect. Therefore, realizing the reconfigurable function of nonreciprocity is very crucial for precise wave manipulation, as the demonstration of reconfigurable photonic crystal (PhC) circuits~\cite{Ben}.
\begin{figure}[htb]
\includegraphics[width=\columnwidth]{fig2}
\caption{\label{fig:fig2}
(Color online) Linear transmission of the system presented in Fig.~\ref{fig:fig1}. Without loss of general, we put here $ N=1 $, $ E_{d1,d2}=0 $ and $ V_{1,2}=0.6 $.
}
\end{figure}
One of the best ways for achieving the high transmission contrast can rely on the Fano resonances, where the transmission can completely vanish due to destructive interference~\cite{fan,aem_rev,B_review,LPR}. Thus, diodes based on the Fano resonances may benefit from their typical resonant reflection and thus ultrahigh-contrast-ratio~\cite{ultra}. It was shown that coupling two nonlinear Fano defects can induce the symmetry breaking of the system~\cite{maes}. Such system reveals rich nonlinear dynamics and thus facilitates one's manipulation of the optical bistable scattering~\cite{bound1,bound2}. In this paper, we suggest that an even simpler system, consisting of a waveguide with two side-coupled Fano defects, one is with cubic nonlinearity while the other one is linear, exhibits quite interest physical property, namely, {\em dynamical reconfigurable nonreciprocity} (DRN). We demonstrate that the interaction between a linear and a nonlinear Fano resonances has a pronounced DRN. It is shown that the unidirectional transmission of nonreciprocity can be further manipulated not only by choosing the operating frequency and geometry parameters, but also dynamically with the input signal.
\section{Model}
The transmission properties of proposed nonreciprocal system shown in Fig.~\ref{fig:fig1} can be studied by using the modified Fano-Anderson model~\cite{aem1}. It allows us to derive analytical solutions of the nonlinear transmission at reversal incident directions. The equations describing the nonlinear dynamics of the scattering are:
\begin{eqnarray}
\label{eq:eq1}
&i\dot{\phi}_n&=C(\phi_{n-1}+\phi_{n+1})+\delta_{n,N} V_{1}\psi_{1} +\delta_{n,M} V_{2}\psi_{2} \nonumber\\
&i\dot{\psi}_{1}&=V_{1}\phi_{N}+E_{d1}\psi_{1} \nonumber\\
&i\dot{\psi}_{2}&=V_{2}\phi_{M}+E_{d2}\psi_{2}+\lambda\vert\psi_{2}\vert^{2} \nonumber\\
\end{eqnarray}
where the overdot stands for the derivative in time, $ \phi_{n} $ and $ \psi_{1,2} $ represent the complex fields of the sites in the chain and side coupled defects, respectively. $ E_{d1,d2} $ is the defect energy, $ \lambda $ is the nonlinearity parameter, $ C $ is the nearest-neighbour coupling constants, $ \delta_{nm} $ is the Kronecker delta symbol and $ V_{1,2} $ is the side-coupled strength between defects and the chain. The transmission matrix for the system, where $ \varepsilon_{1,2}=V^{2}_{1,2}/(\omega-E_{d1,d2}) $, $ \omega=2C\cos q $ and $ C_{q}=2C\sin q $, is shown in Eq.~\ref{eq:eq2} with $ \phi_{n}=A_{n}e^{-i\omega t} $ and $ \psi_{1,2}=B_{1,2}e^{-i\omega t} $, $ A_{n} $ and $ B_{1,2} $ are complex numbers~\cite{aem1}.
\begin{eqnarray}
\label{eq:eq2}
K=\dfrac{-1}{C_{q}^{2}}\left[
\begin{array}{ccc}
\varepsilon_{1}+iC_{q}&\varepsilon_{1}e^{-i2Nq}\\
-\varepsilon_{1}e^{i2Nq}&-\varepsilon_{1}+iC_{q}\\
\end{array}
\right]
\left[
\begin{array}{ccc}
\varepsilon_{2}+iC_{q}&\varepsilon_{2}e^{-i2Mq}\\
-\varepsilon_{2}e^{i2Mq}&-\varepsilon_{2}+iC_{q}\\
\end{array}
\right] \nonumber\\
\end{eqnarray}
Linear transmission of the system can be obtained by $ T=\vert 1/K(2,2)\vert^{2} $. Four representative examples are plotted in Fig.~\ref{fig:fig2}. As can be seen from these results, when $ M-N=2n+1 $ and $ n=1,2,3\cdots $ the transmission forms a nearly flat bottom stop band with sharp edge near the Fano resonances of the side-coupled defects.
\begin{figure}[htb]
\includegraphics[width=\columnwidth]{fig3}
\caption{\label{fig:fig3}
(Color online) (a) Nonlinear transmission of opposite incident directions. Here $ N=1 $, $ M=4 $, $ E_{d1,d2}=0 $, $ V_{1,2}=0.6 $, $ \lambda=1 $ and incident power $ \vert I\vert^{2}=0.05 $. (b) The same plot as (a) except for $ \vert I\vert^{2}=0.08 $. Blue solid/dashed red color indicates the forward/backward incident direction throughout this paper, as is indicated in Fig. 1 and Fig. 3.
}
\end{figure}
When we add a cubic nonlinearity to one of the Fano defects, above certain threshold the reciprocity can be broken. Moreover, the indirect interaction between the linear and nonlinear Fano defects via the chain suggests the possibility of the {\em reversible nonreciprocal effect}~\cite{aem_non}. In particular, incident wave propagating in opposite directions would excite the nonlinear Fano defect with different rates, as is shown in Fig.~\ref{fig:fig1}. Because the nonlinear Fano resonance is the input power dependent~\cite{aem1}, different power injection would lead to distinct shift of the nonlinear Fano resonance, which, in turn, would interact differently with the linear Fano resonance. Furthermore, the bistability of the nonlinear Fano resonance offers another opportunity to control the interaction by dynamically choosing different branches in the hysteresis loop of the bistable state~\cite{aem1}. Such tunable interaction gives rise the possibility for dynamically manipulating the physical properties of wave scattering. The nonlinear transmission of the system can be written as follow
\begin{figure*}[!t]
\includegraphics[width=2\columnwidth]{fig4}
\caption{\label{fig:fig4}
(Color online) (a) The dependence of nonlinear transmission on the input power of the system for opposite incident directions (frequency $ \omega=0.28 $). Here $ N=1 $, $ M=4 $, $ E_{d1,d2}=0 $, $ \lambda=1 $ and $ V_{1,2}=0.6 $. Solid lines represent the forward excitation while dashed lines indicate the backward one. (b) The corresponding $ T_{LR}/T_{RL} $ contrast-ratio versus input power in (a). (c) and (d) The same plots as (a) and (b) except for $ \omega=0.45 $. (a)-(d) share similar $ x $ label. The arrows indicate the transition directions between two stable states.
}
\end{figure*}
\begin{widetext}
\begin{eqnarray}
\label{eq:eq3}
T_{LR/RL}=\dfrac{\alpha_{q1}^{2}x^{2}}{2(1-\cos 2ql)(1-\alpha_{q1}x)+\alpha_{q1}^{2}x^{2}-2\sin(2ql)(\alpha_{q1}+x)+(\alpha_{q1}+x)^{2}},\nonumber\\
\end{eqnarray}
\end{widetext}
where $ \alpha_{q1}=C_{q}(\omega-E_{d1})/V_{1}^{2} $, $l=M-N$ and $ x=-\cot \theta $, $ \theta $ is the scattering phase of the nonlinear Fano defect. While $ x $ is the real solutions of the following cubic equation for the elastic scattering problem~\cite{QM}:
\begin{eqnarray}
\label{eq:eq4}
ax^{3}+bx^{2}+cx+d_{LR/RL}=0
\end{eqnarray}
with $ a=\alpha_{q1}^{2}+1$, $ b=(2\cos 2ql+1)\alpha_{q1}-2\sin 2ql-\alpha_{q2}(\alpha_{q1}^{2}+1) $, $ c=2-2\cos 2ql-2\alpha_{q1}\sin 2ql+\alpha_{q1}^{2}-2\alpha_{q2}(\alpha_{q1}\cos 2ql-\sin 2ql)$, $ d_{LR}=-\alpha_{q2}\gamma+\lambda C_{q}^{3}\alpha_{q1}^{2}I^{2}/V_{2}^{4} $ and $ d_{RL}=-\alpha_{q2}\gamma+\lambda C_{q}^{3}\alpha_{q1}^{2}I^{2}\eta/V_{2}^{4} $, $ \gamma=2-2\cos 2ql-2\alpha_{q1}\sin 2ql+\alpha_{q1}^{2} $ and $ \eta=\vert e^{3iq}-\omega V_{1}^{2}+V_{1}^{2}/\omega\vert^{2} $ for the case $ l=3 $ and $ C=1 $. Here LR means the forward transmission (blue color) and RL denotes the backward case (red color) as are indicated in Fig. 1 (a).
As can be seen from the Eq.~\ref{eq:eq4}, the asymmetry scattering of the system arises when $ d_{LR}\neq d_{RL}$ and thus different roots of the cubic equation. Intuitively, such asymmetry ($ \eta\neq 1 $) can be fulfilled by choosing suitable frequency $ \omega $ and coupling strength $ V_{1} $. Fig.~\ref{fig:fig2} (a) and (b) are two examples with different input powers $ \vert I\vert^{2}=0.05 $ and $ \vert I\vert^{2}=0.08 $, respectively. Other parameters can be found in the caption. It can be seen from these figures that the system exhibits asymmetric response for opposite incident directions. Furthermore, given certain $ \omega $ and $ V_{1} $ preserving $ \eta\neq 1 $, the input power $I^2 $ dependence of both $ d_{LR}$ and $ d_{RL}$ offers us another degree of freedom to realize advanced manipulation of the nonreciprocity. When $ d_{LR}\neq d_{RL}$, the nonlinear Fano resonances of the system occur at different input power for reversal incident directions. Therefore, ultrahigh-contrast-ratio Fano diodes with reversible nonreciprocity are presented at these two corresponding transmission dips, as shown in Fig.~\ref{fig:fig4} (a)-(d). Fig.~\ref{fig:fig4} (a) presents the dependence of the forward and backward transmissions on the input power when $ \omega=0.28 $. There are distinct transmissions for reversal incident waves. Fig.~\ref{fig:fig4} (b) shows the corresponding $ T_{LR}/T_{RL} $ in Fig.~\ref{fig:fig4} (a). These results show that we can flip the unidirectional transmission by simply controlling the input power. Fig.~\ref{fig:fig4} (c) and (d) show the case when we use suitable frequency detuning $ \omega=0.45 $, i.e. different $ \eta $, to trigger bistability. Changing the input power is corresponding to tune the interaction between the linear and nonlinear Fano resonance into distinct bistable state and thus the rectification capacity of the system. It is demonstrated that the input power dependence of $ T_{LR}/T_{RL} $ can be engineered from tunable single value to tunable bistability by choosing different $ \eta $, which would facilitate the advanced control of the rectification, as are shown in Fig.~\ref{fig:fig4} (b) and (d).
\begin{figure}[htb]
\includegraphics[width=\columnwidth]{fig5}
\caption{\label{fig:fig5}
(Color online) (a)Dynamics of the system at reversal incident direction with a Gaussian pulse with $ I=I_{0}\exp(-(t-t_{0})^{2}/W^{2})\sin(\omega t) $, where $ W=1400 $ and $ \omega=0.14 $. The geometry parameters are similar with the one in Fig.~\ref{fig:fig4} (c). The forward directions of the diode at specified pulse times are indicated by the insets. (b)Transmissions (thin lines) derived from (a) and the corresponding analytical results (thick lines). (c)Time evolutions of the nonlinear cavity excitations. (d)The effective pumping rates of the cavity. Solid lines stand for the forward excitation while dashed lines represent the backward case.
}
\end{figure}
\begin{figure}[htb]
\includegraphics[width=6.5cm]{fig6}
\caption{\label{fig:fig6}
(Color online) (a) The FDTD results of the nonlinear transmission at reversal incident directions. A Gaussian pulse with a carrier frequency $ f=0.378$ $ $ $2\pi c/a$ and $ 20 $ ps duration is used as an input probe signal. (b)-(e) are the corresponding instantaneous electric field $ \vert E\vert^{2} $ distribution marked b-e in (a), respectively. The structure details of the PhCs are outlined by the green lines. The side-coupled cavity on the right is with Kerr nonlinearity. (b)-(e) are normalized with each maximum and is saturated for better visualization. The FDTD grid is nonuniform. The grid size near the defect is $ a/100 $ to tell the small feature while the others are $ a/50 $. The computation domain is surrounded by perfect matching layers to absorb outgoing wave.
}
\end{figure}
\begin{figure*}[!t]
\includegraphics[width=2\columnwidth]{fig7}
\caption{\label{fig:fig7}
(Color online) (a) The FDTD results of the pulse dynamics at reversal incident directions ($ \omega=0.378 2\pi c/a $). The input signal, which is used as a probe pulse to access the bistability, is shown by the dotted line. (b)Reconfigurable nonreciprocity driven by a pump pulse whose duration is comparable to the nonlinear cavity's lifetime. The maximum amplitude of the pump pulse is one third of the probe one. The pump pulse is superimposed upon the input Gaussian pulse, as is shown by the dotted line in (b).
}
\end{figure*}
It is demonstrated that nonlinear Fano defect is suffered from modulational instability near the resonance under continuous-wave excitation~\cite{dynamic}, which resemble the scenario of wave scattering by a nonlinear center first addressed in Ref.~\cite{mi}. The modulation instability of our system is similar to the previous reported one~\cite{dynamic}. So, there is still a way to access the bistable hysteresis loop by a Gaussian pulses with suitable duration. Crank-Nicolson method~\cite{CK} with absorption boundary condition~\cite{dtbc} are used to solve Eqs. 1. Fig.~\ref{fig:fig5} (a) and (c) present the dynamics of the system under the excitation of a pulse where we only show part of pulse for better visualization. As can be seen from the marked forward directions of the diode in Fig.~\ref{fig:fig5} (a), the unidirectionality of the nonreciprocity is dynamically reconfigured at different time of the pulse's rising edge and falling edge, respectively. It is because the threshold power, which triggers the transition between two stable states at the cases of the increasing and decreasing input power, are both different referring to reversal incident wave, as are shown in Fig.~\ref{fig:fig5} (b). They valid that DRN could be realized by shaping the excitation condition to switch between the high-transmission states and the nonlinear Fano resonant low-transmission states. Fig.~\ref{fig:fig5} (b) also provides the analytical results (thick curves) which agrees with the dynamical simulations(thin curves). The oscillation between two bistable states indicates the transition between them and it is the property of the dynamical bistability~\cite{dynamic}. Fig.~\ref{fig:fig5} (c) and (d) present the dynamics of the power in the nonlinear Fano defect and the corresponding tunable power in nonlinear cavity, which demonstrates the incident direction dependence of the pumping rate in the nonlinear Fano defect. It is consistent with the idea of which we use directional dependent nonlinear Fano resonance to manipulate the wave transmission.
\section{Photonic crystal realization}
Now we provide a realistic example of reconfigurable nonreciprocity which is consisted of a linear photonic crystal (PhC) waveguide, with a side-coupled linear and nonlinear cavities pair. The PhC is formed by dielectric rods arranging in square lattice. The radius of the rods are $r=0.18a $, where $ a $ is the lattice constant and the refractive index $ n=3.4 $. Air is the background medium. The waveguide is created by removing one row of rods. And the nonlinear defect on the right is made by replacing one rod with a polymer rod with $r_{d}=0.1a $, $ n_{l}=1.6 $, $ n_{2}=1.14\times 10^{-12} cm^{2}/W $ and $ n=n_{l}+n_{2}I $, while the linear defect is introduced by removing one dielectric rod. The nearest rods of the linear defect are shifted $ 0.028a $ away from the center to keep the eigen-resonance frequencies of two defects the same and thus obtain flat stop band with sharp edge. The structure details are outlined by the green lines in Fig.~\ref{fig:fig6} (b)-(e). It was demonstrated that the nonlinear dynamics of this photonic structure can be described by a discrete model similar to the one presented in Sec. II~\cite{discrete}.
The finite different time domain (FDTD) simulation results [see Fig.~\ref{fig:fig6} (a) and Fig.~\ref{fig:fig7} (a)] obtained by a Gaussian pulse, show good agreement with the theoretical model. They prove that we can dynamically manipulate the nonreciprocity of the system by using the rising and falling edges of a pulse with suitable duration. The flipping of unidirectional transmission are indicated by b,c and d,e in Fig.~\ref{fig:fig6} (a), in which the forward direction of the Fano diode is reverse. We can use pump-assisting method to access low transmission case c and e marked in Fig.~\ref{fig:fig6} (a)~\cite{fan,lin}. The corresponding instantaneous electric field distributions $ \vert E\vert^{2} $ in the dynamic manipulation of the DRN are shown in Fig.~\ref{fig:fig6} (b)-(e). They manifest themselves as distinct interfering effects because different modes of the system are excited. For the Fano diode blocking the forward propagating wave, the interaction between two Fano defects is crucial [See Fig.~\ref{fig:fig6} (e)] while the nonlinear Fano defect dominates the resonant reflection when the forward direction is flipped [See Fig.~\ref{fig:fig6} (c)]. FDTD modelling detail can be found in the caption of Fig.~\ref{fig:fig6}. Furthermore, launching a suitable pump pulse (duration is comparable to the nonlinear cavity's lifetime) together with the probe one, it is possible to change the properties of the nonreciprocity. Because of distinct nonlinear feedbacks obtained at reversal excitations, their corresponding transmissions drop at different time referring to the same pump pulse, as is shown in Fig.~\ref{fig:fig7} (b). We can conclude from Fig.~\ref{fig:fig6} and Fig.~\ref{fig:fig7} that for the reversal incident waves, the nonreciprocity of the nonlinear Fano resonance system has distinct scenarios with respect to the input signal, which open up the possibility of dynamical wave manipulation.
\section{Conclusions}
We suggested a concept of the dynamical reconfigurable nonreciprocity based on a nonlinear Fano resonance system. Tunable rectification, which manifests itself via tunable bistability, and dynamical reconfiguration of diode's forward direction is theoretically investigated. The interaction between the linear and nonlinear Fano resonances plays an important role in manipulating the nonreciprocity. Numerical experiments confirm our motivation via a realistic photonic example. Our results could pave the way for the advanced manipulation of wave rectification. Because of the similarity between the modified Fano-Anderson model and the discrete nonlinear Schr$ \ddot{o} $dinger equation which is shown to be relevant in many other physical contexts, we believe our results can be generalized to similar physical system such as the coupled optical waveguide arrays system~\cite{photo_wave_array} and many others.
\section*{ACKNOWLEDGEMENT}
The work of A. E. Miroshnichenko was supported by the Australian Research Council through Future Fellowship program (FT110100037). Y. Xu acknowledges the support from the National Natural Science Foundation (Grant No.11304047), Natural Science Foundation of Guangdong Province (Grant No.S201310014807) and the Xin Miao Science Foundation of Guangzhou University.
|
\section{Introduction}
In our previous work \cite{LSZ}, the notion of a strongly Higgs semistable bundle has been introduced and it has been conjectured that a semistable Higgs bundle is strongly Higgs semistable. In this note, we give a proof of the conjecture when the rank of Higgs bundle is smaller than the characteristic $p$ of the base field.
The rank-$2$ case has been verified in \cite{LSZ}; the rank-$3$-and -$4$ case has been proved by Lingguang Li \cite{Li}.
The key point of our proof is to show the existence of a Griffiths-transverse filtration of a flat bundle such that the associated-graded Higgs bundle is semistable.
In the case of a complex curve, the existence has been proved by Simpson \cite[Theorem 2.5]{Simpson}. We apply similar techniques and establish the existence of such a filtration in Theorem \ref{goodfil}.
Early this May, Adrian Langer kindly informed us that he could prove that every semistable Higgs bundle is strongly Higgs semistable. Since his manuscript, to which we have been looking forward, has not been available in any form, we decide to write down our own proof.
\section{A gr-semistable filtration in arbitrary characteristic}
In this section, we generalize in Theorem \ref{goodfil} a result of Simpson \cite[Theorem 2.5]{Simpson}, which states that over a smooth complex projective curve, every vector bundle with an integrable holomorphic connection admits a Griffiths-transverse filtration such that the associated-graded Higgs bundle is semistable. The generalization is proved by applying the same technique as in \cite[Theorem 2.5]{Simpson}.
In this note, vector (Higgs) bundles always mean torsion-free (Higgs) sheaves; filtrations only mean those consisting of saturated torsion-free subsheaves;
$X$ is a smooth projective variety over an algebraically closed field $k$; the semistability of vector (Higgs) bundles is referred to the $\mu$-semistability with respect to an ample invertible sheaf ${\mathcal H}$ over $X$.
Let $(V, \nabla)$ be a flat vector bundle over $X$, i.e. the connection $\nabla: V\to V\otimes_{{\mathcal O}_X}\Omega^1_X$ is integrable.
We first recall some notations from \cite{Simpson}. A Griffiths-transverse filtration is a decreasing filtration of $V$ by strict subbundles
\begin{align*}
V=F^0\supset F^1\supset F^2\cdots\supset F^n\supset F^{n+1}=0
\end{align*}
which satisfies the Griffiths transversality condition
\begin{align*}
\nabla: F^i\to F^{i-1}\otimes_{{\mathcal O}_X}\Omega^1_X, \forall i.
\end{align*}
Put $E:=\text{Gr}_F(V):=\oplus_{i=0}^n E^i$ with $E^i=F^i/F^{i+1}$. By using $\nabla$, one can define an ${\mathcal O}_X$-linear map
\begin{align*}
\theta: E^i\to E^{i-1}\otimes_{{\mathcal O}_X}\Omega^1_X, \forall i.
\end{align*}
$(E, \theta)$ is called to be the associated-graded Higgs bundle corresponding to $(V, \nabla, F^*)$. We say that $(V, \nabla, F^*)$ is gr-semistable if $(E, \theta)$ is semistable as a Higgs bundle. Moreover, we need the notion of $\nabla$-semistability.
\begin{definition}
A flat bundle $(V, \nabla)$ is called to be $\nabla$-semistable if for every subbundle $V_1\subset V$ with $\nabla(V_1)\subset V_1\otimes_{{\mathcal O}_X}\Omega^1_X$, we have $\mu(V_1)\leq \mu(V)$.
\end{definition}
The goal of this section is to prove the following theorem.
\begin{theorem}\label{goodfil}
Let $(V, \nabla)$ be a $\nabla$-semistable flat bundle over a smooth projective variety $X$ defined over an algebraically closed field $k$. Then there exists a Griffiths-transverse filtration $F^*$ such that the associated-graded Higgs bundle corresponding to $(V, \nabla, F^*)$ is semistable, or equivalently, $(V, \nabla)$ admits a gr-semistable Griffiths-transverse filtration.
\end{theorem}
\begin{remark}
By \cite{Weil}, over a complex curve, every holomorphic vector bundle that admits a connection is of degree $0$, thus in this case every flat bundle $(V,\nabla)$ is naturally $\nabla$-semistable, and \cite[Theorem 2.5]{Simpson} is a special case of Theorem \ref{goodfil} when $k={\mathbb C}$. When $k$ is of characteristic $p>0$, the degree of a vector bundle that admits a connection is $0$ after modulo $p$ but not necessarily $0$, thus not every flat bundle is $\nabla$-semistable.
\end{remark}
To prove the above theorem, we need some lemmas.
\begin{lemma}\label{maxsubbundle}
Let $(E, \theta)$ be a system of Hodge bundle, i.e. $E=\oplus_{i=0}^n E^i$ with $\theta: E^i\to E^{i-1}\otimes_{{\mathcal O}_X}\Omega^1_X$. If $(E, \theta)$ is unstable as a Higgs bundle, then its maximal semistable Higgs subbundle $I\subset E$ is itself a system of Hodge bundles, that is $I=\oplus_{i=0}^n I^i$ with $I^i:=I\cap E^{i}$.
\end{lemma}
\begin{proof}
Choose $t\in k$ such that $t^i\neq 1 $ for $0<i\leq n$. Note that there is an isomorphism $f: (E, \theta)\to (E, t\theta)$ giving by $f|_{E^i}=t^i\text{id}$. Because of the uniqueness of the maximal destabilizing subobject, we see that $f(I)=I$.
Let $s$ be any local section of $I$. Write $s$ as $\sum_{i=0}^ns^i$, where $s^i$ is a local section of $E^i$. Then
\begin{align*}
\text{for }j\geq0,
f^{j} (s)=\sum_{i=0}^nt^{ji}s^i \in I.
\end{align*}
Consider
$$
\begin{pmatrix}
s\\
f(s)\\
\vdots\\
f^n(s)
\end{pmatrix} =
\begin{pmatrix}
1 & 1 &1 &\cdots &1 \\
1 & t &t^2&\cdots &t^n\\
\vdots & \vdots& \vdots& \ddots & \vdots \\
1 & t^n& t^{2n}& \cdots &t^{n^2}
\end{pmatrix} \cdot
\begin{pmatrix}
s^0\\
s^1\\
\vdots\\
s^n
\end{pmatrix}.
$$
By assumption on $t$, the coefficient matrix is invertible; thus all $s^i$'s are local sections of $I$ and $I =\oplus_{i=0}^n I\cap E^{i}$.
\end{proof}
Let $(V, \nabla)$ be a flat bundle over $X$. Start with an arbitrary Griffiths-transverse filtration $F^*$ of weight $n$ and consider the associated Higgs bundle $(\text{Gr}_F(V), \theta)$. If $(\text{Gr}_F(V), \theta)$ is unstable, let $I_F$ be the maximal semistable Higgs subbundle, which is also called the maximal destabilizing subobject; otherwise let $I_F=\text{Gr}_F(V)$. By Lemma \ref{maxsubbundle},
\begin{align*}
I_F=\oplus_{i=0}^nI_F^{i}, \quad I_F^i\subset F^i/F^{i+1} \subset V/F^{i+1}.
\end{align*}
We define an operation $\xi$ on the set of Griffiths-transverse filtrations. The new filtration $\xi(F)^*$ of $V$ is given by
\begin{align*}
\xi(F)^{i+1}:=\text{Ker}(V\to \frac{ V/F^{i+1}}{I_F^i} ), \text{ for } 0\leq i\leq n; \quad \xi(F)^0=V.
\end{align*}
Note that $F^i\supset \xi(F)^{i+1}\supset F^{i+1}$ and $\xi(F)^*$ is again Griffiths-transverse. And there is an exact sequence
\begin{align}
0\to \text{Gr}^i_F(V)/I_F^i\to \text{Gr}^i_{\xi(F)}(V) \to I_F^{i-1} \to 0, \quad \text{for } 0\leq i\leq n+1.
\end{align}
By adding all together, we obtain an exact sequence of systems of Hodge bundles
\begin{align}\label{exactseq}
0\to \text{Gr}_F(V)/I_F\to \text{Gr}_{\xi(F)}(V) \to I_F^{[1]} \to 0,
\end{align}
where $E^{[k]}$ denotes the system of Hodge bundle $E$ with Hodge index shifted so that $(E^{[k]})^i=E^{i-k}$. If $(E, \theta)$ is unstable, let $\mu_\text{max}(E)$ and $r_\text{max}(E)$ denote respectively the slope and rank of the maximal destabilizing subobject of $E$; otherwise, let $\mu_\text{max}(E)=\mu(E)$ and $r_\text{max}(E)=\text{rk}(E)$. By (\ref{exactseq}), we have
\begin{lemma}\label{invariants}
\begin{enumerate}
\item $\mu_\text{max}(\text{Gr}_{\xi(F)}(V)) \leq \mu_\text{max}(\text{Gr}_F(V)) $.
\item If $\mu_\text{max}(\text{Gr}_{\xi(F)}(V)) = \mu_\text{max}(\text{Gr}_F(V)) $, then $r_\text{max}(\text{Gr}_{\xi(F)}(V)) \leq r_\text{max}(\text{Gr}_F(V)) $.
\item Moreover, if $r_\text{max}(\text{Gr}_{\xi(F)}(V)) = r_\text{max}(\text{Gr}_F(V)) $, then (\ref{exactseq}) is split as Higgs bundles and $ I_F^{[1]}$ is the maximal destabilizing subobject of $\text{Gr}_{\xi(F)}(V)$, denoted by $ I_{\xi(F)}= I_F^{[1]}$.
\end{enumerate}
\end{lemma}
We will see that either $\mu_\text{max}$ or $r_\text{max}$ drops after applying $\xi$ finitely many times in the case of a flat $\nabla$-semistable vector bundle.
\begin{lemma}\label{terminal}
Let $F^*$ be a weight-$n$ Griffiths-transverse filtration of a flat $\nabla$-semistable vector bundle $(V,\nabla)$ over $X$. Assume that $(\text{Gr}_F(V), \theta)$ is an unstable Higgs bundle. Then at least one of the following two inequalities holds:
\begin{align*}
(1) \mu_\text{max}(\text{Gr}_{\xi^n(F)}(V))< \mu_\text{max}(\text{Gr}_F(V)); \quad (2) r_\text{max}(\text{Gr}_{\xi^n(F)}(V)) < r_\text{max}(\text{Gr}_F(V)).
\end{align*}
\end{lemma}
\begin{proof}
Let $\mu_k=\mu_\text{max}(\text{Gr}_{\xi^k(F)}(V))$ and $r_k=r_\text{max}(\text{Gr}_{\xi^k(F)}(V))$ for $k\geq 0$. By Lemma \ref{invariants}, $\{(\mu_k, r_k) | k\in {\mathbb N}\}$ decreases in the lexicographic ordering. Argue by contradiction. Suppose on the contrary that $\mu_n=\mu_0$ and $r_n=r_0$. Then for $0\leq k<n$, $\mu_{k+1}=\mu_{k}$ and $r_{k+1}=r_k$. By Lemma \ref{invariants}, $I_{\xi^{k+1}(F)}=I_{\xi^{k}(F)}^{[1]}=I_F^{[k+1]}$. It is not hard to check that for some $0< n_0\leq n$,
the exact sequence (\ref{exactseq}) for $\text{Gr}_{\xi^{n_0}(F)}(V))$
\begin{align} \label{seq3}
0\to \text{Gr}_{\xi^{n_0-1}(F)}(V)/I_F^{[n_0-1]}\to \text{Gr}_{\xi^{n_0}(F)}(V) \to I_F^{[n_0]} \to 0,
\end{align}
as graded vector bundles is of the form
\begin{align} \label{seq4}
0\to \oplus_{i=0}^{n'}E^i \to \oplus_{i=0}^{n'+k}E^i \to \oplus_{i=n'+1}^{n'+k} E^{i} \to 0.
\end{align}
Since $\mu_{n_0}=\mu_{n_0-1}$ and $r_{n_0}=r_{n_0-1}$, (\ref{seq3})
is split as Higgs bundles; as (\ref{seq3}) is of the form (\ref{seq4}), $ I_F^{[n_0]}$ corresponds to a subbundle of $V$ preserved by $\nabla$ with slope $\mu_{n_0} >\mu(V)$, contradiction with the $\nabla$-semistability of $(V,\nabla)$.
\end{proof}
\begin{proof}[Proof of Theorem \ref{goodfil}] Choose an arbitrary Griffiths-transverse filtration $F^*$ of $(V, \nabla)$.
For $ k\in {\mathbb N} $, let $\mu_k=\mu_\text{max}(\text{Gr}_{\xi^k(F)}(V))$ and $r_k=r_\text{max}(\text{Gr}_{\xi^k(F)}(V))$.
By Lemma \ref{invariants}, $\{(\mu_k, r_k)| k\in {\mathbb N} \}$ decreases in the lexicographic ordering; By Lemma \ref{terminal}, for every $(\mu_k, r_k)$, if $\mu_k>\mu(V)$, there exists some $N>0$ such that $(\mu_{k+N}, r_{k+N})< (\mu_k, r_k)$. As both $\mu_k$ and $ r_k$ can take on only finitely many values, there exists some $k_0>0$ such that the subsequence $\{(\mu_k, r_k)| k\geq k_0 \}$ is constant, which implies that $\mu_{k_0}=\mu(V)$. In other words, $( \text{Gr}_{ \xi^{k_0}(F)}(V), \theta)$ is a semistable Higgs bundle. The Griffiths-transverse filtration $\xi^{k_0}(F)^*$ of $(V, \nabla)$ is as desired.
\end{proof}
\section{Semistable Higgs bundles are strongly Higgs semistable}
We give a proof of the conjecture that semistable Higgs bundles are strongly Higgs semistable in small ranks.
In this section, $k=\overline{{\mathbb F}}_p$ for some odd prime $p$.
We first recall some notations from \cite{LSZ}.With a smooth $W_2$-lifting of $X$, one can define an inverse Cartier functor $C_0^{-1}$ from the category of (semistable) Higgs bundles over $X$ with nilpotent Higgs field of exponent $<p$ to the category of ($\nabla$-semistable) flat bundles.
A Higgs-de Rham flow is a sequence of the following form:
$$
\xymatrix{
& (H_0,\nabla_0)\ar[dr]^{Gr_{F_0}} && (H_1,\nabla_1)\ar[dr]^{Gr_{F_1}} \\
(E_0,\theta_0) \ar[ur]^{C_0^{-1}} & & (E_1,\theta_1) \ar[ur]^{C_0^{-1}}&&\ldots. }
$$
If all $(H_i, \nabla_i)$'s and $(E_i, \theta_i)$'s in a flow are defined over a finte field $k_0\supset {\mathbb F}_p$, then we say that the flow is defined over $k_0 $.
A Higgs bundle is called to be strongly Higgs semistable, if it appears as the leading term of a Higgs-de Rham flow defined over a finite field.
\begin{lemma}\label{unique fil} Let $(V,\nabla)$ be a flat vector bundle of degree zero. If there exists a Griffiths-transverse filtration $F^*$ such that the associated-graded Higgs bundle $(E,\theta)$ is stable, then such a filtration is unique up to index shifting.
\end{lemma}
\begin{proof}
Suppose that there is another gr-semistable Hodge filtration $\bar{F}^*$ different from $F^*$ up to index shifting.\\
Case 1: Suppose that there exist an integer $N$, such that for every $i$, $F^{i+N}\subset \bar{F}^i$. Then the inclusion induces a natural map $f:(E,\theta)\to (\bar{E},\bar{\theta})$. As the two filtrations are different, $f$ is neither injective nor a zero map. On one side, $Im(f)$ is a quotient object of $(E,\theta)$, $\mu(Im(f))>0$; on the other side, $Im(f)$ is also a subobject of $(\bar{E},\bar{\theta})$, $\mu(Im(f))\leq0$, contradiction.
Case 2: otherwise,
let $a$ be the largest integer such that $F^a$ is not contained in $\bar{F}^a$. And $b$ be the largest integer such that $F^{a-i}$ is contained in $F^{b-i}$ for all $i\geq 0$. Then $b<a$. And we can define a map $f:(E,\theta)\to (\bar{E},{\theta})$ as follows: for $i>0$, $f|_{E^{a+i}}=0$ , and for $j\leq 0$ $f|_{E^{a-j}}$ is
$$
E^{a-j}=F^{a-j}/F^{a-j+1}\to \bar{F}^{b-j}/\bar{F}^{b-j+1}.
$$
We see that $f$ is neither injective nor a zero map, similarly as in Case 1, we obtain a contradiction.
\end{proof}
\begin{theorem}
Let $(E_0,\theta_0)$ be a semistable Higgs bundle defined over a finite field $k_0$ with nilpotent Higgs field of exponent $<p$. Assume that $\text{rk}(E_0)<p$, then $(E_0,\theta_0)$ is strongly Higgs semistable.
\end{theorem}
\begin{proof} Let $(H_0,\nabla_0):=C_0^{-1}(E_0,\theta_0)$. Note that $(H_0,\nabla_0)$ is $\nabla$-semistable and also defined over $k_0$. By Theorem \ref{goodfil}, there exists a gr-semistable Griffiths-transverse filtration $F^*$ of $(H_0,\nabla_0)$. Because $\text{rk}(E_0)<p$, $(E_1, \theta_1):=(\text{Gr}_F(H_0), \theta)$ satisfies the same assumptions as that on $(E_0,\theta_0)$. We can apply the two functors $C_0^{-1}$ and $\text{Gr}$ alternatively and obtain a Higgs-de Rham flow with $(E_0,\theta_0)$ to be the leading term.
It suffices to o show that such a Higgs-de Rham flow is defined over a finite field. Argue by induction on $\text{rk}(E_0)$.
Assume that the statement is true when $\text{rk}(E_0)<r$ for $r>1$.
Now suppose that $(E_0,\theta_0)$ is of rank $r$.
Case 1: $(E_1, \theta_1)$ is not stable. The filtration $F^*$ is defined over a finite extension $k_1$ of $k_0$, so is $(E_1,\theta_1)$. Since $(E_1,\theta_1)$ is not stable, there exists an extension of Higgs bundles
$$0\to (E_1',\theta_1')\to (E_1,\theta_1)\to (E'',\theta'')\to 0 $$
over a finite extension $k_2$ of $k_1$ with the same slope (=0).\
By induction assumption we can assume that both $(E_1',\theta_1')$ and $(E_1, \theta_1'')$ are leading terms of flows defined over a finite field $k_3\supset k_2$.
Applying inverse Cartier functor over the extension we obtain an extension of flat bundles over $k_3$.
$$0\to (H_1',\nabla')\to (H_1,\nabla_1)_1\to (H_1'',\nabla_1'')\to 0.$$
Let $(F')^*$ (resp. $(F'')^*$) denote the gr-semistable hodge filtraion defined over $k_3$ of $H_1'$ (resp. $H_1''$) in the flows, and the associated graded Higgs bundle is $(E_2',\theta_2')$ (resp. $(E_2'',\theta_2'')$).
One can now make extension of $(F')^*$ and $(F'')^*$ to obtain
a gr-semistable Hodge filtration on $H_1$ as follows:
pulled back $(F'')^*$ via the projection $H_1\to H''_1$ one obtains
a hodge filtration $(\tilde{F}'')^*$ of $H_1$, and each term of $(\tilde {F}'')^*$ contains $H_1'$. So we can add the $(F')^*$ to $(\tilde{F}'')^*$ and form a new Hodge filtration $(F_1)^*$ of $H_1$ defined over $k_3$ .
One checks that it is a gr-semistable Hodge Filtraion on $H_1$, and the associated graded Higgs bundle $(E_2,\theta_2)$ is a direct sum of $(E_2',\theta_2')$ and $(E_2'',\theta_2'')$.
As invere cartier preserve direct sum, we can produce further terms using the datum of the two flows for $(E_1',\theta_1')$ and $(E_1'', \theta_1'')$, and all the terms are defined over $k_3$.
Case 2: $(E_1, \theta_1)$ is stable. By lemma(\ref{unique fil}) it has only one gr-semistable Hodge filtration, so the filtration must be defined over $k_0$. One go on the inverse Cartier-grading process, if in some step it run into case 1, then we are done. If not, then all the Higgs terms are stable, and the filtration are all defined over $k_0$, so the flows is already defined over $k_0$.
\end{proof}
$ $\\
As already mentioned in \cite{Li} , we have the following corollary.
\begin{corollary}
Let k be the algebraic closure of finite fields of characteristic $p >0$, and $X$ a smooth projective varieties over $k$ which has a $W_2(k)$-lifting. If $(E_i,\theta_i)$ for $i=1,2$ are semistable Higgs bundles of degree $0$, and $rank(E_1)+rank(E_2)<p$, then the tensor product $(E_1\otimes E_2, \theta_1\otimes1 +1\otimes \theta_2)$ is also semistable.
\end{corollary}
|
\section{Introduction}
We develop mathematical models for the dissemination of information on directed graphs and investigate the
influence of parameters such as the degree distribution on the dynamics of the dissemination process.
The directed graph model, as opposed to the undirected model, is better suited for networks like the
one of Twitter because of the asymmetric relationship that exists between different users. Specifically,
in the Twitter network, a user can choose to receive the tweets---in other words, become a follower---of one or more
other users by subscribing to their accounts. Certain users, celebrities for example, have several millions of
followers who follow their tweets. These users do not necessarily follow the tweets of all of their followers
which results in an asymmetric relationship between users. This asymmetry is modelled by a directed graph in
which an outgoing edge is drawn from a user to each of its followers. An edge in the opposite direction from
the follower to the user need not always exist and is drawn only if the user subscribes to the channel of this follower.
A hashtag is a word or a phrase prefixed by \# and used in social networks as a keyword. The prefix facilitates
the search for conversations related to the prefixed word or phrase. The typical life cycle of a hashtag closely
resembles an epidemic. In the first phase the interest in the hashtag grows as users
generate tweets containing this hashtag. These tweets are received by followers who then either retweet them
or generate new tweets with this hashtag. The number of users tweeting this hashtag (``infected'' users)
grows as a function of time during this phase. At a certain point in time, the interest reaches its zenith
and starts to wane as users move on and get interested in other events. The second phase begins at this point
in time as users stop tweeting this hashtag (or, ``recover''), and the number of infected users decreases.
In this work we use the mean-field approach to derive the differential equations which describe the process
of information dissemination. We obtain a couple of differential equations which describe the evolution
of the fractions of infected and recovered persons. As a model for the underlying network we take
the Configuration-type model for directed graphs \cite{CO-C13}. While epidemics have been widely studied on
undirected graphs, there is a hardly any analysis of the epidemic-type processes on directed networks.
In \cite{PS01,M02}, the mean-field approach has been applied to the analysis of epidemics on an undirected
configuration-type graph model. In \cite{GMT05}, the effect of network topology has been analysed in
the case of undirected graphs. In particular, the authors of \cite{GMT05} applied their general results
to analyse the Erd\"{o}s-R\'enyi and preferential attachment random graph models.
An interesting approach combining a decomposition approach with two-state primitive Markov chain has been
proposed in \cite{vMOK09} for undirected networks with general topology. For an overview of various results
about epidemic processes on undirected networks we refer the interested reader to the books \cite{D07,BBV08,DM10}.
\section{The mean-field model}
\label{sec:meanfield}
Consider a network of $N$ nodes structured according to the Configuration-type model for directed graphs \cite{CO-C13}.
The in-degree and out-degree of the nodes are drawn from a distribution $f(k,l) = \P(K = k, L = l)$, defined on the
bounded set $\mathcal{D} = \{(k,l) : 0 \leq k \leq \hat{K}, 0\leq l \leq \hat{L}, (k,l) \neq (0,0)\}$, where the first
(resp. second) index corresponds to the in-degree (resp. out-degree) and $\hat{K}$ (resp. $\hat{L}$) is the maximal
in-degree (resp. out-degree). In the remainder, we always assume that $\mathbb{E} K = \mathbb{E} L$; in a network every outgoing link
is an incoming link of some other node. A generic node with in-degree $k$ and out-degree $l$ shall be referred to as a $(k,l)$-node.
Each node in the network can be in one of the three states : infected, recovered, or susceptible. An infected node infects its susceptible
neighbours after an exponentially distributed time with intensity $\lambda$. Note that, the infection is spread simultaneously along
all the outgoing edges and not just one edge at a time. The simultaneous dissemination along all outgoing edges models the spread of
tweets on Twitter. An infected node recovers after an exponentially distributed time of rate $\nu$, at which time it stops spreading
information in the network.
We shall be mainly interested in a
large-population model, that is when $N\to\infty$. This assumption simplifies considerably the analysis of the dissemination process
while being realistic\footnote{Twitter has approximately $200$ million registered users
(source: Wikipedia).}.
Let $i_{k,l}(t)$ (resp. $r_{k,l}(t)$) denote the fraction of infected (resp. recovered) $(k,l)$ nodes
at time $t$. The following result describes the dynamics of these two quantities.
\begin{theorem}
Let $i_{k,l}(0) > 0$ for some $(k,l)\in\mathcal{D}$. Then, $\forall (k,l) \in \mathcal{D}$,
\begin{equation}
\frac{di_{k,l}(t)}{dt} = \lambda k (f(k,l)-i_{k,l}(t) - r_{k,l}(t))\frac{\sum_{k',l'}l'i_{k',l'}(t)}{\sum_{k',l'}l'f(k',l')}
- i_{k,l}(t)\nu,
\label{eqn:alt_master}
\end{equation}
and
\begin{equation}
\frac{dr_{k,l}(t)}{dt} = i_{k,l}(t)\nu.
\label{eqn:alt_master_r}
\end{equation}
\end{theorem}
\begin{proof}[Sketch of proof]
Let $I^{N}_{k,l}(t)$ (resp. $R^{N}_{k,l}(t)$) be the number of infected (resp. recovered) $(k,l)$ nodes in
a network of $N$ nodes. Then in a small time interval $\Delta$,
\begin{align*}
I^{(N)}_{k,l}(t+\Delta) &= I^{(N)}_{k,l}(t) + \mbox{number of $(k,l)$ nodes infected in time $\Delta$} \\
&- \mbox{number of $(k,l)$ infected $(k,l)$ nodes that recover in time $\Delta$}.
\end{align*}
Since each infected node recovers after an exponentially distributed time of rate $\nu$, the number of $(k,l)$
infected nodes that recover in $\Delta$ will be approximately $I^{(N)}_{k,l}(t)\nu\Delta$. There will be additional
terms containing $\Delta^2$ which we neglect.
Let us compute the number of $(k,l)$ nodes that get infected in time $\Delta$. There are
$N^{(N)}_{k,l} - (I^{(N)}_{k,l}(t) + R^{(N)}_{k,l}(t))$ susceptible $(k,l)$ nodes. Assume that each $(k,l)$ node has a
probability $p_{k,l}$ to get infected in the interval $\Delta$. Then, expected number of infected $(k,l)$ nodes in time
$\Delta$ will be $(N^{(N)}_{k,l}-(I^{(N)}_{k,l} + R^{(N)}_{k,l}))p_{k,l}$.
Each $(k,l)$ node has $k$ incoming edges. Assuming that the edges are connected independently,
$p_{k,l} = (1 - (1-q_{k,l})^k)$, where $q_{k,l}$ is the probability that the infection is transmitted along one of the edges in
$\Delta$.
The number of $(k,l)$ nodes infected in $\Delta$ is thus
\[
(N^{(N)}_{k,l}-(I^{(N)}_{k,l}+R^{(N)}_{k,;}))\cdot(1-(1-q_{k,l})^k),
\]
which, for $\Delta$ sufficiently small, can be approximated as
\[
(N^{(N)}_{k,l}-(I^{(N)}_{k,l} + R^{(N)}_{k,l}))kq_{k,l}.
\]
Finally, we compute $q_{k,l}$. In an interval $\Delta$, each infected node transmits the infection with probability
$\lambda \Delta$. Thus, there are $\sum_{k','l}l'I^{(N)}_{k',l'}\lambda\Delta$ edges that are infected and that transmit the
infection in $\Delta$. There are a total of $\sum_{k',l'}l'N^{(N)}_{k',l'}$. Assuming that an incoming edge is connected
uniformly at random to an outgoing edge, we obtain
\[
q_{k,l} = \frac{\sum_{k',l'}l'I^{(N)}_{k',l'}\lambda\Delta}{\sum_{k',l'}l'N^{(N)}_{k',l'}}.
\]
Consequently,
\[
I^{(N)}_{k,l}(t+\Delta) - I^{(N)}_{k,l}(t) = (N^{(N)}_{k,l}-(I^{(N)}_{k,l}
+ R^{(N)}_{k,l}))k\frac{\sum_{k',l'}l'I^{(N)}_{k',l'}\lambda\Delta}{\sum_{k',l'}l'N^{(N)}_{k',l'}}
- I^{(N)}_{k,l}\nu\Delta.
\]
If the initial number of nodes is large, then we can divide the two sides of the above equation to obtain the following difference
equation in terms of the fraction of the infected and the recovered nodes:
\[
i_{k,l}(t+\Delta) - i_{k,l}(t) = (f(k,l)-(i_{k,l}(t)
+ r_{k,l}(t)))k\frac{\sum_{k',l'}l'i_{k',l'}(t)\lambda\Delta}{\sum_{k',l'}l'f(k',l')}
- i_{k,l}(t)\nu\Delta.
\]
To complete the picture, we take the limit $\Delta \to 0$, and obtain the differential equations \eqref{eqn:alt_master}
and \eqref{eqn:alt_master_r}.
\qed
\end{proof}
\begin{remark}
In the above equation $i_{k,l}$ is the fraction of the $(k,l)$ nodes that are infected. This fraction varies between $0$ and $f(k,l)$. If
instead, we want to look at the evolution of the fraction of infected nodes and recovered nodes conditioned on them being
$(k,l)$ nodes, then the corresponding differential equations for these fractions will be
\begin{align}
\frac{di_{k,l}(t)}{dt} &= \lambda k (1-i_{k,l}(t)
- r_{k,l}(t))\frac{\sum_{k',l'}l'f(k'.l')i_{k',l'}(t)}{\sum_{k',l'}l'f(k',l')}
- i_{k,l}(t)\nu, \label{eqn:alt_master2} \\
\frac{dr_{k,l}(t)}{dt} &= i_{k,l}(t)\nu.
\label{eqn:alt_master2_r}
\end{align}
\end{remark}
\section{Epidemics without recovery}
The solution of \eqref{eqn:alt_master} and \eqref{eqn:alt_master_r} can be computed numerically. In some specific case
we can obtain explicit solutions to these equations. In particular, this is the case when there is no recovery: $\nu = 0$,
or in the language of Twitter, they keep generating new tweets with the same hashtag. That is, a hashtag never gets out
of mode. This can well represent the case for the topics or personalities that can sustain popularity over a long period of time.
Since there are no recovered nodes, $r_{k,l}(t) = 0, \, \forall t$, and \eqref{eqn:alt_master} takes the form
\begin{align}
\frac{di_{k,l}(t)}{dt} &= \lambda k (f(k,l)-i_{k,l}(t))\frac{\sum_{k',l'}l'i_{k',l'}(t)}{\sum_{k',l'}l'f(k',l')}
\label{eqn:master_nu0}.
\end{align}
The differential equation \eqref{eqn:master_nu0} can be solved in terms of a reference value of $(k,l)$, say
$(k,l) = (1,1)$ by noting that
\[
\frac{1}{k (f(k,l)-i_{k,l}(t))}\frac{di_{k,l}(t)}{dt} = \frac{1}{(f(1,1)-i_{1,1}(t))}\frac{di_{1,1}(t)}{dt},
\]
whence
\begin{align}
f(k,l)-i_{k,l}(t) &= \frac{f(k,l)-i_{k,l}(0)}{(f(1,1) - i_{1,1}(0))^k}(f(1,1)-i_{1,1}(t))^k =: c_{k,l}(f(1,1)-i_{1,1}(t))^k.
\label{eqn:par_soln_alt_mas_nu0}
\end{align}
The fraction of infected nodes of degree $(1,1)$ can be obtained by substuting the value of $i_{k,l}(t)$ in
\eqref{eqn:master_nu0} and solving it:
\begin{equation}
\frac{di_{1,1}(t)}{dt} = \lambda (f(1,1)-i_{1,1}(t)) \frac{\sum_{k',l'}l'
\left(f(k',l') - c_{k', l'}(f(1,1)-i_{1,1}(t))^{k'}\right)}{\sum_{k',l'}l'f(k',l')}.
\label{eqn:soln_alt_master}
\end{equation}
\subsection*{Deterministic in-degree}
Assume that the in-degree $K$ is deterministic and is equal to $d$.
Then, equation \eqref{eqn:master_nu0} becomes
\[
\frac{di_{d,l}(t)}{dt} = \lambda d (f(d,l)-i_{d,l}(t))\sum_{l'} \frac{l'i_{d,l'}(t)}{\sum_j jf(d,j)}.
\]
Since the expected in-degree and the expected out-degree coincide, $\sum_j jf(d,j) = d.$
Denote $\Theta(t) = \sum_j \frac{ji_{d,j}(t)}{d}$, and rewrite the above equation as:
\begin{equation}
\frac{di_{d,l}}{dt} = \lambda d (f(d,l)-i_{d,l}(t))\Theta(t).
\label{eqn:mast_fix_in_theta}
\end{equation}
Multiplying the above equation by $\frac{l}{d}$ and summing over all values $l$, we obtain the following equation for $\Theta$:
\[
\frac{d\Theta}{dt} = \lambda d(1-\Theta)\Theta,
\]
which upon integration yields:
\begin{equation}
i_{d,l}(t) = f(d,l) - c_1 e^{-\lambda d \int \Theta(t) dt}
= f(d,l) - \frac{f(d,l)-i_{d,l}(0)}{1-\Theta(0)+\Theta(0) e^{-\lambda d \cdot t}}.
\label{eqn:mast_soln_fixin}
\end{equation}
\section{Numerical experiments}
In this section, we validate the mean-field model developed in Section \ref{sec:meanfield}.
In the numerical experiments, first the in-degree and out-degree sequences are generated according to the given
degree distributions. So as to have the same number of incoming stubs as outgoing stubs, the difference between the two is
added to the smaller quantity. A configuration-type graph is then created by matching an incoming stub with an outgoing
stub chosen uniformly at random. It was shown in \cite{CO-C13} that this procedure does indeed approximates closely the
configuration model. The information diffusion process is then simulated on this graph.
For computing the solution of the system of differential equations \eqref{eqn:alt_master} and \eqref{eqn:alt_master_r}
numerically, the empirical degree distributions from the graph generated previously are given as input.
The results of two such experiments with $20000$ nodes is shown in Figure \ref{fig:experiments}. The in-degree and the
out-degree sequences were taken to be independent of each other. The out-degree sequence was drawn from a Uniform distribution
in the set $\{1,20\}$ in the two simulations. For the figure on
the left, the in-degree distribution was taken to be deterministic with parameter $10$, and for the figure on the right it was
the Zipf law on $\{1,71\}$ and exponent $1.2$. In both experiments, $\lambda = 1$ and $\nu = 0.5$, and $5$ percent of all nodes
were assumed to be infected at time $0$.
\begin{figure}[ht]
\begin{minipage}{0.45\linewidth}
\centering\includegraphics[width=1.1\linewidth]{dete10_unif20_v0.5_plot.eps}
\end{minipage}
\hspace{1cm}
\begin{minipage}{0.45\linewidth}
\centering\includegraphics[width=1.1\linewidth]{zipf71_unif20_v0.5_plot.eps}
\end{minipage}
\caption{Fraction of all nodes infected as a function of time for Deterministic in-degree distribution (left) and Zipf in-degree
distribution (right).}
\label{fig:experiments}
\end{figure}
\subsection*{Observations}
It is observed that the dissemination process is faster when the variance of the in-degree distribution is smaller. This
observation was reinforced by other experiments in which the in-degree was drawn from a Uniform distribution. In several
other experiments that we conducted, it was also observed that the out-degree distribution does not have
any noticeable effect of the dynamics of the epidemics.
Our on-going work is oriented towards investigating the influence of the variance of the in-degree distribution
and giving a theoretical foundation to the above observations.
\section{Acknowledgments}
This work was partially supported by the Parternariat Hubert Curien PHC
TOURNESOL FR 2013 29053SF between France and the Flemish community
of Belgium, by the Inria Alcatel-Lucent Joint Lab ARC ``Network Science'', and by the European Commission within
the framework of the CONGAS project FP7-ICT-2011-8-317672.
\bibliographystyle{plain}
|
\section{Supplemental Material}
\label{sec:supplement}
In this supplemental material, we derive new evolution equations for the radii of a hydrodynamically expanding cloud as a function of interaction strength near a collisional (Feshbach) resonance. The evolution equations, which are based on a scaling approximation, include the change in the pressure $\Delta p$, relative to the resonant regime, which breaks conformal symmetry, as well as the forces and heating arising from viscosity. We justify a single component hydrodynamic description by estimating the Knudsen number and investigating the transition from hydrodynamic to ballistic flow as the cloud expands. We show that by measuring the transverse aspect ratio $\sigma_x/\sigma_y$, as a function of time after release of the cloud, we are able to precisely determine the shear viscosity. We also show that $\sigma_x/\sigma_y$ is insensitive to both $\Delta p$ and the bulk viscosity, enabling a single parameter fit to the transverse aspect ratio data to determine the shear viscosity coefficient. Finally, we provide a detailed description of the data analysis and the complete results.
\subsection{Hydrodynamic Theory}
\label{sec:hydro}
To determine the viscosity on and away from the Feshbach resonance, we employ a hydrodynamic description for a single-component fluid. The single fluid description is justified in Ref.~\cite{StringariBulk} for the expansion of the resonantly interacting gas in the normal fluid regime and below the superfluid transition where the normal and superfluid components expand together. As noted in the main text, for our experiments below resonance, a finite dimer pair fraction can exist. At low temperatures, the dimers can condense into a BEC and a two-fluid description would be required. We avoid this complication by working in the normal fluid regime, where we do not observe a condensate fraction. Further justification for the single component hydrodynamic description based on the Knudsen number can be found in Sec.\ref{sec:Assumptions}.
\subsubsection{Hydrodynamic Equations}
For a single component fluid, the velocity field $\mathbf{v}(\mathbf{r},t)$ is determined by the scalar pressure and the viscous stress tensor,
\begin{equation}
n\,m\left(\partial_t +\mathbf{v}\cdot\nabla\right)v_i=-\partial_i p + \sum_j \partial_j (\eta\,\sigma_{ij}+\zeta_B\,\sigma^{'}\delta_{ij})-n\,\partial_i U_{total}.
\label{eq:force}
\end{equation}
Here $p$ is the scalar pressure and $m$ is the atom mass. $U_{total}$ is the total trapping potential energy arising from the optical trap $U_{opt}$ and the bias magnetic field curvature $U_{mag}$, as described in the main text. The second term on the right describes the friction forces
arising from both shear $\eta$ and bulk $\zeta_B$ viscosities, where $\sigma_{ij}=\partial
v_i/\partial x_j+\partial v_j/\partial
x_i-2\delta_{ij}\nabla\cdot\mathbf{v}/3$ and $\sigma^{'}\equiv\nabla\cdot\mathbf{v}$. Current conservation for the density $n({\mathbf{r}},t)$ requires
\begin{equation}
\frac{\partial n}{\partial t}+\nabla\cdot(n{\mathbf{v}})=0.
\label{eq:ncons}
\end{equation}
Finally, consistent with Eq.~\ref{eq:force} and Eq.~\ref{eq:ncons}, conservation of the total energy is described by
\begin{equation}
\frac{d}{d t}\int d^3{\mathbf{r}}\left(n\frac{1}{2}m{\mathbf{v}}^2+{\cal E}+n\,U_{total}\right)=0.
\label{eq:energycons}
\end{equation}
The first term in Eq.~\ref{eq:energycons} is the kinetic energy of the velocity field and ${\cal E}$ is the internal energy density of the cloud.
For each direction $i=x,y,z$, the mean square size $\langle x_i^2\rangle\equiv\frac{1}{N}\int d^3{\mathbf{r}}\,n({\mathbf{r}},t)\,x_i^2$ obeys
\begin{eqnarray}
\frac{d\langle x_i^2\rangle}{d t}&=&\frac{1}{N}\int d^3{\mathbf{r}}\,\frac{\partial n}{\partial t}x_i^2=\frac{1}{N}\int d^3{\mathbf{r}}\,[-\nabla\cdot(n{\mathbf{v}})]x_i^2=\frac{1}{N}\int d^3{\mathbf{r}}\,n\,{\mathbf{v}}\cdot\nabla x_i^2\nonumber\\
&=&2\langle x_i\,v_i\rangle,
\label{eq:xsq}
\end{eqnarray}
where $N$ is the total number of atoms. We have used integration by parts and $n=0$ for $x_i\rightarrow\pm\infty$ to obtain the second line. Here, and throughout the discussion, $\langle ...\rangle\equiv\int d^3{\mathbf{r}}\,(...) \,n({\mathbf{r}},t)/N$ denotes the cloud average with respect to the normalized density. Similarly,
\begin{eqnarray}
\frac{d\langle x_iv_i\rangle}{d t}&=&\frac{1}{N}\int d^3{\mathbf{r}}\,n\,x_i\frac{\partial v_i}{\partial t}+\frac{1}{N}\int d^3{\mathbf{r}}\,\frac{\partial n}{\partial t}\,x_i v_i=\frac{1}{N}\int d^3{\mathbf{r}}\,n\,x_i\frac{\partial v_i}{\partial t}+\frac{1}{N}\int d^3{\mathbf{r}}\,n\,{\mathbf{v}}\cdot\nabla (x_iv_i)\nonumber\\
&=&\langle x_i(\partial_t+{\mathbf{v}}\cdot\nabla)v_i\rangle+\langle v_i^2\rangle.
\label{eq:xv}
\end{eqnarray}
Combining Eq.~\ref{eq:xsq} and Eq.~\ref{eq:xv}, we obtain,
\begin{equation}
\frac{d^2}{d t^2}\frac{\langle x_i^2\rangle}{2}=\langle x_i(\partial_t+{\mathbf{v}}\cdot\nabla)v_i\rangle+\langle v_i^2\rangle.
\label{eq:xsqddot1}
\end{equation}
To proceed, we use Eq.~\ref{eq:force}, which yields
\begin{equation*}
\int d^3{\mathbf{r}}\,n\,x_i(\partial_t+{\mathbf{v}}\cdot\nabla)v_i= \frac{1}{m}\int d^3{\mathbf{r}}\,x_i(-\partial_i p-n\,\partial_i U_{total})+\frac{1}{m}\sum_j\int d^3{\mathbf{r}}\,x_i\partial_j (\eta\,\sigma_{ij}+\zeta_B\,\sigma^{'}\delta_{ij})
\end{equation*}
Integrating by parts on the right hand side, assuming that the surface terms vanish, we obtain
\begin{equation}
\langle x_i(\partial_t+{\mathbf{v}}\cdot\nabla)v_i\rangle=\frac{1}{Nm}\int d^3{\mathbf{r}}\,p-\frac{1}{m}\langle x_i\partial_iU_{total}\rangle-\frac{1}{Nm}\int d^3{\mathbf{r}}\,(\eta\,\sigma_{ii}+\zeta_B\,\sigma')
\end{equation}
with $\sigma'\equiv\nabla\cdot{\mathbf{v}}$.
Using $\hbar\,n$ as the natural scale of viscosity, we define the shear and bulk viscosity coefficients $\alpha_S$ and $\alpha_B$ by $\eta\equiv\alpha_S\,\hbar\,n$ and $\zeta_B\equiv\alpha_B\,\hbar\,n$, respectively. Then,
\begin{equation}
\langle x_i(\partial_t+{\mathbf{v}}\cdot\nabla)v_i\rangle=\frac{1}{Nm}\int d^3{\mathbf{r}}\,p-\frac{1}{m}\langle x_i\partial_iU_{total}\rangle-\frac{\hbar}{m}\langle\alpha_S\,\sigma_{ii}+\alpha_B\,\sigma'\rangle,
\label{eq:2.4a}
\end{equation}
where
\begin{equation}
\langle\alpha_S\,\sigma_{ii}+\alpha_B\,\sigma'\rangle\equiv\frac{1}{N}\int d^3{\mathbf{r}}\,n\,(\alpha_S\,\sigma_{ii}+\alpha_B\,\sigma').
\end{equation}
Using Eq.~\ref{eq:2.4a} in Eq.~\ref{eq:xsqddot1}, we then obtain for one direction $x_i$,
\begin{equation}
\frac{d^2}{dt^2}\frac{\langle x_i^2\rangle}{2}=\frac{1}{Nm}\int d^3{\mathbf{r}}\,p+\langle v_i^2\rangle-\frac{1}{m}\langle x_i\partial_i U_{total}\rangle-\frac{\hbar}{m}\langle\alpha_S\,\sigma_{ii}+\alpha_B\,\sigma'\rangle.
\label{eq:xsqddot2}
\end{equation}
Eq.~\ref{eq:xsqddot2} determines the evolution of the mean square cloud radii along each axis, $\langle x_i^2\rangle$, which depends on the conservative forces arising from the scalar pressure and the trap potential, as well as the viscous forces arising from the shear and bulk viscosities.
\subsection{Scaling Solution}
\label{sec:scaling}
We determine the viscosity by measuring the cloud radii and the transverse aspect ratio as a function of time after the cloud is released from the trap. To analyze the aspect ratio data, we employ a scaling solution of Eq.~\ref{eq:xsqddot2}, where the density is given by
\begin{equation}
n({\mathbf{r}},t)=\frac{n_0(x/b_x,y/b_y,z/b_z)}{\Gamma},
\label{eq:density}
\end{equation}
where $b_i(t)$, $i=x,y,z$ is a time dependent scale factor, with $b_i(0)=1$ and $\dot{b}_i(0)=0$. $n_0$ is the density profile of the trapped cloud in equilibrium. Here, $\Gamma(t)\equiv b_xb_yb_z$ is the volume scale factor, which is independent of the spatial coordinates in the scaling approximation. With Eq.~\ref{eq:density} and a velocity field that is linear in the spatial coordinates, $v_i=x_i\,\dot{b}_i/b_i$, Eq.~\ref{eq:ncons} is automatically satisfied.
We note that $\langle x_i^2\rangle=\langle x_i^2\rangle_0\,b_i^2(t)$, and $\langle v_i^2\rangle=\langle x_i^2\rangle\,\dot{b_i}^2/b_i^2=\langle x_i^2\rangle_0\,\dot{b_i}^2(t)$, where $\langle x_i^2\rangle_0$ is the mean-square cloud radius of the trapped cloud in the $i^{th}$ direction, just before release. Then, with these scaling assumptions, Eq.~\ref{eq:xsqddot2} yields
\begin{equation}
\langle x_i^2\rangle_0\,b_i\,\ddot{b}_i=\frac{1}{Nm}\int d^3{\mathbf{r}}\,p-\frac{1}{m}\langle x_i \partial_i U_{total}\rangle-\frac{\hbar}{m}\langle\alpha_S\,\sigma_{ii}+\alpha_B\nabla\cdot{\mathbf{v}}\rangle.
\label{eq:2.4}
\end{equation}
We see that Eq.~\ref{eq:2.4} contains the pressure only in a volume integral. To determine the evolution equation for the pressure integral, we use energy conservation. We begin by defining
\begin{equation}
\Delta p\equiv p-\frac{2}{3}\,{\cal E}.
\label{eq:deltap}
\end{equation}
As noted in our previous study of scale invariance~\cite{ElliottScaleInv}, $\Delta p$ is the conformal symmetry breaking pressure change, which vanishes at resonance, where $p=\frac{2}{3}\,{\cal E}$. Analogous to the methods used to derive Eq.~\ref{eq:xsqddot1}, we move the time derivatives of the velocity field and density inside the integral in Eq.~\ref{eq:energycons} and use Eq.~\ref{eq:force} and Eq.~\ref{eq:ncons} to obtain an evolution equation for the volume integral of the energy density,
\begin{equation}
\frac{d}{d t}\int d^3{\mathbf{r}}\,{\cal E}+\int d^3{\mathbf{r}}(\nabla\cdot{\mathbf{v}})\,p\, +
\int d^3{\mathbf{r}}\,n\,\frac{\partial U_{total}}{\partial t}=\dot{Q},
\label{eq:3.1e}
\end{equation}
where $\dot{Q}\equiv\int d^3{\mathbf{r}}\,\dot{q}$ is the total heating rate arising from the friction forces and $\dot{q}$ is the heating rate per unit volume,
\begin{equation}
\dot{q}=\frac{1}{2}\eta\sum_{ij}\sigma_{ij}^2+\zeta_B(\nabla\cdot{\mathbf{v}})^2.
\label{eq:2.1}
\end{equation}
Just after release of the cloud, the trap potential is constant in time and $\partial_t U_{total}$ vanishes. Then for each volume element $d^3{\mathbf{r}}$, Eq.~\ref{eq:3.1e} is just $dE_{int}=dQ-p\,dV$, where $dE_{int}$ is the internal energy in the volume element, $dQ$ is the heat added to the volume element in a time $dt$, and the work done by the volume element arises from expansion, $dV=d^3{\mathbf{r}}\,(\nabla\cdot{\mathbf{v}})\,dt$.
Using $\partial_t U_{total}=0$ and Eq.~\ref{eq:deltap} to eliminate ${\cal E}$, Eq.~\ref{eq:3.1e} takes the form
\begin{equation}
\frac{d}{d t}\int d^3{\mathbf{r}}\,p+\frac{2}{3}\int d^3{\mathbf{r}}\,(\nabla\cdot{\mathbf{v}})\,p=\frac{2}{3}\,\dot{Q}+\frac{d}{dt}\int d^3{\mathbf{r}}\,\Delta p.
\label{eq:3.2e}
\end{equation}
As we intend to explore small deviations from the scale invariant regime, the last term on the right of Eq.~\ref{eq:3.2e} can be evaluated using suitable approximations, as discussed below. It vanishes if the volume integral of $\Delta p$ is time independent.
With the scaling assumptions, $\nabla\cdot{\mathbf{v}}=\dot\Gamma/\Gamma$ is independent of the spatial coordinates, and Eq.~\ref{eq:3.2e} reduces to
\begin{equation}
\frac{d}{d t}\int d^3{\mathbf{r}}\,p+\frac{2}{3}\frac{\dot{\Gamma}}{\Gamma}\,\int d^3{\mathbf{r}}\,p=\frac{2}{3}\dot{Q}+\frac{d}{dt}\int d^3{\mathbf{r}}\,\Delta p.
\label{eq:2.9e}
\end{equation}
Using the integrating factor $\Gamma^{2/3}$, integration of Eq.~\ref{eq:2.9e} from $t=0$ to $t$ yields
\begin{equation}
3\Gamma^{2/3}\int d^3{\mathbf{r}}\,p=3\int d^3{\mathbf{r}}\,p_0+2\int_0^t dt\,\Gamma^{2/3}\dot{Q}+3\left[\Gamma^{2/3}\int d^3{\mathbf{r}}\,\Delta p-\int d^3{\mathbf{r}}\,\Delta p_0\right]-
2\int_1^\Gamma\frac{d\Gamma}{\Gamma^{1/3}}\int d^3{\mathbf{r}}\,\Delta p,
\label{eq:3.4e}
\end{equation}
where we have used $\Gamma(0)=1$. Here $p_0$ and $\Delta p_0$ denote the initial pressure and the conformal symmetry breaking pressure just after release.
Eq.~\ref{eq:3.4e} is a general consequence of energy conservation. Although it can be used to determine the evolution of $\int d^3{\mathbf{r}}\,p$ in general, it is particularly well-suited to a perturbative treatment of $\Delta p$ in the near scale-invariant regime. In that case, we can approximate the time-dependence of the temperature in $\Delta p$ as adiabatic, i.e., $T=T_0\,\Gamma^{-2/3}$, where $T_0$ is the initial temperature of the trapped cloud. Then the volume integral of $\Delta p$ becomes a known function of time, as discussed below in more detail in \S~\ref{sec:highT}.
We find the initial condition $\int d^3{\mathbf{r}}\,p_0$ from
\begin{equation}
\frac{3}{N}\int d^3{\mathbf{r}}\,p_{\,0}=\langle {\mathbf{r}}\cdot\nabla U_{total}\rangle_0\equiv\widetilde{E},
\label{eq:1.2}
\end{equation}
with the energy scale $\widetilde{E}$ defined and measured as described in the main text.
Using Eq.~\ref{eq:3.4e}, we write the time-dependent volume integral of the pressure using Eq.~\ref{eq:3.4e} in the form,
\begin{equation}
\frac{1}{N}\int d^3{\mathbf{r}}\,p=\frac{\langle{\mathbf{r}}\cdot\nabla U_{total}\rangle_0}{3\,\Gamma^{2/3}}\left[1+ C_Q(t)+ C_{\Delta p}(t)\right].
\label{eq:pressint}
\end{equation}
Here, the fractional change in the pressure integral due to viscous heating is given by $C_Q(t)$, which is determined from
\begin{equation}
\dot{C}_Q(t)\equiv\frac{\Gamma^{2/3}(t)\frac{2\dot{Q}}{N}}{\langle{\mathbf{r}}\cdot\nabla U_{total}\rangle_0},
\label{eq:6.2e}
\end{equation}
with the initial condition $C_Q(0)=0$. Using Eq.~\ref{eq:2.1} with the velocity field $v_i=x_i\,\dot{b}_i/b_i$, where $\partial_j v_i=\delta_{ij}\dot{b}_i/b_i$ is spatially constant, it is straightforward to obtain
\begin{equation}
\frac{2\dot{Q}}{N}=\hbar\,\langle \alpha_S \rangle\,\sum_i\sigma_{ii}^2+2\hbar\,\langle \alpha_B \rangle\,
\frac{\dot{\Gamma}^2}{\Gamma^2}.
\label{eq:heat2}
\end{equation}
The trap averaged-viscosity coefficients, which appear in Eq.~\ref{eq:heat2}, are defined by
\begin{eqnarray}
\langle \alpha_S \rangle&\equiv&\int d^3{\mathbf{r}}\,\eta/(N\hbar)\nonumber\\
\langle \alpha_B \rangle&\equiv&\int d^3{\mathbf{r}}\,\zeta_B/(N\hbar).
\label{eq:viscoeff}
\end{eqnarray}
In general, the trap-averaged viscosity coefficients are dependent on the scattering length $a$ and are time-dependent, as described in the main text.
In Eq.~\ref{eq:heat2},
\begin{equation}
\frac{\dot{\Gamma}}{\Gamma}=\frac{\dot{b}_x}{b_x}+\frac{\dot{b}_y}{b_y}+\frac{\dot{b}_z}{b_z}
\end{equation}
and
\begin{equation}
\sigma_{ii}=2\frac{\dot{b}_i}{b_i}-\frac{2}{3}\frac{\dot{\Gamma}}{\Gamma}.
\end{equation}
Then,
\begin{equation}
\sum_i\sigma_{ii}^2=4\sum_i\frac{\dot{b}_i^2}{b_i^2}-\frac{4}{3}\frac{\dot{\Gamma}^2}{\Gamma^2}.
\end{equation}
The time-dependent $\Delta p$ terms in Eq.~\ref{eq:3.4e} give the net fractional change in the pressure integral arising from the conformal symmetry breaking pressure $\Delta p$,
\begin{equation}
C_{\Delta p}(t)\equiv C_F(t)-C_F(0)-C_p(t),
\label{eq:Cdeltap}
\end{equation}
where
\begin{equation}
C_F(t)\equiv \frac{\Gamma^{2/3}(t)\frac{3}{N}\int d^3{\mathbf{r}}\,\Delta p}{\langle{\mathbf{r}}\cdot\nabla U_{total}\rangle_0}.
\label{eq:6.4e}
\end{equation}
and
\begin{equation}
C_p(t)\equiv\frac{2\int_1^{\Gamma(t)}\frac{d\Gamma}{\Gamma^{1/3}}\frac{1}{N}\int d^3{\mathbf{r}}\,\Delta p}{\langle{\mathbf{r}}\cdot\nabla U_{total}\rangle_0}.
\label{eq:6.5e}
\end{equation}
From Eq.~\ref{eq:6.4e} and Eq.~\ref{eq:6.5e}, we see that $C_{\Delta p}$ of Eq.~\ref{eq:Cdeltap} vanishes at $t=0$, and also when $\Delta p$ is time independent, as it should.
With Eq.~\ref{eq:pressint} for the volume integral of the pressure, Eq.~\ref{eq:2.4} yields our central result for the scale factor evolution,
\begin{equation}
\ddot{b}_i=\frac{\overline{\omega_i^2}}{\Gamma^{2/3}b_i}\left[1+ C_Q(t)+ C_{\Delta p}(t)\right]-\frac{\hbar\left(\langle \alpha_S \rangle\,\sigma_{ii}
+\langle \alpha_B \rangle\,\frac{\dot{\Gamma}}{\Gamma}\right)}{m\langle x_i^2\rangle_0 b_i}-\frac{\langle x_i \partial_i U_{mag}\rangle}{m\langle x_i^2\rangle_0 b_i}.
\label{eq:6.1e}
\end{equation}
In the last term of Eq.~\ref{eq:6.1e}, note that $U_{total}$ is replaced by the magnetic potential, $U_{mag}$ defined in the main text, as we are interested in expansion of the cloud after the optical part of the potential is extinguished.
Further, we have defined the mean square ballistic frequency for an arbitrary trapping potential, which need not be harmonic,
\begin{equation}
\overline{\omega_i^2}\equiv\frac{\langle x_i\partial_i U_{total}\rangle_0}{m\langle x_i^2\rangle_0}
=\frac{\langle{\mathbf{r}}\cdot\nabla U_{total}\rangle_0}{3m\langle x_i^2\rangle_0}.
\label{eq:4.6e}
\end{equation}
Here, $U_{total}$ is the total trap potential {\it prior} to release of the cloud. The second form follows from force balance in equilibrium, $\partial_i p+n\partial_i U_{total}=0$. Multiplying by $x_i$ and integrating by parts requires that $\langle x_i\partial_i U_{total}\rangle_0$ be the same for all directions. We determine $\overline{\omega_i^2}$ from the measured cloud profile and trap parameters, which are given in the main text.
Eq.~\ref{eq:6.1e} determines the expansion factors $b_i$ with the initial conditions $b_i(0)=1$ and $\dot{b}_i(0)=0$, using the known trap parameters and a suitable approximation for $\Delta p$ in the off-resonance case. The trap-averaged (generally time-dependent) shear and bulk viscosity coefficients, $\langle \alpha_S \rangle$ and $\langle \alpha_B \rangle$ are used as fit parameters, as described in the main text. In the experiments, we determine $\langle \alpha_S \rangle$ by fitting the predicted aspect ratios to the aspect ratio data, neglecting the much smaller $\langle \alpha_B \rangle$. The bulk viscosity coefficient $\langle \alpha_B \rangle$ is measured by observing the mean square cloud radius $\langle{\mathbf{r}}^2\rangle$, which is a scalar, as a function of time after release, as described in Ref.~\cite{ElliottScaleInv}.
\subsection{Basic Assumptions}
\label{sec:Assumptions}
Here we elucidate the basic assumptions underlying the data analysis. First, we address the question of the validity of the hydrodynamic model both on and off resonance. Next we show that for measuring the shear viscosity, we can ignore the effects of the bulk viscosity $\langle \alpha_B \rangle$ and the conformal symmetry breaking pressure $\Delta p$. Qualitatively, these scalar parameters uniformly slow or accelerate the expansion, affecting each direction in the same way. Hence, the change in the aspect ratio $\sigma_x/\sigma_y$ is suppressed. In contrast, the transverse aspect ratio $\sigma_x/\sigma_y$ is very sensitive to the shear viscosity, which directs momentum from the more rapidly expanding direction $x$ into the less rapidly expanding direction $y$.
\subsubsection{Knudsen Number and Validity of a Hydrodynamic Model}
To investigate a possible breakdown of hydrodynamics, we consider the high temperature regime, where the system is most likely to deviate from hydrodynamic flow, since the initial density $n$ and collision cross section $\sigma$ are smallest. Assuming that the cloud comprises a normal fluid mixture of atoms and dimers, the gas will be hydrodynamic and will expand as a single fluid if the Knudsen number $K_n$ is small for all species. Using a classical scattering description, we take for $K_n$ the ratio of the collisional mean free path $\lambda_{mfp}$ to the smallest diameter of the cloud $2R_x$, analogous to our previous treatment~\cite{NJPReview}. When the Knudsen number for all species is small, one expects that the dimer component of the cloud will move together with the atom component. In this case, a single-component hydrodynamic description is valid. As first shown in Ref.~\cite{PetrovDimerDimerScattLength}, for a two-component Fermi gas near a Feshbach resonance, the dimer-dimer scattering length is $0.6\,a$, where $a$ is the scattering length for a collision between a spin-up and a spin-down atom. For dimer-atom collisions, the scattering length is $1.2\,a$~\cite{PetrovAtomDimerScattLength}. We expect a small dimer fraction at high temperature, so that dimer-atom collisions are predominant in determining the hydrodynamic behavior of the dimers. Since the dimer-atom scattering cross section is larger than that for atom-atom scattering and the dimers scatter from both atomic species, i.e., from the total atomic density, the Knudsen number for the dimers will be smaller than that for the atoms. Hence, a conservative estimate of the relevant Knudsen number can be based on the mean free path for the atom component.
We assume that the dimer fraction remains constant as the cloud expands, because changing the molecular population requires three-body collisions, which occur with negligible probability during the expansion time. We take the mean free path to be $\lambda_{mfp}=1/(n_\uparrow\sigma)$, where $n_\uparrow=n/2$ is the central density in one spin state, for a 50-50 mixture. To be conservative, we take $\sigma$ to be the average {\it transport} cross section, $\sigma_{trans}$, with suppressed forward scattering, as used to estimate the viscosity in Ref.~\cite{BruunViscosity2}. For s-wave scattering, we have $\sigma(k) =4\pi a^2/(1+k^2a^2)$. Using a Maxwell-Boltzmann distribution of relative wave vectors $k$, the average transport cross section is then $\bar{\sigma}_{trans}=(2\lambda_T^2/3)\,F(q)$, where $\lambda_T=h/\sqrt{2\pi mk_BT}$ is the thermal wavelength. $F(q)\equiv \frac{1}{2}\int_0^\infty \frac{dy\,y^3\,e^{-y}}{y+q^2}$, with $q=\lambda_T/(|a|\sqrt{2\pi})$. For $\lambda_T>>|a|$, $\bar{\sigma}_{trans}\rightarrow 4\pi a^2$. Note that at resonance, $\sigma_{trans}$ is a factor of $6$ smaller than than the thermal average unitary collision cross section and therefore increases the Knudsen number by a factor of $6$ compared to an estimate based on the collisional mean free path~\cite{NJPReview}. Using $m\omega_x^2 R_x^2=2k_BT$, $E=3k_BT$, and $E_F=(3N\omega_x\omega_y\omega_z)^{1/3}\equiv k_B T_{FI}$ (Fermi energy of an ideal gas at the trap center), we find the initial Knudsen number at the cloud center, just after release
\begin{equation}
K_n=\frac{\sqrt{\pi}}{2(3N\lambda_x)^{1/3}F(q)}\left(\frac{E}{E_F}\right)^2.
\label{eq:Knudsen}
\end{equation}
Here, $q\equiv \frac{1}{k_{FI}|a|}\sqrt{\frac{6 E_F}{E}}$ and $\lambda_x\equiv \omega_y\omega_z/\omega_x^2$.
\begin{figure}
\begin{center}\
\includegraphics[width=4.0in]{knhighenergy.eps}
\end{center}
\caption{Knudsen number at the cloud center as a function of expansion time for different scattering lengths $a$ at an energy $\widetilde{E}/E_F=1.6$. Curves from top to bottom for $1/(k_{FI}|a|)=0.9,0.6,0.2,0$, respectively. \label{fig:Knudsen}}
\end{figure}
As the gas expands, several factors causes the Knudsen number $K_n=\lambda_{mfp}/(2R_x)$ to change. First, the cloud radius $R_x$ increases as $b_x(t)$, with $b_x$ a time-dependent scale factor, as described in detail in \S~\ref{sec:scaling}. Second, the mean free path $\lambda_{mfp}=2/(n\sigma)$ changes, since the density $n$ decreases as $1/\Gamma(t)$, where $\Gamma$ is the volume scale factor. Using an adiabatic approximation, the temperature decreases as $1/\Gamma^{2/3}(t)$, which causes the cross section $\sigma$ to change as well. For the resonantly interacting gas, where $q=0$ and $F(0)=1$, the net effect is that the Knudsen number {\it decreases} as the gas expands, as shown in Fig.~\ref{fig:Knudsen} for $1/(k_{FI}|a|)=0$ by the lowest (blue) curve. This is easy to understand: At resonance, the cross section increases as $\lambda_T^2\propto\Gamma^{2/3}$, so that the Knudsen number then decreases as $K_n(t)=K_n(0)\Gamma^{1/3}(t)/b_x(t)$. Hence, the resonantly interacting gas becomes {\it more} hydrodynamic as the gas expands. For finite scattering length, the Knudsen number decreases with expansion time until $\lambda_T>|a|$, when the cross section becomes constant, $\rightarrow 4\pi a^2$, and then increases as the density decreases.
\begin{figure}
\begin{center}\
\includegraphics[width=4.0in]{FigA_ARvst.eps}
\end{center}
\caption{Breakdown of hydrodynamic expansion: Aspect ratio versus time for $\widetilde{E}/E_F=1.6$ where the viscosity coefficient $\alpha_{S0}\simeq 3.0$. Blue solid curve: Hydrodynamic theory; Red dashed curve: Ballistic expansion with initial conditions set by the hydrodynamic theory at $0.5$ ms. \label{fig:HydroToBallistic}}
\end{figure}
We see from Fig.~\ref{fig:Knudsen} that for times $t>0.5$ ms, only the largest $1/(k_{FI}|a|)=0.9$ may deviate from hydrodynamic flow at the highest energies. To investigate the time-dependent breakdown of hydrodynamics, we simulate hydrodynamic flow and abruptly switch to ballistic flow for $t\geq 0.5$ ms and compare the result to hydrodynamic flow for the entire time period. Fig.~\ref{fig:HydroToBallistic} shows that there is very little difference between switching to ballistic flow for $t\geq 0.5$ ms (red dashed curve) and hydrodynamic flow at all times (blue solid curve). Hence, the asymptotic aspect ratio is determined by the hydrodynamic expansion at short times. If the expansion deviated from hydrodynamic flow by becoming ballistic, we would expect an apparent increase in the shear viscosity, in contrast to the suppression that is observed near $1/(k_{FI}|a|)=0.9$. Further, for a small dimer fraction, we would expect the Knudsen number to be smaller than that of a cloud comprised solely of atoms. We do not observe any abrupt changes in the aspect ratio versus time data for the normal fluid regime studied in the expansion experiments. Hence, we assume that a hydrodynamic description is satisfactory and that dimer-atom mixtures expand as a single fluid.
\subsubsection{Effects of Bulk Viscosity $\langle \alpha_B \rangle$ and $\Delta p$ on the Aspect Ratio}
\label{sec:highT}
The bulk viscosity and the conformal symmetry breaking pressure affect the expansion of the gas in a similar way,
as both involve scalar quantities. In Ref.~\cite{ElliottScaleInv}, we measured the bulk viscosity to be $\langle \alpha_B \rangle \leq 0.04\,\hbar n$ on resonance. Further, we found that the difference between the off-resonance and on-resonance expansion dynamics is dominated by the conformal symmetry breaking pressure $\Delta p$. The effect of the off-resonance bulk viscosity is much smaller. For this reason, we focus in this section on the $\Delta p$ correction.
When the bias magnetic field is tuned away from the Feshbach resonance, the pressure deviates from the unitary limit $\Delta p=p-\frac{2}{3}{\cal E}$. We show that $\Delta p$ has a negligible effect on the transverse aspect ratio, $\sigma_x/\sigma_y$, compared to the shear viscosity. This is accomplished using a simple model. Based on dimensional analysis, to first order in in $1/(k_Fa)$, $\Delta p$ has a natural scale $n\epsilon_F(n)/(k_Fa)$, where $\epsilon_F(n)\propto k_F^2$ is the local Fermi energy. Hence, $\Delta p\propto k_F$ requires the time dependence $\Gamma^{-1/3}$, so we take
\begin{equation}
\frac{1}{N}\int d^3{\mathbf{r}}\,\Delta p=C\,\frac{\langle{{\mathbf{r}}}\cdot\nabla U_{total}\rangle_0}{3k_{FI}a}\,\Gamma^{-1/3}(t),
\label{eq:Deltap}
\end{equation}
where $C$ is a constant. As shown in Ref.~\cite{ElliottScaleInv}, the next order (quadratic) term in $1/(k_Fa)$ is time independent, and has no effect on the expansion dynamics. Using Eq.~\ref{eq:Deltap} in Eqs.~\ref{eq:Cdeltap}-\ref{eq:6.5e}, we then obtain
\begin{equation}
C_{\Delta p}(t)=-\frac{C}{k_{FI}a}\,[\Gamma^{1/3}(t)-1]
\label{eq:C}
\end{equation}
We determine $C=0.07$ from the measured expansion of the mean square cloud radius $\langle{\mathbf{r}}^2\rangle=\langle x^2+y^2+z^2\rangle$ both on and off resonance, with $1/(k_{FI}a)=0,\pm 0.6$ and $\tilde{E}/E_F =1.0$, as shown in Fig.~4 of Ref.~\cite{ElliottScaleInv}. We then use Eq.~\ref{eq:C} in Eq.~\ref{eq:6.1e} to determine the cloud radii and the transverse aspect ratio as a function of time after release.
\begin{figure}[htb]
\begin{center}\
\hspace*{-0.25in}\includegraphics[width=5.0in]{FigureS2.eps}
\end{center}
\caption{Transverse aspect ratio $\sigma_x/\sigma_y$ for $1/(k_{FI}a)=+0.6$ as a function of time after release, calculated using Eq.~\ref{eq:6.1e} for $\widetilde{E}/E_F=1.0$, where the measured viscosity at resonance is $\langle \alpha_S \rangle_0=1.0$. Black solid line unitary result: The change in the shear viscosity relative to the unitary value $\Delta \langle \alpha_S \rangle=0$, $C_{\Delta p}=0$; Red solid line, off-resonance for $\Delta p=0$: $\Delta \langle \alpha_S \rangle=0.5$, $C_{\Delta p}=0$. Blue dashed line, off-resonance result with $\Delta p\neq 0$: $\Delta\langle \alpha_S \rangle = 0.5$, $C_{\Delta p}$ determined from Eq.~\ref{eq:C} with $C=0.07$ (as measured from the expansion of the corresponding mean square cloud radius), showing negligible effect on $\sigma_x/\sigma_y$.\label{fig:A1a}}
\end{figure}
Fig.~\ref{fig:A1a} shows that the effect of $\Delta p$ on the $\sigma_x/\sigma_y$ aspect ratio. As in the main text, we define $\Delta \langle \alpha_S \rangle$ as the change in the shear viscosity relative to the unitary value $\langle\alpha_S\rangle_0$ at the same $\widetilde{E}$. We see that the effect of $\Delta p$ is negligible compared to the effect from the change in the shear viscosity. For this reason, we neglect both $\Delta p$ and the bulk viscosity in our determination of the shear viscosity in the off-resonance regime.
\subsection{Measurement of the Shear Viscosity}
In the experiments the cloud radii are measured as a function of time after release in all three dimensions, using two simultaneous probe pulses interacting with different spin states to obtain independent absorption images on two CCD cameras~\cite{ElliottScaleInv}. The parameters measured in our experiments are the cloud radii $\sigma_i=\sqrt{2\langle x_i^2 \rangle}$ and number of atoms $N$. Using Eq.~\ref{eq:4.6e} in Eq.~\ref{eq:6.1e}, and ignoring the effects of the conformal symmetry breaking pressure and bulk viscosity as discussed above, the equations of motion for the scale factors reduce to
\begin{equation}
\ddot{b}_i=\frac{\langle x_i\partial_i U\rangle_0}{m\langle x_i^2\rangle_0\Gamma^{2/3}b_i\,}\left[1+ C_Q(t)\right]
-\frac{\hbar\langle \alpha_S \rangle\sigma_{ii}}{m\langle x_i^2\rangle_0 b_i},
-\omega^2_{i\,mag}b_i,
\label{eq:ShearEq1}
\end{equation}
where $\Gamma\equiv b_xb_yb_z$. The last term arises from the bias magnetic field curvature, where $\omega^2_{y\,mag}=\omega^2_{z\,mag}\equiv\omega^2_{mag}$ and $\omega^2_{x\,mag}=-2\,\omega^2_{mag}$, with $\omega^2_{mag}$ given in the main text. The heating term is determined from
\begin{equation}
\dot{C}_Q\equiv\frac{\hbar\,\langle \alpha_S \rangle\Gamma^{2/3}}{\langle {\mathbf{r}}\cdot\nabla U\rangle_0}\sum_i\sigma_{ii}^2,
\label{eq:ShearEq2}
\end{equation}
with
\begin{eqnarray*}
\sigma_{ii} &=&2\frac{\dot{b}_i}{b_i}-\frac{2}{3}\frac{\dot{\Gamma}}{\Gamma}\\
\sum_i\sigma_{ii}^2&=&4\sum_i\frac{\dot{b}_i^2}{b_i^2}-\frac{4}{3}\frac{\dot{\Gamma}^2}{\Gamma^2}.
\end{eqnarray*}
Eq.~\ref{eq:ShearEq1} and Eq.~\ref{eq:ShearEq2} constitute a set of four differential equations that can be solved numerically with initial values $b_i(0) = 1$, $\dot{b}_i(0) = 0$, and $C_Q(0) = 0$, providing the expansion factors $b_i$ as a function of time after release of the cloud.
In the experiments, the shear viscosity is parameterized by the scattering length-independent energy scale $\widetilde{E}\equiv\langle{\mathbf{r}}\cdot\nabla U\rangle_0$, as described in the main text. For a given value of $\langle \alpha_S \rangle$, we can then determine $\langle x^2_i \rangle_0 = \langle x^2_i\rangle / b_i^2$ as well as $\langle x_i\partial_i U_{total}\rangle_0$ and $\langle \mathbf{r}\cdot\nabla U\rangle_0$. Consistency is tested by checking that $\langle x_i\partial_iU\rangle_0$ is the same for all directions $i=x,y,z$, which follows from force balance in the trap for a scalar pressure. In the harmonic approximation this requires
\begin{equation*}
\omega_x^2 \langle x^2 \rangle_0 = \omega_y^2 \langle y^2 \rangle_0.
\end{equation*}
With an anharmonic trap, we require instead
\begin{equation*}
\omega_x^2 \langle x^2 \rangle_0[1+f_x(\widetilde{E})] = \omega_y^2 \langle y^2 \rangle_0[1+f_y(\widetilde{E})],
\end{equation*}
where the $f_{x,y}(\widetilde{E})$ are anharmonic correction factors. For the transverse directions $x$ and $y$, we assume identical correction factors, $f_x(\widetilde{E})=f_y(\widetilde{E})$. Then, for both harmonic and anharmonic traps, the transverse aspect ratio as a function of time is determined by
\begin{equation}
\sqrt{\frac{\langle x^2 \rangle}{\langle y^2 \rangle}} = \sqrt{\frac{\langle x^2 \rangle_0 b_x^2}{\langle y^2 \rangle_0 b_y^2}} = \frac{b_x}{b_y} \frac{\omega_y}{\omega_x}.
\end{equation}
Eq.~\ref{eq:ShearEq1} and Eq.~\ref{eq:ShearEq2} provide expansion factors $b_i$ that depend upon $\langle \alpha_S \rangle$ and $\langle x^2_i\rangle_0$, which in turn depends upon the expansion factors. These parameters are determined self consistently by iterative fits to the aspect ratio data, using initial guesses and then reiterating until the desired precision is achieved. Fig.\ref{fig:S1} shows curves corresponding to the model aspect ratio $\left(b_x/b_y\right)/\left(\omega_x/\omega_y\right)$ at $1.2$ $ms$ after the trap has been turned off and the corresponding measured $\widetilde{E}$ as a function of a test $\langle \alpha_S \rangle$ for a single data point. Note the sensitivity of aspect ratio to shear viscosity and the relative insensitive of energy to shear viscosity.
\begin{figure}[htb]
\begin{center}\
\includegraphics[width=5in]{FigureS1.eps}
\end{center}
\caption{Variation of the self-consistently computed energy and aspect ratio $\sigma_x/\sigma_y$ with the trial shear viscosity coefficient $\langle\alpha_S \rangle_0$. The aspect ratio (blue curve) is determined for a fixed time of $1.2$ ms after release. The energy $\widetilde{E}$ (red curve) is found from the initial cloud radii, which are determined from the cloud radii measured at 1.2 ms using the calculated expansion factors, which vary with the shear viscosity. Note that the aspect ratio is very sensitive to the shear viscosity, while the energy is not. The dots with error bars are the corresponding self-consistent values for a single measurement of the three cloud radii.
\label{fig:S1}}
\end{figure}
As described in the main text, the shear viscosity coefficient varies with time as
\begin{equation}
\langle \alpha_S \rangle = \langle \alpha_S \rangle_0 +\frac{c_1}{k_{FI} a}\, \Gamma ^{1/3}+\frac{c_2}{\left(k_{FI} a \right)^2}\, \Gamma^{2/3}
\label{eq:ShearFull}
\end{equation}
To determine the coefficients $c_1$ and $c_2$ from the $\widetilde{E}$ and $1/k{FI}$ dependent aspect ratio data, we fit the data globally.
\begin{figure}[htb]
\begin{center}\
\includegraphics[width=5in]{FigureS3.eps}
\end{center}
\caption{Unitary shear viscosity data (points) $\langle \alpha_S \rangle_0$ as a function of $\widetilde{E}/E_F$. 301 data points are The data is divided into energy bins with $0.1 \widetilde{E}/E_F$ spacing and averaged. Blue line is a polynomial $d_0 + d_1 \widetilde{E}/E_F + d_2 (\widetilde{E}/E_F)^2$ where $d_0 = -0.31$, $d_1 = 0.35$, and $d_2 = 1.14$ provides the best fit to the data.
\label{fig:S3}}
\end{figure}
First, we find the shear viscosity coefficient at resonance $\langle \alpha_S \rangle_0$ as a smooth function of energy $\widetilde{E}/E_F$. To measure the energy dependence, 301 data points are divided into energy bins with $0.1 \widetilde{E}/E_F$ spacing and averaged. Then the resulting averaged data, Fig.~\ref{fig:S3}, are fit with the polynomial $d_0 + d_1 \widetilde{E}/E_F + d_2 (\widetilde{E}/E_F)^2$, where the best fit gives $d_0 = -0.31$, $d_1 = 0.35$, and $d_2 = 1.14$, which is shown as the smooth curve.
Next, we use Eq.~\ref{eq:ShearFull} in Eq.~\ref{eq:ShearEq1} and Eq.~\ref{eq:ShearEq2} to find $c_1$ and $c_2$, where the polynomial fit for the viscosity at resonance determines $\langle \alpha_S \rangle_0$ as a function of $\widetilde{E}/E_F$. $c_1$ and $c_2$ are determined using a $\chi ^2$ fit to the aspect ratio data. As noted in the text, off-resonance, the change in the shear viscosity depends on both $\widetilde{E}$ and $1/(k_{FI} a)$ and the data deviates from the shifted parabolic fit at extreme values of $1/(k_{FI} a)$. Therefore we divide the data into discrete energy ranges as in the resonant case and also limit the range of interaction strength to $-0.5 < 1/(k_{FI} a) < 0.7$ to avoid the extreme values of $1/(k_{FI} a)$. The results of the 2-parameter $\chi^2$ fit for $c_1$ and $c_2$ are summarized in Table~\ref{tab:T1}. Errors $\Delta c_1$ (holding $c_2$ constant) and $\Delta c_2$ (holding $c_1$ constant) are estimated from the range where the normalized $\chi^2$ increases by the inverse of the number of data points (\# pts) in the fit. $\Delta \widetilde{E}/E_F$ is simply the standard deviation for the measured energy range.
\begin{table}[htb]
\large
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|}
\hline
Energy Range & \# pts & $\widetilde{E}/E_F$ & $\Delta\widetilde{E}/E_F$ & $\chi^2$ & $c_1$ & $\Delta c_1$ & $c_2$ & $\Delta c_2$ & $\frac{-c_1}{2 c_2}$ \\ \hline
0.75-0.90 & 78 & 0.83 & 0.05 & 2.2 & -1.43 & 0.07 & 2.49 & 0.10 & 0.29 \\ \hline
0.90-1.10 & 71 & 0.97 & 0.06 & 1.1 & -1.22 & 0.05 & 2.42 & 0.09 & 0.25 \\ \hline
1.10-1.30 & 123 & 1.18 & 0.05 & 2.0 & -1.11 & 0.07 & 2.19 & 0.09 & 0.25 \\ \hline
1.30-1.65 & 57 & 1.40 & 0.08 & 1.4 & -0.87 & 0.12 & 2.05 & 0.18 & 0.21 \\ \hline
1.65-2.50 & 19 & 1.90 & 0.25 & 1.4 & -0.36 & 0.12 & 1.52 & 0.22 & 0.11 \\ \hline
\end{tabular}
\label{tab:T1}
\caption{Determination of $c_1$ and $c_2$ in Eq.~\ref{eq:ShearFull}. The first column shows the range of each energy bin. $\chi^2$ is the total $\chi^2$ normalized by the number of data points in each bin. The last column gives the location of the center of the parabolic fit versus $1/(k_{FI}a)$ for each energy bin.}
\end{table}
\end{document}
|
\section{Introduction\vspace{-0.1cm}}
\label{sec:Introduction}
The minimization of the sum of a smooth function, $F$, and of a nonsmooth
(separable) convex one, $G$,
\begin{equation}\label{eq:problem 1}
\min_{\mathbf{x}\in X} V(\mathbf{x}) \triangleq F(\mathbf{x}) + G(\mathbf{x}),
\end{equation}
is an ubiquitous problem that arises in many fields of
engineering, so diverse as compressed sensing, basis pursuit denoising,
sensor networks, neuroelectromagnetic imaging, machine learning,
data mining, sparse logistic regression, genomics, metereology, tensor factorization and completion,
geophysics, and radio astronomy.
Usually the nonsmooth term is used to promote sparsity of the optimal solution,
that often corresponds to a parsimonious representation of some phenomenon at hand.
Many of the mentioned applications
can give rise to extremely large problems so that standard optimization techniques are hardly applicable.
And indeed, recent years have witnessed a flurry of research activity aimed at developing
solution methods that are simple (for example based solely on matrix/vector multiplications) but yet capable to
converge
to a good approximate solution in reasonable time. It is hardly possible here to even summarize the huge amount
of work done in this field; we refer the reader to the recent works
\cite{bach2011optimization,bradley2011parallel,buhlmann2011statistics,byrd2013inexact,fountoulakis2013second,necoara2013efficient,nesterov2012gradient,
nesterov2012efficiency,qin2010efficient,rakotomamonjy2011surveying,razaviyayn2013unified,richtarik2012iteration,
richtarik2012parallel,Sra-Nowozin-Wright_book11,tseng2009coordinate,xu2012block,yin2013parallel,
yuan2010comparison,wright2012accelerated}
as entry points to the literature.
It is clear however that if one wants to solve really large problems,
parallel methods exploiting the computational power of multi-core processors have to be employed.
It is then surprising that while serial solutions methods for Problem \eqref{eq:problem 1} have been widely investigated,
the analysis of parallel algorithms suitable to large-scale implementations lags behind.
Gradient-type methods can of course be easily parallelized. However, beyond that,
we are only aware of very few papers, all very recent,
that deal with parallel solution methods \cite{bradley2011parallel,necoara2013efficient, richtarik2012parallel,yin2013parallel}.
These papers analyze both randomized and deterministic block Coordinate Descent Methods (CDMs) that, essentially,
are still (regularized) gradient-based methods. One advantage of the analyses in \cite{bradley2011parallel,necoara2013efficient, richtarik2012parallel,yin2013parallel}
is that they provide an interesting (global) rate of convergence. On the other hand they apply only to convex problems and
are not flexible enough to include, among other things, very natural Jacobi-type methods (where at each iteration
a partial minimization of the original function is performed with respect to a block variable while all other variables are kept fixed) and the possibility to deal with a nonconvex $F$.
In this paper, building on the approach proposed in
\cite{scutari_facchinei_et_al_icassp13,scutari_facchinei_et_al_tsp13}, we present a broad, deterministic algorithmic framework for the solution of Problem \eqref{eq:problem 1} with the following novel features:
i) it is parallel, with a degree of parallelism that can be chosen by the user and that can go from a complete parallelism (each variable is updated
in parallel to all the others) to the sequential (one variable only is updated at each iteration); ii) it can tackle a nonconvex $F$; iii)
it is very flexible and includes, among others, updates based on gradient- or Newton-type methods; and iv) it easily allows for inexact
solutions.
Our framework allows us to define different schemes, \emph{all converging under the same conditions}, that can accommodate different problem features and algorithmic requirements.
Even in the most studied case in which $F$ is convex and $G(\mathbf{x}) \equiv 0$ our results compare favourably to existing ones and the numerical results
show our
approach to be very promising.
\vspace{-0.1cm}
\section{Problem Definition\vspace{-0.1cm}}
We consider Problem \eqref{eq:problem 1},
where the feasible set $X=X_1\times \cdots \times X_N$ is a cartesian product of lower dimensional convex sets $X_i\subseteq \Re^{n_i}$, and $\mathbf x\in \Re^n$ is partitioned accordingly to $\mathbf x = (\mathbf x_1, \ldots, \mathbf x_N)$, with each $\mathbf x_i \in \Re^{n_i}$. $F$ is smooth (and not necessarily convex)
and $G$ is convex and possibly nondifferentiable, with $ G(\mathbf{x}) = {\scriptstyle \sum_{i=1}^N} g_i(\mathbf{x}_i)$ with $\mathbf{x}_i \in X_i$.
This format is very general and includes problems of great interest.
Below we list some instances of Problem \eqref{eq:problem 1}.
\noindent $\bullet$
$G(\mathbf{x})=0$; in this case the problem reduces to the minimization of a smooth, possibly nonconvex problem with convex constraints.
\noindent $\bullet$
$ F(\mathbf{x}) = \|\mathbf{A}\mathbf{x} -\mathbf{b}\|^2 $ and $G(\mathbf{x})= c \|\mathbf{x}\|_1$, $X=\Re^n$, with $\mathbf{A}\in\Re^{m \times n}$, $\mathbf{b} \in \Re^m$, and $c\in \Re_{++}$ given constants; this is the very famous and much studied Lasso problem \cite{tibshirani1996regression}.
\noindent $\bullet$
$ F(\mathbf{x}) = \|\mathbf{A}\mathbf{x} -\mathbf{b}\|^2 $ and $G(\mathbf{x})= c \sum_{i=1}^N\|\mathbf{x}_i\|_2$, $X=\Re^n$, with $\mathbf{A}\in\Re^{m \times n}$, $\mathbf{b} \in \Re^m$, and $c\in \Re_{++}$
given constants; this is the group Lasso problem \cite{yuan2006model}.
\noindent $\bullet$
$ F(\mathbf{x}) = \sum_{j=1}^m \log(1 + e^{-a_i \mathbf{y}_i^T\mathbf{x} }) $ and $G(\mathbf{x})= c \|\mathbf{x}\|_1$ (or $G(\mathbf{x})= c \sum_{i=1}^N\|\mathbf{x}_i\|_2$), with $\mathbf{y}_i\in \Re^n$, $a_i\in \Re$, and $c\in \Re_{++}$ given
constants; this is the sparse logistic regression problem \cite{shevade2003simple,meier2008group}.
\noindent $\bullet$
$F(\mathbf{x}) = \sum_{j=1}^m\max\{0, 1- a_i \mathbf{y}_i^T\mathbf{x}\}^2$ and $G(\mathbf{x})= c \|\mathbf{x}\|_1$, with $a_i\in \{-1,1\}$, $\mathbf{y}_i\in \Re^n$, and $c\in \Re_{++}$ given; this is
the $\ell_1$-regularized
$\ell_2$-loss Support Vector Machine problem, see e.g. \cite{yuan2010comparison}.
\noindent $\bullet$ Other problems that can be cast in the form \eqref{eq:problem 1} include the Nuclear Norm Minimization problem, the Robust Principal Component Analysis problem, the Sparse Inverse Covariance Selection problem, the Nonnegative Matrix (or Tensor) Factorization problem, see e.g. \cite{xu2012block,goldfarb2012fast} and references therein.
\smallskip
\noindent
Given \eqref{eq:problem 1}, we make the following standard, blanket assumptions:
\begin{description}[topsep=-2.0pt,itemsep=-2.0pt]
\item[\rm (A1)] Each $X_i$ is nonempty, closed, and convex;
\item[\rm (A2)] $F$ is $C^1$ on an open set containing $X$;
\item[\rm (A3)] $\nabla F$ is Lipschitz continuous
on $X$ with constant $L_{F}$;
\item[\rm (A4)] $G(\mathbf{x}) = \sum_{i=i}^N g_i(\mathbf{x}_i)$, with all $g_i$ continuous and convex on $X_i$;
\item[\rm (A5)] $V$ is coercive.
\end{description}
\vspace{-0.2cm}
\section{Main Results\label{sec:Main Results}\vspace{-0.2cm}}
We want to develop {\em parallel} solution methods for Problem \eqref{eq:problem 1} whereby operations can be carried out on some or (possibly) all (block) variables $\mathbf{x}_i$ at
the \emph{same} time. The most natural parallel (Jacobi-type) method one can think of is updating \emph{all} blocks simultaneously: given $\mathbf{x}^k$, each (block) variable $\mathbf{x}^{k+1}_i$ is computed as the solution of
$\min_{\mathbf{x}_i}$ $[F(\mathbf{x}_i, \mathbf{x}_{-i}^k) + g_i(\mathbf{x}_i)] $ (where $\mathbf{x}_{-i}$ denotes the vector obtained from $\mathbf{x}$ by deleting the block $\mathbf{x}_i$).
Unfortunately this method converges only under very restrictive conditions \cite{Bertsekas_Book-Parallel-Comp} that are seldom verified in practice. To cope with this issue we introduce some ``memory"
and set the new point to be a convex combination of $\mathbf{x}^k$ and the solutions of $\min_{\mathbf{x}_i} [F(\mathbf{x}_i, \mathbf{x}^k_{-i}) + g_i(\mathbf{x}_i)] $.
However our framework has many additional features, as discussed next.
\noindent \textbf{Approximating $F$}: Solving each $\min_{\mathbf{x}_i} [F(\mathbf{x}_i, \mathbf{x}^k_{-i}) + g_i(\mathbf{x}_i)] $ may be too costly or difficult in
some
situations. One may then
prefer to approximate this problem, in some suitable sense, in order to facilitate the task of computing the new iteration.
To this end,
we assume that for all $i\in {\cal N} \triangleq \{1, \dots, N\}$ we
can define a function $P_{i} (\mathbf{z};\mathbf{w}) : X_i \times X \to \Re$ having the following properties (we denote by $\nabla P_{i}$
the partial gradient of $P_i$ with respect to $\mathbf{z}$):
\begin{description}[topsep=-1.0pt,itemsep=-2.0pt]
\item[\rm (P1)]
$P_{i} (\mathbf{\bullet}; \mathbf{w})$ is convex and continuously differentiable
on $ X_i$ for all $\mathbf{w}\in X$;
\item[\rm (P2)] $\nabla P_{i} (\mathbf{x}_i;\mathbf{x}) = \nabla_{{\mathbf{x}}_i} F(\mathbf{x})$ for all $\mathbf{x} \in X$;
\item[\rm (P3)] $\nabla P_{i} (\mathbf{z};\mathbf{\bullet})$ is Lipschitz continuous
on $ X$
for all $\mathbf{z} \in X_i$.
\end{description}
\smallskip
Such a function $P_i$ should be regarded as a (simple) convex approximation of $F$ at the point $\mathbf{x}$ with respect to the block of variables $\mathbf{x}_i$, that
preserves the first order properties of $F$
with respect to $\mathbf{x}_i$.
Based on this approximation we can define at any point $\mathbf{x}^k\in X$ a {\em regularized} approximation $\widetilde{h}_i (\mathbf{x}_{i};\mathbf{x}^k)$ of $V$ with respect
to $\mathbf{x}_i$ where $F$ is replaced by $P_i$ while the nondifferentiable term is preserved, and a quadratic
regularization is added to make the overall approximation strongly convex. More formally, we have
\[
\widetilde{h}_i (\mathbf{x}_{i}; \mathbf{x}^k) \! \triangleq \!\underbrace{ P_{i}(\mathbf{x}_{i}; \mathbf{x}^k)\!
+\!\dfrac{{\tau}_{i}}{2} \! \left(\mathbf{x}_{i}- \mathbf{x}_{i}^k\right)^{T}\!\!\!\!\mathbf{Q}_{i}( \mathbf{x}^k)\!\left(\mathbf{x}_{i}- \mathbf{x}^k_{i}\right)}_{\triangleq h_i(\mathbf{x}_{i}; \mathbf{x}^k) }
\!+ g_{i}(\mathbf{x}_{i}),\label{eq:convex_approx_of_fi_on_Ci}\vspace{-0.2cm}
\]
where $\mathbf{Q}_{i}(\mathbf{x}^k)$ is an $n_{i}\times n_{i}$
positive definite matrix (possibly dependent on $\mathbf{x}^k)$, satisfying the following conditions. \smallski
\begin{description}[leftmargin=*,topsep=-2.0pt,itemsep=-2.0pt]
\item[\rm (A6)] All matrices $\mathbf{Q}_{i}(\mathbf{x}^k)$ are uniformly positive definite with a common positive definiteness constant $q > 0$; furthermore, $\mathbf{Q}_{i}(\bullet)$ is Lipschitz continuous on $X$.\smallskip
\end{description}
Note that in most cases (and in all our numerical experiments) the $\mathbf{Q}_{i}$ are constant and equal to the identity matrix, so that (A6) is automatically satisfied.
Associated with each $i$ and point $ \mathbf{x}^k \in X$ we can define the following optimal solution map:\vspace{-0.1cm}
\begin{equation}\label{eq:decoupled_problem_i}
\widehat{\mathbf{x}}_{i}(\mathbf{x}^k,\tau_{i})\triangleq\underset{\mathbf{x}_{i}\in X_{i}}{\mbox{argmin}\,}\tilde{h}_{i}(\mathbf{x}_{i};\mathbf{x}^{k}).\vspace{-0.1cm}
\end{equation}
Note that $\widehat{\mathbf{x}}_{i}(\mathbf{x}^{k},\tau_{i})$
is always well-defined, since the optimization problem in (\ref{eq:decoupled_problem_i})
is strongly convex. Given (\ref{eq:decoupled_problem_i}),
we can then introduce
\[
X\ni\mathbf{y}\mapsto\widehat{\mathbf{x}}(\mathbf{y},\boldsymbol{{\tau}})\triangleq\left(\widehat{\mathbf{x}}_{i}(\mathbf{y},\tau_{i})\right)_{i=1}^{N}.
\]
The algorithm we are about to described is based on the computation of $\widehat{\mathbf{x}}$. Therefore the approximating functions $P_i$ should
lead to as easily computable functions $\widehat{\mathbf{x}}$ as possible. An appropriate choice depends on the problem at hand and on computational
requirements. We discuss some possible choice for the approximations $P_i$ after introducing the main algorithm (Algorithm \ref{alg:general}).
\noindent \textbf{Inexact solutions:} In many situations (especially in the case of large-scale problems), it can be useful to
further reduce the computational effort needed to solve the subproblems in \eqref{eq:decoupled_problem_i} by allowing \emph{inexact} computations $\mathbf{z}^k$ of $\widehat{\mathbf{x}}_{i}(\mathbf{x}^k,\tau_{i})$, i.e., $\|\mathbf{z}_{i}^{k}-\widehat{\mathbf{x}}_i\left(\mathbf{x}^{k},\boldsymbol{{\tau}}\right)\|\leq\varepsilon_{i}^{k}$, where $\varepsilon_{i}^{k}$ measures the accuracy in computing the solution.
\noindent \textbf{Updating only some blocks:} Another important feature of our algorithmic framework is the possibility of updating only some of the variables at each iteration.
Essentially we prove convergence assuming that at each iteration only a subset of
the variables is updated under the condition that this subset contains at least one (block) component
which is within a factor $\rho \in (0,1]$ from being ``sufficiently far'' from optimality, in the sense explained next.
First of all $\mathbf{x}^k_i$ is optimal for $\tilde{h}_{i}(\mathbf{x}_{i};\mathbf{x}^k)$ if and only if
$\widehat{\mathbf{x}}_{i}(\mathbf{x}^k,\tau_{i})=\mathbf{x}^k_i$. Ideally we would like then to select the indices to update according to the optimality measure $\|\widehat{\mathbf{x}}_{i}(\mathbf{x}^k,
\tau_{i})-\mathbf{x}^k_i\|$; but in some situations this could be computationally too expensive. In order to be able to develop alternative choices, based
on the same idea, we suppose one can compute an {\em error bound}, i.e., a function $E_i(\mathbf{x})$ such that\vspace{-0.1cm}
\begin{equation}\label{eq:error bound}
\underbar s_i\|\widehat{\mathbf{x}}_{i}(\mathbf{x}^k,\tau_{i})-\mathbf{x}^k_i\| \le E_i(\mathbf{x}^k) \leq \bar s_i\|\widehat{\mathbf{x}}_{i}(\mathbf{x}^k,\tau_{i})-\mathbf{x}^k_i\|,\vspace{-0.1cm}
\end{equation}
for some $0< \underbar s_i \le \bar s_i$. Of course we can always set $ E_i(\mathbf{x}^k) = \|\widehat{\mathbf{x}}_{i}(\mathbf{x}^k,\tau_{i})-\mathbf{x}^k_i\| $, but other choices are also possible; we discuss
some of them after introducing the algorithm.
We are now ready to formally introduce our algorithm, Algorithm 1, that enjoys all the features discussed above. Its convergence properties are given in Theorem \ref{Theorem_convergence_inexact_Jacobi}, whose proof is omitted because of space limitation, see \cite{FacchineiSagratellaScutariMPsub13}
\vspace{-0.4cm}
\begin{algo}{\textbf{Inexact Parallel Algorithm}} S$\textbf{Data}:$ $\{\varepsilon_{i}^{k}\}$
for $i\in\mathcal{N}$, $\boldsymbol{{\tau}}\geq\mathbf{0}$, $\{\gamma^{k}\}>0$,
$\mathbf{x}^{0}\in X$, $\rho \in (0,1]$.
\hspace{1.24cm}Set $k=0$.
\texttt{$\mbox{(S.1)}:$}$\,\,$If $\mathbf{x}^{k}$ satisfies a termination
criterion: STOP;
\texttt{$\mbox{(S.2)}:$} For all $i\in\mathcal{N}$, solve (\ref{eq:decoupled_problem_i})
with accuracy $\varepsilon_{i}^{k}:$
\hspace{1.24cm}Find $\mathbf{z}_{i}^{k}\in X_i$ s.t. $\|\mathbf{z}_{i}^{k}-\widehat{\mathbf{x}}_i\left(\mathbf{x}^{k},\boldsymbol{{\tau}}\right)\|\leq\varepsilon_{i}^{k}$;
\texttt{$\mbox{(S.3)}:$} Set $M^k \triangleq \max_i \{E_i(\mathbf{x}^k)\}$.
\hspace{1.24cm}Choose a set $S^k$ that contains at least one index $i$
\hspace{1.24cm}for which
$E_i(\mathbf{x}^k) \geq \rho M^k.$
\hspace{1.24cm}Set $\widehat{\mathbf{z}}^k_i = \mathbf{\mathbf{z}}_{i}^k$ for $i\in S^k$ and
$\widehat{\mathbf{z}}^k_i = \mathbf{x}^k_i$ for $i\not \in S^k$
\texttt{$\mbox{(S.4)}:$} Set $\mathbf{x}^{k+1}\triangleq\mathbf{x}^k+\gamma^{k}\,(\widehat{\mathbf{z}}^{k}-\mathbf{x}^{k})$;
\texttt{$\mbox{(S.5)}:$} $k\leftarrow k+1$, and go to \texttt{$\mbox{(S.1)}.$}
\label{alg:general}
\end{algo}\vspace{-0.4cm}
\begin{theorem} \label{Theorem_convergence_inexact_Jacobi}Let
$\{\mathbf{x}^{k}\}$ be the sequence generated by
Algorithm \ref{alg:general}, under A1-A6.
Suppose that $\{\gamma^{k}\}$
and $\{\varepsilon_{i}^{k}\}$ satisfy the following conditions: i)
$\gamma^{k}\in(0,1]$; ii) $\gamma^{k}\rightarrow0$; iii) $\sum_{k}\gamma^{k}=+\infty$;
iv) $\sum_{k}\left(\gamma^{k}\right)^{2}<+\infty$;
and v) $\varepsilon_{i}^{k}\leq \gamma^k \alpha_1\min\{\alpha_2, 1/\|\nabla_{\mathbf{x}_i} F(\mathbf{x}^k)\| \}$
for all $i\in {\cal N}$ and some nonnegative constants $\alpha_1$ and $\alpha_2$.
Additionally, if inexact solutions are used in Step S.2, i.e., $\varepsilon_{i}^{k}>0$ for some $i$ and infinite $k$, then
assume also that $G$ is globally Lipschitz on $X$.
Then, either Algorithm \ref{alg:general} converges in a finite number of iterations to a stationary solution
of \eqref{eq:problem 1} or every limit point of
$\{\mathbf{x}^{k}\}$ (at least
one such points exists) is a stationary solution of \eqref{eq:problem 1}.\vspace{-0.1cm}
\end{theorem}
In the theorem we obtain convergence to stationary points $\mathbf{x}^*$, i.e. points for which a subgradient $\xi \in \partial G(\mathbf{x}^*)$ exists such that
$(\nabla F(\mathbf{x}^*) $ $+ \xi)^T(\mathbf{y}-\mathbf{x}^*) \geq 0$ for all $\mathbf{y}\in X$. Of course, if $F$ is convex, stationary points coincide with global minimizers.
\smallskip
\noindent \textbf{On Algorithm 1.} The proposed algorithm is extremely flexible. We can always choose $S^k ={\cal N}$ resulting in the simultaneous update of all the (block) variables (full Jacobi scheme); or, at the other extreme, one can update a single (block) variable per time, thus obtaining a Gauss-Southwell kind of method.
One can also compute inexact solutions (Step 2) while preserving convergence, provided that the error term $\varepsilon_{i}^{k}$ and the step-size $\gamma^{k}$\textquoteright{}s are chosen according
to Theorem 1. We emphasize that the Lipschitzianity of $G$ is required only if $\widehat{\mathbf{x}}(\mathbf{x}^k, \tau)$ is not computed exactly for infinite iterations. At any rate this Lipschitz conditions is automatically satisfied if $G$ is a norm (and therefore in Lasso and group Lasso problems for example) or if
$X$ is bounded.\smallskip
\noindent \textbf{On the choice of the stepsize $\gamma^k$}\emph{.}
An example of step-size rule satisfying i-iv in Theorem \ref{Theorem_convergence_inexact_Jacobi} is:
given $\gamma^{0}=1$, let
\begin{equation}\label{eq:gamma}
\gamma^{k}=\gamma^{k-1}\left(1-\theta\,\gamma^{k-1}\right),\quad k=1,\ldots,\vspace{-0.1cm}
\end{equation}
where $\theta\in(0,1)$ is a given constant; see \cite{scutari_facchinei_et_al_tsp13}
for others rules. This is actually the rule we used in our practical experiments, see next section. Notice that while this rule
may still require some tuning for optimal behaviour, it is quite reliable, since in general we are not
using a (sub)gradient direction, so that many of the well-known practical drawbacks associated
with a (sub)gradient method with diminishing step-size are mitigated in our setting. Furthermore, this choice of step-size does not require any form of centralized coordination, which is a favourable feature in a parallel environment.
We remark that it is possible to
prove convergence of Algorithm 1 also using other step-size rules, such as a standard Armijo-like line-search procedure or a (suitably small) constant step-size; see \cite{FacchineiSagratellaScutariMPsub13} for more details.
We omit the discussion of these options because of lack of space, but also because
the former is not in line with our parallel approach while the latter is numerically less efficient.
\smallskip
\noindent \textbf{On the choice of $E_i(\mathbf{x})$}\emph{.}
\noindent $\bullet$
As we mentioned, the most obvious choice is to take $E_i(\mathbf{x}) = \|\widehat{\mathbf{x}}_{i}(\mathbf{x}^k,\tau_{i})-\mathbf{x}^k_i\|$.
This is a valuable choice if the computation of $\widehat{\mathbf{x}}_{i}(\mathbf{x}^k,\tau_{i})$ can be easily accomplished. For instance, in the
Lasso problem with ${\cal N} = \{1, \ldots, n\}$ (i.e., when each block reduces to a scalar variable),
it is well-known that $\widehat{\mathbf{x}}_{i}(\mathbf{x}^k,\tau_{i})$ can be computed in closed form using
the soft-thresholding operator.
\noindent $\bullet$ In situations where the computation of $\|\widehat{\mathbf{x}}_{i}(\mathbf{x}^k,\tau_{i})-\mathbf{x}^k_i\|$ is not possible or advisable,
we can resort to estimates. To make the discussion simple, assume momentarily that $G\equiv 0$. Then it is known \cite{Facchinei-Pang_FVI03} that
$\|\Pi_{X_i} (\mathbf{x}^k_i - \nabla_{\mathbf{x}_i}F(\mathbf{x}^k)) - \mathbf{x}^k_i\|$ is an error bound for the minimization problem in \eqref{eq:decoupled_problem_i} and therefore
satisfies \eqref{eq:error bound}. In this situation we can choose $E_i(\mathbf{x}^k) = \|\Pi_{X_i} (\mathbf{x}^k_i - \nabla_{\mathbf{x}_i}F(\mathbf{x}^k)) - \mathbf{x}^k_i\|$.
If $G(\mathbf{x}) \not \equiv 0$ we can easily reduce to the case $G\equiv 0$ by a simple transformation; the details are omitted for lack of space, see \cite{scutari_facchinei_et_al_tsp13}.
\noindent $\bullet$ It is interesting to note that the computation of $E_i$ is only needed if a partial update of the (block) variables is
performed. However, an option that is always feasible is to take $S^k = {\cal N}$ at each iteration, i.e., update all (block) variables at each iteration.
With this choice we can dispense with the computation of $E_i$ altogether.\smallskip
\noindent \textbf{On the choice of $P_i(\mathbf{x}_i;\mathbf{x} )$}.
\noindent $\bullet$ The most obvious choice for $P_i$ is the linearization of $F$ at $\mathbf{x}^k$ with respect to $\mathbf{x}_i$:
$ P_i(\mathbf{x}_i; \mathbf{x}^k) = F(\mathbf{x}^k) + \nabla_{\mathbf{x}_i}F(\mathbf{x}^k)^T(\mathbf{x}_i - \mathbf{x}_i^k)$. With this choice, and taking for simplicity $\mathbf{Q}_i(\mathbf{x}^k) = \mathbf I$,
$\widehat {\mathbf x}_i(\mathbf{x}^k, \tau_i) $ is given by\vspace{-0.2cm}
\begin{equation}\label{eq:proposal 1}
\underset{\mathbf{x}_{i}\in X_{i}}{\mbox{argmin}\,} F(\mathbf{x}^k) + \nabla_{\mathbf{x}_i}F(\mathbf{x}^k)^T(\mathbf{x}_i - \mathbf{x}_i^k) + \frac{\tau_i}{2}
\| \mathbf{x}_i- \mathbf{x}_i^k\|^2 + g_i(\mathbf{x}_i).\vspace{-0.1cm}
\end{equation}
This is essentially the way a new iteration is computed in most {\em sequential} (block-)CDMs for the
solution of (group) Lasso problems and its generalizations. Note that contrary to most existing schemes, our algorithm is {\em parallel}.
\noindent $\bullet$ At another extreme we could just take $P_i(\mathbf{x}_i; \mathbf{x}^k) = F(\mathbf{x}_i, \mathbf{x}_{-i}^k)$. Of course, to
have (P1) satisfied, we must assume that $F(\mathbf{x}_i, \mathbf{x}_{-i}^k)$ is convex.
With this choice, and setting for simplicity $\mathbf{Q}_i(\mathbf{x}^k) =\mathbf I$, $\widehat \mathbf{x}_i(\mathbf{x}^k, \tau_i) $ is given by\vspace{-0.1cm}
\begin{equation}\label{eq:proposal 2}
\underset{\mathbf{x}_{i}\in X_{i}}{\mbox{argmin}\,} F(\mathbf{x}_i, \mathbf{x}^k_{-i}) + \frac{\tau_i}{2}
\| \mathbf{x}_i- \mathbf{x}_i^k\|^2 + g_i(\mathbf{x}_i),
\end{equation}
thus giving rise to a parallel nonlinear Jacobi type method for the constrained minimization of $V(\mathbf{x})$.
\noindent $\bullet$ Between the two ``extreme'' solutions proposed above one can consider ``intermediate'' choices. For example,
If $F(\mathbf{x}_i, \mathbf{x}_{-i}^k)$ is convex, we can take $P_i(\mathbf{x}_i;\mathbf{x}^k)$ as a second order approximation of $F(\mathbf{x}_i, \mathbf{x}_{-i}^k)$, i.e.,
$ P_i(\mathbf{x}_i; \mathbf{x}^k) = F(\mathbf{x}^k) + \nabla_{\mathbf{x}_i}F(\mathbf{x}^k)^T$ $(\mathbf{x}_i - \mathbf{x}_i^k) + \frac{1}{2} (\mathbf{x}_i - \mathbf{x}_i^k)^T \nabla^2_{\mathbf{x}_i\mathbf{x}_i} F(\mathbf{x}^k) (\mathbf{x}_i - \mathbf{x}_i^k) $. When $g_i(\mathbf{x}_i) \equiv 0$, this essentially corresponds to taking a Newton step in minimizing the ``reduced'' problem $\min_{\mathbf{x}_i\in X_i}F(\mathbf{x}_i, \mathbf{x}_{-i}^k)$. The resulting $\widehat \mathbf{x}_i(\mathbf{x}^k, \tau_i)$ is
\begin{align*}
&\underset{\mathbf{x}_{i}\in X_{i}}{\mbox{argmin}\,} F( \mathbf{x}^k) +\, \nabla_{\mathbf{x}_i}F( \mathbf{x}^k)^T(\mathbf{x}_i - \mathbf{x}_i^k) +
\\&\;
+\, \frac{1}{2} (\mathbf{x}_i - \mathbf{x}_i^k)^T \nabla^2_{\mathbf{x}_i\mathbf{x}_i} F( \mathbf{x}^k) (\mathbf{x}_i - \mathbf{x}_i^k) +
\frac{\tau_i}{2}\| \mathbf{x}_i- \mathbf{x}_i^k\|^2 + g_i(\mathbf{x}_i).
\end{align*}\vspace{-0.4cm}
The framework described in Algorithm \ref{alg:general} can give rise
to very different instances, according to the choices one makes for the many
variable features it contains, some of which have been detailed above. For lack of space, we cannot fully discuss here all possibilities.
We provide next just a few instances of possible algorithms that fall in our framework; more examples can be found in \cite{FacchineiSagratellaScutariMPsub13}.
\noindent {\bf Example \#1$-$(Proximal) Jacobi algorithms for
convex functions: }
Consider the simplest problem falling in our setting: the unconstrained minimization of a continuously differentiable convex function, i.e.,
assume that $F$ is convex, $G(\mathbf{x}) \equiv 0$, and $X= \Re^n$. Although this is possibly the best studied problem in nonlinear optimization, classical
parallel methods for this problem \cite[Sec. 3.2.4]{Bertsekas_Book-Parallel-Comp} require very strong contraction conditions. In our framework we can take $S^k = {\cal N}$,
$P_i(\mathbf{x}_i;\mathbf{x}^k) = F(\mathbf{x}_i, \mathbf{x}_{-i}^k)$, resulting in a fully parallel Jacobi-type method which does not need any additional assumptions. Furthermore our theory shows that
we can even dispense with the convexity assumption and still get convergence of a Jacobi-type method to a stationary point.
\noindent {\bf Example \# 2$-$Parallel coordinate descent method for Lasso}
\noindent
Consider the Lasso problem, i.e., $ F(\mathbf{x}) = \|\mathbf{A}\mathbf{x} -\mathbf{b}\|^2 $, $G(\mathbf{x})= c \|\mathbf{x}\|_1$, and $X=\Re^n$. Probably, to date, the most succesful class of methods for this problem is
that of CDMs, whereby at each iteration a single variable is updated using \eqref{eq:proposal 1}.
We can easily obtain a parallel version for this method by taking $n_i =1$, $S^k = {\cal N}$ and
still using \eqref{eq:proposal 1}. Alternatively, instead of linearizing $F(\mathbf{x})$, we can better exploit the convexity of $F(\mathbf{x})$ and use (\ref{eq:proposal 2}). Furthermore, we can easily consider similar methods for the group Lasso problem (just take $n_i >1$).
As a final remark, we observe that convergence conditions of existing (deterministic) fully distributed parallel versions of CDMs such as
\cite{bradley2011parallel,yin2013parallel} impose a constraint on the maximum number of variables that can be simultaneously updated (linked to the spectral radius of some matrices), a constraint that in many large scale problems is likely not satisfied. A key feature of the proposed scheme is that we can parallelize over (possibly) all variables while guaranteeing convergence. \vspace{-0.3cm}
\section{Numerical Results{\small \vspace{-0.2cm}}}
In this section we report some preliminary numerical results that not only show viability of our approach, but also seem to
indicate that our algorithmic framework can lead to practical methods that exploit well parallelism and compare favorably to existing schemes, both
parallel and sequential.
The tests were carried out on Lasso problems, the most studied case of Problem (\ref{eq:problem 1}) and, arguably, the most important one.
We generate four instances of problems using the random generation technique proposed by Nesterov in \cite{nesterov2012gradient}, that permits to control the sparsity of
the solution. For the first three groups, we considered problems with 10,000 variables with the matrix $\mathbf{A}$ having 2,000 rows. The three groups differ in
the number of non zeros of the solution, which is 20\% (low sparsity), 10\% (medium sparsity), and 5\% (high sparsity) respectively. The last group is
an instance with 100,000 variables and 5000 rows, and solutions having 5\% of non zero variables (high sparsity).
We implemented the instance of Algorithm 1 that we described in Example \# 2 in the previous section, with the only difference that we used
\eqref{eq:proposal 2} instead of the proximal-linear choice \eqref{eq:proposal 1}.
Note that in the case of Lasso problems, the unique solution \eqref{eq:proposal 2} can be computed
in closed form using the soft-thresholding operator, see e.g. \cite{beck_teboulle_jis2009}. The free parameters of the algorithm are chosen as follows.
The proximal parameters are initially set to $\displaystyle \tau_i
= \text{tr}(\mathbf{A}^T\mathbf{A})/2n$ for all $i$, where $n$ is the total number of variables. Furthermore, we allowed a finite number of possible changes to $\tau_i$ according to the following rules:
(i) all $\tau_i$ are doubled if at a certain iteration the objective function does not decrease; and (ii) they are all halved if the objective function decreases
for ten consecutive iterations. We updated $\gamma^k$ according to \eqref{eq:gamma} with $\gamma^0 = 0.9$ and $\theta = 1e-5$. Note that since $\tau_i$ are changed only a finite number of times, conditions of Theorem \ref{Theorem_convergence_inexact_Jacobi} are satisfied, and thus this instance of Algorithm 1 is guaranteed to converge.
Finally we choose not to update all variables but set $E_i(\mathbf{x}^k) =\|\widehat \mathbf{x}_i (\mathbf{x}^k,\tau_i) - \mathbf{x}_i^k\|$ and $\rho = 0.5$ in Algorithm \ref{alg:general}.
We compared our algorithm above, termed FPA (for Flexible Parallel Algorithm), with a parallel implementation of FISTA \cite{beck_teboulle_jis2009}, that can be regarded as the benchmark algorithm for Lasso problems, and Grock, a parallel algorithm proposed in \cite{yin2013parallel} that seems to perform extremely well on sparse problems. We actually tested two instances of Grock; in the first only one variable is updated at each iteration while in the second the number of updated variables is equal to the number of
parallel processors used (16 for the first three set of test problems, 32 for the last). Note that the theoretical convergence properties of Grock are in jeopardy as the number of updated variables increases and theoretical convergence conditions for this method are likely to hold only if the columns of $\mathbf A$ are ``almost'' orthogonal, a feature enjoyed by our test problems, which however is not satisfied in most applications. As benchmark, we also implemented two classical sequential schemes: (i) a Gauss-Seidel (GS) method computing $\hat \mathbf{x}_i$, and
then updating $\mathbf{x}_i$ using unitary step-size, in a sequential fashion, and (ii) a classical Alternating Method of Multipliers (ADMM) \cite{boyd2011distributed} in the form of \cite{luo2012linear}. Note that ADMM can be parallelized, but they are known not to scale well and therefore we did not consider this possibility.
All codes have been written in C++ and use the Message Passing Interface for parallel operations. All algebra is performed by using the GNU Scientific Library
(GSL). The algorithms were tested on a cluster at the State University of New York at Buffalo. All computations were done on one 32 core node composed of four
8 core CPUs with 96GB of RAM and Infiniband card. The 10,000 variables problems were solved using
$16$ parallel processes while for the 100,000 variables problems $32$ parallel processes were used. GS and ADMM were always run on a
single process. Results of our experiments are reported in Fig. \ref{fig}. The curves are averaged over ten random realizations for each of the 10,000 variables groups, while for large 100,000 variables problems the average is over 3 realizations.
\begin{figure}[h]
\vspace{-0.2cm}
\centering
\begin{subfigure}[b]{0.25\textwidth}
\includegraphics[width=\textwidth]{low_spar_error_time2.eps}
\caption{}
\label{fig:low}
\end{subfigure}
\hspace{-0.5cm}
\begin{subfigure}[b]{0.25\textwidth}
\hspace{0.3cm}
\includegraphics[width=\textwidth]{med_spar_error_time2.eps}
\caption{}
\label{fig:med}
\end{subfigure}
\begin{subfigure}[b]{0.25\textwidth}
\includegraphics[width=\textwidth]{high_spar_error_time2.eps}
\caption{}
\label{fig:high}
\end{subfigure}
\hspace{-0.5cm}
\begin{subfigure}[b]{0.25\textwidth}
\hspace{0.3cm}
\includegraphics[width=\textwidth]{big2_error_time2.eps}
\caption{}
\label{fig:big}
\end{subfigure}\vspace{-0.4cm}
\caption{Relative error vs. time (in seconds, logarithmic scale):
(a) medium size and low sparsity - (b) medium size and sparsity - (c) medium size and high sparsity - (d) large size and high sparsity \label{fig}}\vspace{-0.3cm}
\end{figure}
Note that in Fig.\ref{fig} the CPU time includes communication times (for distributed algorithms) and the initial time needed by the methods to perform all pre-iterations computations (this explains why the plot of FISTA starts after the others; in fact FISTA requires some nontrivial initializations based on the computation of $\|\mathbf{A}\|_2^2$).
Some comments are in order. Fig \ref{fig} shows that on the tested problems FPA outperforms in a consistent manner all other implemented
algorithms.
Sequential methods behave strikingly well on the 10,000 variables problems, if one keeps in mind that they only use one process; however, as expected,
they cannot compete with parallel methods when the dimensions increase. FISTA is capable to approach relatively fast low accuracy solutions, but has difficulties in
reaching high accuracies. The parallel version of Grock is the closest match to FPA, but only when the problems are very sparse and the dimensions not too large.
This is consistent
with the fact that at each iteration Grock updates only a very limited number of variables, and also with the fact that its convergence properties are at stake
when the problems are quite dense. Our experiments also suggest that, differently from what one could think (and often claimed in similar situations when using gradient-like methods), updating only a (suitably chosen) subset of blocks rather than all variables may lead to faster algorithms.
In conclusion, we believe the results overall indicate that our approach can lead to very efficient practical methods for the solution of
large problems, with the flexibility to adapt to many different problem characteristics.
\section{Conclusions
We proposed a highly parallelizable algorithmic scheme for the minimization of the sum of a differentiable function and a block-separable nonsmooth one.
Our framework easily allows us to analyze parallel versions of well-known sequential methods and leads to entirely new algorithms.
When applied to large-scale Lasso problems, our algorithm was shown to outperform
existing methods.
\newpage
\section{Appendix: Proof of Theorem 1}
We first introduce some preliminary results instrumental to prove the theorem. Hereafter, for notational simplicity, we will omit the dependence of $\widehat{\mathbf{x}}(\mathbf y,\boldsymbol{\tau})$ on $\boldsymbol\tau$ and write $\widehat{\mathbf{x}}(\mathbf y)$. Given $S\subseteq \mathcal N$ and $\mathbf{x}\triangleq (x_i)_{i=1}^N$, we will also denote by $(\mathbf x)_S$ (or interchangeably $\mathbf x_S$) the vector whose component $i$ is equal to $x_i$ if $i\in S$, and zero otherwise.
\subsection{Intermediate results}
\begin{lemma}\label{Lemma_f_x_y_properties} Set $H(\mathbf{x}; \mathbf{y}) \triangleq \sum_i h_i (\mathbf{x}_i; \mathbf{y})$. Then, the following hold:
\noindent (i) $H(\mathbf{\bullet};\mathbf{y})$
is uniformly strongly convex on $X$ with constant $c_{\boldsymbol{{\tau}}}>0$,
i.e., \vspace{-0.2cm}
\begin{equation}
\begin{array}{l}
\left(\mathbf{x}-\mathbf{w}\right)^{T}\left(\nabla_{\mathbf{x}}H\left(\mathbf{x};\mathbf{y}\right)-\nabla_{\mathbf{x}}H\left(\mathbf{w};\mathbf{y}\right)\right)\geq c_{{\boldsymbol{\tau}}}\left\Vert \mathbf{x}-\mathbf{w}\right\Vert ^{2}\end{array},\label{eq:strong_cvx_f_tilde}
\end{equation}
for all $\mathbf{x},\mathbf{w}\in X$ and given $\mathbf{y}\in X$;
\noindent (ii) $\nabla_{\mathbf{x}}H(\mathbf{x};\mathbf{\bullet})$
is uniformly Lipschitz continuous on $X$, i.e., there exists
a $0<L_{\nabla_H}<\infty$ independent on $\mathbf{x}$ such
that
\begin{equation}
\begin{array}{l}
\left\Vert \nabla_{\mathbf{x}}H\left(\mathbf{x};\mathbf{y}\right)-\nabla_{\mathbf{x}}H\left(\mathbf{x};\mathbf{w}\right)\right\Vert \end{array}\leq\, L_{\nabla H}\,\left\Vert \mathbf{y}-\mathbf{w}\right\Vert ,\label{eq:Lip_grad_L_f}
\end{equation}
for all $\mathbf{y},\mathbf{w}\in X$ and given $\mathbf{x}\in X$. \vspace{-0.3cm}\end{lemma}
\noindent \textbf{Proof.} The proof is standard and thus is omitted.
\begin{proposition}\label{Prop_x_y} Consider Problem (\ref{eq:problem 1})
under (A1)-(A6).
Then the mapping $X\ni\mathbf{y}\mapsto\widehat{\mathbf{x}}(\mathbf{y})$
has the following properties:
\noindent (a)\emph{ $\widehat{\mathbf{x}} (\mathbf{\bullet})$}
is Lipschitz continuous on\emph{ $X$, }i.e., there
exists a positive constant $\hat{{L}}$ such that\emph{
\begin{equation}
\left\Vert \widehat{\mathbf{x}}(\mathbf{y})-\widehat{\mathbf{x}}(\mathbf{z})\right\Vert \leq\,\hat{{L}}\,\left\Vert \mathbf{y}-\mathbf{z}\right\Vert ,\quad\forall\mathbf{y},\mathbf{z}\in X;\vspace{-0.3cm}\label{eq:Lipt_x_map}
\end{equation}
}
\noindent(b) the set of the fixed-points of\emph{ }$\widehat{\mathbf{x}}(\mathbf{\bullet})$\emph{
}coincides with the set of stationary solutions of Problem
(\ref{eq:problem 1}); therefore $\widehat{\mathbf{x}}(\mathbf{y})$ has a fixed-point;
\noindent(c) for every given $\mathbf{y}\in X$ and for any set $S \subseteq {\cal N}$, it holds that
\begin{align}
\left(\widehat{\mathbf{x}}(\mathbf{y})-\mathbf{y}\right)_S^{T}\,\nabla_{\mathbf{x}}F(\mathbf{y})_S +
& \sum_{i\in S} g_i(\widehat{\mathbf{x}}_i (\mathbf{y})) - \sum_{i\in S} g_i(\mathbf{y}_i)\label{eq:descent_direction}
\\ &\leq-c_\tau\,\left\Vert (\widehat{\mathbf{x}}(\mathbf{y})-\mathbf{y})_S\right\Vert ^{2}, \nonumber \vspace{-0.3cm}
\end{align}
with $c_\tau \triangleq q\, \min_i \tau_i$.
\end{proposition}
\noindent \textbf{Proof.} We prove the proposition in the following order: (c), (a), (b).
\noindent (c): Given $\mathbf{y}\in X$, by definition,
each $\widehat{\mathbf{x}}_{i}(\mathbf{y})$ is the unique solution
of problem (\ref{eq:decoupled_problem_i}); then it is not difficult to see that the following holds:
for all $\mathbf{z}_{i}\in X_{i}$,\vspace{-0.2cm}
\begin{equation}\label{eq:VI_i}
\left( \mathbf{z}_i-\widehat{\mathbf{x}}_i(\mathbf{y}) \right)^T
\nabla_{\mathbf{x}_i}h_i(\widehat{\mathbf{x}}_{i}(\mathbf{y}); \mathbf{y}) + g_i(\mathbf{z}_i) - g_i(\widehat{\mathbf{x}}_{i}(\mathbf{y}))
\geq 0.
\end{equation}
Summing and subtracting $\nabla_{\mathbf{x}_{i}}P_i\left(\mathbf{y}_{i};\,\mathbf{y}\right)$
in (\ref{eq:VI_i}), choosing $\mathbf{z}_{i}=\mathbf{y}_{i}$, and
using (P2),
we get
\begin{equation}
\begin{array}[t]{l}
\left(\mathbf{y}_{i}-\widehat{\mathbf{x}}_{i}(\mathbf{y})\right)^{T}\left(\nabla_{\mathbf{x}_{i}}P_{i}\!\left(\widehat{\mathbf{x}}_{i}(\mathbf{y});
\,\mathbf{y}\right)-\nabla_{\mathbf{x}_{i}}P_i\!\left(\mathbf{y}_{i};\,\mathbf{y}\right)\right)\smallskip\\
\quad+\left(\mathbf{y}_{i}-\widehat{\mathbf{x}}_{i}(\mathbf{y})\right)^{T}\nabla_{\mathbf{x}_{i}}F(\mathbf{y})\smallskip
+ g_i(\mathbf{y}_i) - g_i(\widehat{\mathbf{x}}_{i}(\mathbf{y}))
\\
\quad-\tau_{i}\,(\widehat{\mathbf{x}}_{i}(\mathbf{y})-\mathbf{y}_{i})^{T}\,\mathbf{Q}_{i}(\mathbf{y})\,(\widehat{\mathbf{x}}_{i}(\mathbf{y})-\mathbf{y}_{i})\geq0,
\end{array}\label{eq:VI_i_row2}
\end{equation}
for all $i\in\mathcal{N}$. Using (\ref{eq:VI_i_row2}) and observing that the term on the first line of
\eqref{eq:VI_i_row2} is non positive by (P1),
we obtain \vspace{-0.1cm}
\begin{equation}
\left(\mathbf{y}_{i}-\widehat{\mathbf{x}}_{i}(\mathbf{y})\right)^{T}\nabla_{\mathbf{x}_{i}}F(\mathbf{y})
+ g_i(\mathbf{y}_i) - g_i(\widehat{\mathbf{x}}_{i}(\mathbf{y}))
\geq c_{\boldsymbol{{\tau}}}\left\Vert \widehat{\mathbf{x}}_{i}(\mathbf{y})-\mathbf{y}_{i}\right\Vert ^{2},\label{eq:VI_i_row4}
\end{equation}
for all $i\in\mathcal{N}$. Summing (\ref{eq:VI_i_row4}) over $i\in S$
we obtain (\ref{eq:descent_direction}).
\noindent(a): We use the notation introduced in Lemma \ref{Lemma_f_x_y_properties}.
Given $\mathbf{y},\mathbf{z}\in X$, by optimality and using \eqref{eq:VI_i}, we have
\[ \begin{array}{l}
\left(\mathbf{v}-\widehat{\mathbf{x}}(\mathbf{y})\right)^{T}\nabla_{\mathbf{x}}H\left(\widehat{\mathbf{x}}(\mathbf{y});\mathbf{y}\right)
+G(\mathbf{v}) - G(\widehat{\mathbf{x}}(\mathbf{y}))
\geq 0\;\forall\mathbf{v}\in X \\
\left(\mathbf{w}-\widehat{\mathbf{x}}(\mathbf{z})\right)^{T}\nabla_{\mathbf{x}}H\left(\widehat{\mathbf{x}}(\mathbf{z});\mathbf{z}\right)
+ G(\mathbf{w}) - G(\widehat{\mathbf{x}}(\mathbf{z})) \geq 0\;\forall\mathbf{w}\in X.
\end{array}
\]
Setting $\mathbf{v}=\widehat{\mathbf{x}}(\mathbf{z})$ and $\mathbf{w}=\widehat{\mathbf{x}}(\mathbf{y})$,
summing the two inequalities above, and adding and subtracting $\nabla_{\mathbf{x}}H\left(\widehat{\mathbf{x}}(\mathbf{y});\mathbf{z}\right)$,
we obtain:
\begin{equation}
\begin{array}{l}
\left(\widehat{\mathbf{x}}(\mathbf{z})-\widehat{\mathbf{x}}(\mathbf{y})\right)^{T}\left(\nabla_{\mathbf{x}}H\left(\widehat{\mathbf{x}}(\mathbf{z});
\mathbf{z}\right)-\nabla_{\mathbf{x}}H\left(\widehat{\mathbf{x}}(\mathbf{y});\mathbf{z}\right)\right)\\
\leq\left(\widehat{\mathbf{x}}(\mathbf{y})-\widehat{\mathbf{x}}(\mathbf{z})\right)^{T}\left(\nabla_{\mathbf{x}}H\left(\widehat{\mathbf{x}}(\mathbf{y});
\mathbf{z}\right)-\nabla_{\mathbf{x}}H\left(\widehat{\mathbf{x}}(\mathbf{y});\mathbf{y}\right)\right).
\end{array}\label{eq:minimum_principle_Lip}
\end{equation}
Using (\ref{eq:strong_cvx_f_tilde}) we can now lower bound the left-hand-side
of (\ref{eq:minimum_principle_Lip}) as
\begin{equation}
\begin{array}{l}
\left(\widehat{\mathbf{x}}(\mathbf{z})-\widehat{\mathbf{x}}(\mathbf{y})\right)^{T}\left(\nabla_{\mathbf{x}}H\left(\widehat{\mathbf{x}}(\mathbf{z});\mathbf{z}\right)-
\nabla_{\mathbf{x}}H\left(\widehat{\mathbf{x}}(\mathbf{y});\mathbf{z}\right)\right)\\
\,\,\geq c_{{\boldsymbol{\tau}}}\left\Vert \widehat{\mathbf{x}}(\mathbf{z})-\widehat{\mathbf{x}}(\mathbf{y})\right\Vert ^{2},
\end{array}\label{eq:lipschtz_map_2}
\end{equation}
whereas the right-hand side of (\ref{eq:minimum_principle_Lip}) can
be upper bounded as
\begin{equation}
\begin{array}{l}
\left(\widehat{\mathbf{x}}(\mathbf{y})-\widehat{\mathbf{x}}(\mathbf{z})\right)^{T}\left(\nabla_{\mathbf{x}}H\left(\widehat{\mathbf{x}}(\mathbf{y});\mathbf{z}\right)
-\nabla_{\mathbf{x}}H\left(\widehat{\mathbf{x}}(\mathbf{y});\mathbf{y}\right)\right)\\
\,\,\leq\, L_{\nabla H}\,\left\Vert \widehat{\mathbf{x}}(\mathbf{y})-\widehat{\mathbf{x}}(\mathbf{z})\right\Vert \,\left\Vert \mathbf{y}-\mathbf{z}\right\Vert ,
\end{array}\label{eq:lipschtz_map_1}
\end{equation}
where the inequality follows from the Cauchy-Schwartz inequality and
(\ref{eq:Lip_grad_L_f}). Combining (\ref{eq:minimum_principle_Lip}),
(\ref{eq:lipschtz_map_2}), and (\ref{eq:lipschtz_map_1}), we obtain
the desired Lipschitz property of $\widehat{\mathbf{x}}(\bullet)$.
\noindent (b): Let $\mathbf{x}^{\star}\in X$ be a fixed
point of $\widehat{\mathbf{x}}(\mathbf{y})$, that is $\mathbf{x}^{\star}=\widehat{\mathbf{x}}(\mathbf{x}^{\star})$.
Each $\widehat{\mathbf{x}}_{i}(\mathbf{y})$ satisfies (\ref{eq:VI_i}) for any given $\mathbf{y}\in X$.
For some $\boldsymbol\xi_i \in \partial g_i(\mathbf{x}^*)$, setting
$\mathbf{y}=\mathbf{x}^{\star}$ and using $\mathbf{x}^{\star}=\widehat{\mathbf{x}}(\mathbf{x}^{\star})$ and the convexity of $g_i$,
(\ref{eq:VI_i}) reduces to
\begin{equation}
\left(\mathbf{z}_{i}-\mathbf{x}_{i}^{\star}\right)^{T}(\nabla_{\mathbf{x}_{i}}F(\mathbf{x}^{\star}) + \boldsymbol\xi_i) \geq 0,\label{eq:fixed_point_min_principle}
\end{equation}
for all $\mathbf{z}_{i}\in X_{i}$ and $i\in\mathcal{N}$.
Taking into account the Cartesian structure of $X$, the separability of $G$, and
summing (\ref{eq:fixed_point_min_principle}) over $i\in\mathcal{N}$
we obtain $\begin{array}[t]{l}
\left(\mathbf{z}-\mathbf{x}^{\star}\right)^{T}(\nabla_{\mathbf{x}}F(\mathbf{x}^{\star}) + \boldsymbol\xi) \geq0,\end{array}$ for all $\mathbf{z}\in X,$ with $\mathbf{z}\triangleq(\mathbf{z}_{i})_{i=1}^{N}$ and $\boldsymbol\xi \triangleq(\boldsymbol\xi_i)_{i=1}^{N} \in \partial G(\mathbf{x}^*)$;
therefore $\mathbf{x}^{\star}$ is a stationary solution of (\ref{eq:problem 1}).
The converse holds because i) $\widehat{\mathbf{x}}(\mathbf{x}^{\star})$
is the unique optimal solution of (\ref{eq:decoupled_problem_i})
with $\mathbf{y}=\mathbf{x}^{\star}$, and ii) $\mathbf{x}^{\star}$
is also an optimal solution of (\ref{eq:decoupled_problem_i}), since
it satisfies the minimum principle.\hfill $\square$
\begin{lemma} \emph{\cite[Lemma 3.4, p.121]{Bertsekas-Tsitsiklis_bookNeuro11}}\label{lemma_Robbinson_Siegmunt}
Let $\{X^{k}\}$, $\{Y^{k}\}$, and $\{Z^{k}\}$ be three sequences
of numbers such that $Y^{k}\geq0$ for all $k$. Suppose that
\[
X^{k+1}\leq X^{k}-Y^{k}+Z^{k},\quad\forall k=0,1,\ldots
\]
and $\sum_{k=0}^\infty Z^{k}<\infty$. Then either $X^{k}\rightarrow-\infty$
or else $\{X^{k}\}$ converges to a finite value and $\sum_{k=0}^\infty Y^{k}<\infty$.
\hfill $\Box$\end{lemma}
\begin{lemma}\label{lemma on errors} Let $\{\mathbf x^k\}$ be the sequence generated by Algorithm 1. Then, there is a positive constant $\tilde c$ such that the following holds: for all $k\geq 1$,
\begin{align*}
\left (\nabla_{\mathbf{x}}F (\mathbf{\mathbf{x}}^{k})\right)^{T}_{\tiny {S^k}} \left(\widehat{x}(\mathbf{x}^k)-\mathbf{x}^k\right)_{\tiny {S^k}}
+\sum_{i \in S^k} g_i(\widehat{\mathbf{x}}_i(\mathbf{x}^k)) -\sum_{i \in S^k} g_i(\mathbf{x}_i^k)\\ \hfill
\leq
-\tilde c \,\| \widehat{\mathbf{x}}(\mathbf{x}^k)-\mathbf{x}^k \|^2.
\end{align*}\vspace{-0.4cm}\end{lemma}
\noindent \textbf{Proof.} Let $j_k$ be an index in $S^k$ such that $E_{j_k}(\mathbf{x}^k) \geq \rho \max_i E_i(\mathbf{x}^k)$ (Step 3 of the algorithm). Then, using the aforementioned bound and (\ref{eq:error bound}),
it is easy to check that the following chain of inequalities holds:
\begin{align*}
\bar s_{j_k} \| \widehat{\mathbf{x}}_{S^k}(\mathbf{x}^k) - \mathbf{x}^k_{S^k}\| & \geq \bar s_{j_k} \| \widehat{\mathbf{x}}_{j_k}(\mathbf{x}^k) - \mathbf{x}^k_{j_k}\| \\ & \geq E_{j_k}(\mathbf{x}^k) \\ &
\geq \rho \max_i E_i(\mathbf{x}^k) \\ & \geq \left( \rho \min_i \underbar s_i\right) \left( \max_i \{ \|\widehat{\mathbf{x}}_i(\mathbf{x}^k) - \mathbf{x}^k_i\| \} \right) \\ &
\geq \left( \frac{\rho \min_i \underbar s_i}{N}\right) \|\widehat{\mathbf{x}}(\mathbf{x}^k) - \mathbf{x}^k\|.
\end{align*}
Hence we have for any $k$,
\begin{equation}\label{lower_bound_error_S_k}
\| \widehat{\mathbf{x}}_{S^k}(\mathbf{x}^k) - \mathbf{x}^k_{S^k}\| \geq \left(\frac{\rho \min_i \underbar s_i}{N \bar s_{j_k}}\right) \|\widehat{\mathbf{x}}(\mathbf{x}^k) - \mathbf{x}^k\|.
\end{equation}
Invoking now \ref{Prop_x_y} (c) with $S=S^k$ and $\mathbf{y}=\mathbf{x}^k$, and using \eqref{lower_bound_error_S_k}, the lemma holds, with $\tilde c \triangleq c_\tau \left(\frac{\rho \min_i \underbar s_i}{N \max_j \bar s_j}\right)^2$. \hfill $\Box$
\subsection{Proof of Theorem \ref{Theorem_convergence_inexact_Jacobi}} We are now ready to prove the theorem.
For any given $k\geq 0$, the Descent Lemma \cite{Bertsekas_NLPbook99}
yields
{\color{black}
\begin{equation}
\begin{array}{lll}
F\left(\mathbf{x}^{k+1}\right) & \leq & F\left(\mathbf{x}^{k}\right)+\gamma^{k}\,\nabla_{\mathbf{x}}F\left(\mathbf{x}^{k}\right)^{T}\left(\widehat{\mathbf{z}}^{k}-\mathbf{x}^{k}\right)\smallskip\\
& & +\dfrac{\left(\gamma^{k}\right)^{2}{L_{\nabla F}}}{2}\,\left\Vert \widehat{\mathbf{z}}^{k}-\mathbf{x}^{k}\right\Vert ^{2},
\end{array}\label{eq:descent_Lemma}
\end{equation}
with $\widehat{\mathbf{z}}^{k}\triangleq(\widehat{\mathbf{z}}_{i}^{k})_{i=1}^{N}$ and
$\mathbf{z}^{k}\triangleq(\mathbf{z}_{i}^{k})_{i=1}^{N}$ defined in Step 3 and 4 (Algorithm \ref{alg:general}).
Observe that
\begin{equation}\label{eq:49bis} \begin{array}{rcl}
\left\Vert \widehat{\mathbf{z}}^{k}-\mathbf{x}^{k}\right\Vert ^{2}&\leq&
\left\Vert \mathbf{z}^{k}-\mathbf{x}^{k}\right\Vert ^{2}\\
& \leq &
2\left\Vert \widehat{\mathbf{x}}(\mathbf{x}^{k})-\mathbf{x}^{k}\right\Vert ^{2}+2\sum_{i\in \cal N}\left\Vert \mathbf{z}_{i}^{k}-\widehat{\mathbf{x}}_{i}(\mathbf{x}^{k})\right\Vert ^{2}\\
& \leq& 2\left\Vert \widehat{\mathbf{x}}(\mathbf{x}^{k})-\mathbf{x}^{k}\right\Vert ^{2}+2\sum_{i \in \cal N}(\varepsilon_{i}^{k})^{2},
\end{array}
\end{equation}
where the first inequality follows from the definition of $ \mathbf{z}^{k}$ and $ \widehat{\mathbf{z}}^{k}$ and
in the last inequality we used $\left\Vert \mathbf{z}_{i}^{k}-\widehat{\mathbf{x}}_{i}(\mathbf{x}^{k})\right\Vert \leq\varepsilon_{i}^{k}$.
Denoting by $\overline{S}^k$ the complement of $S$, we also have, for $k$ large enough,
\begin{equation}
\begin{array}{l}
\nabla_{\mathbf{x}}F\left(\mathbf{x}^{k}\right)^{T}\left(\widehat{\mathbf{z}}^{k}-\mathbf{x}^{k}\right)\smallskip\\
\qquad\qquad =
\nabla_{\mathbf{x}}F\left(\mathbf{x}^{k}\right)^{T}\left(\widehat{\mathbf{z}}^{k}-\widehat{\mathbf{x}}(\mathbf{x}^{k})
+\widehat{\mathbf{x}}(\mathbf{x}^{k})-\mathbf{x}^{k}\right)\smallskip\\
\qquad\qquad = \nabla_{\mathbf{x}}F\left(\mathbf{x}^{k}\right)^{T}_{S^k} (\mathbf{z}^k - \widehat{\mathbf{x}}(\mathbf{x}^k))_{S^k}\smallskip\\ \qquad\qquad\quad
+
\nabla_{\mathbf{x}}F\left(\mathbf{x}^{k}\right)^{T}_{\overline{S}^k} (\mathbf{x}^k - \widehat{\mathbf{x}}(\mathbf{x}^k))_{\overline{S}^k}\smallskip\\
\qquad \qquad \quad + \nabla_{\mathbf{x}}F\left(\mathbf{x}^{k}\right)^{T}_{S^k} (\widehat{\mathbf{x}}(\mathbf{x}^k)-\mathbf{x}^k)_{S^k}\smallskip\\ \qquad\qquad\quad
+ \nabla_{\mathbf{x}}F\left(\mathbf{x}^{k}\right)^{T}_{\overline{S}^k} (\widehat{\mathbf{x}}(\mathbf{x}^k)-\mathbf{x}^k)_{\overline{S}^k}\smallskip\\
\qquad\qquad = \nabla_{\mathbf{x}}F\left(\mathbf{x}^{k}\right)^{T}_{S^k} (\mathbf{z}^k - \widehat{\mathbf{x}}(\mathbf{x}^k))_{S^k} \smallskip\\ \qquad\qquad +
\nabla_{\mathbf{x}}F\left(\mathbf{x}^{k}\right)^{T}_{S^k} (\widehat{\mathbf{x}}(\mathbf{x}^k)-\mathbf{x}^k)_{S^k},
\end{array}
\end{equation}
where in the second equality we used the definition of $\widehat{\mathbf{z}}^k$ and of the set $S^k$.
Now, using the above identity and
Lemma \ref{lemma on errors}, we can write
\begin{equation}\label{eq:descent_at_x_n}
\begin{array}{l}
\nabla_{\mathbf{x}}F\left(\mathbf{x}^{k}\right)^{T}\left(\widehat{\mathbf{z}}^{k}-\mathbf{x}^{k}\right) + \sum_{i\in S^k} g_i(\widehat{\mathbf{z}}_i^{k})-\sum_{i\in S^k} g_i(\mathbf{x}^{k}_i) \\[0.3em]
= \nabla_{\mathbf{x}}F\left(\mathbf{x}^{k}\right)^{T}\left(\widehat{\mathbf{z}}^{k}-\mathbf{x}^{k}\right) +
\sum_{i\in S^k} g_i(\widehat{\mathbf{x}}_i(\mathbf{x}^{k})) -\sum_{i\in S^k} g_i(\mathbf{x}^{k}_i) \\[0.3em]
\qquad \qquad +
\sum_{i \in S^k} g_i(\widehat{\mathbf{z}}_i^k) -\sum_{i \in S^k} g_i(\widehat{\mathbf{x}}_i(\mathbf{x}^{k}))
\\[0.3em]
\leq -\tilde c\left\Vert \widehat{\mathbf{x}}(\mathbf{x}^{k})-\mathbf{x}^{k}\right
\Vert ^{2}+\sum_{i \in S^k}\varepsilon_{i}^{k}\left\Vert \nabla_{\mathbf{x}_{i}}F(\mathbf{x}^{k})\right\Vert
+ L_G\sum_{i \in S^k}\varepsilon_{i}^{k}.
\end{array}
\end{equation}
Finally, from the definition of
$\widehat{\mathbf{z}}^k$ and of the set $S^k$, we have for all $k$ large enough,
\begin{equation}
\begin{array}{l}
V(\mathbf{x}^{k+1}) = F(\mathbf{x}^{k+1}) + \sum_{ {i \in \cal N}}g_i (\mathbf{x}_i^{k+1}) \\[0.3em]
= F(\mathbf{x}^{k+1}) + \sum_{i\in {\cal N}}g_i (\mathbf{x}_i^{k} + \gamma^k(\widehat{\mathbf{z}}^k_i - \mathbf{x}_i^{k} ))\\[0.3em]
\leq F(\mathbf{x}^{k+1}) + \sum_{i \in \cal N}g_i (\mathbf{x}^k_i) + \gamma^k \left(\sum_{i \in S^k} (g_i (\widehat{\mathbf{z}}^k_i)
- g_i (\mathbf{x}^k_i)) \right)\\[0.3em]
\leq
V\left(\mathbf{x}^{k}\right)-\gamma^{k}\left(\tilde c -\gamma^{k}{L_{\nabla U}}\right)\left\Vert \widehat{\mathbf{x}}(\mathbf{x}^{k})-\mathbf{x}^{k}\right\Vert ^{2}+T^{k},\end{array}\label{eq:descent_Lemma_2}
\end{equation}
where in the first inequality we used the the convexity of the $g_i$'s, whereas the second follows from
\eqref{eq:descent_Lemma}, \eqref{eq:49bis} and \eqref{eq:descent_at_x_n}, with
$$T^{k}\triangleq\gamma^{k}\,
\sum_{i \in S^k}\varepsilon_{i}^{k}\left( L_G +
\left\Vert \nabla_{\mathbf{x}_{i}}F(\mathbf{x}^{k})\right\Vert\right) +\left(\gamma^{k}\right)^{2}{L_{\nabla F}}\,\sum_{i\in \cal N}(\varepsilon_{i}^{k})^{2}.$$
Using assumption (v),
we can bound $T^k$ as
\[
T^{k}\leq
(\gamma^k)^2 \left[ N \alpha_1 (\alpha_2 L_G +1) + (\gamma^k)^2 L_{\nabla F} \left(N\alpha_1\alpha_2\right)^2 \right],
\]
which, by assumption (iv) implies
$\sum_{k=0}^{\infty}T^{k}<\infty$.
Since $\gamma^{k}\rightarrow 0$, it follows from (\ref{eq:descent_Lemma_2}) that there exist some positive constant
$\beta_{1}$ and a sufficiently large $k$, say $k\geq\bar{{k}}$, such that
\begin{equation}
V(\mathbf{x}^{k+1})\leq V(\mathbf{x}^{k})-\gamma^{k}\beta_{1}\left\Vert \widehat{\mathbf{x}}(\mathbf{x}^{k})-\mathbf{x}^{k}\right\Vert ^{2}+T^{k}.\label{eq:descent_Lemma_3_}
\end{equation}
Invoking Lemma \ref{lemma_Robbinson_Siegmunt} with the identifications
$X^{k}=V(\mathbf{x}^{k+1})$, $Y^{k}=\gamma^{k}\beta_{1}\left\Vert \widehat{\mathbf{x}}(\mathbf{x}^{k})-\mathbf{x}^{k}\right\Vert ^{2}$
and $Z^{k}=T^{k}$ while using $\sum_{k=0}^\infty T^{k}<\infty$, we deduce
from (\ref{eq:descent_Lemma_3_}) that either $\{V(\mathbf{x}^{k})\}\rightarrow-\infty$
or else $\{V(\mathbf{x}^{k})\}$ converges to a finite
value and\vspace{-0.1cm}
\begin{equation}
\lim_{k\rightarrow\infty}\sum_{t=\bar{{k}}}^{k}\gamma^{t}\left\Vert \widehat{\mathbf{x}}(\mathbf{x}^{t})-\mathbf{x}^{t}\right\Vert ^{2}<+\infty.\vspace{-0.1cm}\label{eq:finite_sum_series}
\end{equation}
Since $V$ is coercive, $V(\mathbf{x})\geq\min_{\mathbf{y}\in X}V(\mathbf{y})>-\infty$,
implying that $\{V\left(\mathbf{x}^{k}\right)\}$ is convergent;
it follows from (\ref{eq:finite_sum_series}) and $\sum_{k=0}^{\infty}\gamma^{k}=\infty$
that $\liminf_{k\rightarrow\infty}\left\Vert \widehat{\mathbf{x}}(\mathbf{x}^{k})-\mathbf{x}^{k}\right\Vert =0.$
Using Prop. \ref{Prop_x_y}, we show next that $\lim_{k\rightarrow\infty}\left\Vert \widehat{\mathbf{x}}(\mathbf{x}^{k})-\mathbf{x}^{k}\right\Vert =0$;
for notational simplicity we will write $\triangle\widehat{\mathbf{x}}(\mathbf{x}^{k})\triangleq\widehat{\mathbf{x}}(\mathbf{x}^{k})-\mathbf{x}^{k}$.
Suppose, by contradiction, that $\limsup_{k\rightarrow\infty}\left\Vert \triangle\widehat{\mathbf{x}}(\mathbf{x}^{k})\right\Vert >0$.
Then, there exists a $\delta>0$ such that $\left\Vert \triangle\widehat{\mathbf{x}}(\mathbf{x}^{k})\right\Vert >2\delta$
for infinitely many $k$ and also $\left\Vert \triangle\widehat{\mathbf{x}}(\mathbf{x}^{k})\right\Vert <\delta$
for infinitely many $k$. Therefore, one can always find an infinite
set of indexes, say $\mathcal{K}$, having the following properties:
for any $k\in\mathcal{K}$, there exists an integer $i_{k}>k$ such
that
\begin{eqnarray}
\left\Vert \triangle\widehat{\mathbf{x}}(\mathbf{x}^{k})\right\Vert <\delta, & & \left\Vert \triangle\widehat{\mathbf{x}}(\mathbf{x}^{i_{k}})\right\Vert >2\delta\medskip\label{eq:outside_interval}\\
\delta\leq\left\Vert \triangle\widehat{\mathbf{x}}(\mathbf{x}^{j})\right\Vert \leq2\delta & & k<j<i_{k}.\label{eq:inside_interval}
\end{eqnarray}
Given the above bounds, the following holds: for all $k\in\mathcal{K}$,
\begin{eqnarray}
\delta & \overset{(a)}{<} & \left\Vert \triangle\widehat{\mathbf{x}}(\mathbf{x}^{i_{k}})\right\Vert -\left\Vert \triangle\widehat{\mathbf{x}}(\mathbf{x}^{k})\right\Vert \medskip\nonumber \\
& \leq & \left\Vert \widehat{\mathbf{x}}(\mathbf{x}^{i_{k}})-\widehat{\mathbf{x}}(\mathbf{x}^{k})\right\Vert +\left\Vert \mathbf{x}^{i_{k}}-\mathbf{x}^{k}\right\Vert \\
& \overset{(b)}{\leq} & (1+\hat{{L}})\left\Vert \mathbf{x}^{i_{k}}-\mathbf{x}^{k}\right\Vert \\
& \overset{(c)}{\leq} & {\color{black} (1+\hat{{L}})\sum_{t=k}^{i_{k}-1}\gamma^{t}\left(\left\Vert \triangle\widehat{\mathbf{x}}(\mathbf{x}^{t})_{S^t}\right\Vert +\left\Vert (\mathbf{z}^{t}-\widehat{\mathbf{x}}(\mathbf{x}^{t}))_{S^t}\right\Vert \right)}\vspace{-0.3cm}\nonumber \\
& \overset{(d)}{\leq} & (1+\hat{{L}})\,(2\delta+\varepsilon^{\max})\sum_{t=k}^{i_{k}-1}\gamma^{t},\label{eq:lower_bound_sum}
\end{eqnarray}
where (a) follows from (\ref{eq:outside_interval});
(b) is due to Prop. \ref{Prop_x_y}(a); (c) comes from the triangle
inequality, the updating rule of the algorithm {\color{black} and the definition of $\widehat{\mathbf{z}}^k$}; and in (d) we used
(\ref{eq:outside_interval}), (\ref{eq:inside_interval}), and $\left\Vert \mathbf{z}^{t}-\widehat{\mathbf{x}}(\mathbf{x}^{t})\right\Vert \leq\sum_{i\in \cal N}\varepsilon_{i}^{t}$,
where $\varepsilon^{\max}\triangleq\max_{k}\sum_{i\in\cal N}\varepsilon_{i}^{k}<\infty$.
It follows from (\ref{eq:lower_bound_sum}) that
\begin{equation}
\liminf_{k\rightarrow\infty}\sum_{t=k}^{i_{k}-1}\gamma^{t}\geq\dfrac{{\delta}}{(1+\hat{{L}})(2\delta+\varepsilon^{\max})}>0.\label{eq:lim_inf_bound}
\end{equation}
\noindent
We show next that (\ref{eq:lim_inf_bound}) is in contradiction with
the convergence of $\{V(\mathbf{x}^{k})\}$. To do that, we preliminary
prove that, for sufficiently large $k\in\mathcal{K}$, it must be
$\left\Vert \triangle\widehat{\mathbf{x}}(\mathbf{x}^{k})\right\Vert \geq\delta/2$.
Proceeding as in (\ref{eq:lower_bound_sum}), we have: for any given
$k\in\mathcal{K}$,
\[
\begin{array}{l}
\left\Vert \triangle\widehat{\mathbf{x}}(\mathbf{x}^{k+1})\right\Vert -\left\Vert \triangle\widehat{\mathbf{x}}(\mathbf{x}^{k})\right\Vert \leq(1+\hat{{L}})\left\Vert \mathbf{x}^{k+1}-\mathbf{x}^{k}\right\Vert \smallskip\\
\qquad\qquad\qquad\qquad\qquad\leq(1+\hat{{L}})\gamma^{k}\left(\left\Vert \triangle\widehat{\mathbf{x}}(\mathbf{x}^{k})\right\Vert +\varepsilon^{\max}\right).
\end{array}
\]
It turns out that for sufficiently large $k\in\mathcal{K}$ so that
$(1+\hat{{L}})\gamma^{k}<\delta/(\delta+2\varepsilon^{\max})$, it
must be
\begin{equation}
\left\Vert \triangle\widehat{\mathbf{x}}(\mathbf{x}^{k})\right\Vert \geq\delta/2;\label{eq:lower_bound_delta_x_n}
\end{equation}
otherwise the condition $\left\Vert \triangle\widehat{\mathbf{x}}(\mathbf{x}^{k+1})\right\Vert \geq\delta$
would be violated {[}cf. (\ref{eq:inside_interval}){]}. Hereafter
we assume w.l.o.g. that (\ref{eq:lower_bound_delta_x_n}) holds for
all $k\in\mathcal{K}$ (in fact, one can alway restrict $\{\mathbf{x}^{k}\}_{k\in\mathcal{K}}$
to a proper subsequence).
We can show now that (\ref{eq:lim_inf_bound}) is in contradiction
with the convergence of $\{V(\mathbf{x}^{k})\}$. Using (\ref{eq:descent_Lemma_3_})
(possibly over a subsequence), we have: for sufficiently large $k\in\mathcal{K}$,
\begin{eqnarray}
V(\mathbf{x}^{i_{k}}) & \leq & V(\mathbf{x}^{k})-\beta_{2}\sum_{t=k}^{i_{k}-1}\gamma^{t}\left\Vert \triangle\widehat{\mathbf{x}}(\mathbf{x}^{t})\right\Vert ^{2}+\sum_{t=k}^{i_{k}-1}T^{t}\nonumber \\
& \overset{(a)}{<} & V(\mathbf{x}^{k})-\beta_{2}(\delta^{2}/4)\sum_{t=k}^{i_{k}-1}\gamma^{t}+\sum_{t=k}^{i_{k}-1}T^{t}\label{eq:liminf_zero}
\end{eqnarray}
where in (a) we used (\ref{eq:inside_interval}) and (\ref{eq:lower_bound_delta_x_n}),
and $\beta_{2}$ is some positive constant. Since $\{V(\mathbf{x}^{k})\}$
converges and $\sum_{k=0}^{\infty}T^{k}<\infty$, (\ref{eq:liminf_zero})
implies $\lim_{\mathcal{K}\ni k\rightarrow\infty}\,\sum_{t=k}^{i_{k}-1}\gamma^{t}=0,$
which contradicts (\ref{eq:lim_inf_bound}).
Finally, since the sequence $\{\mathbf{x}^{k}\}$ is bounded {[}due
to the coercivity of $V$ and the convergence of $\{V(\mathbf{x}^{k})\}${]},
it has at least one limit point $\bar{{\mathbf{x}}}$ that must belong
to $X$. By the continuity of $\widehat{\mathbf{x}}(\bullet)$
{[}Prop. \ref{Prop_x_y}(a){]} and $\lim_{k\rightarrow\infty}\left\Vert \widehat{\mathbf{x}}(\mathbf{x}^{k})-\mathbf{x}^{k}\right\Vert =0$,
it must be $\widehat{\mathbf{x}}(\bar{{\mathbf{x}}})=\bar{{\mathbf{x}}}$.
By Prop. \ref{Prop_x_y}(b) $\bar{{\mathbf{x}}}$ is also a stationary
solution of Problem (\ref{eq:problem 1}).
As a final remark, note that if $\varepsilon_i^k=0$ for every $i$ and for every $k$ large enough, i.e. if eventually
$\widehat{\mathbf{x}}(\mathbf{x}^k)$ is computed exactly, there is no need to assume that $G$ is
globally Lipschitz. In fact in \eqref{eq:descent_at_x_n} the term containing $L_G$ disappears, and
actually all the terms $T^k$ are zero and all the subsequent derivations independent of the Lipschitzianity of $G$.
\hfill$\square$
\begin{spacing}{0.050000000000000003}
\bibliographystyle{IEEEtran}
\section{Introduction\vspace{-0.1cm}}
\label{sec:Introduction}
The minimization of the sum of a smooth function, $F$, and of a nonsmooth
(separable) convex one, $G$,
\begin{equation}\label{eq:problem 1}
\min_{\mathbf{x}\in X} V(\mathbf{x}) \triangleq F(\mathbf{x}) + G(\mathbf{x}),
\end{equation}
is an ubiquitous problem that arises in many fields of
engineering, so diverse as compressed sensing, basis pursuit denoising,
sensor networks, neuroelectromagnetic imaging, machine learning,
data mining, sparse logistic regression, genomics, metereology, tensor factorization and completion,
geophysics, and radio astronomy.
Usually the nonsmooth term is used to promote sparsity of the optimal solution,
that often corresponds to a parsimonious representation of some phenomenon at hand.
Many of the mentioned applications
can give rise to extremely large problems so that standard optimization techniques are hardly applicable.
And indeed, recent years have witnessed a flurry of research activity aimed at developing
solution methods that are simple (for example based solely on matrix/vector multiplications) but yet capable to
converge
to a good approximate solution in reasonable time. It is hardly possible here to even summarize the huge amount
of work done in this field; we refer the reader to the recent works
\cite{bach2011optimization,bradley2011parallel,buhlmann2011statistics,byrd2013inexact,fountoulakis2013second,necoara2013efficient,nesterov2012gradient,
nesterov2012efficiency,qin2010efficient,rakotomamonjy2011surveying,razaviyayn2013unified,richtarik2012iteration,
richtarik2012parallel,Sra-Nowozin-Wright_book11,tseng2009coordinate,xu2012block,yin2013parallel,
yuan2010comparison,wright2012accelerated}
as entry points to the literature.
It is clear however that if one wants to solve really large problems,
parallel methods exploiting the computational power of multi-core processors have to be employed.
It is then surprising that while serial solutions methods for Problem \eqref{eq:problem 1} have been widely investigated,
the analysis of parallel algorithms suitable to large-scale implementations lags behind.
Gradient-type methods can of course be easily parallelized. However, beyond that,
we are only aware of very few papers, all very recent,
that deal with parallel solution methods \cite{bradley2011parallel,necoara2013efficient, richtarik2012parallel,yin2013parallel}.
These papers analyze both randomized and deterministic block Coordinate Descent Methods (CDMs) that, essentially,
are still (regularized) gradient-based methods. One advantage of the analyses in \cite{bradley2011parallel,necoara2013efficient, richtarik2012parallel,yin2013parallel}
is that they provide an interesting (global) rate of convergence. On the other hand they apply only to convex problems and
are not flexible enough to include, among other things, very natural Jacobi-type methods (where at each iteration
a partial minimization of the original function is performed with respect to a block variable while all other variables are kept fixed) and the possibility to deal with a nonconvex $F$.
In this paper, building on the approach proposed in
\cite{scutari_facchinei_et_al_icassp13,scutari_facchinei_et_al_tsp13}, we present a broad, deterministic algorithmic framework for the solution of Problem \eqref{eq:problem 1} with the following novel features:
i) it is parallel, with a degree of parallelism that can be chosen by the user and that can go from a complete parallelism (each variable is updated
in parallel to all the others) to the sequential (one variable only is updated at each iteration); ii) it can tackle a nonconvex $F$; iii)
it is very flexible and includes, among others, updates based on gradient- or Newton-type methods; and iv) it easily allows for inexact
solutions.
Our framework allows us to define different schemes, \emph{all converging under the same conditions}, that can accommodate different problem features and algorithmic requirements.
Even in the most studied case in which $F$ is convex and $G(\mathbf{x}) \equiv 0$ our results compare favourably to existing ones and the numerical results
show our
approach to be very promising.
\vspace{-0.1cm}
\section{Problem Definition\vspace{-0.1cm}}
We consider Problem \eqref{eq:problem 1},
where the feasible set $X=X_1\times \cdots \times X_N$ is a cartesian product of lower dimensional convex sets $X_i\subseteq \Re^{n_i}$, and $\mathbf x\in \Re^n$ is partitioned accordingly to $\mathbf x = (\mathbf x_1, \ldots, \mathbf x_N)$, with each $\mathbf x_i \in \Re^{n_i}$. $F$ is smooth (and not necessarily convex)
and $G$ is convex and possibly nondifferentiable, with $ G(\mathbf{x}) = {\scriptstyle \sum_{i=1}^N} g_i(\mathbf{x}_i)$ with $\mathbf{x}_i \in X_i$.
This format is very general and includes problems of great interest.
Below we list some instances of Problem \eqref{eq:problem 1}.
\noindent $\bullet$
$G(\mathbf{x})=0$; in this case the problem reduces to the minimization of a smooth, possibly nonconvex problem with convex constraints.
\noindent $\bullet$
$ F(\mathbf{x}) = \|\mathbf{A}\mathbf{x} -\mathbf{b}\|^2 $ and $G(\mathbf{x})= c \|\mathbf{x}\|_1$, $X=\Re^n$, with $\mathbf{A}\in\Re^{m \times n}$, $\mathbf{b} \in \Re^m$, and $c\in \Re_{++}$ given constants; this is the very famous and much studied Lasso problem \cite{tibshirani1996regression}.
\noindent $\bullet$
$ F(\mathbf{x}) = \|\mathbf{A}\mathbf{x} -\mathbf{b}\|^2 $ and $G(\mathbf{x})= c \sum_{i=1}^N\|\mathbf{x}_i\|_2$, $X=\Re^n$, with $\mathbf{A}\in\Re^{m \times n}$, $\mathbf{b} \in \Re^m$, and $c\in \Re_{++}$
given constants; this is the group Lasso problem \cite{yuan2006model}.
\noindent $\bullet$
$ F(\mathbf{x}) = \sum_{j=1}^m \log(1 + e^{-a_i \mathbf{y}_i^T\mathbf{x} }) $ and $G(\mathbf{x})= c \|\mathbf{x}\|_1$ (or $G(\mathbf{x})= c \sum_{i=1}^N\|\mathbf{x}_i\|_2$), with $\mathbf{y}_i\in \Re^n$, $a_i\in \Re$, and $c\in \Re_{++}$ given
constants; this is the sparse logistic regression problem \cite{shevade2003simple,meier2008group}.
\noindent $\bullet$
$F(\mathbf{x}) = \sum_{j=1}^m\max\{0, 1- a_i \mathbf{y}_i^T\mathbf{x}\}^2$ and $G(\mathbf{x})= c \|\mathbf{x}\|_1$, with $a_i\in \{-1,1\}$, $\mathbf{y}_i\in \Re^n$, and $c\in \Re_{++}$ given; this is
the $\ell_1$-regularized
$\ell_2$-loss Support Vector Machine problem, see e.g. \cite{yuan2010comparison}.
\noindent $\bullet$ Other problems that can be cast in the form \eqref{eq:problem 1} include the Nuclear Norm Minimization problem, the Robust Principal Component Analysis problem, the Sparse Inverse Covariance Selection problem, the Nonnegative Matrix (or Tensor) Factorization problem, see e.g. \cite{xu2012block,goldfarb2012fast} and references therein.
\smallskip
\noindent
Given \eqref{eq:problem 1}, we make the following standard, blanket assumptions:
\begin{description}[topsep=-2.0pt,itemsep=-2.0pt]
\item[\rm (A1)] Each $X_i$ is nonempty, closed, and convex;
\item[\rm (A2)] $F$ is $C^1$ on an open set containing $X$;
\item[\rm (A3)] $\nabla F$ is Lipschitz continuous
on $X$ with constant $L_{F}$;
\item[\rm (A4)] $G(\mathbf{x}) = \sum_{i=i}^N g_i(\mathbf{x}_i)$, with all $g_i$ continuous and convex on $X_i$;
\item[\rm (A5)] $V$ is coercive.
\end{description}
\vspace{-0.2cm}
\section{Main Results\label{sec:Main Results}\vspace{-0.2cm}}
We want to develop {\em parallel} solution methods for Problem \eqref{eq:problem 1} whereby operations can be carried out on some or (possibly) all (block) variables $\mathbf{x}_i$ at
the \emph{same} time. The most natural parallel (Jacobi-type) method one can think of is updating \emph{all} blocks simultaneously: given $\mathbf{x}^k$, each (block) variable $\mathbf{x}^{k+1}_i$ is computed as the solution of
$\min_{\mathbf{x}_i}$ $[F(\mathbf{x}_i, \mathbf{x}_{-i}^k) + g_i(\mathbf{x}_i)] $ (where $\mathbf{x}_{-i}$ denotes the vector obtained from $\mathbf{x}$ by deleting the block $\mathbf{x}_i$).
Unfortunately this method converges only under very restrictive conditions \cite{Bertsekas_Book-Parallel-Comp} that are seldom verified in practice. To cope with this issue we introduce some ``memory"
and set the new point to be a convex combination of $\mathbf{x}^k$ and the solutions of $\min_{\mathbf{x}_i} [F(\mathbf{x}_i, \mathbf{x}^k_{-i}) + g_i(\mathbf{x}_i)] $.
However our framework has many additional features, as discussed next.
\noindent \textbf{Approximating $F$}: Solving each $\min_{\mathbf{x}_i} [F(\mathbf{x}_i, \mathbf{x}^k_{-i}) + g_i(\mathbf{x}_i)] $ may be too costly or difficult in
some
situations. One may then
prefer to approximate this problem, in some suitable sense, in order to facilitate the task of computing the new iteration.
To this end,
we assume that for all $i\in {\cal N} \triangleq \{1, \dots, N\}$ we
can define a function $P_{i} (\mathbf{z};\mathbf{w}) : X_i \times X \to \Re$ having the following properties (we denote by $\nabla P_{i}$
the partial gradient of $P_i$ with respect to $\mathbf{z}$):
\begin{description}[topsep=-1.0pt,itemsep=-2.0pt]
\item[\rm (P1)]
$P_{i} (\mathbf{\bullet}; \mathbf{w})$ is convex and continuously differentiable
on $ X_i$ for all $\mathbf{w}\in X$;
\item[\rm (P2)] $\nabla P_{i} (\mathbf{x}_i;\mathbf{x}) = \nabla_{{\mathbf{x}}_i} F(\mathbf{x})$ for all $\mathbf{x} \in X$;
\item[\rm (P3)] $\nabla P_{i} (\mathbf{z};\mathbf{\bullet})$ is Lipschitz continuous
on $ X$
for all $\mathbf{z} \in X_i$.
\end{description}
\smallskip
Such a function $P_i$ should be regarded as a (simple) convex approximation of $F$ at the point $\mathbf{x}$ with respect to the block of variables $\mathbf{x}_i$, that
preserves the first order properties of $F$
with respect to $\mathbf{x}_i$.
Based on this approximation we can define at any point $\mathbf{x}^k\in X$ a {\em regularized} approximation $\widetilde{h}_i (\mathbf{x}_{i};\mathbf{x}^k)$ of $V$ with respect
to $\mathbf{x}_i$ where $F$ is replaced by $P_i$ while the nondifferentiable term is preserved, and a quadratic
regularization is added to make the overall approximation strongly convex. More formally, we have
\[
\widetilde{h}_i (\mathbf{x}_{i}; \mathbf{x}^k) \! \triangleq \!\underbrace{ P_{i}(\mathbf{x}_{i}; \mathbf{x}^k)\!
+\!\dfrac{{\tau}_{i}}{2} \! \left(\mathbf{x}_{i}- \mathbf{x}_{i}^k\right)^{T}\!\!\!\!\mathbf{Q}_{i}( \mathbf{x}^k)\!\left(\mathbf{x}_{i}- \mathbf{x}^k_{i}\right)}_{\triangleq h_i(\mathbf{x}_{i}; \mathbf{x}^k) }
\!+ g_{i}(\mathbf{x}_{i}),\label{eq:convex_approx_of_fi_on_Ci}\vspace{-0.2cm}
\]
where $\mathbf{Q}_{i}(\mathbf{x}^k)$ is an $n_{i}\times n_{i}$
positive definite matrix (possibly dependent on $\mathbf{x}^k)$, satisfying the following conditions. \smallski
\begin{description}[leftmargin=*,topsep=-2.0pt,itemsep=-2.0pt]
\item[\rm (A6)] All matrices $\mathbf{Q}_{i}(\mathbf{x}^k)$ are uniformly positive definite with a common positive definiteness constant $q > 0$; furthermore, $\mathbf{Q}_{i}(\bullet)$ is Lipschitz continuous on $X$.\smallskip
\end{description}
Note that in most cases (and in all our numerical experiments) the $\mathbf{Q}_{i}$ are constant and equal to the identity matrix, so that (A6) is automatically satisfied.
Associated with each $i$ and point $ \mathbf{x}^k \in X$ we can define the following optimal solution map:\vspace{-0.1cm}
\begin{equation}\label{eq:decoupled_problem_i}
\widehat{\mathbf{x}}_{i}(\mathbf{x}^k,\tau_{i})\triangleq\underset{\mathbf{x}_{i}\in X_{i}}{\mbox{argmin}\,}\tilde{h}_{i}(\mathbf{x}_{i};\mathbf{x}^{k}).\vspace{-0.1cm}
\end{equation}
Note that $\widehat{\mathbf{x}}_{i}(\mathbf{x}^{k},\tau_{i})$
is always well-defined, since the optimization problem in (\ref{eq:decoupled_problem_i})
is strongly convex. Given (\ref{eq:decoupled_problem_i}),
we can then introduce
\[
X\ni\mathbf{y}\mapsto\widehat{\mathbf{x}}(\mathbf{y},\boldsymbol{{\tau}})\triangleq\left(\widehat{\mathbf{x}}_{i}(\mathbf{y},\tau_{i})\right)_{i=1}^{N}.
\]
The algorithm we are about to described is based on the computation of $\widehat{\mathbf{x}}$. Therefore the approximating functions $P_i$ should
lead to as easily computable functions $\widehat{\mathbf{x}}$ as possible. An appropriate choice depends on the problem at hand and on computational
requirements. We discuss some possible choice for the approximations $P_i$ after introducing the main algorithm (Algorithm \ref{alg:general}).
\noindent \textbf{Inexact solutions:} In many situations (especially in the case of large-scale problems), it can be useful to
further reduce the computational effort needed to solve the subproblems in \eqref{eq:decoupled_problem_i} by allowing \emph{inexact} computations $\mathbf{z}^k$ of $\widehat{\mathbf{x}}_{i}(\mathbf{x}^k,\tau_{i})$, i.e., $\|\mathbf{z}_{i}^{k}-\widehat{\mathbf{x}}_i\left(\mathbf{x}^{k},\boldsymbol{{\tau}}\right)\|\leq\varepsilon_{i}^{k}$, where $\varepsilon_{i}^{k}$ measures the accuracy in computing the solution.
\noindent \textbf{Updating only some blocks:} Another important feature of our algorithmic framework is the possibility of updating only some of the variables at each iteration.
Essentially we prove convergence assuming that at each iteration only a subset of
the variables is updated under the condition that this subset contains at least one (block) component
which is within a factor $\rho \in (0,1]$ from being ``sufficiently far'' from optimality, in the sense explained next.
First of all $\mathbf{x}^k_i$ is optimal for $\tilde{h}_{i}(\mathbf{x}_{i};\mathbf{x}^k)$ if and only if
$\widehat{\mathbf{x}}_{i}(\mathbf{x}^k,\tau_{i})=\mathbf{x}^k_i$. Ideally we would like then to select the indices to update according to the optimality measure $\|\widehat{\mathbf{x}}_{i}(\mathbf{x}^k,
\tau_{i})-\mathbf{x}^k_i\|$; but in some situations this could be computationally too expensive. In order to be able to develop alternative choices, based
on the same idea, we suppose one can compute an {\em error bound}, i.e., a function $E_i(\mathbf{x})$ such that\vspace{-0.1cm}
\begin{equation}\label{eq:error bound}
\underbar s_i\|\widehat{\mathbf{x}}_{i}(\mathbf{x}^k,\tau_{i})-\mathbf{x}^k_i\| \le E_i(\mathbf{x}^k) \leq \bar s_i\|\widehat{\mathbf{x}}_{i}(\mathbf{x}^k,\tau_{i})-\mathbf{x}^k_i\|,\vspace{-0.1cm}
\end{equation}
for some $0< \underbar s_i \le \bar s_i$. Of course we can always set $ E_i(\mathbf{x}^k) = \|\widehat{\mathbf{x}}_{i}(\mathbf{x}^k,\tau_{i})-\mathbf{x}^k_i\| $, but other choices are also possible; we discuss
some of them after introducing the algorithm.
We are now ready to formally introduce our algorithm, Algorithm 1, that enjoys all the features discussed above. Its convergence properties are given in Theorem \ref{Theorem_convergence_inexact_Jacobi}, whose proof is omitted because of space limitation, see \cite{FacchineiSagratellaScutariMPsub13}
\vspace{-0.4cm}
\begin{algo}{\textbf{Inexact Parallel Algorithm}} S$\textbf{Data}:$ $\{\varepsilon_{i}^{k}\}$
for $i\in\mathcal{N}$, $\boldsymbol{{\tau}}\geq\mathbf{0}$, $\{\gamma^{k}\}>0$,
$\mathbf{x}^{0}\in X$, $\rho \in (0,1]$.
\hspace{1.24cm}Set $k=0$.
\texttt{$\mbox{(S.1)}:$}$\,\,$If $\mathbf{x}^{k}$ satisfies a termination
criterion: STOP;
\texttt{$\mbox{(S.2)}:$} For all $i\in\mathcal{N}$, solve (\ref{eq:decoupled_problem_i})
with accuracy $\varepsilon_{i}^{k}:$
\hspace{1.24cm}Find $\mathbf{z}_{i}^{k}\in X_i$ s.t. $\|\mathbf{z}_{i}^{k}-\widehat{\mathbf{x}}_i\left(\mathbf{x}^{k},\boldsymbol{{\tau}}\right)\|\leq\varepsilon_{i}^{k}$;
\texttt{$\mbox{(S.3)}:$} Set $M^k \triangleq \max_i \{E_i(\mathbf{x}^k)\}$.
\hspace{1.24cm}Choose a set $S^k$ that contains at least one index $i$
\hspace{1.24cm}for which
$E_i(\mathbf{x}^k) \geq \rho M^k.$
\hspace{1.24cm}Set $\widehat{\mathbf{z}}^k_i = \mathbf{\mathbf{z}}_{i}^k$ for $i\in S^k$ and
$\widehat{\mathbf{z}}^k_i = \mathbf{x}^k_i$ for $i\not \in S^k$
\texttt{$\mbox{(S.4)}:$} Set $\mathbf{x}^{k+1}\triangleq\mathbf{x}^k+\gamma^{k}\,(\widehat{\mathbf{z}}^{k}-\mathbf{x}^{k})$;
\texttt{$\mbox{(S.5)}:$} $k\leftarrow k+1$, and go to \texttt{$\mbox{(S.1)}.$}
\label{alg:general}
\end{algo}\vspace{-0.4cm}
\begin{theorem} \label{Theorem_convergence_inexact_Jacobi}Let
$\{\mathbf{x}^{k}\}$ be the sequence generated by
Algorithm \ref{alg:general}, under A1-A6.
Suppose that $\{\gamma^{k}\}$
and $\{\varepsilon_{i}^{k}\}$ satisfy the following conditions: i)
$\gamma^{k}\in(0,1]$; ii) $\gamma^{k}\rightarrow0$; iii) $\sum_{k}\gamma^{k}=+\infty$;
iv) $\sum_{k}\left(\gamma^{k}\right)^{2}<+\infty$;
and v) $\varepsilon_{i}^{k}\leq \gamma^k \alpha_1\min\{\alpha_2, 1/\|\nabla_{\mathbf{x}_i} F(\mathbf{x}^k)\| \}$
for all $i\in {\cal N}$ and some nonnegative constants $\alpha_1$ and $\alpha_2$.
Additionally, if inexact solutions are used in Step S.2, i.e., $\varepsilon_{i}^{k}>0$ for some $i$ and infinite $k$, then
assume also that $G$ is globally Lipschitz on $X$.
Then, either Algorithm \ref{alg:general} converges in a finite number of iterations to a stationary solution
of \eqref{eq:problem 1} or every limit point of
$\{\mathbf{x}^{k}\}$ (at least
one such points exists) is a stationary solution of \eqref{eq:problem 1}.\vspace{-0.1cm}
\end{theorem}
In the theorem we obtain convergence to stationary points $\mathbf{x}^*$, i.e. points for which a subgradient $\xi \in \partial G(\mathbf{x}^*)$ exists such that
$(\nabla F(\mathbf{x}^*) $ $+ \xi)^T(\mathbf{y}-\mathbf{x}^*) \geq 0$ for all $\mathbf{y}\in X$. Of course, if $F$ is convex, stationary points coincide with global minimizers.
\smallskip
\noindent \textbf{On Algorithm 1.} The proposed algorithm is extremely flexible. We can always choose $S^k ={\cal N}$ resulting in the simultaneous update of all the (block) variables (full Jacobi scheme); or, at the other extreme, one can update a single (block) variable per time, thus obtaining a Gauss-Southwell kind of method.
One can also compute inexact solutions (Step 2) while preserving convergence, provided that the error term $\varepsilon_{i}^{k}$ and the step-size $\gamma^{k}$\textquoteright{}s are chosen according
to Theorem 1. We emphasize that the Lipschitzianity of $G$ is required only if $\widehat{\mathbf{x}}(\mathbf{x}^k, \tau)$ is not computed exactly for infinite iterations. At any rate this Lipschitz conditions is automatically satisfied if $G$ is a norm (and therefore in Lasso and group Lasso problems for example) or if
$X$ is bounded.\smallskip
\noindent \textbf{On the choice of the stepsize $\gamma^k$}\emph{.}
An example of step-size rule satisfying i-iv in Theorem \ref{Theorem_convergence_inexact_Jacobi} is:
given $\gamma^{0}=1$, let
\begin{equation}\label{eq:gamma}
\gamma^{k}=\gamma^{k-1}\left(1-\theta\,\gamma^{k-1}\right),\quad k=1,\ldots,\vspace{-0.1cm}
\end{equation}
where $\theta\in(0,1)$ is a given constant; see \cite{scutari_facchinei_et_al_tsp13}
for others rules. This is actually the rule we used in our practical experiments, see next section. Notice that while this rule
may still require some tuning for optimal behaviour, it is quite reliable, since in general we are not
using a (sub)gradient direction, so that many of the well-known practical drawbacks associated
with a (sub)gradient method with diminishing step-size are mitigated in our setting. Furthermore, this choice of step-size does not require any form of centralized coordination, which is a favourable feature in a parallel environment.
We remark that it is possible to
prove convergence of Algorithm 1 also using other step-size rules, such as a standard Armijo-like line-search procedure or a (suitably small) constant step-size; see \cite{FacchineiSagratellaScutariMPsub13} for more details.
We omit the discussion of these options because of lack of space, but also because
the former is not in line with our parallel approach while the latter is numerically less efficient.
\smallskip
\noindent \textbf{On the choice of $E_i(\mathbf{x})$}\emph{.}
\noindent $\bullet$
As we mentioned, the most obvious choice is to take $E_i(\mathbf{x}) = \|\widehat{\mathbf{x}}_{i}(\mathbf{x}^k,\tau_{i})-\mathbf{x}^k_i\|$.
This is a valuable choice if the computation of $\widehat{\mathbf{x}}_{i}(\mathbf{x}^k,\tau_{i})$ can be easily accomplished. For instance, in the
Lasso problem with ${\cal N} = \{1, \ldots, n\}$ (i.e., when each block reduces to a scalar variable),
it is well-known that $\widehat{\mathbf{x}}_{i}(\mathbf{x}^k,\tau_{i})$ can be computed in closed form using
the soft-thresholding operator.
\noindent $\bullet$ In situations where the computation of $\|\widehat{\mathbf{x}}_{i}(\mathbf{x}^k,\tau_{i})-\mathbf{x}^k_i\|$ is not possible or advisable,
we can resort to estimates. To make the discussion simple, assume momentarily that $G\equiv 0$. Then it is known \cite{Facchinei-Pang_FVI03} that
$\|\Pi_{X_i} (\mathbf{x}^k_i - \nabla_{\mathbf{x}_i}F(\mathbf{x}^k)) - \mathbf{x}^k_i\|$ is an error bound for the minimization problem in \eqref{eq:decoupled_problem_i} and therefore
satisfies \eqref{eq:error bound}. In this situation we can choose $E_i(\mathbf{x}^k) = \|\Pi_{X_i} (\mathbf{x}^k_i - \nabla_{\mathbf{x}_i}F(\mathbf{x}^k)) - \mathbf{x}^k_i\|$.
If $G(\mathbf{x}) \not \equiv 0$ we can easily reduce to the case $G\equiv 0$ by a simple transformation; the details are omitted for lack of space, see \cite{scutari_facchinei_et_al_tsp13}.
\noindent $\bullet$ It is interesting to note that the computation of $E_i$ is only needed if a partial update of the (block) variables is
performed. However, an option that is always feasible is to take $S^k = {\cal N}$ at each iteration, i.e., update all (block) variables at each iteration.
With this choice we can dispense with the computation of $E_i$ altogether.\smallskip
\noindent \textbf{On the choice of $P_i(\mathbf{x}_i;\mathbf{x} )$}.
\noindent $\bullet$ The most obvious choice for $P_i$ is the linearization of $F$ at $\mathbf{x}^k$ with respect to $\mathbf{x}_i$:
$ P_i(\mathbf{x}_i; \mathbf{x}^k) = F(\mathbf{x}^k) + \nabla_{\mathbf{x}_i}F(\mathbf{x}^k)^T(\mathbf{x}_i - \mathbf{x}_i^k)$. With this choice, and taking for simplicity $\mathbf{Q}_i(\mathbf{x}^k) = \mathbf I$,
$\widehat {\mathbf x}_i(\mathbf{x}^k, \tau_i) $ is given by\vspace{-0.2cm}
\begin{equation}\label{eq:proposal 1}
\underset{\mathbf{x}_{i}\in X_{i}}{\mbox{argmin}\,} F(\mathbf{x}^k) + \nabla_{\mathbf{x}_i}F(\mathbf{x}^k)^T(\mathbf{x}_i - \mathbf{x}_i^k) + \frac{\tau_i}{2}
\| \mathbf{x}_i- \mathbf{x}_i^k\|^2 + g_i(\mathbf{x}_i).\vspace{-0.1cm}
\end{equation}
This is essentially the way a new iteration is computed in most {\em sequential} (block-)CDMs for the
solution of (group) Lasso problems and its generalizations. Note that contrary to most existing schemes, our algorithm is {\em parallel}.
\noindent $\bullet$ At another extreme we could just take $P_i(\mathbf{x}_i; \mathbf{x}^k) = F(\mathbf{x}_i, \mathbf{x}_{-i}^k)$. Of course, to
have (P1) satisfied, we must assume that $F(\mathbf{x}_i, \mathbf{x}_{-i}^k)$ is convex.
With this choice, and setting for simplicity $\mathbf{Q}_i(\mathbf{x}^k) =\mathbf I$, $\widehat \mathbf{x}_i(\mathbf{x}^k, \tau_i) $ is given by\vspace{-0.1cm}
\begin{equation}\label{eq:proposal 2}
\underset{\mathbf{x}_{i}\in X_{i}}{\mbox{argmin}\,} F(\mathbf{x}_i, \mathbf{x}^k_{-i}) + \frac{\tau_i}{2}
\| \mathbf{x}_i- \mathbf{x}_i^k\|^2 + g_i(\mathbf{x}_i),
\end{equation}
thus giving rise to a parallel nonlinear Jacobi type method for the constrained minimization of $V(\mathbf{x})$.
\noindent $\bullet$ Between the two ``extreme'' solutions proposed above one can consider ``intermediate'' choices. For example,
If $F(\mathbf{x}_i, \mathbf{x}_{-i}^k)$ is convex, we can take $P_i(\mathbf{x}_i;\mathbf{x}^k)$ as a second order approximation of $F(\mathbf{x}_i, \mathbf{x}_{-i}^k)$, i.e.,
$ P_i(\mathbf{x}_i; \mathbf{x}^k) = F(\mathbf{x}^k) + \nabla_{\mathbf{x}_i}F(\mathbf{x}^k)^T$ $(\mathbf{x}_i - \mathbf{x}_i^k) + \frac{1}{2} (\mathbf{x}_i - \mathbf{x}_i^k)^T \nabla^2_{\mathbf{x}_i\mathbf{x}_i} F(\mathbf{x}^k) (\mathbf{x}_i - \mathbf{x}_i^k) $. When $g_i(\mathbf{x}_i) \equiv 0$, this essentially corresponds to taking a Newton step in minimizing the ``reduced'' problem $\min_{\mathbf{x}_i\in X_i}F(\mathbf{x}_i, \mathbf{x}_{-i}^k)$. The resulting $\widehat \mathbf{x}_i(\mathbf{x}^k, \tau_i)$ is
\begin{align*}
&\underset{\mathbf{x}_{i}\in X_{i}}{\mbox{argmin}\,} F( \mathbf{x}^k) +\, \nabla_{\mathbf{x}_i}F( \mathbf{x}^k)^T(\mathbf{x}_i - \mathbf{x}_i^k) +
\\&\;
+\, \frac{1}{2} (\mathbf{x}_i - \mathbf{x}_i^k)^T \nabla^2_{\mathbf{x}_i\mathbf{x}_i} F( \mathbf{x}^k) (\mathbf{x}_i - \mathbf{x}_i^k) +
\frac{\tau_i}{2}\| \mathbf{x}_i- \mathbf{x}_i^k\|^2 + g_i(\mathbf{x}_i).
\end{align*}\vspace{-0.4cm}
The framework described in Algorithm \ref{alg:general} can give rise
to very different instances, according to the choices one makes for the many
variable features it contains, some of which have been detailed above. For lack of space, we cannot fully discuss here all possibilities.
We provide next just a few instances of possible algorithms that fall in our framework; more examples can be found in \cite{FacchineiSagratellaScutariMPsub13}.
\noindent {\bf Example \#1$-$(Proximal) Jacobi algorithms for
convex functions: }
Consider the simplest problem falling in our setting: the unconstrained minimization of a continuously differentiable convex function, i.e.,
assume that $F$ is convex, $G(\mathbf{x}) \equiv 0$, and $X= \Re^n$. Although this is possibly the best studied problem in nonlinear optimization, classical
parallel methods for this problem \cite[Sec. 3.2.4]{Bertsekas_Book-Parallel-Comp} require very strong contraction conditions. In our framework we can take $S^k = {\cal N}$,
$P_i(\mathbf{x}_i;\mathbf{x}^k) = F(\mathbf{x}_i, \mathbf{x}_{-i}^k)$, resulting in a fully parallel Jacobi-type method which does not need any additional assumptions. Furthermore our theory shows that
we can even dispense with the convexity assumption and still get convergence of a Jacobi-type method to a stationary point.
\noindent {\bf Example \# 2$-$Parallel coordinate descent method for Lasso}
\noindent
Consider the Lasso problem, i.e., $ F(\mathbf{x}) = \|\mathbf{A}\mathbf{x} -\mathbf{b}\|^2 $, $G(\mathbf{x})= c \|\mathbf{x}\|_1$, and $X=\Re^n$. Probably, to date, the most succesful class of methods for this problem is
that of CDMs, whereby at each iteration a single variable is updated using \eqref{eq:proposal 1}.
We can easily obtain a parallel version for this method by taking $n_i =1$, $S^k = {\cal N}$ and
still using \eqref{eq:proposal 1}. Alternatively, instead of linearizing $F(\mathbf{x})$, we can better exploit the convexity of $F(\mathbf{x})$ and use (\ref{eq:proposal 2}). Furthermore, we can easily consider similar methods for the group Lasso problem (just take $n_i >1$).
As a final remark, we observe that convergence conditions of existing (deterministic) fully distributed parallel versions of CDMs such as
\cite{bradley2011parallel,yin2013parallel} impose a constraint on the maximum number of variables that can be simultaneously updated (linked to the spectral radius of some matrices), a constraint that in many large scale problems is likely not satisfied. A key feature of the proposed scheme is that we can parallelize over (possibly) all variables while guaranteeing convergence. \vspace{-0.3cm}
\section{Numerical Results{\small \vspace{-0.2cm}}}
In this section we report some preliminary numerical results that not only show viability of our approach, but also seem to
indicate that our algorithmic framework can lead to practical methods that exploit well parallelism and compare favorably to existing schemes, both
parallel and sequential.
The tests were carried out on Lasso problems, the most studied case of Problem (\ref{eq:problem 1}) and, arguably, the most important one.
We generate four instances of problems using the random generation technique proposed by Nesterov in \cite{nesterov2012gradient}, that permits to control the sparsity of
the solution. For the first three groups, we considered problems with 10,000 variables with the matrix $\mathbf{A}$ having 2,000 rows. The three groups differ in
the number of non zeros of the solution, which is 20\% (low sparsity), 10\% (medium sparsity), and 5\% (high sparsity) respectively. The last group is
an instance with 100,000 variables and 5000 rows, and solutions having 5\% of non zero variables (high sparsity).
We implemented the instance of Algorithm 1 that we described in Example \# 2 in the previous section, with the only difference that we used
\eqref{eq:proposal 2} instead of the proximal-linear choice \eqref{eq:proposal 1}.
Note that in the case of Lasso problems, the unique solution \eqref{eq:proposal 2} can be computed
in closed form using the soft-thresholding operator, see e.g. \cite{beck_teboulle_jis2009}. The free parameters of the algorithm are chosen as follows.
The proximal parameters are initially set to $\displaystyle \tau_i
= \text{tr}(\mathbf{A}^T\mathbf{A})/2n$ for all $i$, where $n$ is the total number of variables. Furthermore, we allowed a finite number of possible changes to $\tau_i$ according to the following rules:
(i) all $\tau_i$ are doubled if at a certain iteration the objective function does not decrease; and (ii) they are all halved if the objective function decreases
for ten consecutive iterations. We updated $\gamma^k$ according to \eqref{eq:gamma} with $\gamma^0 = 0.9$ and $\theta = 1e-5$. Note that since $\tau_i$ are changed only a finite number of times, conditions of Theorem \ref{Theorem_convergence_inexact_Jacobi} are satisfied, and thus this instance of Algorithm 1 is guaranteed to converge.
Finally we choose not to update all variables but set $E_i(\mathbf{x}^k) =\|\widehat \mathbf{x}_i (\mathbf{x}^k,\tau_i) - \mathbf{x}_i^k\|$ and $\rho = 0.5$ in Algorithm \ref{alg:general}.
We compared our algorithm above, termed FPA (for Flexible Parallel Algorithm), with a parallel implementation of FISTA \cite{beck_teboulle_jis2009}, that can be regarded as the benchmark algorithm for Lasso problems, and Grock, a parallel algorithm proposed in \cite{yin2013parallel} that seems to perform extremely well on sparse problems. We actually tested two instances of Grock; in the first only one variable is updated at each iteration while in the second the number of updated variables is equal to the number of
parallel processors used (16 for the first three set of test problems, 32 for the last). Note that the theoretical convergence properties of Grock are in jeopardy as the number of updated variables increases and theoretical convergence conditions for this method are likely to hold only if the columns of $\mathbf A$ are ``almost'' orthogonal, a feature enjoyed by our test problems, which however is not satisfied in most applications. As benchmark, we also implemented two classical sequential schemes: (i) a Gauss-Seidel (GS) method computing $\hat \mathbf{x}_i$, and
then updating $\mathbf{x}_i$ using unitary step-size, in a sequential fashion, and (ii) a classical Alternating Method of Multipliers (ADMM) \cite{boyd2011distributed} in the form of \cite{luo2012linear}. Note that ADMM can be parallelized, but they are known not to scale well and therefore we did not consider this possibility.
All codes have been written in C++ and use the Message Passing Interface for parallel operations. All algebra is performed by using the GNU Scientific Library
(GSL). The algorithms were tested on a cluster at the State University of New York at Buffalo. All computations were done on one 32 core node composed of four
8 core CPUs with 96GB of RAM and Infiniband card. The 10,000 variables problems were solved using
$16$ parallel processes while for the 100,000 variables problems $32$ parallel processes were used. GS and ADMM were always run on a
single process. Results of our experiments are reported in Fig. \ref{fig}. The curves are averaged over ten random realizations for each of the 10,000 variables groups, while for large 100,000 variables problems the average is over 3 realizations.
\begin{figure}[h]
\vspace{-0.2cm}
\centering
\begin{subfigure}[b]{0.25\textwidth}
\includegraphics[width=\textwidth]{low_spar_error_time2.eps}
\caption{}
\label{fig:low}
\end{subfigure}
\hspace{-0.5cm}
\begin{subfigure}[b]{0.25\textwidth}
\hspace{0.3cm}
\includegraphics[width=\textwidth]{med_spar_error_time2.eps}
\caption{}
\label{fig:med}
\end{subfigure}
\begin{subfigure}[b]{0.25\textwidth}
\includegraphics[width=\textwidth]{high_spar_error_time2.eps}
\caption{}
\label{fig:high}
\end{subfigure}
\hspace{-0.5cm}
\begin{subfigure}[b]{0.25\textwidth}
\hspace{0.3cm}
\includegraphics[width=\textwidth]{big2_error_time2.eps}
\caption{}
\label{fig:big}
\end{subfigure}\vspace{-0.4cm}
\caption{Relative error vs. time (in seconds, logarithmic scale):
(a) medium size and low sparsity - (b) medium size and sparsity - (c) medium size and high sparsity - (d) large size and high sparsity \label{fig}}\vspace{-0.3cm}
\end{figure}
Note that in Fig.\ref{fig} the CPU time includes communication times (for distributed algorithms) and the initial time needed by the methods to perform all pre-iterations computations (this explains why the plot of FISTA starts after the others; in fact FISTA requires some nontrivial initializations based on the computation of $\|\mathbf{A}\|_2^2$).
Some comments are in order. Fig \ref{fig} shows that on the tested problems FPA outperforms in a consistent manner all other implemented
algorithms.
Sequential methods behave strikingly well on the 10,000 variables problems, if one keeps in mind that they only use one process; however, as expected,
they cannot compete with parallel methods when the dimensions increase. FISTA is capable to approach relatively fast low accuracy solutions, but has difficulties in
reaching high accuracies. The parallel version of Grock is the closest match to FPA, but only when the problems are very sparse and the dimensions not too large.
This is consistent
with the fact that at each iteration Grock updates only a very limited number of variables, and also with the fact that its convergence properties are at stake
when the problems are quite dense. Our experiments also suggest that, differently from what one could think (and often claimed in similar situations when using gradient-like methods), updating only a (suitably chosen) subset of blocks rather than all variables may lead to faster algorithms.
In conclusion, we believe the results overall indicate that our approach can lead to very efficient practical methods for the solution of
large problems, with the flexibility to adapt to many different problem characteristics.
\section{Conclusions
We proposed a highly parallelizable algorithmic scheme for the minimization of the sum of a differentiable function and a block-separable nonsmooth one.
Our framework easily allows us to analyze parallel versions of well-known sequential methods and leads to entirely new algorithms.
When applied to large-scale Lasso problems, our algorithm was shown to outperform
existing methods.
\newpage
\section{Appendix: Proof of Theorem 1}
We first introduce some preliminary results instrumental to prove the theorem. Hereafter, for notational simplicity, we will omit the dependence of $\widehat{\mathbf{x}}(\mathbf y,\boldsymbol{\tau})$ on $\boldsymbol\tau$ and write $\widehat{\mathbf{x}}(\mathbf y)$. Given $S\subseteq \mathcal N$ and $\mathbf{x}\triangleq (x_i)_{i=1}^N$, we will also denote by $(\mathbf x)_S$ (or interchangeably $\mathbf x_S$) the vector whose component $i$ is equal to $x_i$ if $i\in S$, and zero otherwise.
\subsection{Intermediate results}
\begin{lemma}\label{Lemma_f_x_y_properties} Set $H(\mathbf{x}; \mathbf{y}) \triangleq \sum_i h_i (\mathbf{x}_i; \mathbf{y})$. Then, the following hold:
\noindent (i) $H(\mathbf{\bullet};\mathbf{y})$
is uniformly strongly convex on $X$ with constant $c_{\boldsymbol{{\tau}}}>0$,
i.e., \vspace{-0.2cm}
\begin{equation}
\begin{array}{l}
\left(\mathbf{x}-\mathbf{w}\right)^{T}\left(\nabla_{\mathbf{x}}H\left(\mathbf{x};\mathbf{y}\right)-\nabla_{\mathbf{x}}H\left(\mathbf{w};\mathbf{y}\right)\right)\geq c_{{\boldsymbol{\tau}}}\left\Vert \mathbf{x}-\mathbf{w}\right\Vert ^{2}\end{array},\label{eq:strong_cvx_f_tilde}
\end{equation}
for all $\mathbf{x},\mathbf{w}\in X$ and given $\mathbf{y}\in X$;
\noindent (ii) $\nabla_{\mathbf{x}}H(\mathbf{x};\mathbf{\bullet})$
is uniformly Lipschitz continuous on $X$, i.e., there exists
a $0<L_{\nabla_H}<\infty$ independent on $\mathbf{x}$ such
that
\begin{equation}
\begin{array}{l}
\left\Vert \nabla_{\mathbf{x}}H\left(\mathbf{x};\mathbf{y}\right)-\nabla_{\mathbf{x}}H\left(\mathbf{x};\mathbf{w}\right)\right\Vert \end{array}\leq\, L_{\nabla H}\,\left\Vert \mathbf{y}-\mathbf{w}\right\Vert ,\label{eq:Lip_grad_L_f}
\end{equation}
for all $\mathbf{y},\mathbf{w}\in X$ and given $\mathbf{x}\in X$. \vspace{-0.3cm}\end{lemma}
\noindent \textbf{Proof.} The proof is standard and thus is omitted.
\begin{proposition}\label{Prop_x_y} Consider Problem (\ref{eq:problem 1})
under (A1)-(A6).
Then the mapping $X\ni\mathbf{y}\mapsto\widehat{\mathbf{x}}(\mathbf{y})$
has the following properties:
\noindent (a)\emph{ $\widehat{\mathbf{x}} (\mathbf{\bullet})$}
is Lipschitz continuous on\emph{ $X$, }i.e., there
exists a positive constant $\hat{{L}}$ such that\emph{
\begin{equation}
\left\Vert \widehat{\mathbf{x}}(\mathbf{y})-\widehat{\mathbf{x}}(\mathbf{z})\right\Vert \leq\,\hat{{L}}\,\left\Vert \mathbf{y}-\mathbf{z}\right\Vert ,\quad\forall\mathbf{y},\mathbf{z}\in X;\vspace{-0.3cm}\label{eq:Lipt_x_map}
\end{equation}
}
\noindent(b) the set of the fixed-points of\emph{ }$\widehat{\mathbf{x}}(\mathbf{\bullet})$\emph{
}coincides with the set of stationary solutions of Problem
(\ref{eq:problem 1}); therefore $\widehat{\mathbf{x}}(\mathbf{y})$ has a fixed-point;
\noindent(c) for every given $\mathbf{y}\in X$ and for any set $S \subseteq {\cal N}$, it holds that
\begin{align}
\left(\widehat{\mathbf{x}}(\mathbf{y})-\mathbf{y}\right)_S^{T}\,\nabla_{\mathbf{x}}F(\mathbf{y})_S +
& \sum_{i\in S} g_i(\widehat{\mathbf{x}}_i (\mathbf{y})) - \sum_{i\in S} g_i(\mathbf{y}_i)\label{eq:descent_direction}
\\ &\leq-c_\tau\,\left\Vert (\widehat{\mathbf{x}}(\mathbf{y})-\mathbf{y})_S\right\Vert ^{2}, \nonumber \vspace{-0.3cm}
\end{align}
with $c_\tau \triangleq q\, \min_i \tau_i$.
\end{proposition}
\noindent \textbf{Proof.} We prove the proposition in the following order: (c), (a), (b).
\noindent (c): Given $\mathbf{y}\in X$, by definition,
each $\widehat{\mathbf{x}}_{i}(\mathbf{y})$ is the unique solution
of problem (\ref{eq:decoupled_problem_i}); then it is not difficult to see that the following holds:
for all $\mathbf{z}_{i}\in X_{i}$,\vspace{-0.2cm}
\begin{equation}\label{eq:VI_i}
\left( \mathbf{z}_i-\widehat{\mathbf{x}}_i(\mathbf{y}) \right)^T
\nabla_{\mathbf{x}_i}h_i(\widehat{\mathbf{x}}_{i}(\mathbf{y}); \mathbf{y}) + g_i(\mathbf{z}_i) - g_i(\widehat{\mathbf{x}}_{i}(\mathbf{y}))
\geq 0.
\end{equation}
Summing and subtracting $\nabla_{\mathbf{x}_{i}}P_i\left(\mathbf{y}_{i};\,\mathbf{y}\right)$
in (\ref{eq:VI_i}), choosing $\mathbf{z}_{i}=\mathbf{y}_{i}$, and
using (P2),
we get
\begin{equation}
\begin{array}[t]{l}
\left(\mathbf{y}_{i}-\widehat{\mathbf{x}}_{i}(\mathbf{y})\right)^{T}\left(\nabla_{\mathbf{x}_{i}}P_{i}\!\left(\widehat{\mathbf{x}}_{i}(\mathbf{y});
\,\mathbf{y}\right)-\nabla_{\mathbf{x}_{i}}P_i\!\left(\mathbf{y}_{i};\,\mathbf{y}\right)\right)\smallskip\\
\quad+\left(\mathbf{y}_{i}-\widehat{\mathbf{x}}_{i}(\mathbf{y})\right)^{T}\nabla_{\mathbf{x}_{i}}F(\mathbf{y})\smallskip
+ g_i(\mathbf{y}_i) - g_i(\widehat{\mathbf{x}}_{i}(\mathbf{y}))
\\
\quad-\tau_{i}\,(\widehat{\mathbf{x}}_{i}(\mathbf{y})-\mathbf{y}_{i})^{T}\,\mathbf{Q}_{i}(\mathbf{y})\,(\widehat{\mathbf{x}}_{i}(\mathbf{y})-\mathbf{y}_{i})\geq0,
\end{array}\label{eq:VI_i_row2}
\end{equation}
for all $i\in\mathcal{N}$. Using (\ref{eq:VI_i_row2}) and observing that the term on the first line of
\eqref{eq:VI_i_row2} is non positive by (P1),
we obtain \vspace{-0.1cm}
\begin{equation}
\left(\mathbf{y}_{i}-\widehat{\mathbf{x}}_{i}(\mathbf{y})\right)^{T}\nabla_{\mathbf{x}_{i}}F(\mathbf{y})
+ g_i(\mathbf{y}_i) - g_i(\widehat{\mathbf{x}}_{i}(\mathbf{y}))
\geq c_{\boldsymbol{{\tau}}}\left\Vert \widehat{\mathbf{x}}_{i}(\mathbf{y})-\mathbf{y}_{i}\right\Vert ^{2},\label{eq:VI_i_row4}
\end{equation}
for all $i\in\mathcal{N}$. Summing (\ref{eq:VI_i_row4}) over $i\in S$
we obtain (\ref{eq:descent_direction}).
\noindent(a): We use the notation introduced in Lemma \ref{Lemma_f_x_y_properties}.
Given $\mathbf{y},\mathbf{z}\in X$, by optimality and using \eqref{eq:VI_i}, we have
\[ \begin{array}{l}
\left(\mathbf{v}-\widehat{\mathbf{x}}(\mathbf{y})\right)^{T}\nabla_{\mathbf{x}}H\left(\widehat{\mathbf{x}}(\mathbf{y});\mathbf{y}\right)
+G(\mathbf{v}) - G(\widehat{\mathbf{x}}(\mathbf{y}))
\geq 0\;\forall\mathbf{v}\in X \\
\left(\mathbf{w}-\widehat{\mathbf{x}}(\mathbf{z})\right)^{T}\nabla_{\mathbf{x}}H\left(\widehat{\mathbf{x}}(\mathbf{z});\mathbf{z}\right)
+ G(\mathbf{w}) - G(\widehat{\mathbf{x}}(\mathbf{z})) \geq 0\;\forall\mathbf{w}\in X.
\end{array}
\]
Setting $\mathbf{v}=\widehat{\mathbf{x}}(\mathbf{z})$ and $\mathbf{w}=\widehat{\mathbf{x}}(\mathbf{y})$,
summing the two inequalities above, and adding and subtracting $\nabla_{\mathbf{x}}H\left(\widehat{\mathbf{x}}(\mathbf{y});\mathbf{z}\right)$,
we obtain:
\begin{equation}
\begin{array}{l}
\left(\widehat{\mathbf{x}}(\mathbf{z})-\widehat{\mathbf{x}}(\mathbf{y})\right)^{T}\left(\nabla_{\mathbf{x}}H\left(\widehat{\mathbf{x}}(\mathbf{z});
\mathbf{z}\right)-\nabla_{\mathbf{x}}H\left(\widehat{\mathbf{x}}(\mathbf{y});\mathbf{z}\right)\right)\\
\leq\left(\widehat{\mathbf{x}}(\mathbf{y})-\widehat{\mathbf{x}}(\mathbf{z})\right)^{T}\left(\nabla_{\mathbf{x}}H\left(\widehat{\mathbf{x}}(\mathbf{y});
\mathbf{z}\right)-\nabla_{\mathbf{x}}H\left(\widehat{\mathbf{x}}(\mathbf{y});\mathbf{y}\right)\right).
\end{array}\label{eq:minimum_principle_Lip}
\end{equation}
Using (\ref{eq:strong_cvx_f_tilde}) we can now lower bound the left-hand-side
of (\ref{eq:minimum_principle_Lip}) as
\begin{equation}
\begin{array}{l}
\left(\widehat{\mathbf{x}}(\mathbf{z})-\widehat{\mathbf{x}}(\mathbf{y})\right)^{T}\left(\nabla_{\mathbf{x}}H\left(\widehat{\mathbf{x}}(\mathbf{z});\mathbf{z}\right)-
\nabla_{\mathbf{x}}H\left(\widehat{\mathbf{x}}(\mathbf{y});\mathbf{z}\right)\right)\\
\,\,\geq c_{{\boldsymbol{\tau}}}\left\Vert \widehat{\mathbf{x}}(\mathbf{z})-\widehat{\mathbf{x}}(\mathbf{y})\right\Vert ^{2},
\end{array}\label{eq:lipschtz_map_2}
\end{equation}
whereas the right-hand side of (\ref{eq:minimum_principle_Lip}) can
be upper bounded as
\begin{equation}
\begin{array}{l}
\left(\widehat{\mathbf{x}}(\mathbf{y})-\widehat{\mathbf{x}}(\mathbf{z})\right)^{T}\left(\nabla_{\mathbf{x}}H\left(\widehat{\mathbf{x}}(\mathbf{y});\mathbf{z}\right)
-\nabla_{\mathbf{x}}H\left(\widehat{\mathbf{x}}(\mathbf{y});\mathbf{y}\right)\right)\\
\,\,\leq\, L_{\nabla H}\,\left\Vert \widehat{\mathbf{x}}(\mathbf{y})-\widehat{\mathbf{x}}(\mathbf{z})\right\Vert \,\left\Vert \mathbf{y}-\mathbf{z}\right\Vert ,
\end{array}\label{eq:lipschtz_map_1}
\end{equation}
where the inequality follows from the Cauchy-Schwartz inequality and
(\ref{eq:Lip_grad_L_f}). Combining (\ref{eq:minimum_principle_Lip}),
(\ref{eq:lipschtz_map_2}), and (\ref{eq:lipschtz_map_1}), we obtain
the desired Lipschitz property of $\widehat{\mathbf{x}}(\bullet)$.
\noindent (b): Let $\mathbf{x}^{\star}\in X$ be a fixed
point of $\widehat{\mathbf{x}}(\mathbf{y})$, that is $\mathbf{x}^{\star}=\widehat{\mathbf{x}}(\mathbf{x}^{\star})$.
Each $\widehat{\mathbf{x}}_{i}(\mathbf{y})$ satisfies (\ref{eq:VI_i}) for any given $\mathbf{y}\in X$.
For some $\boldsymbol\xi_i \in \partial g_i(\mathbf{x}^*)$, setting
$\mathbf{y}=\mathbf{x}^{\star}$ and using $\mathbf{x}^{\star}=\widehat{\mathbf{x}}(\mathbf{x}^{\star})$ and the convexity of $g_i$,
(\ref{eq:VI_i}) reduces to
\begin{equation}
\left(\mathbf{z}_{i}-\mathbf{x}_{i}^{\star}\right)^{T}(\nabla_{\mathbf{x}_{i}}F(\mathbf{x}^{\star}) + \boldsymbol\xi_i) \geq 0,\label{eq:fixed_point_min_principle}
\end{equation}
for all $\mathbf{z}_{i}\in X_{i}$ and $i\in\mathcal{N}$.
Taking into account the Cartesian structure of $X$, the separability of $G$, and
summing (\ref{eq:fixed_point_min_principle}) over $i\in\mathcal{N}$
we obtain $\begin{array}[t]{l}
\left(\mathbf{z}-\mathbf{x}^{\star}\right)^{T}(\nabla_{\mathbf{x}}F(\mathbf{x}^{\star}) + \boldsymbol\xi) \geq0,\end{array}$ for all $\mathbf{z}\in X,$ with $\mathbf{z}\triangleq(\mathbf{z}_{i})_{i=1}^{N}$ and $\boldsymbol\xi \triangleq(\boldsymbol\xi_i)_{i=1}^{N} \in \partial G(\mathbf{x}^*)$;
therefore $\mathbf{x}^{\star}$ is a stationary solution of (\ref{eq:problem 1}).
The converse holds because i) $\widehat{\mathbf{x}}(\mathbf{x}^{\star})$
is the unique optimal solution of (\ref{eq:decoupled_problem_i})
with $\mathbf{y}=\mathbf{x}^{\star}$, and ii) $\mathbf{x}^{\star}$
is also an optimal solution of (\ref{eq:decoupled_problem_i}), since
it satisfies the minimum principle.\hfill $\square$
\begin{lemma} \emph{\cite[Lemma 3.4, p.121]{Bertsekas-Tsitsiklis_bookNeuro11}}\label{lemma_Robbinson_Siegmunt}
Let $\{X^{k}\}$, $\{Y^{k}\}$, and $\{Z^{k}\}$ be three sequences
of numbers such that $Y^{k}\geq0$ for all $k$. Suppose that
\[
X^{k+1}\leq X^{k}-Y^{k}+Z^{k},\quad\forall k=0,1,\ldots
\]
and $\sum_{k=0}^\infty Z^{k}<\infty$. Then either $X^{k}\rightarrow-\infty$
or else $\{X^{k}\}$ converges to a finite value and $\sum_{k=0}^\infty Y^{k}<\infty$.
\hfill $\Box$\end{lemma}
\begin{lemma}\label{lemma on errors} Let $\{\mathbf x^k\}$ be the sequence generated by Algorithm 1. Then, there is a positive constant $\tilde c$ such that the following holds: for all $k\geq 1$,
\begin{align*}
\left (\nabla_{\mathbf{x}}F (\mathbf{\mathbf{x}}^{k})\right)^{T}_{\tiny {S^k}} \left(\widehat{x}(\mathbf{x}^k)-\mathbf{x}^k\right)_{\tiny {S^k}}
+\sum_{i \in S^k} g_i(\widehat{\mathbf{x}}_i(\mathbf{x}^k)) -\sum_{i \in S^k} g_i(\mathbf{x}_i^k)\\ \hfill
\leq
-\tilde c \,\| \widehat{\mathbf{x}}(\mathbf{x}^k)-\mathbf{x}^k \|^2.
\end{align*}\vspace{-0.4cm}\end{lemma}
\noindent \textbf{Proof.} Let $j_k$ be an index in $S^k$ such that $E_{j_k}(\mathbf{x}^k) \geq \rho \max_i E_i(\mathbf{x}^k)$ (Step 3 of the algorithm). Then, using the aforementioned bound and (\ref{eq:error bound}),
it is easy to check that the following chain of inequalities holds:
\begin{align*}
\bar s_{j_k} \| \widehat{\mathbf{x}}_{S^k}(\mathbf{x}^k) - \mathbf{x}^k_{S^k}\| & \geq \bar s_{j_k} \| \widehat{\mathbf{x}}_{j_k}(\mathbf{x}^k) - \mathbf{x}^k_{j_k}\| \\ & \geq E_{j_k}(\mathbf{x}^k) \\ &
\geq \rho \max_i E_i(\mathbf{x}^k) \\ & \geq \left( \rho \min_i \underbar s_i\right) \left( \max_i \{ \|\widehat{\mathbf{x}}_i(\mathbf{x}^k) - \mathbf{x}^k_i\| \} \right) \\ &
\geq \left( \frac{\rho \min_i \underbar s_i}{N}\right) \|\widehat{\mathbf{x}}(\mathbf{x}^k) - \mathbf{x}^k\|.
\end{align*}
Hence we have for any $k$,
\begin{equation}\label{lower_bound_error_S_k}
\| \widehat{\mathbf{x}}_{S^k}(\mathbf{x}^k) - \mathbf{x}^k_{S^k}\| \geq \left(\frac{\rho \min_i \underbar s_i}{N \bar s_{j_k}}\right) \|\widehat{\mathbf{x}}(\mathbf{x}^k) - \mathbf{x}^k\|.
\end{equation}
Invoking now \ref{Prop_x_y} (c) with $S=S^k$ and $\mathbf{y}=\mathbf{x}^k$, and using \eqref{lower_bound_error_S_k}, the lemma holds, with $\tilde c \triangleq c_\tau \left(\frac{\rho \min_i \underbar s_i}{N \max_j \bar s_j}\right)^2$. \hfill $\Box$
\subsection{Proof of Theorem \ref{Theorem_convergence_inexact_Jacobi}} We are now ready to prove the theorem.
For any given $k\geq 0$, the Descent Lemma \cite{Bertsekas_NLPbook99}
yields
{\color{black}
\begin{equation}
\begin{array}{lll}
F\left(\mathbf{x}^{k+1}\right) & \leq & F\left(\mathbf{x}^{k}\right)+\gamma^{k}\,\nabla_{\mathbf{x}}F\left(\mathbf{x}^{k}\right)^{T}\left(\widehat{\mathbf{z}}^{k}-\mathbf{x}^{k}\right)\smallskip\\
& & +\dfrac{\left(\gamma^{k}\right)^{2}{L_{\nabla F}}}{2}\,\left\Vert \widehat{\mathbf{z}}^{k}-\mathbf{x}^{k}\right\Vert ^{2},
\end{array}\label{eq:descent_Lemma}
\end{equation}
with $\widehat{\mathbf{z}}^{k}\triangleq(\widehat{\mathbf{z}}_{i}^{k})_{i=1}^{N}$ and
$\mathbf{z}^{k}\triangleq(\mathbf{z}_{i}^{k})_{i=1}^{N}$ defined in Step 3 and 4 (Algorithm \ref{alg:general}).
Observe that
\begin{equation}\label{eq:49bis} \begin{array}{rcl}
\left\Vert \widehat{\mathbf{z}}^{k}-\mathbf{x}^{k}\right\Vert ^{2}&\leq&
\left\Vert \mathbf{z}^{k}-\mathbf{x}^{k}\right\Vert ^{2}\\
& \leq &
2\left\Vert \widehat{\mathbf{x}}(\mathbf{x}^{k})-\mathbf{x}^{k}\right\Vert ^{2}+2\sum_{i\in \cal N}\left\Vert \mathbf{z}_{i}^{k}-\widehat{\mathbf{x}}_{i}(\mathbf{x}^{k})\right\Vert ^{2}\\
& \leq& 2\left\Vert \widehat{\mathbf{x}}(\mathbf{x}^{k})-\mathbf{x}^{k}\right\Vert ^{2}+2\sum_{i \in \cal N}(\varepsilon_{i}^{k})^{2},
\end{array}
\end{equation}
where the first inequality follows from the definition of $ \mathbf{z}^{k}$ and $ \widehat{\mathbf{z}}^{k}$ and
in the last inequality we used $\left\Vert \mathbf{z}_{i}^{k}-\widehat{\mathbf{x}}_{i}(\mathbf{x}^{k})\right\Vert \leq\varepsilon_{i}^{k}$.
Denoting by $\overline{S}^k$ the complement of $S$, we also have, for $k$ large enough,
\begin{equation}
\begin{array}{l}
\nabla_{\mathbf{x}}F\left(\mathbf{x}^{k}\right)^{T}\left(\widehat{\mathbf{z}}^{k}-\mathbf{x}^{k}\right)\smallskip\\
\qquad\qquad =
\nabla_{\mathbf{x}}F\left(\mathbf{x}^{k}\right)^{T}\left(\widehat{\mathbf{z}}^{k}-\widehat{\mathbf{x}}(\mathbf{x}^{k})
+\widehat{\mathbf{x}}(\mathbf{x}^{k})-\mathbf{x}^{k}\right)\smallskip\\
\qquad\qquad = \nabla_{\mathbf{x}}F\left(\mathbf{x}^{k}\right)^{T}_{S^k} (\mathbf{z}^k - \widehat{\mathbf{x}}(\mathbf{x}^k))_{S^k}\smallskip\\ \qquad\qquad\quad
+
\nabla_{\mathbf{x}}F\left(\mathbf{x}^{k}\right)^{T}_{\overline{S}^k} (\mathbf{x}^k - \widehat{\mathbf{x}}(\mathbf{x}^k))_{\overline{S}^k}\smallskip\\
\qquad \qquad \quad + \nabla_{\mathbf{x}}F\left(\mathbf{x}^{k}\right)^{T}_{S^k} (\widehat{\mathbf{x}}(\mathbf{x}^k)-\mathbf{x}^k)_{S^k}\smallskip\\ \qquad\qquad\quad
+ \nabla_{\mathbf{x}}F\left(\mathbf{x}^{k}\right)^{T}_{\overline{S}^k} (\widehat{\mathbf{x}}(\mathbf{x}^k)-\mathbf{x}^k)_{\overline{S}^k}\smallskip\\
\qquad\qquad = \nabla_{\mathbf{x}}F\left(\mathbf{x}^{k}\right)^{T}_{S^k} (\mathbf{z}^k - \widehat{\mathbf{x}}(\mathbf{x}^k))_{S^k} \smallskip\\ \qquad\qquad +
\nabla_{\mathbf{x}}F\left(\mathbf{x}^{k}\right)^{T}_{S^k} (\widehat{\mathbf{x}}(\mathbf{x}^k)-\mathbf{x}^k)_{S^k},
\end{array}
\end{equation}
where in the second equality we used the definition of $\widehat{\mathbf{z}}^k$ and of the set $S^k$.
Now, using the above identity and
Lemma \ref{lemma on errors}, we can write
\begin{equation}\label{eq:descent_at_x_n}
\begin{array}{l}
\nabla_{\mathbf{x}}F\left(\mathbf{x}^{k}\right)^{T}\left(\widehat{\mathbf{z}}^{k}-\mathbf{x}^{k}\right) + \sum_{i\in S^k} g_i(\widehat{\mathbf{z}}_i^{k})-\sum_{i\in S^k} g_i(\mathbf{x}^{k}_i) \\[0.3em]
= \nabla_{\mathbf{x}}F\left(\mathbf{x}^{k}\right)^{T}\left(\widehat{\mathbf{z}}^{k}-\mathbf{x}^{k}\right) +
\sum_{i\in S^k} g_i(\widehat{\mathbf{x}}_i(\mathbf{x}^{k})) -\sum_{i\in S^k} g_i(\mathbf{x}^{k}_i) \\[0.3em]
\qquad \qquad +
\sum_{i \in S^k} g_i(\widehat{\mathbf{z}}_i^k) -\sum_{i \in S^k} g_i(\widehat{\mathbf{x}}_i(\mathbf{x}^{k}))
\\[0.3em]
\leq -\tilde c\left\Vert \widehat{\mathbf{x}}(\mathbf{x}^{k})-\mathbf{x}^{k}\right
\Vert ^{2}+\sum_{i \in S^k}\varepsilon_{i}^{k}\left\Vert \nabla_{\mathbf{x}_{i}}F(\mathbf{x}^{k})\right\Vert
+ L_G\sum_{i \in S^k}\varepsilon_{i}^{k}.
\end{array}
\end{equation}
Finally, from the definition of
$\widehat{\mathbf{z}}^k$ and of the set $S^k$, we have for all $k$ large enough,
\begin{equation}
\begin{array}{l}
V(\mathbf{x}^{k+1}) = F(\mathbf{x}^{k+1}) + \sum_{ {i \in \cal N}}g_i (\mathbf{x}_i^{k+1}) \\[0.3em]
= F(\mathbf{x}^{k+1}) + \sum_{i\in {\cal N}}g_i (\mathbf{x}_i^{k} + \gamma^k(\widehat{\mathbf{z}}^k_i - \mathbf{x}_i^{k} ))\\[0.3em]
\leq F(\mathbf{x}^{k+1}) + \sum_{i \in \cal N}g_i (\mathbf{x}^k_i) + \gamma^k \left(\sum_{i \in S^k} (g_i (\widehat{\mathbf{z}}^k_i)
- g_i (\mathbf{x}^k_i)) \right)\\[0.3em]
\leq
V\left(\mathbf{x}^{k}\right)-\gamma^{k}\left(\tilde c -\gamma^{k}{L_{\nabla U}}\right)\left\Vert \widehat{\mathbf{x}}(\mathbf{x}^{k})-\mathbf{x}^{k}\right\Vert ^{2}+T^{k},\end{array}\label{eq:descent_Lemma_2}
\end{equation}
where in the first inequality we used the the convexity of the $g_i$'s, whereas the second follows from
\eqref{eq:descent_Lemma}, \eqref{eq:49bis} and \eqref{eq:descent_at_x_n}, with
$$T^{k}\triangleq\gamma^{k}\,
\sum_{i \in S^k}\varepsilon_{i}^{k}\left( L_G +
\left\Vert \nabla_{\mathbf{x}_{i}}F(\mathbf{x}^{k})\right\Vert\right) +\left(\gamma^{k}\right)^{2}{L_{\nabla F}}\,\sum_{i\in \cal N}(\varepsilon_{i}^{k})^{2}.$$
Using assumption (v),
we can bound $T^k$ as
\[
T^{k}\leq
(\gamma^k)^2 \left[ N \alpha_1 (\alpha_2 L_G +1) + (\gamma^k)^2 L_{\nabla F} \left(N\alpha_1\alpha_2\right)^2 \right],
\]
which, by assumption (iv) implies
$\sum_{k=0}^{\infty}T^{k}<\infty$.
Since $\gamma^{k}\rightarrow 0$, it follows from (\ref{eq:descent_Lemma_2}) that there exist some positive constant
$\beta_{1}$ and a sufficiently large $k$, say $k\geq\bar{{k}}$, such that
\begin{equation}
V(\mathbf{x}^{k+1})\leq V(\mathbf{x}^{k})-\gamma^{k}\beta_{1}\left\Vert \widehat{\mathbf{x}}(\mathbf{x}^{k})-\mathbf{x}^{k}\right\Vert ^{2}+T^{k}.\label{eq:descent_Lemma_3_}
\end{equation}
Invoking Lemma \ref{lemma_Robbinson_Siegmunt} with the identifications
$X^{k}=V(\mathbf{x}^{k+1})$, $Y^{k}=\gamma^{k}\beta_{1}\left\Vert \widehat{\mathbf{x}}(\mathbf{x}^{k})-\mathbf{x}^{k}\right\Vert ^{2}$
and $Z^{k}=T^{k}$ while using $\sum_{k=0}^\infty T^{k}<\infty$, we deduce
from (\ref{eq:descent_Lemma_3_}) that either $\{V(\mathbf{x}^{k})\}\rightarrow-\infty$
or else $\{V(\mathbf{x}^{k})\}$ converges to a finite
value and\vspace{-0.1cm}
\begin{equation}
\lim_{k\rightarrow\infty}\sum_{t=\bar{{k}}}^{k}\gamma^{t}\left\Vert \widehat{\mathbf{x}}(\mathbf{x}^{t})-\mathbf{x}^{t}\right\Vert ^{2}<+\infty.\vspace{-0.1cm}\label{eq:finite_sum_series}
\end{equation}
Since $V$ is coercive, $V(\mathbf{x})\geq\min_{\mathbf{y}\in X}V(\mathbf{y})>-\infty$,
implying that $\{V\left(\mathbf{x}^{k}\right)\}$ is convergent;
it follows from (\ref{eq:finite_sum_series}) and $\sum_{k=0}^{\infty}\gamma^{k}=\infty$
that $\liminf_{k\rightarrow\infty}\left\Vert \widehat{\mathbf{x}}(\mathbf{x}^{k})-\mathbf{x}^{k}\right\Vert =0.$
Using Prop. \ref{Prop_x_y}, we show next that $\lim_{k\rightarrow\infty}\left\Vert \widehat{\mathbf{x}}(\mathbf{x}^{k})-\mathbf{x}^{k}\right\Vert =0$;
for notational simplicity we will write $\triangle\widehat{\mathbf{x}}(\mathbf{x}^{k})\triangleq\widehat{\mathbf{x}}(\mathbf{x}^{k})-\mathbf{x}^{k}$.
Suppose, by contradiction, that $\limsup_{k\rightarrow\infty}\left\Vert \triangle\widehat{\mathbf{x}}(\mathbf{x}^{k})\right\Vert >0$.
Then, there exists a $\delta>0$ such that $\left\Vert \triangle\widehat{\mathbf{x}}(\mathbf{x}^{k})\right\Vert >2\delta$
for infinitely many $k$ and also $\left\Vert \triangle\widehat{\mathbf{x}}(\mathbf{x}^{k})\right\Vert <\delta$
for infinitely many $k$. Therefore, one can always find an infinite
set of indexes, say $\mathcal{K}$, having the following properties:
for any $k\in\mathcal{K}$, there exists an integer $i_{k}>k$ such
that
\begin{eqnarray}
\left\Vert \triangle\widehat{\mathbf{x}}(\mathbf{x}^{k})\right\Vert <\delta, & & \left\Vert \triangle\widehat{\mathbf{x}}(\mathbf{x}^{i_{k}})\right\Vert >2\delta\medskip\label{eq:outside_interval}\\
\delta\leq\left\Vert \triangle\widehat{\mathbf{x}}(\mathbf{x}^{j})\right\Vert \leq2\delta & & k<j<i_{k}.\label{eq:inside_interval}
\end{eqnarray}
Given the above bounds, the following holds: for all $k\in\mathcal{K}$,
\begin{eqnarray}
\delta & \overset{(a)}{<} & \left\Vert \triangle\widehat{\mathbf{x}}(\mathbf{x}^{i_{k}})\right\Vert -\left\Vert \triangle\widehat{\mathbf{x}}(\mathbf{x}^{k})\right\Vert \medskip\nonumber \\
& \leq & \left\Vert \widehat{\mathbf{x}}(\mathbf{x}^{i_{k}})-\widehat{\mathbf{x}}(\mathbf{x}^{k})\right\Vert +\left\Vert \mathbf{x}^{i_{k}}-\mathbf{x}^{k}\right\Vert \\
& \overset{(b)}{\leq} & (1+\hat{{L}})\left\Vert \mathbf{x}^{i_{k}}-\mathbf{x}^{k}\right\Vert \\
& \overset{(c)}{\leq} & {\color{black} (1+\hat{{L}})\sum_{t=k}^{i_{k}-1}\gamma^{t}\left(\left\Vert \triangle\widehat{\mathbf{x}}(\mathbf{x}^{t})_{S^t}\right\Vert +\left\Vert (\mathbf{z}^{t}-\widehat{\mathbf{x}}(\mathbf{x}^{t}))_{S^t}\right\Vert \right)}\vspace{-0.3cm}\nonumber \\
& \overset{(d)}{\leq} & (1+\hat{{L}})\,(2\delta+\varepsilon^{\max})\sum_{t=k}^{i_{k}-1}\gamma^{t},\label{eq:lower_bound_sum}
\end{eqnarray}
where (a) follows from (\ref{eq:outside_interval});
(b) is due to Prop. \ref{Prop_x_y}(a); (c) comes from the triangle
inequality, the updating rule of the algorithm {\color{black} and the definition of $\widehat{\mathbf{z}}^k$}; and in (d) we used
(\ref{eq:outside_interval}), (\ref{eq:inside_interval}), and $\left\Vert \mathbf{z}^{t}-\widehat{\mathbf{x}}(\mathbf{x}^{t})\right\Vert \leq\sum_{i\in \cal N}\varepsilon_{i}^{t}$,
where $\varepsilon^{\max}\triangleq\max_{k}\sum_{i\in\cal N}\varepsilon_{i}^{k}<\infty$.
It follows from (\ref{eq:lower_bound_sum}) that
\begin{equation}
\liminf_{k\rightarrow\infty}\sum_{t=k}^{i_{k}-1}\gamma^{t}\geq\dfrac{{\delta}}{(1+\hat{{L}})(2\delta+\varepsilon^{\max})}>0.\label{eq:lim_inf_bound}
\end{equation}
\noindent
We show next that (\ref{eq:lim_inf_bound}) is in contradiction with
the convergence of $\{V(\mathbf{x}^{k})\}$. To do that, we preliminary
prove that, for sufficiently large $k\in\mathcal{K}$, it must be
$\left\Vert \triangle\widehat{\mathbf{x}}(\mathbf{x}^{k})\right\Vert \geq\delta/2$.
Proceeding as in (\ref{eq:lower_bound_sum}), we have: for any given
$k\in\mathcal{K}$,
\[
\begin{array}{l}
\left\Vert \triangle\widehat{\mathbf{x}}(\mathbf{x}^{k+1})\right\Vert -\left\Vert \triangle\widehat{\mathbf{x}}(\mathbf{x}^{k})\right\Vert \leq(1+\hat{{L}})\left\Vert \mathbf{x}^{k+1}-\mathbf{x}^{k}\right\Vert \smallskip\\
\qquad\qquad\qquad\qquad\qquad\leq(1+\hat{{L}})\gamma^{k}\left(\left\Vert \triangle\widehat{\mathbf{x}}(\mathbf{x}^{k})\right\Vert +\varepsilon^{\max}\right).
\end{array}
\]
It turns out that for sufficiently large $k\in\mathcal{K}$ so that
$(1+\hat{{L}})\gamma^{k}<\delta/(\delta+2\varepsilon^{\max})$, it
must be
\begin{equation}
\left\Vert \triangle\widehat{\mathbf{x}}(\mathbf{x}^{k})\right\Vert \geq\delta/2;\label{eq:lower_bound_delta_x_n}
\end{equation}
otherwise the condition $\left\Vert \triangle\widehat{\mathbf{x}}(\mathbf{x}^{k+1})\right\Vert \geq\delta$
would be violated {[}cf. (\ref{eq:inside_interval}){]}. Hereafter
we assume w.l.o.g. that (\ref{eq:lower_bound_delta_x_n}) holds for
all $k\in\mathcal{K}$ (in fact, one can alway restrict $\{\mathbf{x}^{k}\}_{k\in\mathcal{K}}$
to a proper subsequence).
We can show now that (\ref{eq:lim_inf_bound}) is in contradiction
with the convergence of $\{V(\mathbf{x}^{k})\}$. Using (\ref{eq:descent_Lemma_3_})
(possibly over a subsequence), we have: for sufficiently large $k\in\mathcal{K}$,
\begin{eqnarray}
V(\mathbf{x}^{i_{k}}) & \leq & V(\mathbf{x}^{k})-\beta_{2}\sum_{t=k}^{i_{k}-1}\gamma^{t}\left\Vert \triangle\widehat{\mathbf{x}}(\mathbf{x}^{t})\right\Vert ^{2}+\sum_{t=k}^{i_{k}-1}T^{t}\nonumber \\
& \overset{(a)}{<} & V(\mathbf{x}^{k})-\beta_{2}(\delta^{2}/4)\sum_{t=k}^{i_{k}-1}\gamma^{t}+\sum_{t=k}^{i_{k}-1}T^{t}\label{eq:liminf_zero}
\end{eqnarray}
where in (a) we used (\ref{eq:inside_interval}) and (\ref{eq:lower_bound_delta_x_n}),
and $\beta_{2}$ is some positive constant. Since $\{V(\mathbf{x}^{k})\}$
converges and $\sum_{k=0}^{\infty}T^{k}<\infty$, (\ref{eq:liminf_zero})
implies $\lim_{\mathcal{K}\ni k\rightarrow\infty}\,\sum_{t=k}^{i_{k}-1}\gamma^{t}=0,$
which contradicts (\ref{eq:lim_inf_bound}).
Finally, since the sequence $\{\mathbf{x}^{k}\}$ is bounded {[}due
to the coercivity of $V$ and the convergence of $\{V(\mathbf{x}^{k})\}${]},
it has at least one limit point $\bar{{\mathbf{x}}}$ that must belong
to $X$. By the continuity of $\widehat{\mathbf{x}}(\bullet)$
{[}Prop. \ref{Prop_x_y}(a){]} and $\lim_{k\rightarrow\infty}\left\Vert \widehat{\mathbf{x}}(\mathbf{x}^{k})-\mathbf{x}^{k}\right\Vert =0$,
it must be $\widehat{\mathbf{x}}(\bar{{\mathbf{x}}})=\bar{{\mathbf{x}}}$.
By Prop. \ref{Prop_x_y}(b) $\bar{{\mathbf{x}}}$ is also a stationary
solution of Problem (\ref{eq:problem 1}).
As a final remark, note that if $\varepsilon_i^k=0$ for every $i$ and for every $k$ large enough, i.e. if eventually
$\widehat{\mathbf{x}}(\mathbf{x}^k)$ is computed exactly, there is no need to assume that $G$ is
globally Lipschitz. In fact in \eqref{eq:descent_at_x_n} the term containing $L_G$ disappears, and
actually all the terms $T^k$ are zero and all the subsequent derivations independent of the Lipschitzianity of $G$.
\hfill$\square$
\begin{spacing}{0.050000000000000003}
\bibliographystyle{IEEEtran}
|
\section{Introduction}\label{intro}
Let $(B_t,~t\geq 0)$ be a standard Brownian motion, $T_1$ its first hitting time of level one, and
$U$ a uniform random variable on $[0,1]$, independent of $B$. In \cite{elie2013expectation}, it is first shown that the random variable $\alpha$ defined by
\begin{equation}\label{variable}
\alpha=\frac{B_{UT_1}}{\sqrt{T_1}}
\end{equation}
is centered. Intrigued by this property, we determined the distribution of this variable, which is expressed in \cite{rosenbaum2013law} under the following form, where $\underset{\mathcal{L}}{=}$ denotes equality in law:
\begin{equation}\label{law}
\alpha\underset{\mathcal{L}}{=}\Lambda L_1-\frac{1}{2}|B_1|,\end{equation}
with $L_t$ the local time at point $0$ of $B$ at time $t$ and $\Lambda$ a uniform random variable on $[0,1]$, independent of $(|B_1|,L_1)$. The centering property is easily recovered from \eqref{law} since
$$\mathbb{E}[\Lambda L_1-\frac{1}{2}|B_1|]=\frac{1}{2}\mathbb{E}[L_1-|B_1|]=0.$$
\noindent In fact, in \cite{rosenbaum2013law}, a preliminary to the proof of \eqref{law} is to obtain the law of a triplet of random variables defined in terms of the pseudo-Brownian bridge introduced in \cite{biane1987processus}, see Section \ref{sec_prelim} below. In this paper, we show that the law of this triplet enables us to derive several unexpected simple formulas for various quantities related to some very classical Brownian type processes, namely the Brownian bridge, the Brownian meander and the three dimensional Bessel process. More precisely, we focus on distributional properties of these processes when sampled with an independent uniform random variable. Thus, this work can be viewed as a modest complement to the seminal paper by Pitman, see \cite{pitman1999brownian}, where the laws of these processes when sampled with (several) independent uniform random variables are already studied.\\
\noindent The paper is organized as follows. In Section \ref{sec_prelim}, we give some preliminary results related to the law of $(|B_1|,L_1)$. Indeed, they play an important role in the proofs. Distributional properties for the Brownian bridge are established in Section \ref{sec_bri} whereas the Brownian meander and the three dimensional Bessel process are investigated in Section \ref{sec_mea}. Finally, in Section \ref{sec_filt}, we reinterpret the fact that $\alpha$ is centered through the lenses of an enlargement formula for the Brownian motion with the time $T_1$ due to Jeulin, see \cite{jeulin1979grossissement}. In particular we show that this centering property can be translated in terms of the expectation of the random variable
$U/(R_UR_1^2)$, where $R$ is a three dimensional Bessel process and $U$ a uniform random variable on $[0,1]$ independent of $R$.
\section{Some preliminary results about the law of $(|B_1|,L_1)$}\label{sec_prelim}
Before dealing with the Brownian bridge, the Brownian meander and the three dimensional Bessel process, we give here some preliminary
results related to the distribution of the couple $(|B_1|,L_1)$. These results will play an important role in the proofs of our main theorems.\\
\noindent It is well known that the law of $(B_1,L_1)$ admits a density on $\mathbb{R}\times\mathbb{R}^+$. Its value at point $(x,l)$ is given by
\begin{equation}\label{dens}
\frac{1}{\sqrt{2\pi}}(|x|+l)\text{exp}\big(-\frac{(l+|x|)^2}{2}\big).
\end{equation}
For $l\geq 0$, we set
$$H(l)=\text{e}^{l^2/2}\int_l^{+\infty}dx\text{e}^{-x^2/2}.$$
The following consequences of \eqref{dens} shall be useful in the sequel.
\begin{proposition}\label{prop_prelim}
Let $l\geq 0$. We have the double identity
\begin{equation}\label{double}
\mathbb{E}[L_1||B_1|=l]=\mathbb{E}[|B_1||L_1=l]=H(l).
\end{equation}
Furthermore, one has
\begin{equation}\label{double2}
H(l)=l\mathbb{E}[\frac{1}{N^2+l^2}],\end{equation}
where $N$ denotes a standard Gaussian random variable.
\end{proposition}
\begin{proof}
We start with the proof of \eqref{double}. Of course, it can be deduced from \eqref{dens} at the cost of some integrations. We
prefer the following arguments. First, the equality on the left hand side of \eqref{double} stems from the symmetry of the law of
$(|B_1|,L_1)$, which is obvious from \eqref{dens}. Thus, we now have to show the second equality in \eqref{double}. This easily follows from the identity
\begin{equation}\label{balayage}
\mathbb{E}[\phi(L_1)|B_1|]=\mathbb{E}[\int_0^{L_1}dx\phi(x)],
\end{equation}
which is valid for any bounded measurable function $\phi$. Indeed, assuming \eqref{balayage} for a moment, using the fact that
$$L_1\underset{\mathcal{L}}{=}|B_1|,$$
we may write \eqref{balayage} as
$$\int_0^{+\infty}dl\text{e}^{-l^2/2}\phi(l)\mathbb{E}[|B_1||L_1=l]=\int_0^{+\infty}dl\phi(l)\int_l^{+\infty}dx\text{e}^{-x^2/2}.$$
Hence, since this is true for every bounded measurable function $\phi$, we get
$$\text{e}^{-l^2/2}\mathbb{E}[|B_1||L_1=l]=\int_l^{+\infty}dx\text{e}^{-x^2/2},$$
which is the desired result for \eqref{double}.\\
\noindent It remains to prove \eqref{balayage} for a generic bounded measurable function $\phi$. Remark that the formula
$$\phi(L_t)|B_t|=\int_0^tdB_s\phi(L_s)\text{sign}(B_s)+\int_0^tdL_s\phi(L_s)$$
is a very particular case of the balayage formula, see \cite{revuz1999continuous}, page 261. It now suffices to take expectation on both sides of this last equality to obtain \eqref{balayage}.\\
\noindent We now give the proof of the second part of Proposition \ref{prop_prelim}. First, note that
$$\mathbb{E}[\frac{1}{N^2+l^2}]=\int_0^{+\infty}dv\text{e}^{-vl^2}\mathbb{E}[\text{e}^{-vN^2}].$$
Using the Laplace transform of $N^2$, we obtain
\begin{equation}\label{form}
\mathbb{E}[\frac{1}{N^2+l^2}]=\int_0^{+\infty}dv\frac{\text{e}^{-vl^2/2}}{2\sqrt{1+v}}.
\end{equation}
Then remark that thanks to the change of variable $x^2=(1+v)l^2$, we get
$$H(l)=\text{e}^{l^2/2}\int_l^{+\infty}dx\text{e}^{-x^2/2}=\int_0^{+\infty}dv\frac{l}{2\sqrt{1+v}}\text{e}^{-vl^2/2}.$$
This together with $\eqref{form}$ gives the second part of Proposition \ref{prop_prelim}.
\end{proof}
\section{The Brownian bridge under uniform sampling}\label{sec_bri}
Before giving our theorem on the uniformly sampled Brownian bridge, we recall a result related to the pseudo-Brownian bridge established in \cite{rosenbaum2013law}, and which is the key to most of the proofs in this paper. The pseudo-Brownian bridge was introduced in \cite{biane1987processus} and is defined by
$$(\frac{B_{u\tau_1}}{\sqrt{\tau_1}},~u\leq 1),$$ with
$(\tau_l,~l>0)$ the inverse local time process:
$$\tau_l=\text{inf}\{t,~L_t>l\}.$$
This pseudo-Brownian bridge is equal to $0$ at time $0$ and time $1$ and has the same quadratic variation as the Brownian motion. Thus, it shares some similarities with the Brownian bridge, which explains its name. Let $U$ be a uniform random variable on $[0,1]$ independent of $B$. The following theorem is proved in \cite{rosenbaum2013law}.
\begin{theorem}\label{theo1}
There is the identity in law
$$(\frac{B_{U\tau_1}}{\sqrt{\tau_1}},\frac{1}{\sqrt{\tau_1}},L_{U\tau_1})\underset{\mathcal{L}}{=}(\frac{1}{2}B_1,L_1,\Lambda),$$
with $\Lambda$ a uniform random variable on $[0,1]$, independent of $(B_1,L_1)$.
\end{theorem}
\noindent In other words, $L_{U\tau_1}$ is a uniform random variable on $[0,1]$, independent of the pair $$(\frac{B_{U\tau_1}}{\sqrt{\tau_1}},\frac{1}{\sqrt{\tau_1}}),$$ which is distributed as $(\frac{1}{2}B_1,L_1)$.\\
\noindent To deduce some properties of the Brownian bridge from Theorem \ref{theo1}, the idea is to use an absolute continuity relationship between the law of the pseudo-Brownian bridge and that of the Brownian bridge shown by Biane, Le Gall and Yor in \cite{biane1987processus}. More precisely, for $F$ a non negative measurable function on $\mathbb{C}([0,1],\mathbb{R})$, we have
\begin{equation}\label{bly}
\mathbb{E}\big[F\big(\frac{B_{u\tau_1}}{\sqrt{\tau_1}},~u\leq 1\big)\big]=\sqrt{\frac{2}{\pi}}\mathbb{E}\big[F\big(b(u),~u\leq 1\big)\frac{1}{\lambda_1^0}\big],\end{equation}
where $\big(b(u),~u\leq 1\big)$ denotes the Brownian bridge and $(\lambda_u^x,~u\leq 1,~x\in \mathbb{R})$ its family of local times.
Let $U$ be again a uniform random variable on $[0,1]$ independent of $b$. The following theorem is easily deduced from Theorem \ref{theo1} together with Equation \ref{bly}.
\begin{theorem}\label{theo2}
For any non negative measurable functions $f$ and $g$, we have
\begin{equation}\label{pont}
\mathbb{E}\big[f\big(b(U),\lambda_1^0\big)g\big(\frac{\lambda_U^0}{\lambda_1^0}\big)\big]=\sqrt{\frac{\pi}{2}}\mathbb{E}\big[f(\frac{1}{2}B_1,L_1)L_1\big]\mathbb{E}[g(\Lambda)],\end{equation}
with $\Lambda$ a uniform random variable on $[0,1]$, independent of $(B_1,L_1)$.
\end{theorem}
\noindent Thus, $\lambda_U^0/\lambda_1^0$ is a uniform random variable on $[0,1]$, independent of the pair $\big(b(U),\lambda_1^0\big)$ which is distributed according to \eqref{pont} with $g=1$.\\
\noindent The following corollary of Theorem \ref{theo2} provides some surprisingly simple expressions for some densities and (conditional) expectations of quantities related to the Brownian bridge.
\begin{corollary}\label{cortheo2}
The following properties hold:\\
\noindent $\bullet~$The variable $\lambda_1^0$ admits a density on $\mathbb{R}^+$. Its value at point $l\geq 0$ is given by
$$l\emph{exp}(-l^2/2).$$
Hence, $\lambda_1^0$ has the same law as $\sqrt{2\mathcal{E}}$, with $\mathcal{E}$ an exponential random variable. Therefore,
$\lambda_1^0$ is Rayleigh distributed.\\
\noindent $\bullet~$The density of $b(U)$ at point $y$ given $\lambda_1^0=l$ is given by
$$\mathbb{E}[\lambda_1^y|\lambda_1^0=l]=(2|y|+l)\emph{exp}\big(-(2y^2+2|y|l)\big).$$
Consequently, there is the formula
$$\mathbb{E}\big[\frac{\lambda_1^y}{\lambda_1^0}\big]=\emph{exp}(-2y^2).$$
\noindent $\bullet~$The density of $b(U)$ at point $y$ is given by
$$\mathbb{E}[\lambda_1^y]=\int_{2|y|}^{+\infty}dz\emph{exp}(-z^2/2).$$
Thus, we have $b(U)\underset{\mathcal{L}}{=}\sqrt{2\mathcal{E}}(V/2)$, with $\mathcal{E}$ an exponential variable independent of $V$ which is uniformly distributed on $[-1,1]$.
\end{corollary}
\noindent The first part of Corollary \ref{cortheo2} is obviously deduced from Theorem \ref{theo2}
and is in fact a very classical result, see \cite{biane1987processus,biane1988quelques,imhof1984density,revuz1999continuous}. We now prove the second part.
\begin{proof}
\noindent Let $f$ be a non negative measurable function. First note that
$$\mathbb{E}\big[f\big(b(U)\big)|\lambda_1^0=l\big]=\mathbb{E}\big[\int_0^1 du f\big(b(u)\big)|\lambda_1^0=l\big]=\int_{\mathbb{R}} dy f(y)\mathbb{E}[\lambda_1^y|\lambda_1^0=l].$$
Hence the density of $b(U)$ at point $y$ given $\lambda_1^0=l$ is equal to
$$\mathbb{E}[\lambda_1^y|\lambda_1^0=l].$$
Now, let $h$ denote the density of the couple $(B_1,L_1)$ given in Equation \eqref{dens}. From Theorem \ref{theo2}, we easily get that the density of $b(U)$ at point $y$ given $\lambda_1^0=l$ is equal to
$$2\sqrt{\frac{\pi}{2}}\frac{lh(2y,l)}{l\text{exp}(-l^2/2)}=\sqrt{2\pi}h(2y,l)\text{exp}(l^2/2).$$
The first statement in the second part of Corollary \ref{cortheo2} readily follows from Equation \eqref{dens}. For the second statement, we use the fact that
$$
\mathbb{E}\big[\frac{\lambda_1^y}{\lambda_1^0}\big]=\mathbb{E}\big[\int_0^{+\infty}dl\frac{1}{l}\mathbb{E}[\lambda_1^y|\lambda_1^0=l]\big]l\text{exp}(-l^2/2)=\sqrt{2\pi}\int_0^{+\infty}dlh(2|y|,l).
$$
Using the definition of $h$, this last expression is equal to
$$\text{exp}(-2y^2).$$
\end{proof}
\noindent The last identity in Corollary \ref{cortheo2} is easily deduced from Theorem \ref{theo2} together with Proposition \ref{prop_prelim}. Note that this formula can also be found in \cite{shorack2009empirical}, page 400.
\section{The Brownian meander and the three dimensional Bessel process under uniform sampling}\label{sec_mea}
In this section, we reinterpret Theorem \ref{theo1} in terms of the Brownian meander and the three dimensional Bessel process.
\subsection{The Brownian meander}
We first turn to the translation of Theorem \ref{theo1} in terms of the Brownian meander, denoted by $\big(m(u),~u\leq 1\big)$. To do so, we use an equality in law shown by Biane and Yor in \cite{biane1988quelques}. More precisely we have
$$\big((m(u),i_u),~u\leq 1\big)\underset{\mathcal{L}}{=}\big((|b(u)|+\lambda_u^0,\lambda_u^0),~u\leq 1\big),$$
where $$i_u=\underset{u\leq t\leq 1}{\text{inf}} m_t.$$ Thus, we can reinterpret Theorem \ref{theo2} as follows.
\begin{theorem}\label{theo3}
For any non negative measurable functions $f$ and $g$, we have
\begin{equation*}
\mathbb{E}\big[f\big(m(U),m(1)\big)g\big(\frac{i_U}{m(1)}\big)\big]=\sqrt{\frac{\pi}{2}}\mathbb{E}\big[f(\frac{1}{2}|B_1|+\Lambda L_1,L_1)L_1g(\Lambda)\big],\end{equation*}
with $\Lambda$ a uniform random variable on $[0,1]$, independent of $(B_1,L_1)$.
\end{theorem}
\noindent Let $(\tilde{\lambda}_1^y,~y\geq 0)$ denotes the family of local times of $m$ at time $1$. Similarly to what we have done for the Brownian bridge, we are able to retrieve from Theorem \ref{theo3} simple expressions for the laws of $m(1)$ and $m(U)$. We state these results in the following corollary.
\begin{corollary}\label{cortheo3}
The following properties hold:\\
\noindent $\bullet~$The variable $m(1)$ is Rayleigh distributed.\\
\noindent $\bullet~$The density of $m(U)$ at point $y\geq 0$ is given by
$$\mathbb{E}[\tilde{\lambda}_1^y]=2\int_y^{2y}\emph{exp}(-z^2/2)dz.$$
Thus, we have $m(U)\underset{\mathcal{L}}{=}\sqrt{2\mathcal{E}}W$, with $\mathcal{E}$ an exponential variable independent of $W$ which is uniformly distributed on $[1/2,1]$.
\end{corollary}
\begin{proof}
The proof of the first part of Corollary \ref{cortheo3} is obvious from Theorem \ref{theo3}. We now consider the second part. Let $f$ be a non negative measurable function. Using Theorem \ref{theo3} together with Equation \eqref{dens}, we get that
$$\mathbb{E}\big[f\big(m(U)\big)\big]=\int_0^{+\infty}dx\int_0^{+\infty}dl l(x+l)\text{e}^{-(x+l)^2/2}\mathbb{E}\big[f(\frac{x}{2}+\Lambda l)\big].$$
Now remark that
$$\mathbb{E}\big[f(\frac{x}{2}+\Lambda l)\big]=\frac{1}{l}\int_{x/2}^{x/2+l}d\nu f(\nu).$$
Therefore, by Fubini's theorem, we get
\begin{align*}
\mathbb{E}\big[f\big(m(U)\big)\big]&=\int_0^{+\infty}dx\int_0^{+\infty}dl (x+l)\text{e}^{-(x+l)^2/2}\int_{x/2}^{x/2+l}d\nu f(\nu)\\
&=\int d\nu f(\nu)\int_0^{2\nu}dx\int_{\nu-x/2}^{+\infty}dl (x+l)\text{e}^{-(x+l)^2/2}.
\end{align*}
Thus, the density of $m(U)$ at point $\nu$ is given by
\begin{align*}
\int_0^{2\nu}dx\int_{\nu-x/2}^{+\infty}dl (x+l)\text{e}^{-(x+l)^2/2}&=\int_0^{2\nu}dx\text{exp}\big(-\frac{(x/2+\nu)^2}{2}\big)\\
&=2\int_{\nu}^{2\nu}dz\text{e}^{-z^2/2}.
\end{align*}
This ends the proof.
\end{proof}
\noindent In fact, as it is the case for the Brownian bridge, we can give explicit formulas for several other quantities related to the Brownian meander, for example the law of $m(U)$ given $m(1)$. However, these expressions are not so simple and therefore probably less interesting than those obtained for the Brownian bridge.
\subsection{The three dimensional Bessel process}
Finally, let $(R_t,~t\geq 0)$ be a three dimensional Bessel process starting from $0$ and
$$J_u=\underset{u\leq t\leq 1}{\text{inf}}R_t.$$
Using Imhof's absolute continuity relationship between the law of the meander and that of the three dimensional Bessel process, see \cite{biane1987processus,imhof1984density}, we may rewrite Theorem \ref{theo3} as follows.
\begin{theorem}\label{theo4}
For any non negative measurable functions $f$ and $g$, we have
\begin{equation*}
\mathbb{E}\big[f\big(R(U),R(1)\big)g\big(\frac{J_U}{R(1)}\big)\big]=\mathbb{E}\big[f(\frac{1}{2}|B_1|+\Lambda L_1,L_1)L_1^2g(\Lambda)\big],\end{equation*}
with $\Lambda$ a uniform random variable on $[0,1]$, independent of $(B_1,L_1)$.
\end{theorem}
\noindent We finally give the following corollary.
\begin{corollary}\label{cortheo4}
The following properties hold:\\
\noindent $\bullet~$The density of $R(1)$ at point $y\geq 0$ is given by
$$\sqrt{\frac{2}{\pi}}y^2\emph{exp}(-y^2/2).$$
\noindent $\bullet~$ $R(U)\underset{\mathcal{L}}{=}\sqrt{U}R(1)$ and its density at point $y\geq 0$ is given by
$$2\sqrt{\frac{2}{\pi}}y\int_y^{+\infty}\emph{exp}(-z^2/2)dz.$$
\noindent $\bullet~$The law of $R(U)$ given $R(1)$ is the same as the law of $m(U)$ given $m(1)$.
\end{corollary}
\noindent The first two parts of Corollary \ref{cortheo4} are in fact easily deduced from basic properties of the three dimensional Bessel process. The last part is a consequence of Imhof's relation.
\section{The centering property of $\alpha$ revisited through an enlargement of filtration formula}\label{sec_filt}
In this last section, we revisit the centering property of the variable
$$\alpha=\frac{B_{UT_1}}{\sqrt{T_1}},$$ which is proved in \cite{elie2013expectation} and leads to various developments in \cite{rosenbaum2013law}.
Our goal here is to show that this result can be recovered from simple properties of the three dimensional Bessel process sampled at uniform time, together with an enlargement of filtration formula for the Brownian motion with the time $T_1$ due to Jeulin, see \cite{jeulin1979grossissement}.
\subsection{Some preliminary remarks on the uniformly sampled Bessel process}
Let $U$ be a uniform random variable on $[0,1]$, independent of the considered Bessel process $R$. We start with the two following lemmas on the conditional expectation of the uniformly sampled Bessel process.
\begin{lemma}\label{lem1}
We have
\begin{equation}\label{lem1_1}
\mathbb{E}[R_U|R_1=r]=\frac{1}{2}\big(r+\mathbb{E}[\frac{U}{R_U}|R_1=r]\big).
\end{equation}
Consequently,
$$\mathbb{E}\big[\frac{R_U}{R_1^2}\big]=\frac{1}{2}\big(\sqrt{\frac{2}{\pi}}+\mathbb{E}[\frac{U}{R_UR_1^2}]\big).$$
\end{lemma}
\begin{lemma}\label{lem2}
We have
$$\mathbb{E}[\frac{U}{R_U}|R_1=r]=H(r).$$
Consequently,
$$\mathbb{E}\big[\frac{R_U}{R_1^2}\big]=\mathbb{E}[\frac{U}{R_UR_1^2}]=\sqrt{\frac{2}{\pi}}.$$
\end{lemma}
\noindent Remark that we already proved the equality
$$\mathbb{E}\big[\frac{R_U}{R_1^2}\big]=\sqrt{\frac{2}{\pi}}$$
in \cite{elie2013expectation}. This was in fact the cornerstone of our first proof of the centering property of $\alpha$. In the enlargement of filtration approach used here, we will see that instead of $R_U/R_1^2$, the random variable $U/(R_UR_1^2)$ appears naturally.
\subsection{Proofs of Lemma \ref{lem1} and Lemma \ref{lem2}}
We now give the proofs of Lemma \ref{lem1} and Lemma \ref{lem2}.
\subsubsection{Proof of Lemma \ref{lem1}}
The first part of Lemma \ref{lem1} follows from the identity
\begin{equation}\label{returnbes}
\mathbb{E}[\frac{R_t}{t}|R_1]=R_1+\mathbb{E}[\int_t^1\frac{dv}{vR_v}|R_1],~t\leq 1,
\end{equation}
after multiplying both sides by $t$ and integrating in $t$ from $0$ to $1$. To show \eqref{returnbes}, we use time inversion with $t=1/w$ and $$R'_w=wR_{1/w},$$ another three dimensional Bessel process. With this notation, using the Ito representation of the Bessel process, we get
$$\mathbb{E}[R'_w-R'_1|R'_1]=\mathbb{E}[\int_1^w\frac{dt}{R'_t}|R'_1],$$ from which \eqref{returnbes} is easily obtained. The second statement in Lemma \ref{lem1} readily follows.
\subsubsection{Proof of Lemma \ref{lem2}}
We start with the proof of the first part of Lemma \ref{lem2}. Let
$$\rho=\mathbb{E}[\int_0^1dv\frac{v}{R_v}|R_1=r].$$
Using the same time inversion trick as for the proof of Lemma \ref{lem1} together with the Markov property, we get
$$\rho=\int_1^{+\infty}\frac{dw}{w^2}\mathbb{E}[\frac{1}{R_w'}|R_1'=r]=\int_0^{+\infty}\frac{dt}{(1+t)^2}\mathbb{E}_r[\frac{1}{R'_t}],$$
where $\mathbb{P}_r$ denotes the law of a Bessel process $R'$ starting from $r$. We then use the Doob's absolute continuity relationship, that is
$$\mathbb{P}_r\big|_{\mathcal{F}_t}=\frac{X_{t\wedge T_0}}{r}W_r\big|_{\mathcal{F}_t},$$
where $W_r$ is the Wiener measure associated to a Brownian motion starting at point $r$, $X$ is the canonical process and $T_0$
is the first hitting time of $0$ by $X$, see for example \cite{revuz1999continuous}, Chapter XI. This together with the fact that
$$\frac{X_{t\wedge T_0}}{X_t}=\mathrm{1}_{T_0>t}$$
gives
$$\mathbb{E}_r[\frac{1}{R'_t}]=\frac{1}{r}W_r[T_0>t]=\frac{1}{r}W_0[T_r>t].$$
Therefore,
$$\rho=\frac{1}{r}\mathbb{E}^{W_0}\big[\int_0^{T_r}\frac{dt}{(1+t)^2}\big]=\frac{1}{r}\mathbb{E}^{W_0}[\frac{T_r}{1+T_r}]=r\mathbb{E}[\frac{1}{N^2+r^2}].$$
Using Equation \eqref{double2}, this is equal to $H(r)$. This ends the proof of the first part of Lemma \ref{lem2}. Using the expression for the density of $R_1$ given in Corollary \ref{cortheo4}, the proof of the second part readily follows remarking that
$$\mathbb{E}\big[\frac{U}{R_UR_1^2}\big]=\sqrt{\frac{2}{\pi}}\int_{0}^{+\infty}dr\int_{r}^{+\infty}dx\text{e}^{-x^2/2}=\sqrt{\frac{2}{\pi}}\int_{0}^{+\infty}dx x\text{e}^{-x^2/2}=\sqrt{\frac{2}{\pi}}.$$
\subsection{An enlargement of filtration approach to the centering property of $\alpha$}
We now revisit the centering property of $\alpha$ through an enlargement of filtration formula. Let $(\mathcal{F}_t)$ denote the filtration of the Brownian motion $(B_t)$ and $(\mathcal{F}'_t)$ the filtration obtained by initially enlarging $(\mathcal{F}_t)$ with $T_1$. It is shown in \cite{jeulin1979grossissement} that $(B_t)$ is a $(\mathcal{F}'_t)$ semi-martingale. More precisely,
\begin{equation}\label{dec}
B_t=\beta_t-\int_0^{t\wedge T_1}\frac{ds}{1-B_s}+\int_0^{t\wedge T_1}ds\frac{1-B_s}{T_1-s},
\end{equation}
where $(\beta_t)$ is a $(\mathcal{F}'_t)$ Brownian motion (in particular it is independent of $T_1$).
Taking expectation on both sides of \eqref{dec} at time $t=UT_1$, we get
$$\mathbb{E}[\alpha]=-\mathbb{E}\big[\frac{1}{\sqrt{T_1}}\int_0^{UT_1}\frac{ds}{1-B_s}\big]+\mathbb{E}\big[\frac{1}{\sqrt{T_1}}\int_0^{UT_1}ds\frac{1-B_s}{T_1-s}\big].$$
Using the change of variable $s=uT_1$ in both integrals, we get
$$\mathbb{E}[\alpha]=-\mathbb{E}\big[\sqrt{T_1}\int_0^{U}\frac{du}{1-B_{uT_1}}\big]+\mathbb{E}\big[\frac{1}{\sqrt{T_1}}\int_0^{U}du\frac{1-B_{uT_1}}{1-u}\big].$$
Since $U$is independent of $B$ and uniformly distributed on $[0,1]$, we get
$$\mathbb{E}[\alpha]=-\mathbb{E}\big[\sqrt{T_1}\int_0^{1}du\frac{(1-u)}{1-B_{uT_1}}\big]+\mathbb{E}\big[\frac{1}{\sqrt{T_1}}\int_0^{1}du(1-B_{uT_1})\big].$$
Thus,
$$2\mathbb{E}[\alpha]=-\mathbb{E}\big[\sqrt{T_1}\int_0^{1}dv\frac{v}{1-B_{T_1(1-v)}}\big]+\mathbb{E}\big[\frac{1}{\sqrt{T_1}}\big].$$
We now use Williams time reversal result:
$$\Big(T_1,\big(1-B_{T_1(1-v)},~v\leq 1\big)\Big)\underset{\mathcal{L}}{=}\Big(\gamma_1,\big(R_{v\gamma_1},~v\leq 1\big)\Big),$$
where
$$\gamma_1=\text{sup}\{s>0,~R_s=1\}.$$
Hence we obtain
$$2\mathbb{E}[\alpha]=-\mathbb{E}\big[\frac{V}{R_{V\gamma_1}/\sqrt{\gamma_1}}\big]+\mathbb{E}\big[\frac{1}{\sqrt{T_1}}\big],$$ with $V$ a uniform random variable on $[0,1]$, independent of $R$. From the absolute continuity relationship between the laws of
$$\big(R_{v\gamma_1}/\sqrt{\gamma_1},~v\leq 1\big)$$ and $(R_{v},~v\leq 1),$ see \cite{biane1987processus}, we get
$$2\mathbb{E}[\alpha]=\sqrt{\frac{2}{\pi}}-\mathbb{E}[\frac{V}{R_VR_1^2}].$$
Hence $\mathbb{E}[\alpha]=0$ if and only if
$$E[\frac{V}{R_VR_1^2}]=\sqrt{\frac{2}{\pi}}.$$ From Lemma \ref{lem2}, the last equality holds. Moreover, it has been shown without the help of our previous results \cite{elie2013expectation,rosenbaum2013law}. Thus, the use of the enlargement formula of \cite{jeulin1979grossissement} provides an alternative proof of the centering property of $\alpha$.
\section{A few words of conclusion}
Together with \cite{elie2013expectation} and \cite{rosenbaum2013law}, this paper is our third work where various aspects of the law of
\begin{equation*}
\alpha=\frac{B_{UT_1}}{\sqrt{T_1}}
\end{equation*} are investigated.
For example, we have considered its centering property, the explicit form of its density, which may be directly deduced from Equation \eqref{law} and Equation \eqref{dens}, and its Mellin transform. In the present paper, starting from the pseudo-Brownian bridge, we obtain some results relative to the Brownian bridge, the Brownian meander and the three dimensional Bessel process. Imhof type relations between these processes allow to go from one to another.
\bibliographystyle{abbrv}
|
\section{Introduction}
The AdS/CFT correspondence is a powerful tool for computing observables in strongly coupled systems with conformal symmetry by mapping them to weakly coupled dual gravitational theories. However, our ability to exploit the correspondence is limited by our ability to compute in the weakly coupled theory itself. For example, on the bulk AdS side, theories of practical use are not only weakly coupled, but also ``well behaved," in the sense that they are effective theories describing the dynamics of only a few fields below some cutoff scale $\Lambda$. The cutoff scale suppresses non-renormalizable operators generated when fields above the cutoff scale are integrated out.
This leads to the line of enquiry: what is the class of CFTs that we can explore by mapping them to weakly coupled, well-behaved AdS duals? Put another way, what are the necessary and sufficient conditions needed for a CFT to have a weakly coupled, well behaved AdS dual?
Explorations along those lines gave rise to the idea of Effective Conformal Theories (ECT) \cite{Fitzpatrick:2010zm}. The idea of ECTs is that the strongly coupled CFTs that can be described through weakly coupled, effective AdS bulk theories are characterized by two conditions: (1) There is a large dimension gap in the spectrum of the dilatation operator. (2) There is a small parameter that suppresses higher point connected correlation functions. These conditions are naturally satisfied in large-$N$ models where $1/N$ plays the role of the small parameter.
In this paper we explore one consequence of such effective conformal descriptions. Assuming that such an effective description is valid for a strongly coupled condensed matter system with non-Abelian global symmetry, the three-point current correlation function $\langle J^a_i(t_1,x)J^b_j(t_2,y)J^c_k(0)\rangle$ admits a perturbative expansion in the parameter $\Delta = (\Lambda R_{AdS})$. The successive terms in the series carry different polarization structures. In the bulk effective AdS, the dominant contribution to the three point current correlation function comes from the renormalizable (for $d\ge4$) bulk operator $(F_{\mu\nu})^2$. The second contribution comes from a non-renormalizable $(F_{\mu\nu})^3$ operator. In this paper we will refer to these two operators as $F^2$ and $F^3$ respectively. The latter operator is suppressed by the mass scale $\Lambda$. The suppression in the boundary dual is by the parameter $\Delta = \Lambda R_{AdS}$. We will show that generally the $F^3$ operator leads to a different polarization structure for the three-point current correlation function. This difference can be exploited to experimentally measure the expansion parameter $\Delta$ through the framework of ECTs.
The outline of the paper is as follows. An overview of ECTs is given in Section 2. In Section 3, we give the derivation of the contribution of the bulk $F^3$ term to the boundary three-point current correlation function. This contribution is compared to the dominant contribution coming from $F^2$ term, which is computed in \cite{Freedman:1998tz}. Generalizations of the conformal tensors $D_{ijk}(x,y,z)$ and $C_{ijk}(x,y,z)$ used in $d=4$ dimensions in \cite{Freedman:1998tz} is given to general $d>2$ dimensions. In Section 4, we will outline a possible experimental measurement that can be performed to test the validity of ECT for condensed matter systems.
\section{Effective Conformal Thoeries}
We begin with the question, ``what are the necessary and sufficient conditions needed for a CFT to have a weakly coupled, well behaved AdS dual?"
The necessary conditions were first motivated by locality considerations in type IIB string theory on $AdS_5\times S^5$/$\m{N}=4$ SYM. The regime where the 10D supergravity is a good description (i.e., the regime where there is an approimate 10D flat spacetime in the neighborhood of every point) requires the mass of string excitations, of order inverse string length $l_s^{-1}$, to be hierarchically larger than those of the supergravity modes of order inverse AdS length $R_{AdS}^{-1}$ \cite{Aharony:1999ti}. At energies much smaller than $l_s^{-1}$ the theory will look like a local field theory. Since $R_{AdS}=\lambda^{1/4}l_s$, where the 't Hooft coupling $\lambda = g_{YM}^2N$, the condition that $R_{AdS} >> l_s$ implies that the 't Hooft coupling must be large, $\lambda >> 1$. Applying S-duality, which maps type IIB string theory to itself under $g_s \rightarrow g'_s = 1/g_s$, and demanding that string modes remain heavy in the S-dual of the type IIB, we find another condition. Under S-duality,
\begin{eqnarray}
1<< \lambda & = & g_{YM}^2N \\
& \underrightarrow{\hbox{\tiny{S-duality}}} & \lambda' = g_{YM}'^2N = \frac{1}{g_{YM}^2}N = \frac{N^2}{\lambda}
\end{eqnarray}
The requirement that string modes should remain heavy in both sides of the duality is the statement that both $\lambda>>1$ and $\lambda' >> 1$. We find the simultaneous requirements that $\lambda>>1$ and $N^2/\lambda>>1$, which are satisfied for $N^2>>\lambda$, i.e, $N$ very large. But since $R_{AdS}/l_p \sim N^{1/4}$, where $l_{p}$ is the Planck length, $N >> 1$ implies that $R_{AdS} >> l_{p}$ as well. Then we can ignore supergravity quantum corrections and consider classical or tree level supergravity.
Therefore, the gravitational bulk theory is an effective field theory with a large mass gap between the fields of mass of order $R_{AdS}^{-1}$ and high mass string and quantum gravitational excitations with masses of order $l_s^{-1}$ and $l_{p}^{-1}$ respectively. The effective theory has a perturbative expansion in the inverse mass gaps which suppress non-renormalizable interactions. In particular, gravitational interactions are suppressed by powers of $M_{p}^{-1}$, so we can ignore graviton exchanges.
In the dual $\m{N}=4$ Super Yang-Mills theory, the large mass gap in the effective AdS translates to a large gap in operator dimensions. Further, the conformal theory has an expansion in $1/N$, since $N$ is large. This is what mirrors the suppression by factors of $M_p^{-1}$ of gravitational interactions in the AdS bulk. The $1/N$ expansion suppresses higher point connected correlation functions compared to two point functions. Based on this result, Heemskerk, Penedones, Polchiniski, and Sully \cite{Heemskerk:2009pn} put forward the conjecture that any CFT with a large-$N$ like expansion and large gap in the operator dimensions has a local bulk dual AdS theory \footnote{ We also need all single trace operators of spin greater than two to have large dimensions since there is no known local bulk theory of particles of spin greater than 2}. The large $N$ - like expansion parameter is needed to suppress higher point connected functions compared to two point ones, which in the bulk dual corresponds to suppression of gravitational interactions. Fitzpatrick and Kaplan \cite{Fitzpatrick:2012cg} have shown that with the added condition that the Mellin amplitudes of the CFT correlators have an effective theory-type expansion, we obtain the full set of necessary and sufficient conditions for a CFT to have a well behaved weakly coupled bulk AdS dual.
The picture we obtain is that the weakly coupled, well-behaved AdS duals have a double expansion in $l_s^{-1}$, and $l_p^{-1}$. The question is, what do these expansions correspond to on the CFT side? From the above paragraphs it is clear that one of these expansions is a $1/N$ expansion which suppresses higher point connected correlation functions. But what does the expansion in the inverse dimension gap imply? Is there a concept of ``Effective Conformal Theory (ECT)?" that describes the dynamics of operators whose dimension lies below the cutoff dimension? If so, how does such a theory distinguish between ``renormalizable" vs ``non-renormalizable" interactions? What suppresses the ``non-renormalizable" operators (since conformal symmetry means that there are no mass scales)? What conditions set the range of validity for such an effective conformal theory, and where does it break down?
To address these questions, Fitzpatrick, Katz, Poland and Simmons-Duffin \cite{Fitzpatrick:2010zm} identified these two expansions with those involving a large parameter $N$ and a large dimension gap $\Delta_{gap} = \Delta_{Heavy} - \Delta_{low}$. Such a theory is an effective conformal theory that captures the dynamics of the low-lying spectrum of the dilatation operator. Let $\Delta_{low}$ be the typical dimension of the low-lying operators, and let all other primary operators have dimension above $\Delta_{Heavy}$ which is hierarchically larger. Then there is a perturbative expansion in both $1/\Delta_{Heavy}$ \cite{Sundrum:2011ic} and $1/N$. The $1/N$ suppresses all interactions, and the $1/\Delta_{Heavy}$ suppresses higher dimensional operators in the OPE.
There is a direct parallel with effective quantum field theories. In that familiar context, there is an expansion in the small coupling constant of the effective QFT in addition to an expansion in $1/M$, where $M$ is the scale where the effective QFT begins to break down. Analogously, in effective CFTs, the large $N$ (playing the role of the small coupling constant) ensures that connected pieces of higher point correlation functions are suppressed compared to two-point functions, whereas the small $\Delta^{-1}$ (playing the role of small $M^{-1}$ in QFTs) suppresses contributions of higher dimensional operators to the correlation function.
The schematic picture obtained is therefore the following. The dilatation operator of the CFT has a perturbative expansions in both $1/N$ and $1/\Delta_{Heavy}$:
\bea
D^{eff} = D^0 + \frac{1}{N}\left( V^{(1)} + \frac{1}{\Delta_{Heavy}}V^{(2)} + \dots \right) + \mathcal{O}(\frac{1}{N^2}),
\end{eqnarray}
where $D^0$ is the mean field dilatation operator and $V^{(1)}, V^{(2)}, \dots$ are perturbations of the dilatation that preserve conformal symmetry.
The next question is then, ``what sets the range of validity of the effective description?" The answer is again analogous to the situation in effective field theories where imposing perturbative unitarity on the Hamiltonian sets the range of validity of the effective theory. In our case, perturbative unitarity is imposed on the dilatation operator \cite{Fitzpatrick:2010zm}. Assume $\m{O}$ is the only single trace primary operator below the cutoff dimension $\Delta_{Heavy}$. Then the low dimensional spectrum of the dilatation consists of double trace primary operators of the type $\m{O}_{n,l}=\m{O}(\partial^2)^n(\partial)^l\m{O}$. These operators receive an order $1/N$ correction to their dimension coming from the $V^{(1)}$ term; $\Delta_{n,l} = 2\Delta + 2n+l+\frac{1}{N}\gamma(n,l)$. Imposing perturbative unitarity gives a bound $|\gamma(n,l)| < 4$ on the anomalous dimension $\gamma(n,l)$. However, operators $V^{(1)}$ dual to bulk interactions of mass (or scaling dimension) $\Lambda^p$ (hence forth refered to as ``non-renormalizable" operators) lead to growth in $\gamma(n,l)$ as $n^{p-(d+1)}$ \cite{Heemskerk:2009pn, Fitzpatrick:2010zm}. Even though $\gamma(n,l)$ is an $O(1/N)$ correction, it leads to violation of the unitarity bound for $p>d+1$ and sufficiently large $n$ no matter how small $1/N$ may be. As $n$ approaches $\Delta_{Heavy}$, the new operators must be integrated in to moderate the growth of $\gamma(n,l)$ and restore unitarity. This will indeed be the case if the non-renormalizable operators $V$ of dimension $p$ are suppressed by $\Delta_{Heavy}^{p-(d+1)}$. In this case, $\gamma(n,l)$ grows as $\left(n/\Delta_{Heavy}\right)^{p-(d+1)}$, the unitarity bound is satisfied as long as $n < \Delta_{Heavy}$, and the ECT breaks down when $n \sim \Delta_{Heavy}$.
This idea to use perturbative unitarity as the condition to set the range of validity of the effective description was suggested by the authors of \cite{Fitzpatrick:2010zm} as a solution to the observation made in \cite{Hofman:2008ar} that in correlation functions involving conserved currents, only certain polarization structures, those arising from the lowest dimension bulk operators appear. Demanding perturbative unitarity on all operators below the cutoff dimension $\Delta < \Delta_{Heavy}$ translates to demanding that the scale suppressing non-renormalizable operators in the bulk satisfy $\Lambda > \left(\Delta_{Heavy}/R_{AdS}\right)$. By explicitly computing the contribution of the bulk operator $F^3$ to the three-point current correlation function, we will show that, in addition to giving a polarization structure different from that of the $F^2$, the contribution is suppressed by the appropriate power of the cutoff dimension $\Delta_{Heavy}$.
\section{Three-point Current Correlation Function}
Armed with the above perturbative expansion, we can compute the three-point current correlation function resulting from the operator $F^3$ and compare the result to the contribution of the $F^2$ operator. It is important to note here that the system remains conformally invariant in the presence of the non-renormalizable $F^3$ operator. This is guaranteed by the fact that in the bulk $AdS$ the operator is invariant under the $AdS$ isometry. The theory we are describing thus models movement along a line of second order phase transition of a system with non-Abelian global symmetry. The movement is parameterized by $\Lambda R_{AdS}$.
We begin with the bulk action
\bea
\label{bulkaction}
S &=& \frac{1}{g^2_{SG}} \int d^{d+1} x \sqrt{g} \left( \frac{1}{2}F^2 + \Lambda^{-p}F^3 \right).
\end{eqnarray}
$g_{SG}$ is the gauge coupling constant for the bulk $AdS$ Yang-Mills theory. $\Lambda$ has mass dimension $+1$. The explicit form of the operator $F^3$ that we will be using is
\bea
F^3 = f^{abc}F_{\mu\alpha}^aF_{\nu\beta}^bF_{\rho\gamma}^c g^{\alpha \nu}g^{\beta \rho}g^{\gamma \mu}.
\end{eqnarray}
Note however that we do not need non-Abelian global symmetry to get $F^3$ term. If there are three $U(1)$ global currents in the boundary CFT, we will get bulk interaction terms of the form $F_{\mu\alpha}F^{\alpha\nu}F_{\nu}^{\mu}$. However, in this case there are no renormalizable bulk interactions that contribute to the three-point current correlation function, the first non-vanishing contribution being the $F^3$.
Through out this paper we will be working in Euclidean $AdS$ and have rescaled the gauge fields so that $A_{\mu}\to (i/g_{SG})A_{\mu}, F_{\mu\nu} \to (i/g_{SG})F_{\mu\nu}.$ Further, the gauge group generators have the commutation relation $[T^a,T^b] = f^{abc}T^c$. With these modifications we have $$F_{\mu\nu} = \partial_{\mu}A_{\nu} - \partial_{\nu}A_{\mu} + [A_{\mu}, A_{\nu}].$$
Dimensional analysis gives the following mass dimensions:
\bea
\left[ g_{SG} \right] &=& \frac{3-d}{2}\nonumber\\
\left [F \right] &=& 2\nonumber\\
p &=& 2 \nonumber
\end{eqnarray}
Let us write the action as $S = S_2 + S_3$ where $S_2$ is the $F^2$ integral and $S_3$ is the $F^3$. We study contributions to the three-point current correlation function $\left \langle J_i^a(x)J_j^b(y)J_k^c(z) \right \rangle$ coming from each of the actions $S_2$ and $S_3$. $i,j,k$ are $d-$dimensional Euclidean spacetime indices and $a,b,c$ label global current indices. The points $x,y,z \in \mathbb{R}^{d}$ are points in $d$-dimensional Euclidean spacetime. In this paper we adapt the notation of \cite{Freedman:1998tz}, where the contribution of $S_2$ has been computed. Let us first begin with a review of the conformal structures of the two and three-point current correlation functions.
\subsection{Review of conformal structures}
The two-point current correlation function in $d-$dimensions is fully determined by conformal invariance up to a normalization constant. It is given by
\bea
\label{2point}
\left \langle J_i^a(x)J_i^b(y)\right \rangle &=& B\delta^{ab}\frac{2(d-1)(d-2)}{(2\pi)^d}\frac{J_{ij}(x-y)}{|x-y|^{2(d-1)}}.
\end{eqnarray}
$B$ is a positive constant and $$J_{ij}(x) = \delta_{ij} - 2\frac{x_ix_j}{x^2}.$$
The coefficient $B$ is computed from the bulk $F^2$ term in \cite{Freedman:1998tz},
\bea
B = \frac{1}{g^2_{SG}}\frac{2^{d-2}\pi^{\frac{d}{2}}\Gamma(d)}{(d-1)\Gamma(\frac{d}{2})}.
\end{eqnarray}
The three-point current correlation function is also determined completely by conformal symmetry up to two constants. In $d=4$ dimensions, the normal parity three-point function is given as the superposition of two permutation-odd conformal tensor structures, $D^{sym}_{ijk}, C^{sym}_{ijk}$ \cite{Freedman:1992tz}.
\bea
\left \langle J_i^a(x)J_j^b(y)J_k^c(z) \right \rangle_{+} &=& f^{abc}\left( k_1 D_{ijk}^{sym} + k_2 C_{ijk}^{sym}\right)
\end{eqnarray}
where
\bea
D^{sym}_{ijk}(x,y,z) &=& D_{ijk}(x,y,z)+D_{ijk}(z,x,y)+D_{ijk}(y,z,x)\\
C^{sym}_{ijk}(x,y,z) &=& C_{ijk}(x,y,z)+C_{ijk}(z,x,y)+C_{ijk}(y,z,x)\\
\end{eqnarray}
The tensors $D_{ijk}(x,y,z),$ and $C_{ijk}(x,y,z)$ are given by
\bea
D_{ijk}(x,y,z) &=& \frac{1}{(x-y)^2(y-z)^2(z-x)^2}\frac{\partial}{\partial x_{i}}\frac{\partial}{\partial y_{j}}\ln\left( (x-y)^2\right)\frac{\partial}{\partial z_{k}}\ln\left( \frac{(x-z)^2}{(y-z)^2}\right)\\
&=& \frac{4}{(x-y)^2(y-z)^2(z-x)^2}J_{ij}(x-y)\frac{\tilde{t}_{k}}{(x-y)^2}
\end{eqnarray}
\bea
C_{ijk}(x,y,z) &=& \frac{1}{(x-y)^4}\frac{\partial}{\partial x_{i}}\frac{\partial}{\partial z_{l}}\ln\left( (x-z)^2\right)\frac{\partial}{\partial y_{j}}\frac{\partial}{\partial z_{l}}\ln\left( (y-z)^2\right)\frac{\partial}{\partial z_{k}}\ln\left( \frac{(x-z)^2}{(y-z)^2}\right)\\
&=& \frac{-8}{(x-y)^2(y-z)^2(z-x)^2}J_{il}(x-z)J_{jl}(y-z)\frac{\tilde{t}_{k}}{(x-y)^2},
\end{eqnarray}
where,
\bea
\tilde{t}_k = \frac{(x-z)_k}{(x-z)^2}-\frac{(y-z)_k}{(y-z)^2},\quad t_k = \frac{(y-x)_k}{(y-x)^2}-\frac{(z-x)_k}{(z-x)^2},\quad\hat{t}_k = \frac{(z-y)_k}{(z-y)^2}-\frac{(x-y)_k}{(x-y)^2}.
\end{eqnarray}
The vectors $t$ and $\hat{t}$ are introduced here for later convenience since they appear in the symmetric sums of $D_{ijk}$, and $C_{ijk}$.
In $d=4, C^{sym}_{ijk}$ satisfies $\frac{\partial}{\partial z_k}C^{sym}_{ijk} = 0$ everywhere, whereas $D^{sym}_{ijk}$ has terms proportional to $\delta^4(z-x)$ and $\delta^4(z-y)$. Therefore, the Ward identity in $d=4$ relates the coefficient $k_1$ to the coefficient $B$ in (\ref{2point}) as
\bea
\label{d4wardid}
k_1 = \frac{B}{16\pi^6}.
\end{eqnarray}
The coefficient $k_2$ is undetermined.
The contribution to the three-point function coming from the bulk action $S_2$ is calculated for general $d$ in \cite{Freedman:1998tz}.
\bea
\label{fijkS2}
\left\langle J^a_i(x)J^b_j(y)J^c_k(z)\right\rangle^{S_2}
&=&\frac{f^{abc}}{2g^2_{SG}\pi^4} 2\left[ \mathcal{F}^{(2)}_{ijk}(x,y,z)+\mathcal{F}^{(2)}_{kij}(z,x,y)+ \mathcal{F}^{(2)}_{jki}(y,z,x)\right],\\
\mathcal{F}^{(2)}_{ijk}(x,y,z) &=& -\kappa \frac{J_{jl}(y-x)}{|y-x|^{2(d-1)}}\frac{J_{km}(z-x)}{|z-x|^{2(d-1)}}\nonumber\\
&&\times\frac{1}{|t|^d}\left[ \delta_{lm}t_i+(d-1)\delta_{il}t_m+(d-1)\delta_{im}t_l-d\frac{t_it_lt_m}{|t|^2}\right]\nonumber
\end{eqnarray}
where,
\bea
\kappa
&=& \pi^{d/2}(C^d)^3\frac{(d-2)}{(d-1)}\frac{\big[ \Gamma(\frac{d}{2})\big]^3}{\big[ \Gamma(d)\big]^2}, \quad C^d = \frac{\Gamma(d)}{2\pi^{d/2}\Gamma(\frac{d}{2})}\nonumber
\end{eqnarray}
In terms of the conformal tensors $D^{sym}_{ijk}, C^{sym}_{ijk}$, the above result takes the elegant form
\bea
\label{jjjS2}
\left\langle J^a_i(x)J^b_j(y)J^c_k(z)\right\rangle^{S_2}
&=&\frac{f^{abc}}{2g^2_{SG}\pi^4}\left( D^{sym}_{ijk}-\frac{1}{8}C^{sym}_{ijk}\right).
\end{eqnarray}
Let us digress here to comment on the comparison between this bulk result for the lowest renormalizable operator $F^2$ in $d=4,$ with the 1-loop exact two and three-point correlation function in the boundary $\mathcal{N}=4$ super-Yang-Mills theory. With the replacement $4\pi/N \rightarrow g_{SG}$ we find that both the two-point and three-point correlation functions agree exactly.
In the two-point function, from the boundary super-Yang-Mills perspective, there are no higher order corrections than the 1-loop result because of powerful non-renormalization theorems \cite{Anselmi:1997am}. But on the bulk side, we would expect that bulk operators of the form
\bea
\label{nonrenorm}
\sum_n \frac{1}{g^2_{SG}\Lambda^{2n}}\left( (\partial_{\rho}\partial^{\rho})^n F_{\mu\nu}F^{\mu\nu} + \partial_{\mu_1}\partial_{\mu_2}\dots\partial_{\mu_n}F_{\mu\nu}\partial^{\mu_1}\partial^{\mu_2}\dots\partial^{\mu_n}F^{\mu\nu} \right)
\end{eqnarray}
would lead to contributions. These are all operators of the same order in $1/N$ expansion compared to the leading $F^2$ term. Since supergravity is an effective theory that starts to break down when we get near the string scale, we will in fact have the above non-renormalizable operators below the string scale. It must then be the case that the $\m{N}=1$ supergravity of the $AdS_5\times S^5$ is responsible for the vanishing all such contributions
If we remove supersymmetry from both sides of the duality, non-renormalizable operators of the form (\ref{nonrenorm}) will lead to corrections to $B$. Similar corrections arise for the three-point function. The claim of \cite{Fitzpatrick:2010zm} is that effective bulk theories where non-renormalizable operators of the form (\ref{nonrenorm}) are suppressed by appropriate mass scales are dual to effective conformal theories where perturbative unitarity is imposed on the dilatation operator. By computing the contribution of the $S_3$ action to the three-point correlation function, we will demonstrate that contributions to $k_1$ and $k_2$ coming from the non-renormalizable bulk operator $F^3$ will be suppressed by $\Delta_{gap}^2 = (R_{AdS}\Lambda_{cutoff})^2$ as required by perturbative unitarity on the dilatation on the CFT side. In addition, we will see that the contribution of the $F^3$ operator has different polarization structure, which could be exploited to experimentally measure the suppression parameter $\Delta$.
\subsection{Generalization in $d>2$}
In $d > 2$ dimensions, the symmetric tensor $J_{ij}$ which appears in the two-point function in (\ref{2point}) remains the same since it comes from general requirements of covariance under the conformal algebra \cite{Osborn:1993cr}.
The tensors $D_{ijk}(x,y,z),$ and $C_{ijk}(x,y,z)$ can be generalize as follows.
\bea
D_{ijk}(x,y,z) &=& \frac{1}{\big( |x-y||y-z||z-x|\big)^{d-2}}\frac{\partial}{\partial x_{i}}\frac{\partial}{\partial y_{j}}\ln\left( |x-y|^{d-2}\right)\frac{\partial}{\partial z_{k}}\ln\left( \frac{|x-z|^{d-2}}{|y-z|^{d-2}}\right)\\
&=& \frac{(d-2)^2}{\big(|x-y||y-z||z-x|\big)^{d-2}}J_{ij}(x-y)\frac{\tilde{t}_{k}}{|x-y|^2}\\
\end{eqnarray}
\bea
C_{ijk}(x,y,z) &=& \frac{1}{|x-y|^d}\frac{\partial}{\partial x_{i}}\frac{\partial}{\partial z_{l}}\ln\left( |x-z|^{d-2}\right)\frac{\partial}{\partial y_{j}}\frac{\partial}{\partial z_{l}}\ln\left( |y-z|^{d-2}\right)\frac{\partial}{\partial z_{k}}\ln\left( \frac{|x-z|^{d-2}}{|y-z|^{d-2}}\right)\\
&=& \frac{-(d-2)^3}{\big(|x-y||y-z||z-x|\big)^{d-2}}J_{il}(x-z)J_{jl}(y-z)\frac{\tilde{t}_{k}}{|x-y|^2},
\end{eqnarray}
The symmetric sums of the tensors, $D^{sym}_{ijk}, C^{sym}_{ijk}$ have the following property:
\bea
\frac{\partial}{\partial z_{k}}D^{sym}_{ijk}&=&(d-2)^2 S_d\left(\frac{d+2}{d}\right)\frac{J_{ij}(x-y)}{|x-y|^{2(d-1)}}\Big(\delta^d(z-y) - \delta^d(z-x) \Big)\nonumber\\
\frac{\partial}{\partial z_{k}}C^{sym}_{ijk}&=&-(d-2)^3S_d\left( \frac{d-4}{d}\right)\frac{J_{ij}(x-y)}{|x-y|^{2(d-1)}}\Big(\delta^d(z-y) - \delta^d(z-x) \Big),
\end{eqnarray}
where, $$S_d = \frac{2\pi^{\frac{d}{2}}}{\Gamma(\frac{d}{2})}.$$ We have used the following formulae to derive the above result:
\bea
\lim_{x\to 0} \frac{x_ix_j}{x^2} = \frac{1}{d}\delta_{ij}, \qquad \lim_{z\to x}\frac{\partial}{\partial z_k}\left( \frac{(z-x)_k}{|z-x|^d}\right) = S_d \delta^d(z-x).
\end{eqnarray}
The Ward identity in $d-$dimensions relates one linear combination of $k_1$ and $k_2$ to $B$.
\bea
\label{ddwardid}
B &=& \frac{(2\pi)^dS_d}{2}\frac{(d-2)}{(d-1)} \left( \frac{(d+2)}{d}k_1 -\frac{(d-2)(d-4)}{d}k_2\right).
\end{eqnarray}
In $d=4$ we recover (\ref{d4wardid}).
To compare the contribution of the $F^3$ operator to the three-point function with that coming from the $F^2$ operator in general $d > 2$ dimensions, it is helpful to find an expression to (\ref{fijkS2}) analogous to (\ref{jjjS2}) for general $d>2$ dimensions. This can be achieved using the formulae
\bea
J_{km}(z-x)t_m &=& -\frac{(y-z)^2}{(y-x)^2}\tilde{t}_k, \; \text{and}\nonumber\\
J_{jl}(y-x)t_l &=& -\frac{(z-y)^2}{(z-x)^2}\hat{t}
\end{eqnarray}
We then find
\bea
\left\langle J^a_i(x)J^b_j(y)J^c_k(z)\right\rangle^{S_2}
&=&\frac{f^{abc}\kappa}{2g^2_{SG}}\frac{(3d-4)}{(d-2)^2}\left( D^{sym}_{ijk}-\frac{1}{(3d-4)}C^{sym}_{ijk}\right)
\end{eqnarray}
\subsection{Contribution of the $F^3$ operator}
From the AdS/CFT ansatz for correlation functions \cite{Witten:1998qj}, we have
\bea
\left \langle \exp \int J^a_iA_0^{ai} \right \rangle_{CFT} &=& Z_S(A_0)
\end{eqnarray}
where $Z_S(A_0)$ is the bulk path integral for the gauge field $A(x_0,x)$ expressed in terms of the boundary value $A_0(x)$. In the limit where the bulk gravitational theory is weakly coupled, the path integral is approximately the classical path integral,
\[
Z_S(A_0) \simeq \exp (-I_s(A_0)),
\]
where $I_s(A_0)$ is the action expressed in terms of the boundary value of the field $A$ at boundary coordinates, $x,y,z$. In the following, Latin indices i,j,k run from 1 to $d$, and Greek letters $\mu,\nu$ run from 0 to $d$, where 0 is the extra $AdS$ coordinate.
We are interested in the connected three point correlator,
\bea
\label{connected3pt}
\left \langle J_i^a(x)J_j^b(y)J_k^c(z) \right \rangle_{connected}
&=&
\frac{\delta}{\delta A_0^{ai}(x)} \frac{\delta}{\delta A_0^{bj}(y)} \frac{\delta}{\delta A_0^{ck}(z)} \log(Z_S(A_0))\nonumber\\
&=&
\frac{\delta}{\delta A_0^{ai}(x)} \frac{\delta}{\delta A_0^{bj}(y)} \frac{\delta}{\delta A_0^{ck}(z)}(-I_s(A_0)) \arrowvert_{A_0 = 0}
\end{eqnarray}
To compute the contribution of the $F^3$ operator, we begin by expressing the $S_3$ part of the action in terms of the boundary value of the gauge field and the boundary-to-bulk Greens function $G^{ab}_{\mu i}(w_0,x;0,\tilde{x})$, where $x,\tilde{x}$ are the $d$-dimensional boundary coordinates and $w_0$ is the perpendicular bulk coordinate.
\bea
A_{\mu}^a(w_0, \tilde{x})&=&
\int d^dxG^{ab}_{\mu i}(w_0,\tilde{x};0,x)A_0^{ib}(0,x), \hspace{3mm} \hbox{where}
G^{ab}_{\mu i} = G_{\mu i}\delta^{ab} \hspace{3mm} \hbox{and so}\nonumber\\
A_{\mu}^a(w_0, \tilde{x})&=&
\int d^dxG_{\mu i}(w_0,\tilde{x};0,x)A_0^{ia}(0,x)
\end{eqnarray}
Plugging this into the $S_3$ part of the bulk action in (\ref{bulkaction}) and evaluating (\ref{connected3pt}) we find the following expression.
\bea
\label{JJJ}
\left \langle J_i^a(x)J_j^b(y)J_k^c(z) \right \rangle^{S_3}_{connected}
&=&\frac{\delta}{\delta A_0^{ai}(x)}\frac{\delta}{\delta A_0^{bj}(y)} \frac{\delta}{\delta A_0^{ck}(z)}
(-S_3)\arrowvert_{A_0 = 0}\nonumber\\
&=&\frac{1}{\Lambda^p g^2_{SG}}2f^{abc}[\mathcal{F}^{(3)}_{ijk}+\mathcal{F}^{(3)}_{jki}+\mathcal{F}^{(3)}_{kij}],
\end{eqnarray}
where
\bea
\label{f3ijk}
\mathcal{F}^{(3)}_{ijk}
&=&\int d^{d+1}w\sqrt{g}\hspace{1mm}\partial_{[ \mu}G_{\alpha ]i}(w,x)\partial_{[ \nu}G_{\beta ]j}(w,y)
\partial_{[ \rho}G_{\gamma ]k}(w,z)g^{\alpha \nu}g^{\beta \rho}g^{\gamma \mu}.
\end{eqnarray}
We evaluate $\mathcal{F}^{(3)}_{ijk}$ in Euclidean $AdS$, in the parameterization of $AdS$ as the Lobachevsky upper half space with the metric
\bea
\label{metric}
ds^2 = \frac{R_{AdS}^2}{w_0^2}\left( dw_0^2+\sum_{\mu = 1}^d dx_{\mu}^2\right).
\end{eqnarray}
We set $R_{AdS} = 1$ in the following computation and restore it in the final answer by dimensional analysis.
The boundary-to-bulk propagator of the gauge field from the boundary point $x^{\mu}=(0,x)^{\mu}$ to the bulk point $w^{\mu}=(w_0,\tilde{x})^{\mu}$ is given explicitly in \cite{Freedman:1998tz}
\bea
\label{gaugepropagator}
G_{\mu i}(w_0,\tilde{x};0,x)
&=&C^d\frac{w_0^{d-2}}{[w_0^2+(\tilde{x}-x)^2]^{d-1}}J_{\mu i}(w-x).
\end{eqnarray}
We will use the technique described by Freedman, Mathur, Matusis, and Rastelli \cite{Freedman:1998tz} to evaluate $\mathcal{F}^{(3)}_{ijk}.$ Their technique takes advantage of the fact that the Green function has translation invariance in the boundary coordinates.
\[
\left \langle J_i^a(x)J_j^b(y)J_k^c(z) \right \rangle = \left \langle J_i^a(0)J_j^b(y-x)
J_k^c(z-x) \right \rangle
\]
Evaluating $\langle J(0)J(y-x)J(z-x) \rangle$ is easier because there are only two terms in the denominator of (\ref{f3ijk}). We begin by calculating $\left \langle J_i^a(0)J_j^b(y)J_k^c(z) \right \rangle.$ Using the metric (\ref{metric}) in the formula for $\mathcal{F}_{ijk}$ we find,
\bea
\label{Fijk}
\mathcal{F}^{(3)}_{ijk}
&=&\int d^dx' dw_0\frac{w_0^6}{w_0^{d+1}}\partial_{[ \mu}G_{\nu ]i}(x',0)\partial_{[ \nu}G_{\rho ]j}(x',y)\partial_{[ \rho}G_{\mu ]k}
(x',z)
\end{eqnarray}
To simplify the above integral further we will take advantage of the inversion isometry of the $AdS$ metric. The transformation
\bea
\label{inversion}
w_0 = \frac{w'_0}{w'^2_0+x'^2}, \quad x^{\mu} = \frac{x'^{\mu}}{w'^2_0+x'^2}
\end{eqnarray}
on the AdS coordinates leaves the metric (\ref{metric}) invariant. On the other hand, such a transformation acts as conformal isometry on the boundary coordinates; the flat boundary metric $ds^2 = \sum_i dx^idx^i \rightarrow \frac{1}{|x|^4}\sum_i dx^idx^i$ under
\bea
\label{boundaryinversion}
x^i &=& \frac{x'^i}{x'^2}.
\end{eqnarray}
The Jacobian of the inversion transformation inherits the tensor structure of $J_{\mu\nu}$
\bea
\frac{\partial w'_{\mu}}{\partial w_{\nu}}&=&w'^2\left( \delta_{\mu \nu}-2\frac{w'_{\mu}w'_{\nu}}{w'^2}\right)\\
&=&w'^2J_{\mu \nu}(w') = \frac{1}{w^2}J_{\mu\nu}(w)
\end{eqnarray}
$J_{\mu \nu}$ satisfies the following identities:
\bea
\label{jmunuproperties}
J_{\mu \nu}(w-u)
&=& J_{\mu \rho}(w')J_{\rho \sigma}(w'-u')J_{\sigma \nu}(u')\\
J_{\mu \nu}(w)J_{\nu \rho}(w)
&=&\delta_{\mu \rho}
\end{eqnarray}
Using these identities and explicit formula for $G_{\mu \nu}$ we can show that it transforms as a covariant rank 2 tensor with scaling dimension $d-2$ under the simultaneous bulk and boundary inversions.
\bea
G_{\mu i}(w_0,\tilde{x};0,x)
&=&C^d\frac{1}{w_0}\left(\frac{w_0}{w_0^2+(\tilde{x}-x)^2}\right)^{d-1}J_{\mu i}(w-x)\nonumber\\
&=&C^d \frac{w'^2}{w'_0}\left(\frac{w'_0}{{w'_0}^2+(\tilde{x}'-x')^2}\right)^{d-1}|x'|^{2(d-1)}J_{\mu \rho}(w')J_{\rho k}(w'-x')J_{ki}(x')\nonumber\\
&=&w'^2J_{\mu \rho}(w')|x'|^2J_{ki}(x') |x'|^{2(d-2)}G_{\mu i}(w',x')\nonumber\\
&=&\frac{\partial w'_{\nu}}{\partial w_{\mu}}\frac{\partial x'_k}{\partial x_i}|x'|^{2(d-2)}G_{\nu k}(w', x')\nonumber\\
&=& \frac{\partial w'_{\nu}}{\partial w_{\mu}}\frac{\partial x'_k}{\partial x_i}G'_{\nu k}(w', x').
\end{eqnarray}
In the second line, and $w'^{\mu} =(w'_0,\tilde{x}')^{\mu} $. Similarly, $\partial_{[\mu}G_{\nu ]i}(w,x)$ transforms covariantly as
\bea
\partial_{[\mu}G_{\nu ]i}(w,x)
&=&w'^2J_{\mu \alpha}(w')w'^2J_{\nu \beta}(w')|x'|^2J_{ik}(x')|x'|^{2(d-2)}\partial'_{[\alpha}G_{\beta]k}
(w',x'),\hspace{2mm}\hbox{where}\\
\partial'
&=&\frac{\partial}{\partial w'}\nonumber.
\end{eqnarray}
When we set $x$ to zero and do an inversion transformation, we find the following simpler forms
\bea
G_{\mu i}(w, 0)
&=& C^d (w'_0)^{d-2}w'^2J_{\mu i}(w')\\
\partial_{[\mu}G_{\nu ]i}(w, 0)
&=&(d-2)C^d(w'_0)^{d-3}(w')^4J_{0[\mu}(w')J_{\nu]i}(w'),\hspace{2mm}
\end{eqnarray}
Applying the inversion on (\ref{Fijk}) and simplifying, we find
\bea
\label{Fijk2}
\mathcal{F}^{(3)}_{ijk}(0,y,z)&=&(d-2)^3(C^d)^3|y|^{2(d-1)}J_{aj}(y')|z|^{2(d-1)}J_{bk}(z) \nonumber\\
&& \int d^dw'dw'_0\frac{(w'_0)^{2d-4}}{[w'^2_0+(\tilde{x}'-y')^2]^{d-1}[w'^2_0+(\tilde{x}'-z')^2]^{d-1}}\nonumber\\
&&\hspace{1.7cm} \bigg(J_{0[i}(w'-y')J_{\gamma]a}(w'-y')J_{0[\gamma}(w'-z')J_{0]b}(w'-z')\nonumber\\
&&\hspace{1.8cm} +J_{0[\gamma}(w'-y')J_{0]a}(w'-y')J_{0[\gamma}(w'-z')J_{i]b}(w'-z')\bigg)
\end{eqnarray}
After performing the integral and expressing the result in terms of the tensors $D_{ijk}, C_{ijk}$, we find the following simple form:
\bea
\mathcal{F}^{(3)}_{ijk}(x,y,z) &=& -\frac{\kappa d}{2}\Big(D_{jki}(y,z,x) + \frac{1}{d}C_{jki}(y,z,x)\Big), \nonumber
\end{eqnarray}
The intermediate steps are included in the appendix. The symmetric sum then becomes
\bea
\mathcal{F}^{(3)sym}_{ijk}(x, y, z)
&=&-\frac{\kappa d}{2}\Big(D^{sym}_{ijk}(y,z,x) + \frac{1}{d}C^{sym}_{ijk}(y,z,x)\Big).
\end{eqnarray}
For comparison, the contribution of the $F^2$ operator to three-point current correlation function, given in (\ref{fijkS2}) is
\bea
\mathcal{F}^{(2)sym}_{ijk}
&=&\frac{\kappa(3d-4)}{2(d-2)^2}\left( D^{sym}_{ijk}(x,y,z)-\frac{1}{(3d-4)}C^{sym}_{ijk}(x,y,z)\right).
\end{eqnarray}
As expected, the polarization structure resulting from the $F^3$ operator is different from the $F^2$ contribution.
After restoring the correct factor of $R_{AdS}$ by dimensional analysis, and letting $R_{AdS}\Lambda = \Delta$, the three-point current contributions of each of the operators $F^2$ and $F^3$ are
\bea
\left\langle J^a_i(x)J^b_j(y)J^c_k(z)\right\rangle^{S_2}
&=&f^{abc}\kappa\left(\frac{(R_{AdS})^{d-3}}{g^2_{SG}}\right)\left(\frac{(3d-4)}{2(d-2)^2}\right)\left( D^{sym}_{ijk}(x,y,z)-\frac{1}{(3d-4)}C^{sym}_{ijk}(x,y,z)\right)\nonumber\\
\left\langle J^a_i(x)J^b_j(y)J^c_k(z)\right\rangle^{S_3}
&=&-f^{abc}\kappa \left(\frac{(R_{AdS})^{d-3}}{\Delta^2g_{SG}^2}\right)d\Big(D^{sym}_{ijk}(y,z,x) + \frac{1}{d}C^{sym}_{ijk}(y,z,x)\Big).
\end{eqnarray}
The sum of the two contributions is,
\bea
\left\langle J^a_i(x)J^b_j(y)J^c_k(z)\right\rangle^{S_2+S_3}
=f^{abc}\kappa\left(\frac{ (R_{AdS})^{d-3}}{g_{SG}^2}\right)\left( \frac{3d-4}{2(d-2)^2}\right)
\hspace{-7mm}&&\Bigg[\bigg(1- \frac{2d(d-2)^2}{(3d-4)\Delta^2}\bigg)D^{sym}_{ijk}\nonumber\\
&&-\frac{1}{3d-4}\bigg(1+\frac{2(d-2)^2}{\Delta^2} \bigg)C^{sym}_{ijk}\Bigg].
\end{eqnarray}
In particular, for $d=3$,
\bea
\mathcal{F}^{(3)sym}_{ijk}
&=&-\frac{1}{2^9}\left( D^{sym}_{ijk}+\frac{1}{3}C^{sym}_{ijk}\right)\\
\mathcal{F}^{(2)sym}_{ijk}
&=&\frac{5}{2^{10}}\left( D^{sym}_{ijk}-\frac{1}{5}C^{sym}_{ijk}\right)\nonumber
\end{eqnarray}
\bea
\left\langle J^a_i(x)J^b_j(y)J^c_k(z)\right\rangle
&=&f^{abc}\left(\frac{5}{2^{10}g_{SG}^2}\right)\left(\big(1- \frac{6}{5\Delta^2}\big)D^{sym}_{ijk}-\frac{1}{5}\big(1+\frac{2}{\Delta^2} \big)C^{sym}_{ijk}\right).
\end{eqnarray}
In $d=4$, the combined three-point current correlation function is
\bea
\mathcal{F}^{(3)sym}_{ijk}
&=&-\frac{1}{\pi^4}\left( D^{sym}_{ijk}+\frac{1}{4}C^{sym}_{ijk}\right)\\
\mathcal{F}^{(2)sym}_{ijk}
&=&\frac{1}{2\pi^4}\left( D^{sym}_{ijk}-\frac{1}{8}C^{sym}_{ijk}\right)\nonumber
\end{eqnarray}
\bea
\left\langle J^a_i(x)J^b_j(y)J^c_k(z)\right\rangle
&=&f^{abc}\left(\frac{ R_{AdS}}{2\pi^4g_{SG}^2}\right)\left(\big(1- \frac{1}{4\Delta^2}\big)D^{sym}_{ijk} -\frac{1}{8}\big(1+\frac{8}{\Delta^2} \big)C^{sym}_{ijk}\right).
\end{eqnarray}
These results give the two lowest order results to the three-point current correlation function in the $1/\Delta$ expansion and leading order in $1/N$ expansion. The first $\mathcal{O}(1/\Delta^2)$ correction to the three-point current correlation function comes from the non-renormalizable $F^3$ operator.
\section{Physical measurement}
Measuring the three-point spin-current in condensed matter systems directly is near impossible through existing technologies. However, measurements that look for non-linear Ohm's-law type effects in induced spin-currents contain data about the three-point current correlation function. In the presence of an external field $\vec{E}$ the induced current will take the form,
\bea
J_k^c &=& \sigma^{ac}_{ik}E_a^i + d^{abc}_{ijk}E_a^iE_b^j + \mathcal{O}(E^3),
\end{eqnarray}
With $a,b,c$ indices of global currents, and $i,j,k$ indices of $d-$dimensional Euclidean spacetime coordinates. $\sigma_{ij}^{ab}$ and $d_{ijk}^{abc}$ are the 2 and 3-rank conductivity tensors. The fact that the two operators lead to different polarization structures will be exploited. Consider the special points
\bea
z &=& (0,0,0,...,0)\nonumber\\
x &=& (\tau, r, 0, ...,0)\nonumber\\
y &=& (\tau, -r, 0,...,0).
\end{eqnarray}
The $i=j=k$ component of the tensor $D^{sym}_{ijk}$ automatically vanishes, whereas the $ijk=122$ component of $D^{sym}_{122}$ is just a rescaling of $C^{sym}_{ijk}$. However, the $ijk=112$ component of the symmetric tensors $D^{sym}_{112}$ and $C^{sym}_{112}$ are linearly independent, and take the values
\bea
D^{sym}_{112}&=& \frac{(d-2)^2}{\Big[2r(\tau^2+r^2)\Big]^{(d-1)}}\left(\frac{r^4-\tau^4+8\tau^2r^2}{(\tau^2+r^2)^2} \right),\nonumber\\
C^{sym}_{112}&=& - \frac{(d-2)^3}{\Big[2r(\tau^2+r^2)\Big]^{(d-1)}} \left( 1-\frac{16\tau^2r^2}{(\tau^2+r^2)^2}\right).
\end{eqnarray}
Then, the two different linear combinations corresponding to the contribution of the $F^2$ operator verses the $F^3$ operator vanish for different values of $\tau$ and $r$. For example, for $d=3$ Euclidean dimensions,
\bea
\left\langle J^a_1(x)J^b_1(y)J^c_2(z) \right\rangle &=& \frac{f^{abc}}{2^9g_{SG}^2}\frac{1}{\big[ 2r(\tau^2+r^2)\big]^4}\left((3r^4-2\tau^4+13\tau^2r^2)-\frac{2}{\Delta^2}(r^4-2\tau^4+19\tau^2r^2) \right) \nonumber
\end{eqnarray}
Comparing the to measurements at the two different set of points where either contribution vanishes, we can not only test the validity of the effective approach, but also find the dimension gap $\Delta$ suppressing higher order corrections.
To conclude, in this paper we computed the three-point current correlation function in the framework of Effective Conformal Field Theory. This describes the dynamics of all operators with dimensions below the cutoff dimension $\Delta_{Heavy}$. In systems with large dimension gap $\Delta_{gap}\approx \Delta_{heavy}$ and a $1/N$ like suppression, there is double expansion in both $1/N$ and $1/\Delta_{gap}$. The contributions to the three-point current correlation function coming from the lowest non-renormalizable bulk operator $F^3$ is computed and compared to the contribution coming from the renormalizable $F^2$ bulk operator already computed in the literature. It is shown that the two operators give rise to different polarization structure of the three-point current correlation function. The polarization structure coming from the non-renormalizable bulk $F^3$ term is suppressed by powers of the cutoff dimension $\Delta_{Heavy}$ prescribed by demanding perturbative unitarity.
By measuring the non-linear response to external fields, it is possible to test the effective description for strongly coupled condensed matter systems. In systems with global non-Abelian symmetry and large hierarchy in operator dimensions at second order phase transition, we can expect new terms of order $1/\Delta_{Heavy}^2$ in the three-point current correlation function with a different polarization structure to the leading effect.
\section*{Acknowledgment}
This humble effort would not have been possible without the help of persons whom I would like to acknowledge in this section. First and foremost I thank God to whom this work, the first fruit of my labor in graduate school, is dedicated.
I thank Jared Kaplan for suggesting the project and for many helpful discussions, both physics related and otherwise. I thank Michael Peskin for being very supportive of my efforts. I also thank Srinivas Raghu and Maissam Barkeshli for helpful conversations about the experimental aspects of the work, and Laim Fitzpatrick for helpful feedback. The work was supported by DOE under contract DE--AC02--76SF00515.
|
\section{Intro}
Molecular simulations provide way to look at water motion and structure at the nanoscale with atomistic details. Notwithstanding classical models demonstrated to correctly reproduce a wide range of observables \cite{Vega2011}, several properties of this fascinating element are not well described yet, including the nucleation mechanism \cite{Matsumoto2002, Moore2011} and the correct proportion between melting and density maximum temperatures to name a few \cite{Vega2005-2}.
Classical water potentials are attractive because they are cheap to compute as compared to purely ab-initio approaches. However, to be fast they went through a series of approximations including rigid molecular structure and on-site fixed partial charges. Like a too short blanket, parametrization of those classical models allowed the correct prediction of some properties leaving behind other ones and vice versa: models like TIP4P-ICE better fit the properties of ice \cite{Abascal2005ice}, others the density anomaly (TIP5P \cite{Mahoney2000}, TIP4P-Ew \cite{Horn2004} and TIP4P/2005 \cite{Abascal2005}) or the diffusion constant (SPC-E \cite{Mark2001}), but all of them fail to reproduce the broader spectrum of water properties. Although many have agreed that four site potentials might represent the best compromise for classical water models \cite{Vega2005,Vega2011,Horn2004}, still some important ingredients are missing in these representations: one being \emph{polarization}. In common words, the latter refers to the ability of an atom to change its charge in response to the environment, reflecting a redistribution of the electronic cloud. This effect is pretty obvious and omnipresent when dealing with charged atoms. Think for example of the effect of an ion on the charge distribution of the surrounding molecules \cite{Lamoureux2006_ion}. The drawback for the introduction of polarizability in a classical potential is however two fold. First, given the increased number of degrees of freedom a fully fledged polarizable molecular model is much more computationally expensive to calculate. Second, parametrization of such a model is non-trivial \cite{Brooks2009}.
In recent years, we saw the rise of several polarizable water models such as for example BK \cite{Baranyai2010}, AMOEBA \cite{Ren2003} and SWM4 \cite{Lamoureux2003}. They differ among each other in the way polarization is implemented. In BK the charge distribution is represented by Gaussian functions while polarizability is introduced via a charge-on-spring method \cite{Baranyai2010}. In AMOEBA, polarization effects are treated via mutual induction of dipoles with experimentally derived polarizabilities and a 14-7 potential to treat Van der Waals interactions \cite{Halgren1992}. A new version of this potential called iAMOEBA (where the ''i'' stands for inexpensive) \cite{Wang2013} makes this model only four times slower compared to a conventional water model. Finally, another way to introduce polarization is to use a Drude oscillator potential \cite{Vanmaaren2001,Lamoureux2003}. In this case a point charge is connected via a classical spring to the oxygen atom via a dummy atom where an external field displacing the dummy particle would in turn induce a dipole \cite{Lamoureux2003}. A fairly adopted implementation of this solution is represented by the SWM4-NDP model \cite{Lamoureux2006}. Being a pairwise based potential and resembling the architecture of a regular four site model like TIP4P, this model seems to fit better into a conventional molecular force field framework. The model is certainly promising at ambient conditions showing better agreement with experiments for viscosity \cite{Stukan2013} and hydration of the calcium carbonate \cite{Bruneval2007}.
Here, we make an effort to further explore the behavior of the SWM4-NDP model on a wider temperature range. Focusing on some basic properties of bulk water, extensive molecular dynamics simulations were performed for temperatures ranging from 170~K to 340~K. We aimed at the characterization of the density curve as well as at the hydrogen bond propensities and tetrahedral order. The model does not seem to perform very well in terms of density, especially in the supercooled regime where the relaxation times became very long. On the other hand, hydrogen-bond connectivity and tetrahedrality agree to optimized four sites classical water models. Our results provide an interesting starting point to improve on the behavior of Drude based water models beyond ambient conditions.
\section{Methods}
\subsection{Simulation details}
All molecular dynamics simulations were run with the NAMD program \cite{NAMD} with an integration step of 1~fs. The system contained 1024 water molecules in a cubic box. Temperature and pressure were controlled with a Langevin thermostat and Berendsen barostat with 1~ps and 100~fs relaxation time, respectively. The temperature of the Drude particles were set to 1~K at all conditions as suggested in the paper implementing the model into NAMD \cite{Jiang2010}. Non-covalent interactions were treated with a 1.2~nm cut-off and PME. Molecular trajectories of 50~ns in length were calculated for temperatures from 170~K to 260~K with steps of 10~K while from 260~K to 340~K with steps of 20~K. At the higher temperatures (T$>$260~K) the simulations length was of only 10~ns per trajectory because of the rapid equilibration times.
TIP4P/2005 simulations \cite{Abascal2005} were run with the program GROMACS
\cite{gromacs} with an integration time-step of 2~fs. The water box
consisted of 1024 molecules in the NPT ensemble with pressure of
1 atm and temperatures ranging from 180~K to 350~K with steps of 10~K.
The Berendsen barostat \cite{Berendsen1984}, velocity rescale
thermostat \cite{Bussi2007} and PME \cite{Darden1993} were used for
pressure coupling, temperature coupling and long-range electrostatics,
respectively.
The data was obtained from 1~ns long
simulations after 10~ns of equilibration in the NPT ensemble for T$>$240~K. For temperatures lower than 240~K 20~ns of equilibration was adopted.
\subsection{Hydrogen-bond propensities}
A maximum of four
hydrogen-bonds per molecule was considered with a bond being formed if the distance between oxygens and the angle O-O-H was smaller than 3.5~\AA\ and 30~degrees, respectively~\cite{Luzar1996}. Water structures were grouped into four archetypal configurations of population P$_i^{(*)}$ \cite{Shevchuk2012}: the fully coordinated first and second solvation shells for a total of 16 hydrogen-bonds (P$_4$); the fully coordinated first shell, in which one or more hydrogen bonds between the first and the second shells are missing or loops are formed (P$_4^*$); the three coordinated water molecule (P$_3$) and the rest (P$_{210}$). Within this representation the sum over the four populations is equal to one for each temperature.
\subsection{Tetrahedral order parameter}
The tetrahedral order parameter for a water molecule $i$ was calculated as
\begin{equation}
q_{i}=1-\frac{3}{8}\sum_{j=1}^{3}\sum_{k=j+1}^{4}\big(cos\psi_{jik}+\frac{1}{3}\big)^{2},
\end{equation}
where $j$ and $k$ are any of the four nearest water molecules of $i$ and $\psi_{jik}$ is the angle formed by their oxygens\cite{Errington2001}. The averaged value of the order parameter is denoted as Q.
\section{Results}
\subsection{The density maximum of the SWM4-NDP model}
\begin{figure}
\includegraphics[width=80mm]{fig1}
\caption{Timeseries of the potential energy of the SWM4-NDP water model for three different temperatures. Below 190~K relaxation times become very long as depicted by the 170~K trajectory (red) and the gray region. }
\label{fig:pot_en_ts}
\end{figure}
\begin{figure}
\includegraphics[width=80mm]{fig2}
\caption{The density curve at 1 atm. Data for SWM4-NDP, TIP4P/2005 and experiments \cite{Kell1975} are shown in red, orange and black, respectively.}
\label{fig:density_ts}
\end{figure}
Molecular dynamics simulations of the Drude-based polarizable water model SWM4-NDP were performed at several temperatures spanning from 170~K to 340~K. Running simulations for temperatures as low as 180~K, Kiss and Baranyai \cite{Kiss2012} recently showed that this model presents no density maximum. Independently from them we were also looking at the same problem. One important difference in our work is that simulations were run for much longer times: 50 ns per trajectory opposed to 5 ns in their case.
Our results strongly indicate that long runs of several ns are needed to characterize SWM4-NDP in the deeply supercooled regime. This becomes clear when looking at the time series of the potential energy. In Fig. 1A traces for different temperatures from 250~K to 170~K are shown. It was found that for temperatures lower than 200~K the relaxation time of the system dramatically slows down. The red line corresponding to 170~K shows that the system required at least 20~ns to equilibrate (gray region). This is a much longer time than the simulation length used in Ref.~\cite{Kiss2012}.
With the longer trajectories at hand, the density curve did present a maximum at around 200~K (red points in Fig.~2), a value that is similar to what was found for TIP3P (182~K \cite{Vega2005-2}). However, this maximum is not a global one as in experiments (black line) or in other classical models like TIP4P/2005 (orange points). In fact, at lower temperatures (T$<$190~K) density grows again, making the density peak difficult to emerge from the statistical error, especially when using short trajectories. An increase of the density passed the density maximum is a feature of several water models. For example, this happens as well for TIP4P/2005 below 200~K (Fig.~2). What makes SWM4-NDP peculiar is the fact that the value of the density in this regime becomes even higher than the density maximum, making the latter a relative maximum (not an absolute one).
\subsection{Hydrogen-bond propensities and temperature-shifts}
\begin{figure}
\includegraphics[width=80mm]{fig3}
\caption{Hydrogen bond propensities including the second solvation shell. $P_4$, $P^*_4$, $P_3$ and $P_{210}$ are shown in red, dark blue, blue and light blue, respectively (see Methods for details). Data for SWM4-NDP and TIP4P/2005 are shown as filled circles and empty squares, respectively.}
\label{fig:macro}
\end{figure}
Complementary information was obtained by investigating hydrogen-bond propensities. As done recently for seven classical water models \cite{Shevchuk2012} we calculated the probability to form fully coordinated hydrogen-bond configurations up to the second shell ($P_4$, red in Fig.~3; see Methods) as well as fully coordinated first shells with a disordered second shell ($P^*_4$, dark blue), three coordinated ($P_3$, blue) and less ($P_{210}$, light blue) first solvation shells. Results for the SWM4-NDP and TIP4P/2005 for comparison are shown in Fig.~3 as filled circles and empty squares, respectively. Contrary to the density analysis, hydrogen-bond propensities between the two models look much more similar (e.g. TIP3P showed a much more drastic temperature shift of 60~K \cite{Shevchuk2012}). The two sets of curves would nicely overlap if a shift of approximately 20~K is applied to the data. This observation suggests that while spatial rearrangement responsible for the density is dramatically different between the two models (and when compared to experiments), hydrogen-bond connectivity is similar. Such a discrepancy was already observed when comparing three-sites with four-sites models where a 10~K difference between temperature shifts estimated from hydrogen bonds or the position of the density maximum was observed \cite{Shevchuk2012}. But in this case the discrepancy is much larger being the temperature shifts respectively of 20~K and 80~K, i.e. a 40~K difference between the two approaches.
\subsection{Tetrahedral order parameter}
\begin{figure}
\includegraphics[width=80mm]{fig4.eps}
\caption{Tetrahedral order parameter. (Top) Average value of the parameter as a function of temperature. (Bottom) parameter distribution at 300~K. Data for SWM4-NDP and TIP4P/2005 are shown in red and orange, respectively.}
\label{fig:tetra_temp}
\end{figure}
Temperature shifts were observed as well when comparing the two models on the base of the average value of the tetrahedral order parameter $Q$ (see Methods). The top panel of Fig.~4 shows this quantity as a function of temperature for both SWM4-NDP (red line) and TIP4P/2005 (orange). As for the case of the hydrogen-bond propensities $P_{i}^{(*)}$, the two models do not differ very much. At ambient conditions the temperature-shift is of about 30~K, a number that is in line with what observed for the hydrogen-bond propensities (Fig.~3). For the sake of comparison the distribution of $Q$ at 300~K for the two models is shown at the bottom of Fig.~4. As it could have been expected from the behavior of the average value of the order parameter, TIP4P/2005 has a slightly larger fraction of molecules in a tetrahedral configuration but the overall shape of the distribution is similar for the two cases. This is even clearer when presenting the data for TIP4P/2005 at a 30~K higher temperature (gray curve): now the distribution for SWM4-NDP and the temperature-shifted TIP4P/2005 nicely overlap on top of each other with good approximation.
\section{Discussion}
In the present work, we performed extensive molecular dynamics simulations of the Drude-based polarizable water model SWM4-NDP as a function of temperature. Contrary to what was reported in a recent paper \cite{Kiss2012}, it was found that the model do present a density maximum which was found to be around 200~K. The density curve was not easy to calculate because of the intrinsic slowing down of the system for temperatures lower than 200~K that hindered the detection of the maximum. To overcome this problem simulation runs of 50~ns each were performed, finding that at temperatures below 200~K the system required at least 20~ns to have the potential energy relaxing to a stationary average value without drifts.
However, the density maximum we found is not as pronounced as other classical water models or in experiments. This was somewhat unexpected. As system temperature was lowered below 190~K, the density started to increase again. This is only in principle similar to what was observed for other models, like for example TIP4P/2005. In fact, in the present case the density value increased to a value that is larger than the density maximum, making the latter a \emph{relative} maximum instead of an absolute one. The raising of the density at a such low temperature is probably due to some sort of frustration into the system leading to glassy behavior. This idea would also explain the dramatic slowing down of the relaxation kinetics of the model below 200~K.
In comparison to other classical models, SWM4-NDP performed very poorly in reproducing the density curve. This is somewhat disappointing given the success of other models in this respect, especially the reparametrized versions of the four-site model, TIP4P/2005 \cite{Abascal2005} and TIP4P-Ew \cite{Horn2004} as well as the newly presented iAMOEBA polarizable model \cite{Wang2013}.
Despite the position of the density maximum of SWM4-NDP is shifted by roughly 80~K, the behavior of the hydrogen-bond propensities and tetrahedrality are very well in line to what the best models in the field predict. This behavior differs from what we found in the past for other non-polarizable water models, i.e. that a temperature-shift in the density maximum corresponds to a similar shift in the hydrogen-bond propensities. The presence of polarizability instead completely decouples these two aspects, giving in principle a wider space to match experimental data, at least in principle.
In conclusion, our work shed some further light on the behavior of the SWM4-NDP polarizable model in temperature space. The great advantage of this model with respect to other approaches is the easy integration in all modern force-fields for biomolecular simulations. However, our results suggest that to make this model fully effective, a new parametrization able to reproduce the density curve and other quantities in temperature space is required.
|
\section{Introduction}
Theoretically modeling the spreading dynamics of computer virus (or malware such as worm and bot)
is important for deepening our understanding and for
designing effective, if not optimal, defenses.
We observe, however, that the utility of theoretical modeling in this context is not well understood yet
because existing models are often adapted from biological epidemic ones.
As a consequence, many existing models of computer virus spreading dynamics made the so-called {\em homogeneity}
assumption, which roughly says that the nodes are equally powerful in infecting others.
Realizing the limitation of the assumption, there have been investigations
that aim to weaken the assumption by considering heterogeneous network
topology (where different nodes may have different infection capabilities
because they have different degrees).
Along this line of study, the present paper moves a step further by
exploring models that accommodate realistic scenarios where
the model parameters may change over time (i.e., the parameters are some
functions of time), which captures the fact that both attack and defense are dynamically evolving
or under dynamical adjustment and reflects the persistence of virus spreading.
This allows us to investigate an
important and novel perspective of virus spreading-defense dynamics,
namely that of {\em adaptive} defense against computer virus spreading.
\subsection{Our Contributions}
We investigate a non-homogeneous Susceptible-Infectious-Susceptible (SIS) model in
arbitrary networks (i.e., there is no restriction on the topology of the spreading networks
and the nodes may have different defense or cure capabilities).
The model can accommodate both {\em semi-adaptive} defense and {\em fully-adaptive} defense.
In the semi-adaptive defense scenario, the input parameters in the model are known and can vary
with respect to time (e.g., according to some
deterministic functions of time or according to some stochastic process, but we do not impose
any practical restrictions on the types of functions).
For this scenario, we present a set of sufficient conditions, from general to specific (but more succinct),
under which the virus spreading will die out. We note that
such sufficient conditions are also known as {\em epidemic thresholds} in the literature.
In the fully-adaptive defense scenario, some input parameters are not known and thus
the aforementioned sufficient conditions are not applicable. Nevertheless,
the defender might be able to observe the outcome of virus spreading (i.e., which nodes are
infected at a point in time).
For this scenario, we present adaptive control strategies
under which the virus spreading will die out or will be contained to a desired level
(which is important when, for example, the price to kill the virus spreading may be too high).
Because of the above, our model supersedes previous homogeneous and non-homogeneous models
that offered relevant analytical insights;
the concrete connection will be made when the need arises.
Our analytical results are confirmed via simulation, from which we draw additional observations
that serve as hints for future modeling studies. We discuss the
practical implications of our model and the derived insights as well.
Finally, we note that the present paper is meant to explore
theoretical characterizations of spreading-defense dynamics
while assuming certain parameters can be observed or measured
(e.g., based on extensive data and possibly expert knowledge).
This may not be feasible some times.
Regardless, we believe that such studies are important on their own and represent a necessary
step towards the ultimate characterization of virus spreading-defense dynamics
(which in turn helps design more effective or even optimal defenses).
\subsection{Related Work}
To the best of our knowledge, there are no existing studies on modeling
adaptive spreading-defense dynamics in arbitrary networks.
The work that is most closely related to ours was due to Chakrabarti et al. \cite{WangTISSEC08},
who considered computer virus spreading in arbitrary networks --- a scenario also investigated
in \cite{WangSRDS03,TowsleyInfocom05}. The most important contribution of these studies is
the identification of a sufficient condition (i.e., epidemic threshold) under which the virus spreading will die out;
we will discuss the relationship between their result and ours when the need arises.
Earlier studies either made the homogeneity assumption
(e.g., \cite{KephartOkland91,KephartOkland93}) as in
biological epidemic models (see, for example, \cite{McKendrick1926,Kermack1927,Bailey1975,Anderson1991,HethcoteSIAMRew00}),
or considered specific non-homogeneous networks \cite{WangTISSEC08}.
We should mention prior work that is conceptually or spiritually relevant.
The concept of ``adaptable robust computer systems" was investigated by Bhargava et al. \cite{BhargavaSIGOPS86},
which however has a very different meaning and is for very different purposes.
Also for a different purpose, Zou et al. \cite{ZouSRUTI05} explored the concept of ``adaptive defense" based on cost optimization,
where cost was introduced by false positives and false negatives. In particular, they considered
optimal adaptive defense against worm infection, but is from the perspective of
decision whether or not to block/allow some specific host traffic. As such, it may be possible to
combine their studies and ours because we do not consider cost.
\smallskip
\noindent{\bf Outline}: In Section \ref{sec:model-alpha-eq-0} we present our model as well as the
analytical insights. We report our simulation study in Section \ref{sec:simulation-study-alpha-eq-0}.
We conclude the paper in Section \ref{sec:conclusion} with open problems.
\section{Adaptive Epidemic Dynamics: Model and Analysis}
\label{sec:model-alpha-eq-0}
\subsection{The model}
\noindent{\bf Primary parameters}.
Because we want to accommodate spreading in arbitrary networks, we assume that virus spreads over
a series of finite, dynamical graphs $G(t)=(V,E(t))$, where
$V$, $|V|=n$, is the set of
nodes or vertices and $E(t)$ is the set of (possibly changing) edges or arcs at time $t \ge 0$
(i.e., the topology may change with respect to time).
At any time $t$, an infected node $u$ can directly infect node $v$ if $(u,v)\in E(t)$.
Denote by $A(t)=[a_{vu}(t)]$ the adjacency matrix of $G(t)$, where
$a_{vv}(t)=0$ for all $v\in V$, and $a_{vu}(t)=1$ if and only if $(u,v) \in E(t)$.
Note that this representation naturally accommodates both directed and undirected topologies,
and thus our results equally apply to them.
A node $v \in V$ is \susceptible\ if $v$ is secure but
vulnerable, and \infected\ if $v$ is successfully attacked (i.e., infected and infectious).
At any time $t$, a node $v \in V$ is either \susceptible\ or \infected.
Moreover, a \susceptible\ node may become
\infected\ because of some \infected\ node $u$ where
$(u,v)\in E(t)$, and an \infected\ node may become \susceptible\ because of cure. Since an
\infected\ node may become \susceptible\ again, our model falls into the category of the
so-called SIS models, but our model has the unique feature that
values of the parameters can change with respect to time.
We consider two dependent variables:
$s_v{(t)}$, the probability $v \in V$ is \susceptible\ at time
$t$; $\cc_{v}{(t)}$, the probability $v\in V$ is \infected\ at
time $t$. We consider a continuous-time model, which preserves the invariant
$s_v{(t)} + \cc_{v}{(t)}=1$. The model's input parameters are:
\begin{itemize}
\item $\beta_v(t)$: The probability an \infected\ node $v$ becomes \susceptible\ at time $t$.
\item $\gamma_{uv}(t)$: The probability an \infected\ node $u$ successfully infects a \susceptible\
node $v$ over edge $(u,v)\in E(t)$ at time $t$.
For simplicity, we assume that $\gamma_{uv}(t) = \gamma(t)$ for all $(u,v) \in E(t)$.
\end{itemize}
For the sake of mathematical rigorousness, the $\beta_v(t)$'s and the $\gamma_{uv}(t)$'s, which are probabilities, should be ``measurable"
so as to ensure the existence of solutions to system (\ref{Eq.2.2}) and be
``bounded" so as to ensure the proof of Theorem \ref{thm4} can get through.
To avoid any unnecessary mathematical subtleties, we simply assume that
these parameters are ``boundedly measurable," which has no consequence in practice.
Note that $A(t)$ is naturally bounded.
\smallskip
\noindent{\bf Other parameters and notations}.
Below is a summary of the major notations used in the paper;
notations only occasionally used are explained when the need arises.
\begin{center}
\begin{tabular}{|r|p{.6\textwidth}|}
\hline
model input parameters: & \\
$A(t)=[a_{vu}(t)]$ & the adjacency matrix of graph $G(t)=(V,E(t))$ where
$|V|=n$, and $a_{vu}(t)=1$ if and only if $(u,v)\in E(t)$. Moreover, $a_{vv}(t)=0$ for all $v \in V$.\\
$\beta_{v}(t)$ & the cure capability of node $v$ at time $t$\\
$\gamma(t)$ & the edge infection capability at time $t$\\
\hline
dependent variables: &\\
$s_v{(t)}$ & the probability node $v\in V$ is \secure\ at time $t$ \\
$\cc_{v}{(t)}$ & the probability node $v\in V$ is \compromised\ at time $t$ \\
\hline
intermediate variables: &\\
$\delta_{v}{(t)}$ & the probability \susceptible\ $v\in V$ becomes \infected\ at time $t$ because of \compromised\
neighbors $\{u: (u,v) \in E(t)\}$ \\
\hline
other notations: &\\
$\lambda_1$ & the largest (in modulus) eigenvalue of adjacency matrix $A$\\
$\|\cdot\|$ & the 1-norm of vector or matrix \\
$B(t)=\diag[\beta_{1}(t),\cdots,\beta_{n}(t)]$ & the cure probability diagonal matrix \\
\hline
\end{tabular}
\end{center}
\noindent{\bf The state transition diagram and master equation}.
Figure \ref{fig:state-diagram-research-plan-adaptive-alpha=0} depicts the state transition diagram of a node, where
the probability $\delta_v{(t)}$ is given by
\begin{equation}
\label{compromise}
\delta_v {(t)}=1-\prod\limits_{(u,v)\in E(t)}\left[1-\gamma(t) \cdot i_u {(t)}\right].
\end{equation}
\begin{figure}[ht]
\centering
\includegraphics[width=.8\textwidth]{research-plan-adaptive-alpha=0.eps}
\caption{State transition diagram of node $v \in V$ at time $t$}
\label{fig:state-diagram-research-plan-adaptive-alpha=0}
\end{figure}
Note that in the derivation of Eq. (\ref{compromise}), we assumed that the events that \infected\ neighbors infect
a node are independent.
Note also that $s_v(t) + \cc_v(t)=1$ for any $t$.
Based on the state transition diagram we obtain the following master equation of dynamics:
\begin{eqnarray}\label{Eq.2}
\frac{d\cc_{v}(t)}{dt} =\bigg[1-\prod_{u\in V}[1-\gamma(t)
a_{vu}(t) \cc_{u} {(t)}]\bigg][1-\cc_{v} {(t)}]-\beta_{v}(t)\cc_{v}
{(t)},
\end{eqnarray}
where $a_{vu}(t)=1$ if and only if $(u,v)\in E(t)$.
\subsection{Sufficient conditions for dying out in the scenario of semi-adaptive defense}
\label{sec:implicit-adaptive-case}
In this subsection we present a set of sufficient conditions under which the virus spreading will die out.
The sufficient conditions are applicable when the model's input parameters,
namely the $\beta_v(t)$'s and $\gamma(t)$ are given. Moreover, it is possible
that $\beta_v(t)$ relies on $\gamma(t)$; for example, the former is an (implicit) function of the latter.
This explains why we call this scenario the semi-adaptive defense.
\begin{theorem}
\label{theorem-1}
\emph{(a general sufficient condition under which virus spreading dies out)}
Consider the following comparison (and linearization) system of Eq. (\ref{Eq.2}):
\begin{eqnarray}
\label{Eq.2.1}
\frac{d x_{v}(t)}{dt} =\sum_{u=1}^{n}a_{vu}(t)\gamma(t)x_{u}(t) -\beta_{v}(t)x_{v} {(t)}.
\end{eqnarray}
Let
$x(t)=[x_{1}(t),\cdots,x_{n}(t)]^{\top}$, we obtain the following compact form of Eq. (\ref{Eq.2.1}):
\begin{eqnarray}\label{Eq.2.2}
\frac{d x(t)}{dt} =\bigg[\gamma(t)A(t) -B(t)\bigg]x(t).
\end{eqnarray}
Denote by $U(t,t')$ the solution matrix of linear system
(\ref{Eq.2.2}), namely that each solution of
linear system (\ref{Eq.2.2}), $x(t)$, with initial condition
$x(t_{0})=x_{0}$, can be written as $x(t)=U(t,t_{0})x_{0}$. Because
the solution $x(t)$ is dependent upon the initial value $x_0$ but
the solution matrix $U(t,t')$ is not, the corresponding
maximum Lyapunov exponent (MLE) is determined by
the solution matrix (rather than by the solution $x(t)$) and can be defined as \cite{Ose}:
\begin{eqnarray*}
\mu=\lim_{t\to\infty}\frac{1}{t}\ln\|U(t,0)\|
\end{eqnarray*}
where $\|\cdot\|$ is (for specificality) the 1-norm of matrix (because $\mu$ is
independent of the choice of the norm).
If $\mu<0$, the virus spreading will die out regardless of the initial infection
configuration;
if $\mu>0$ and system (\ref{Eq.2.2}) is ergodic \cite{Ose}, the virus spreading will not die out in some initial infection
configurations (i.e., ``the equilibrium of $i^*=0$ is unstable" in mathematical terms).
\end{theorem}
\begin{proof}
Note that
\begin{eqnarray}
1-\prod_{u\in V}[1-\gamma(t)a_{vu}(t)\cc_{u}(t)]\le
\gamma(t)\sum_{u\in V}a_{vu}(t)\cc_{u}(t).\label{ineq}
\end{eqnarray}
If $\cc_{v}(0)=x_{v}(0)$ for all $v\in V$, then the
comparison system (\ref{Eq.2.1}) satisfies that
$\cc_{v}(t)\le x_{v}(t)$ holds for all $t\ge 0$ and $v\in V$.
Since Eq. (\ref{Eq.2.1}) is actually the
linear system of the system (\ref{Eq.2}) at $\cc_{v}=0$ for all
$v\in V$. Therefore, we can conclude that the stability
of Eq. (\ref{Eq.2}) is equivalent to that of Eq. (\ref{Eq.2.1}). In
other words, if all $x_{v}(t)$'s of system (\ref{Eq.2.1}) converge
to zero, then system (\ref{Eq.2}) is stable regardless of the
initial values; on the other hand, if system (\ref{Eq.2.1}) is
unstable, then system (\ref{Eq.2}) is also unstable. If system
(\ref{Eq.2.2}) is ergodic \cite{Ose}, then the limit $\mu$ exists;
otherwise, we can alternatively define
$\mu=\overline{\lim}_{t\to\infty}\frac{1}{t}\ln\|U(t,0)\|$,
where $\overline\lim_{t\to\infty} z(t)$ represents the upper bound
of the limit of $z(t)$ as $t$ goes to infinity, is also guaranteed
to exist. In any case, by applying the definition of MLE, we obtain
the theorem immediately.
\end{proof}
\noindent{\bf Discussion}. The above sufficient condition $\mu<0$
for the virus spreading to die out is actually close to being
necessary, meaning that if $\mu >0$, then the virus spreading will
not die out in {\em most}, rather than just {\em some}, initial
infection configurations. According to the Lyapunov exponent and smooth
ergodic theory developed by \cite{Pesin77} and many
others, $\mu>0$ means that the system (\ref{Eq.2}) possesses an
unstable manifold, which implies that the stable manifold, i.e., the
set of points (i.e., the initial values) starting from which the system (\ref{Eq.2})
converges to the origin, has dimension less than $n$. Therefore,
the stable manifold has Lesbegue measure $0$. That is, except a set
with Lesbegue measure $0$, the virus spreading
never dies out with respect to any initial infection configuration.
The sufficient condition given in Theorem \ref{theorem-1} is very
general because in its derivation no ``amplification" is used and
dynamical topology $E(t)$ is accommodated. However, it requires to,
among other things, solve a system of $n$ linear equations of $n$
variables (equivalently, diagonalizing a $n \times n$ matrix), which
can be quite time-consuming for large $n$ (the number of nodes).
In what follows we give two succinct sufficient conditions,
which can be easily connected to previous state-of-the-art results.
\begin{theorem}
\label{thm4}
\emph{(a succinct sufficient condition)} Suppose
$\beta_v(t)=\beta(t)$ for all $v \in V$ (i.e., all nodes have the
same cure capability) and $E(t)=E$ for any time $t$, meaning that the topology
does not change over time and $A=A(t)$ for any $t$. Consider the system
\begin{eqnarray}\label{Eq.2.3}
\frac{d x(t)}{dt} =\bigg[\gamma(t)A -\beta_v(t)I_{n}\bigg]x(t),
\end{eqnarray}
where $I_n$ is the $n \times n$ identity matrix. Let
\begin{eqnarray*}
\bar{\gamma}=\lim_{t\to\infty}\frac{1}{t}\int_{t_{0}}^{t+t_{0}}\gamma(\tau)d\tau,
~~~\bar{\beta}=\lim_{t\to\infty}\frac{1}{t}\int_{t_{0}}^{t+t_{0}}\beta(\tau)d\tau,
\end{eqnarray*}
and suppose the limits exist and are uniform with respect to
$t_{0}$. Let
$\lambda_{1}$ be the largest (in modulus) eigenvalue of $A$.
If
\begin{eqnarray}
\lambda_{1}<\frac{\bar{\beta}}{\bar{\gamma}},\label{stable}
\end{eqnarray}
then the virus spreading will die out regardless of the initial infection configuration;
if
\begin{eqnarray}
\lambda_{1}>\frac{\bar{\beta}}{\bar{\gamma}},\label{unstable}
\end{eqnarray}
then the virus spreading will not die out in some initial infection configurations.
\end{theorem}
\begin{proof}
Consider system (\ref{Eq.2.2}). Let
$A=S^{-1}JS$ be the Jordan canonical form with
\begin{eqnarray*}
J=\left[\begin{array}{lllll}J_{1}&&&&\\
&J_{2}&&&\\&&J_{3}&\\&&&\ddots&\\&&&&J_{K}\end{array}\right],
~J_{k}=\left[\begin{array}{lllll}\lambda_{k}&0&0&\cdots&0\\
1&\lambda_{k}&0&\cdots&0\\0&1&\lambda_{k}&\cdots&0\\
\vdots&\vdots&\vdots&\ddots&\vdots\\0&0&\cdots&1&\lambda_{k}\end{array}\right],
\end{eqnarray*}
where $\lambda_{k}$, $k=1,\cdots,K$, are the distinct eigenvalues of
$A$. Recall that $\lambda_{1}$ is the largest eigenvalue in modulus.
From the Perron-Frobenius theorem \cite{Berman03}, $\lambda_{1}$ is
a real number. Then, letting $y(t)=Sx(t)$, we have
\begin{eqnarray*}
\frac{d y(t)}{dt}=[\gamma(t) J-\beta(t) I_{n}]y(t).
\end{eqnarray*}
Namely,
\begin{eqnarray*}
\frac{d y_{v}(t)}{dt}=[\gamma(t)\lambda_{k_{v}}-\beta(t)]y_{v}(t)+\xi_{v}(t),
\end{eqnarray*}
where $\lambda_{k_{v}}$ is the eigenvalue of $A$ corresponding to
the Jordan block $J_{k_{v}}$ that contains column $v$, and
$\xi_{v}(t)=0$ if the $v$-th row of $S$ is an eigenvector of $A$ and
$\xi_{v}(t)=\gamma(t)y_{v-1}(t)$ otherwise. First, consider $v=1$
corresponding to the eigenvalues $\lambda_{1}$. We have
\begin{eqnarray}
y_{1}(t)=y_{1}(0)\exp\left(\int_{0}^{t}[\gamma(\tau)\lambda_{1}-\beta(\tau)]d\tau\right).\label{comp1-1}
\end{eqnarray}
One can see that the Lyapunov exponent of system (\ref{comp1-1})
is calculated as
\begin{eqnarray*}
\lim_{t\to\infty}\frac{1}{t}\ln||y_{1}(t)||&=&
\lim_{t\to\infty}\frac{1}{t}\ln||y_{1}(0)||+\lim_{t\to\infty}\frac{1}{t}\int_{0}^{t}
[\gamma(\tau)\lambda_{1}-\beta(\tau)]d\tau\\
&=&0+\bar{\gamma}\lambda_{1}-\bar{\beta}<0,
\end{eqnarray*}
which implies $\lim_{t\to\infty}y_{1}(t)=0$.
Assuming $\lim_{t\to\infty}y_{v}(t)=0$ already proved, consider $y_{v+1}(t)$. We have
\begin{eqnarray*}
y_{v+1}(t)=y_{v+1}(0)\exp\left(\int_{0}^{t}k(\tau)d\tau\right)
+\int_{0}^{t}\xi_{v+1}(a)\exp\left(\int_{a}^{t}k(\tau)d\tau\right)
d a
\end{eqnarray*}
where $k(\tau)=\gamma(\tau)\lambda_{k_{v+1}}-\beta(\tau)$.
From condition (\ref{stable}), there exists a sufficiently
small $\epsilon>0$ such that
$(\bar{\gamma}+\epsilon)\lambda_{1}-(\bar{\beta}+\epsilon)<0$.
Let
$\varphi=-(\bar{\gamma}+\epsilon)\lambda_{1}+(\bar{\beta}+\epsilon)$.
One can see that $\varphi>0$. Since the limits
$\int_{t_{0}}^{t_{0}+t}\beta(\tau)d\tau$ and
$\int_{t_{0}}^{t_{0}+t}\gamma(\tau)d\tau$ are uniform with respect
to $t_{0}$, there exists $T>0$ such that
\begin{eqnarray*}
\frac{1}{t-a}\int_{a}^{t}\beta(\tau)d\tau<\bar{\beta}+\epsilon,
~\frac{1}{t-a}\int_{a}^{t}\gamma(\tau)d\tau<\bar{\gamma}+\epsilon
\end{eqnarray*}
hold for any $a$ and $t$ with $t-a>T$. Let ${\mathcal Re} (z)$ denote the real part of a complex number
$z$. Then, we have
\begin{eqnarray*}
\mathcal
Re\bigg(\frac{1}{t-a}\int_{a}^{t}k(\tau)d\tau\bigg)&=&\frac{1}{t-a}\int_{a}^{t}
[\gamma(\tau)\mathcal Re(\lambda_{k_{v+1}})-\beta(\tau)]d\tau\\
&\le&\frac{1}{t-a}\int_{a}^{t} [\gamma(\tau)\lambda_{1}-\beta(\tau)]d\tau \\
&\le&(\bar{\gamma}+\epsilon)\lambda_{1}-(\bar{\beta}+\epsilon)=-\varphi<0
\end{eqnarray*}
for all $t$ and $a$ with $t-a>T$. This implies that the
first term $y_{v+1}(0)\exp(\int_{0}^{t}k(\tau)d\tau)$ converges to
zero. Let $M>0$ be a constant such that
$\sup\limits_{\tau,j}|\gamma(\tau)\lambda_{j}-\beta(\tau)|<M$. For the
second term, we have
\begin{eqnarray*}
&&\bigg|\int_{0}^{t}\xi_{v+1}(a)\exp\left(\int_{a}^{t}k(\tau)d\tau\right)da\bigg| \\
&=&\bigg|\int_{0}^{t-T}\xi_{v+1}(a)\exp\left(\int_{a}^{t}k(\tau)d\tau\right)da
+\int_{t-T}^{t}\xi_{v+1}(a)\exp\left(\int_{a}^{t}k(\tau)d\tau\right)da\bigg|\\
&\le&\int_{0}^{t-T}|\xi_{v+1}(a)|\exp\left(-\varphi
(t-a)\right)da+\int_{t-T}^{t}|\xi_{v+1}(a)|\exp(MT)da.
\end{eqnarray*}
In the case of $\xi_{v+1}(t)=0$, we immediately conclude that
$\lim_{t\to\infty}y_{v+1}(t)=0$. Otherwise, according to the
condition and the L'Hospital principle, we have
\begin{eqnarray*}
\lim_{t\to\infty}\int_{0}^{t-T}|\xi_{v+1}(a)|\exp\big[-\varphi (t-a)\big]da &=&
\lim_{t\to\infty}\frac{\int_{0}^{t-T}|\xi_{v+1}(a)|\exp(\varphi a)da}{\exp(\varphi t)} \\
&=&\lim_{t\to\infty}\frac{|\xi_{v+1}(t-T)|\exp(\varphi(t-T))}{\varphi\exp(\varphi t)}\\
&=&\lim_{t\to\infty}\frac{|y_{v}(t-T)\gamma(t-T)|\exp(-\varphi T))}{\varphi}=0,
\end{eqnarray*}
due to the assumption $\lim_{t\to\infty}y_{v}(t)=0$. We also have
\begin{eqnarray*}
\lim_{t\to\infty}\int_{t-T}^{t}|\xi_{v+1}(a)|\exp(MT)da=0.
\end{eqnarray*}
Therefore, we can conclude $\lim_{t\to\infty}y_{v+1}(t)=0$.
Note that condition (\ref{unstable}) implies that system
(\ref{Eq.2.2}) is unstable. Since system (\ref{Eq.2.2}) is in
fact the linearization system of Eq. (\ref{Eq.2}), we can conclude
that system (\ref{Eq.2}) is unstable under condition
(\ref{unstable}). This completes the proof.
\qed
\end{proof}
If $\beta(t)$ and $\gamma(t)$ are ergodic stochastic processes, from
the multiplicative ergodic theory of the random dynamical
systems \cite{Arn}, we have
the following result as a corollary of Theorem \ref{thm4}.
\begin{corollary}
\label{theorem-main-result-2}
\emph{(another succinct sufficient condition)}
Suppose $\beta_v(t)=\beta(t)$ for all $v \in V$ (i.e., all nodes
have the same cure capability) and $E(t)=E$ for any time $t$ (i.e.,
topology does not change over time). Suppose $\{\beta(t)\}_{t \ge
0}$ and $\{\gamma(t)\}_{t \ge 0}$ are ergodic stochastic processes
(i.e., $\beta(t)$ and $\gamma(t)$ are some random variables). Let
$\E(\beta(0))$ and $\E(\gamma(0))$ be the expectations with respect
to the stationary distributions of the respective ergodic stochastic
process. Suppose the convergences
\begin{eqnarray*}
\E(\beta(0))=\frac{1}{t}\int_{t_{0}}^{t_{0}+t}\beta(\tau)d\tau,~
\E(\gamma(0))=\frac{1}{t}\int_{t_{0}}^{t_{0}+t}\gamma(\tau)d\tau,
\end{eqnarray*}
are both uniform with respect to $t_{0}$ {\em almost surely}.
If
\begin{eqnarray*}
\lambda_{1}<\frac{\E(\beta(0))}{\E(\gamma(0))},
\end{eqnarray*}
the spreading will die out almost surely regardless of the initial infection configuration;
if
\begin{eqnarray*}
\lambda_{1}>\frac{\E(\beta(0))}{\E(\gamma(0))},
\end{eqnarray*}
the spreading will not die out in some initial infection configurations.
\end{corollary}
\noindent{\bf Discussion}. The state-of-the-art sufficient
condition for the dying out of virus spreading in an arbitrary network is
$\lambda_1 < \frac{\beta}{\gamma}$, which was given in \cite{WangTISSEC08}.
In the setting of \cite{WangTISSEC08}, the parameters satisfy that $\beta_v(t)=\beta$ for all $v\in V$ and all $t$,
and $\gamma(t)=\gamma$ for all $t$.
As such, their result is clearly a special case of the above Corollary \ref{theorem-main-result-2}
(note that it is guaranteed that $\E(\gamma(0))\neq 0$),
and thus of the above Theorem \ref{thm4}.
\subsection{Adaptive control in the scenario of fully-adaptive scenario}
In the semi-adaptive defense scenario investigated above, we assumed
that the parameters $\gamma(t)$ and $\beta_v(t)$ are given.
What if they are not given?
In what follows we investigate a representative scenario, where
the defender is not given $\gamma(t)$ but can observe $i_v(t)$.
Specifically, we consider two sufficient conditions of adaptive control:
one under which the virus spreading will die out (Section \ref{sec:adaptive-control-1}),
and another under which the virus spreading
will not die out but will be contained to a desired level of infection (Section \ref{sec:adaptive-control-2}).
\subsubsection{Sufficient condition under which the virus spreading dies out under adaptive control}
\label{sec:adaptive-control-1}
The question we ask is: How should the defender adjust the defense,
namely how $\beta_v(t)$ should depend upon $i_v(t)$, so
that the virus spreading will die out? We assume for concreteness that
$\beta_v(0)=0$ for all $v\in V$; this accounts for the worst-case scenario.
\begin{theorem}
\label{thm_stablize1}
\emph{(characterization of adaptive control strategy under which the virus spreading will die out)}
Suppose without loss of generality $\beta_v(0)=0$ for all $v\in V$.
If
\begin{eqnarray*}
\frac{d\beta_{v}(t)}{dt}=\rho\cc_{v}(t),
\end{eqnarray*}
where $\rho$ is an (almost) arbitrary positive constant, then the virus spreading will die out
regardless of the initial infection configuration.
\end{theorem}
\begin{proof}
Define a candidate Lyapunov function with respect to the infectious
probabilities $\cc=[\cc_{1},\cdots,\cc_{n}]^{\top}$ and the cure
capabilities
$\tilde{\beta}=[\beta_{1},\cdots,\beta_{n}]^{\top}$:
\begin{eqnarray*}
V(\cc,\tilde{\beta})=\sum_{v=1}^{n}i_{v}(t)+\frac{1}{2\rho}\sum_{v=1}^{n}(\beta_{v}(t)-\beta_{0,v})^{2}.
\end{eqnarray*}
Let $\beta_{0,v}$, $v=\oneton$, be positive constants satisfying
\begin{eqnarray*}
\beta_{0,v}>(n-1)\sup_{t}\gamma(t)+1,~\forall~v=\oneton
\end{eqnarray*}
owing to the fact that $1\ge \gamma(t)\ge 0$.
Due to inequality (\ref{ineq}), differentiating $V(\cc,\tilde{\beta})$ gives
\begin{eqnarray*}
\frac{dV(i,\tilde{\beta})}{dt}&=&\sum_{v=1}^{n}\left[1-\prod_{(u,v)\in
E{(t)}}(1-\gamma(t)\cc_{u}(t))\right](1-\cc_{v}(t))-\sum_{v=1}^{n}\beta_{v}(t)\cc_{v}(t) \\
&&+\sum_{v=1}^{n}(\beta_{v}(t)-\beta_{0,v})\cc_{v}(t)\\
&\le&\sum_{v=1}^{n}\gamma(t)\sum_{u=1}^{n}a_{vu}(t)\cc_{u}(t)-\sum_{v=1}^{n}\beta_{v}(t)\cc_{v}(t)
+\sum_{v=1}^{n}(\beta_{v}(t)-\beta_{0,v})\cc_{v}(t)\\
&\le&\sum_{v=1}^{n}\gamma(t)\sum_{u=1}^{n}a_{vu}(t)\cc_{u}(t)-\sum_{v=1}^{n}\beta_{0,v}\cc_{v}(t)\\
&\le&\gamma(t)(n-1)\sum_{u=1}^{n}\cc_{u}(t)-\sum_{v=1}^{n}\beta_{0,v}\cc_{v}(t) \\
&\le& -\sum_{v=1}^{n}\cc_{v}(t).
\end{eqnarray*}
According to the LaSalle principle \cite{LaSalle1960}, the
system converges to the largest invariant set
$\{(\cc,\tilde{\beta}):\sum_{v=1}^{n}\cc_{v}=0\}$, which implies
$\lim_{t\to\infty}\cc_{v}(t)=0$ for all $v=\oneton$. \qed
\end{proof}
The above Theorem \ref{thm_stablize1} has the following implications.
\begin{proposition}
\label{proposition-adaptive-1}
We can bound from above the accumulated number of infected nodes in the long run (a node
is counted multiple times if it is infected at multiple points in time) as follows:
\begin{eqnarray}
\label{est1} \int_{0}^{\infty}\sum_{v=1}^{n}\cc(\tau)d\tau\le
\frac{(n-1)^{2}\gamma_{m}}{\rho}+\frac{n-1}{\rho}\sqrt{\sum_{v=1}^{n}\cc_{v}(0)
+\frac{(n-1)^{3}\gamma_{m}^{2}}{2\rho}}
\end{eqnarray}
where $\gamma_{m}=\sup\limits_{t}\gamma(t)$ and
$\beta_{0,m}=\min\limits_{v}\beta_{0,v}$.
\end{proposition}
\begin{proof}
From the proof of Theorem \ref{thm_stablize1}, we have
\begin{eqnarray*}
\frac{dV(i,B)}{dt}\le
[\gamma_{m}(n-1)-\beta_{0,m}]\sum_{v=1}^{n}\cc_{v}(t),
\end{eqnarray*}
which implies
\begin{eqnarray*}
\int_{0}^{t}\sum_{v=1}^{n}\cc_{v}(\tau)d\tau\le\frac{1}{\beta_{0,m}-\gamma_{m}(n-1)}[V(0)-V(t)]\le
\frac{1}{\beta_{0,m}-\gamma_{m}(n-1)}V(0).
\end{eqnarray*}
If $\beta_{v}(0)=0$ for all $v=\oneton$ and all
$\beta_{v,0}$ are the same and are thus denoted by $\beta_{0}$, we have the
following estimation:
\begin{eqnarray*}
\int_{0}^{\infty}\sum_{v=1}^{n}\cc(\tau)d\tau\le\frac{1}{\beta_{0}-\gamma_{m}(n-1)}\big[\sum_{v=1}^{n}
\cc_{v}(0)+\frac{1}{2\rho}(n-1)\beta_{0}^{2}\big].
\end{eqnarray*}
By setting
$$\beta_{0}=a+\frac{\sqrt{\sum_{v=1}^{n}\cc_{v}(0)+1/(2\rho)(n-1)^{3}\gamma_{m}^{2}}}{1/(2\rho)(n-1)},$$
we obtain the minimum of the right-hand side and thus complete the proof. \qed
\end{proof}
\noindent{\bf Physical meanings of Proposition \ref{proposition-adaptive-1}}.
The term at the left-hand side of inequality (\ref{est1}) captures,
in addition to the aforementioned estimation of the total number of infected nodes
in the network (counting repetition) over time, the convergence rate of the adaptive control strategy.
This allows us to draw the following insights:
(i) The larger $\rho$, the faster the
virus spreading will die out;
(ii) the larger degree of initial infection,
the slower the virus spreading will die out;
(iii) the larger edge infection probability $\gamma$,
the slower the virus spreading will die out.
\subsubsection{Adaptive control under which the virus spreading will not die out but will be contained to a desired level}
\label{sec:adaptive-control-2}
In the above we have given some sufficient condition on adjusting the defense
or $\beta_v(t)$ so that the virus spreading will die out.
What if the required $\beta_v(t)$ cannot be achieved, meaning that
we may not expect that the virus spreading die out?
This is possible because
the defense may not be as good as one may wish or because of budget limitation.
In this case, we ask an alternative interesting question:
What it takes so that $i_v(t)$ can converge or be contained
to some pre-determined level of infection $\cc_{v}^{*}$?
\begin{theorem}
\label{theorem-control-to-specified-state}
\emph{(characterization of adaptive control strategy under which the virus spreading will be contained to a desired level of infection)}
Consider the following variant of the master equation Eq. (\ref{Eq.2}),
\begin{eqnarray}
\label{eq-adaptive-control-2}
\frac{d\cc_{v}(t)}{dt}&=&\bigg[1-\prod_{(u,v)\in E}[1-\gamma
\cdot \cc_{u} {(t)}]\bigg](1-\cc_{v} {(t)})-\beta_{v}(t)\cc_{v} {(t)}+w_{v}
\end{eqnarray}
For any $1\ge \cc_{v}^{*}> 0$, $v=\oneton$, letting
\begin{eqnarray*}
\beta_{v}^{*}=\frac{[1-\prod_{(u,v)\in E}(1-\gamma \cdot
\cc_{u}^{*})](1-\cc_{v}^{*})}{\cc_{v}^{*}},
\end{eqnarray*}
if we use the following adaptive control strategy
\begin{eqnarray*}
\frac{d\beta_{v}(t)}{dt}=\rho[\cc_{v}(t)-\cc_{v}^{*}]\cc_{v}(t), ~~~~~~
w_{v}=\eta[\cc_{v}^{*}-\cc_{v}(t)],
\end{eqnarray*}
where $\rho$ is an (almost) arbitrary positive constant, and $\eta$
is a positive constant with
$\eta+\min\limits_{v}\beta_{v}^{*}>1+(1-\gamma)\lambda_{1}$ where
$\lambda_{1}$ is the largest (in modulus) eigenvalue of the
adjacency matrix $A$, we have
$\lim_{t\to\infty}\cc_{v}(t)=\cc_{v}^{*}$.
\end{theorem}
\begin{proof}
From Perron-Frobenius theorem \cite{Berman03}, we have that there
exist some positive constants $P_{1},\cdots,P_{n}$ such that $P={\rm
diag}[P_{1},\cdots,P_{n}]$ satisfies
$[PA+A^{\top}P]\le\lambda_{1}P$. Consider the candidate Lyapunov
function with respect to $\cc=[\cc_{1},\cdots,\cc_{n}]^{\top}$ and
$\tilde{\beta}=[\beta_{1},\cdots,\beta_{n}]^{\top}$:
\begin{eqnarray*}
V(\cc,\tilde{\beta})=\frac{1}{2}\sum_{v=1}^{n}P_{v}[\cc_{v}-\cc_{v}^{*}]^{2}+\frac{1}{2\rho}\sum_{v=1}^{n}P_{v}
(\beta_{v}-\beta_{v}^{*})^{2}.
\end{eqnarray*}
Differentiating $V(x,\tilde{\beta})$ gives
\begin{eqnarray*}
&&\frac{d{V}(\cc,\tilde{\beta})}{dt}\\
&=&\sum_{v=1}^{n}P_{v}[\cc_{v}(t)-\cc_{v}^{*}]\bigg\{[1-\prod_{(u,v)\in
E}(1-\gamma \cdot \cc_{u} {(t)})](1-\cc_{v}
{(t)})-\beta_{v}(t)\cc_{v} {(t)}-\eta(\cc_{v}-\cc_{v}^{*})\bigg\}\\
&&+\sum_{v=1}^{n}P_{v}(\beta_{v}(t)-\beta_{v}^{*})(\cc_{v}(t)-\cc_{v}^{*})\cc_{v}(t)\\
&=&\sum_{v=1}^{n}P_{v}[\cc_{v}(t)-\cc_{v}^{*}]\bigg\{\big[1-\prod_{(u,v)\in
E}(1-\gamma \cdot \cc_{u} {(t)})\big](1-\cc_{v} {(t)}) \\
&&-\big[1-\prod_{(u,v)\in E}(1-\gamma \cdot \cc_{u}^{*}
)\big](1-\cc_{v}^{*})
-\beta_{v}(t)\cc_{v} {(t)}-\beta_{v}^{*}\cc_{v}^{*}+\eta(\cc_{v}-\cc_{v}^{*})\bigg\}\\
&&+\sum_{v=1}^{n}P_{v}(\beta_{v}(t)+\beta_{v}^{*})(\cc_{v}(t)-\cc_{v}^{*})\cc_{v}(t).
\end{eqnarray*}
Note that
\begin{eqnarray*}
&&\bigg|[1-\prod_{(u,v)\in E}(1-\gamma \cdot \cc_{u}
{(t)})](1-\cc_{v} {(t)})-[1-\prod_{(u,v)\in E}(1-\gamma
\cdot \cc_{u}^{*} )](1-\cc_{v}^{*})\bigg|\\
&\le&\bigg|(\cc_{v}(t)-\cc_{v}^{*})\prod_{(u,v)\in
E}(1-\gamma\cc_{v}(t))\bigg| +\bigg|(1-\cc_{v}^{*})\big[\prod_{(u,v)\in
E}(1-\gamma\cc_{u}(t))-\prod_{(u,v)\in
E}(1-\gamma\cc_{u}^{*})\big]\bigg|\\
&\le&|\cc_{v}(t)-\cc_{v}^{*}|+\sum_{(u,v)\in
E}|\cc_{u}(t)-\cc_{u}^{*}|\prod_{(u_{1},v)\in
E,~u_{1}>u}|1-\gamma\cc_{u_{1}}^{*}|\prod_{(u_{2},v)\in
E,~u_{2}<u}|1-\gamma\cc_{u_{2}}(t)|\\
&\le&|\cc_{v}(t)-\cc_{v}^{*}|+\gamma\sum_{(u,v)\in
E}|\cc_{u}(t)-\cc_{u}^{*}|.
\end{eqnarray*}
Thus, we have
\begin{eqnarray*}
\frac{d{V}(\cc,\tilde{\beta})}{dt} &\le&\sum_{v=1}^{n}
P_{v}|\cc_{v}(t)-\cc_{v}^{*}|^{2}+\sum_{v=1}^{n}\sum_{u=1}^{n}\gamma
P_{v}a_{vu}|\cc_{v}(t)-\cc_{v}^{*}||\cc_{u}(t)-\cc_{u}^{*}|\\
&&-\sum_{v=1}^{n}\beta_{v}^{*}P_{v}(\cc_{v}(t)-\cc_{v}^{*})^{2}
-\eta\sum_{v=1}^{n}P_{v}[\cc_{v}(t)-\cc_{v}^{*}]^{2}.
\end{eqnarray*}
Let $z(t)=[z_{1}(t),\cdots,z_{n}(t)]^{\top}$ with
$z_{v}(t)=|\cc_{v}(t)-\cc_{v}^{*}|$. Since $\beta_{v}^{*}\ge 0$, we have
\begin{eqnarray*}
\frac{d{V}(\cc,\tilde{\beta})}{dt}&\le& z^{\top}\big[
P+\gamma(PA+A^{\top}P)/2-\eta P-B^{*}P\big]z\\
&\le&
z^{\top}\big[1+\gamma\lambda_{1}-\eta-\min_{v}\beta_{v}^{*}\big]P
z<0~,\forall~z\ne 0,
\end{eqnarray*}
where $B^{*}={\rm diag}[\beta_{1},\cdots,\beta_{n}]$. Due to the
LaSalle principle \cite{LaSalle1960}, we have that the system will
converge to the largest invariant set in
$\{(\cc,\tilde{\beta}):~\sum_{v=1}^{n}[\cc_{v}-\cc_{v}^{*}]^{2}=0\}$,
which implies that $\lim_{t\to\infty}\cc_{v}(t)=\cc_{v}^{*}$ for all
$v=\oneton$. \qed
\end{proof}
\noindent{\bf Discussion}. The above theorem is quite general because of the term $w_v$
in Eq. (\ref{eq-adaptive-control-2}). If $w_v \neq 0$, the adaptive control strategy must be used with caution
because we must guarantee that its value does have physical meanings. In general, $w_v \le 0$ would be reasonable;
in our simulation study (Section \ref{sec:simulation-study-alpha-eq-0}),
we set $w_v = 0$ for simplicity. On the other hand, the theorem is necessarily
based on the premise that $\gamma(t)=\gamma$, namely that $\gamma(t)$ does not vary
with respect to time, because of the way $\beta_v^*$ is defined.
In our simulation study (Section \ref{sec:simulation-study-alpha-eq-0}), we will show that the result is quite robust, meaning
that even if $\gamma(t)$ varies with respect to time (as we considered in Section \ref{sec:implicit-adaptive-case}),
the result is still valid.
This is very important because the fixed $\gamma$ can be seen as, in a sense, the average of the unknown
$\gamma(t)$ over time.
Similar to Proposition \ref{proposition-adaptive-1}, we can have
\begin{proposition}
\label{prop:adaptive-control-2}
We have
\begin{eqnarray}
\int_{0}^{\infty}\sum_{v=1}^{n}|\cc_{v}(\tau)-\cc_{v}^{*}|d\tau
&\le&\frac{1}{\min_{v}P_{v}(\eta-1-\gamma\lambda_{1})}
\big[\frac{1}{2}\sum_{v=1}^{n}P_{i}|\cc_{v}(0)-\cc_{v}^{*}|^{2} \nonumber\\
&&+\frac{1}{2\rho}\sum_{v=1}^{n}
|\beta_{v}(0)-\beta_{v}^{*}|^{2}\big]\label{est3}.
\end{eqnarray}
\end{proposition}
\noindent{\bf Physical meanings of Proposition \ref{prop:adaptive-control-2}}. The above proposition offers the following insights:
The larger $\rho$, the larger $\eta$, the smaller $\lambda_{1}$,
the smaller $\gamma$, the smaller differential between $\cc_{v}(0)$ and
$\cc_{v}^{*}$, or the smaller the differential between $\beta_{v}(0)$ and $\beta_{v}^{*}$,
the faster the virus spreading will die out.
\section{Simulation Study}
\label{sec:simulation-study-alpha-eq-0}
We conduct simulation to complement our analytic study for two purposes.
First, we want to confirm our analytical results offered in Section \ref{sec:model-alpha-eq-0}.
Second, we want to draw some relevant observations that are not offered by our analytic results.
Such observations may guide future studies of analytic models (e.g., how to enhance them so
that other useful insights may be obtained analytically).
As mentioned above, our model is very general because it accommodates dynamical graph topology $G(t)=(V,E(t))$.
However, it's not clear at this stage
how to appropriately define a physically meaningful way according to which
the topology changes. Therefore we leave the full-fledged characterization (beyond what is implied
by our analytical results) of the impact of dynamical topology to future work.
We conducted simulations using both synthetic (regular, random, and power-law) graphs
and a real network graph. Due to space limitation, here we report the simulation results in the latter
case (but all the simulation results are consistent). The real network graph $G=(V,E)$
is based on the Oregon router views (available from {\tt http://topology.eecs.umich.edu/data.html}),
where $|V|=11,461$ representing AS peers, $|E|=32,730$ representing links between the AS peers.
The largest eigenvalue of the corresponding adjacency matrix is $\lambda_1=75.2407$.
\subsection{Methodology}
Our simulation is conducted in an event-driven fashion.
For the purpose of studying the dynamics under our adaptive control strategies, we need to measure $i_v(t)$,
the probability that node $v\in V$ is infected at time $t$. This parameter can be
obtained in a tedious way (i.e., by conducting for example 100 simulation runs in parallel, rather than in sequence,
because we need to count the number of times each node $v$ is infected at each time step).
A much more simpler way however is to use our model
formula Eq. (\ref{Eq.2}) to compute $i_v(t)$ instead, as long as the model is accurate.
To confirm the accuracy of the model, we compare it with simulation.
For simplicity, we let both $\beta(t)$ and $\gamma(t)$ be some periodic functions with period $T$, which means that
both attack and defense vary with respect to time. We consider
three settings: $\beta(t)$ and $\gamma(t)$ being synchronous, asynchronous, or anti-synchronous
because we want to observe whether, and if so to what extent, the degree of (a)synchrony has an impact on the outcome.
Specifically, we consider two sets of parameters:
$\beta(t)\in \{0.3,0.5\}$ and $\gamma(t) \in \{0.003,0.007\}$
according to the functions shown in Figure \ref{fig:beta-gamma-t-model};
$\beta(t)\in \{0.005,0.015\}$ and $\gamma(t) \in \{0.003,0.007\}$ in the same fashion.
In the asynchronous case, we let $\beta_v(t)$ is $T/4$ behind $\gamma_v(t)$
because cure often comes after attack is identified.
To draw insights into whether the period $T$ has an impact on the outcome, we
consider $T = 8, ~16$, respectively.
In any case, it is clear that $\beta(t)$ is an implicit function of $\gamma(t)$.
\begin{figure*}[ht]
\centering
\subfigure[synchronous setting]
{\label{fig:beta-gamma-t-model-synchronous-case}
\includegraphics[width=.32\textwidth]{beta-gamma-t-model-synchronous-case-new.eps}}
\subfigure[asynchronous setting]
{\label{fig:beta-gamma-t-model-asynchronous-case}
\includegraphics[width=.32\textwidth]{beta-gamma-t-model-asynchronous-case-new.eps}}
\subfigure[anti-synchronous setting]
{\label{fig:beta-gamma-t-model-antisynchronous-case}
\includegraphics[width=.32\textwidth]{beta-gamma-t-model-antisynchronous-case-new.eps}}
\caption{Examples of synchronous, asynchronous, and counter-synchronous $\beta(t)$ and $\gamma(t)$}
\label{fig:beta-gamma-t-model}
\end{figure*}
Figure \ref{fig:Model-accuracy} plots the curves obtained by simulation and by model computing
in the case of both $\beta_v(t)$ and $\gamma_v(t)$ have period $T=8$ and $T=16$ (as shown in Figure \ref{fig:beta-gamma-t-model}).
In each graph, we let the virus initially infect 2,292 or 20\% vertices that are randomly selected;
note that the degree of initial infection does not impact whether the virus spreading will die out or not.
Since the model computing and simulation results (obtained as the average of 50 simulation runs)
match almost perfectly no matter the virus spreading will die out or not,
we will use simulation and model computing interchangeably. Since the same phenomenon applies to
both cases of $T=8$ and $T=16$, in what follows we only report the case of $T=8$.
Again, the accuracy result allows us to obtain $i_v(t)$ via model computing in the process of confirming
the analytical results of our adaptive control strategies.
\begin{figure*}[ht]
\centering
\subfigure[T=8]
{\label{fig:model-accuracy-T=8}
\includegraphics[width=.48\textwidth]{oregon_Model_compare_T=8.eps}}
\subfigure[T=16]
{\label{fig:model-accuracy-T=16}
\includegraphics[width=.48\textwidth]{oregon_Model_compare_T=16.eps}}
\caption{Model accuracy}
\label{fig:Model-accuracy}
\end{figure*}
\subsection{Confirmation of the sufficient conditions in the semi-adaptive scenario}
In this section, we use the aforementioned Oregon graph to confirm
our analytical results presented in Section \ref{sec:implicit-adaptive-case}.
We confirm Theorem \ref{thm4} and Corollary \ref{theorem-main-result-2} because they
offer succinct sufficient conditions under which the virus spreading dies out.
\subsubsection{Confirmation of Theorem \ref{thm4}}
For the graph, we let the virus initially infect 2,292 or 20\% randomly selected nodes,
and consider three cases between the model's input parameter --- synchrony, asynchrony, and anti-synchrony
as illustrated in Figure \ref{fig:beta-gamma-t-model}.
Since Theorem \ref{thm4} has two parts, we confirm them respectively.
\begin{figure*}[ht]
\centering
\subfigure[Parameter set (i)]
{\label{fig:global-analysis-regular-graph-case-1-virus-th4-part1-rl4-1}
\includegraphics[width=.48\textwidth]{oregon_Model_th4_part1_rl4_1.eps}}
\subfigure[Parameter set (ii)]
{\label{fig:global-analysis-random-graph-case-1-virus-th4-part1-rl4-2}
\includegraphics[width=.48\textwidth]{oregon_Model_th4_part1_rl4_2.eps}}
\caption{Confirmation of sufficient condition under which virus spreading dies out ($T = 8$)}
\label{fig:J-1-Theorem-4-case1-T=8}
\end{figure*}
\noindent{\bf Case 1: Confirmation of the sufficient condition under which the virus spreading will die out}.
We consider two sets of parameters:
(i) $\beta(t)\in \{0.3,0.5\}$ and $\gamma(t) \in \{0.003,0.007\}$;
(ii) $\beta(t)\in \{0.1,0.22\}$ and $\gamma(t) \in \{0.001,0.003\}$.
Both functions, $\beta(t)$ and $\gamma(t)$, have period $T=8$.
Both parameter sets satisfy the sufficient condition of Theorem \ref{thm4}, namely $\lambda_1 < \bar{\beta}/\bar{\gamma}$,
which means that the virus spreading will die out.
Figure \ref{fig:J-1-Theorem-4-case1-T=8} plots the dynamics of the numbers of infected nodes with respect to time.
From Figure \ref{fig:J-1-Theorem-4-case1-T=8} we can draw the following observations.
First, the virus spreading does die out at about the 50th and 25th step, respectively, which confirms the sufficient condition
under which the virus spreading will die out. It is an interesting future work to quantitatively
characterize how the speed of convergence (i.e., dying out) depends upon functions $\beta(t)$ and $\gamma(t)$.
Second, it is counter-intuitive and interesting that the virus spreading is somewhat more effectively
defended against in the anti-synchronous case than in the synchronous case, which is in turn more
effectively defended against than in the asynchronous case. More studies
are needed in order to explain this phenomenon.
Third, the curves are convex, meaning that cure is more effective in the early stage
of the attack-defense dynamics than in the later stage. For example, it takes
a shorter period of time to reduce the infection from 2,292 nodes to 100 nodes
than to reduce the infection from 100 nodes to zero nodes (i.e., dying out).
\smallskip
\noindent{\bf Case 2: Confirmation of the condition under which the virus spreading may not die out}.
We consider two sets of parameters:
(i) $\beta(t)\in \{0.2,0.4\}$ and $\gamma(t) \in \{0.003,0.007\}$;
(ii) $\beta(t)\in \{0.05,0.15\}$ and $\gamma(t) \in \{0.001,0.003\}$.
Both functions, $\beta(t)$ and $\gamma(t)$, have period $T=8$.
Both parameter sets do not satisfy the sufficient condition of Theorem \ref{thm4}
because $\lambda_1 > \bar{\beta}/\bar{\gamma}$, which means that the virus spreading does not die out
in some initial infection configurations.
\begin{figure*}[ht]
\centering
\subfigure[Parameter set (i)]
{\label{fig:global-analysis-regular-graph-case-1-virus-th4-part2-rl4-1}
\includegraphics[width=.48\textwidth]{oregon_Model_th4_part2_rl4_1.eps}}
\subfigure[Parameter set (ii)]
{\label{fig:global-analysis-random-graph-case-1-virus-th4-part2-rl4-2}
\includegraphics[width=.48\textwidth]{oregon_Model_th4_part2_rl4_2.eps}}
\caption{Confirmation of virus spreading not dying out ($T = 8$)}
\label{fig:J-1-Theorem-4-case2-T=8}
\end{figure*}
Figure \ref{fig:J-1-Theorem-4-case2-T=8} plots the dynamics, from which we draw the following observations.
First, the virus spreading does not die out, which confirms Theorem \ref{thm4}.
Second, all the curves exhibit periodic behaviors, but the extent of oscillation in the
case of anti-synchrony is more significant than in the case of asynchrony, which
in turn is more significant than in the case of synchrony. This means that the degree of
synchrony between $\beta(t)$ and $\gamma(t)$ will impact the outcome when the virus spreading does not
die out. Third, comparing Figures \ref{fig:global-analysis-regular-graph-case-1-virus-th4-part2-rl4-1}
and \ref{fig:global-analysis-random-graph-case-1-virus-th4-part2-rl4-2}, we observe that,
under the same synchrony, the outcome will depend on functions $\beta(t)$ and $\gamma(t)$.
More studies are needed to characterize these dependence relationships.
\subsubsection{Confirmation of Corollary \ref{theorem-main-result-2}}
Corollary \ref{theorem-main-result-2} gives an even more succinct sufficient condition
under which the virus spreading will die out. For the graph, we let the virus initially infect
2,291 or 20\% randomly selected nodes.
For the case the sufficient condition in Corollary \ref{theorem-main-result-2} is satisfied,
we consider two sets of parameters that are uniformly chosen at random from certain intervals:
(i) $\beta(t)\in [0.7, 0.9]$ and $\gamma(t) \in [0.006, 0.0014]$;
(ii) $\beta(t) \in [0.1, 0.3]$ and $\gamma(t) \in [0.0015, 0.0035]$.
In each parameter setting, the sufficient condition stated in Corollary \ref{theorem-main-result-2} is satisfied,
namely $\lambda_1 < \E(\beta(0))/\E(\gamma(0))$, which means that the virus spreading will die out.
For the case the sufficient condition in Corollary \ref{theorem-main-result-2} is not satisfied,
we consider two sets of parameters that are uniformly chosen at random from certain intervals:
(i) $\beta(t) \in [0.4, 0.6]$ and $\gamma(t) \in [ 0.006, 0.014]$;
(ii) $\beta(t) \in [0.05, 0.15]$ and $\gamma(t) \in [0.0015, 0.0035]$.
In each parameter setting, the sufficient condition stated in Corollary \ref{theorem-main-result-2} is
not satisfied because $\lambda_1 > \E(\beta(0))/\E(\gamma(0))$,
and thus the analytical result says that the virus spreading does not die out
in some initial infection configurations.
\begin{figure*}[ht]
\centering
\subfigure[The sufficient condition is satisfied]
{\label{fig:global-analysis-regular-graph-case-1-virus-th5-part1-J1-rl4}
\includegraphics[width=.48\textwidth]{oregon_Model_th5_part1_rl4.eps}}
\subfigure[The sufficient condition is not satisfied]
{\label{fig:global-analysis-random-graph-case-1-virus-th5-part2-J1-rl4}
\includegraphics[width=.48\textwidth]{oregon_Model_th5_part2_rl4.eps}}
\caption{Confirmation of Corollary \ref{theorem-main-result-2} ($T = 8$)}
\label{fig:J-1-corollary-1}
\end{figure*}
Figure \ref{fig:global-analysis-regular-graph-case-1-virus-th5-part1-J1-rl4} plots the
dynamics of the number of infected nodes with respect to time when the sufficient condition is satisfied.
From it we can draw the following observations.
First, the virus spreading does die out, as predicted by Corollary \ref{theorem-main-result-2}.
Second, the larger the $\beta(t)$, the more effective the defense against the virus spreading.
Third, all the curves are convex, meaning that it takes a shorter period of time
to significantly reduce the number of infected nodes (e.g., from 2,291 to 50)
than to making the virus spreading die out (e.g., from 50 to zero).
Figure \ref{fig:global-analysis-random-graph-case-1-virus-th5-part2-J1-rl4}
plots the dynamics of the number of infected nodes with respect to time when the sufficient condition is not satisfied.
From it we can draw the following observations.
First, the virus spreading does not die out, which confirms Corollary \ref{theorem-main-result-2}.
Second, the larger (in a stochastic sense) the $\beta(t)$, the earlier the system will converge to the steady state.
However, the ultimate degree of infection does not depend on $\beta(t)$, but rather on
$\E(\beta(0))/\E(\gamma(0))$, which means that $\E(\beta(0))/\E(\gamma(0))$ may be used as an indicator of
steady-state infection when the virus spreading does not die out. It is an interesting future work
to rigorously characterize this phenomenon.
\subsection{Confirmation of the controllability in the fully-adaptive scenario}
\label{sec:confirmation-adaptive-control}
\subsubsection{Confirmation of Theorem \ref{thm_stablize1}}
Theorem \ref{thm_stablize1} states that even if we do not know $\gamma(t)$ but we
may be able to observe $i_v(t)$ and may be able to adjust the defense as needed,
following its control strategy will cause the dying out of the virus spreading.
To compare the effects of adaptive control and semi-adaptive control,
in our simulation study, we also used the periodical functions $\beta(t) \in \{0.375, 0.40\}$
and $\gamma(t) \in \{0.003, 0.007\}$
with period $T=8$ as illustrated in Figure \ref{fig:beta-gamma-t-model}.
These parameters satisfy the sufficient condition in Theorem \ref{thm4}, which means
that the virus spreading will die out as we discussed above.
To ensure comparability, we also let the virus initially infect 2,291 or 20\% randomly selected nodes.
As mentioned before, since the anti-synchronous defense is somewhat
more effective than the synchronous and asynchronous defenses,
we will compare it with the outcome of the adaptive control strategy.
\begin{figure*}[ht]
\centering
\subfigure[Fully-adaptive control]
{\label{fig:fully-adaptive control-th4}
\includegraphics[width=.48\textwidth]{oregon_Model_th6_rl4.eps}}
\subfigure[Comparison of cure capabilities]
{\label{fig:cure capabilities-th4}
\includegraphics[width=.48\textwidth]{oregon_th6_beta.eps}}
\caption{Confirmation of semi-adaptive dying out vs. fully-adaptive control}
\label{fig:confirming-adaptive-control-1}
\end{figure*}
Figure \ref{fig:fully-adaptive control-th4} plots the dynamics of the number of infected nodes
with respect to time in the following four cases: the adaptive control parameter $\rho=0.005$;
the adaptive control parameter $\rho=0.01$; the adaptive control parameter $\rho=0.02$;
the comparison dynamics corresponding to the anti-synchronous case
with $T=8$ periodical function $\beta(t) \in \{0.375, 0.40\}$.
We draw the following observations.
First, $\rho$ plays a crucial role in indicating the rate at which the virus spreading dies out.
For example, for $\rho=0.02$, it takes only about 80 steps to reduce the number of infected nodes from 2,291 to 160
(nevertheless it takes another 60 steps to kill the virus spreading, namely to reduce the number of infected nodes from 160 to zero);
for $\rho=0.01$, it takes about 130 steps to reduce the number of infected nodes from 2,291 to 160
(nevertheless it takes about another 100 steps to kill the virus spreading, namely to reduce the number of infected
nodes from 160 to zero).
This also confirms the physical meanings of Proposition \ref{proposition-adaptive-1} discussed above.
Second, Figure \ref{fig:fully-adaptive control-th4} indicates that
for all $\rho=0.005$, $\rho=0.01$ and $\rho=0.02$, the fully-adaptive defenses are less effective than
the semi-adaptive defense represented by $\beta(t) \in \{0.375, 0.40\}$.
As we show in Figure \ref{fig:cure capabilities-th4},
this is caused by the fact that the semi-adaptive $\beta(t)$ is much larger than the adaptive $\beta(t)$.
This means that the sufficient condition in the semi-adaptive case, under which the virus spreading dies out,
may be significantly beyond being necessary. In contrast, the fully-adaptive control strategy is much more
``cost-effective" because larger $\beta(t)$ will likely cause a higher cost.
\subsubsection{Confirmation of Theorem \ref{theorem-control-to-specified-state}}
Theorem \ref{theorem-control-to-specified-state} states that even if we do not know $\gamma(t)$ but we
may be able to observe $i_v(t)$ and may be able to adjust the defense (but cannot kill the virus spreading),
then following its control strategy will cause the containment of the virus spreading.
In our simulation study, we used the periodical function $\gamma(t) \in \{0.0005,0.001\}$
with period $T=8$ similar to what was shown in Figure \ref{fig:beta-gamma-t-model}. This input parameter is not used
in our adaptive control algorithm, rather it is merely for the purpose of comparison
to the sufficient condition in Theorem \ref{thm4}, which requires the $\beta_v(t)$
satisfy certain property (for example, we use $\beta(t) \in \{0.01, 0.02\}$, meaning that the virus spreading does not
die out as predicated),
and the $\beta_v(t)$ derived from our adaptive control strategy.
To ensure the comparability, we also let the virus initially infect 2,291 or 20\% randomly selected nodes.
As mentioned before, since the anti-synchronous defense is more effective than the synchronous and asynchronous defenses,
we will compare it with the outcome of the adaptive control algorithm.
\begin{figure*}[ht]
\centering
\subfigure[Fully-adaptive control]
{\label{fig:fully-adaptive control-th5}
\includegraphics[width=.48\textwidth]{oregon_Model_th9_rl4.eps}}
\subfigure[Comparison of cure capabilities]
{\label{fig:cure capabilities-th5}
\includegraphics[width=.48\textwidth]{oregon_th9_beta.eps}}
\caption{Confirmation of the virus spreading containment via fully-adaptive control (the red-color curves are
hidden behind the green-color curves)}
\label{fig:confirming-adaptive-control-2}
\end{figure*}
In our simulation we set $i^*=0.1$ (i.e., we want to contain the degree of infection to 10\%)
and $u_v=0$ (i.e., a special case of Theorem \ref{theorem-control-to-specified-state}).
Figure \ref{fig:fully-adaptive control-th5} plots the dynamics of the number of infected nodes
with respect to time in the following three cases: the adaptive control parameter $\rho=0.001$;
the adaptive control parameter $\rho=0.001$ with fixed $\bar{\gamma}(t) = (0.0005+0.001)/2=0.00075$
as specified in Theorem \ref{theorem-control-to-specified-state});
the comparison dynamics corresponding to the anti-synchronous case
with $T=8$ periodical function $\beta(t) \in \{0.01, 0.02\}$.
We draw the following observations.
First, the control strategy does contain the infection to the pre-determined level of
$i^*=0.1$ or 10\% infection. Moreover,
the adaptive control strategy is robust because perturbation in $\gamma$ does not
fundamentally change the dynamics behavior.
This also confirms the physical meanings of Proposition \ref{prop:adaptive-control-2} discussed above.
Second, Figure \ref{fig:fully-adaptive control-th5} indicates
that the adaptive defenses are slightly less effective in defending against the virus spreading
than the defense of $T=8$ periodical function $\beta(t) \in \{0.01, 0.02\}$.
As we show in Figure \ref{fig:cure capabilities-th5}, the adaptive
control strategy can be much more ``cost-effective" because it leads to significantly smaller $\beta(t)$.
\section{Conclusion}
\label{sec:conclusion}
We have presented a novel dynamical systems model for studying both semi-adaptive and fully-adaptive
defenses against virus spreading. For semi-adaptive defense, we give general as well as succinct
sufficient conditions under which the virus spreading will die out. For fully-adaptive defense, we characterize two
adaptive control strategies under which the virus spreading will die out or will be contained
to a desired level of infection.
Our analytical results are confirmed with simulation study.
This paper brings a range of open questions for future research. In addition to those
mentioned in the body of the paper, here are more examples:
What are the necessary conditions under which the virus spreading will die out?
What are the optimal adaptive control strategies?
\smallskip
\noindent{\bf Acknowledgement}. We thank the anonymous reviewers for their useful comments,
and Raj Boppana for helpful discussion on the simulation.
This work was supported in part by AFOSR, AFOSR MURI, ONR, and UTSA.
The views and conclusions contained in the article are those of the authors and should
not be interpreted as, in any sense, the official policies or endorsements of the government or
the agencies.
|
\section{Introduction\label{sec:intro}}
A composition sum operator $T$ is a linear operator acting on a vector
space $F$ of functions such that, for each function $f\in F$, the
image vector $Tf$, evaluated at $x$, is given by
\begin{equation}
Tf(x)=\sum_{i=1}^{\ell}a_{i}(x)f(\alpha_{i}(x))\,,\label{eq:CSOs}
\end{equation}
where $\ell\ge1$, $\alpha_{1},\alpha_{2},\dots,\alpha_{\ell}$ are
affine contractions, and $a_{1}$, $a_{2}$, \dots, $a_{\ell}$ are
a fixed sequence of coefficients associated with $\alpha_{1},\alpha_{2},\dots,\alpha_{\ell}$.
A full definition is given in Section~\ref{sec:CSOs}. Although written
here as functions $a_{i}(x)$, in many applications the $a_{i}$ are
constants.
Composition sum operators have been studied extensively by Kuczma
and his co-workers. In particular, in the seminal monograph~\cite{Kuczma-et-al1990},
CSOs are discussed in detail in Chapter~6 ``Higher order equations
and linear systems'', principally in the real domain and in the ``cyclic
equation'' case, in which the $\alpha_{i}$ are iterates of a single
function. Kuczma \emph{et al} give some important existence and uniqueness
theorems in this context and we refer the reader to~\cite{Kuczma-et-al1990}
and to the references contained therein.
CSOs arise in several contexts, including the application of renormalisation
techniques to quasi-periodic non-linear dynamical systems and a toy-model
of magnetic flux growth in kinematic dynamo theory~\cite{Gilbert2002,Gilbert2005}.
Quasi-periodic systems are an important class of non-linear dynamical
systems which find application in many areas of the physical sciences.
In the simplest case, the dynamics are governed by an irrational number
$\omega\not\in\bbbq$, often called the rotation number or winding
number in the literature. It can often be identified as the ratio
of two incommensurate frequencies in the underlying system. Studies
of quasi-periodic systems often focus on the time-correlations between
system variables. These correlations, and, indeed other properties
of quasi-periodic systems, typically depend on the number-theoretic
properties of $\omega$, and, in particular, on the continued-fraction
expansion of $\omega$ and the associated rational convergents $p_{n}/q_{n}\to\omega$.
Examples of quasi-periodic systems include strange non-chaotic attractors,
the Harper equation and its generalisations, and other quantum mechanical
models depending on an underlying irrational rotation of the circle.
The correlation structure of quasi-periodic systems may be understood
by renormalisation analysis, leading to dynamical functional equations
which relate correlations at time $t$ to those at time $t+q_{n}$,
the dynamical properties of which depend on the dynamical behaviour
of the Gauss map applied to $\omega$ or, equivalently, on the action
of the shift map on the entries in the continued-fraction expansion
of $\omega$. In such studies the case of the golden-mean rotation
number, for which $\omega=(\sqrt{5}-1)/2$, often plays a pivotal
role. This is perhaps not surprising given the simplicity of its continued
fraction $[1,1,1,\dots]$, with all entries equal to $1$. For the
golden-mean, renormalisation analysis frequently leads to fixed-point
functional equations.
For example, renormalisation of correlations for the golden-mean Harper
equation leads to the so-called strong-coupling fixed point, satisfying
the functional equation
\begin{equation}
f(z)=f(-\omega x)\, f(\omega^{2}x+\omega)\label{eq:SCFP}
\end{equation}
where $f$ is an analytic function with a pole of order $2$ at $z=1$
and $\omega=(\sqrt{5}-1)/2$. The construction of the strong-coupling
fixed point involves first studying fixed points of the composition
sum operator $M$ given by
\begin{equation}
Mf(z)=f(-\omega z)+f(\omega^{2}z+\omega)\label{eq:M}
\end{equation}
where $f$ is analytic with a logarithmic singularity at $1$. (Here,
of course, $\ell=2$ and $a_{1}=a_{2}=1$ in equation~\eqref{eq:CSOs}.)
Note that equation~\eqref{eq:SCFP} is the fixed point case $f_{n}=f_{n-1}=f_{n-2}=f$
of the second-order multiplicative functional recurrence
\begin{equation}
f_{n}(z)=f_{n-1}(-\omega x)\, f_{n-2}(\omega^{2}x+\omega)\,.\label{eq:multiplicative recurrence}
\end{equation}
Similarly, the associated linear recurrence
\begin{equation}
f_{n}(z)=f_{n-1}(-\omega x)+f_{n-2}(\omega^{2}x+\omega)\label{eq:additive recurrence}
\end{equation}
leads in turn to the operator $M$. The functional recurrences~\eqref{eq:multiplicative recurrence}
and~\eqref{eq:additive recurrence} arise in several contexts involving
the golden-mean rotation number, in particular in connection with
the Ketoja-Satija orchid flower for the generalised Harper equation~\cite{Ketoja1995,Mestel2004a},
and, with piecewise constant functions $f_{n}$, in the analysis of
quantum two-level systems~\cite{Feudel-Pikovsky-Zaks1995}, barrier
billiards~\cite{Chapman2003} and strange non-chaotic attractors~\cite{Feudel-Pikovsky-Politi1996,Mestel2002}.
We refer to the comprehensive book by Feudel \emph{et al}~\cite{Feudel-Kuznetsov-Pikovsky2006}
for a full discussion of the applications of renormalisation theory
in strange non-chaotic attractors and to~\cite{Osbaldestin-Mestel2003}
for an overview of applications of equations~\eqref{eq:multiplicative recurrence}
and~\eqref{eq:additive recurrence}.
For other quadratic irrationals with constant continued-fraction expansion,
say $\omega=[a,a,a,a,\dots]$ where the integer $a\ge1$, we obtain
the multiplicative and additive fixed-point equations
\begin{equation}
f(x)=\left(\prod_{i=0}^{a-1}f(-\omega x-i)\right)\; f(\omega^{2}x+a\omega)\,,\quad f(x)=\left(\sum_{i=0}^{a-1}f(-\omega x-i)\right)+f(\omega^{2}x+a\omega)\label{eq:multip_and_add_general_a}
\end{equation}
again with associated functional recurrences. See~\cite{Mestel2004b,Dalton-Mestel2003}
for applications in this case.
We now return to the golden-mean case. In~\cite{Mestel2000}, a rigorous
analysis established \emph{inter alia} the existence and properties
of a unique solution of~\eqref{eq:SCFP} under the constraints of
that physical situation, and provided an explicit expansion for the
strong-coupling fixed point. Indeed, writing $\phi_{1}(z)=-\omega z$
and $\phi_{2}(z)=\omega^{2}z+\omega$, this fixed point of $M$ has
the form:
\begin{eqnarray*}
f(z) & = & \lambda\left[\log\left(\frac{1-z}{1-\omega}\right)+\sum_{k=1}^{\infty}\;\sum_{\substack{i_{1},...,i_{k}\\
i_{1}=1
}
}\log\frac{1-\phi_{i_{1}}\circ...\circ\phi_{i_{k}}(z)}{1-\phi_{i_{1}}\circ...\circ\phi_{i_{k}}(\omega)}\right]
\end{eqnarray*}
where $i_{j}\in\{1,2\}$, $\lambda\in\mathbb{C},$ $\log$ is the
principal branch of the logarithm, and $\circ$ signifies functional
composition.
The proof in~\cite{Mestel2000} is non-trivial and also depends on
properties of the golden mean so that its generalisation is not obvious.
The theory presented in this paper provides a general framework for
the construction of fixed-points of composition sum operators, which
not only illuminates the results of~\cite{Mestel2000}, but enables
the construction of fixed-points of other renormalisation operators
in a simplified and unified manner.
The organisation of this paper is as follows. We first describe in
Section~2 a general theory for the construction of fixed points of
linear operators on vector spaces. In Section~3 we introduce formally
the Composition Sum Operators (CSOs), describe their properties, and
define the function spaces of analytic functions on which we shall
work. We introduce the idea of a seed function, which we will use
extensively in the construction of fixed points of these operators.
In Section~4, we apply the theory in Section~2 to construct fixed
points of CSOs in the constant coefficient affine case (which we simply
call affine CSOs). Finally, in Section~5, we show how the theory
can be applied to construct fixed-points arising in the renormalisation
theory of quasi-periodic systems. This final section uses a construction
method derived from the methods in~\cite{Mestel2000}. We conclude
with some directions for future research.
\section{Seeded fixed point theory on vector spaces \label{sec:seeded-fixed-point-thy}}
In this section we describe the formal abstract setting for our construction
of fixed-points of linear operators on vector spaces. Our goal is
to derive from a given linear operator $T$, a \emph{fixed point operator}
$\widehat{T}$ which maps its domain (called the \emph{seed space}
of $T$) to the \emph{fixed point space} of $T$ (denoted $FP(T)$).
Although we have in mind applications to composition operators on
spaces of complex analytic functions with singularities, the theory
is quite general and may used in cases in which linear operators act
on a vector space that may be decomposed into a direct sum of subspaces,
one with a well defined Banach space structure and no non-zero fixed
points, and the other consisting of unbounded or singular vectors,
but which are prototype fixed points or ``seeds''. Although, set
in this general context, the theory is straightforward, its power
lies in its application to construct fixed points of renormalisation
operators and other operators, in which the fundamental structure
of the fixed points are evident, but the precise detail is not.
Let $F$ be a vector space and let $G\subset F$ be a proper non-zero
subspace of $F$. In many cases $G$ is equipped with norm $\|\cdot\|$
which endows $G$ with a Banach space structure, but this is not necessary
for the general theory. Let $T:F\to F$ be a linear operator and let
$\overline{T}$ denote the operator $I-T$ on $F$. We note that $f\in F$
is a fixed point of $T$ if, and only if, it is in the kernel of $\overline{T}$.
We assume that $T$ satisfies the following two properties.
\begin{enumerate}
\item P1 $\overline{T}(F)\subset G$, so that $\overline{T}$ maps the whole
of $F$ into the subspace $G$.
\item P2 The restricted operator $\overline{T}_{|G}$ is invertible on $G$
so that $T^{+}=\overline{T}_{|G}^{-1}$ exists and maps $G$ to $G$.
\end{enumerate}
We note three points. First, when $G$ has a Banach space structure
with norm $\|\cdot\|$, then the second condition is satisfied when
$T_{|G}$ is a contraction with $\|T\|<1$ (but this is not a necessary
condition). Second, although these two conditions are very general,
one can often think of $G$ as the well-behaved non-singular part
of $F$ and $F\backslash G$ as being the singular or unbounded part
of $F$, which provides seeds for the construction of non-zero fixed
points of $T$. Third, writing $F$ explicitly as the direct sum $F=G\oplus S$,
we will see below that the vector space $S$ is a subspace of the
seed space which is mapped one to one to the fixed points of $T$
by the fixed point operator.
The following is the principal result for the construction of fixed
points of the operator $T$.
\begin{theorem}\label{thm:fixed_point_thm} Let $F$ be a vector space
and $T:F\to F$ be a linear operator satisfying the conditions P1
and P2 above. Then the linear operator $\widehat{T}:F\to F$ given
by
\[
\widehat{T}=I-T^{+}\overline{T}
\]
maps $F$ to the subspace $FP(T)$ of fixed points of $T$ and $\widehat{T}$
induces a vector space isomorphism from the factor space $F/G$ to
$FP(T)$. In the case when $F=G\oplus S$, $\widehat{T}$ induces
an isomorphism from $S$ to $FP(T)$. \end{theorem}
The proof of this theorem is quite straightforward and belies the
utility and power of the theorem itself.
\begin{proof}
Let $f\in F$ and consider $\widehat{T}f$. Then $\overline{T}(\widehat{T}f))$
$=$ $\overline{T}f-\overline{T}(T^{+}\overline{T})f$ = $\overline{T}f-\overline{T}f=0$
so that $\widehat{T}f$ is a fixed point of $T$. Conversely, let
$f\in F$ be a fixed point of $T$. Then $\overline{T}f$ $=$ $0$
so that $\widehat{T}f=f$, and so $f\in\widehat{T}F$. It is straightforward
to show that $FP(T)$ is a linear subspace of $F$.
Now let us abuse notation slightly and also denote by $\widehat{T}$
the map $\widehat{T}:F/G\to FP(T)$ given by $\widehat{T}[f]=\widehat{T}f$,
for $[f]$ an element of $F/G$. It is straightforward to show that
$\widehat{T}$ is a linear map and we note that this map is well defined
because $G$ is in the kernel of $\widehat{T}$. It is immediate that
$\textrm{Im}~\widehat{T}$ $=FP(T)$ . Now let $[f]\in F/G$ be in
$\textrm{ker}~\widehat{T}$. Then $\widehat{T}f=0$, so that $f=T^{+}\overline{T}f$
whence $f\in G$ and so $[f]=[0]$. It follows that $\widehat{T}$
is a vector space isomorphism.
Finally, in the case $F=G\oplus S$, $S$ is isomorphic to $F/G$
via the natural inclusion, and so $\widehat{T}$ induces a vector
space isomorphism $S$ to $FP(T)$. \end{proof}
Note that $\widehat{T}$ is now our fixed point operator derived from
$T$, and its seed space (domain) is $\overline{T}^{-1}G$, ie $f$
is a seed if, and only if, $f\in\overline{T}^{-1}G$. This also means
$\widehat{T}f=f+g$ for $g=-T^{+}\overline{T}f\in G$.
There is also a straightforward but important extension which allows
us to extend the operator $\widehat{T}$.
\begin{corollary}\label{cor:FixedPointThmExtension} Suppose that condition
P2 holds, but not P1, so that we do not necessarily have $T(F)\subseteq G$.
Suppose, instead, that for some $f\in F$, $\overline{T}f\not\in G$,
but $\overline{T}(T^{k}f)\in G$ for some integer $k\ge1$. Writing
$f_{k}=\sum_{i=0}^{k-1}\overline{T}(T^{i}f)$, then $\overline{T}(f-f_{k})\in G$
and $\widehat{T}(f-f_{k})$ is a fixed point of $T$. Moreover, the
fixed point is independent of the choice of $k$. \end{corollary} \begin{proof}Since
$f_{k}=\sum_{0}^{k-1}\overline{T}(T^{i}f)=f-T^{k}f$, it is immediate
that $\overline{T}(f-f_{k})=\overline{T}(T^{k}f)\in G$, and so $f_{k}-f$
is a seed and we may now apply Theorem~\ref{thm:fixed_point_thm}.
The final statement follows from the observation that, if $\tilde{k}\ge k$,
then $\overline{T}T^{\tilde{k}}f$ $=$ $T^{\tilde{k}-k}\overline{T}T^{k}f\in G$,
since $T(G)\subseteq G$. It follows that $\widehat{T}(f-f_{\tilde{k}})-\widehat{T}(f-f_{k})$
$\widehat{T}(f_{k}-f_{\tilde{k}})$ $=0$, since $f_{k}-f_{\tilde{k}}\in G$.
\end{proof} We call a vector satisfying the hypotheses of Corollary\ref{cor:FixedPointThmExtension}
a \emph{generalised seed}.
As a simple application of Theorem~\ref{thm:fixed_point_thm}, we
consider the operator $T$ on real functions $c(x)$ defined on $[-1,1]$
\[
Tc(x)=c\left(\frac{x-1}{2}\right)-c\left(\frac{1-x}{2}\right)\,.
\]
This operator arises from the zero-shear base case of a the Stretch-Fold-Shear
toy model in kinematic dynamo theory, studied in detail by Gilbert~\cite{Gilbert2002,Gilbert2005}.
Now, for integer $n\ge1$, let $P_{2n-1}^{o}$ denote the real vector
space of odd polynomials of degree at most $2n-1$. Then, evidently,
$TP_{2n-1}^{o}\subseteq P_{2n-1}^{o}$, and, indeed, $T$ has an upper-triangular
matrix with respect to the standard basis $\{x,x^{3},\dots,x^{2n-1}\}$,
from which the spectrum of $T$ restricted to $P_{2n-1}^{o}$ is readily
obtained. Let us now consider the operator $T$ from the viewpoint
of Theorem~\ref{thm:fixed_point_thm}. Writing $P_{2n-1}^{o}=P_{2n-3}^{o}\oplus<x^{2n-1}>$
and $T_{2}=4^{n-1}T$, then it is straightforward to verify that the
hypotheses of Theorem~\ref{thm:fixed_point_thm} are satisfied with
$G=P_{2n-3}^{o}$ and $S=<x^{2n-1}>$, from which the spectrum and
eigenfunctions of $T$ on $P_{2n-1}^{o}$ may be calculated. In fact,
this is also the spectrum of $T$ acting on a more general space of
analytic functions on which $T$ is compact. See~\cite{Gilbert2002,Gilbert2005}
for details.
\section{Composition sum operators, analytic function spaces, and seed functions
\label{sec:CSOs}}
We can now apply the general theory of the previous section to help
us identify fixed points of the particular class of operators we call
\emph{Composition Sum Operators}. The operators $M$ and $T$ introduced
above are examples of this class.
Let $\alpha:D\longrightarrow D$ be a map of a complex domain into
itself, and let $F$ be a ring of complex-valued functions defined
on $D$. In practical applications, $D$ is frequently a disc and
$F$ a space of analytic functions on $D$, possibly with singularities.
For any $f\in F$ we denote by $\alpha^{*}f$ the function $f\alpha$,
the composition of $f$ with $\alpha$. Then $\alpha^{*}$ is an operator
on $F$, which we call a \emph{Composition Operator} on $F$.
A \emph{Composition Sum Operator (CSO)} on $F$ is an operator $T=\sum_{i=1}^{\ell}a_{i}\alpha_{i}^{*}$
on $F$ where $a_{i}\in F$, $a_{i}\not=0$, $\alpha_{i}^{*}$ is
a Composition Operator on $F$, and $\left(\sum a_{i}\alpha_{i}^{*}\right)f=\sum a_{i}.f\alpha_{i}$.
We call the positive integer $\ell$ the \emph{length} of the CSO,
and we assume that $\alpha_{i}\not=\alpha_{j}$, for $i\not=j$. When
it is clear from the context, we suppress the explicit range $i=1$,
$\dots$, $\ell$.
In the case when $F$ is Banach space with norm $\left\Vert \cdot\right\Vert $,
we have $\left\Vert Tf\right\Vert =\left\Vert \sum a_{i}.f\alpha_{i}\right\Vert \le\sum\left\Vert a_{i}\right\Vert \left\Vert f\right\Vert $
so $\left\Vert T\right\Vert \le\sum\left\Vert a_{i}\right\Vert <\infty$
so a CSO is also a bounded linear operator on these spaces. However,
the most interesting cases occur when $T$ is an operator on a space
of functions with singularities.
Although CSOs may be defined for general maps $\alpha_{i}$, in most
applications they are affine contractions and the coefficients $a_{i}$
are constant functions. We say that the CSO $T=\sum_{i=1}^{\ell}a_{i}\alpha_{i}^{*}$
is \emph{affine} if each $a_{i}$ is a constant, and each $\alpha_{i}$
is an affine contraction, ie., $\alpha_{i}(z)=s_{i}(z-z_{i})+z_{i}$,
where $z_{i}\in D$ and $|s_{i}|<1$, $s_{i}\in\bbbc$.
Affine CSOs are the principal application of the seeded fixed point
theory described above. In the next section we develop the theory
of seed functions to construct fixed points of CSOs with singularities
of pole and logarithmic type. We concentrate on these as they currently
seem the most significant; however the techniques presented are readily
extended to other types of singularity such as removable and algebraic
singularities.
\subsection{Seed functions over spaces of bounded functions}
In this section we look at the conditions under which a complex function
$f$ with various types of singularities can be a seed for a CSO $T$
over a Banach space of analytic functions $G$. In particular this
requires that $\overline{T}f\in G$. This is a strong condition as
we shall see. In this context we refer to seeds as \emph{seed functions}.
We also recall the important result that that $\widehat{T}f=f+g$
for some $g\in G$.
We will show that, apart from a small set of CSOs which admit polynomials
as fixed points, there are no analytic non-zero fixed points of an
affine CSO in $G$. All other non-zero fixed points therefore have
singularities of some sort. For some CSOs with real coefficients the
singularities can be discontinuities (see~\cite{Mestel2002} for
an example), but in the context of CSOs defined on spaces of analytic
functions (with singularities) it is logarithmic and unbounded isolated
singularities which are of greatest interest, and we will analyse
these below.
To be precise, by unbounded isolated singularities we mean poles and
essential singularities, ie not the singularities of multivalued functions
such as $\log z$, or $z^{\alpha}$ for non-integral $\alpha$. And
by logarithmic singularities, we mean the unbounded singularities
of functions which can be written $\log g$ with $g$ analytic, ie
not composed log functions such as $\log\log z$ or $(\log z)^{\alpha}$.
We will refer to functions analytic apart from these singularities
as\label{PESL} (Pole/Essential/Simple Log) functions.
Consider a fixed affine CSO acting on functions on a domain $D\subseteq\bbbc$
and $G$ a Banach space of analytic functions on $D$. We consider
a complex valued function defined and analytic on an open dense subset
$D'$ of $D$. The set of points on which $f$ fails to be analytic
might include, for example, points on branch cuts other than branch
points. We therefore restrict our discussion to unbounded singularities
ie points at which $f$ fails to be bounded on any neighbourhood.
We say that $z_{0}\in D-D'$ is an \emph{unbounded singularity} if
$z_{0}$ has no neighbourhood $U\subseteq D$ on which $f_{|U\bigcap D'}$
is bounded, ie $\sup_{z\in U\bigcap D'}|f(z)|=\infty$ for any neighbourhood
$U$ of $z_{0}$. in $D$. The \emph{unbounded singularity set} of
$f$ in $D$, written $\unb_{D}(f)$ is the (possibly empty) set of
points in $D$ at which $f$ has an unbounded singularity.
Recall that $f$ is a seed function if $\overline{T}f\in G$. It follows
that, if $f$ is a seed function, then $\unb_{D}(f)$ $=$ $\unb_{D}(Tf)$,
because each function in $G$ has no unbounded singularities in $D$.
In principle, $\unb_{D}(f)$ may be quite large and composition operators
may act and interact on the set in intricate ways which are beyond
the scope of this paper. We will restrict our attention to \emph{simple\label{simple}}
actions which we define as follows: we will say that $\unb_{D}(f)$
is \emph{simple} under $T$ if (i) each $\alpha_{i}$ acts unstably
on $\unb_{D}(f)$, (ie for $z\in unb_{D}(f)$, $\alpha_{i}(z)\not\in unb_{D}(f)$
unless $\alpha_{i}(z)=z$) and (ii) if $\alpha_{i}(z)=z$ then $\alpha_{j}(z)\ne z$
for $j\ne i$. We also say $f$ itself is simple under $T$ if $\unb_{D}(f)$
is simple under $T$.
From these results the following lemma follows readily. \begin{lemma}
\label{lem:UnboundedPoints}Let $T=\sum_{i}a_{i}\alpha_{i}^{*}$ be
an affine CSO, and let $f$ be a seed over a Banach space $G$ of
analytic functions. If the unbounded set of $f$ is simple under $T$,
then $\unb_{D}(f)\subseteq\bigcup_{i}FP(\alpha_{i})$, and each unbounded
point of $f$ is a fixed point of precisely one $\alpha_{i}$. \end{lemma}
\begin{proof} If $z_{0}\in\unb(f)$ then $z_{0}\in\unb(Tf)$ so $\sum_{i}a_{i}.f\alpha_{i}$
is unbounded at $z_{0}$, so, for at least one $i$, $f\alpha_{i}$
is unbounded at $z_{0}$. Since the unbounded set of $f$ is simple,
this means $\alpha_{i}(z_{0})=z_{0}$ and the $i$ is unique. \end{proof}
As an example, consider the CSO $M$ above. Its fixed points are $\{0,1\}$.
So any simple seed or fixed point of $M$ is unbounded on at most
$\{0,1\}$.
\section{Fixed-point construction for affine CSOs}
Recall that an affine CSO is a CSO for which the functions $a_{1}$,
$a_{2}$, $\dots$, $a_{\ell}$ are non-zero constants and the maps
$\alpha_{1}$, $\alpha_{2}$, $\dots$, $\alpha_{\ell}$ are affine
contractions $\alpha_{i}(z)=z_{i}+s_{i}(z-z_{i})$ on the complex
plane, where the fixed points $z_{i}$ and and contraction rates $s_{i}$
are all complex constants, with $0\le|s_{i}|<1$. (Here, and in what
follows, $i$ ranges from $1,\dots,\ell$.) In many applications the
$z_{i}$ and the $s_{i}$ are real, but the theory may be just as
easily developed for complex $z_{i}$ and $s_{i}$. With affine CSOs
we are able to obtain a good theory for the construction of fixed
points, drawing on the work of the previous sections.
We define here an additional constraint which which will prove useful
in this section. Given a CSO $T=\sum a_{i}\alpha_{i}^{*}$ on a domain
$D$, we will say $\alpha_{i}$ is \emph{fixed point independent on
$D$} if $z_{i}\not\in\bigcup_{j\ne i}\overline{\alpha_{j}(D)}$.
This means that if $z_{i}$ is a singularity of a function $f$, $f\alpha_{j}$
has a singularity at $z_{i}$ if, and only if, $j=i$.
We shall work in a fixed disc in the complex plane. Let $D=D_{r}$,
the open disc of radius $r$ about $0$ in $\bbbc$, and let $G(D_{r})$
be the complex Banach space of functions $g$ analytic on $D_{r}$
with finite supremum norm $||g||_{\infty,r}=\sup\{|g(z)|\;:\; z\in D_{r}\}$.
Let $R>0$ be chosen so that for some $\delta>0$, $\overline{\alpha_{i}(D_{R+\delta})}\subseteq D_{R-\delta}$
for all $i=1,\dots,\ell$. Because the $\alpha_{i}$ are contractions,
this condition holds provided we take $R$ sufficiently large. We
write $D=D_{R}$, and $G=G(D_{R})$ and $||g||_{\infty}$ $=$ $||g||_{\infty,R}$,
for $g\in G(D_{R})$.
Let us consider the affine Composition Sum Operator $Tf(z)=\sum_{i=1}^{\ell}a_{i}f(\alpha_{i}(z))$.
It is straightforward to verify that $T$ is a linear operator on
the complex Banach Space $G$. Moreover, for $m\ge0$, we may differentiate
$m$ times the function $Tg$:
\begin{equation}
(Tg)^{(m)}(z)=\sum_{i=1}^{\ell}a_{i}s_{i}^{m}g^{(m)}(\alpha_{i}(z))\,.
\end{equation}
We now define an induced operator on $G$, which we denote by $T^{(m)}$,
given by
\begin{equation}
T^{(m)}\tilde{g}(z)=\sum_{i=1}^{\ell}a_{i}s_{i}^{m}\tilde{g}(\alpha_{i}(z))\,,
\end{equation}
for $\tilde{g}\in G$. We have that
\begin{equation}
||T^{(m)}\tilde{g}||_{\infty}\le\sum_{i=1}^{\ell}|a_{i}||s_{i}|^{m}||\tilde{g}||_{\infty}\,,
\end{equation}
so that the operator norm $||T^{(m)}||\le\sum_{i=1}^{\ell}|a_{i}||s_{i}|^{m}$.
An immediate consequence is that there exists $m\ge0$ such that $||T^{(m)}||<1$,
a contraction. Finally, using Cauchy estimates, we see that if $g\in G$,
then $g^{(m)}\alpha_{i}$ is also in $G$ and, moreover, $||g^{(m)}\alpha_{i}||_{\infty}\le K||g||_{\infty}$,
where $K=m!$ $\delta^{-(m+1)}$, for $i=1,\dots,\ell$.
From these results, we may readily show that all non-trivial fixed
points of $T$ in $G$ are polynomials. The proof is rather elegant.
Indeed, suppose $g\in G$ is a fixed point of $T$, with $g\not=0$.
Then, for some $m\ge0$, $0$ $\le$ $||g^{(m)}||_{\infty}$ $=$
$||T^{(m)}g^{(m)}||_{\infty}$ $\le$ $||T^{(m)}||\,||g^{(m)}||_{\infty}$
$<$ $||g^{(m)}||_{\infty}$, a contradiction. It follows that $g$
is zero or a polynomial of degree at most $m-1$.
Whether or not $T$ has a polynomial fixed point depends on the precise
values of the $a_{i}$ and $\alpha_{i}$. Indeed, for a polynomial
$p(x)=p_{0}+p_{1}x+\dots+p_{m}x^{m}$, $p_{m}\not=0$, it is clear
that $Tp$ is a polynomial of degree at most $m$. Inspecting the
coefficient of $x^{m}$ in $Tp(x)=p(x)$, we have
\begin{equation}
a_{1}s_{1}^{m}+a_{2}s_{2}^{m}+\dots+a_{\ell}s_{\ell}^{m}=1\label{eq:degree-m-relation}
\end{equation}
which is clearly a necessary condition for a polynomial fixed point
of degree $m$. Conversely, suppose that~\eqref{eq:degree-m-relation}
holds. Then if $p(x)$ is of degree $m$, $\overline{T}p$ is of degree
at most $m-1$, and so $\overline{T}$ is degenerate and has non-trivial
kernel. If $q$ is in the kernel, then $Tq=q$. Hence $T$ has a non-trivial
space of polynomial fixed points if, and only if, (\ref{eq:degree-m-relation})
holds for one or more $m\ge0$. Note that there is some $N>0$ such
that the condition does not hold for any $m\ge N$, and so the space
of polynomial fixed points of $T$ is of bounded maximum degree.
We now assume that there are no polynomial fixed points, ie that
\begin{equation}
a_{1}s_{1}^{j}+a_{2}s_{2}^{j}+\dots+a_{\ell}s_{\ell}^{j}\not=1\,,\quad\textrm{for all \ensuremath{j\ge0}.}\label{eq:no-poly-soln}
\end{equation}
It is now evident that the only non-zero fixed points are necessarily
singular on $D$. In what follows we restrict ourselves to unbounded
isolated and logarithmic singularities.
Let us consider first simple seeds $f$ with unbounded isolated singularities.
Since $f$ is simple, every point of $\unb_{D}{f}$ is a fixed point
of a unique $\alpha_{i}$. Without loss of generality, we let $i=1$
and we suppose that $f$ has an isolated singularity at $z_{1}$,
so that $f_{|D-\{z_{1}\}}$ is analytic on some neighbourhood of $z_{1}$.
Let $C_{\epsilon}$ be a circle of radius $\epsilon>0$ about $z_{1}$.
Then if $\alpha_{i}$ is fixed point independent of the other $\alpha_{j}$
on $D$, and for $\epsilon$ sufficiently small, $f$ is analytic
inside and on $\alpha_{i}(C_{\epsilon})$ for $i=2,\dots,\ell$. Using
the fact that $f-Tf$ is analytic and integrating along $C_{\epsilon}$,
we have, for integer $k\ge0$,
\begin{align}
0 & =\int_{C_{\epsilon}}(z-z_{1})^{k}(f(z)-Tf(z))dz=\int_{C_{\epsilon}}(z-z_{1})^{k}(f(z)-a_{1}f(\alpha_{1}(z)))dz\\
& =\left(\int_{C_{\epsilon}}(z-z_{1})^{k}f(z)dz-\int_{C_{\epsilon}}(z-z_{1})^{k}a_{1}f(\alpha_{1}(z))dz\right)\\
& =\left(\int_{C_{\epsilon}}(z-z_{1})^{k}f(z)dz-\int_{\alpha_{1}^{-1}C_{\epsilon}}s_{1}^{-(k+1)}(w-z_{1})^{k}a_{1}f(w)dw\right)\\
& =2\pi if_{-(k+1)}\left(1-s_{1}^{-(k+1)}a_{1}\right)\,.
\end{align}
In this calculation we have used Cauchy's integral theorem, together
with a change of variable $w=\alpha_{1}(z)$. We have denoted by $f_{-(k+1)}$
the $(k+1)-$th coefficient in the Laurent expansion of $f$ about
$z_{1}$.
We conclude for $k\ge0$ that either $f_{-(k+1)}=0$ or $a_{1}=s_{1}^{(k+1)}$.
If the latter condition holds, then we may have $f_{-(k+1)}\not=0$
and $f(z)=(z-z_{1})^{-(k+1)}$ is a seed function, from which a fixed
point of $T$ may be constructed, provided equation~\eqref{eq:no-poly-soln}
holds. The construction is omitted here as it is similar to that given
below for the logarithmic case.
The result also shows that essential singularities do not lead to
fixed points of affine CSOs. For $\left(1-s_{1}^{-(k+1)}a_{1}\right)=0$
cannot hold for more than one $k\ge0$, ruling out a non-finite principal
part of $f$ at $z_{1}$.
We now give the construction of a fixed point of $T$ in the case
when the seed function $f$ has a single logarithmic singularity of
the form $f(z)=\log(z-z_{i})+g(z)$, where $g\in G$$ $ and $z_{i}\in D$,
and where $i$ is one of $1$, $2$, $\dots$, $\ell$. Again, without
loss of generality, we take $i=1$.
Our first observation is that we may take $f(z)=\log(z-z_{1})$, since
if $\tilde{f}(z)=\log(z-z_{1})+g(z)$, where $g\in G$, then $\widehat{T}f=\widehat{T}\tilde{f}$.
(Any convenient branch of the logarithm function may be taken, although,
to be specific, we choose the principal branch.) Again we let $\alpha_{1}$
be fixed point independent on $D$ so that, for $j\not=1$, $z_{1}\not\in\overline{\alpha_{j}(D)}$.
Therefore $f(z)-Tf(z)=\log(z-z_{1})-a_{1}\log(\alpha_{1}(z)-z_{1})+g(z)$,
where $g\in G$, and, since $\log(z-z_{1})-a_{1}\log(\alpha_{1}(z)-z_{1})$
$=\log(z-z_{1})-a_{1}\log(z_{1}+s_{1}(z-z_{1})-z_{1})$ $=(1-a_{1})\log(z-z_{1})-a_{1}\log s_{1}$,
it follows that $f-Tf\in G$ if, and only if, $a_{1}=1$.
Let us now assume that $a_{1}=1$ and $f(z)=\log(z-z_{1})$. For convenience
we consider separately the cases when $T$ is a contraction on $G$
and when $T$ is not a contraction on $G$.
The first case is easily handled directly by appealing to Theorem~\ref{thm:fixed_point_thm}.
Let $F=\langle\log(z-z_{1})\rangle\oplus G$. Then $T:F\to F$ satisfies
the hypothesis of Theorem~\ref{thm:fixed_point_thm}, from which
we conclude immediately the construction of a one-dimensional subspace
of fixed-points of $F$ in $F$ with logarithmic singularity $\log(z-z_{1})$.
The second case may be handled by differentiating the operator $T$,
say $m$ times, until it is a contraction, appealing to Theorem~\ref{thm:fixed_point_thm}
for the induced operator $T^{(m)}$, and then integrating up to obtain
a fixed point of $T$. Specifically, let $m\ge1$ be such that $\sum_{i=1}^{\ell}|a_{i}||s_{i}|^{m}<K<1$
and let $f_{m}(z)=(m-1)!(-1)^{m-1}(z-z_{1})^{-m}$. Then
\begin{equation}
T^{(m)}f_{m}(z)-f_{m}(z)=\sum_{i=1}^{\ell}a_{i}s_{i}^{m}f_{m}(\alpha_{i}(z))-f_{m}(z)=\sum_{i=2}^{\ell}a_{i}s_{i}^{m}f_{m}(\alpha_{i}(z))\,,
\end{equation}
as may readily be ascertained by direct calculation. The right-hand
side is in $G$ (since $z_{1}\not\in\overline{\alpha_{i}(D)}$) for
$i=2,\dots,\ell$, so that $f_{m}$ is a seed function for $T^{(m)}$.
Moreover, $I-T^{(m)}$ is invertible in $G$ because $||T^{(m)}||\le K<1$.
We may therefore apply Theorem~\ref{thm:fixed_point_thm} with $F$
$=$ $<f_{m}>$ $\oplus$ $G$ to obtain a one-dimensional subspace
of fixed points $<\widehat{f_{m}}>$ of $T^{(m)}$ in $F$.
To obtain a fixed point of $T$, we integrate $m$ times, although
we must then handle a polynomial of degree at most $m-1$ that arises
from the constants of integration. Specifically, let us define the
integration operator $I:G\to G$ by the integral on the line segment
$[0,z]$ for $z\in D$:
\begin{equation}
I(g)(z)=\int_{0}^{z}g(w)dw\,.
\end{equation}
Denoting the $m-$th iterate of $I$ by $I^{m}$, and noting that
$\widehat{f_{m}}-f_{m}\in G$, we may define the function $f+I^{m}(\widehat{f_{m}}-f_{m})$
which we denote $\hat{f}$. The function $\hat{f}$ is not necessarily
a fixed point of $T$. However, differentiating $T\hat{f}-\hat{f}$
$m$ times, we obtain
\begin{align}
\left(T\hat{f}-\hat{f}\right)^{(m)} & =\left(Tf-f+TI^{m}(\widehat{f_{m}}-f_{m})-I^{m}(\widehat{f_{m}}-f_{m})\right)^{(m)}\\
& =(T^{(m)}f_{m}-f_{m})+T^{(m)}(\widehat{f_{m}}-f_{m})-(\widehat{f_{m}}-f_{m})\\
& =0\,,
\end{align}
since $T^{(m)}\widehat{f_{m}}=\widehat{f_{m}}$. It follows that $T\hat{f}-\hat{f}=q_{m}$,
where $q_{m}$ is a polynomial of degree at most $m-1$. Now let $p_{m}=(I-T)^{-1}q_{m}$,
a polynomial of degree at most $m-1$, the inverse existing because
of~\eqref{eq:no-poly-soln}. Then we have immediately that $T(\widehat{f}+p_{m})-(\widehat{f}+p_{m})=0$,
so that $\widehat{f}+p_{m}$ is a fixed point of $T$.
We have therefore proved the following theorem: \begin{theorem} Let $T$
be an affine Composition Sum Operator given by
\[
T(f)(z)=\sum_{i=1}^{\ell}a_{i}f(\alpha_{i}(z))\,,
\]
where $\ell\ge2$ is an integer, and for $i=1,\dots,\ell$, $a_{i}$
$\in\bbbc$, and $\alpha_{i}(z)=s_{i}(z-z_{i})+z_{i}$ are affine
contractions. Let $R>0$ be such that there exists $\delta>0$ with
$\overline{\alpha_{i}(D_{R+\delta})}\subseteq D_{R-\delta}$ for $i=1,\dots,\ell$
Then
\begin{enumerate}
\item $T$ has a fixed point which is a non-zero polynomial if and only
if $a_{1}s_{1}^{m}+a_{2}s_{2}^{m}+\dots+a_{\ell}s_{\ell}^{m}=1$ for
some integer $m\ge0$. If there are polynomial fixed points, there
is also a maximum integer $m$ satisfying the constraint, and all
the fixed points are then of degree at most $m$.
\item If there are no polynomial fixed point
\footnote{This condition guarantees the invertibility of $I-T$, and hence the
existence of PESL fixed points. However if polynomial fixed points
do exist, the possibility of PESL fixed points is not ruled out, and
if they do exist they will satisfy the conditions given above for
$a_{i}$.
}, but for some $1\le i\le l$, $\alpha_{i}$ is fixed point independent
on $D_{R}$ (ie $z_{i}\not\in\bigcup_{j\ne i}\overline{\alpha_{j}(D_{R})}$),
then:
\begin{enumerate}
\item If $a_{i}=s_{i}^{k}$, for some $k\ge1$, then $T$ has a fixed point
$f$ of the form
\[
f(z)=(z-z_{i})^{-k}+g(z)
\]
where $g\in G$ is analytic and bounded in $D_{R}$.
\item If $a_{i}=1$, then $T$ has a fixed point $f$ of the form
\[
f(z)=\log(z-z_{i})+g(z)
\]
where $g\in G$ is analytic and bounded in $D_{R}$.
\end{enumerate}
\end{enumerate}
Moreover if every $\alpha_{i}$ of $T$ is fixed point independent
on $D_{R}$, then every simple fixed point of $T$ which is of PES
\footnote{Having only Pole/Essential/Simple Log singularities - see \ref{PESL
} type is necessarily a linear combination of fixed points satisfying
the conditions above.
\end{theorem}
\section{\label{sub:CaseStudy}Applications to problems arising from renormalisation
theory}
We now consider further the operator $M$ given in~\eqref{eq:M}
above and we discuss the construction of fixed points of $M$. The
approach we adopt differs from that in the previous section in that
we work directly from the operator $M$, modifying it by subtracting
a constant CSO to obtain a contraction. This is more in the spirit
of the work in \cite{Mestel2000,Dalton-Mestel2003}. We start by developing
a general theory of CSOs acting on $\ell_{1}$ spaces of analytic
functions, a theory which is complementary to that developed in Sections~3
and~4.
\subsection{CSOs acting on $\ell_{1}$ spaces of analytic functions}
For $R>0$, let $G_{R}$ denote the complex Banach space of analytic
functions on the open disc $D_{R}=\{z:\,|z|<R\}$ with finite $\ell_{1}-$norm
$\left\Vert \sum_{n=0}^{\infty}c_{n}z^{n}\right\Vert _{R}=\sum_{n=0}^{\infty}|c_{n}|R^{n}$.
For $n\ge0$, we denote by $Z_{n}$ the basis function $Z_{n}:\, z\longmapsto z^{n}$,
which has norm $R^{n}$. The set $\{Z_{n}\;:\; n=0,1,2,\dots\}$ forms
a basis for $G_{R}$. We note the following standard lemma, which
we include for completeness.
\begin{lemma} \label{R:TonZn} Let $T$ be a bounded linear operator
on the Banach space $G_{R}$ of analytic functions and let $K>0$.
Then the induced operator norm $\bigl\Vert T\bigr\Vert_{R}\le K$
if, and only if, $\bigl\Vert TZ_{k}\bigr\Vert_{R}\le K\left\Vert Z_{k}\right\Vert $
for all $k\ge0$. The result also holds when the inequality is replaced
with a strict inequality. \end{lemma}
It follows that $T$ is a contraction on $G_{R}$ with contraction
rate $K<1$ if, and only if, it contracts each basis function $Z_{k}$
with contraction rate $K$.
\begin{proof} First suppose $\bigl\Vert T\bigr\Vert_{R}\le K$. Since
$Z_{k}\in G_{R}$, $\bigl\Vert TZ_{k}\bigr\Vert_{R}$ $\le$ $\bigl\Vert T\bigr\Vert_{R}\bigl\Vert Z_{k}\bigr\Vert_{R}$
$\le$ $K.R^{k}$, as required. We now prove the converse. Let $f=\sum_{r=0}^{\infty}a_{r}Z_{r}\in G_{R}$.
Since $T$ is bounded, hence continuous, $Tf=T\lim_{n\to\infty}\sum_{r=0}^{n}a_{r}Z_{r}=\lim_{n\to\infty}T\sum_{r=0}^{n}a_{r}Z_{r}=\lim_{n\to\infty}\sum_{r=0}^{n}a_{r}TZ_{r}$
and so it follows that $\bigl\Vert Tf\bigr\Vert_{R}$ $=$ $\lim_{n\to\infty}\bigl\Vert\sum_{r=0}^{n}a_{r}TZ_{r}\bigr\Vert_{R}$
$\le$ $\lim_{k\to\infty}\sum_{r=0}^{k}|a_{r}|KR^{r}$ $=$ $K\bigl\Vert f\bigr\Vert_{R}$
whence $\bigl\Vert T\bigr\Vert_{R}\le K$, as claimed. This completes
the proof. \end{proof}
One particular feature of affine CSOs is that they are contractions
on the basis functions $Z_{k}$ for $k$ sufficiently large, as is
shown by the following result.
\begin{lemma} Let $T=\sum_{i=1}^{\ell}a_{i}\alpha_{i}^{*}$ be a CSO
with $a_{i}$ constant (and non-zero), $\alpha_{i}(z)=s_{i}z+t_{i}$
where $|s_{i}|<1$. Let $s=\max_{i}\{|s_{i}|\}$, and let $\mu\in\bbbr$
satisfy $s<\mu\le1$.
Then there exists $R_{0}\ge0$, and integer $N>1$ such that
\begin{enumerate}
\item $\bigl\Vert TZ_{k}\bigr\Vert_{R}<\mu^{k}R^{k}$ for all $R>R_{0}$
and all $n\ge N$.
\item For $0\le n<N$, $\left\Vert TZ_{n}\right\Vert _{R}<\mu R^{n}$ for
all $R>R_{0}$, whenever $\left|\sum_{i=1}^{\ell}a_{i}s_{i}^{k}\right|<\mu$.
\end{enumerate}
\end{lemma}
\begin{proof} We have $(TZ_{n})(z)$ $=$ $\sum_{i}a_{i}(s_{i}z+t_{i})^{n}$
$=$ $\sum_{r=0}^{n}z^{r}\sum_{i}a_{i}\binom{n}{r}s_{i}^{r}t_{i}^{n-r}$,
hence it follows that $\left\Vert TZ_{n}\right\Vert _{R}$ $\le$
$\sum_{r=0}^{n}\left|\sum_{i}a_{i}\binom{n}{r}s_{i}^{r}t_{i}^{n-r}\right|.R^{r}$
$=$ $R^{n}\sum_{r=0}^{n}\binom{n}{r}\left|\sum_{i}a_{i}s_{i}^{r}\left(\frac{t_{i}}{R}\right)^{n-r}\right|$.
Therefore $\left\Vert TZ_{n}\right\Vert _{R}\le R^{n}\sum_{i}\left|a_{i}\right|.(\left|s_{i}\right|+\left|\frac{t_{i}}{R}\right|)^{n}$
from which we see that if $\left|s_{i}\right|+\left|\frac{t_{i}}{R}\right|<\mu\le1$
for all $i$, then we can find $N\ge0$ so that $\left\Vert TZ_{n}\right\Vert <\mu^{n}R^{n}$
for $n\ge N$. The condition on $R$ equates to $R_{0}=\max_{i}\{\frac{|t_{i}|}{\mu-|s_{i}|}\}$.
We now consider $n<N$ for $N>1$. From above $\left\Vert TZ_{n}\right\Vert _{R}\le R^{n}\left(\left|\sum_{i}a_{i}s_{i}^{n}\right|+\sum_{r=0}^{n-1}\binom{n}{r}\left|\sum_{i}a_{i}s_{i}^{r}\left(\frac{t_{i}}{R}\right)^{n-r}\right|\right)$.
Since $n$ is now bounded, for any $\epsilon>0$, we can choose $R$
large enough to give $\left\Vert TZ_{n}\right\Vert _{R}\le R^{n}\left(\left|\sum_{i}a_{i}s_{i}^{n}\right|+\epsilon\right)$.
Hence if $\left|\sum_{i}a_{i}s_{i}^{n}\right|<\mu\le1$ for $0\le n<N$,
we will have $\left\Vert TZ_{n}\right\Vert _{R}<\mu R^{n}$ for large
enough $R$. \end{proof}
The following corollaries are immediate. \begin{corollary}\label{cor:TcontractionforlargeR}
If $T$ is an affine CSO with $|s_{i}|<1$ and, for each $n\ge0$,
$\left|\sum_{i}a_{i}s_{i}^{n}\right|<1$, then $T$ is a contraction
on $G_{R}$ for large enough $R$. \end{corollary}
\begin{corollary}\label{cor:TcontractionforlargeR-1} If $T$ is an affine
CSO with $|s_{i}|<1$ and, for each $n\ge1$, $\left|\sum_{i}a_{i}s_{i}^{n}\right|<1$,
then $T_{c}$ is a contraction on $G_{R}$ for large enough $R$,
where $c$ is a constant and $T_{c}f=Tf-c^{*}(Tf)=\sum_{i=1}^{\ell}a_{i}f\alpha_{i}-\sum_{i=1}^{\ell}a_{i}f\alpha_{i}(c)$.
\end{corollary} \begin{proof} Note that for any constants $c$ and $a$,
$ac^{*}$ is a degenerate affine CSO with $s=0$, so that for $n\ge1$
the sum $\left|\sum_{i}a_{i}s_{i}^{n}\right|$ is unchanged between
$T$ and $T_{c}.$ But the sum is precisely $0$ for $n=0$, and so
the previous corollary can be applied to $T_{c}$.\end{proof}
For a seed function $f$, it may happen that the function $g=Tf-f\in G_{R}$
only for $R$ in a restricted range. In these circumstances it may
not be possible to apply Corollaries~\ref{cor:TcontractionforlargeR}
and~\ref{cor:TcontractionforlargeR-1} directly. Instead we may have
to iterate $T$ several times so that the domain on which $g$ is
defined is extended to include $D_{R}$ for $R$ sufficiently large
for Corollaries~\ref{cor:TcontractionforlargeR} and~\ref{cor:TcontractionforlargeR-1}
to apply. That it is possible to do this follows from the fact that
the $\alpha_{i}$ contract the whole of $\bbbc$ uniformly.
Let us first note that there exists $R_{0}\ge0$ such that $\alpha_{i}(D_{R})\subseteq(D_{R})$
for each $i=1,\dots,\ell$ and each $R\ge R_{0}$. The following domain
expansion lemma is straightforward to prove. \begin{lemma} \label{lem:Extending_T}
Let $g \in G_{R_{1}}$ for
some $R_{1}>R_{0}$. Then for each $R\ge R_{1}$ there exists an integer
$K\ge0$ such that $T^{k}g\in G_{R}$ for all $k\ge K$. \end{lemma} \begin{proof} The
proof is a straightforward consequence of the fact that the $\alpha_{i}$
contract uniformly. If $g\in G_{R}$, then the result holds with $k=0$.
Otherwise, let $k\ge1$ and consider a composition of $k$ contractions
chosen from the $\alpha_{i}$, possibly with repetition. The resulting
composition $\alpha_{i_{1}}\dots\alpha_{i_{k}}$ is an affine map
so we may write $\alpha_{i_{1}}\dots\alpha_{i_{k}}(z)=s_{i_{1}\dots i_{k}}z+t_{i_{1}\dots i_{k}}$,
for $i_{1},\dots,i_{k}\in\{1,\dots,\ell\}$. Now $t_{i_{1}\dots i_{k}}=\alpha_{i_{1}}\dots\alpha_{i_{k}}(0)\in D_{R_{0}}$.
Moreover, since the $\alpha_{i}$ contract uniformly on $\bbbc$,
the sequence $s_{i_{1}\dots i_{k}}\to0$ uniformly in $k$ as $k\to\infty$.
It follows immediately, that there must exist $k\ge1$ such that $\alpha_{i_{1}}\dots\alpha_{i_{k}}(D_{R})\subseteq D_{R_{1}}$
for all $i_{1},\dots,i_{k}\in\{1,\dots,\ell\}$. It follows that for all
$k$ large enough, we have $T^{k}g\in G_{R}$, as claimed. \end{proof}
In the application we shall consider in the next subsection, the operator
$T$ fails to be a contraction, because it is not a contraction on
constant functions. To solve this problem, for any affine CSO $T$,
we introduce a new derived operator which is a contraction, and which
shares certain fixed points with $T$. We give the construction for
general $\ell\ge2$, although in our application we shall specialise
to the binary case $\ell=2$.
For $1\le j\le\ell$, we define the operator $T_{j}$ by
\begin{equation}
T_{j}f=Tf-\frac{1}{L}\sum_{i=1,i\not=j}^{\ell}a_{i}Tf(\alpha_{i}(z_{j}))\,,\quad L=\sum_{i=1,i\not=j}^{\ell}a_{i}\label{eq:Tj}
\end{equation}
provided $L\not=0$. We note that, if $a_{j}=1$, a fixed point of
$T_{j}$ is also a fixed point of $T$. For, suppose $T_{j}f=f$.
Then
\begin{equation}
f(z)=Tf(z)-\frac{1}{L}\sum_{i=1,i\not=j}^{\ell}a_{i}Tf(\alpha_{i}(z_{j}))\,.\label{eq:Tjf=00003D00003D00003Df}
\end{equation}
Taking a weighted sum of this equation evaluated at $\alpha_{i}(z_{j})$,
gives
\begin{equation}
\sum_{i=1,i\not=j}^{\ell}a_{i}f(\alpha_{i}(z_{j}))=\sum_{i=1,i\not=j}^{\ell}a_{i}\left(Tf(\alpha_{i}(z_{j}))-\frac{1}{L}\sum_{k=1,k\not=j}^{\ell}a_{k}Tf(\alpha_{k}(z_{j}))\right)=0\,.\label{eq:sum_ai_f_alpha_zi}
\end{equation}
Hence, using $\alpha_{j}(z_{j})=z_{j}$ we obtain
\begin{equation}
\begin{split}
f(z_{j})=Tf(z_{j})-\frac{1}{L}\sum_{i=1,i\not=j}^{\ell}a_{i}Tf(\alpha_{i}(z_{j})) \\
=a_{j}f(z_{j})+\sum_{i=1,i\not=j}^{\ell}a_{i}f(\alpha_{i}(z_{j}))-\frac{1}{L}\sum_{i=1,i\not=j}^{\ell}a_{i}Tf(\alpha_{i}(z_{j})),
\end{split}
\end{equation}
whence, since $a_{j}=1$,
\begin{equation}
\frac{1}{L}\sum_{i=1,i\not=j}^{\ell}a_{i}Tf(\alpha_{i}(z_{j}))=\sum_{i=1,i\not=j}^{\ell}a_{i}f(\alpha_{i}(z_{j}))=0\,,
\end{equation}
from \eqref{eq:sum_ai_f_alpha_zi}. It follows that $Tf=f$, as claimed.
If $T$ is a binary CSO, then we can write the operator $T_{j}$ as
$T_{c}$ where the modified operator $T_{c}$ is given by $T_{c}f=Tf-Tf(c)$
with $c=\alpha_{i}(z_{j})$ and where now $\{i,j\}$ is precisely
$\{1,2\}$. It is immediate that, for a binary CSO, any fixed point
$f$ of $T_{c}$ is a fixed point of $T$ and $f(c)=0$ by~\eqref{eq:sum_ai_f_alpha_zi}.
\subsection{Fixed points of the operator $M$}
We now apply the theory we have developed to find fixed points of
the operator $M$ introduced in Section~\ref{sec:seeded-fixed-point-thy}.
Recall that $M=\phi_{1}^{*}+\phi_{2}^{*}$ with $\phi_{1}(z)=-\omega z,\phi_{2}(z)=\omega^{2}z+\omega$,
where $\omega=\frac{1}{2}(\sqrt{5}-1)$. For consistency with~\cite{Mestel2000,Dalton-Mestel2003},
we use the notation $\phi_{i}=\alpha_{i}$ for $i=1,2$.
Now, in the notation used above, $s=\max(\omega,\omega^{2})=\omega$.
Let $\omega<\mu<1$. It follows that, for $N$ sufficiently large,
$M(Z_{n})\le\mu^{n}R^{n}$, for $n\ge N$ and $R\ge R_{0}$. In fact
it is readily seen that $N=2$ suffices when $R\ge1.9009$. For $0\le n\le1$,
we calculate as follows. If $n=1$, $\left|\sum_{i}\lambda_{i}s_{i}^{n}\right|=\left|(-\omega)+\omega^{2}\right|=\omega^{3}<1$
and we have a contraction for $R$ sufficiently large. It is straightforward
to verify that is sufficient to take $R\ge1.619$. We therefore need
to choose $R\ge1.9009$. However, for $n=0$, $\left|\sum_{i}\lambda_{i}s_{i}^{n}\right|=2>1$
and we do not have a contraction.
Let us consider the operator $M_{c}$ introduced at the end of the
last section and given by $M_{c}f=Mf-Mf(c)$. Since $a_{1}=a_{2}=1$,
in this case, we can take in turn $j=1,2$ and choose, in turn, $c=c_{1},c_{2}$,
where $c_{1}=\phi_{1}(1)=-\omega$ and $c_{2}=\phi_{2}(0)=\omega$.
We note that, for any $c$, $M_{c}$ is itself a (degenerate) CSO.
Indeed, $M_{c}=\phi_{1}^{*}+\phi_{2}^{*}-(\phi_{1}c)^{*}-(\phi{}_{2}c)^{*}=\sum_{1}^{4}a_{i}\alpha_{i}$.
The last two terms $(i=3,4)$ are constants, so are degenerate affine
contractions.
Let us now construct fixed points of $M_{c}$ and hence of $M$. First
we note that $M_{c}Z_{0}=0$, and, from the above calculations, we
see that $M_{c}$ contracts the functions $Z_{n}:\, z\longmapsto z^{n}$
for $n\ge1$. We deduce from Corollary~\ref{cor:TcontractionforlargeR-1}
that $M_{c}$ is a contraction on $G_{R}$ for large enough $R$.
Hence $M_{c}$ has no non-trivial analytic fixed points in $G_{R}$,
for $R$ large enough.
Our first task is to consider a space of seed functions for $M$.
Since $a_{1}$ $=$ $a_{2}$ $=$ $1$, we can look for seeds with
logarithmic singularities at $0$, $1$, the fixed points of $\phi_{1}$
and $\phi_{2}$ respectively. Indeed for simple unbounded singularities,
the space of seed functions for $M_{c}$ is the span $<\log z,\log(z-1)>$.
Hence, to find the fixed points of $M_{c}$ we apply the previous
theory to obtain a fixed point for each of the two basis seed functions
$\log z$ and $\log(z-1)$.
We calculate $M$ acting on the two basis (generalised) seed functions.
We have $M\log z=$ $\log(-\omega z)+\log(\omega^{2}z+\omega)$ $=$
$\log z+\log(1+\omega z)+b_{1}$, where $b_{1}$ is a constant. Similarly
$M\log(z-1)=\log(-\omega z-1)+\log(\omega^{2}z+\omega-1))$ $=$ $\log(z-1)+\log(1+\omega z)+b_{2}$,
$b_{2}$ constant (where we have used $1-\omega=\omega^{2}$). It
follows readily that in both cases we have (using the notation of Lemma \ref{lem:Extending_T}) $\overline{M_{c}}\log z$,
$\overline{M_{c}}\log(z-1)$ $\in G_{R_1}$, provided $1=R_0<R_1<\omega^{-1}$.
Because of the restriction on $R_1$, we cannot necessarily use Theorem~\ref{thm:fixed_point_thm}
directly, but if not we can use Corollary~\ref{cor:FixedPointThmExtension}
instead. For convenience, in what follows, we write $f(z)$ to represent one
of the generalised seed functions $\log z$ and $\log(z-1)$ and we
write $\tilde{f}(z)=\log(1+\omega z)$.
Using Corollary~\ref{cor:TcontractionforlargeR-1} we choose $R$
sufficiently large so that $M_{c}$ is a contraction on $G_{R}$. If we can choose $1<R<\omega^{-1}$, we can use
Theorem~\ref{thm:fixed_point_thm} directly to obtain a fixed point of $M_{c}$. Otherwise
using Lemma~\ref{lem:Extending_T}, we set $k\ge0$ such that
$M_{c}^{k}\overline{M_{c}}f\in G_{R}$. We note that in fact $k\ge1$ since
$\tilde{f}\not\in G_{R}$, because $R\ge\omega^{-1}$. It also follows
that $M_{c}^{k}(f)$ is a seed function since $\overline{M_{c}}M_{c}^{k}f\in G_{R}$.
A fixed point of $M_{c}$ is now obtained from this seed function by applying
Corollary~\ref{cor:FixedPointThmExtension}.
We can obtain some explicit expansions for the fixed points. From
the above calculations, we have $Mf=f+\tilde{f}+b$, where $b$ is
a constant, whence $M_{c}f=I_{c}f+I_{c}\tilde{f}$. Hence we readily obtain for $k\ge1$
that $M_{c}^{k}f=I_{c}f+I_{c}\tilde{f}+\sum_{n=1}^{k-1}M_{c}^{n}\tilde{f}$,
and so $\overline{M_{c}}M_{c}^{k}f=-M_{c}^{k}\tilde{f}$. We conclude
that $M_{c}^{k}\tilde{f}\in G_{R}$.
Let
\begin{equation}
f_{*}=I_{c}f+I_{c}\tilde{f}+\sum_{n=1}^{\infty}M_{c}^{n}\tilde{f}\,.\label{eq:f*}
\end{equation}
The infinite sum converges because $M_{c}$ is a contraction on $G_{R}$
and $M_{c}^{k}\tilde{f}\in G_{R}$. Now,
\begin{equation}
M_{c}f_{*}=M_{c}f+M_{c}\tilde{f}+\sum_{n=2}^{\infty}M_{c}^{n}\tilde{f}=I_{c}f+I_{c}\tilde{f}+\sum_{n=1}^{\infty}M_{c}^{n}\tilde{f}=f_{*}\,.
\end{equation}
since $M_{c}f=I_{c}(f+\tilde{f})$ $=$ $I_{c}f+I_{c}\tilde{f}$.
As remarked above, we choose in turn $f(z)=\log z$ and $c=$ $c_{1}=\phi_{1}(1)=-\omega$,
$f(z)=\log(z-1)$, $c_{2}=\phi_{2}(0)=\omega$. From~\eqref{eq:f*}
and noting that $M_{c}^{n}g=M^{n}(g)-(M^{n}g)(c)$, this gives a fixed
point space for $M$ of $<f_{1},f_{2}>$, where
\begin{equation}
f_{1}(z)=\log\frac{z}{-\omega}+\sum_{n=0}^{\infty}\sum_{\underline{i}\in I^{n}}\log\frac{1+\omega\phi_{\underline{i}}z}{1+\omega\phi_{\underline{i}}(-\omega)}
\end{equation}
and
\begin{equation}
f_{2}(z)=\log\frac{z-1}{\omega-1}+\sum_{n=0}^{\infty}\sum_{\underline{i}\in I^{n}}\log\frac{1+\omega\phi_{\underline{i}}z}{1+\omega\phi_{\underline{i}}(\omega)}\,,
\end{equation}
where $\underline{i}=(i_{1},i_{2},\ldots,i_{n}),\ I^{n}=\{1,2\}^{n},\ \phi_{\underline{i}}=\circ_{j=1}^{n}\phi_{i_{j}}$
$=$ $=\phi_{i_{1}}\dots\phi_{i_{n}}$, for $n>0$, and the identity
map when $n=0$. We note that $f_{2}$ is the fixed point already
reported by Mestel \emph{et al} in~\cite{Mestel2000}.
A particularly elegant example of a fixed point of $M$ is obtained
by putting $f_{3}=f_{1}-f_{2}$ to give the fixed point $f_{3}(z)$
$=$ $\left[\log\frac{z}{z-1}.\frac{\omega-1}{-\omega}+W\right]$
where $W=\sum_{n=0}^{\infty}\sum_{\underline{i}\in I^{n}}\log\frac{1+\omega\phi_{\underline{i}}\omega}{1+\omega\phi_{\underline{i}}(-\omega)}$.
Since this is a fixed point we also have $f_{3}=Mf_{3}=\left[M\log\frac{z}{z-1}.\frac{\omega-1}{-\omega}+2W\right]$
$=$ $\log\frac{-\omega z}{-\omega z-1}.\frac{\omega^{2}z+\omega}{\omega^{2}z+\omega-1}.\left(\frac{\omega-1}{-\omega}\right)^{2}+2W$
$=$ $\left[\log\frac{z}{z-1}.\omega^{2}+2W\right]$. Hence $W=-\log\omega$
$=\log(1+\omega)$, by the properties of $\omega$, and $f_{3}(z)=\lambda\log\frac{z}{z-1}$,
where $\lambda$ is constant. This gives us the subspace of fixed
points $<\log\frac{z}{z-1}>$ and also the identity $e^{W}=$
\begin{equation}
\prod_{n=0}^{\infty}\prod_{\underline{i}\in I^{n}}\left(\frac{1+\omega\phi_{\underline{i}}\omega}{1+\omega\phi_{\underline{i}}(-\omega)}\right)=1+\omega\,.
\end{equation}
Clearly we can take exponentials of the fixed points $f_{1},f_{2}$
to obtain instead fixed points of the \emph{multiplicative} functional
equation $f(z)=f(-\omega z).f(\omega^{2}z+\omega)$. The singularities
are removable and can be replaced with zeroes to obtain entire functions.
The real parts of $\exp f_{1},\exp f_{2}$ are shown in Figure \ref{fig:Graph},
and it can be seen that this is consistent with the identity $\frac{\exp f_{1}}{\exp f_{2}}=\frac{z}{z-1}$.
\begin{figure}
\noindent \begin{centering}
\includegraphics[scale=0.2]{Figure1}
\par\end{centering}
\caption{\label{fig:Graph}Graph of the real parts of the two multiplicative
fixed points}
\end{figure}
\section{Conclusion and further directions for research}
In this paper we have introduced a new theory of non-zero fixed points
of linear operators $T$, showing the existence of a fixed point operator
$\widehat{T}$ derived from $T$ whose image is all the fixed points
of $T$. We have applied this theory to ``affine composition sum
operators'', a class of operators whose fixed points are important
in the study of the renormalisation of a variety of physical problems.
In particular this has enabled us, under a simple set of constraints,
to show the necessary form of fixed points which are of PESL (Pole/Essential/Simple
Log) type (see \eqref{PESL} for the full definition), these being
currently seen as the most important type. The techniques are readily
extended to functions with other types of singularity such as removable
of algebraic singularities. In addition we developed simple tests
for their existence. Finally we have applied the theory to the much
studied operator $Mf(z)=f(-\omega x)+f(\omega^{2}x+\omega)$ (see
\cite{Mestel2000,Dalton-Mestel2003} and subsequent papers) to deduce
for the first time the \emph{complete} set of \emph{simple} (see \eqref{simple})
fixed points of PESL type, including previously unknown solutions.
There are several directions for further research in this area. First,
we may extend our study to cover the full spectrum of affine CSOs.
Considered as linear operators on function spaces of analytic functions,
CSOs are compact operators and thus have discrete non-zero spectrum.
It is likely that the techniques developed in this paper may be adapted
to construct more general eigenfunctions of affine CSOs, with a view
to obtaining a full description of their spectra.
Second, it is likely that the approach of, for example,~\cite{Mestel2004a}
may be applied to understand fixed points of an affine CSO with non-simple
unbounded singularity set and, more generally, all periodic points
of an affine CSO. An full understanding of the latter is indeed necessary
for a complete description of all the fixed points of a CSO. For let
$f_{1}$, $f_{2}$ be a periodic orbit of period-2 of a CSO $T$.
Then $Tf_{1}=f_{2}$ and $Tf_{2}=f_{1}$ so that $f=f_{1}+f_{2}$
is generally a non-simple fixed point of $T$, a construction that
clearly generalises to other periods.
Third, an important future direction is to consider more general CSOs
than affine CSOs. Of course, explicit construction of fixed points
(and more general eigenfunctions) may not be in general possible for
non-affine CSOs. However a general theory may well be possible and
it may be possible to make extensive progress for special important
cases. An analogy may be drawn here with the theory of linear differential
equations. The theory of constant coefficient linear differential
equations is complete, while that for general linear equations is
less well developed except in special cases of particular interest.
Nevertheless, non-constant coefficient CSOs are of considerable interest.
For example, the full Stretch-Fold-Shear toy model studied by Gilbert~\cite{Gilbert2002,Gilbert2005}
involves a study of the spectrum of the CSO $T$ on complex-valued
$c$ functions of a real variable $x$ given by
\begin{equation}
Tc(x)=e^{i\alpha(x-1)/2}\; c\left(\frac{x-1}{2}\right)-e^{i\alpha(1-x)/2}\; c\left(\frac{1-x}{2}\right)\,,
\end{equation}
where $\alpha\ge0$ is a real parameter, corresponding to the level
of shear in the map.
Recall that a CSO given by~\eqref{eq:CSOs} is affine if each of
the coefficients $a_{i}$ is constant and each of the maps $\alpha_{i}$
is an affine contraction. While it would certainly be interesting
to relax each of these conditions, a theory for non-constant $a_{i}$
would be of immediate application is several areas including the kinematic
dynamo theory discussed in~\cite{Gilbert2002,Gilbert2005} and in
the study of non-chaotic strange attractors~\cite{Feudel-Kuznetsov-Pikovsky2006}.
Finally, a promising area for research is to study CSOs in a wider
context than spaces of analytic functions with singularities. Indeed,
work on CSOs in spaces of piecewise constant real functions has already
found fruitful application in several fields, as detailed above. It
would be very interesting to develop a general theory of CSOs for
spaces of functions with discontinuities either of the function or
its derivatives.
\bibliographystyle{gDEA}
|
\subsection{Model of active groups}
Links of the network are influenced not only by nodes changing memberships to groups but also by the birth and death of groups themselves. New groups can be born \rev{and} old ones can die. However, without explicitly modeling group birth and death there exists ambiguity between group membership change and the birth/death of groups. For example, consider two disjoint groups $k$ and $l$ such that their lifetimes and members do not overlap. In other words, group $l$ is born after group $k$ dies out. However, if group birth and death dynamics is not explicitly modeled, then the model could interpret that the two groups correspond to a single latent group where all the members of $k$ leave the group before the members of $l$ join the group. To resolve this ambiguity we devise an explicit model of birth/death dynamics of groups by introducing a notion of \textit{\alive} groups.
Under our model, a group can be in one of two states: it can be either active (alive) or inactive (not yet born or dead). However, once a group becomes inactive, it can never be active again. That is, once a group dies, it can never be alive again. To ensure coherence of group's state over time, we build on the idea of distance-dependent Indian Buffet Processes (dd-IBP)~\cite{ddibp}. The IBP is named after a metaphorical process that gives rise to \rev{a} probability distribution, where customers enter an Indian Buffet restaurant and sample some subset of an infinitely long sequence of dishes. In the context of networks, nodes usually correspond to `customers' and latent features/groups correspond to `dishes'. However, we apply dd-IBP in a different way. We regard each time step \rev{$t$} as a `customer' that samples a set of active groups \rev{$\mathcal{K}_{t}$}.
So, at the first time step $t = 1$, we have
\rev{$Poisson(\lambda)$ number of groups}
that are initially {\alive}, \rev{\ie, $|\mathcal{K}_{1}| \sim Poisson(\lambda)$}. To account for death of groups we \rev{then} consider that each {\alive} group at time $t - 1$ can become inactive at the next time step $t$ with probability $\gamma$.
On the other hand,
\rev{$Poisson(\gamma\lambda)$ new groups}
are \rev{also} born at time $t$.
Thus, at each time currently active groups can die, while new ones can also be born.
The hyperparameter $\gamma$ controls for how often new groups are born and how often old ones die. For instance, there will be almost no newborn or dead groups if $\gamma \approx 1$, while there would be no temporal \rev{group} coherence and practically all the groups would die between consecutive time steps if $\gamma = 0$.
Figure~\ref{fig:model}(a) gives an example of the above process.
\rev{Black circles indicate {\alive} groups and white circles denote inactive (not yet born or dead) groups.}
Groups 1 and 3 exist at $t = 1$ and Group 2 is born at $t = 2$.
\rev{At $t = 3$, Group 3 dies but Group 4 is born.
Without our group activity model,
Group 3 could have been reused with a completely new set of members and Group 4 would have never been born.
Our model can distinguish these two disjoint groups.
}
Formally, we denote the number of active groups at time $t$ by $K_{t} = |\mathcal{K}_{t}|$. We also denote the state (active/inactive) of group $k$ at time $t$ by $W_{k}^{(t)} = \mathbf{1}\{k \in \mathcal{K}_{t}\}$. For convenience, we also define a set of newly active groups at time $t$ be $\mathcal{K}_{t}^{+} = \{k | W_{k}^{(t)} = 1, W_{k}^{(t')} = 0 ~\forall t' < t\}$ and $K_{t}^{+} = |\mathcal{K}_{t}^{+}|$.
Putting it all together we can now fully describe the process of group birth/death as follows:
\begin{align}
K_{t}^{+} & \sim
\begin{cases}
Poisson\left(\lambda\right), & \mbox{for } t = 1 \\
Poisson\left(\gamma\lambda\right), & \mbox{for } t > 1
\end{cases} \nonumber \\
W_{k}^{(t)} & \sim
\begin{cases}
Bernoulli(1 - \gamma) & \mbox{if } W_{k}^{(t-1)} = 1 \\
1, & \mbox{if } \sum_{t' = 1}^{t-1} K_{t'}^{+} < k \leq \sum_{t' = 1}^{t} K_{t'}^{+} \\
0, & \mbox{otherwise} \, .
\end{cases}
\label{eq:groupmodel}
\end{align}
Note that under this model an infinite number of active groups can exist. This means our model automatically determines the right number of active groups and each node can belong to many groups simultaneously. We now proceed by describing the model of node group membership dynamics.
\begin{figure}[t]
\centering
\begin{tabular}{cc}
\includegraphics[width=0.45\textwidth]{FIG/FeatureWindow} &
\raisebox{3mm}{\includegraphics[width=0.5\textwidth]{FIG/LinkProbability}} \\
(a) Group activity model & (b) Link function model
\end{tabular}
\caption{\small {\bf (a) Birth and death of groups:} Black circles represent active and white circles represent inactive (unborn or dead) groups. A dead group can never become active again. {\bf (b) Link function:} $z_i^{(t)}$ denotes binary node group memberships. Entries of link affinity matrix $\Theta_k$ denotes linking parameters between all 4 combinations of members ($z_i^{(t)}=1$) and non-members ($z_i^{(t)}=0$). To obtain link probability $p_{ij}^{(t)}$, individual affinities $\Theta_k[z_j^{(t)}, z_j^{(t)}]$ are combined using a logistic function $g(\cdot)$}.
\label{fig:model}
\vspace{-5mm}
\end{figure}
\subsection{Dynamics of node group memberships}
We capture the dynamics of nodes joining and leaving groups by assuming that latent node group memberships \rev{form} a Markov chain. In this framework, node memberships to active groups evolve through time according to Markov dynamics:
$$P(z_{ik}^{(t)} | z_{ik}^{(t-1)}) = Q_{k} = \left( \begin{array}{cc} 1 - a_{k} & a_{k} \\ b_{k} & 1 - b_{k} \end{array}\right),$$
where matrix $Q_{k}[r,s]$ denotes a Markov transition from state $r$ to state $s$, which can be a fixed parameter, group specific, or otherwise domain dependent as long as it defines a Markov transition matrix.
Thus, the transition of node's $i$ membership to active group $k$ can be defined as follows:
\rev{
\begin{align}
a_{k}, b_{k} \sim Beta(\alpha, \beta) ,\, \,
z_{ik}^{(t)} \sim W_{k}^{(t)} \cdot Bernoulli\left(a_{k}^{1-z_{ik}^{(t-1)}}\left(1-b_{k}\right)^{z_{ik}^{(t-1)}}\right) \,.
\end{align}
}
Typically, $\beta > \alpha$, which ensures that group's memberships are not too volatile over time.
\subsection{Relationship between node group memberships and links of the network}
Last, we describe the part of the model that establishes the connection between node's memberships to groups and the links of the network. We achieve this by defining a link function $f(i,j)$, which for given a pair of nodes $i, j$ determines their interaction probability $p_{ij}^{(t)}$ based on their group memberships.
We build on the Multiplicative Attribute Graph model~\cite{mh11uai,mh12im}, where each group $k$ is associated with a link affinity matrix $\Theta_{k} \in \mathcal{R}^{2 \times 2}$. Each of the four entries of the link affinity matrix captures the tendency of linking between group's members, members and non-members, as well as non-members themselves. While traditionally link affinities were considered to be probabilities, we relax this assumption by allowing affinities to be arbitrary real numbers and then combine them through a logistic function to obtain a final link probability.
The model is illustrated in Figure~\ref{fig:model}(b). Given group memberships $z_{ik}^{(t)}$ and $z_{jk}^{(t)}$ of nodes $i$ and $j$ at time $t$ the binary indicators ``select'' an entry $\Theta_{k}[z_{ik}^{(t)}, z_{jk}^{(t)}]$ of matrix $\Theta_{k}$. This way linking tendency from node $i$ to node $j$ is reflected based on their membership to group $k$. We then determine the overall link probability $p_{ij}^{(t)}$ by combining the link affinities via a logistic function $g(\cdot)$\footnote{\rev{$g(x) = \exp(x) / (1 + \exp(x))$}}. \rev{Thus,}
\rev{
\begin{align}
p_{ij}^{(t)} = f(z_{i\cdot}^{(t)}, z_{j\cdot}^{(t)}) = g \left( \epsilon_{t} + \sum_{k=1}^{\infty} \Theta_{k}[z_{ik}^{(t)}, z_{jk}^{(t)}] \right) , \,\,\,
Y_{ij} \sim Bernoulli(p_{ij}^{(t)})
\end{align}
}
where
$\epsilon_{t}$ is a density parameter that reflects the varying link density of network over time.
Note that due to potentially infinite number of groups the sum of an infinite number of link affinities may not be tractable. To resolve this, we notice that for a given $\Theta_{k}$ subtracting $\Theta_{k}[0, 0]$ from all its entries and then adding this value to $\epsilon_{t}$ does not change the overall linking probability $p_{ij}^{(t)}$. Thus, we can set $\Theta_{k}[0, 0] = 0$ and then only a finite number \rev{of} affinities selected by $z_{ik}^{(t)}$ have to be considered. For all other entries of $\Theta_{k}$ we use $\mathcal{N}(0, \nu^{2})$ as a prior distribution.
To sum up, Figure~\ref{fig:plate} illustrates the three components of the {\model} in a plate notation. Group's state $W_{k}^{(t)}$ is determined by the dd-IBP process and each node-group membership $z_{ik}^{(t)}$ is defined as the FHMM over active groups. Then, the link between nodes $i$ and $j$ is determined based on the groups they belong to and the corresponding group link affinity matrices $\Theta$.
\begin{figure}[t]
\centering
\includegraphics[width=0.47\textwidth]{FIG/PlateModel}
\caption{\small \fullmodel. Network $Y$ depends on each node's group memberships $Z$ and active groups $W$. Links of $Y$ appear via link affinities $\Theta$.}
\label{fig:plate}
\vspace{-5mm}
\end{figure}
\subsection{Experimental setup}
For the two prediction experiments, we use the following three datasets.
First, the \emph{NIPS co-authorships network} connects two people if they appear on the same publication in the NIPS conference in a given year. Network spans $T$=17 years (1987 to 2003). Following~\cite{heaukulani13icml} we focus on a subset of 110 most connected people over all time periods.
Second, the \emph{DBLP co-authorship network} is obtained from 21 Computer Science conferences from 2000 to 2009 ($T$ = 10)~\cite{dblpdata}. We focus on 209 people by taking 7-core of the aggregated network for the entire time.
Third, the \emph{INFOCOM} dataset represents the \rev{physical} proximity interactions between 78 students at the 2006 INFOCOM conference, recorded by wireless detector remotes given to each attendee~\cite{infocomdata}. As in \cite{heaukulani13icml} we use the processed data that removes inactive time slices to have $T$=50.
To evaluate the predictive performance of our model, we compare it to three baseline models.
For a naive baseline model, we regard the relationship between each pair of nodes as the instance of independent Bernoulli distribution with $Beta(1, 1)$ prior. Thus, for a given pair of nodes, the link probability at each time equals to the expected probability from the posterior distribution given network data.
Second baseline is LFRM~\cite{lfrm}, a model of static networks. For missing link prediction, we independently fit LFRM to each snapshot of dynamic networks. For network forecasting task, we fit LFRM to the most recent snapshot of a network. Even though LFRM does not capture time dynamics, we consider this to be a strong baseline model.
Finally, for the comparison with dynamic network models, we consider two recent state of the art models. The DRIFT model~\cite{foulds11aistats} is based on an infinite factorial HMM and authors kindly shared their implementation. We also consider the LFP model~\cite{heaukulani13icml} for which we were not able to obtain the implementation, but since we use the same datasets, we compare performance numbers directly with those reported in~\cite{heaukulani13icml}.
To evaluate predictive performance, we use various standard evaluation metrics.
First, to assess goodness of inferred probability distributions, we report the log-likelihood of \rev{held-out} edges.
Second, to verify the predictive performance, we compute the area under the ROC curve (AUC). Last, we also report the maximum F1-score (F1) by scanning over all possible precision/recall thresholds.
\subsection{Task 1: Predicting missing links}
To generate the datasets for the task of missing link prediction, we randomly hold out 20\% of node pairs (\ie, either link or non-link) throughout the entire time period. We then run each model to obtain 400 samples after 800 burn-in samples for each of 10 MCMC chains. Each sample gives a link probability for a given missing entry, so the final link probability of a missing entry is computed by averaging the corresponding link probability over all the samples. This final link probability provides the evaluation metric for a given missing data entry.
\begin{table}[t]
\centering
\small
\begin{tabular}{|c||c|c|c||c|c|c||c|c|c|}
\hline
\multirow{2}{*}{Model} & \multicolumn{3}{c||}{NIPS} & \multicolumn{3}{c||}{DBLP} & \multicolumn{3}{c|}{INFOCOM} \\
\cline{2-10}
& TestLL & AUC & F1 & TestLL & AUC & F1 & TestLL & AUC & F1 \\
\hline \hline
Naive & -2030 & 0.808 & 0.177
& -12051 & 0.814 & 0.300
& -17821 & 0.677 & 0.252 \\
\hline
LFRM & -880 & 0.777 & 0.195
& -3783 & 0.784 & 0.146
& -8689 & 0.946 & 0.703 \\
\hline
DRIFT & -758 & 0.866 & 0.296
& -3108 & 0.916 & 0.421
& -6654 & 0.973 & 0.757 \\
\hline \hline
{\model} & $\threestar{\mathbf{-624}}$ & $\threestar{\mathbf{0.916}}$ & $\threestar{\mathbf{0.434}}$
& $\threestar{\mathbf{-2684}}$ & $\threestar{\mathbf{0.939}}$ & $\threestar{\mathbf{0.492}}$
& $\threestar{\mathbf{-6422}}$ & $\twostar{\mathbf{0.976}}$ & $\threestar{\mathbf{0.764}}$ \\
\hline
\end{tabular}
\caption{\small Missing link prediction. We bold the performance of the best scoring method. Our {\model} performs the best in all cases. All improvements are statistically significant at 0.01 significance level.}
\label{tbl:linkprediction}
\vspace{-5mm}
\end{table}
Table~\ref{tbl:linkprediction} shows average evaluation metrics for each model and dataset over 10 runs. We also compute the $p$-value on the difference between two best results for each dataset and metric. Overall, our {\model} model significantly outperforms the other models in every metric and dataset.
Particularly in terms of F1-score we gain up to 46.6\% improvement over the other models.
By comparing the naive model and LFRM, we observe that LFRM performs \rev{especially} poorly compared to the naive model in two \rev{networks with few edges} (NIPS and DBLP).
Intuitively this makes sense because due to the network sparsity we can obtain more information from the temporal trajectory of each link than from each snapshot of network. However, both DRIFT and {\model} successfully combine the temporal and the network information which results in better predictive performance. Furthermore, we note that {\model} outperforms \rev{the other models} by a larger margin as networks get sparser.
{\model} makes better use of temporal information
because it can explicitly model temporally local links through {\alive} groups.
Last, we also compare our model to the LFP model. The LFP paper reports AUC ROC score of $\sim$0.85 for NIPS and $\sim$0.95 for INFOCOM on the same task of missing link prediction with 20\% held-out missing data~\cite{heaukulani13icml}. Performance of our {\model} on these same networks under the same conditions is 0.916 for NIPS and 0.976 for INFOCOM, which is a strong improvement over LFP.
\newcommand{T_{obs}}{T_{obs}}
\subsection{Task 2: Future network forecasting}
Here we are given a dynamic network up to time $T_{obs}$ and the goal is to predict the network at the next time $T_{obs} + 1$. We follow the experimental protocol described in~\cite{foulds11aistats,heaukulani13icml}\rev{: We train} the models on first $T_{obs}$ networks, fix the parameters, and then for each model we run MCMC sampling one time step into the future.
For each model and network, we obtain 400 samples with 10 different MCMC chains, resulting in 400K network samples. These network samples provide a probability distribution over links at time $T_{obs} + 1$.
\begin{table}[t]
\centering
\small
\begin{tabular}{|c||c|c|c||c|c|c||c|c|c|}
\hline
\multirow{2}{*}{Model} & \multicolumn{3}{c||}{NIPS} & \multicolumn{3}{c||}{DBLP} & \multicolumn{3}{c|}{INFOCOM} \\
\cline{2-10}
& TestLL & AUC& F1 & TestLL & AUC& F1 & TestLL & AUC& F1 \\
\hline \hline
Naive & -547 & 0.524 & 0.130
& -3248 & $\mathbf{0.668}$ & 0.243
& -774 & 0.673 & 0.270 \\
\hline
LFRM & -356 & 0.398 & 0.011
& -1680 & 0.492 & 0.024
& -760 & 0.640 & 0.248 \\
\hline
DRIFT & $\threestar{\mathbf{-148}}$ & 0.672 & 0.084
& $\mathbf{-1324}$ & 0.650 & 0.122
& -661 & 0.782 & 0.381 \\
\hline \hline
{\model} & -170 & $\threestar{\mathbf{0.732}}$
& $\threestar{\mathbf{0.196}}$& -1347 & 0.652 & $\mathbf{0.245}$
& $\onestar{\mathbf{-625}}$ & $\onestar{\mathbf{0.804}}$ & $\onestar{\mathbf{0.392}}$ \\
\hline
\end{tabular}
\caption{\small Future network forecasting. {\model} performs best on NIPS and INFOCOM while results on DBLP are mixed.}
\label{tbl:futureprediction}
\vspace{-3mm}
\end{table}
Table~\ref{tbl:futureprediction} shows performance averaged over different $T_{obs}$ values ranging from 3 to $T$-1.
Overall, {\model} generally exhibits the best performance,
but performance results seem to depend on the dataset.
{\model} performs the best at 0.001 significance level in terms of AUC and F1 for the NIPS dataset, and at 0.05 level for the INFOCOM dataset.
While {\model} improves performance on AUC (9\%) and F1 (133\%), DRIFT achieves the best log-likelihood on the NIPS dataset.
In light of our previous observations, we conjecture that this is due to change in network edge density between different snapshots.
On the DBLP dataset, DRIFT gives the best log-likelihood, the naive model performs best in terms of AUC, and {\model} is the best on F1 score. However, in all cases of DBLP dataset, the differences are not statistically significant.
Overall, {\model} performs the best on NIPS and INFOCOM and provides \rev{comparable} performance on DBLP.
\begin{figure}
\centering
\begin{tabular}{ccc}
\includegraphics[width=0.32\textwidth]{FIG/Ring-F2} &
\includegraphics[width=0.32\textwidth]{FIG/Ring-F1} &
\includegraphics[width=0.32\textwidth]{FIG/Ring-F0} \\
(a) Group 1 & (b) Group 2 & (c) Group 3 \\
\end{tabular}
\caption{\small Group arrival and departure dynamics of different characters in the Lord of the Rings. Dark areas in the plots correspond to a give node's (y-axis) membership to each group over time (x-axis)}.
\label{fig:lotr}
\vspace{-5mm}
\end{figure}
\subsection{Task 3: Case study of ``The Lord of the Rings: The Two Towers'' social network}
Last, we also investigate groups identified by our model on a dynamic social network of characters in a \rev{movie, \textit{The Lord of the Rings: The Two Towers}}. Based on the transcript of the movie we created a dynamic social network on 21 characters and $T$=5 time epochs, where we connect a pair of characters if they co-appear inside some time window.
We fit our model to this network and examine the results in Figure~\ref{fig:lotr}. Our model identified three dynamic groups, which all nicely correspond to the Lord of the Rings storyline. For example, the core of Group 1 corresponds to Aragorn, elf Legolas, dwarf Gimli, and people in Rohan who in the end all fight against \rev{the} Orcs. Similarly, Group 2 corresponds to hobbits Sam, Frodo and Gollum on their mission to destroy the ring in Mordor, and are later joined by Faramir and ranger Madril.
Interestingly, Group 3 evolving around Merry and Pippin only forms at $t$=2 when they start their journey with Treebeard and later fight against wizard Saruman. While the fight occurs in two separate places we find that some scenes are not distinguishable, so it looks as if Merry and Pippin fought together with Rohan's army against Saruman's army.
\section{Sampling group memberships $Z$}
To sample node group membership $z_{ik}^{(t)}$, we use the forward-backward recursion algorithm~\cite{scott02} that samples the whole Markov chain $z_{ik}^{(1:T)}$ at once. Since we focus only on active groups, we only need to sample $z_{ik}^{(T_{k}^{B}:T_{k}^{D})}$ where $T_{k}^{B}$ and $T_{k}^{D}$ indicates the birth time and the death time of group $k$, respectively.
Suppose that all the other variables but $Z$ are given.
For the sample of each group membership $z_{ik}^{(t)}$, we use the forward-backward recursion algorithm~\cite{scott02} that sample the whole Markov chain $z_{ik}^{(1:T)}$ together.
Moreover, since the {\alive} groups are fixed, \ie, the birth time $T_{k}^{B}$ and death time $T_{k}^{D}$ of group $k$ is given, we only need to sample its sub-chain $z_{ik}^{(T_{k}^{B}:T_{k}^{D})}$.
The algorithm consists two passes: forward and backward passes.
In the forward pass, for each time $t$, we compute the posterior transition probability of $z_{ik}^{(\cdot)}$ from $t-1$ to $t$ given the links upto time $t$.
Once the forward pass is done, we sample the latent feature $z_{ik}^{(\cdot)}$ backward from $T_{k}^{D}$ to $T_{k}^{B}$, with consideration of the posterior transition probability computed in the forward pass.
To be concrete, let $\Omega$ be the states of all the other variables except for $z_{ik}^{(\cdot)}$.
For the forward pass, we define the following variables:
\begin{align}
P_{trs} = P \left(z_{ik}^{(t-1)} = r, z_{ik}^{(t)} = s | Y^{(T_{k}^{B}:t)}, \Omega \right), \quad
\pi_{ts} = P\left(z_{ik}^{(t)} = s | Y^{(T_{k}^{B}:t)}, \Omega \right) \,.
\end{align}
Then, we can find the value of each $P_{trs}$ and $\pi_{ts}$ by dynamic programming:
\begin{align}
\pi_{ts} = \sum_{r} P_{trs}, \quad
P_{trs} \propto \pi_{t-1, s} Q_{k}[r, s] P \left( Y^{(t)} | z_{ik}^{(t)} = s, \Omega \right)
\end{align}
where $Q_{k} = \left( \begin{array}{cc} 1 - a_{k} & a_{k} \\ 1 - b_{k} & b_{k} \end{array}\right)$ and $\sum_{r, s} P_{trs} = 1$.
Now given each $P_{trs}$ and $\pi_{ts}$, $z_{ik}^{T_{k}^{D}}$ can be sampled according to $\pi_{T_{k}^{D}}$, and then the backward pass samples the $z_{ik}^{\cdot}$ chain backwards:
\begin{align}
P\left(z_{ik}^{(t)} = r | z_{ik}^{(t+1)} = s, Y^{(T_{k}^{B}:T_{k}^{D})}, \Omega \right) \propto P_{(t+1)rs}\, .
\end{align}
\hide{
\clearpage
\begin{table}
\begin{tabular}{|c||c|c|c||c|c|c|}
\hline
\multirow{2}{*}{Model} & \multicolumn{3}{c||}{NIPS(40766)} & \multicolumn{3}{c|}{DBLP(86940)} \\
\cline{2-7}
& TestLL & AUC-ROC & F1 & TestLL & AUC-ROC & F1 \\
\hline
Baseline & -2030 $\pm$ 39 & 0.808 $\pm$ 0.025 & 0.177 $\pm$ 0.027
& -12051 $\pm$ 63 & 0.814 $\pm$ 0.008 & 0.300 $\pm$ 0.013 \\
\hline
LFRM & -880 $\pm$ 57 & 0.777 $\pm$ 0.026 & 0.195 $\pm$ 0.055
& -3783 $\pm$ 136 & 0.784 $\pm$ 0.010 & 0.146 $\pm$ 0.012 \\
\hline
DRIFT & -758 $\pm$ 58 & 0.866 $\pm$ 0.018 & 0.296 $\pm$ 0.027
& -3108 $\pm$ 150 & 0.916 $\pm$ 0.006 & 0.421 $\pm$ 0.016 \\
\hline
{\model} & -624 $\pm$ 53 & 0.916 $\pm$ 0.017 & 0.434 $\pm$ 0.060
& -2684 $\pm$ 152 & 0.939 $\pm$ 0.009 & 0.492 $\pm$ 0.022 \\
\hline
\multirow{2}{*}{Model} & \multicolumn{3}{c||}{Infocom(60372)} & \multicolumn{3}{c|}{Senator Vote (49500)} \\
\cline{2-7}
& TestLL & AUC-ROC & F1 & TestLL & AUC-ROC & F1 \\
\hline
Baseline & -17821 $\pm$ 122 & 0.677 $\pm$ 0.002 & 0.252 $\pm$ 0.004
& -30125 $\pm$ 115 & 0.716 $\pm$ 0.002 & 0.645 $\pm$ 0.003 \\
\hline
LFRM & -8689 $\pm$ 112 & 0.946 $\pm$ 0.002 & 0.703 $\pm$ 0.005
& -8143 $\pm$ 146 & 0.987 $\pm$ 0.001 & 0.926 $\pm$ 0.002 \\
\hline
DRIFT & -6654 $\pm$ 105 & 0.973 $\pm$ 0.001 & 0.757 $\pm$ 0.006
& -5091 $\pm$ 202 & 0.995 $\pm$ 0.001 & 0.955 $\pm$ 0.002 \\
\hline
{\model} & -6422 $\pm$ 130 & 0.976 $\pm$ 0.001 & 0.764 $\pm$ 0.006
& -9333 $\pm$ 694 & 0.984 $\pm$ 0.003 & 0.928 $\pm$ 0.007 \\
\hline
\end{tabular}
\caption{Missing Link Prediction}
\end{table}
\begin{figure}
\centering
\begin{tabular}{cc}
\includegraphics[width=0.45\textwidth]{FIG/NIPSFuture-LL} &
\includegraphics[width=0.45\textwidth]{FIG/DBLPFuture-LL} \\
(a) NIPS & (b) DBLP \\
\includegraphics[width=0.45\textwidth]{FIG/InfocomFuture-LL} &
\includegraphics[width=0.45\textwidth]{FIG/SVoteFuture-LL} \\
(c) Infocom & (d) Senator Vote
\end{tabular}
\caption{Forcast LL}
\end{figure}
\begin{figure}
\centering
\begin{tabular}{cc}
\includegraphics[width=0.45\textwidth]{FIG/NIPSFuture-ROC} &
\includegraphics[width=0.45\textwidth]{FIG/DBLPFuture-ROC} \\
(a) NIPS & (b) DBLP \\
\includegraphics[width=0.45\textwidth]{FIG/InfocomFuture-ROC} &
\includegraphics[width=0.45\textwidth]{FIG/SVoteFuture-ROC} \\
(c) Infocom & (d) Senator Vote
\end{tabular}
\caption{Forcast AUC-ROC}
\end{figure}
\begin{figure}
\centering
\begin{tabular}{cc}
\includegraphics[width=0.45\textwidth]{FIG/NIPSFuture-F1} &
\includegraphics[width=0.45\textwidth]{FIG/DBLPFuture-F1} \\
(a) NIPS & (b) DBLP \\
\includegraphics[width=0.45\textwidth]{FIG/InfocomFuture-F1} &
\includegraphics[width=0.45\textwidth]{FIG/SVoteFuture-F1} \\
(c) Infocom & (d) Senator Vote
\end{tabular}
\caption{Forcast F1}
\end{figure}
}
\section{Introduction}
\input{010intro}
\label{sec:intro}
\section{Models of Dynamic Networks}
\input{020dynamicmodel}
\label{sec:dynamicmodel}
\section{\fullmodel}
\input{030model}
\label{sec:model}
\section{Related Work}
\input{020related}
\label{sec:related}
\vspace{-2mm}
\section{Model Inference via MCMC}
\vspace{-1mm}
\label{sec:inference}
\input{040inference}
\section{Experiments}
\label{sec:experiments}
\input{050experiments}
\subsubsection*{Acknowledgments}
\rev{We thank Creighton Heaukulani and Zoubin Ghahramani for sharing data and code.}
This research has been supported in part by NSF
IIS-1016909,
CNS-1010921,
IIS-1149837,
IIS-1159679,
IARPA AFRL FA8650-10-C-7058,
Okawa Foundation,
Docomo,
Boeing,
Allyes,
Volkswagen,
Intel,
Alfred P. Sloan Fellowship and
the Microsoft Faculty Fellowship.
\subsubsection*{References}
{ \small
\bibliographystyle{abbrv}
\renewcommand{\refname}{\vspace{-5mm}}
|
\section{Introduction}
Modern data collection has allowed us to collect not just more data, but more complex data. In particular, complex objects like sets, distributions, and functions are becoming prevalent in many domains. It would be beneficial to perform machine learning tasks using these complex objects. However, many existing techniques can not handle complex, possibly infinite dimensional, objects; hence one often resorts to the ad-hoc technique of representing these complex object by arbitrary summary statistics.
In this paper, we look to perform a regression task when dealing with functional data. Specifically, we look to regress a mapping that takes in many functional input covariates and outputs a real value. Moreover, since we are considering many functional covariates (possibly many more than the number of instances of one's data), we look to find an estimator that performs feature selection by only regressing on a subset of all possible input functional covariates. To this end we present the Functional Shrinkage and Selection Operator (FuSSO), for performing sparse functional regression in a principled, semi-parametric manner.
Indeed, there are a multitude of applications and domains where the study of a mapping that takes in a functional input and outputs a real-value is of interest. That is, if $\mathcal{I}$ is some class of input functions with domain $\Psi \subseteq \mathbb{R}$ and range $\mathbb{R}$, then one may be interested in a mapping $h:\mathcal{I}\mapsto\mathbb{R}$: $h(f)=Y$ (Figure \ref{fig:simp}). Examples include: a mapping that takes in the time-series of a commodity's price in the past ($f$ is a function with the domain of time and range of price) and outputs the expected price of the commodity in the nearby future; also, a mapping that takes a patient's cardiac monitor's time-series and outputs a health index. Recently, work by \cite{poczos2012distfree} has explored this type of regression problem when the input function is a distribution. Furthermore, the general case of an arbitrary functional input is related to functional analysis \cite{ferraty2006nonparametric}.
However, often it is expected that the response one is interested in regressing is dependent on not just one, but many functions. That is, it may be fruitful to consider a mapping $h:\mathcal{I}_1\times\ldots\times\mathcal{I}_p\mapsto\mathbb{R}$: $h(f_1,\ldots,f_p)=Y$ (Figure \ref{fig:mult}). For instance, this is likely the case in regressing the price of a commodity in the future, since the commodity's future price is not only dependent on the history of it own price, but also the history of other commodities' prices as well. A response's dependence on multiple functional covariates is especially common in neurological data, where thousands of voxels in the brain may each contain a corresponding function. In fact, in such domains it is not uncommon to have a number of input functional covariates that far exceeds the number of training instances one has in a data-set. Thus, it would be beneficial to have an estimator that is sparse in the number of functional covariates used to regress the response against. That is, find an estimate, $\hat{h}_s$, that depends on a small subset $\{i_1,\ldots,i_{S}\} \subset \{1,\ldots,p\}$, such that $\hat{h}(f_1,\ldots,f_p)=\hat{h}_s(f_{i_1},\ldots,f_{i_{S}})$ (Figure \ref{fig:sparse}).
\begin{figure*}[t!]
\centering
\subfigure[Single Functional Covariate]{\includegraphics[width=.15\textwidth]{simple-model.pdf}\label{fig:simp}}\quad\quad\quad
\subfigure[Multiple Functional Covariates]{\includegraphics[width=.2\textwidth]{full-model.pdf}\label{fig:mult}}\quad\quad\quad
\subfigure[Sparse Model]{\includegraphics[width=.2\textwidth]{sparse-model.pdf}\label{fig:sparse}}
\caption{(a) Model where mapping takes in a function $f$ and produces a real $Y$. (b) Model where response $Y$ is dependent on multiple input functions $f_{1},\ldots,f_{p}$. (c) Sparse model where response $Y$ is dependent on a sparse subset of input functions $f_{1},\ldots,f_{p}$.}
\vspace{-0.2cm}
\end{figure*}
Here we present a semi-parametric estimator to perform sparse regression with multiple input functional covariates and a real-valued response, the FuSSO: Functional Shrinkage and Selection Operator. No parametric assumptions are made on the nature of input functions. We shall assume that the response is the result of a sparse set of linear combinations of input functions and other non-paramteric functions $\{g_i\}$: $Y = \sum_j \langle f_j, g_j \rangle$. The resulting method is a LASSO-like \cite{tibshirani1996regression} estimator that effectively zeros out entire functions from consideration in regressing the response.
Our contributions are as follows. We introduce the FuSSO, an estimator for performing regression with many functional covariates and a real-valued response. Furthermore, we provide a theoretical backing of the FuSSO estimator via proof of asymptotic sparsistency under certain conditions. We also illustrate the estimator with applications on synthetic data as well as in regressing the age of a subject when given orientation distribution function (dODF) \cite{yeh2011ntu} data for the subject's white matter.
\section{Related Work}
As previously mentioned, recently \cite{poczos2012distfree} explored regression with a mapping that takes in a probability density function and outputs a real value. Furthermore, \cite{olivadistribution} studies the case when both the input and outputs are distributions. In addition, functional analysis relates to the study for functional data \cite{ferraty2006nonparametric}. In all these works, the mappings studied take in only one functional covariate. Based on them, it is not immediately evident how to expand on these ideas to develop an estimator that simultaneously performs regression and feature selection with multiple function covariates.
To the best of our knowledge, there has been no prior work in studying sparse mappings that take multiple functional inputs and produce a real-valued output. LASSO-like regression estimators that work with functional data include the following. In \cite{mingotti2013lasso}, one has a functional output and several real-valued covariates. Here, the estimator finds a sparse set of functions to scale by the real valued covariates to produce the functional response. Also, \cite{zhao2012wavelet,james2009functional} study the case when one has one functional covariate $f$ and one real valued response that is linearly dependent on $f$ and some function $g$: $Y = \langle f,g\rangle = \int f g$. In \cite{zhao2012wavelet} the estimator searches for sparsity across wavelet basis projection coefficients. In \cite{james2009functional}, sparsity is in achieved in the time (input) domain of the $d^{\mathrm{th}}$ derivative of $g$; i.e. $[D^dg](t)=0$ for many values of $t$ where $D^d$ is the differential operator. Hence, roughly speaking, \cite{zhao2012wavelet,james2009functional} look for sparsity across frequency and time domains respectively, for the regressing function $g$. However, these methods do not consider the case where one has many input functional covariates $\{f_1,\ldots,f_p\}$, and needs to choose among them. That is, \cite{zhao2012wavelet,james2009functional} do not provide a method to select among function covariates in an analogous fashion to how the LASSO selects among real-valued covariates.
Lastly, it is worth noting that in our estimator we will have an additive linear model, $\sum_j \langle f_j, g_j \rangle$ where we search for $\{g_i\}$ in a broad, non-parametric family such that many $g_j$ are the zero function. Such a task is similar in nature to the SpAM estimator \cite{ravikumar2009sparse}, in which one also has an additive model $\sum_j g_j(X_j)$ (in the dimensions of a real vector $X$) and searches for $\{g_i\}$ in a broad, non-parametric family such that many $g_j$ are the zero function. Note though, that in the SpAM model, the $\{g_i\}$ functions are applied to real covariates via a function evaluation. In the FuSSO model, $\{g_i\}$ are applied to functional covariates via an inner product; that is, FuSSO works over functional, not real-valued covariates, unlike SpAM.
\section{Model}
To better understand FuSSO's model we draw several analogies to real-valued linear regression and Group-LASSO \cite{yuan2006model}. Note that although for simplicity we focus on functions working over a one dimensional domain, it is straightforward to extend the estimator and results to the multidimensional case. Consider a model for typical real-valued linear regression with a data-set of input-output pairs $\{(X_i,Y_i) \}_{i=1}^{N}$:
\begin{align*}
Y_i = \langle X_i, w \rangle + \epsilon_i,
\end{align*}
where $Y_i \in \mathbb{R}$,$\ X_i \in \mathbb{R}^d$, $w\in\mathbb{R}^d,\ \epsilon_i \overset{iid}{\sim} \mathcal{N}(0,\sigma^2)$, and $\langle X_i, w \rangle=\sum_{j=1}^d X_{ij}w_{j}$. If instead one were working with functional data $\{(f^{(i)},Y_i)\}_{i=1}^{N}$, where $f^{(i)}:[0,1] \mapsto \mathbb{R}$ and $f^{(i)} \in L_2[0,1]$, one may similarly consider a linear model:
\begin{align*}
Y_i = \langle f^{(i)}, g \rangle + \epsilon_i,
\end{align*}
where, $g:[0,1] \mapsto \mathbb{R}$, and $\langle f^{(i)}, g \rangle=\int_0^1 f^{(i)}(t)g(t) \mathrm{d} t$. If $\Phi = \{\varphi_m\}_{m=1}^{\infty}$ is an orthonormal basis for $L_2[0,1]$ \cite{tsybakov2008introduction} then we have that
\begin{align}
f^{(i)}(x)=\sum_{m=1}^{\infty}\alpha^{(i)}_{m}\varphi_m(x), \label{eq:projcoef}
\end{align}
where, $\alpha^{(i)}_{m} =\int_0^1 f^{(i)}(t)\varphi_m(t)\mathrm{d} t$. Similarly, $g(x)=\sum_{m=1}^{\infty}\beta_{m}\varphi_m(x)$ where $\beta_{m} =\int_0^1 g(t)\varphi_m(t)\mathrm{d} t$. Thus,
\begin{align*}
Y_i &= \langle f^{(i)}, g \rangle + \epsilon_i \\
&= \langle \sum_{m=1}^{\infty}\alpha^{(i)}_{m}\varphi_m(x), \sum_{k=1}^{\infty}\beta_{k}\varphi_k(x) \rangle + \epsilon_i \\
&= \sum_{m=1}^{\infty}\sum_{k=1}^{\infty} \alpha^{(i)}_{m}\beta_{k} \langle \varphi_m(x), \varphi_k(x) \rangle + \epsilon_i \\
&= \sum_{m=1}^{\infty} \alpha^{(i)}_{m}\beta_{m} + \epsilon_i,
\end{align*}
where the last step follows from orthonormality of $\Phi$.
Going back to the real-valued covariate case, if instead of having one feature vector per data instance: $X_i \in \mathbb{R}^d$, one had $p$ feature vectors associated to each data instance: $\{X_{ij}\ |\ 1\leq j \leq p,\ X_{ij} \in \mathbb{R}^d \}$, an additive linear model may be used for regression:
\begin{align*}
Y_i = \sum_{j=1}^p\langle X_{ij}, w_j \rangle + \epsilon_i, \mathrm{where}\ w_1,\ldots,w_p \in \mathbb{R}^d.
\end{align*}
Similarly, in the functional case one may have $p$ functions associated with data instance $i$: $\{f^{(i)}_{j}\ |\ \ 1\leq j \leq p,\ f^{(i)}_{j}\in L_2[0,1] \}$. Then, an additive linear model would be:
\begin{align}
Y_i &= \sum_{j=1}^p\langle f^{(i)}_j, g_j \rangle + \epsilon_i = \sum_{j=1}^p\sum_{m=1}^{\infty} \alpha^{(i)}_{jm}\beta_{jm} + \epsilon_i \label{eq:fgmodel},
\end{align}
where $g_1,\ldots,g_p \in L_2[0,1]$, and $\alpha^{(i)}_{jm}$ and $\beta_{jm}$ are projection coefficients for $f^{(i)}_j$ and $g_j$ respectively.
Suppose that one has few observations relative to the number of features ($N\ll p$). In the real-valued case, in order to effectively find a solution for $ w = (w_1^{T},\ldots,w_p^{T})^T$ one may search for a group sparse solution where many $w_j=0$. To do so, one may consider the following Group-LASSO regression:
\begin{align}
w^\star = \argmin_{w} {\frac{1}{2N}\norm{Y-\sum_{j=1}^p X_j w_j}^2 + \lambda_{\scalebox{.6}{N}} \sum_{j=1}^p \|w_j\| },
\label{eq:grplasso}
\end{align}
where $X_j$ is the $ N \times d $ matrix $X_j = [X_{1j} \ldots X_{Nj}]^T$, $Y=(Y_1,\ldots,Y_N)^T$, and $\|\cdot\|$ is the Euclidean norm.
If in the functional case (\ref{eq:fgmodel}) one also has that $N\ll p$, one may set up a similar optimization to (\ref{eq:grplasso}), whose direct analogue is:
\begin{align}
g^\star = \argmin_{g}& \frac{1}{2N} \sum_{i=1}^N \left(Y_i - \sum_{j=1}^p \langle f^{(i)}_j, g_j \rangle \right)^2 \\
&+ \lambda_{\scalebox{.6}{N}} \sum_{j=1}^p \norm{g_j} \label{eq:fgrplasso};
\end{align}
equivalently,
\begin{align}
\beta^\star = \argmin_{\beta}& \frac{1}{2N} \sum_{i=1}^N \left(Y_i - \sum_{j=1}^p\sum_{m=1}^{\infty} \alpha^{(i)}_{jm}\beta_{jm} \right)^2 \\
&+ \lambda_{\scalebox{.6}{N}} \sum_{j=1}^p \sqrt{\sum_{m=1}^{\infty} \beta_{jm}^2}, \label{eq:bgrplasso}
\end{align}
where $g = \{g_i\}_{i=1}^p =\{\sum_{m=1}^{\infty} \beta_{im}\varphi_m ,\}_{i=1}^p $.
However, it is unfeasible to directly observe functional inputs $\{f^{(i)}_j\ |\ 1\leq i \leq N, 1\leq j \leq p \}$ . Thus, we shall instead assume that one observes $\{\vec{y}^{\,(i)}_j\ |\ 1\leq i \leq N, 1\leq j \leq p \}$ where
\begin{align}
\vec{y}^{\,(i)}_j =& \vec{f}^{\,(i)}_j + \xi^{(i)}_j, \label{eq:fvec}\\
\vec{f}^{\,(i)}_j =& \left(f^{(i)}_j(1/n),\ f^{(i)}_j(2/n),\ \ldots,\ f^{(i)}_j(1)\right)^T, \\
\xi^{(i)}_j \overset{iid}{\sim} &\mathcal{N}(0,\sigma_\xi^2I_n).
\end{align}
That is, we observe a grid of $n$ noisy values for each functional input. Then, one may estimate $\alpha^{(i)}_{jm}$ as:
\begin{align}
\tilde{\alpha}^{(i)}_{jm} = \frac{1}{n} \vec{\varphi}_m^T \vec{y}^{\,(i)}_j = \frac{1}{n} \vec{\varphi}_m^T (\vec{f}^{\,(i)}_j + \xi^{(i)}_j ) = \bar{\alpha}^{(i)}_{jm} + \eta^{(i)}_{jm} \label{eq:talpha}
\end{align}
where $\vec{\varphi}_m = \left(\varphi_m(1/n),\ \varphi_m(2/n),\ \ldots,\ \varphi_m(1)\right)^T$. Furthermore, we may truncate the number of basis functions used to express ${f}^{\,(i)}_j$ to $M_n$, estimating it as:
\begin{align}
\tilde{f}^{(i)}_j(x) = \sum_{m=1}^{M_n} \tilde{\alpha}^{(i)}_{jm} \varphi_m(x). \label{eq:trunc}
\end{align}
Using the truncated estimate (\ref{eq:trunc}), one has:
\begin{align*}
\langle \tilde{f}^{(i)}_j(x), g_j \rangle =& \sum_{m=1}^{M_n} \tilde{\alpha}^{(i)}_{jm} \beta_{jm}, \mathrm{and}\ \\
\norm{\tilde{f}^{(i)}_j(x)} =& \sqrt{\sum_{m=1}^{M_n} (\tilde{\alpha}^{(i)}_{jm})^2}.
\end{align*}
Hence, using the approximations (\ref{eq:trunc}), (\ref{eq:bgrplasso}) becomes:
\begin{align}
\hat{\beta} = \argmin_{\beta}& \frac{1}{2N} \sum_{i=1}^N \left(Y_i - \sum_{j=1}^p\sum_{m=1}^{M_n} \tilde{\alpha}^{(i)}_{jm}\beta_{jm} \right)^2 \label{eq:MSElasso}\\
&+ \lambda_{\scalebox{.6}{N}} \sum_{j=1}^p \sqrt{\sum_{m=1}^{M_n} \beta_{jm}^2} \label{eq:normlasso}\\
= \argmin_{\beta}& \frac{1}{2N} \norm{Y - \sum_{j=1}^p\tilde{A}_{j}\beta_{j} }^2 + \lambda_{\scalebox{.6}{N}} \sum_{j=1}^p \norm{\beta_{j}} \label{eq:btrunkgrplasso},
\end{align}
where $\tilde{A}_{j}$ is the $N \times M_n$ matrix with values $\tilde{A}_{j}(i,m) = \tilde{\alpha}^{(i)}_{jm}$ and $\beta_j = (\beta_{j1},\ldots,\beta_{jM_n})^T$. Note that one need not consider projection coefficients $\beta_{jm}$ for $m>M_n$ since such projection coefficients will not decrease the MSE term in \eqref{eq:MSElasso} (because $\tilde{\alpha}^{(i)}_{jm}=0$ for $m>M_n$), and $\beta_{jm}\neq0$ for $m>M_n$ increases the norm penalty term in (\ref{eq:normlasso}). Hence we see that our sparse functional estimates are a Group-LASSO problem on the projection coefficients.
\paragraph{Extensions} It is useful to note that there are several straightforward extensions to the FuSSO as presented. First, we would like to note that it may be possible to estimate the inner product of a function $f^{(i)}_j$, and $g_j$ as $\int f^{(i)}_j g_j \approx \langle \vec{y}^{\,(i)}_j, \tfrac{1}{n}\vec{g}_j \rangle$, where $\vec{g}_j = (g_j(1/n),\ldots,g_j(1))^T$. This effectively allows one to use a naive approach of simply using Group-LASSO on the $\vec{y}^{\,(i)}_j$ feature vectors directly (we'll refer to this method as Y-GL). It is important to note, however, that Y-GL will be less robust to noise, and adaptive (and efficient) to smoothness than the FuSSO. Furthermore, we note that it is not necessary to have observations for input functions that are on a grid for the FuSSO, since one may estimate projection coefficients in the case of an irregular design \cite{tsybakov2008introduction}. Moreover, we may also estimate projection coeffincients for density functions with samples drawn from the pdf. Note that the Y-GL would fail to estimate our model in the irregular design case, and would not be possible in the case were functions are pdf. Also, a two-stage estimator as described in \cite{meinshausen2007relaxed}, where one first uses the regularization penalty with a large $\lambda$ to find the support, then solves the optimization problem with a smaller $\lambda$ on just the estimated support to estimate the response, may be more efficient at estimating the response. Furthermore, an analogous problem as (\ref{eq:btrunkgrplasso}) may be framed to perform logistic regression and classification.
\section{Theory}
Next, we show that the FuSSO is able to recover the correct sparsity pattern asymptotically; i.e., that the FuSSO estimate is sparsistent. In order to do so, we shall show that with high probability there is an optimal solution to our optimization problem (\ref{eq:btrunkgrplasso}) with the correct sparsity pattern. We follow a similar argument to \cite{wainwright2006sharp, ravikumar2009sparse}. We shall use a ``witness" technique to show that there is a coefficient/subgradient pair $(\hat{\beta},\hat{u})$ such that $\supp(\hat{\beta})=\supp(\beta^*)$, for true response generating $\beta^{*}$. Let $\Omega(\beta) = \sum_{j=1}^{p} \norm{\beta_j}_2$, be our penalty term (\ref{eq:normlasso}). Let $S$ denote the true set of non-zero functions; i.e.\ $S = \{ j\ |\ \beta^*_j\neq0 \}$, with $s = |S|$. First, we fix $\hat{\beta}_{S^c}=0$, and set $\hat{u}_{S} = \partial\Omega(\cdot)(\beta^*)_{S}$. Note that for a vector $\beta'$, $\partial\Omega(\cdot)(\beta') = \{ u \}$ where: $u_j = \beta'_j/\norm{\beta'}_2$, if $\beta'_j\neq0$; $u_j =\norm{u_j}_2 \leq 1$ if $\beta'_j=0$.
We shall show that with high probability, $\forall j \in S,\ \hat{\beta}_{j}\neq0$ and $\forall j \in S^c,\ \norm{u_j}_2 < 1$, thus showing that there is an optimal solution to our optimization problem (\ref{eq:btrunkgrplasso}) that has the true sparsity pattern with high probability.
First, we elaborate on our assumptions.
\subsection{Assumptions}
Let $\Phi$ be the trigonometric basis, $\varphi_1(x) \equiv 1$, $k\geq2$:
\begin{align*}
\varphi_{2k}(x) \equiv \sqrt{2}\cos(2\pi k x),\
\varphi_{2k+1}(x) \equiv \sqrt{2}\sin(2\pi k x) .
\end{align*}
Let $\mathcal{D}= \{ (\{\vec{y}^{\,(i)}_j\}_{j=1}^p, Y_i)\}_{i=1}^N$, where $\vec{y}^{\,(i)}_j$ is as (\ref{eq:fvec}), and $Y_i = \sum_{j=1}^p\sum_{m=1}^{\infty} \alpha^{(i)}_{jm}\beta^*_{jm} + \epsilon_i$ as in (\ref{eq:fgmodel}). Assume that $\forall\ 1\leq i \leq N,\ 1\leq j \leq p$: $\alpha^{(i)}_j\in \Theta(\gamma,Q)$, where:
\begin{align*}
\Theta(\gamma,Q) =& \{ \theta\ :\ \sum_{k=1}^\infty c_k^{2} \theta_k^2 \leq Q \},\\
c_k =& k^\gamma\ \mathrm{if}\ k\ \mathrm{even\ or\ one},\ (k-1)^\gamma\ \mathrm{otherwise},\\
\alpha^{(i)}_j =& \{\alpha^{(i)}_{jm} \in \mathbb{R}\ |\ \alpha^{(i)}_{jm} = \int_0^1 f^{(i)}_j \varphi_m,\ m\in\mathbb{N}^{+} \}
\end{align*}
for $0<Q<\infty$ and $\tfrac{1}{2}<\gamma<\infty$. Furthermore, assume that that for the true $\beta$ generating the observed responses $Y_i$, $\beta^{*}$, $\forall\ 1\leq j \leq p$: $\beta^{*}_j \in \Theta(\gamma,Q)$.
Let $A_j$ be the $N \times M_n$ matrix with entries $A_{j}(i,m) = \alpha^{(i)}_{jm}$. Let $A_S$ denote the matrix made up from horizontally concatenating the $A_j$ matrices with $j\in S$; i.e.\ $A_S = [A_{j_1}\ldots A_{j_s}]$, where $\{j_1,\ldots,j_s\}=S$ and $j_i<j_k$ for $i<k$. Suppose the following:
\begin{align}
\Lambda_{\max} \left(\tfrac{1}{N}A_S^TA_S\right) \leq C_{\max} <& \infty \label{eq:maxeig} \\
\Lambda_{\min} \left(\tfrac{1}{N}A_S^TA_S\right) \geq C_{\min} >& 0 \label{eq:mineig}.
\end{align}
Also, suppose $\exists\delta\in(0,1]$ s.t. $\forall j\in S^c$
\begin{align}
\Lambda_{\max} \left(\tfrac{1}{N}A_j^TA_j\right) \leq C_{\max} <& \infty \\
\Norm{(\tfrac{1}{N}A_{j}^TA_S)(\tfrac{1}{N}A_S^TA_S)^{-1}}_2 \leq& {1-\delta}/{\sqrt{s}}
\end{align}
Let $\bar{A}_j$ be the $N \times M_n$ matrix with entries $\bar{A}_j(i,m) = \bar{\alpha}^{(i)}_{jm} = \frac{1}{n} \vec{\varphi}_m^T \vec{f}^{\,(i)}_j $. Let $H_j$ be the $N \times M_n$ matrix with entries $H_{j}(i,m) = \eta^{(i)}_{jm} = \frac{1}{n} \vec{\varphi}_m^T \xi^{(i)}_j $. Thus, $\tilde{A}_j = \bar{A}_j + H_j$. Furthermore, let $E_j = \bar{A}_j - A_j$. Then, $\tilde{A}_j =A_j +E_j+ H_j$.
In addition to the aforementioned assumptions, we shall further assume the following:
\begin{gather}
\exists a<1/2 \quad \mathrm{s.t.}\quad pM_n n^{a - \sfrac{1}{2}} e^{-n^{1-2a}} \rightarrow 0 \label{eq:bndH_S}\\
\rho^*_N \equiv \min_{j\in S}\norm{\beta^*_j}_\infty>0 \label{eq:supportaway}\\
\sqrt{s M_n} \left(n^{-\gamma +\sfrac{1}{2}}+ n^{-a} \right) \rightarrow 0 \\
\frac{1}{\rho^*_N}\left( s^{\sfrac{3}{2}}M_n^{\sfrac{1}{2}-2\gamma} + \sqrt{{\log(sM_n)}/{N}} \right) \rightarrow 0 \\
\lambda_{\scalebox{.6}{N}} \sqrt{sM_n}/\rho^*_N \rightarrow 0 \label{eq:lambdaroot}\\
\frac{1}{\lambda_{\scalebox{.6}{N}}}\left({s\sqrt{M_n}} n^{-\gamma +1/2} + \sqrt{\tfrac{s\log(N)}{n}} \right) \rightarrow 0 \\
\frac{1}{\lambda_{\scalebox{.6}{N}}} \left( \frac{s M_n}{ n^{\gamma+a -\sfrac{1}{2}}} + \frac{\sqrt{sM_n\log(N)}}{ n^{a +\sfrac{1}{2}}} \right) \rightarrow 0\\
\frac{1}{\lambda_{\scalebox{.6}{N}}} \sqrt{{M_n\log((p-s)M_n)}/{N}} \rightarrow 0 \\
{s}/{(\lambda_{\scalebox{.6}{N}} N M_n^{2\gamma-\sfrac{1}{2}})} \rightarrow 0 \label{eq:l-ass},
\end{gather}
and we assume $\gamma \geq 1$ for the sake of simplification. We may further simplify our assumptions if we take $n=N^{1/2}$ and choose $M_n$ optimally for function estimation: $M_n\asymp n^{1/(2\gamma +1)}=N^{1/(4\gamma +2)}$. Furthermore, take $s = O(1)$, $\rho^*_N \asymp 1$, and $\gamma=2$. Under these conditions, our assumptions reduce to $\frac{1}{10}<a$ and taking the follow to go to zero:
\begin{gather*}
pN^{\frac{10a-3}{20}}e^{-N^{\frac{1}{2}-a}},
\lambda_{\scalebox{.6}{N}} N^{\sfrac{1}{20}}, N^{-\frac{7}{10}}/\lambda_{\scalebox{.6}{N}},\\
\tfrac{1}{\lambda_{\scalebox{.6}{N}}^2} N^{\frac{1}{2}}\log(N),\tfrac{1}{\lambda_{\scalebox{.6}{N}}^2} N^{-\sfrac{9}{10}}\log(pN) .
\end{gather*}
\subsection{Sparsistency}
\thmfirst{thm:sparsistent}: $\mathbb{P}\left(\hat{S}_N=S\right) \rightarrow 1$.
First, we state some lemmas, whose proofs may be found in the supplementary materials.
\subsubsection{Lemmata}
\lemmafirst{thm:exp_bnd} Let $X$ be a non-negative r.v. and $\mathcal{C}$ be an measurable event, then $\EE{X | \mathcal{C} }\mathbb{P}(\mathcal{C}) \leq \EE{X} $.
\lemmafirst{thm:sumkrondelt} $\frac{1}{n}\sum_{k=1}^n \varphi_{m}(k/n)\varphi_{l}(k/n) = \mathbb{I}\{l=m\}$, for $1\leq l,m \leq n-1$.
\lemmafirst{thm:noisedist} Let $H_{j}^{(i)}$ be the rows of $H_{j}$, then $H_{j}^{(i)} \overset{iid}{\sim} \mathcal{N}(0,\frac{\sigma_\xi^2}{n}I)$, and $H_{S}^{(i)} \overset{iid}{\sim} \mathcal{N}(0,\frac{\sigma_\xi^2}{n}I)$.
\lemmafirst{thm:bndHmax} $\mathbb{P}\left(\norm{H}_{\max} \geq n^a \right) \leq 2 \sigma_\xi pM_n n^{a - \sfrac{1}{2}} e^{-\frac{n^{1-2a}}{2\sigma_\xi^2}}$
\lemmafirst{thm:biasmax} $\norm{E_j}_{\max}\leq C_Q n^{-\gamma+1/2}$, where $C_Q\in(0,\infty)$ is a constant depending on $Q$.
\lemmafirst{thm:beta} $\norm{\beta^{*}_S}_2^2 \leq Qs$.
\lemmafirst{thm:noiseeigs} $\exists N_0,n_0,\tilde{C}_{\min}, \tilde{C}_{\max}$, $ 0<\tilde{C}_{\min}\leq \tilde{C}_{\max} < \infty$, $0<\tilde{\delta}\leq1$ s.t. if $\norm{H}_{\max}<n^{-a}$, and $N>N_0$, $n>n_0$ then
\begin{align}
\Lambda_{\max} \left(\tfrac{1}{N}\tilde{A}_S^T\tilde{A}_S\right) \leq \tilde{C}_{\max} <& \infty \\
\Lambda_{\min} \left(\tfrac{1}{N}\tilde{A}_S^T\tilde{A}_S\right) \geq \tilde{C}_{\min} >& 0\\
\forall j \in S^c,\ \Norm{(\tfrac{1}{N}\tilde{A}_{j}^T\tilde{A}_S)(\tfrac{1}{N}\tilde{A}_S^T\tilde{A}_S)^{-1}}_2 \leq& \frac{1-\tilde{\delta}}{\sqrt{s}}
\end{align}
\subsubsection{Proof of Theorem 1}
\propfirst{thm:infsupp} $\mathbb{P}\left(\forall j\in S\ \hat{\beta}_j\neq0\right) \rightarrow 1$.
\begin{proof}
Recall that by (\ref{eq:supportaway}), $\rho^*_N = \min_{j\in S}\norm{\beta^*_j}_\infty>0$. Thus to prove that $\forall j\in S\ \hat{\beta}_j\neq0$, it suffices to show that :
$\norm{\hat{\beta}_S-\beta^{*}_S}_\infty \leq \frac{\rho^*_N}{2}$.
To do so we show $\mathbb{P}\left(\norm{\hat{\beta}_S-\beta^{*}_S}_\infty > \frac{\rho^*_N}{2}\right) \rightarrow 0$. Let $\mathcal{B}$ be the event that $\norm{H}_{\max}<n^{-a}$. Note that:
\begin{align*}
&\mathbb{P}\left(\norm{\hat{\beta}_S-\beta^{*}_S}_\infty > \frac{\rho^*_N}{2}\right)\\
&\leq \mathbb{P}\left(\norm{\hat{\beta}_S-\beta^{*}_S}_\infty > \frac{\rho^*_N}{2} \Big| \mathcal{B} \right) \mathbb{P}(\mathcal{B}) +\mathbb{P}\left(\mathcal{B}^c \right).
\end{align*}
Furthermore,
\begin{align*}
&\mathbb{P}\left(\norm{\hat{\beta}_S-\beta^{*}_S}_\infty > \frac{\rho^*_N}{2} \Big| \mathcal{B} \right) \leq \tfrac{2}{\rho^*_N}\mathbb{E}\left[\norm{\hat{\beta}_S-\beta^{*}_S}_\infty \Big| \mathcal{B}\right].
\end{align*}
Then, looking at the stationarity condition for the support $S$:
\begin{align}
\tfrac{1}{N}\tilde{A}_S^T\left(\tilde{A}_S\hat{\beta}_S-Y\right)+\lambda_{\scalebox{.6}{N}}\hat{u}_S=0. \label{eq:supp-station}
\end{align}
Let $V$ be the $N \times 1$ vector with entries $V_i= \sum_{j\in S}\sum_{m=M_n+1}^{\infty}\alpha^{(i)}_{jm}\beta^{*}_{jm}$; i.e.\ the error from truncation. Then, using \eqref{eq:supp-station} $Y=A_S\beta^{*}_S +V+\epsilon \implies$
\begin{align*}
&\tfrac{1}{N}\tilde{A}_S^T\left(\tilde{A}_S\hat{\beta}_S-A_S\beta^{*}_S -V-\epsilon\right)+\lambda_{\scalebox{.6}{N}}\hat{u}_S=0 \implies \\
&\tfrac{\tilde{A}_S^T}{N}\left(\tilde{A}_S(\hat{\beta}_S-\beta^{*}_S)-(A_S-\tilde{A}_S)\beta^{*}_S -V-\epsilon\right)=
-\lambda_{\scalebox{.6}{N}}\hat{u}_S
\end{align*}
Thus,
\begin{align}
\tfrac{1}{N}\tilde{A}_S^T\tilde{A}_S(\hat{\beta}_S-\beta^{*}_S)= &-\tfrac{1}{N}\tilde{A}_S^T(E_S+H_S)\beta^{*}_S +\tfrac{1}{N}\tilde{A}_S^TV \nonumber \\
&+\tfrac{1}{N}\tilde{A}_S^T\epsilon-\lambda_{\scalebox{.6}{N}}\hat{u}_S \label{eq:supp-diff}.
\end{align}
Let $\tilde{\Sigma}_{SS}^{-1}=(\tfrac{1}{N}\tilde{A}_S^T\tilde{A}_S)^{-1}$; we see that,
\begin{align*}
\norm{\hat{\beta}_S-\beta^{*}_S}_\infty & \leq \norm{\tilde{\Sigma}_{SS}^{-1}(\tfrac{1}{N}\tilde{A}_S^T)(E_S+H_S)\beta^{*}_S}_\infty \\
& +\norm{\tilde{\Sigma}_{SS}^{-1}(\tfrac{1}{N}\tilde{A}_S^T)V}_\infty +\norm{\tilde{\Sigma}_{SS}^{-1}(\tfrac{1}{N}\tilde{A}_S^T)\epsilon}_\infty\\
& +\norm{\tilde{\Sigma}_{SS}^{-1}\lambda_{\scalebox{.6}{N}}\hat{u}_S}_\infty.
\end{align*}
Thus, we proceed to bound each term on the LHS in expectation. First, note that $\norm{\tilde{\Sigma}_{SS}^{-1}(\tfrac{1}{N}\tilde{A}_S^T)(E_S+H_S)\beta^{*}_S}_\infty
\leq \norm{\tilde{\Sigma}_{SS}^{-1}}_\infty\norm{(\tfrac{1}{N}\tilde{A}_S^T)(E_S+H_S)\beta^{*}_S}_\infty
= \norm{\tilde{\Sigma}_{SS}^{-1}}_\infty \norm{\tfrac{1}{N}(A_S^T +E_S^T+ H_S^T)(E_S+H_S)\beta^{*}_S}_\infty
\leq \frac{\norm{\tilde{\Sigma}_{SS}^{-1}}_\infty}{N} \Big( \norm{A_S^T(E_S+H_S)\beta^{*}_S}_\infty
+ \norm{(E_S+H_S)^T(E_S+H_S)\beta^{*}_S}_\infty \Big)$.
Moreover, given that $\mathcal{B}$ occurs:
\begin{align*}
\norm{\tilde{\Sigma}_{SS}^{-1}}_\infty \leq \sqrt{sM_n} \norm{\tilde{\Sigma}_{SS}^{-1}}_2 \leq \frac{\sqrt{sM_n} }{\tilde{C}_{\min}}.
\end{align*}
Thus, $\mathbb{E}\left[ \tfrac{\norm{\tilde{\Sigma}_{SS}^{-1}}_\infty}{N} \norm{A_S^T(E_S+H_S)\beta^{*}_S}_\infty \big| \mathcal{B} \right]\mathbb{P}(\mathcal{B})$
\begin{align*}
&\leq \frac{\sqrt{sM_n} }{\tilde{C}_{\min}N} \norm{A_S^T}_\infty \mathbb{E}\left[ \norm{(E_S+H_S)\beta^{*}_S}_\infty \big| \mathcal{B} \right] \mathbb{P}(\mathcal{B})\\
&\leq \frac{Q\sqrt{sM_n} }{\tilde{C}_{\min}} \left(\norm{E_S\beta^{*}_S}_\infty + \mathbb{E}\left[ \norm{H_S\beta^{*}_S}_\infty \big| \mathcal{B} \right]\mathbb{P}(\mathcal{B}) \right),
\end{align*}
noting that $\norm{A_S^T}_{\infty}\leq NQ$. Moreover, by Lemma \ref{thm:exp_bnd}:
\begin{align*}
\mathbb{E}\left[ \norm{H_S\beta^{*}_S}_\infty \big| \mathcal{B} \right]\mathbb{P}(\mathcal{B}) \leq \mathbb{E}\left[ \norm{H_S\beta^{*}_S}_\infty \right].
\end{align*}
Also, $H_S\beta^{*}_S$ is normally distributed and $\Var[H^{(i)T}_S\beta^{*}_S]$
\begin{align*}
= \sum_{j=1}^{sM_n} \Var[H^{(i)}_{Sj}\beta^{*}_{Sj}] = \frac{\sigma_\xi^2}{n} \norm{\beta^{*}_{S}}_2^2
\leq \frac{\sigma_\xi^2Q s}{n} .
\end{align*}
Hence, by a Gaussian inequality (e.g. \cite{wasserman10705}) we have:
\begin{align*}
\mathbb{E}\left[ \norm{H_S\beta^{*}_S}_\infty \right] \leq \sqrt{{2\sigma_\xi^2Q s \log(N)}/{n} }.
\end{align*}
Unless otherwise specified, let $X^{(i)}$ be the $\jth{i}$ row of matrix $X$ and $X_j$ be the $\jth{j}$ column. Also,
\begin{align*}
\norm{E_S\beta^{*}_S}_\infty &= \max_{1\leq i \leq N} |E_S^{(i)T}\beta^{*}_S| \leq \norm{\beta^{*}_S}_2\max_{1\leq i \leq N} \norm{E_S^{(i)}}_2 \\
& \leq\sqrt{Qs}\left( C_Q \sqrt{sM_n} n^{-\gamma +1/2} \right) \\
& = \sqrt{Q}C_Q s\sqrt{M_n} n^{-\gamma +1/2}
\end{align*}
Thus, $\mathbb{E}\left[ \tfrac{\norm{\Sigma_{SS}^{-1}}_\infty}{N} \norm{A_S^T(E_S+H_S)\beta^{*}_S}_\infty \big| \mathcal{B} \right]$
\begin{align*}
= O\left( \sqrt{sM_n} \left( s\sqrt{M_n} n^{-\gamma +1/2} + \sqrt{\frac{s\log(N)}{n} } \right) \right).
\end{align*}
Furthermore, $\mathbb{E}\left[ \norm{(E_S+H_S)^T(E_S+H_S)\beta^{*}_S}_\infty \big| \mathcal{B} \right]\mathbb{P}(\mathcal{B})$
\begin{align*}
& =\mathbb{E}\left[ \max_{j \leq sM_n} |(E_{Sj}+H_{Sj})^T\left((E_{S}+H_{S})\beta_S^*\right)| \big| \mathcal{B} \right]\mathbb{P}(\mathcal{B}) \\
& \leq \mathbb{E}\left[ \max_{j\leq sM_n} \norm{E_{Sj}+H_{Sj}}_1 \norm{(E_{S}+H_{S})\beta_S^*}_\infty \big| \mathcal{B} \right]\mathbb{P}(\mathcal{B}) \\
& = \mathbb{E}\left[ \norm{E_{S}+H_{S}}_{1} \norm{(E_{S}+H_{S})\beta_S^*}_\infty \big| \mathcal{B} \right]\mathbb{P}(\mathcal{B}).
\end{align*}
Then, given that $\mathcal{B}$ occurs $\norm{E_{S}+H_{S}}_{1}$
\begin{align*}
\leq \norm{E_{S}}_{1}+\norm{H_{Sj}}_{1}
\leq N(C_Q n^{-\gamma+1/2}+n^{-a}),
\end{align*}
and, as before: $\EE{\norm{(E_{S}+H_{S})\beta_S^*}_\infty | \mathcal{B}}$
\begin{align*}
&\leq C_2 s\sqrt{M_n} n^{-\gamma +1/2} + C_3 \sqrt{\frac{s\log(N)}{n} }.
\end{align*}
Hence, $\EE{\norm{\tilde{\Sigma}_{SS}^{-1}(\tfrac{1}{N}\tilde{A}_S^T)(E_S+H_S)\beta^{*}_S}_\infty \Big| \cal B}\mathbb{P}(\mathcal{B})$
\begin{align*}
=& O\Big(\sqrt{sM_n}\left( s\sqrt{M_n} n^{-\gamma +1/2} + \sqrt{{s\log(N)}/{n} } \right)\\
&\left(1+ n^{-\gamma+1/2}+n^{-a}\right)\Big)\\
=& O\left( \sqrt{sM_n} \left( s\sqrt{M_n} n^{-\gamma +1/2} + \sqrt{{s\log(N)}/{n} } \right)\right)
\end{align*}
The next terms are bounded as follows\footnote{See Supplemental Materials for proof.}.
\lemmafirst{thm:Vterm} $\mathbb{E}\left[\norm{\tilde{\Sigma}_{SS}^{-1}(\tfrac{1}{N}\tilde{A}_S^T)V}_\infty \Big| \mathcal{B} \right]\mathbb{P}(\mathcal{B}) = O\left( \frac{s^{\sfrac{3}{2}}}{M_n^{2\gamma-\sfrac{1}{2}}} \right)$ .
\lemmafirst{thm:epsterm} $\EE{\norm{\tilde{\Sigma}_{SS}^{-1}(\tfrac{1}{N}\tilde{A}_S^T)\epsilon}_\infty \Big| \mathcal{B}}\mathbb{P}(\mathcal{B}) = O\left(\sqrt{\log(sM_n)/N}\right)$.
Lastly,
\begin{align*}
&\norm{\hat{u}_S}_\infty = \max_{j\in S} \norm{\hat{u}_j}_\infty \leq \max_{j\in S} \norm{\hat{u}_j}_2 \leq 1 \implies\\
&\norm{\tilde{\Sigma}_{SS}^{-1}\lambda_{\scalebox{.6}{N}}\hat{u}_S}_\infty \leq \lambda_{\scalebox{.6}{N}}\norm{\tilde{\Sigma}_{SS}^{-1}}_\infty \norm{\hat{u}_S}_\infty \leq \frac{\lambda_{\scalebox{.6}{N}}\sqrt{sM_n} }{\tilde{C}_{\min}}.
\end{align*}
Keeping only leading terms, $\EE{\norm{\hat{\beta}_S-\beta^{*}_S}_\infty \Big| \mathcal{B}}\mathbb{P}(\mathcal{B}) $
\begin{align*}
=&O\left( s^{\sfrac{3}{2}} M_n n^{-\gamma +1/2} + s\sqrt{{M_n\log(N)}/{n}} \right)\\
&+O\left({s^{\sfrac{3}{2}}}/{M_n^{2\gamma-\sfrac{1}{2}}}+\sqrt{{\log(sM_n)}/{N}} + \lambda_{\scalebox{.6}{N}}\sqrt{sM_n}\right).
\end{align*}
Hence, by assumptions (\ref{eq:bndH_S})-(\ref{eq:lambdaroot}) we have $\mathbb{P}\left(\norm{\hat{\beta}_S-\beta^{*}_S}_\infty > \frac{\rho^*_N}{2}\right) \rightarrow 0$
\end{proof}
One may similarly look at the stationarity for $j\in S^c$ to analyze $\hat{u}_j$: $0=\tfrac{1}{N}\tilde{A}_j^T\left(\tilde{A}_S\beta_S-Y\right)+\lambda_{\scalebox{.6}{N}}\hat{u}_j$
\begin{align*}
=&\tfrac{\tilde{A}_j^T}{N}\left(\tilde{A}_S(\hat{\beta}_S-\beta^{*}_S)-(A_S-\tilde{A}_S)\beta^{*}_S -V-\epsilon\right) +\lambda_{\scalebox{.6}{N}}\hat{u}_j.
\end{align*}
Thus,
\begin{align*}
\hat{u}_j =&\tfrac{1}{\lambda_{\scalebox{.6}{N}} N}\tilde{A}_j^T\tilde{A}_S(\beta^{*}_S-\hat{\beta}_S)+\tfrac{1}{\lambda_{\scalebox{.6}{N}} N}\tilde{A}_j^T(A_S-\tilde{A}_S)\beta^{*}_S \\
&+ \tfrac{1}{\lambda_{\scalebox{.6}{N}} N} \tilde{A}_j^T (V+\epsilon)\\
=&\tfrac{1}{\lambda_{\scalebox{.6}{N}}}\tilde{\Sigma}_{jS}\tilde{\Sigma}_{SS}^{-1}\Big( \tfrac{1}{N}\tilde{A}_S^T(E_S+H_S)\beta^{*}_S -\tfrac{1}{N}\tilde{A}_S^TV\\ &-\tfrac{1}{N}\tilde{A}_S^T\epsilon+\lambda_{\scalebox{.6}{N}}\hat{u}_S\Big)-\tfrac{1}{\lambda_{\scalebox{.6}{N}} N}\tilde{A}_j^T(E_S+H_S)\beta^{*}_S \\
&+ \tfrac{1}{\lambda_{\scalebox{.6}{N}} N} \tilde{A}_j^T (V+\epsilon),
\end{align*}
where $\tilde{\Sigma}_{jS}=\tfrac{1}{N}\tilde{A}_j^T\tilde{A}_S$ and using \eqref{eq:supp-diff}. We wish to show that $\forall j\in S^c$ $\hat{u}_j$ satisfies the KKT conditions, that is:
\propfirst{thm:muH} $\mathbb{P}\left(\max_{j\in S^c} \norm{\hat{u}_j}_2 < 1 \right) \rightarrow 1$.
\begin{proof}
Let $\mu_j^H \equiv \EE{\hat{u}_j \big| H}$. We proceed as follows:
\begin{align*}
&\mathbb{P}\left(\max_{j\in S^c} \norm{u_j}_2 < 1 \right) \\
&\ge \mathbb{P}\left(\max_{j\in S^c} \norm{\mu_j^H}_2 + \norm{u_j - \mu_j^H}_2 < 1 \right)\\
&\ge \mathbb{P}\left(\max_{j\in S^c} \norm{\mu_j^H}_2 + \sqrt{M_n} \norm{u_j - \mu_j^H}_\infty < 1 \right) \\
&\ge \mathbb{P} \left(\max_{j\in S^c} \norm{\mu_j^H}_2< 1-\tfrac{\tilde{\delta}}{2},\max_{j\in S^c} \norm{u_j - \mu_j^H}_\infty < \tfrac{\tilde{\delta}}{2\sqrt{M_n}} \right) \\
&\ge 1 - \mathbb{P}\left(\max_{j\in S^c} \norm{\mu_j^H}_2 \geq 1-\tilde{\delta} \right) \\
&- \mathbb{P}\left(\max_{j\in S^c} \norm{u_j - \mu_j^H}_\infty \geq \frac{\tilde{\delta}}{2\sqrt{M_n}} \right).
\end{align*}
We obtain the following results:
\lemmafirst{thm:uH} $\mathbb{P}\left(\max_{j\in S^c} \norm{\mu_j^H }_2 \geq 1-\frac{\tilde{\delta}}{2} \right) \rightarrow 0 $
\lemmafirst{thm:uj} $\mathbb{P}\left(\max_{j\in S^c} \norm{\hat{u}_j - \mu_j^H }_\infty \geq \frac{\tilde{\delta}}{2\sqrt{M_n}} \right) \rightarrow 0$
Hence, we have that $\mathbb{P}\left(\max_{j\in S^c} \norm{\hat{u}_j}_2 < 1 \right) \rightarrow 1$.
\end{proof}
\section{Experiments}
\subsection{Synthetic Data}
We tested the FuSSO on synthetic data-sets of $\mathcal{D}= \{ (\{\vec{y}^{\,(i)}_j\}_{j=1}^p,Y_i)\}_{i=1}^N$ (where $\vec{y}^{\,(i)}_j$ as in (\ref{eq:fvec})). The experiments performed were as follows. First, we fix $N,n,p,$ and $s$. For $i=1,\ldots,N$, $j=1,\ldots,p$ we create random functions using a maximum of $M$ projection coefficients as follows: 1) Set $a_{jm}\overset{iid}{\sim} \text{Unif}[-1,1]$ for $m=1,\ldots,M$; 2) set $a_{jm}= a_{ji}/c_m^2$, where $c_m = m$ if $m=1$ or is even, $c_m = m-1$ if $m$ is odd; 3) set $a_{jm} = a_{jm}/\norm{a_j}$; 4) set $\alpha^{(i)}_j = a_j$. (See Figures \ref{fig:func_n5},\ref{fig:func_n25} for typical functions.) Similarly, we generate $\beta_j^*$ for $j=1,\ldots,s$; for $j=s+1,\ldots,p$, we set $\beta^*_j = 0$. Then, we generate $Y_i$ as $Y_i = \sum_{j=1}^p \langle\beta^*_j,\alpha^{(i)}_j\rangle + \epsilon_i = \sum_{j=1}^s \langle\beta^*_j,\alpha^{(i)}_j\rangle + \epsilon_i $, where $\epsilon\overset{iid}{\sim} \mathcal{N}(0,.1)$. Also, a grid of $n$ noisy function evaluations were generated to make $\vec{y}^{\,(i)}_j$ as in (\ref{eq:fvec}), with $\sigma_\xi=.1$. These were then used to compute $\tilde{\alpha}^{(i)}_{jm}$ for $m=1,\ldots,M_n$ as in (\ref{eq:talpha}), $M_n$ was chosen by cross validation. (See Figures \ref{fig:func_n5}, \ref{fig:func_n25} for typical noisy observations and function estimates for $n=5$ and $n=25$ respectively.)
We fixed $s=5$ and chose the following configurations for the other parameters: $(p,N,n)\in\{(100,50,5),(1000,500,25),(20000,500,25)\}$. For each tuple of $(p,N,n)$ configurations, 100 random trails were performed. We recorded, $r$, the fraction of the trails that a $\lambda$ value was able recover the correct sparsity pattern (i.e.\ that only the first 5 functions are in the support). We also recorded the mean length of the range of $\lambda$, $\Delta_\lambda$, that were able to recover the correct support; i.e.\ $\Delta_\lambda=\tfrac{1}{t}\sum_{t=1}^{100}\Delta^{(t)}_\lambda$, where $\Delta^{(t)}_\lambda= (\lambda^{(t)}_f-\lambda^{(t)}_l)/\lambda^{(t)}_{\max}$, $\lambda^{(t)}_f$ is the largest $
\lambda$ value found to recover the correct support in the $\jth{t}$ trails, $\lambda^{(t)}_l$ the smallest such $\lambda$, and $\lambda^{(t)}_{\max}$ is the smallest $\lambda$ to produce $\hat{\beta}=0$ ($\Delta^{(t)}_\lambda$ is taken to be zero if no $\lambda$ recovered the correct support). The results were as follows:
\begin{center}
\begin{tabular}{l{c}{c}}
$(p,N,n)$ & $r$ & $\Delta$ \\
\hline
(100,50,5) & .68 & .2125 \\
(1000,500,25) & 1 & .4771 \\
(20000,500,25) & 1 & .4729 \\
\end{tabular}
\end{center}
Hence we see that even when the number of observations per function is small ($5$ or $25$) and the number of total number of input functional covariates is large (we were able to test up to $20000$), the FuSSO can recover the correct support.
Also, to illustrate this point that running Group-LASSO on the $\vec{y}^{\,(i)}_j$ features (Y-GL) is less robust to noise and adaptive to smoothness, we ran noisier trails using the configuration of $(p,N,n)=(1000,500,25)$. We increased the standard deviation of the noise on grid function observation and on the response to be $5$ and $1$ respectively. Under these conditions the FuSSO was able to recover the support in $49\%$ of the trails were as Y-GL recovered the support in $32\%$ of the trails. Furthermore the FuSSO had a $\Delta^{(t)}_\lambda=.0743$ compared to $\Delta^{(t)}_\lambda=.0254$ for Y-GL.
\begin{figure}[t]
\centering
\subfigure[Function at $n=5$]{\includegraphics[width=.23\textwidth]{n5_fest.pdf}}
\subfigure[Reg. Path at $p=100$, $n=50$, $n=5$ of $\norm{\hat{\beta}_j}$]{\includegraphics[width=.23\textwidth]{p100N50n5_freg}\label{fig:func_n5}}
\subfigure[Function at $n=25$]{\includegraphics[width=.23\textwidth]{n25_fest.pdf}\label{fig:func_n25}}
\subfigure[Reg. Path at $p=1000$, $n=500$, $n=25$ of $\norm{\hat{\beta}_j}$]{\includegraphics[width=.23\textwidth]{p1000N500n25_freg.pdf}}
\caption{(a)(c) Two typical functions, noisy observations, and estimates. (b)(d) Regularization paths showing the norms of $\hat{\beta}_j$ (in red for $j$ in support, blue otherwise) for a range of $\lambda$; rightmost vertical line indicates largest $\lambda$ able to recover the support, leftmost line for smallest such $\lambda$}
\vspace{-0.3cm}
\end{figure}
\subsection{Neurological Data}
We also tested the FuSSO estimator with a neurological data-set, using a total of 89 subjects \cite{yeh2011ntu}. Subjects ranged in age from 18 to 60 years old (Figure \ref{fig:ages}). Our goal was to learn a regression that maps the dODFs at each white matter voxel for each subject to the subject's age. The dODF is a function represents the amount of water molecules, or spins, undergoing diffusion in different orientations over the $S^2$ sphere\cite{yeh2010generalized}. I.e., each dODF is a function with a $2d$ domain (of azimuth, elevation spherical coordinates) and a range of reals representing the strength of water diffusion at the given orientations (see Figure \ref{fig:odf}). Data was provided for each subject in a template space for white-matter voxels; a total of over 25 thousand voxels' dODFs were regressed on (i.e.\ $p\approx 25000$). We also compared regression using the FuSSO and functional covariates to using the LASSO and real valued covariates. We used the non-functional collection of quantitative anisotropy (QA) values for the same white matter voxels as with dODF functions. QA values are the estimated amount of spins that undergo diffusion in the direction of the principle fiber orientation, i.e., the peak of the dODF; QAs have been used as a measure of white matter integrity in the underlying voxel hence making for a descriptive and effective summary statistic of an dODF function for age regression \cite{yeh2010generalized}.
The projection coefficients for the dODFs at each voxel were estimated using the cosine basis. The FuSSO estimator gave a cross-validated MSE of $70.855$, where the variance for age was $156.4265$; selected voxels in the support may be seen in Figure \ref{fig:active}. The LASSO estimate using QA values gave a cross-validated MSE of $77.1302$. Thus, one may see that considering the entire functional data gave us better results for age regression. We note that we were unable to use the naive approach of Y-GL in this case because of memory constraints and the fact that function evaluation points did not lie on a $2d$ square grid.
\begin{figure}[h]
\centering
\subfigure[Example ODF]{\includegraphics[width=.23\textwidth]{original_odf.png}\label{fig:odf}}
\subfigure[Ages]{\includegraphics[width=.225\textwidth]{ages.pdf}\label{fig:ages}}
\subfigure[Voxels in support]{\includegraphics[width=.21\textwidth]{brain_lin_reg.png}\label{fig:active}}
\subfigure[Errors]{\includegraphics[width=.225\textwidth]{errs.pdf}}
\caption{(a) An example ODF for a voxel. (b) Histogram of ages for subjects. (c) Voxels in the support of model shown in blue. (d) Histogram of held out error magnitudes.}
\vspace{-0.3cm}
\end{figure}
\vspace{-0mm}
\section{Conclusion}
In conclusion, this paper presents the FuSSO, a functional analogue to the LASSO. The FuSSO allows one to efficiently find a sparse set of functional input covariates to regress a real-valued response against. The FuSSO makes no parametric assumptions about the nature of input functional covariates and assumes a linear form to the mapping of functional covariates to the response. We provide a statistical backing for use of the FuSSO via proof of asymptotic sparsistency.
\subsubsection*{Acknowledgements}
This work is supported in part by NSF grants IIS1247658 and IIS1250350.
\clearpage
|
\section{Introduction}
We shall be concerned with finite families of finite sets; as often, we shall assume that such a family is a subset of $\mathcal P(n)=\mathcal P([n])$ for some $n$, where $\mathcal P$ denotes the powerset and $[n]=\{1,\dots,n\}$. For $\mathcal A \subseteq \mathcal P(n)$, we call $\mathcal A$ \emph{union-closed} if for any two elements $A$ and $B$ of $\mathcal A$ the set $A\cup B$ is also in $\mathcal A$. For $i\in \mathbb N$, the \emph{degree of $i$ in $\mathcal A$}, denoted $\mathrm{deg}_{\mathcal A}(i)$, is simply
\[
|\{A\in \mathcal A: i\in A\}|.
\]
The \emph{union-closed conjecture}, often attributed to Frankl \cite{Duf}, states that if $\mathcal A$ is a union-closed family other than $\{\emptyset\}$ then there is some $i$ with $\mathrm{deg}_{\mathcal A}(i)\ge |\mathcal A|/2$.
A related problem is the \emph{union-closed size problem}, which asks how small the sets of a union-closed family can be. For a finite family $\mathcal A\subseteq \mathcal P(n)$, we define the \emph{total size} of $\mathcal A$ to be
\[
||\mathcal A|| = \sum_{A\in \mathcal A}|A|.
\]
Then the union-closed size problem asks what is the value of
\[
f(m) = \min ||\mathcal A||,
\]
where the minimum runs over union-closed families which consist of $m$ sets. This problem was first addressed by Reimer \cite{Rei} in 2003, who proved that
\[
f(m) \ge \frac{m}{2} \log_2 m.
\]
Recently, Balla, Bollob\'as and Eccles \cite{BaBoEc} settled the union-closed size problem entirely, determining the exact value of $f(m)$ for all $m$. We denote by $\mathcal I(m)$ the initial segment of the colex order on $\mathbb N^{(<\infty)}$ of length $m$; this order shall be defined fully in Section \ref{sec_defs}.
\begin{theorem}\label{thm_colex_bound}
Let $m$ be a positive integer, and let $n$ be the unique integer with $2^{n-1} < m \le 2^n$. Set $m' = 2^n-m$. Then
\[
f(m) = ||\mathcal P(n)|| - ||\mathcal I(m')|| - m'.
\]
In particular, if $\mathcal A$ is a counterexample to the union-closed conjecture in $\mathcal P(n)$ with $|\mathcal A|=m$ then $f(m)<nm/2$, and so
\[
||\mathcal I(m')||+m' > n2^n/2 - nm/2 = nm'/2.
\]
\end{theorem}
\noindent
The extremal family $\mathcal A$ for the first part of the theorem has $\mathcal P(n)\setminus \mathcal A = \{B\cup \{n\}: B\in \mathcal I(m')\}$. Through bounding $||\mathcal I(m')||$, this result is sufficient to prove the union-closed conjecture if $|\mathcal A|$ is large --- in fact the following bound is given in \cite{BaBoEc}.
\begin{corollary}\label{cor_old_bound}
The union-closed conjecture holds for all union-closed families $\mathcal A\subseteq \mathcal P(n)$ with $|\mathcal A| \ge \frac{2}{3}2^n$.
\end{corollary}
However, this is as far as one can go considering only averaging arguments --- if $m < \frac{2}{3}2^n$, then $f(m) < mn/2$. From this, one might reasonably assume that the constant $\frac{2}{3}$ in Corollary \ref{cor_old_bound} is hard to improve to any constant $\frac{2}{3}-\epsilon$. But in the extremal examples for $f(m)$, the family $\mathcal A$ is very asymmetric --- indeed, there is a single element which is in every set of $\mathcal P(n)\setminus \mathcal A$ --- and so $\mathcal A$ is in a sense far away from being a counterexample to the union-closed conjecture. In this paper, we prove a stability result for the union-closed size problem for union-closed families $\mathcal A\subseteq \mathcal P(n)$ with $|\mathcal A|\ge 2^{n-1}$. Roughly speaking, we show that if $||\mathcal A||$ is close to the maximum possible then $\mathcal P(n)\setminus \mathcal A$ has an element of high degree --- this result is Theorem \ref{thm_stability}. This enables us to extend Theorem \ref{thm_colex_bound}.
\begin{theorem}\label{thm_main}
There is a positive constant $c_1$ such that if $\mathcal A$ is a counterexample to the union-closed conjecture in $\mathcal P(n)$, and $\mathcal B= \mathcal P(n)\setminus \mathcal A$ with $|\mathcal B| = m$, then
\[
||\mathcal I(m)|| > m(n/2-1+c_1).
\]
\end{theorem}
\noindent
Using simple bounds on $||\mathcal I(m)||$, this extends slightly the range where we can prove the union-closed conjecture.
\begin{corollary}\label{cor_main}
There is a positive constant $c_2$ such that the union-closed conjecture holds for all union-closed familes $\mathcal A\subseteq \mathcal P(n)$ with $|\mathcal A| \ge 2^n(2/3 - c_2)$.
\end{corollary}
In fact, we shall prove these theorems with bounds of $c_1\ge 1/24$ and $c_2 \ge 1/104$.
The rest of the paper is organised as follows. In Section \ref{sec_defs}, we define the concepts needed in the proofs of our main theorems --- in particular down-compressions and simply rooted families, which shall be at the heart of our argument. In Section \ref{sec_state} we state Theorem \ref{thm_stability}, our stability result for Theorem \ref{thm_colex_bound}. In Section \ref{sec_pf_thm}, we prove Theorem \ref{thm_stability}, and use it to prove Theorem \ref{thm_main}. In Section \ref{sec_pf_cor} we bound $||\mathcal I(m)||$, proving Corollary \ref{cor_main} from Theorem \ref{thm_main}. In Section \ref{sec_refine} we prove a slightly stronger form of Theorem \ref{thm_stability}, which improves the constants $c_1$ and $c_2$ a little --- this is left out of the main proof for the sake of clarity.
\section{Definitions}\label{sec_defs}
In this section, we recall some concepts used by Reimer \cite{Rei} and Balla, Bollob\'as and Eccles \cite{BaBoEc} in their work on the union-closed size problem. Central to both of those papers are \emph{compressions}. Up- and down-compressions are by now standard; see for example Bollob\'as and Leader \cite{BoLe}. For a family $\mathcal B\subseteq \mathcal P(n)$ and $i\in [n]$, we define the \emph{down-compression of $\mathcal B$ in direction $i$}, denoted $d_i(\mathcal B)$, by defining
\[
d_{(i,\mathcal B)}(B)=
\begin{cases}
B-i: i\in B,\, B-i\notin \mathcal B\\
B: \mathrm{ otherwise,}
\end{cases}
\]
and $d_i(\mathcal B) = \{d_{(i,\mathcal B)}(B): B\in \mathcal B\}$. A down-compression of a family $\mathcal B$ is equivalent to an up-compression on its complement in $\mathcal P(n)$, in that
\begin{equation}\label{eq_comp_equiv}
\mathcal P(n)\setminus d_i(\mathcal B) = u_i(\mathcal P(n)\setminus \mathcal B),
\end{equation}
where $u_i$ is the up-compression in direction $i$, defined analogously to $d_i$. Also, for $\mathcal B\subseteq \mathcal P(n)$ we define $d(\mathcal B)$ to be $d_n\dots d_1(\mathcal B)$, the compression obtained by applying the compressions $d_i$ to $\mathcal B$ for $1\le i \le n$, starting with $d_1$.
For $B\in \mathcal B$ we define $d_{\mathcal B}(B)$ to be the image of $B$ under the down-compression $d_\mathcal B$; that is, letting $\mathcal B_i$ = $d_i\dots d_1 (\mathcal B)$, we define
\[
d_\mathcal B(B) = d_{(n, \mathcal B_{n-1})}\dots d_{(2, \mathcal B_1)}d_{(1,\mathcal B)}(B);
\]
so $d_{\mathcal B}(B)$ is the set we get by following $B$ through the compressions $d_i$. Similarly, we shall often want to consider the family $\mathcal B$ after some of the compressions $d_i$ have been applied; to this end we define $D_k(\mathcal B) = d_k\dots d_1(\mathcal B)$, the family after compressing in directions $i$ for $1\le i \le k$, and for $B\in \mathcal B$ we define $D_{(\mathcal B,k)}(B) = d_{(k, \mathcal B_{k-1})}\dots d_{(1,\mathcal B)}(B)$, the image of the set $B$ in $D_k(\mathcal B)$.
Following the approach of \cite{BaBoEc}, we shall view the complement of a union-closed family as a simply rooted family --- this perspective is crucial for our proof of Theorem \ref{thm_main}. We call a family $\mathcal B\subseteq \mathcal P(n)$ \emph{simply rooted} if for every $\emptyset \neq B\in \mathcal B$, there is some $b\in B$ with $[\{b\},B]\subseteq \mathcal B$. The following simple observation was made in \cite{BaBoEc}.
\begin{observation}\label{obs_rooted}
Let $\mathcal A\subseteq \mathcal P(n)$, and $\mathcal B = \mathcal P(n)\setminus \mathcal A$. Then $\mathcal B$ is a simply rooted family if and only if $\mathcal A$ is a union-closed family.
\end{observation}
\begin{proof}
The family $\mathcal A$ is union-closed exactly when for every $B\notin \mathcal A$ we have
\[
\bigcup_{B'\subseteq B,\,B'\in \mathcal A}B'\neq B,
\]
which is in turn true exactly when $[\{b\},B]\subseteq \mathcal B$ for some $b\in B$.
\end{proof}
Finally, we recall the colex order on $\mathbb N^{(<\infty)}$, the collection of finite sets of positive integers, and some of its standard properties. Given $A$ and $B$ sets in $\mathbb N^{(<\infty)}$, we define the colex order $<$ by
\[
A<B \iff \max(A\triangle B) \in B.
\]
This is a linear order on $\mathbb N^{(<\infty)}$. We write $\mathcal I(m)$ for the initial segment of this order of length $m$; so, for example, $\mathcal I(9) = \{\emptyset, 1, 2, 12, 3, 13, 23, 123, 4\}$, where we write $13$ for the set $\{1,3\}$. Also, a family of sets $\mathcal D$ is called a \emph{down-set} if for every $A\in \mathcal D$ we have $\mathcal P(A)\subseteq \mathcal D$. The following result is a well-known consequence of the fundamental theorem of Kruskal \cite{Kru} and Katona \cite{Kat}.
\begin{lemma}\label{lem_KK}
If $\mathcal D$ is a down-set, then
$||\mathcal D||\le ||\mathcal I(|\mathcal D|)||$. \hfill\qed
\end{lemma}
The other fact which we shall need about initial segments of colex is the following lemma, which is a simple corollary of Lemma \ref{lem_KK} -- see for example \cite{BaBoEc}.
\begin{lemma}\label{lem_colex_sums}
Let $m_1$ and $m_2$ be positive integers. Then
\[
||\mathcal I(m_1)||+||\mathcal I(m_2)|| \le ||\mathcal I(m_1+m_2)|| - \min(m_1,m_2).
\]
\end{lemma}
This can be proved for $m_1\ge m_2$ by applying Lemma \ref{lem_KK} to the down-set $\mathcal I(m_1)\cup\{A+N:A\in \mathcal I(m_2)\}$, for a sufficiently large integer $N$. \hfill\qed
\section{Stability for sizes of simply rooted families}\label{sec_state}
For a simply rooted family $\mathcal B\subseteq \mathcal P(n)$, a set $B\in \mathcal B$ and an element $i\in [n]$, we say that $B$ is \emph{$\mathcal B$-rooted at $i$} if $i\in B$ and the cube $[\{i\},B]$ is contained in $\mathcal B$. Then for a set $S\subseteq [n]$, we define $\mathcal B_S$ to be those sets of $\mathcal B$ which are $\mathcal B$-rooted at some $i\in S$.
Let $\mathcal B\subseteq \mathcal P(n)$ be a simply rooted family. By Observation \ref{obs_rooted} the family $\mathcal P(n)\setminus \mathcal B$ is union-closed, and so Theorem \ref{thm_colex_bound} gives us
\begin{equation}\label{eq_rooted_bound}
||\mathcal B|| = ||\mathcal P(n)|| - ||\mathcal A|| \le ||\mathcal P(n)||-f(|\mathcal A|) = ||\mathcal I(|\mathcal B|)||+|\mathcal B|.
\end{equation}
For any $m$, the family $\mathcal B=\{B+n:B\in \mathcal I(m)\}$ makes this inequality tight for $n=\lceil log_2(m)\rceil +1$ --- every set in this family is $\mathcal B$-rooted at $n$. In fact, up to isomorphism this is the only simply rooted family of $m$ sets for which equality holds; this is a consequence of the uniqueness of extremal families for $f(m)$, which was proved in \cite{BaBoEc}. In particular, if $||\mathcal B|| = ||\mathcal I(|\mathcal B|)||+|\mathcal B|$ then $\mathcal B_{\{i\}} = \mathcal B$ for some $i\in [n]$. The following result extends this, showing that if $||\mathcal B||$ is close to $||\mathcal I(m)||+m$ then $|\mathcal B_{\{i\}}|$ is large for some $i$.
\begin{theorem}\label{thm_stability}
Let $\mathcal B$ be a simply rooted family in $\mathcal P(n)$ with $|\mathcal B| = m$, and $p\in [0,1]$. Suppose that $|\mathcal B_{\{i\}}| \le p m$ for all $i\in [n]$. Then
\begin{equation*}
||\mathcal B|| \le ||\mathcal I(m)|| + m - m^2(1/12 - p^2/12)/2^n.
\end{equation*}
\end{theorem}
Theorem \ref{thm_stability} provides a stability result for Theorem \ref{thm_colex_bound} for union-closed families $\mathcal A \subseteq \mathcal P(n)$ with $|\mathcal A|\ge 2^{n-1}$. Indeed, let $\mathcal A$ be such a family and set $\mathcal B = \mathcal P(n)\setminus \mathcal A$ with $|\mathcal B|=m$. Since $\mathcal B$ is a simply rooted family by Observation \ref{obs_rooted}, if $|\mathcal B_{\{i\}}| \le p m$ for all $i$ we have
\begin{align*}
||\mathcal A|| &= ||\mathcal P(n)||-||\mathcal B||\\
& \ge ||\mathcal P(n)||-||\mathcal I(m)|| - m + m^2(1/12 - p^2/12)/2^n\\
&= f(|\mathcal A|) + m^2(1/12 - p^2/12)/2^n.
\end{align*}
Hence if $||\mathcal A||$ is close to $f(|\mathcal A|)$, some element of $[n]$ appears in nearly all the sets of $\mathcal B$.
\section{Proofs of main theorems}\label{sec_pf_thm}
Now we turn to the proofs of Theorems \ref{thm_stability} and \ref{thm_main}. First we give two definitions we shall need in the proof of Theorem \ref{thm_stability}. For a finite set $B$, let $\delta B = \{B-i : i \in B\}$ be the \emph{shadow of $B$}. Given a simply rooted family $\mathcal B\subseteq \mathcal P(n)$, we call a set $B\in \mathcal B$ a \emph{bad set of $\mathcal B$} if either $\delta B\subseteq \mathcal B$ or $d_{\mathcal B}(B)=B$. We call a set $B\in \mathcal B$ that is not bad a \emph{good set of $\mathcal B$}.
We now sketch the proof of Theorem \ref{thm_stability}. In Lemmas \ref{lem_no_falls} and \ref{lem_full_sh} we shall show that if $\mathcal B$ has many bad sets then $||\mathcal B||$ is much less than $||\mathcal I(m)||+m$; as a result, it is enough to show that a simply rooted family satisfying the condition of Theorem \ref{thm_stability} has many bad sets. Given such a simply rooted family $\mathcal B$, we shall then write $\mathcal B$ as $\mathcal B_S\cup \mathcal B_T$, where $S\cup T$ is a partition of $[n]$. Since no $\mathcal B_{\{i\}}$ is too large, we can do this so both $\mathcal B_S$ and $\mathcal B_T$ are fairly large. If their intersection $|\mathcal B_S\cap \mathcal B_T|$ is large, then we conclude that $\mathcal B$ has many bad sets, since all the sets of $\mathcal B_S\cap \mathcal B_T$ are $\mathcal B$-rooted at two elements of $[n]$, and so are bad sets of $\mathcal B$. If, on the other hand, $|\mathcal B_S\cap \mathcal B_T|$ is small, we shall show in Corollary \ref{cor_lower_b} that $\mathcal B$ still has many bad sets. We prove this by considering the down-sets $d(\mathcal B_S)$ and $d(\mathcal B_T)$; since these are large down-sets in $\mathcal P(n)$, they have a large intersection, and in Lemma \ref{lem_split_rooted} we shall show that sets in this intersection correspond to sets in either $\mathcal B_S \cap \mathcal B_T$ or bad sets of $\mathcal B$.
\subsection{Applying down-compressions to simply rooted families}
In Lemma \ref{lem_no_falls}, we shall show that if $\mathcal B$ has many sets with $d_\mathcal B(B)=B$ then $||\mathcal B||$ is small. In order to prove this lemma, we first recall some results of Reimer \cite{Rei} on union-closed families, restating them in terms of simply rooted families.
\begin{lemma}\label{lem_Rei_basics}
Let $\mathcal B\subseteq \mathcal P(n)$ be a simply rooted family. Then
\begin{enumerate}
\item \label{it_end_down} $d(\mathcal B)$ is a down-set,
\item \label{it_rooted}for $1\le k\le n$, $D_k(\mathcal B)$ is a simply rooted family.\hfill\qed
\end{enumerate}
\end{lemma}
We now prove some further basic properties of down-compressions on simply rooted families.
\begin{lemma}\label{lem_rooted_basics}
Let $\mathcal B\subseteq \mathcal P(n)$ be a simply rooted family. Then
\begin{enumerate}
\item \label{it_down-set}for $B\in \mathcal B$ and $1\le k\le n$, if $D_{(\mathcal B, k)}(B)\neq B$ then $\mathcal P(D_{(\mathcal B, k)}(B))\subseteq D_k(\mathcal B)$,
\item \label{it_one_fall}for $B\in \mathcal B$, $|B\setminus d_{\mathcal B}(B)|\le 1$.
\end{enumerate}
\end{lemma}
\begin{proof}
Suppose that $D_{(\mathcal B, j)}(B)\neq B$ for some $j\in [n]$; otherwise both parts of the lemma hold for the set $B$. Let $\ell$ be minimal with $D_{(\mathcal B, \ell)}(B) \neq B$. Then $D_{(\mathcal B, \ell)}(B) = B-\ell$, and $B-\ell\notin D_{\ell-1}(\mathcal B)$. Also, by Part \ref{it_rooted} of Lemma \ref{lem_Rei_basics}, $D_{\ell-1}(\mathcal B)$ is simply rooted, and so there is some $i\in B$ such that $[\{i\},B] \subseteq D_{\ell-1}(\mathcal B)$ --- and since $B-\ell \notin D_{\ell-1}(\mathcal B)$, we must have $i = \ell$. Hence $\mathcal P(B-\ell) \subseteq D_\ell(\mathcal B)$, since if a family $\mathcal F$ contains $S+\ell$ for some set $S$ then $d_\ell(\mathcal F)$ contains $S$.
Now, $\mathcal P(B-\ell)$ is a down-set which is contained in $D_\ell(\mathcal B)$, and so any down-compression of the family $D_\ell(\mathcal B)$ fixes every set in $\mathcal P(B-\ell)$. For Part \ref{it_down-set} of the lemma, if $D_{(\mathcal B, k)}(B)\neq B$ for some $k\in [n]$, $k\ge \ell$, and so $D_k(\mathcal B) = d_k\dots d_{l+1}(D_l(\mathcal B))$. Hence we have $D_{(\mathcal B,k)}(B) = B-\ell$ and
\[
\mathcal P(D_{(\mathcal B, k)}(B)) = \mathcal P(B-\ell)\subseteq D_k(\mathcal B),
\]
so Part \ref{it_down-set} holds. For Part \ref{it_one_fall}, we have $d_{\mathcal B}(B)=D_{(\mathcal B,\ell)}(B)=B-\ell$, so $|B\setminus d_{\mathcal B}(B)|=1$.
\end{proof}
From Lemmas \ref{lem_Rei_basics} and \ref{lem_rooted_basics}, we immediately get a bound on the total size of a simply rooted family.
\begin{lemma}\label{lem_no_falls}
Let $\mathcal B$ be a simply rooted family, let $|\mathcal B| = m$, and let $m'$ be the number of sets $B\in \mathcal B$ with $d_{\mathcal B}(B)= B$. Then
\[
||\mathcal B||\le ||\mathcal I(m)|| + m - m'.
\]
\end{lemma}
\begin{proof}
Note that
\[
||\mathcal B|| = ||d(\mathcal B)|| + \sum_{B\in \mathcal B} |B\setminus d_{\mathcal B}(B)|.
\]
By Part \ref{it_end_down} of Lemma \ref{lem_Rei_basics}, $d(\mathcal B)$ is a down-set, and so by Lemma \ref{lem_KK} $||d(\mathcal B)||$ is at most $||\mathcal I(m)||$. Also, by Part \ref{it_one_fall} of Lemma \ref{lem_rooted_basics}, $\sum_{B\in \mathcal B} |B\setminus d_{\mathcal B}(B)|$ is exactly $m-m'$, and so the result follows.
\end{proof}
Note that if $\mathcal B$ is a simply rooted family and $B \in \mathcal B$, $|\delta B\setminus \mathcal B|\le 1$. We shall now show that if there are many $B\in \mathcal B$ with the entire shadow of $B$ contained in $\mathcal B$ then Theorem \ref{thm_stability} holds.
\begin{lemma}\label{lem_full_sh}
Let $\mathcal B$ be a simply rooted family of size $m$, and set $m'$ to be the number of sets $B\in \mathcal B$ with $\delta B \subseteq \mathcal B$. Then
\[
||\mathcal B|| \le ||I(m)|| + m-m'.
\]
\end{lemma}
We shall deduce this from a more general result. For $\mathcal B$ a finite family of finite sets, we define the \emph{deficiency of $\mathcal B$}, denoted $\mathrm{def}(\mathcal B)$, to be the number of sets in the shadows of sets $B\in \mathcal B$ that are missing from $\mathcal B$ --- that is,
\[
\mathrm{def}(\mathcal B) = \sum_{B\in \mathcal B}|\delta(B)\setminus \mathcal B|.
\]
Then we have the following lemma concerning the total size of a family of given size and deficiency.
\begin{lemma}\label{lem_def}
Suppose that $\mathcal B$ is a finite family of finite sets in $\mathcal P(n)$, with $|\mathcal B|=m$. Then
\begin{equation*}
||\mathcal B|| \le ||\mathcal I(m)|| + \mathrm{def}(\mathcal B).
\end{equation*}
\end{lemma}
We note that Lemma \ref{lem_full_sh} is immediate from this lemma, since if $\mathcal B$ is a simply rooted family then for each $B$ we have $|\delta B \setminus \mathcal B|\le 1$, and there are $m'$ sets $B\in \mathcal B$ with $|\delta B \setminus \mathcal B| = 0$, so $\mathrm{def}(\mathcal B)=m-m'$.
\begin{proof}
We apply induction on $n$. For $n=1$ the result is easily checked. If $n>1$, we define families of sets $\mathcal B_n^+$ and $\mathcal B_n^-$ by
\begin{align*}
\mathcal B^+_n &= \{B\in \mathcal P(n-1):B + n\in \mathcal B\},\,\textrm{and}\\
\mathcal B^-_n &= \{B\in \mathcal P(n-1): B\in \mathcal B\},
\end{align*}
so that $|\mathcal B| = |\mathcal B_n^+|+|\mathcal B_n^-|$, and $||\mathcal B|| = ||\mathcal B_n^+||+||\mathcal B_n^-|| + |\mathcal B_n^+|$. We define $m_n^+=|\mathcal B_n^+|$, and $m_n^-=|\mathcal B_n^-|$. Now we count pairs $(B,i)$ such that $B\in \mathcal B$, $i\in B$ and $B-i \notin \mathcal B$ --- these are the pairs which contribute to $\mathrm{def}(\mathcal B)$. We obtain
\begin{align*}
\mathrm{def}(\mathcal B) = &|\{B\in \mathcal B,\, i\in[n]: i\neq n,\,n\in B,\,B-i\notin \mathcal B\}|+\\
&|\{B\in \mathcal B,\,i\in [n]: i\neq n,\,n \notin B,\,B-i\notin \mathcal B\}|+\\
&|\{B\in \mathcal B: n \in B,\,B-n\notin \mathcal B\}|\\
&= \mathrm{def}(\mathcal B_n^+) + \mathrm{def}(\mathcal B_n^-)+|\mathcal B_n^+\setminus \mathcal B_n^-|.
\end{align*}
By the induction hypothesis and Lemma \ref{lem_colex_sums},
\begin{align*}
||\mathcal B|| &= ||\mathcal B_n^+|| + ||\mathcal B_n^-|| + |\mathcal B_n^+|\\
&\le ||\mathcal I(m_{n,+})|| + ||\mathcal I(m_{n,-})|| + m_{n,+} + \mathrm{def}(\mathcal B_n^+)+ \mathrm{def}(\mathcal B_n^-)\\
&\le ||\mathcal I(m)|| - \min(m_{n,+},m_{n,-}) + m_{n,+} + \mathrm{def}(\mathcal B) - |\mathcal B_n^+\setminus \mathcal B_n^-|.
\end{align*}
If $m_{n,+}\le m_{n,-}$ then $||\mathcal B||\le ||\mathcal I(m)|| + \mathrm{def}(\mathcal B)$, and so we are done. If not, then since $|\mathcal B_n^+\setminus \mathcal B_n^-| \ge m_{n,+} - m_{n,-}$ we have
\begin{align*}
||\mathcal B|| &\le ||\mathcal I(m)|| - m_{n,-} +m_{n,+} + \mathrm{def}(\mathcal B) - (m_{n,+} - m_{n,-})\\
&= ||\mathcal I(m)|| + \mathrm{def}(\mathcal B),
\end{align*}
and so we are also done.
\end{proof}
\begin{remark}
For positive integers $k$ and $m$, there is a family $\mathcal B$ with $|\mathcal B| =m$ and $\mathrm{def}(\mathcal B) = km$ so that the inequality in Lemma \ref{lem_def} is tight --- we can take $\mathcal B = \{A\cup\{N,\dots,N+k-1\}: A\in \mathcal I(m)\}$, for $N$ a sufficiently large integer. For general $|\mathcal B|$ and $\mathrm{def}(\mathcal B)$, there is not always a family $\mathcal B$ so that the inequality is tight; for example if $|\mathcal B|=2$ and $\mathrm{def}(\mathcal B) = 3$, then in fact $||\mathcal B|| \le 3 = ||\mathcal I(2)|| + 2$.
\end{remark}
Together, Lemmas \ref{lem_no_falls} and \ref{lem_full_sh} show that if $\mathcal B$ has many bad sets then $||\mathcal B||$ is small. Indeed, if $\mathcal B$ has $b$ bad sets then either at least $b/2$ sets of $\mathcal B$ have $\delta B\subseteq \mathcal B$, or at least $b/2$ have $d_{\mathcal B}(B)=B$. By Lemma \ref{lem_full_sh} in the first case, and Lemma \ref{lem_no_falls} in the second,
\[
||\mathcal B||\le ||\mathcal I(|\mathcal B|)||+|\mathcal B|-b/2.
\]
Our aim now is to give a lower bound on the number of bad sets of $\mathcal B$. To do this, we shall focus on how the down-compression $d_{\mathcal B}$ affects the sets of a simply rooted family $\mathcal B$.
\begin{lemma}\label{lem_fall_b}
Let $\mathcal B$ be a simply rooted family, and let $B\in \mathcal B$ with $B-b\notin \mathcal B$ for some $b\in B$. Then $d_{\mathcal B}(B) \in \{B, B-b\}$.
\end{lemma}
\begin{proof}
Since $\mathcal B$ is a simply rooted family, for some $a\in B$ we have $[\{a\},B]\subseteq \mathcal B$. But $B-b\notin \mathcal B$, and so $a=b$. We now consider $D_{b-1}(\mathcal B)$, the family obtained by applying the compressions $d_1,\dots,d_{b-1}$ to $\mathcal B$, starting with $d_1$. We note that the cube $[\{b\},B]$ is fixed when we apply any down-compression $d_i$ with $i\neq b$ to $\mathcal B$; indeed, if $A\in [\{b\},B]$ then $A-i \in [\{b\},B]$, so $d_{(\mathcal B,i)}(A)=A$. Hence we have $[\{b\},B] \subseteq D_{b-1}(\mathcal B)$.
We now consider two cases. If $B-b \in D_{b-1}(\mathcal B)$ then it is $D_{(\mathcal B,b-1)}(B')$ for some $B'\in \mathcal B$. Hence $D_{(\mathcal B,b-1)}(B')\neq B'$, and so by Part \ref{it_down-set} of Lemma \ref{lem_rooted_basics} we have $\mathcal P(B-b)\subseteq D_{b-1}(\mathcal B)$. Since $[\{b\},B]\subseteq D_{b-1}(B)$, we then have $\mathcal P(B)\subseteq D_{b-1}(\mathcal B)$ and so $d_\mathcal B(B)=B$. On the other hand, if $B-b \notin D_{b-1}(\mathcal B)$ then $D_{(\mathcal B,b)}(B)= B-b$, and by Part \ref{it_one_fall} of Lemma \ref{lem_rooted_basics} we have $d_\mathcal B(B) = B-b$.
\end{proof}
In the next lemma, we show that if $\mathcal B'\subseteq \mathcal B$ are simply rooted families, then sets in $\mathcal B'$ which are fixed by $d_{\mathcal B'}$ are also fixed by $d_\mathcal B$.
\begin{lemma}\label{lem_smaller_falls}
Let $\mathcal B'\subseteq \mathcal B$ be simply rooted families, and let $B\in \mathcal B'$ with $d_{\mathcal B'}(B)=B$. Then $d_{\mathcal B}(B)=B$.
\end{lemma}
\begin{proof}
By induction on $k$, it is easy to show that for all $1\le k \le n$ we have $D_k(\mathcal B') \subseteq D_k(\mathcal B)$. Indeed, if $\mathcal F'\subseteq \mathcal F\subseteq \mathcal P(n)$, then for any $i\in [n]$ we have $d_i(\mathcal F')\subseteq d_i(\mathcal F)$. Since $d_{\mathcal B'}(B)=B$, for all $k \in B$ we must have $B-k \in D_{k-1}(\mathcal B')$, and so also $B-k \in D_{k-1}(\mathcal B)$. This is exactly the condition we need to guarantee $d_{\mathcal B}(B)=B$.
\end{proof}
From the previous lemmas, we can read out a result on how the good sets of $\mathcal B$ behave under down-compressions $d_{\mathcal B'}$ for simply rooted families $\mathcal B'\subseteq \mathcal B$.
\begin{lemma}\label{lem_good_fall}
Let $\mathcal B'\subseteq \mathcal B$ be simply rooted families, with $B\in \mathcal B'$ a good set of $\mathcal B$. Then $d_\mathcal B(B) = d_{\mathcal B'}(B)$.
\end{lemma}
\begin{proof}
Since $B$ is a good set of $\mathcal B$, for some $b \in B$ we have $B-b\notin \mathcal B$, and by Part \ref{it_one_fall} of Lemma \ref{lem_rooted_basics} we have $d_\mathcal B(B) = B-a$ for some $a\in B$. So by Lemma \ref{lem_smaller_falls}, $d_{\mathcal B'}(B)\neq B$, and so $d_{\mathcal B'}(B) = B-c$ for some $c\in B$. But by Lemma \ref{lem_fall_b}, $a=c=b$, and the result holds.
\end{proof}
\subsection{Unions of simply rooted families}
Now we are in a position to prove a lemma about simply rooted families $\mathcal B$ which can be decomposed as the union of two other simply rooted families $\mathcal B_1 \cup \mathcal B_2$. Specifically, we show that if $\mathcal B_1$ and $\mathcal B_2$ have small intersection, but the down-sets $d(\mathcal B_1)$ and $d(\mathcal B_2)$ have large intersection, then $\mathcal B$ has many bad sets.
\begin{lemma}\label{lem_split_rooted}
Let $\mathcal B$, $\mathcal B_1$ and $\mathcal B_2$ be simply rooted families, with $\mathcal B_1 \cup \mathcal B_2 = \mathcal B$. Let $b$ be the number of bad sets of $\mathcal B$. Then
\[
|d(\mathcal B_1)\cap d(\mathcal B_2)| \le b + |\mathcal B_1 \cap \mathcal B_2|.
\]
\end{lemma}
\begin{proof}
Let $B$ be a set in $d(\mathcal B_1)\cap d(\mathcal B_2)$. Then $B= d_{\mathcal B_1}(B_1)= d_{\mathcal B_2}(B_2)$, for some $B_1\in \mathcal B_1$ and $B_2\in \mathcal B_2$. If both $B_1$ and $B_2$ are good sets of $\mathcal B$ then, applying Lemma \ref{lem_good_fall},
\[
d_{\mathcal B}(B_1) = d_{\mathcal B_1}(B_1)= d_{\mathcal B_2}(B_2)=d_{\mathcal B}(B_2).
\]
But $d_{\mathcal B}:\mathcal B\to d(\mathcal B)$ is injective, and so $B_1=B_2 \in \mathcal B_1 \cap \mathcal B_2$. On the other hand, if $B_1$ and $B_2$ are not both good sets of $\mathcal B$, $B$ is the $d_{\mathcal B_1}$ image of a bad set of $\mathcal B$ in $\mathcal B_1$, or the $d_{\mathcal B_2}$ image of a bad set of $\mathcal B$ in $\mathcal B_2$. Hence the number of sets in $d(\mathcal B_1)\cap d(\mathcal B_2)$ is at most the number of good sets of $\mathcal B$ in $\mathcal B_1 \cap \mathcal B_2$, plus the number of bad sets of $\mathcal B$ in $\mathcal B_1$, plus the number of bad sets of $\mathcal B$ in $\mathcal B_2$ --- which is precisely $b + |\mathcal B_1 \cap \mathcal B_2|$.\end{proof}
This result has an immediate corollary using Harris's Lemma \cite{Har}, which states that down-sets in the cube are positively correlated. Precisely, if $\mathcal D_1$ and $\mathcal D_2$ are down-sets in $\mathcal P(n)$, then $|\mathcal D_1\cap \mathcal D_2|\ge 2^{-n}|\mathcal D_1||\mathcal D_2|$. Applying this to the down-sets $d(\mathcal B_1)$ and $d(\mathcal B_2)$, and using the fact that $|d(\mathcal B_i)|=|\mathcal B_i|$ for $i=1$ and $2$, we get the following result.
\begin{corollary}\label{cor_lower_b}
Let $\mathcal B$, $\mathcal B_1$ and $\mathcal B_2$ be simply rooted families, with $\mathcal B_1 \cup \mathcal B_2 = \mathcal B$. Let $b$ be the number of bad sets of $\mathcal B$. Then
\[
2^{-n}|\mathcal B_1| |\mathcal B_2| \le b + |\mathcal B_1 \cap \mathcal B_2|.
\]
\qed
\end{corollary}
Next we shall choose simply rooted families $\mathcal B_1$ and $\mathcal B_2$ to which we can apply this result to give a lower bound on the number of bad sets in $\mathcal B$. Recall that for a set $S\subseteq [n]$ and a simply rooted family $\mathcal B$, $\mathcal B_S$ is the family consisting those elements of $B$ which are $\mathcal B$-rooted at some element of $S$. We note that $\mathcal B_S$ is a simply rooted family; if $B$ is $\mathcal B$-rooted at $s\in S$, every set of $[\{s\},B]$ is $\mathcal B$-rooted at $s$ and hence is in $\mathcal B_S$, so $B$ is $\mathcal B_S$-rooted at $s$. We restate Corollary \ref{cor_lower_b} for these simply rooted families.
\begin{lemma}\label{lem_many_bad}
Let $(S,T)$ be a partition of $[n]$ into two disjoint sets, and let $\mathcal B$ be a simply rooted family in $\mathcal P(n)$. Let $b_1$ be the number of sets $B\in \mathcal B\setminus (\mathcal B_S\cap\mathcal B_T)$ with $\delta B\subseteq \mathcal B$, let $b_2 = |\mathcal B_S \cap \mathcal B_T|$, and let $b_3$ be the number of sets $B\in \mathcal B$ with $d_{\mathcal B}(B)=B$. Then
\[
2^{-n}|\mathcal B_S||\mathcal B_T|\le b_1+2b_2+b_3.
\]
\end{lemma}
\begin{proof}
Note that since $\mathcal B$ is a simply rooted family, $\mathcal B_S$ and $\mathcal B_T$ are simply rooted families and $\mathcal B_S\cup \mathcal B_T = \mathcal B$. Also, there are at most $b_1+b_3$ bad sets of $\mathcal B$ not in $\mathcal B_S \cap \mathcal B_T$, and at most $b_2$ in $\mathcal B_S\cap \mathcal B_T$. Hence the total number of bad sets of $\mathcal B$ is at most $b_1+b_2+b_3$, and so by Corollary \ref{cor_lower_b} we have
\[
2^{-n}|\mathcal B_S||\mathcal B_T| \le b_1+b_2+b_3 + |\mathcal B_S\cap \mathcal B_T| = b_1+2b_2+b_3,
\]
as required.
\end{proof}
We note that in fact every set in $\mathcal B_S \cap \mathcal B_T$ is bad --- indeed, any set $B$ which is $\mathcal B$-rooted at two distinct integers has $\delta B \subseteq \mathcal B$. Hence this result gives us a lower bound on the number of bad sets in $\mathcal B$.
To use Lemma \ref{lem_many_bad} to prove Theorem \ref{thm_stability}, we shall pick $S$ and $T$ to make $|\mathcal B_S||\mathcal B_T|$ large. In general, we cannot do well; if, for example, $m \le 2^{n-1}$ and $\mathcal B$ is $\{B+n: B\in \mathcal I(m)\}$, then for any partition $[n]=S\cup T$ one of $\mathcal B_S$ and $\mathcal B_T$ is empty --- which is as we expect, because this family has no bad sets. However, if $\mathcal B_{\{i\}}$ --- that is, the family of sets of $\mathcal B$ which are $\mathcal B$-rooted at $i$ --- is not too large for any $i$, we can easily choose $S$ and $T$ to make $|\mathcal B_S||\mathcal B_T|$ large.
\begin{lemma}\label{lem_large_product}
Let $\mathcal B$ be a simply rooted family in $\mathcal P(n)$ with $|\mathcal B| = m$. Suppose that no $i \in [n]$ has $|\mathcal B_{\{i\}}| > p m$. Then there exists a partition $[n]=S\cup T$ such that
\[
|\mathcal B_S||\mathcal B_T| \ge m^2(1/4 - p^2/4).
\]
\end{lemma}
\begin{proof}
Take the partition $[n]=S\cup T$ where the smaller of $|\mathcal B_S|$ and $|\mathcal B_T|$ is as large as possible --- without loss of generality $|\mathcal B_S|\le |\mathcal B_T|$. If $|\mathcal B_S|< m(1/2-p/2)$, we can move an element $t$ of $T$ to $S$ such that $\min(\mathcal B_S,\mathcal B_T)$ increases, a contradiction. Since $|\mathcal B_S|+|\mathcal B_T|\ge m$, $|\mathcal B_S||\mathcal B_T| \ge (m/2-p/2)(m/2+p/2)=m^2(1/4 - p^2/4)$.
\end{proof}
We are now ready to prove Theorem \ref{thm_stability}. Let $\mathcal B$ be a simply rooted family in $\mathcal P(n)$ with $|\mathcal B| = m$, such that no $i \in [n]$ has $|\mathcal B_{\{i\}}| > p m$. By Lemma \ref{lem_large_product}, there exists a partition $[n]=S\cup T$ such that $|\mathcal B_S||\mathcal B_T|\ge m^2(1/4 - p^2/4)$. We let $b_1$ be the number of sets $B\in \mathcal B\setminus (\mathcal B_S\cap\mathcal B_T)$ with $\delta B\subseteq \mathcal B$, $b_2 = |\mathcal B_S \cap \mathcal B_T|$, and $b_3$ be the number of sets $B\in \mathcal B$ with $d_{\mathcal B}(B)=B$. Then from Lemma \ref{lem_many_bad} we have
\[
2^{-n}m^2(1/4 - p^2/4)\le 2^{-n}|\mathcal B_S||\mathcal B_T| \le b_1+2b_2+b_3,
\]
and so either $b_1+b_2\ge m^2(1/12 - p^2/12)/2^n$ or $b_3\ge m^2(1/12 - p^2/12)/2^n$. In the first case, since $b_1+b_2$ is the number of sets in $\mathcal B$ with $\delta B\subseteq \mathcal B$, by Lemma \ref{lem_full_sh} we have $||\mathcal B||\le ||\mathcal I(m)|| + m - m^2(1/12 - p^2/12)/2^n$. In the second case, from Lemma \ref{lem_no_falls} we also have $||\mathcal B||\le ||\mathcal I(m)|| + m - m^2(1/12 - p^2/12)/2^n$, as required.\hfill\qed
\subsection{Proof of Theorem \ref{thm_main}}
We now prove Theorem \ref{thm_main} from Theorem \ref{thm_stability}, giving us a tighter restriction than Theorem \ref{thm_colex_bound} on (hypothetical) counterexamples to the union-closed conjecture. Let $\mathcal A$ be such a counterexample, with $\mathcal B=\mathcal P(n)\setminus \mathcal A$ and $|\mathcal B| = m$. Our task is to show that $||\mathcal I(m)||>m(n/2-1+c_1)$, for some universal constant $c_1$. If $||\mathcal B_{\{i\}}||$ is small for all $i$, we shall prove this using Theorem \ref{thm_stability}, since if $\mathcal A$ is a counterexample to the union-closed conjecture we have $mn/2 \le ||\mathcal B_{\{i\}}||$. To complete the proof of Theorem \ref{thm_main}, we shall show that if $|\mathcal B_{\{i\}}|$ is large for some $i$ then $||\mathcal I(m)||$ is large. For this, we use the following simple observation.
\begin{lemma}\label{lem_low_degrees}
Let $\mathcal A\subseteq \mathcal P(n)$ be a counterexample to the union-closed conjecture, let $\mathcal B = \mathcal P(n)\setminus \mathcal A$, and let $p\in [0,1/2]$. If some element of $[n]$ is in $m(1/2+p)$ sets of $\mathcal B$ then
\[
||\mathcal I(m)|| > m(n/2 - 1 + p).
\]
\end{lemma}
\begin{proof}
From Equation \eqref{eq_rooted_bound}, we have $||\mathcal B|| \le ||\mathcal I(m)|| + m$. Here, since every element of $[n]$ is in more than $m/2$ sets of $\mathcal B$, we must also have $||\mathcal B|| > (n-1)m/2 + m(1/2+p)$, and the result follows.
\end{proof}
Now we can show that if many sets of $\mathcal B$ are $\mathcal B$-rooted at the same $i\in [n]$ then Theorem \ref{thm_main} holds. We shall use Lemma \ref{lem_low_degrees}, and also Theorem 19 of \cite{BaBoEc}, which we state in a slightly different form.
\begin{theorem}\label{thm_down-set}
Let $\mathcal B\subseteq \mathcal P(n)$ be a simply rooted family with $|\mathcal B|=m$. Suppose the largest down-set contained in $\mathcal B$ is $\mathcal D$. Then $||\mathcal B||\le ||\mathcal I(m)|| + m - |\mathcal D|$.\qed
\end{theorem}
In fact, this theorem is an immediate consequence of Lemma \ref{lem_full_sh}, since every $B\in \mathcal D$ has $\delta B \subseteq \mathcal B$.
\begin{lemma}\label{lem_few_with_root}
Suppose that $\mathcal A\subseteq \mathcal P(n)$ is a counterexample to the union-closed conjecture, let $\mathcal B = \mathcal P(n)\setminus \mathcal A$, and let $p\in [0,1]$. If $|\mathcal B_{\{i\}}| \ge 3p m$ for some $i \in [n]$, then
\[
||\mathcal I(m)|| > m(n/2 - 1 + p).
\]
\end{lemma}
\begin{proof}
We may assume $i=n$. We define
\begin{align*}
\mathcal B^+_n &= \{B\subseteq \mathcal P(n-1):B + n\in \mathcal B\},\\
\mathcal B^-_n &= \{B\subseteq \mathcal P(n-1): B\in \mathcal B\}.
\end{align*}
Also, define $m_{n,+} = |\mathcal B^+_n|$, and $m_{n,-} = |\mathcal B^-_n|$. Since $\mathcal A$ is a counterexample to the union-closed conjecture, $m_{n,+}> m_{n,-}$. If $m_{n,+}>m(1/2+p)$, we are done by Lemma \ref{lem_low_degrees}, so we may assume that $m_{n,+}\le m(1/2+p)$, and hence $m_{n,+}-m_{n,-}\le 2pm$. Then, setting $D_+$ to be the largest down-set contained in $\mathcal B_n^+$, we have $\{B-n: [\{n\},B]\subseteq \mathcal B\}\subseteq \mathcal D_+$, and so $|\mathcal D_+|\ge 3pm \ge m_{n,+}-m_{n,-}+pm$. Applying Theorem \ref{thm_down-set} to $\mathcal B^+_n$ now gives us
\begin{align*}
||\mathcal B||&=||\mathcal B^+_n||+||\mathcal B^-_n||+m_{n,+}\\
&\le ||\mathcal I(m_{n,+})||+m_{n,+}-|\mathcal D_+|+||\mathcal I(m_{n,-})||+ m_{n,-}+m_{n,+}\\
&= ||\mathcal I(m_{n,+})||+||\mathcal I(m_{n,-})|| + m +m_{n,+} - |\mathcal D_+|\\
&\le ||\mathcal I(m_{n,+})||+||\mathcal I(m_{n,-})|| + m + m_{n,-} - pm.
\end{align*}
Now, since $m_{n,+}>m_{n,-}$, by Lemma \ref{lem_colex_sums} we have $||\mathcal I(m_{n,+})||+||\mathcal I(m_{n,-})||+ m_{n,-}\le ||\mathcal I(m)||$, and hence
\[
||\mathcal B|| \le ||\mathcal I(m)||+m-pm.
\]
Since $\mathcal B$ is the complement of a counterexample to the union-closed conjecture, we also have $||\mathcal B|| > mn/2$, and the result follows.
\end{proof}
Putting Theorem \ref{thm_stability} and Lemma \ref{lem_few_with_root} together, we can prove Theorem \ref{thm_main}. Indeed, suppose there is a counterexample $\mathcal A$ to the union-closed conjecture in $\mathcal P(n)$, and let $\mathcal B$ be $\mathcal P(n)\setminus \mathcal A$ with $|\mathcal B| = m$. Suppose that $||\mathcal I(m)|| = m(n/2-1+p)$. Then by Lemma \ref{lem_few_with_root} we have $|\mathcal B_{\{i\}}|\le 3pm$ for every $i\in [n]$. The family $\mathcal B$ is the complement of a union-closed family, and so is simply rooted, so by Theorem \ref{thm_stability} we have
\[
||\mathcal B|| \le ||\mathcal I(m)|| + m - m^2(1/12 - 9p^2/12)/2^n.
\]
However, $||\mathcal B|| > mn/2$, since $\mathcal B$ is the complement of a counterexample to the union-closed conjecture. Hence
\[
||\mathcal I(m)|| = m(n/2-1+p) > m \left(n/2 - 1 + \frac{m(1-9p^2)}{12\cdot 2^n}\right).
\]
Now, $\mathcal P(n)\setminus \mathcal B$ is a counterexample to the union-closed conjecture, so by Corollary \ref{cor_old_bound} we have $m\ge 2^n/3$, and so
\[
pm > m(1/36 - 9p^2/36),
\]
and
\[
36p + 9p^2 > 1.
\]
This is false for all $0\le p \le 1/37$, and so we have that
\[
||\mathcal I(m)|| > m(n/2-1+1/37),
\]
proving Theorem \ref{thm_main} with a bound of $c_1\ge 1/37$. \qed
\section{Bounding $||\mathcal I(m)||$}\label{sec_pf_cor}
In this section we bound $||\mathcal I(m)||$, enabling us to prove Corollary \ref{cor_main}. We will use a result of Cz\'edli, Mar\'oti and Schmidt \cite{CzMa}, which states that for a positive integer $r$ we have $||\mathcal I(m)|| > mr/2$ if and only if $m>2^{r+2}/3$. Here, we shall want a more precise bound for general $m$.
\begin{lemma}\label{lem_colex_total}
Let $r$ and $m$ be positive integers with $r\ge 1$ and $2^r/3\le m \le 2^{r+1}/3$, and write $m = 2^r/3 + m'$. Then
\[
||\mathcal I(m)|| \le m(r/2-1) + 3m'/2.
\]
\end{lemma}
\begin{proof}
We prove this by induction on $r$ --- we deduce the assertion for $r$ from those for $r-1$ and $r-2$. For $r=1$ or $2$ the result is easy to check. For $r\ge 3$, first suppose that $m \ge 2^{r-1}$. Since $m \le 2^{r+1}/3$, we have
\begin{align*}
||\mathcal I(m)|| \le m(r/2-1/2)= m(r/2-1) + m/2.
\end{align*}
Also, $m'\ge m/3$, so the result follows. Otherwise, write $m = 2^{r-2} + k$, where $2^{r-2}/3 \le k < 2^{r-2}$. If $k \ge 2^{r-1}/3$, we set $k = 2^{r-1}/3 +k'$ and use the induction hypothesis;
\begin{align*}
||\mathcal I(m)|| &= (r/2-1)2^{r-2} + k +||\mathcal I(k)||\\
&\le (r/2-1)2^{r-2} + k+ k(r/2-3/2) + 3k'/2\\
&= (r/2-1)m - k/2 + 3(k-2^{r-1}/3)/2,
\end{align*}
while $m' = k - 2^{r-2}/3$. Hence we need that for all $2^{r-2}/3 \le k < 2^{r-2}$,
\[
3/2( k - 2^{r-2}/3) \le k/2 + 3(k-2^{r-1}/3)/2,
\]
which does indeed hold. Finally, if $k<2^{r-1}/3$ we have $k = 2^{r-2}/3 +m'$, and by the induction hypothesis we have
\begin{align*}
||\mathcal I(m)|| &= (r/2-1)2^{r-2} + k +||\mathcal I(k)||\\
&\le (r/2-1)2^{r-2} + k + k(r/2-2) + 3m'/2\\
&= (r/2-1)m + 3m'/2,
\end{align*}
as required.
\end{proof}
In fact, we have equality in Lemma \ref{lem_colex_total} whenever $m$ is of the form $2^{a}+2^{a-2}+\dots+2^{a-2j} + 2^{a-2j-1}$ for some integers $a$ and $j$ with $a>0$, $j\ge 0$ and $a-2j-1>0$. We can now prove Corollary \ref{cor_main}. If $\mathcal A$ is a counterexample to the union-closed conjecture in $\mathcal P(n)$, and $\mathcal B = \mathcal P(n)\setminus \mathcal A$ with $|\mathcal B|=m$, then write $m = 2^n/3 + m'$. Then from Theorem \ref{thm_main} and Lemma \ref{lem_colex_total} we have
\begin{align*}
m(n/2-1)+3m'/2 &\ge ||\mathcal I(m)||\\
&\ge m(n/2-1+1/37),
\end{align*}
and so $3m'/2 \ge (2^n/3+m')/37$, which rearranges to $m'\ge \frac{2}{327}2^n$, and Corollary \ref{cor_main} follows with a bound of $c_2\ge \frac{2}{327}$.\qed
\section{Improving the constants}\label{sec_refine}
In this section, we give a modification to the arguments in Section \ref{sec_pf_thm} which improves the constants in our main theorems. To do this, we give stronger versions of Lemmas \ref{lem_split_rooted} and \ref{lem_many_bad}.
For a triple of simply rooted families $\mathcal B$, $\mathcal B_1$, $\mathcal B_2$ with $\mathcal B=\mathcal B_1\cup \mathcal B_2$,
\[
Z(\mathcal B, \mathcal B_1, \mathcal B_2) = \{B\in \mathcal B_1\cap \mathcal B_2: d_{\mathcal B}(B),\, d_{\mathcal B_1}(B) \textrm{ and } d_{\mathcal B_2}(B) \textrm{ are all distinct}\}.
\]
The definition of $Z(\mathcal B, \mathcal B_1, \mathcal B_2)$ is motivated by the proof of Lemma \ref{lem_split_rooted}. The sets in $Z(\mathcal B,\mathcal B_1,\mathcal B_2)$ are those sets $B$ for which $d_{\mathcal B_1}(B)$ and $d_{\mathcal B_2}(B)$ may be distinct sets of $d(\mathcal B_1)\cap d(\mathcal B_2)$, so if we can bound $|Z(\mathcal B,\mathcal B_1,\mathcal B_2)|$ we can improve our bound on $|d(\mathcal B_1)\cap d(\mathcal B_2)|$.
\begin{lemma}\label{lem_split_rooted_2}
Let $\mathcal B$, $\mathcal B_1$ and $\mathcal B_2$ be simply rooted families, with $\mathcal B_1 \cup \mathcal B_2 = \mathcal B$. Let $b$ be the number of bad sets of $\mathcal B$. If every set $B\in \mathcal B_1 \cap \mathcal B_2$ has $\delta B \subseteq \mathcal B$, then
\[
|d(\mathcal B_1)\cap d(\mathcal B_2)| \le b + |Z(\mathcal B, \mathcal B_1, \mathcal B_2)|.
\]
\end{lemma}
\begin{proof}
The proof is similar to that of Lemma \ref{lem_split_rooted}. Letting $\mathcal D_i=d(\mathcal B_i)$, consider an element $S$ of $\mathcal D_1\cap \mathcal D_2$. Then $S=d_{\mathcal B_1}(B_1)=d_{\mathcal B_2}(B_2)$ for some $B_1\in \mathcal B_1$ and $B_2 \in \mathcal B_2$. We now define a function $f:\mathcal D_1\cap \mathcal D_2\to \mathcal B$. If $B_1=B_2$, then we set $f(S)=B_1$ --- note that since $B_1\in \mathcal B_1\cap \mathcal B_2$, $\delta B \subseteq \mathcal B$ and so $B_1$ is a bad set of $\mathcal B$. Otherwise, since $d_{\mathcal B}$ is injective, for $i=1$ or $2$ we have $d_{\mathcal B_i}(B_i)\neq d_{\mathcal B}(B_i)$. In this case, we define $f(S)=B_i$ --- note that since $d_{\mathcal B_i}(B_i)\neq d_{\mathcal B}(B_i)$, by Lemma \ref{lem_good_fall} the set $B_i$ is a bad set of $\mathcal B$. So $f(S)$ is a bad set of $\mathcal B$ for all $S\in \mathcal D_1\cap \mathcal D_2$. Also, for $S\neq T \in \mathcal D_1\cap \mathcal D_2$, if $f(S) = f(T)=B$ then
\[
S=d_{\mathcal B_i}(B)\neq d_{\mathcal B}(B) \neq d_{\mathcal B_j}(B)=T,
\]
where $\{i,j\}=\{1,2\}$. In particular $B\in Z(\mathcal B,\mathcal B_1, \mathcal B_2)$, and there is no $U\in \mathcal D_1\cap \mathcal D_2$ with $S\neq U\neq T$ and $f(U)=B$. Hence the size of the image of $f$ is at least $|\mathcal D_1\cap \mathcal D_2|-|Z(\mathcal B,\mathcal B_1,\mathcal B_2)|$, and since every set in the image is a bad set of $\mathcal B$ the result follows.
\end{proof}
We can now prove a stronger form of Lemma \ref{lem_many_bad}, using Lemma \ref{lem_split_rooted_2} and making sure we do not overcount the bad sets of $\mathcal B$. For a family of sets $\mathcal B$ we define
\[
Y(\mathcal B) = \{B\in \mathcal B: \delta B\subseteq \mathcal B \textrm{ and } d_{\mathcal B}(B)=B\}.
\]
The sets in $Y(\mathcal B)$ are those that satisfy both criteria for a set to be bad; we have often overcounted the number of bad sets of $\mathcal B$ by $|Y(\mathcal B)|$.
\begin{lemma}\label{lem_many_bad_2}
Let $(S,T)$ be a partition of $[n]$ into two disjoint sets, and $\mathcal B$ be a simply rooted family in $\mathcal P(n)$. Let $b_1$ be the number of sets $B\in \mathcal B\setminus (\mathcal B_S\cap\mathcal B_T)$ with $\delta B\subseteq \mathcal B$, $b_2 = |\mathcal B_S \cap \mathcal B_T|$, and $b_3$ be the number of sets $B\in \mathcal B$ with $d_{\mathcal B}(B)=B$. Then
\[
2^{-n}|\mathcal B_S||\mathcal B_T|\le b_1 + b_2 + b_3 + |Z(\mathcal B, \mathcal B_S, \mathcal B_T)| - |Y(\mathcal B)|.
\]
\end{lemma}
\begin{proof}
The proof is identical to that of Lemma \ref{lem_many_bad} --- letting $b$ be the number of bad sets of $\mathcal B$, by Harris's Lemma and Lemma \ref{lem_split_rooted_2} we have
\[
2^{-n}|\mathcal B_S||\mathcal B_T|\le b + |Z(\mathcal B, \mathcal B_S, \mathcal B_T)|,
\]
and $b = b_1+b_2+b_3-|Y(\mathcal B)|$.
\end{proof}
We shall show that in fact $|Y(\mathcal B)| \ge |Z(\mathcal B, \mathcal B_S, \mathcal B_T)|$, improving our bound on the number of bad sets of $\mathcal B$. For this, we shall use the key lemma of Reimer \cite{Rei} on up-compressions of union-closed families. For a family $\mathcal A\subseteq \mathcal P(n)$, a set $A\in \mathcal A$ and an element $i\in [n]$, we define
\[
u_{(i,\mathcal A)}(A)=
\begin{cases}
A+i: i\notin A,\, A+i\notin \mathcal A\\
A: \mathrm{ otherwise.}
\end{cases}
\]
Then $u(\mathcal A)$, $u_{\mathcal A}(A)$, $u_i(\mathcal A)$, $U_i(\mathcal A)$ and $U_{(\mathcal A,i)}(A)$ are defined analagously to in the case of down-compressions. In particular, $u(\mathcal A) = u_1\dots u_n (\mathcal A)$, and $u_{\mathcal A}(A)$ is the image of the set $A$ in $u(\mathcal A)$ under the sequence of up-compressions $u_1\dots u_n$.
\begin{lemma}\label{lem_Rei}
If $\mathcal A$ is a union-closed family, and $A_1\neq A_2$ are sets in $\mathcal A$, the cubes $[A_1,u_{\mathcal A}(A_1)]$ and $[A_2,u_{\mathcal A}(A_2)]$ are disjoint.\qed
\end{lemma}
We make a simple observation about the relationship between sets of a simply rooted family which lose an element under the down-compression $d_{\mathcal B}$, and the sets of the union-closed family $\mathcal P(n)\setminus \mathcal B$.
\begin{lemma}\label{lem_uc_image}
Let $\mathcal B\subseteq \mathcal P(n)$ be a simply rooted family, let $\mathcal A = \mathcal P(n)\setminus \mathcal B$, and let $B\in \mathcal B$. If $d_{\mathcal B}(B)\neq B$ then for some $1\le k \le n$ and some $A\in \mathcal A$ we have $U_{(\mathcal A,k)}(A)=B$.
\end{lemma}
\begin{proof}
Let $k$ be minimal with $D_{(\mathcal B,k)}(B)\neq B$. Then $D_{(\mathcal B,k)}(B) = B-k$, and $B-k \notin D_{k-1}(B)$. Hence $B-k \in \mathcal P(n)\setminus D_{k-1}(B) = U_{k-1}(\mathcal A)$, and so $B-k = U_{(\mathcal A,k-1)}(A)$ for some $A\in \mathcal A$, and $B = U_{(\mathcal A,k)}(A)$.
\end{proof}
For a simply rooted family $\mathcal B$, and a set $B\in \mathcal B$, let $R_{\mathcal B}(B)=\{r\in [n]:[\{r\},B]\subseteq \mathcal B\}$ be the set of roots of $B$ in $\mathcal B$. We now prove that if $B$ is in some cube $[A,U_\mathcal A(A)]$, we must have $A= B\setminus R_{\mathcal B}(B)$.
\begin{lemma}\label{lem_cube_set}
Let $\mathcal B\subseteq \mathcal P(n)$ be a simply rooted family, and let $B\in \mathcal B$. Let $\mathcal A= \mathcal P(n)\setminus \mathcal B$. If $B$ is in the cube $[A,U_\mathcal A(A)]$ for some $A\in \mathcal A$, then $A= B\setminus R_\mathcal B(B)$.
\end{lemma}
\begin{proof}
Let $R= R_\mathcal B(B)$. First we observe that $B\setminus R \in \mathcal A$. Indeed, suppose $B\setminus R\in \mathcal B$; then it is $\mathcal B$-rooted at some $b\in B\setminus R$. But then we have $\{B'\subseteq B: B'\cap R \neq \emptyset\}\subseteq \mathcal B$, and $\{B'\subseteq B\setminus R: b\in B'\}\subseteq \mathcal B$. Hence $B$ is $\mathcal B$-rooted at $b$, and so $b\in R$, a contradiction as $b\in B\setminus R$.
Now, since $B\in [A,U_\mathcal A(A)]$, we have $A\subseteq B$. However, $\{B'\subseteq B: B'\cap R \neq \emptyset\}$ is contained in the family $\mathcal B$, and hence we have $A\subseteq B\setminus R$. In particular, $B\setminus R \in [A,B]\subseteq [A,U_\mathcal A(A)]$. Hence the cubes $[A,U_\mathcal A(A)]$ and $[B\setminus R,U_\mathcal A(B\setminus R)]$ intersect, and so from Theorem \ref{lem_Rei} we have $A= B\setminus R$.
\end{proof}
Using the last two lemmas, it is immediate that if a set $B\in \mathcal B$ loses an element $r$ under the down-compression $d_\mathcal B$, then $B$ is $\mathcal B$-rooted at $r$.
\begin{lemma}\label{lem_root_fall}
Let $\mathcal B\subseteq \mathcal P(n)$ be a simply rooted family, and let $B\in \mathcal B$. Then $d_{\mathcal B}(B)\in \{B\}\cup\{B-r:r\in R_{\mathcal B}(B)\}$.
\end{lemma}
\begin{proof}
Let $\mathcal A = \mathcal P(n)\setminus \mathcal B$. Suppose $d_{\mathcal B}(B)\neq B$ --- then by Lemma \ref{lem_rooted_basics}, $d_{\mathcal B}(B)=B-b$ for some $b\in B$. Also, by Lemma \ref{lem_uc_image} we have that for some $k$ and some $A\in \mathcal A$ we have $U_{(\mathcal A, k)}(A)=B$. In particular, $B\in [A,u_{\mathcal A}(A)]$, and so $A=B\setminus R_{\mathcal B}(B)$. Since $d_{\mathcal B}(B) \in [A,B]$, we then have $b \in B\setminus A = R_{\mathcal B}(B)$, as required.
\end{proof}
In the special case where $B-b\notin \mathcal B$ for some $b\in B$, we must have $R_{\mathcal B}(B)=\{b\}$, and so Lemma \ref{lem_fall_b} is a special case of Lemma \ref{lem_root_fall}. We read out the following corollary on the number of roots of sets in $Z(\mathcal B, \mathcal B_S, \mathcal B_T)$.
\begin{corollary}\label{cor_Z_roots}
Let $\mathcal B$ be a simply rooted family, $S\cup T$ a partition of $[n]$, and $B\in Z(\mathcal B, \mathcal B_S, \mathcal B_T)$. Then $|R_{\mathcal B}(B)|\ge 2$. If $d_{\mathcal B}(B)\neq B$, then $|R_{\mathcal B}(B)|\ge 3$.
\end{corollary}
\begin{proof}
By Lemma \ref{lem_root_fall} the sets $d_{\mathcal B_S}(B)$, $d_{\mathcal B_T}(B)$ and $d_{\mathcal B}(B)$ are all elements of $\{B\}\cup\{B-r:r\in R_{\mathcal B}(B)\}$. But $B\in Z(\mathcal B, \mathcal B_S, \mathcal B_T)$, so these sets are all distinct, and in particular, $|R_{\mathcal B}(B)|\ge 2$. If $d_{\mathcal B}(B)\neq B$, then by Lemma \ref{lem_smaller_falls} we also have $d_{\mathcal B_1}(B)\neq B\neq d_{\mathcal B_2}(B)$, so the sets $d_{\mathcal B_1}(B)$, $d_{\mathcal B_2}(B)$ and $d_{\mathcal B}(B)$ are distinct elements of $\{B-r:r\in R_{\mathcal B}(B)\}$ and $|R_{\mathcal B}(B)|\ge 3$.\end{proof}
Now we shall prove that $|Y(\mathcal B)| \ge |Z(\mathcal B, \mathcal B_S, \mathcal B_T)|$. For a finite set $B$ we define the \emph{$2$nd shadow of $B$} to be $\delta_2 B = \{B'\subseteq B: |B'|=|B|-2\}$.
\begin{lemma}\label{lem_Y_ge_Z}
Let $\mathcal B$ be a simply rooted family, and $S\cup T$ a partition of $[n]$. Then $|Y(\mathcal B)| \ge |Z(\mathcal B, \mathcal B_S, \mathcal B_T)|$.
\end{lemma}
\begin{proof}
We write $Z= Z(\mathcal B, \mathcal B_S, \mathcal B_T)$, and $Y=Y(\mathcal B)$. Let $\mathcal A$ be $\mathcal P(n)\setminus \mathcal B$; $\mathcal A$ is a union-closed family, since $\mathcal B$ is simply rooted. If a set $B\in Z$ is not in a cube $[A,u_{\mathcal A}(A)]$ for some $A\in \mathcal A$, then $B$ is also in $Y$. Indeed, $\delta B\subseteq \mathcal B$ because all sets in $Z$ are $\mathcal B$-rooted at two distinct elements of $[n]$. $d_{\mathcal B}(B)=B$ follows from Lemma \ref{lem_uc_image}; otherwise we must have $B=U_{(\mathcal A,k)}(A)$ for some $A\in \mathcal A$ and $1\le k \le n$, so $B\in [A,u_{\mathcal A}(A)]$, a contradiction. Hence it is enough to show that for every cube $C=[A,u_{\mathcal A}(A)]$,
\[
|C\cap Y| \ge |C\cap Z|,
\]
since by Theorem \ref{lem_Rei} these cubes are disjoint for different $A$. We shall now show this for the cube $C$. If $C\cap Z\subseteq C\cap Y$, we are done. Otherwise, let $B\in (C\cap Z) \setminus Y$. Since $B\in Z$, $B$ has at least two roots in $\mathcal B$ by Corollary \ref{cor_Z_roots}, so $\delta B\subseteq \mathcal B$. Then since $B\notin Y$, $d_{\mathcal B}(B)\neq B$, and $B$ has at least $3$ roots in $\mathcal B$ by Corollary \ref{cor_Z_roots}. Hence $\delta_2(B)\subseteq \mathcal B$, and so $|u_{\mathcal A}(A)\setminus A| \ge |B\setminus A| \ge 3$. We define $r=|u_{\mathcal A}(A)\setminus A|$.
We now count the sets of $C\setminus Y$. Note that by Lemma \ref{lem_Rei}, $C$ contains no set of $\mathcal A$ other than $A$. In $C$, there are $r+1$ sets which are $U_{(\mathcal A,k)}(A)$ for some $0\le k \le n$, one of size $i$ for each $i$ with $|A|\le i \le |u_{\mathcal A}(A)|$. All other sets $B\in C$ are in $\mathcal B$, and have $d_{\mathcal B}(B)=B$. Also, every set $B$ in $C$ of size at least $|A|+2$ has $\delta B\subseteq \mathcal B$. Indeed, if $i\in B\cap A$, $B - i \in C\setminus\{A\}\subseteq \mathcal B$. If $i\in B\setminus A$, then $(B-i)\cup A = B$, and $\mathcal A$ is union-closed, so $B-i\in \mathcal B$.
Hence $|C\cap Y| = 2^r - 2r$ --- the elements of $C\setminus Y$ are precisely the $r+1$ sets of $C$ of size $|A|$ or $|A|+1$, together with one set of size $i$ for each $i$ with $|A|+2\le i \le |A|+r$. To bound $|C\cap Z|$, we note that $A$ is not in $Z$, and nor is $A+i$ for any $i\in (u_{\mathcal A}(A)\setminus A)$, since $A+i$ has only one $\mathcal B$-root. Also, if $A+i+j$ is in $Z$, for $i\neq j$ both in $u_{\mathcal A}(A)\setminus A$, then since $A+i+j$ is not $\mathcal B$-rooted at any element in $A$ by Corollary \ref{cor_Z_roots} we must have $\{d_{\mathcal B}(A+i+j),d_{\mathcal B_S}(A+i+j), d_{\mathcal B_T}(A+i+j)\} = \{A+i+j,A+i,A+j\}$.
However, we must have $d_{\mathcal B}(A+i+j) = A+i+j$; otherwise by Lemma \ref{lem_smaller_falls} $d_{\mathcal B_S}(A+i+j)\neq A+i+j \neq d_{\mathcal B_T}(A+i+j)$, a contradiction. So $\{d_{\mathcal B_S}(A+i+j), d_{\mathcal B_T}(A+i+j)\} = \{A+i,A+j\}$. Without loss of generality, $d_{\mathcal B_S}(A+i+j)=A+i$. Then by Lemma \ref{lem_root_fall} we have $j\in R_{\mathcal B_S}(A+i+j)$, and so $j \in S$. Similarly, $i\in T$.
Now, suppose $A+i+k\in Z$ for some $k\neq j$. Then, as before, $\{d_{\mathcal B_S}(A+i+k), d_{\mathcal B_T}(A+i+k)\} = \{A+i,A+k\}$, and since $i\in T$ we must have $k\in S$ and $d_{\mathcal B_S}(A+i+k)=A+i$, contradicting the injectivity of $d_{\mathcal B_S}$. Hence each element of $u_{\mathcal A}(A)\setminus A$ appears in at most one of size $|A|+2$ in $C\cap Z$. So $C\cap Z$ does not contain $A$, nor any of the $r$ sets of size $|A|+1$ in $C$, and contains at most $\lfloor r/2\rfloor$ of the $\binom{r}{2}$ sets of size $|A|+2$ in $C$. So the total number of sets in $C\cap Z$ is at most $2^r-1-r-\binom{r}{2} +\lfloor r/2\rfloor$. It is easy to see that for $r\ge 3$ this is at most $2^r - 2r$, with equality when $r=3$. Hence $|Y|\ge |Z|$, as required.\end{proof}
\noindent
Combining Lemmas \ref{lem_many_bad_2} and \ref{lem_Y_ge_Z}, we get the following lemma.
\begin{lemma}\label{lem_refinement}
Let $(S,T)$ be a partition of $[n]$ into two disjoint sets. Also, let $\mathcal B$ be a simply rooted family in $\mathcal P(n)$. Let $b_1$ be the number of sets $B\in \mathcal B\setminus (\mathcal B_S\cap\mathcal B_T)$ with $\delta B\subseteq \mathcal B$, let $b_2 = |\mathcal B_S \cap \mathcal B_T|$, and let $b_3$ be the number of sets $B\in \mathcal B$ with $d_{\mathcal B}(B)=B$. Then
\[
2^{-n}|\mathcal B_S||\mathcal B_T|\le b_1 + b_2 + b_3.
\]
\hfill\qed
\end{lemma}
This result is a stronger version Lemma \ref{lem_many_bad}, and using it instead of that lemma improves the constant in Theorem \ref{thm_stability}, giving the following result.
\begin{theorem}\label{thm_stability_2}
Let $\mathcal B$ be a simply rooted family in $\mathcal P(n)$ with $|\mathcal B| = m$, and let $p\in[0,1]$. Suppose that no $i \in [n]$ has $|\mathcal B_{\{i\}}| \ge pm$. Then
\[
||\mathcal B|| \le ||\mathcal I(m)|| + m - m^2(1/8 - p^2/8)/2^n.
\]
\end{theorem}
\noindent
\begin{proof}
Indeed, by Lemma \ref{lem_large_product} we can choose a partition $[n]=S\cup T$ so that $|\mathcal B_S||\mathcal B_T|\ge m^2(1/2-p^2/4)$. Then we have
\[
2^{-n}m^2(1/2-p^2/4)\le 2^{-n}|\mathcal B_S||\mathcal B_T|\le b_1+b_2+b_3,
\]
and so either $b_1+b_2$ or $b_3$ is at least $m^2(1/8-p^2/8)/2^n$. Applying Lemma \ref{lem_full_sh} in the first case or Lemma \ref{lem_no_falls} in the second, we get Theorem \ref{thm_stability_2}.\end{proof}
This in turn improves the constants in Theorem \ref{thm_main} and Corollary \ref{cor_main}. We also note another minor change to the proof of Theorem \ref{thm_main} --- at the end of the proof, we use the fact that a counterexample to the union-closed conjecture in $\mathcal P(n)$ has fewer than $\frac{2}{3}2^n$ elements. Since we now have a better bound, we can use this instead to improve the argument slightly. Applying these improvements together improves our bound in Theorem \ref{thm_main} to $c_1\ge 0.04218\ldots > 1/24$ and in Corollary \ref{cor_main} to $c_2 \ge 0.009646\ldots > 1/104$ --- that is, the union-closed conjecture holds for families in $\mathcal P(n)$ with at least $(\frac{2}{3}-\frac{1}{104})2^n$ elements.
\section{Further Work}
Theorem \ref{thm_stability_2} is a stability result for the total sizes of simply rooted families, which in turn provides a stability result for the union-closed size problem in the case of large union-closed families; if $\mathcal A\subseteq \mathcal P(n)$ is union-closed, with $|\mathcal A|\ge 2^{n-1}$ and $||\mathcal A||$ is close to the minimum possible, then $\mathcal P(n)\setminus \mathcal A$ has an element of high degree. However, we have no stability result for the union-closed size problem in general. It was proved in \cite{BaBoEc} that there is a unique uinon-closed family $\mathcal F_m$ with $|\mathcal F_m|=m$ and $||\mathcal F_m|| = f(m)$, but if $\mathcal A$ is a union-closed family of $m$ sets with $||\mathcal A||$ close to $||\mathcal F_m||$ in a large powerset, we have no result (or even conjecture) which states that $\mathcal A$ is in some sense similar to $\mathcal F_m$.
Another direction would be to improve our stability results for the sizes of simply rooted families. For example, it was conjectured in \cite{BaBoEc} that if $\mathcal B\subseteq \mathcal P(n)$ is a simply rooted family then
\begin{equation}\label{eq:con1}
||\mathcal B|| \le ||\mathcal I(m)|| + \max_{i\in[n]} \mathrm{deg}_\mathcal B(i).
\end{equation}
This remains open, but we conjecture a stronger result still; that we can replace the maximum of the degrees $d_\mathcal B(i)$ with the largest number of elements of $\mathcal B$ rooted at a single element of $[n]$:
\begin{conjecture}\label{con:max_rooted}
Let $\mathcal B$ be a simply rooted family in $\mathcal P(n)$ then
\[
||\mathcal B|| \le ||\mathcal I(m)|| + \max_{i\in[n]} |\mathcal B_{\{i\}}|.
\]
\end{conjecture}
Even if these conjectures do not hold, it seems likely that some version of Theorem \ref{thm_stability} which does not depend on $n$ is true. To be precise, we conjecture that there are some positive constants $\epsilon$ and $\delta$ such that if $\mathcal B\subseteq \mathcal P(n)$ is a simply rooted family of $m$ sets, and $|\mathcal B_{\{i\}}|\le \epsilon m$ for all $i\in [n]$, then
\[
||\mathcal B|| \le ||\mathcal I(m)|| + m(1-\delta).
\]
\section{Acknowledgements}
The author would like to thank B\'ela Bollob\'as for his helpful comments on earlier versions of this paper.
|
\section{Introduction}
The Small Magellanic Cloud (SMC) is a dwarf irregular galaxy connected by a hydrogen gas and stellar bridge to the Large Magellanic Cloud (LMC). Both are nearby satellites of our Milky Way galaxy and because of this proximity they are ideal environments within which to study stellar populations and to calibrate many different stellar standard candles. The SMC is more metal-poor than the LMC, and this provides an important opportunity to check on the metallicity dependencies of the different standard candles used in distance scale work. This makes the SMC very important in establishing a precise distance scale despite its complicated geometrical structure and larger extension in the line-of-sight compared to the LMC. For these reasons, the SMC has played an important role in our Araucaria Project \citep{gie05} whose main goal is to produce a significantly improved calibration of the extragalactic distance scale.
For almost a decade we have been observing late-type eclipsing binary systems in both Magellanic Clouds, systems which have a particularly strong potential to derive precise distances these galaxies \citep{pie09,gra12}. Recent observations have proven this potential: analysis of eight binaries in the LMC has led to a distance to the LMC accurate to 2.2\%, more accurate than any other previous determination of this fundamental parameter \citep{pie13}. Here we extend this work to a measurement of the distance to the SMC using the same method, leading in addition to a better knowledge of the geometrical structure of this galaxy. From our studies of late-type detached eclipsing binary systems in the SMC and LMC we also expect to improve on the determination of the astrophysical parameters of red giant stars and Cepheid variables \citep{pie10,pil13}.
The advantage of using detached late type systems over those containing early type stars comes from utilizing well determined empirical surface brightness relation (e.g.~the Barnes-Evans relation in the optical) derived for late type stars from interferometric measurements (e.g.~\cite{bar76}, \cite{ker04}, \cite{ben05}). The main points are well summarized by \cite{lac77}: "The distances are derived without assumptions about luminosity class or effective temperature, and are ultimately based only on geometrical factors. That means that the only source for systematic error should be from the Barnes-Evans calibration and this is the place where we expect the most significant improvement will occur. Another wirtue is that for stars later than A0 the Barnes-Evans distance is essentially independent of reddening."
The paper is organized as follows. In Section 2 we describe photometric and spectroscopic observations, Section 3 gives details of our method, Section 4 contains a description of model solutions for individual binaries. In section 5 we present a distance determination to the SMC and Section 6 is devoted to a discussion of our results.
\section{Observations and Data Reduction}
\label{obs}
Basic data on our four eclipsing binary systems are given in Table~\ref{tbl:1}. Their position within the body of the SMC is presented in Fig.~\ref{fig0}. With the exception of SMC130.5 4296 our targets were discovered by \cite{uda98} and confirmed to be eclipsing binaries by \cite{wyr04}. Optical photometry in the Johnson-Cousins filters was obtained with the Warsaw 1.3 m telescope at Las Campanas Observatory in the course of the second, third and fourth phase of the OGLE project \citep{uda97,uda03,sos12}. We secured 5314 I-band measurements (over 1000 per system) and 391 V-band measurements (over 70 per system) for five eclipsing binaries. Because of the long orbital periods consecutive epochs were taken usually on different nights. The time span of the I-band observations is 5684 days. The raw data were reduced with the image-subtraction technique \citep{woz00,uda03} and instrumental magnitudes were calibrated onto the standard system using Landolt standards.
Near-infrared photometry was collected with the ESO NTT telescope on La Silla, equipped with the SOFI camera (PI: Pietrzy{\'n}ski). The setup of the instrument, reduction and calibration of the data onto the UKIRT system were essentially the same as described in \cite{pie09}. We collected at least 10 epochs of infrared photometry for every our target outside eclipses. The transformation of our photometry onto the Johnson system was done using the equations given by \cite{car01} and \cite{bes88}. The whole multiband photometry is presented in Table \ref{tbl:phot}.
High resolution echelle spectra were collected with the Clay 6.5 m telescope at Las Campanas, equipped with the MIKE spectrograph and with the 3.6 m telescope in La Silla, equipped with the HARPS spectrograph. We used a $5\times0.7$ arcsec slit with MIKE giving a resolution of about 40000. In the case of HARPS we used the EGGS mode giving a resolution of about R$\sim$80000. The typical S/N at $\sim$5300 \AA~was about 20 and 6 for the MIKE-Red part and HARPS spectra, respectively, and about 12 at $\sim$4700 \AA~for the MIKE-Blue part spectra.
In order to determine radial velocities of the system's components we employed the Broadening Function (BF) formalism introduced by \citet{ruc92,ruc99}. Radial velocities were derived using the RaveSpan software \citep{pil12} using numerous metallic lines in the wavelength regions 4125-4230, 4245-4320, 4350-4840, 4880-5000, 5350-5850, 5920-6250, 6300-6390, 6600-6800 \AA. As templates we used synthetic spectra with [Fe/H]=$-0.5$ from a library computed by \cite{col05}. The templates were chosen to match the atmospheric properties of the stars in a grid of $T_{\rm eff}$ and $\log g$. The resulting velocities are presented in Table \ref{tbl:spec}. We calculated instrumental shifts taking into account residual radial velocities from our best models. They are: MIKE-RED$-$MIKE-BLUE$= +3\pm85$ m s$^{-1}$ and MIKE-BLUE$-$HARPS$= +48\pm114$ m s$^{-1}$. We concluded that there is no significant systematic shift in radial velocities from spectra taken by HARPS and MIKE spectrographs. In the analysis we included separately radial velocities from both the blue and red sides of MIKE, effectively including the MIKE velocities at twice the weight of the HARPS velocities. The reason is that the MIKE spectra usually have significantly better S/N compared to the HARPS spectra.
\begin{deluxetable*}{@{}lcclcccc}
\tabletypesize{\scriptsize}
\tablecaption{The target stars \label{tbl:1}}
\tablewidth{0pt}
\tablehead{
\colhead{OGLE ID} & \colhead{RA} & \colhead{Dec} & \colhead{$V$} & \colhead{$V-I$} & \colhead{$V-K$} & \colhead{$J-K$} & \colhead{$P_{\rm obs}$} \\
\colhead{} & \colhead{h:m:s} & \colhead{deg:m:s} & \colhead{mag} & \colhead{mag} & \colhead{mag} & \colhead{mag} & \colhead{d}
}
\startdata
SMC101.8 14077 & 00:48:22.70 & -72:48:48.6 & 17.177 & 0.935 & 2.099 & 0.489 & 102.90\\
SMC108.1 14904 & 01:00:18.10 & -72:24:07.1 & 15.205 & 0.963 & 2.194 & 0.556 &185.22 \\
SMC126.1 210 & 00:44:02.68 & -72:54:22.5 & 16.771 & 1.249 & 2.941 & 0.765 & 635.00 \\
SMC130.5 4296 & 00:33:47.90 & -73:04:28.0 & 16.783 & 1.207 & 2.788 & 0.754 & 120.47
\enddata
\tablecomments{Given are coordinates, observed magnitudes, colors and orbital periods. Identification numbers are from the OGLE-III database \citep{uda03}.}
\end{deluxetable*}
\begin{deluxetable}{@{}lcrcc}
\tabletypesize{\scriptsize}
\tablecaption{The photometry of target systems \label{tbl:phot}}
\tablewidth{0pt}
\tablehead{
\colhead{Name} & \colhead{Band} & \colhead{HJD} & \colhead{Mag} & \colhead{Err} \\
\colhead{} & \colhead{} & \colhead{HJD-2450000} & \colhead{mag} & \colhead{mag}
}
\startdata
smc101 & I & 621.80276 & 16.228 & 0.010 \\
smc101 & I & 622.87983 & 16.235 & 0.008 \\
smc101 & I & 624.90639 & 16.252 & 0.008 \\
smc101 & I & 625.89383 & 16.265 & 0.008 \\
smc101 & I & 626.86536 & 16.258 & 0.008 \\
smc101 & I & 627.85158 & 16.252 & 0.008
\enddata
\tablecomments{This table is avalaible entirety in machine-readable format in the online journal. A portion is shown here for guidance regarding its form and content.}
\end{deluxetable}
\begin{deluxetable*}{@{}lrccccl}
\tabletypesize{\scriptsize}
\tablecaption{The radial velocities measurements \label{tbl:spec}}
\tablewidth{0pt}
\tablehead{
\colhead{Name} & \colhead{HJD} & \colhead{RV1} & \colhead{Err1} & \colhead{RV2} & \colhead{Err2} &\colhead{Instrument} \\
\colhead{} & \colhead{HJD-2450000} & \colhead{km s$^{-1}$} & \colhead{km s$^{-1}$} & \colhead{km s$^{-1}$} & \colhead{km s$^{-1}$} & \colhead{}
}
\startdata
smc101 & 4328.78143 & 224.764 & 0.300 & 153.453 & 0.300 & MIKE-Red \\
smc101 & 4328.78161 & 224.407 & 0.300 & 153.465 & 0.300 & MIKE-Blue \\
smc101 & 4329.81804 & 225.409 & 0.300 & 152.038 & 0.300 & MIKE-Blue \\
smc101 & 4329.81804 & 225.577 & 0.300 & 151.729 & 0.300 & MIKE-Red \\
smc101 & 4395.57940 & 150.271 & 0.300 & 221.691 & 0.300 & HARPS \\
smc101 & 4424.58869 & 212.609 & 0.300 & 163.422 & 0.300 & MIKE-Blue
\enddata
\tablecomments{This table is avalaible entirety in machine-readable format in the online journal. A portion is shown here for guidance regarding its form and content.}
\end{deluxetable*}
\begin{figure}
\includegraphics[angle=0,scale=.575]{f1.eps}
\caption{ The positions of the eclipsing binaries within the SMC. North is up, east on the left. The filled circles denote four systems from this paper and one system previously analyzed by \cite{gra12} on upper left - SMC113.3 4007. The open circles represent three additional systems for which the analysis is not yet finished. \label{fig0}}
\end{figure}
\section{Modeling approach}
The light and radial velocity curves were analyzed using the Wilson-Devinney (hereafter WD) program, version 2007 \citep{wil71,wil79,wil90,van07}. We follow a methodology of analysis as described in Graczyk et al. (2012) and the Supplementary Information section of Pietrzy{\'n}ski et al. (2013). Here we give some additional details of our approach. The analysis can be divided into "preliminary" and "final" steps. During the "preliminary" analysis we use all light curves, radial velocities, infrared colors and a first estimation of the reddening to set initial constraints on model parameters and to find a preliminary solution. This solution is used to disentangle the spectra. In the "final" analysis we add information from spectral disentangling to improve input model parameters and derive the final solution.
\subsection{Parameter Choice}
\label{par}
\begin{deluxetable}{lcc}
\tabletypesize{\scriptsize}
\tablecaption{The intensity ratios \label{tbl-2}}
\tablewidth{0pt}
\tablehead{
\colhead{ID} & \colhead{Blue ($\sim 4500$ \AA)} & \colhead{Red ($\sim 6000$ \AA)}
}
\startdata
SMC101.8 14077 & $1.37 \pm 0.05$ & $1.61 \pm 0.05$ \\
SMC108.1 14904 & $0.66 \pm 0.02$ & $1.00 \pm 0.02$ \\
SMC126.1 210 & $0.95 \pm 0.02$ & $0.91 \pm 0.02$\\
SMC130.5 4296 & $2.1 \pm 0.1$ & $2.6 \pm 0.3$
\enddata
\tablecomments{The relative strengths of the absorption lines of the secondary with respect to those of the primary.}
\end{deluxetable}
\begin{deluxetable*}{@{}lcccccccc}
\tabletypesize{\scriptsize}
\tablecaption{Atmospheric parameters\label{tbl-4}}
\tablewidth{0pt}
\tablehead{
\colhead{Property} &\multicolumn{2}{c}{{SMC101-8-14077}}& \multicolumn{2}{c}{{SMC108-1-14904}}& \multicolumn{2}{c}{{SMC126-1-210}}& \multicolumn{2}{c}{{SMC130-5-4296}} \\
&\colhead{Primary} & \colhead{Secondary} &\colhead{Primary} & \colhead{Secondary} &\colhead{Primary} & \colhead{Secondary} &\colhead{Primary}&\colhead{Secondary}
}
\startdata
$T_{\rm eff}$ (K) & - & 5080 &5350&5200&4500&4540&5050& 4420\\
$\log{g}$ (cgs) & - & 1.20 &0.40 &2.20 &1.45 &1.10 &2.45 &1.75\\
$[$Fe/H$]$ & - &-1.01 &-0.95&-0.64&-0.94&-0.79&-0.77&-0.99\\
$v_t$ (km $s^{-1}$) & - &1.95 &3.50 &2.40 &1.60 &1.79&1.67 &2.08
\enddata
\end{deluxetable*}
The orbital periods were calculated with the string-length method (Lafler \& Kinman 1965, Clarke 2002). The phased light curves were inspected visually to set a preliminary epoch of the primary minimum and to verify the photometric stability of the systems - a lack of spots, flares, the O'Connell effect (O'Connell 1951) - at the precision level of our photometry.
The effective temperature of the primary component of each system (eclipsed during minimum at orbital phase 0) was estimated as follows. We determined the interstellar extinction in the direction of our targets using \cite{has11} reddening maps - see Section~\ref{red}. Average out-of-eclipse magnitudes were calculated from all observations taken outside photometric minima. In the case of the two systems showing proximity effects between eclipses we used magnitudes only around both quadratures. All out-of-eclipse V-band magnitudes and ($V\!-\!I$) colors are presented in Table~\ref{tbl:1}. These magnitudes were dereddened using the interstellar extinction law given by Cardelli et al. (1989) and O'Donnell (1994) and assuming R$_V=3.1$.
To set the effective temperature scale of each system we ran the WD code initially assuming a temperature of $T_1=5000$ K for the primary and [Fe/H]$\,=-0.5$. The resulting luminosity ratios in the V, I and K bands were combined with dereddened magnitudes to calculate first guesses for the intrinsic ($V\!-\!I$) and ($V\!-\!K$) colors of both components. Then we utilized several calibrations between colors and effective temperatures \citep{ben98,alo99,hou00,ram05,mas06,gon09,cas10,wor11} to set new temperatures for the primaries. We iterated these steps until new derived temperatures changed by less than 10 K. The resulting effective temperatures were used in the "preliminary" solution.
We estimated rotational velocities of the components from a Broadening Function analysis. Rotation in all cases is consistent with synchronous rotation, thus we set the rotation parameter $F=1.0$ for both components. The albedo parameter was set to 0.5 and the gravity brightening to 0.32, both values appropriate for a cool, convective atmosphere. The limb darkening coefficients were calculated internally by WD code (setting LD=$-2$) according to the logarithmic law \citep{kli70} during each iteration of the Differential Correction (DC) subroutine using tabulated data computed by \cite{van93}.
As free parameters of the WD model we chose the orbital period $P_{\rm obs}$, the semimajor axis $a$, the orbital eccentricity $e$, the argument of periastron $\omega$, the epoch of the primary spectroscopic conjunction $T_0$, the systemic radial velocity $\gamma$, the orbital inclination $i$ , the secondary star average surface temperature $T_2$, the modified surface potential of both components ($\Omega_1$ and $\Omega_2$), the mass ratio $q$, and the relative monochromatic luminosity of the primary star in two bands ($L1_V$ and $L1_I$).
\subsection{Fitting Procedure}
\label{fit}
For each star we simultaneously fitted the I-band and V-band light curves and the two radial velocity curves using the DC subroutine from the WD code. The detached configuration (Mode 2) was chosen during all the analyses and a simple reflection treatment (MREF=1, NREF=1) was employed. A stellar atmosphere formulation was selected for both stars (IFAT=1). Level dependent weighting was applied (NOISE=1) and curve dependent weightings (SIGMA) were calculated after each iteration. The grid size was set to N=40 on both components. We did not break the adjustable parameters set into subsets, but instead we adjusted all free parameters at each iteration.
For each system we used information obtained from spectroscopy to derive the relative intensity of absorption lines of both components from the MIKE spectra. We treat these as approximate light ratios in the blue and red part of the optical spectra. These light ratios are reported in Table~\ref{tbl-2}. We do not use them to fix light ratios in our models, however, but use them to rule out alias solutions and to verify if the light curve solution is consistent with the spectroscopy.
\subsection{Spectral disentangling}
\label{specdi}
To constrain the important input parameters of the eclipsing binary components we decided to disentangle the spectra of the individual components and carry out an abundance analysis. The most important parameter is the effective temperature of each component which scales the luminosity ratio of the components and the limb darkening coefficients. Also, by calculating the intrinsic ($V\!-\!I$) colors from the derived temperatures, we can estimate the reddening of each system.
We used the prescription given by Gonzales \& Levato (2005) to disentangle individual spectra of the binary components. The method works in the real wavelength domain. It requires rather high signal to noise ratio spectra to work properly and thus we were restricted to using only the red MIKE spectra. The red MIKE spectra were shifted and stacked using the radial velocities found by using the Broadening Function as described in the previous section. No initial secondary spectrum was used in the disentangling. The range of wavelength used was from 4960 \AA ~to 6800 \AA ~where numerous FeI and FeII lines are present. We emphasise that we do not use the disentangled spectra to derive new radial velocities because they always give higher dispersion than synthetic templates. The reason is that the relatively low S/N ratio of our disentangled spectra reduces the accuracy of radial velocity measurements.
The disentangled spectra have to be renormalized in order to account for the companion's continuum which dilutes the depth of the absorption lines. If the light ratio of the secondary to the primary component at a given wavelength is $k(\lambda)=L_{21}(\lambda)$ then we subtract a value $k/(1+k)$ from the disentangled spectrum of the primary (corresponding to the secondary flux) and we again normalize the final spectrum. In the case of the secondary spectrum we subtract $1/(1+k)$. Here two things must be emphasized: 1) the choice of a proper normalization level before disentangling and 2) the choice of a proper model of a system to calculate an appropriate light ratio function $L_{21}(\lambda)$. Regarding the first point, all spectra have to be normalized and for high S/N spectra the level of normalization is simply the mode value. In the case of lower S/N spectra like those obtained for our binaries in the SMC, the level of normalization is a bit ambiguous and usually we set it to a value of $\sim$90 percentile.
Regarding the second point, we must be aware that there is a large number of possible model solutions to the observed light and radial velocity curves. A model which is the best i.e. it minimizes the residua of synthetic fits does not always predict light ratios in agreement with spectroscopic light ratios. This is because our data have limited S/N ratio and because an optical depth were absorption lines are formed is usually different from an optical depth at which we observe stellar photosphere. If we use an "improper" model for a system to predict the spectroscopic light ratios, the resulting absorption line depths in the disentangled spectra are invalid. An easy way to notice an inappropriate model is to check if the absorption line bottoms in the renormalized spectrum of the primary or the secondary are below zero, which is a level signifying an unphysical model. Another check is based on a comparison of equivalent widths of the same absorption lines in both components. Usually the hotter component should have lines with smaller equaivalent widths. However such a comparison is not unique because components of the same binary may have different abundances \citep{fol10}. Finally, we can calculate the equivalent width ratios for the components for the same lines and determine that this ratio is not wavelength dependent. We expect that for realistic models, this ratio is approximately constant within the spectral range of disentangling.
\subsection{Atmospheric parameters analysis}
\label{atmo}
\begin{deluxetable*}{lccccccc}
\tabletypesize{\scriptsize}
\tablecaption{Reddenings from the Na I D1 line \label{tbl:redd}}
\tablewidth{0pt}
\tablehead{\colhead{ID} &\multicolumn{3}{c}{{Galactic component }} &\multicolumn{3}{c}{{SMC component}} & \colhead{Total} \\
\colhead{} &\colhead{$V_r$} &\colhead{Eqv.} &\colhead{E($B\!-\!V$)} &\colhead{$V_r$} &\colhead{Eqv.} &\colhead{E($B\!-\!V$)}&\colhead{E($B\!-\!V$)}\\
\colhead{} &\colhead{(km s$^{-1}$)} & \colhead{(\AA)} &\colhead{(mag)} &\colhead{(km s$^{-1}$)} & \colhead{(\AA)} &\colhead{(mag)} & \colhead{(mag)}
}
\startdata
SMC101.8 14077 &21&0.155&0.049&148&0.091&0.029&0.078\\
SMC108.1 14904 &10&0.171&0.055&129&0.095&0.030&0.085\\
SMC126.1 210 &13&0.235&0.084&$-$&$ -$&$-$&0.084\\
SMC130.5 4296 &6&0.118&0.037&107&0.082&0.026&0.063
\enddata
\end{deluxetable*}
Atmospheric parameters and the iron content were obtained from the equivalent widths (EWs) of the iron spectral lines. See \cite{mar08} for a more detailed explanation of the method we used to measure the EWs and \cite{vil10} for a description of the line list that was used. We adopted $\log{\epsilon}(Fe)=7.50$ as the solar iron abundance.
Atmospheric parameters were obtained in the following way. As a first step, model atmospheres were calculated using ATLAS9 \citep{kur70} assuming as initial estimations for $T_{\rm eff}$, $\log g$, and $v_t$ values typical for giant stars (4800 K, 2.00 dex, 1.80 km s$^{-1}$), and [Fe/H]$=-0.8$ dex as a typical value for young stars in the SMC \citep{dia10}.
Then $T_{\rm eff}$, $v_t$, and $\log g$ were adjusted and new atmospheric models calculated in an interactive way in order to remove trends in excitation potential (EP) and EWs versus abundance for $T_{\rm eff}$ and $v_t$, respectively, and to satisfy the ionization equilibrium for $\log g$. FeI and FeII were used for this purpose. The [Fe/H] value of the model was changed at each iteration according to the output of the abundance analysis. The local thermodynamic equilibrium program MOOG \citep{sne73} was used for the abundance analysis.
The derived atmospheric parameters for the eclipsing binary components are given in Table~\ref{tbl-4}. For all systems with the exception of SMC101.8 14077 we manage to calculate parameters for both components. For SMC101.8 14077 the iron lines of the primary are too weak to carry out an abundance analysis. Our measurements of abundance are consistent with each component in a binary having the same abundance. The typical accuracy of the parameters are 75 K, 0.4 dex, 0.15 dex and 0.3 km s$^{-1}$ for $T_{\rm eff}$, $\log g$, [Fe/H] and $v_t$, respectively.
\subsection{Reddening}
\label{red}
The interstellar extinction in the direction to each of our target stars was derived in three ways. First, we utilized Magellanic Cloud reddening maps published by Haschke et al. (2011). We calculate an average E($V\!-\!I$) value from all of the reddening estimates within a 2 arc min radius of our stars and divided this average by a factor of 1.28 to get a ($B\!-\!V$) color excess for each system. We then added $\Delta$E($B\!-\!V$)$\,=0.037$ mag, the mean foreground Galactic reddening in the direction of the SMC as derived from the Schlegel et al.(1998) maps. The second column of Table~\ref{tbl-5} gives the resulting reddening to each eclipsing binary. These were used to set the temperature scale of each system and to derive the "preliminary" solution (see Section~\ref{par}).
\begin{deluxetable}{@{}lcccc}
\tabletypesize{\scriptsize}
\tablecaption{Color Excess E($B\!-\!V$) \label{tbl-5}}
\tablewidth{0pt}
\tablehead{\colhead{ID} &\colhead{Haschke} &\colhead{Na I D1} & \colhead{Atmosphere} &\colhead{Adopted}\\
\colhead{} &\colhead{(mag)} &\colhead{(mag)} & \colhead{(mag)} &\colhead{(mag)}
}
\startdata
SMC101.8 14077 &0.076&0.078&0.046&0.067\\
SMC108.1 14904 &0.087&0.085&0.107&0.093\\
SMC126.1 210 &0.060&0.084&0.097&0.080\\
SMC130.5 4296 &0.072&0.063&0.101&0.079
\enddata
\end{deluxetable}
\begin{figure}
\includegraphics[angle=0,scale=.50]{f2.eps}
\caption{The MIKE spectrum of SMC101.8 14077 in the region of the Na I lines (5890.0, 5895.6 \AA) taken close to the third orbital quadrature. Narrow absorption lines arise from galactic (GAL) and SMC (SMC) interstellar matter. The stellar lines of the primary (P) and secondary (S) are also denoted. \label{fig:redd}}
\end{figure}
The second method is based on a calibration of the equivalent width of the interstellar Na I D1 line (5890.0 \AA) and reddening (Munari \& Zwitter 1997). The calibration works best for relatively small values of reddening, E($B\!-\!V$)$\,<0.4$ mag. Figure~\ref{fig:redd} presents the spectrum of SMC101.8 14077 around the sodium doublet. Narrow interstellar absorption lines from the Galaxy and the SMC can be identified together with wider stellar absorption from both stars. We separately measured the equivalent widths of both interstellar components and derived the reddening for each. The results for all our stars is presented in Table~\ref{tbl:redd}. In the system SMC126-1-210 we could not detect any SMC component of the Na I D1 line because of strong blending with stellar lines.
\begin{figure*}
\begin{minipage}[th]{0.5\linewidth}
\includegraphics[angle=0,scale=.51]{f3a.eps} \vspace{-1.17cm}
\mbox{}\\
\includegraphics[angle=0,scale=.51]{f3b.eps}
\end{minipage}\hfill
\begin{minipage}[th]{0.5\linewidth}
\includegraphics[angle=0,scale=.51]{f3c.eps} \vspace{-1.17cm}
\mbox{}\\
\includegraphics[angle=0,scale=.51]{f3d.eps}
\end{minipage}\hfill
\caption{ The radial velocity solutions to four eclipsing binaries in the SMC. Filled circles correspond to the velocities of the primaries. The continuous line is the synthetic fit from the WD analysis. Distortions near the photometric eclipses are caused by the Rossiter-McLaughlin effect. The dashed lines signify the systemic velocities.\label{fig3}}
\end{figure*}
The final method uses the effective temperatures we derived from the atmospheric analysis described in Section~\ref{atmo}. From the effective temperature - ($V\!-\!I$) color calibrations of Houdashelt et al. (2000) and Worthey \& Lee (2011) we estimated the intrinsic ($V\!-\!I$) colors of each component. These colors were compared with the observed colors of the components obtained from the preliminary solution to directly derive E($V\!-\!I$) color excesses. The reddening to each system was calculated as the mean value of the two components with the exception of the system SMC101.8 14077 where we have at our disposal a reddening estimate from only the secondary star.
Each of the three methods has an accuracy of approximately 0.03 - 0.04 mag. We calculated an average reddening for each system from the three estimates, and used this new reddening estimate to update the temperature scale of the components. We then calculated new models and repeated the reddening derivation using the third method. The fourth column of Table~\ref{tbl-5} gives the reddening estimates after two such iterations. The fifth column presents the adopted E($B\!-\!V$) for each system as used in the "final" solution. We assigned a statistical error of 0.02 mag and an additional 0.02 mag systematic error for each estimate of the reddening.
\section{Adopted solutions}
\label{fin}
\begin{figure*}
\begin{minipage}[th]{0.5\linewidth}
\includegraphics[angle=0,scale=.51]{f4a.eps} \vspace{-1.17cm}
\mbox{}\\
\includegraphics[angle=0,scale=.51]{f4b.eps}
\end{minipage}\hfill
\begin{minipage}[th]{0.5\linewidth}
\includegraphics[angle=0,scale=.51]{f4c.eps} \vspace{-1.17cm}
\mbox{}\\
\includegraphics[angle=0,scale=.51]{f4d.eps}
\end{minipage}\hfill
\caption{ The I-band light curve solutions to four eclipsing binaries in the SMC.\label{fig1}}
\end{figure*}
After the last reddening estimate we set the temperature of the primary in the way described in Section \ref{par} and we then recalculated all of the models. We call these the "basic" set of models. The temperature scale consistency of each eclipsing binary was checked by computing the distance to each system resulting from a scaling of the bolometric flux observed at Earth. To calculate the bolometric corrections we used an average from several calibrations \citep{cas10,mas06,alo99,flo96}. These distances were then compared to the distances computed with a surface brightness - color relation. In all cases both estimates of the distances agree within the errors, thus confirming the proper temperature scales.
Finally we computed three additional sets of models by adjusting: 1) the four coefficients of the linear law of limb darkening (LB=+1) for both stars and both light curves, 2) the third light $l_3$ in two bands, and 3) the third light and linear coefficients of limb darkening together - six more free parameters. We compared the model basic set with these new sets to find the physical model with the lowest reduced $\chi^2_{\rm r}$. Models with unphysical values of the third light ($l_3\!<\!0$) and/or with very high limb darkening coefficients ($x\!>\!1.1$) were excluded even if they produced a better formal fit with lower $\chi^2_{\rm r}$. Table~\ref{tbl-3} lists the parameters of the best model for each system. The quoted uncertainties are errors calculated with the DC routine. The quantity $P_{\rm orb}$ signifies the rest frame orbital period.
Solutions to the radial velocity curves are presented in Fig.~\ref{fig3}. The I-band and V-band light curve solutions are presented in Figures~\ref{fig1} and~\ref{fig2}, respectively. Figure~\ref{fig6} shows some details of the shapes of the eclipses. The absolute dimensions are reported in Table~\ref{tbl-6}. These were calculated following \cite{gra12} adopting the same values for physical constants. The spectral type of each component was estimated according to its effective temperature using a calibration by \cite{alo99}. We now comment on individual systems and particular model solutions.
\begin{figure*}
\begin{minipage}[th]{0.5\linewidth}
\includegraphics[angle=0,scale=.51]{f5a.eps} \vspace{-1.17cm}
\mbox{}\\
\includegraphics[angle=0,scale=.51]{f5b.eps}
\end{minipage}\hfill
\begin{minipage}[th]{0.5\linewidth}
\includegraphics[angle=0,scale=.51]{f5c.eps} \vspace{-1.17cm}
\mbox{}\\
\includegraphics[angle=0,scale=.51]{f5d.eps}
\end{minipage}\hfill
\caption{ The V-band light curve solutions to four eclipsing binaries in the SMC.\label{fig2}}
\end{figure*}
\begin{deluxetable*}{@{}lcccc}
\tabletypesize{\scriptsize}
\tablecaption{Model parameters from the Wilson-Devinney code\tablenotemark{a} \label{tbl-3}}
\tablewidth{0pt}
\tablehead{
\colhead{Parameter} & \colhead{SMC101.8 14077} &\colhead{SMC108.1 14904}&\colhead{SMC126.1 210}&\colhead{SMC130.5 4296}
}
\startdata
{\bf Orbital param.} &&&&\\
$P_{\rm obs}$ (days) &$102.8984 \pm 0.0003$ & $185.2176 \pm 0.0017$ &$634.999 \pm 0.009$ &$120.4681 \pm 0.0009$ \\
$T_0$ (HJD-2450000) &$2818.913 \pm 0.006$&$641.081 \pm 0.010$&$2858.30 \pm 0.04$&$2052.976 \pm 0.009$ \\ $a\sin{i}$ (R$_\odot$) &$163.69 \pm 0.47$ &$277.61 \pm 0.45$ &$464.35 \pm 1.90$&$157.10 \pm 0.40$ \\
$q=M_2/M_1$ &$1.0404 \pm 0.0062$ &$1.0028 \pm 0.0030$ &$0.9971 \pm 0.0078$ &$0.9736 \pm 0.0048$ \\
$\gamma$ (km s$^{-1}$) &$187.76 \pm 0.11$ &$178.42 \pm 0.07$&$104.82 \pm 0.08$&$148.70 \pm 0.09$\\
$e$ & $0$ & $0$ &$0.0422 \pm 0.0018$ & 0\\
$\omega$ (deg)& $90$ & $90$ &$54.4 \pm 1.7$& 90\\
&&&&\\
{\bf Photometric param.} &&&&\\
$i$ (deg) & $88.04 \pm 0.23$ & $78.87 \pm 0.10$ &$86.92 \pm 0.09$ &$83.09 \pm 0.13$ \\
$T_2/T_1$ &$0.9269 \pm 0.0014$ &$0.8701 \pm 0.0054$&$1.0065 \pm 0.0009$ &$0.9194 \pm 0.0055$ \\
$r_1$ & $0.1093 \pm 0.0012$ & $0.1660 \pm 0.0016$ &$0.0936 \pm 0.0011 $&$0.1608 \pm 0.0012$\\
$r_2$ & $ 0.1458 \pm 0.0012$ & $0.2265 \pm 0.0014$&$0.0839 \pm 0.0016 $&$0.2908 \pm 0.0015$\\
&&&&\\
$(L2/L1)_V$ & $1.205 \pm 0.010$ & $0.928 \pm 0.013$&$0.838 \pm 0.016$&$1.939 \pm 0.015$ \\
$(L2/L1)_I$ & $1.349 \pm 0.008$ &$ 1.169 \pm 0.013$ &$0.827 \pm 0.016$&$2.370 \pm 0.009$\\
$(L2/L1)_K$ & $1.620$\tablenotemark{b} &$1.572$\tablenotemark{d} &0.810\tablenotemark{b}&2.925\tablenotemark{d}\\
$x_{1,V}$ &- &$0.873 \pm 0.116$&-&$0.764 \pm 0.100$ \\
$x_{2,V}$ &- &$0.874 \pm 0.106$&-&$0.873 \pm 0.048$\\
$x_{1,I}$ &-&$0.461 \pm 0.050$&-&$0.597 \pm 0.070$ \\
$x_{2,I}$ &-&$0.425 \pm 0.045$&-&$0.562 \pm 0.030$ \\
$l_{3,V}$ & $0.005 \pm 0.018$ & 0&0& 0\\
$l_{3,I}$ & $0.043 \pm 0.017$&0&0& 0\\
$l_{3,K}$ &0.043\tablenotemark{c} &0&0&0\\
&&&&\\
{\bf Derived quantities} &&&&\\
$P_{\rm orb}$ (days) &$102.8340 \pm 0.0003$&$185.1076 \pm 0.0017$&$634.777 \pm 0.009$ &$120.4084 \pm 0.0009$\\
$a$ (R$_\odot$) &$163.69 \pm 0.47$&$282.76 \pm 0.45$&$464.86 \pm 1.90$&$158.17 \pm 0.40$ \\
$K_1$ (km s$^{-1}$) &$41.03 \pm 0.20$&$37.96 \pm 0.08$&$18.48 \pm 0.11$&$32.54 \pm 0.12$\\
$K_2$ (km s$^{-1}$) &$39.44 \pm 0.12$&$37.85 \pm 0.09$&$18.54 \pm 0.10$&$33.42 \pm 0.11$\\
rms$_{1}$ (km s$^{-1}$) &0.84& 0.52& 0.44& 0.56 \\
rms$_{2}$ (km s$^{-1}$) &0.52 & 0.57& 0.42& 0.51 \\
&&&&\\
$(r_1+r_2)$ &$ 0.2551$& $0.3926 $&0.1775&0.4516\\
$k=r_2/r_1$& $1.3332$& $ 1.3642 $&0.8961&1.808\\
$(j_2/j_1)_V$ &0.678&0.499&1.044&0.593\\
$(j_2/j_1)_I$ &0.759&0.628&1.029&0.725
\enddata
\tablenotetext{a}{Simultaneous solution of V-band and I-band light curves together with radial velocity curves of both components. Uncertainties quoted are the standard errors from Differential Corrections subroutine.}
\tablenotetext{b}{Extrapolated from the WD model}
\tablenotetext{c}{Assumed}
\tablenotetext{d}{The resulting light ratio assuming the same distance to both stars}
\end{deluxetable*}
\begin{figure*}
\begin{minipage}[th]{0.5\linewidth}
\includegraphics[angle=0,scale=.5]{f6a.eps} \vspace{-.17cm}
\mbox{}\\
\includegraphics[angle=0,scale=.5]{f6b.eps}
\end{minipage}\hfill
\begin{minipage}[th]{0.5\linewidth}
\includegraphics[angle=0,scale=.5]{f6c.eps} \vspace{-.17cm}
\mbox{}\\
\includegraphics[angle=0,scale=.5]{f6d.eps}
\end{minipage}\hfill
\caption{ The close-up of the eclipses of individual eclipsing binaries in the SMC. Continuous red line is a best-fit model from the WD program, vertical dashed lines mark times of mid-eclipse.\label{fig6}}
\end{figure*}
\begin{deluxetable*}{@{\hspace{-12pt}}lcccccccc}
\tabletypesize{\scriptsize}
\tablecaption{Physical Properties of the SMC Eclipsing Binaries \label{tbl-6}}
\tablewidth{0pt}
\tablehead{
\colhead{Property\tablenotemark{a,b}} &\multicolumn{2}{c}{{SMC101.8 14077}}& \multicolumn{2}{c}{{SMC108.1 14904}}& \multicolumn{2}{c}{{SMC126.1 210}}& \multicolumn{2}{c}{{SMC130.5 4296}} \\
&\colhead{Primary} & \colhead{Secondary} &\colhead{Primary} & \colhead{Secondary} &\colhead{Primary} & \colhead{Secondary} &\colhead{Primary} &\colhead{Secondary}
}
\startdata
Spectral Type & G0 III & G4 III &F9 II& G7 II&K2 III&K1 III&G7 III & K1 III\\
$V$ (mag) & $18.041\pm0.013$ & $17.838\pm0.013$ &$15.918\pm0.010$&$15.999\pm0.010$&$17.432\pm0.011$&$17.624\pm0.011$&$17.954\pm0.014$&$17.235\pm0.012$\\
\mbox{$V\!-\!I$} (mag) & $0.825\pm0.016$ & $0.947\pm0.016$ &$0.835\pm0.014$&$1.086\pm0.014$&$1.256\pm0.015$&$1.241\pm0.015$&$1.058\pm0.017$&$1.276\pm0.016$\\
\mbox{$V\!-\!K$} (mag) & $1.828\pm0.020$ & $2.149\pm0.020$ &$1.839\pm0.018$&$2.411\pm0.018$&$2.916\pm0.017$&$2.879\pm0.017$&$2.432\pm0.022$&$2.878\pm0.020$\\
\mbox{$J\!-\!K$} (mag) & $0.481\pm0.025$ & $0.575\pm0.025$ &$0.536\pm0.021$&$0.662\pm0.021$&$0.846\pm0.024$&$0.836\pm0.024$&$0.765\pm0.027$&$0.852\pm0.025$\\
Mass (M$_\sun$) & $2.725\pm 0.034$ & $2.835 \pm 0.055$ &$4.416 \pm 0.041$&$4.429 \pm 0.037$&$1.674 \pm 0.037$&$1.669 \pm 0.039$ &$1.854 \pm 0.025$&$1.805 \pm 0.027$\\
Radius (R$_\sun$) & $17.90 \pm 0.50$& $23.86 \pm 0.31$&$46.95 \pm 0.53$&$64.05 \pm 0.50$&$43.52 \pm 1.02$ &$39.00 \pm 0.98$&$25.44 \pm 0.25$&$46.00 \pm 0.35$\\
$\log g$ (cgs) & $2.368 \pm 0.029$ & $2.136 \pm 0.019$ &$1.740 \pm 0.014$&$1.472 \pm 0.010$ &$1.385 \pm 0.029$ &$1.479 \pm 0.031$ &$1.895 \pm 0.014$&$1.369 \pm 0.013$\\
$T_{\rm eff}$ (K) & $5580 \pm 95$ & $5170 \pm 90$ &$5675 \pm 105$ &$4955 \pm 90$&$4480 \pm 70$ &$4510 \pm 70$ &$4912 \pm 80 $& $4515 \pm 75$\\
Luminosity (L$_\sun$) & $280 \pm 34$& $365 \pm 35$ &$2055 \pm 200$ &$2220 \pm 195$ &$685 \pm 74$ &$565 \pm 63$ &$338 \pm 28$ &$790 \pm 65$\\
$M_{\rm bol}$ (mag) &$-1.38$ & $-1.67$ &$-3.54$&$-3.63$&$-2.35$&$-2.14$&$-1.57$&$-2.50$\\
$M_V$ (mag) &$-1.24$ & $-1.44$ &$-3.42$&$-3.34$&$-1.81$&$-1.62$&$-1.26$&$-1.98$\\
$\left[{\rm Fe/H}\right]$ (dex)& - &$-1.01$&$-0.95$&$-0.64$&$-0.94$&$-0.79$&$-0.77$&$-0.99$ \\
E(\bv) (mag)& \multicolumn{2}{c}{$0.067 \pm 0.020$}&\multicolumn{2}{c}{$0.093 \pm 0.020$}&\multicolumn{2}{c}{$0.080 \pm 0.020$}&\multicolumn{2}{c}{$0.079 \pm 0.020$}
\enddata
\tablenotetext{a}{Absolute dimensions were calculated assuming: $G=6.673\cdot10^{-8}$ cm$^3$g$^{-1}$s$^{-2}$, $R_\sun=695 600$ km, $M_\sun=1.9888\cdot10^{33}$ g, $T_{\rm eff,\sun}=5777$ K, $M_{bol,\sun}=+4.75$.}
\tablenotetext{b}{The magnitudes and colors are observed values.}
\end{deluxetable*}
\begin{figure}
\includegraphics[angle=0,scale=.50]{f8.eps}
\caption{ The distance estimates to four new eclipsing binaries in the SMC and the system SMC113.4 4007 published previously by our team. The continuous line is an unweighted mean distance from all five binaries and the dashed lines
represent a 1-$\sigma$ range. The errorbars signify statistical uncertainties. \label{fig8}}
\end{figure}
\begin{figure*}
\begin{minipage}[th]{0.5\linewidth}
\includegraphics[angle=0,scale=.42]{f7a.eps}
\mbox{}\\
\includegraphics[angle=0,scale=.42]{f7b.eps}
\mbox{}\\
\end{minipage}\hfill
\begin{minipage}[th]{0.5\linewidth}
\vspace{-0.25cm}
\includegraphics[angle=0,scale=.42]{f7c.eps}
\mbox{}
\includegraphics[angle=0,scale=.42]{f7d.eps}
\end{minipage}\hfill
\caption{ Reduced $\chi^2$ maps from the Monte Carlo simulations. We calculated about 30 thousand models for each eclipsing binary using the WD code to investigate the dependence of the distance on the model parameters. The horizontal lines correspond to $1\sigma$, $2\sigma$ and $3\sigma$ confidence intervals in the distance modulus. \label{fig11}}
\end{figure*}
\subsection{SMC101.8 14077}
This variable star is the faintest and bluest analyzed in this paper. The best solution for this system was obtained with a third light contribution included. If we assume that the presence of the third light is real and is not connected with some imperfections in the absolute calibration of the OGLE photometric data then this suggests the presence of a red companion or a blended star with an almost negligible contribution to the optical portion of spectrum. Indeed we could not detect any additional source of absorption lines in our spectra. For the purpose of the determination of the distance we assumed that the third light contribution in the K-band is at least equivalent to that in the I-band. The uncertainty in the amount of third light in the K-band is the largest source of statistical error of the absolute stellar radii and the distance determination to the system.
This eclipsing binary has the identifier SC4 192903 in OGLE-II database. It was previously analyzed by \cite{gra03} who used only the photometric data from the second phase of the OGLE project \citep{uda98} without any spectroscopic observations. A comparison of the photometric parameters presented in our Table~\ref{tbl-3} and in Table~2 from \cite{gra03} shows very similar model solutions with the only difference resulting from our inclusion of a contribution from third light. This results in different measured orbital inclinations. Although the stellar mass estimate in \cite{gra03} seems to be too low (see Tab.~3 in that paper), the stellar radii, temperatures and luminosities show good agreement.
\subsection{SMC108.1 14904}
The system contains the most luminous and massive late type giants in the Magellanic Clouds analyzed by our team. The system was analyzed before by \cite{gra03} using data from OGLE-II. As in the case of SMC101.8 14077, our photometric solution is almost identical with that reported by \cite{gra03}, but a difference appears in the absolute dimensions which are significantly underestimated in the previous study. The results of our preliminary work on this system, reported by \cite{gra13}, are not significantly different from those reported in this paper, and based on more extensive observational data we refine the previous distance estimate to this binary.
An interesting feature of this system is the difference of the center-of-mass radial velocities between the components. The hotter primary has systematically blueshifted radial velocities with respect to the secondary star amounting to 0.78 km s$^{-1}$ - see Fig.~3 in \cite{gra13}. We accounted for this effect by subtracting the difference and solving the radial velocities equating the systemic velocity of the secondary with the barycenter velocity of the system. Some possible reasons for the difference are discussed in \cite{tor09} for Capella, a physically similar but lower mass binary system.
Another irregularity is a discrepancy between the temperature ratio $T_2/T_1$ as obtained from the analysis of the atmospheric parameters (see Section~\ref{atmo} and Table~\ref{tbl-4}) compared to that arising from the light curve analysis (Table~\ref{tbl-3}). The light curve solution and the spectroscopic light ratios suggest a temperature difference of about $\Delta T=720$~K, while the atmospheric analysis suggests a difference of $\Delta T=150$~K. We repeated the atmospheric analysis by fixing the gravities of both stars to reliable values obtained from the dynamical solution reported in Table~\ref{tbl-6}. However, we obtained the same small difference of the effective temperatures. We also performed additional analysis by forcing the same metallicity for both components. However, the difference in effective temperatures do not vary again. Probably the problem with this system is the low S/N of the spectrum of the secondary star. While the primary has a nice spectrum and by consequence the parameters are well established, for the secondary both $T_{eff}$ and $v_t$ are uncertain, and by consequence the metallicity. It is worth noting that the mean temperature of the system $(T_1 + T_2)/2=5275$~K obtained from the atmospheric analysis is the same as the mean temperature calculated from color calibrations (i.e. 5315~K) to within the errors.
\subsection{SMC126.1 210}
This system contains two stars with very similar masses and surface temperatures but with quite different radii. It is the only eccentric system in our sample. The system is very well-detached with no proximity effects visible in the light curve, not surprising given its long orbital period. The best solution is found for a logarithmic law of limb darkening, a marginal improvement to that found using a linear law of limb darkening.
An analysis of the broadening of the absorption lines provides an estimate of the ratio of the projected rotational velocities, $v_1/v_2\approx1.1$. This value is consistent with the ratio of the radii and signifies similar periods of rotation for the components (tidal locking of both stars), in spite of their relatively large separation.
\subsection{SMC130.5 4296}
This is only system in our sample with total eclipses. This allows us to determine the stellar radii with a precision of better than 1\% for both components. This eclipsing binary shows quite large proximity effects because of the relatively large secondary component. An interesting feature is the reversed luminosity ratio of the components: the less massive secondary seems to be more evolved and much brighter than the hotter primary component. Taking into account the relative proximity of the stars, we cannot exclude that there has been an episode of mass transfer in the system when the present secondary star filled its Roche-lobe as it evolved along the red giant branch. The present system is detached, and such past mass exchange does not influence our distance determination.
\section{Distance determination}
\label{distan}
\begin{deluxetable*}{lccccccccc}
\tabletypesize{\scriptsize}
\tablecaption{The distance moduli \label{tbl-7}}
\tablewidth{0pt}
\tablehead{
\colhead{ID} & \colhead{$(m\!-\!M)$} & \colhead{$\sigma{\rm A}$}&\colhead{$\sigma({\rm MonteCarlo})$}&\colhead{$\sigma E(B\!-\!V$)} &\colhead{$\sigma V$}&\colhead{$\sigma K$}& \colhead{$\sigma l_{3,K}$}&\colhead{Combined error} & \colhead{Distance} \\
& \colhead{(mag)} &\colhead{(mag)} & \colhead{(mag)}&\colhead{(mag)} &\colhead{(mag)}&\colhead{(mag)}& \colhead{(mag)} &\colhead{(mag)}& \colhead{(kpc)}
}
\startdata
SMC101.8 14077& 19.057 & 0.006 &0.021& 0.012 & 0.001& 0.005& 0.042&0.049 & $64.79 \pm 1.48$\\
SMC108.1 14904 & 19.032 & 0.004&0.014& 0.011& 0.001& 0.004& -&0.019 & $64.03 \pm 0.55$ \\
SMC126.1 210 & 18.979 & 0.009&0.022& 0.008& 0.001& 0.004& -&0.025 & $62.49 \pm 0.74$\\
SMC130.5 4296 & 18.948 & 0.006&0.018& 0.010& 0.002& 0.009& -&0.023 & $61.60 \pm 0.67$
\enddata
\tablecomments{Distance moduli determinations to individual targets together with an error budget. The errors quoted are only statistical uncertainties. Each measurement has additional an systematic uncertainty equal to 0.048 mag. }
\end{deluxetable*}
Distance estimates were derived by employing a calibration of the relation between $V$-band surface brightness and the \mbox{$V\!-\!K$} color by \cite{ben05}. This relation (SBR) was derived from precision measurements of the angular diameters of a sample of giant and dwarf stars using interferometry. The individual distances were calculated according to equations 4 and 5 in \cite{gra12} and the results are summarized in Table~\ref{tbl-7}. The resulting distance modulus is the average of the measured distances of each star of an eclipsing binary. The differences in distance moduli between the components of the same system are very small. The largest discrepancy is 0.003 mag in the case of SMC101.8 14077. The unweighted mean distance modulus from our four systems is (m$-$M)$=19.004$ mag with a dispersion of only 0.050 mag. The new systems clump quite tightly around this value.
To derive the distance to the main body of the SMC galaxy we combined the present estimates with the distance determination to another late type eclipsing binary published by our team \citep{gra12}. Figure~\ref{fig8} shows all our estimates. The system previously analyzed - SMC113.3 4007 - is the closest one. It lies in the north-east part of the SMC (see Fig.~\ref{fig0}) which is, in fact, reported as being closer to Earth than the main body of the galaxy by a number of studies \citep{mat11,sub12,has12}. The unweighted and weighted mean distance moduli from the five systems are different by $\Delta m=0.023$ mag (0.66 kpc), with the former being longer. The weighted mean is dominated by the estimate to SMC113.3 4007 which has the smallest statistical uncertainty but is significantly off of the rest of our sample. A final distance is estimated as follows. We calculated the weighted mean distance modulus from four new binaries as $(m\!-\!M)=18.998 \pm 0.008$ mag. We then combined this result with the fifth system, obtaining $\mu_{SMC}=18.965 \pm 0.025$ mag. The statistical uncertainty of this final estimate arises by combining the standard deviation of the mean (0.008 mag) and the difference $\Delta m$ in quadrature. The resulting uncertainty is dominated by the uncertain structure of the SMC and possible selection effects. The total uncertainty must include an additional systematic error of 0.048 mag (Section~\ref{error}). Our final distance estimate to the SMC is $62.1 \pm 2.0$ kpc.
\begin{deluxetable}{@{}lcc}
\tabletypesize{\scriptsize}
\tablecaption{Distance modulus differences \label{tbl-8}}
\tablewidth{0pt}
\tablehead{
\colhead{Method} & \colhead{$\Delta\mu$ (mag) } & \colhead{Reference}
}
\startdata
Tip of RGB & $0.44 \pm 0.05$ & \cite{cio00} \\
Cepheids & $0.50 \pm 0.10$ & \cite{gro00} \\
Quattuor\tablenotemark{a}&$0.50 \pm 0.05$ & \cite{uda00}\\
Red Clump & $0.47 \pm 0.02$ & \cite{pie03} \\
Tip of RGB & $0.50 \pm 0.03$ & \cite{pie03} \\
Tip of RGB & $0.40 \pm 0.12$ & \cite{sak04}\\
RR Lyrae & $0.39 \pm 0.04$& \cite{sze09} \\
Cepheids &$0.44 \pm 0.12$& \cite{bon10} \\
Red Pulsators & $0.41 \pm 0.02$ & \cite{tab10} \\
Type II Cepheids & $0.39 \pm 0.05$ & \cite{mat11}\\
Cepheids & $0.43 \pm 0.05$ & \cite{mat11} \\
IRSB Cepheids & $0.47 \pm 0.15$ & \cite{sto11}\\
RR Lyr & $0.61 \pm 0.20 $ & \hspace*{-0.4cm} \cite{kop12} \\
IRSB Cepheids & $0.44 \pm 0.06$ & \cite{gro13} \\
FU Cepheids & $0.48 \pm 0.03$ & \cite{ino13} \\
FO Cepheids & $0.52 \pm 0.03$ & \cite{ino13} \\
Eclipsing Binaries\tablenotemark{b} & $0.472 \pm 0.026$ & this paper
\enddata
\tablenotetext{a}{Four methods: average of Red Clump, Tip of RGB, RR Lyr and classical Cepheids}
\tablenotetext{b}{Late type systems.}
\tablecomments{Recent distance moduli differences between Magellanic Clouds based on different methods.}
\end{deluxetable}
\begin{figure}
\includegraphics[angle=0,scale=.50]{f9.eps}
\caption{ The normal distributions of the relative distance moduli between Magellanic Clouds $\Delta\mu\!=\!\mu_{SMC}-\mu_{LMC}$ taken from Table~\ref{tbl-8} - thin lines. The distribution of their sum is shown as points and it is fitted by a single normal distribution, plotted as the thick line. \label{fig10}}
\end{figure}
\subsection{Error Budget}
\label{error}
The main contributions to the statistical uncertainties for all systems are presented in Tab.~\ref{tbl-7} (columns 3-7). The error in the semimajor axis is given as $\sigma{\rm A}$ and an uncertainty of 1\% in the absolute scale of the system translates to a 0.022 mag error in the distance modulus. We performed Monte Carlo simulations as described in \cite{pie13} for each of our binaries - see Fig.~\ref{fig11}. The resulting uncertainty of the distance is reported in column 4. This uncertainty takes into account the error in the determination of the relative radii, magnitude disentangling, and correlations between model parameters in the WD code. The main contribution to the uncertainty returned by the Monte Carlo simulations is the error in the sum of the relative radii $r_1\!+r_2$. The statistical uncertainty in the extinction was assumed to be 0.020 mag in each case. Columns 6 and 7 give the standard deviations of the mean brightness at maximum light in the V- and K-band, respectively. Remaining sources of statistical uncertainty, i.e.: model atmosphere approximations, disentangling of individual magnitudes, the adopted limb darkening law, contribute at a level of 0.001 mag \citep{gra12,pie13}, and these are insignificant in the total statistical error budget.
The sources of systematic uncertainty in our method are the following: the uncertainty of the empirical calibration of surface brightnesses by \cite{ben05} (0.040 mag), uncertainty in the extinction (0.020 mag, translating to a 0.010 mag error in distance), metallicity dependence on the surface brightness - color relation (0.004 mag) and zero point errors of V- and K-band photometry (0.010 mag and 0.015 mag, respectively). Combining these in quadrature we obtain a systematic error of 0.048 mag in the distance modulus of each eclipsing binary.
\section{Discussion}
\subsection{Relative distance between the Magellanic Clouds}
\label{relativ}
\cite{san68} used observations of classical Cepheids to unambiguously show that the Magellanic Clouds have different distance moduli, corresponding to a distance difference of about 11 kpc. This difference is small enough that in practice almost all methods of distance determination can be employed simultaneously to both objects. In fact the relative distance between the galaxies obtained from the application of a particular method is usually better constrained than the individual distances determined by that method. This is because distance determination methods are dominated by systematic uncertainties which mostly cancel out when used to obtain the distance to each galaxy with the same observational setup and similar quality data.
In Table~\ref{tbl-8} we summarize recent determinations of distance moduli differences between the Magellanic Clouds ($\Delta\mu\!=\!\mu_{SMC}-\mu_{LMC}$) by authors who used exactly the same method for both galaxies. Distance moduli based on old population stars are metallicity corrected. In the case of our method we adopted the distance modulus to the LMC from \cite{pie13}, $\mu_{LMC}=18.493\pm0.006$ mag. The interpretation of Table~\ref{tbl-8} is hindered by the fact that some authors do not clearly separate statistical and systematic uncertainties. The average from the young "metal rich" population containing classical Cepheids and eclipsing binary systems is $\Delta\mu=0.472$ mag and the average from the old "metal poor" population is $\Delta\mu=0.453$ mag. We summarize these distributions in Fig.~\ref{fig10}. The resulting distribution is slightly bimodal with a main peak at the position of the young population average. A gaussian fit to the distribution gives $\Delta\mu=0.458 \pm 0.068$ mag.
\subsection{The geometrical depth of the SMC}
The extended structure of the SMC in the line of sight was suggested by \cite{jon61}. Over the last 30 years substantial evidence has been accumulated about the intricate and extended structure of the SMC from the studies of early type stars and HII regions \citep{mar89}, stellar clusters \citep{cro01}, classical Cepheids \citep{wel87,gro00,has12}, RR Lyr stars \citep{sub12,has12}, red clump stars \citep{gar91,hat93,sub09,sub12}, eclipsing binary stars \citep{hil05,nor10} and red giant branch stars \citep{lah05}. The north-east part of the SMC was shown to be closer to us and have a substantial degree of complexity, and it is reported also to have the largest geometrical depth in line-of-sight \citep{sub12}. The apparent difference in the distance to the eclipsing binary SMC113.3 compared to the rest of our sample may confirm the existence of substructure in the front of the galaxy. We can interpret that the remaining eclipsing binaries are part of the main body of the galaxy.
The small dispersion in the distance moduli of our four binaries (0.05 mag, corresponding to $\sim1.4$ kpc) may signify a small geometrical depth of the main part of the SMC. We can investigate this more by an analysis of the distribution of differences in the distance moduli (Fig.~\ref{fig10}). The cumulative distribution has a dispersion of 0.068 mag. Because most of the systematic effects do not influence $\Delta\mu$, and the geometrical depth of the central part of the LMC is small \citep{pie13}, we can interpret this spread as an imprint of the geometrical depth of the main part of the SMC. At a distance of 62 kpc this dispersion corresponds to a 1-$\sigma$ line-of-sight depth of 4 kpc. \cite{has12} gave a review of recent line-of-sight depth estimates for the SMC which we can use for a comparison. They concluded that the 1-$\sigma$ line-of-sight depth is between 4 and 5 kpc for the old population of stars (RR Lyr and red clump stars). For the Cepheids the estimated depth is larger by up to $\sim 8$ kpc, and in the case of intermediate-age clusters the depth reaches $\sim 10$ kpc.
\subsection{Distance to the SMC}
\cite{gra12,gra13} compiled some recent distance determinations to the SMC as reported up to 2011. Their conclusion was that the canonical value of the distance modulus to the SMC ($\mu_{SMC}=18.90$ mag, \citep{wes97}) remained marginally consistent with most recent determinations (see also Section~\ref{relativ}). However, modern determinations prefer a slightly larger distance moduli of $\mu_{SMC}=18.95 \pm 0.07$ mag. The inclusion of the most recent distance determinations \citep{has12,gro13,ino13} does not affect this finding. In fact, if we adopt our estimate of the relative distance modulus between the Magellanic Clouds and we assume $\mu_{LMC}=18.493$ mag, we derive $\mu_{SMC}=18.493+0.458=18.951$ mag. This result remains in perfect agreement with our distance modulus measured with late type eclipsing binaries, and being based on a number of independent standard candles, we advocate here its adoption as the "canonical" distance modulus to the Small Magellanic Cloud.
\acknowledgments
We are gratefull for financial support from Polish National Science Center grant MAESTRO 2012/06/A/ST9/00269 and the TEAM subsidy from the Foundation for Polish Science (FNP). Support from the BASAL Centro de Astrof{\'i}sica y Tecnolog{\'i}as Afines (CATA) PFB-06/2007 is also acknowledged. Based on observations made with ESO 3.6m and NTT telescopes in La Silla under programme 074.D-0318, 074.D-0505, 082.D-0499, 083.D-0549, 084.D-0591, 086.D-0078, 091,D-0469(A) and CNTAC programme CN2010B-060. We also thank the staffs at La Silla Observatory (ESO) and Las Campanas Observatory (Carnegie) for their excellent support and the anonymous referee who helped us to improve this paper.
The OGLE project has received funding from the European Research Council under the European Community's Seventh Framework Programme (FP7/2007-2013)/ERC grant agreement no. 246678 to AU.
WG gratefully acknowledges the hospitality and support of Professor Ralf Bender during his sabbatical leave at Munich Sternwarte and the Max Planck Institute for Extraterrestrial Physics in Garching.
|
\section{Introduction}
Markov chain Monte Carlo (MCMC) simulations
are used in different branches of
statistics
and science
to estimate an expected value
with respect to a probability measure, say $\pi$,
by the sample average
of the Markov chain.
This procedure is of advantage if random numbers with
distribution $\pi$ are difficult to construct.
When sampling the Markov chain the transitions
are usually modeled as driven by
i.i.d. $\mathcal{U}(0,1)^s$ random variables
for some $s\geq1$.
But in simulations the driver sequences are pseudo-random numbers.
In many applications,
if one uses a
carefully constructed random number generator,
this works well. Instead of modeling the Markov chain with random numbers, or imitating
random numbers, the idea of Markov chain quasi-Monte Carlo is to construct
a finite, deterministic sequence of numbers, $(u_i)_{0\leq i \leq n}$ in $[0,1]^s$
for all $n\in\mathbb{N}$, to generate a deterministic Markov chain sample
and to use it to estimate the desired mean.
The motivation
of this conceptual change is that carefully constructed sequences may
lead to more accurate sample averages.
For example, quasi-Monte Carlo (QMC) points lead to higher order of convergence
compared to plain Monte Carlo, which is a special case of MCMC.
Numerical experiments for QMC versions of MCMC
also show promising results \cite{LeSi06,Li98,OwTr05,So74,Tr07}.
In particular, Owen and Tribble \cite{OwTr05} and Tribble \cite{Tr07} report an
improvement by a factor of up to $10^{3}$ and a better convergence
rate for a Gibbs sampler problem.
In the work of Chen, Dick and Owen \cite{ChDiOw11} and Chen \cite{Ch11}
the first theoretical justification for Markov chain quasi-Monte Carlo
on continuous state spaces is provided.
The authors show a consistency result
if
a contraction assumption is satisfied and
the random sequence is substituted
by a deterministic `completely uniformly distributed'
sequence,
see \cite{ChDiOw11,ChMaNiOw12,TrOw08}.
Thus
the sample average converges to the expected value but we do not
know how fast this convergence takes place.
Recently, in \cite{DiRuZh13} another idea appears.
Namely, the question
is considered whether there exists a good driver sequence such that
an explicit error bound is satisfied.
It is shown that if the Markov chain
is uniformly ergodic,
then for any initial state a deterministic sequence exists such that the sample average
converges to the mean almost with the Monte Carlo rate.
However, in \cite{ChDiOw11} and \cite{DiRuZh13} rather strong conditions,
the contraction assumption and uniform ergodicity, are imposed on the Markov chain.
We substantially extend the results of \cite{DiRuZh13} to Markov chains
which satisfy a much weaker convergence condition.
Namely, we consider variance bounding Markov chains,
introduced by Roberts and Rosenthal in \cite{RoRo08},
and show existence results of good driver sequences.
We also show what property pseudo-random number generators need
to satisfy in order to improve the performance of
Markov chain quasi-Monte Carlo
algorithms, see Definition~\ref{def-pull-back-discrepancy} below.
This property has not been studied in the literature before.
In the following we describe the setting in detail and explain
our main contributions.
\subsection{Main results}
Let $(X_n)_{n\in\mathbb{N}}$ be a Markov chain with transition kernel $K$
and initial distribution $\nu$ on $(G,\mathcal{B}(G))$ with $G\subseteq \mathbb{R}^d$.
For $f\colon G \to \mathbb{R}$
let $\mathbb{E}_\pi(f) = \int_G f(x) \pi({\rm d}x)$ be the desired mean
and $P f(x) = \int_G f(y)K(x,{\rm d}y)$
be the Markov operator induced by the transition kernel
$K$.
We assume that the transition kernel
is reversible with respect to the distribution $\pi$
and that it is variance bounding, see \cite{RoRo08}. Roughly, a
Markov chain is variance bounding if the asymptotic variances for functionals
with unit stationary variance are uniformly bounded.
Equivalent to this is the assumption that
$\Lambda < 1$ with
\begin{equation} \label{eq: lambda_intro}
\Lambda = \sup \{ \lambda \in \text{\rm spec}(P-\mathbb{E}_\pi\mid L_2) \}
\end{equation}
where
$L_2=L_2(\pi)$ is the Hilbert space of functions $f\colon G \to \mathbb{R}$
with \[
\|f\|_{2} = \left(\int_G \abs{f(x)}^2 \,\pi(\mathrm{d} x)\right)^{1/2}<\infty
\]
and
$\text{\rm spec} (P-\mathbb{E}_\pi\mid L_2)$
denotes the spectrum of $P-\mathbb{E}_\pi$ on $L_2$.
Let us point out that the
Markov chain does not need to be
uniformly or geometrically ergodic.
In fact, a variance bounding Markov chain may even be periodic.
Hence the distribution of $X_i$, for $i$ arbitrarily large,
is not necessarily close to $\pi$.
Let $\varphi\colon G \times [0,1]^s \to G$ be an arbitrary \emph{update function}
of $K$ and $\psi \colon [0,1]^s \to G$ be an arbitrary \emph{generator function} of $\nu$
for some $s\in \mathbb{N}$. This means that
the Markov chain $(X_n)_{n\in \mathbb{N}}$ permits the representation
\begin{align*}
X_1 & = \psi(U_0),\\
X_{i+1} & = \varphi(X_i,U_i), \quad i\geq 1,
\end{align*}
where $(U_n)_{n\in \mathbb{N}}$ are i.i.d. with $U_i \sim \mathcal{U}[0,1]^s$.
Using a deterministic sequence $(u_i)_{i\geq0}$ we generate
the deterministic Markov chain $(x_i)_{i\geq1}$ with $x_1=\psi(u_0)$
and $x_{i+1}=\varphi(x_{i},u_{i})$ where $i\geq1$.
The efficiency of this procedure is measured by the star-discrepancy, a generalized Kolmogorov-Smirnov
test, between the stationary measure $\pi$ and the empirical
distribution $\widehat{\pi}_n(A) = \frac{1}{n} \sum_{i=1}^n 1_{x_i\in A}$, where
$1_{x_i\in A}$ is the indicator function of a set $A\subseteq G$.
For a certain set $\mathscr{A}$ of subsets of $G$ we define the
star-discrepancy $D^\ast_{\mathscr{A}, \pi}$ of $S_n=\{x_1,\dots,x_n\}$
as the supremum
of $\abs{\pi(A)-\widehat{\pi}_n(A)}$ over all $A\in\mathscr{A}$, i.e.
\begin{equation*}
D^\ast_{\mathscr{A}, \pi}(S_n)
= \sup_{A \in \mathscr{A}} \left|\widehat{\pi}_n(A) - \pi(A)\right|.
\end{equation*}
By inverting the iterates of the update function
we also define a pull-back discrepancy of the driver sequence (the test sets are
pulled back).
We show that for large $n\in\mathbb{N}$
both discrepancies are close to each other.
The main result, in a general setting, is an
estimate of $D^\ast_{\mathscr{A}, \pi}(S_n)$ (Theorem~\ref{thm_main})
under the
assumption that we have an approximation of
$\mathscr{A}$, for any $\delta>0$, given by a so-called
$\delta$-cover $\Gamma_\delta$ of $\mathscr{A}$ with respect to
$\pi$ (Definition~\ref{def: delta_cover}).
The proof of the main result is based on a Hoeffding inequality for Markov chains.
After that we prove
that a sufficiently
good $\delta$-cover exists if $\pi$ is absolutely continuous with respect
to the Lebesgue measure and the set of test sets is
the set of open boxes restricted to $G$
anchored at $-\infty$, i.e. we consider the set of test sets
\[
\mathscr{B}= \{ (-\infty,x)\cap G \colon x\in \mathbb{R}^d \},
\]
with $(-\infty,x)= \Pi_{i=1}^d (-\infty,x_i)$.
By the Koksma-Hlawka inequality (Theorem~\ref{thm_int_error}) we have
\begin{equation*}
\left|\mathbb{E}_\pi(f) - \frac{1}{n} \sum_{i=1}^n f(x_i)\right|
\le \|f\|_{H_{1}} D^\ast_{\mathscr{B}, \pi}(S_n),
\end{equation*}
with $\|f\|_{H_{1}}$ defined in \eqref{norm_H1} below.
Thus a bound on the discrepancy leads to an error bound for the approximation
of $\mathbb{E}_\pi(f)$.
We show
for any update function $\varphi\colon G \times [0,1]^s \to G$ of $K$, any
generator function $\psi \colon [0,1]^s \to G$ of $\nu$, and
for all $n\geq16$ that there exists a driver sequence $u_0,\dots,u_{n-1} \in [0,1]^s$ such that
$S_n=\{ x_1,\dots,x_n \}$ given by
\begin{align*}
x_1 & =\psi(u_0)\\
x_{i+1}& = \varphi(x_i;u_i), \quad i=1,\dots,n-1,
\end{align*}
satisfies
\begin{equation} \label{eq: disc_bound_boxes}
D^*_{\mathscr{B},\pi}(S_n)
\leq \sqrt{\frac{1+\Lambda_0}{1-\Lambda_0}} \cdot
\frac{\sqrt{2}\,(\log\norm{\frac{d\nu}{d\pi}}_{2}+d \log n + 3d^2 \log(5d))^{1/2}}{\sqrt{n}}
+\frac{8}{n^{3/4}},
\end{equation}
where $\frac{d\nu}{d\pi}$ is the density of
$\nu=\P_{\psi}$ (the probability measure induced by $\psi$) with respect to $\pi$
and $\Lambda_0=\max\{\Lambda,0\}$ with $\Lambda$ defined in \eqref{eq: lambda_intro}.
For the details we refer to Corollary~\ref{coro_main} below.
This implies, by the Koksma-Hlawka inequality, that the sample average converges
to the mean with $\mathcal{O}(n^{-1/2}(\log n)^{1/2})$.
Additionally we might take a burn-in period of $n_0$
steps into account to reduce the
dependence of the initial state in the discrepancy bound.
Roughly, the idea is to generate a sequence
$x_1,\dots,x_{n_0+n}$ by the Markov chain quasi-Monte Carlo
procedure
and to consider the discrepancy of the point set
$S_{[n_0,n]}=\{x_{n_0+1},\dots, x_{n_0+n}\}$.
Under suitable convergence conditions on
the Markov chain, for example the existence of an absolute $L_2$-spectral gap
(see Definition~\ref{def: abs_spec_gap}), the density $\frac{d(\nu P^{n_0})}{d\pi}$
is close to $1$, see Subsection~\ref{subsec_burn_in}.
If we further assume that one can reach every state from every other
state within one step of the Markov chain,
then we prove that there exists a driver sequence such that the discrepancy
converges with $\mathcal{O}(n^{-1} (\log_2 n)^{(3d+1)/2})$.
We call the additional assumption `anywhere-to-anywhere' condition.
The result shows that in principle a higher order of convergence for Markov
chain quasi-Monte Carlo is possible.
Note that, many well studied
Markov chains satisfy such a condition, for example the hit-and-run algorithm,
the independent Metropolis sampler or the slice sampler, see for example \cite{Li08}.
From our work it is not immediately clear how to obtain suitable
driver sequences which yield such an improvement. However,
what our results here show is that the main quality criterion is the pull-back discrepancy
(see Definition~\ref{def-pull-back-discrepancy} below) of the driver sequence.
Since this has previously not been known, the pull-back discrepancy
of explicit constructions of quasi-Monte Carlo point sets or pseudo-random number
generators has not been studied so far. The task of future work is
therefore to explicitly construct point sets with small pull-back discrepancy.
We leave it is an interesting and challenging problem for further research.
We provide an outline of our work in the following.
\subsection{Outline}
In the next section the necessary background information on Markov chains is stated.
Section~\ref{sec_discr} is devoted to the study of the relation of the discrepancies.
The Monte Carlo rate of convergence for
Markov chain-quasi Monte Carlo
is shown in Section~\ref{sec: MC_rate}.
There we also provide results
for the case when a burn-in period is taken into account.
Section~\ref{sec_application}
deals with the set of test sets which consists of axis parallel boxes, see $\mathscr{B}$ above.
We show the existence of a good $\delta$-cover and
how the discrepancy bounds can be used to obtain bounds on the
error for the computation of expected values of smooth functions.
This yields a Koksma-Hlawka inequality for Markov chains.
To illustrate our results, we provide an example of a
Metropolis algorithm with ball walk proposal on the Euclidean unit ball.
A special situation arises when the update function
of the Markov chain has an `anywhere-to-anywhere' property,
see Section~\ref{sec: beyound_MC}.
In this situation we show that a convergence rate
of order almost $n^{-1}$ can be obtained.
\section{Background on Markov chains}\label{sec_mc}
Let $G \subseteq \mathbb{R}^d$ and let $\mathcal{B}(G)$
denote the Borel $\sigma$-algebra of $G$.
In the following we provide a brief introduction to Markov chains on $(G,\mathcal{B}(G))$.
We assume that $K:G \times \mathcal{B}(G) \to [0,1]$
is a transition kernel on $(G,\mathcal{B}(G))$, i.e.
for each $x\in G$ the mapping $A\in\mathcal{B}(G) \mapsto K(x,A)$ is a probability measure and
for each $A\in\mathcal{B}(G)$ the mapping $x\in G \mapsto K(x,A)$ is a $\mathcal{B}(G)$-measurable real-valued function.
Further let $\nu$ be a
probability measure on $(G,\mathcal{B}(G))$.
Then let $(X_n)_{n\in \mathbb{N}}$,
with $X_n$ mapping from some probability space into $(G,\mathcal{B}(G))$,
be a Markov chain with transition kernel $K$ and initial distribution $\nu$.
This might be interpreted as follows:
Let $X_1=x_1 \in G$ be chosen with $\nu$ on $(G,\mathcal{B}(G))$
and let $i\in\mathbb{N}$.
Then for a given $X_{i}=x_{i}$, the random variable $X_{i+1}$ has distribution
$K(x_{i}, \cdot)$, that is, for all $A \in \mathcal{B}(G)$,
the probability that $X_{i+1}\in A$ is given by $K(x_{i}, A)$.
Let $\pi$ be a probability measure on $(G,\mathcal{B}(G))$.
We assume that the transition kernel $K$
is \emph{reversible with respect to $\pi$}, i.e.
\[
\int_A K(x,B)\, \pi({\rm d}x)= \int_B K(x,A)\, \pi({\rm d}x)
\]
for all $A,B \in \mathcal{B}(G)$.
This implies that $\pi$ is a \emph{stationary distribution}
of the transition kernel $K$, i.e.
\begin{equation} \label{eq: stat}
\int_G K(x,A) \,\pi(\mathrm{d} x) = \pi(A)
\end{equation}
for all $A\in \mathcal{B}(G)$.
We assume that the stationary distribution $\pi$ is unique.
Let $L_2=L_2(\pi)$ be the set of all functions $f\colon G\to \mathbb{R}$ with
\[
\norm{f}_{2}=\left(\int_G \abs{f(x)}^2\,\pi(\mathrm{d} x)\right)^{1/2} < \infty.
\]
The transition kernel $K$ induces an operator acting on functions and
an operator acting on measures. For $x\in G$ and $A\in \mathcal{B}(G)$ the operators are given
by
\[
Pf (x) = \int_G f(y) \, K(x,{\rm d}y),
\quad \mbox{and} \quad
\nu P(A) = \int_G K(x,A)\, \nu({\rm d} x),
\]
where $f\in L_2$ and $\nu$ is a signed measure on $(G,\mathcal{B}(G))$ with a density $\frac{d\nu}{d\pi} \in L_2$.
By the reversibility with respect to $\pi$ we have that $P\colon L_2 \to L_2$ is self-adjoint
and $\pi$-almost everywhere holds $P(\frac{d\nu}{d\pi})(x) = \frac{ d (\nu P)}{d \pi}(x)$.
For details we refer to \cite{Ru12}.
In the following we introduce two convergence properties of transition kernels.
Let the expectation with respect to $\pi$
be denoted by $\mathbb{E}_\pi(f)=\int_G f(y) \pi(\mathrm{d} x)$.
Let $L_2^0 = \{ f\in L_2 \colon \mathbb{E}_\pi (f) =0 \}$ and note that
$L_2^0$ is a closed subspace of $L_2$.
We have
\[
\norm{P-\mathbb{E}_\pi}_{L_2 \to L_2}
= \norm{P}_{L_2^0 \to L_2^0}
= \sup_{f\in L_2^0,\, \norm{f}_{2}\leq 1} \norm{Pf}_{2},
\]
for details see \cite[Lemma~3.16, p.~44]{Ru12}.
\begin{definition}[absolute $L_2$-spectral gap] \label{def: abs_spec_gap}
We say that a transition kernel $K$, and its corresponding Markov operator $P$,
has an \emph{absolute $L_2$-spectral gap}
if
\[
\beta = \norm{P}_{L_2^0 \to L_2^0} < 1,
\]
and the absolute spectral gap is $ 1 - \beta $.
\end{definition}
Let us introduce the \emph{total variation distance} of two probability measures $\nu_1, \nu_2$ on $(G,\mathcal{B}(G))$
by
\[
\norm{\nu_1 - \nu_2}_{\text{\rm tv}}
= \sup_{A \in \mathcal{B}(G)} \abs{\nu_1(A) - \nu_2(A)}.
\]
Note that for a Markov chain $(X_n)_{n\in \mathbb{N}}$ with transition kernel $K$
and initial distribution $\nu$ holds $\P_{\nu,K}(X_n \in A) = \nu P^{n-1}(A)$, where
$\nu$ and $K$ in $\P_{\nu,K}$ indicate the initial distribution and transition kernel.
Then we obtain the following relation between the absolute
$L_2$-spectral gap and the total variation distance.
The result is an application of \cite[Corollary~3.15 and Lemma~3.21]{Ru12}.
\begin{proposition} \label{prop: tv_abs_spec_gap}
Let $\nu$ be
a distribution on $(G,\mathcal{B}(G))$
and assume that there exists a density $\frac{d \nu}{d \pi} \in L_2$.
Then
\begin{align*}
\norm{\nu P^n - \pi}_{\text{\rm tv}}
& \leq \beta^n \norm{\frac{d \nu}{d \pi}-1}_{2}, \quad n\in\mathbb{N},
\end{align*}
with $\beta = \norm{P}_{L_2^0 \to L_2^0}$.
\end{proposition}
The next convergence property is weaker than the existence of an absolute spectral gap.
\begin{definition}[Variance bounding or $L_2$-spectral gap]
We say that a reversible transition kernel $K$, and its corresponding Markov operator $P$,
is \emph{variance bounding} or has an \emph{$L_2$-spectral gap} if
\begin{equation} \label{eq: Lambda_variance_bounding}
\Lambda = \sup\{ \lambda \in {\rm spec}(P\mid L_2^0) \} < 1,
\end{equation}
where ${\rm spec}(P\mid L_2^0)$ denotes the spectrum of $P \colon L_2^0 \to L_2^0$.
\end{definition}
For a motivation of the term variance bounding and a
general treatment we refer to \cite{RoRo08}.
In particular, by \cite[Theorem~14]{RoRo08} under the assumption of reversibility our definition
is equivalent to the one stated by Roberts and Rosenthal.
Note that the existence of an absolute $L_2$-spectral gap implies variance bounding, since
\[
\norm{P}_{L_2^0 \to L_2^0} = \sup_{ \lambda \in {\rm spec}(P\mid L_2^0) }\abs{\lambda}.
\]
We have the following relation between variance bounding
and the total variation distance.
\begin{lemma} \label{lem: tv_variance_bounding}
Let the transition kernel
$K$ be reversible with respect to $\pi$ and let $n\in \mathbb{N}$ with $n\geq 2$.
Further, let $P$ be variance bounding.
Then the Markov operator
$P_n = \frac{1}{n} \sum_{j=0}^{n-1} P^j$ has an absolute $L_2$-spectral gap.
In particular, if $\nu$ is a distribution on $(G,\mathcal{B}(G))$ with $\frac{d\nu}{d\pi} \in L_2$, then
\[
\norm{\nu P_n - \pi}_{{\rm tv}} \leq \frac{1-\Lambda_0^n}{n\cdot(1-\Lambda_0)} \norm{\frac{d \nu}{d \pi}-1}_{2},
\]
with $\Lambda_0 = \max\{0,\Lambda\}$, see \eqref{eq: Lambda_variance_bounding}.
\end{lemma}
\begin{proof}
By the spectral theorem for bounded self-adjoint operators we have for a
polynomial
$F\colon {\rm spec}(P \mid L_2^0) \to \mathbb{R}$ that
\[
\norm{F(P)}_{L_2^0 \to L_2^0} = \max_{\a \in \text{\rm spec}(P|L_2^0)} \abs{F(\a)}.
\]
For details see for example \cite{Ru91} or \cite[Theorem~9.9-2]{Kr89}.
In our case $F(\lambda)= \frac{1}{n}\sum_{i=0}^{n-1} \lambda^i $ so that
$F(P)=\frac{1}{n}\sum_{i=0}^{n-1} P^i$.
Thus
\begin{align*}
\norm{\frac{1}{n}\sum_{i=0}^{n-1} P^i}_{L_2^0\to L_2^0}
& = \max_{\lambda \in \text{\rm spec}(P|L_2^0)}
\abs{\frac{1-\lambda^n}{n\cdot(1-\lambda)}}
\leq \frac{1-\Lambda_0^n}{n\cdot(1-\Lambda_0)}.
\end{align*}
The last inequality is proven by $\text{\rm spec}(P|L_2^0) \subseteq[-1,1]$ and the following facts: For $\lambda\in[-1,0]$
holds $\frac{1-\lambda^n}{n\cdot(1-\lambda) }
\leq \frac{1}{n}$ and for $\lambda\in[0,1]$
the function
$\frac{1-\lambda^n}{n\cdot(1-\lambda)}= \frac{1}{n}\sum_{j=0}^{n-1} \lambda^j$ is increasing.
The estimate of the total variation distance follows by
Proposition~\ref{prop: tv_abs_spec_gap}.
\end{proof}
The next part deals with an update function, say $\varphi$, of a given transition kernel $K$.
We state the crucial properties of the transition kernel in terms
of an update function. This is partially based on \cite{DiRuZh13}.
\begin{definition}[Update function]
Let $\varphi:G \times [0,1]^s \to G$
be a measurable
function and
\begin{align*}
B : G \times \mathcal{B}(G) & \to \mathcal{B}([0,1]^s), \\
B(x,A) & = \{u \in [0,1]^s: \varphi(x;u) \in A\}.
\end{align*}
Let $\lambda_s$ denote the Lebesgue measure on $\mathbb{R}^s$.
Then the function $\varphi$ is an update function for the transition kernel $K$ if and only if
\begin{equation}\label{eq_update_prop}
K(x,A) = \P(\varphi(x;U)\in A) = \lambda_s(B(x,A)),
\end{equation}
where $\P$ is the probability measure for the uniform distribution in $[0,1]^s$.
\end{definition}
Note that for any transition kernel on $(G,\mathcal{B}(G))$
there exists an update function with $s=1$, see for example \cite[Lemma~2.22, p.~34]{Ka02}.
For $x\in G$ and $A\in \mathcal{B}(G)$ the set $B(x,A)$ is
the set of all random numbers $u\in [0,1]^s$ which take
$x$ into the set $A$ using the update function $\varphi$
with arguments $x$ and $u$.
We consider the iterated application of an update function.
Let $\varphi_1(x;u) = \varphi(x;u)$ and for $i > 1$ with $i\in \mathbb{N}$ let
\begin{align*}
\varphi_i & : G \times [0,1]^{is} \to G, \\
\varphi_i(x; u_1, u_2, \ldots, u_i) & = \varphi(\varphi_{i-1}(x; u_1, u_2,\ldots, u_{i-1}); u_i).
\end{align*}
Thus, $x_{i+1}=\varphi_i(x; u_1, u_2,\ldots, u_i) \in G$ is the point obtained
via $i$ updates using the sequence $u_1,u_2,\dots,u_i\in[0,1]^s$,
where the starting point is $x \in G$.
\begin{lemma} \label{lem: update_fct}
Let $i,j\in\mathbb{N}$ and $i\geq j$. For any $u_1,\dots,u_i \in [0,1]^s$ and $x\in G$ we have
\begin{equation} \label{eq: iterat_update_fct}
\varphi_i(x;u_1,\dots,u_i) = \varphi_{i-j}(\varphi_j(x;u_1,\dots,u_j);u_{j+1},\dots,u_i).
\end{equation}
\end{lemma}
\begin{proof}
The proof follows by induction on $i$.
\end{proof}
For $i \ge 1$ with $i\in\mathbb{N}$ let
\begin{align*}
B_i & : G \times \mathcal{B}(G) \to \mathcal{B}([0,1]^{i s}), \\
B_i(x,A) & = \{(u_1, u_2, \ldots, u_i) \in [0,1]^{i s}: \varphi_i(x; u_1, u_2,\ldots, u_i) \in A\}.
\end{align*}
Note that $B_1(x,A) = B(x,A)$.
For $x\in G$ and $A\in \mathcal{B}(G)$ the set $B_i(x,A)$ is
the set of all random numbers $u_1,u_2,\dots,u_i\in [0,1]^s$ which take
$x$ into the set $A$ after the $i$th iteration of the update function $\varphi$, i.e.
$\varphi_i$
with arguments $x$ and $u_1,u_2,\dots,u_i$.
In \cite{DiRuZh13}
we considered the case where the initial state is deterministically chosen.
The following definition is useful to work with general initial distributions.
\begin{definition}
For a probability measure $\nu$ on $(G,\mathcal{B}(G))$ we call a measurable
function
$\psi\colon [0,1]^s \to G$ generator function if
\[
\nu(A) = \P ( \psi(U) \in A), \quad A \in \mathcal{B}(G),
\]
where $\P$ is the uniform distribution in $[0,1]^s$.
\end{definition}
Let $\nu$ be a probability measure on $(G,\mathcal{B}(G))$ and let
$\psi \colon [0,1]^s \to G$ be its generator function.
Then, for $i\geq1$ with $i\in \mathbb{N}$ and $A\in \mathcal{B}(G)$, let
\begin{equation} \label{eq: C_i_psi}
\begin{split}
C_{i,\psi}(A) & = \{ (u_0,u_1,\dots,u_{i})\in[0,1]^{(i+1)s}\colon \varphi_{i}(\psi(u_0);u_1,\dots,u_{i})\in A \}\\
& = \{ (u_0,u_1,\dots,u_{i})\in[0,1]^{(i+1)s}\colon (u_1,\dots,u_i)\in B_i(\psi(u_0), A) \}
\end{split}
\end{equation}
and $C_{0,\psi}(A)=\{ u_0\in[0,1]^s \colon \psi(u_0)\in A \}$.
The set $C_{i,\psi}(A)\subseteq[0,1]^{(i+1)s}$ is the set of possible sequences
to get into the set $A$ with starting point $\psi(u_0)$ and $i$ updates of
the update function.
The next lemma is important to understand the relation
between the update function, generator function,
transition kernel and initial distribution.
\begin{lemma} \label{lem: same_expect}
Let $K$ be a transition kernel and $\nu$ be a distribution on $(G,\mathcal{B}(G))$.
Let $(X_n)_{n\in\mathbb{N}}$ be a Markov
chain with transition kernel $K$ and initial distribution $\nu$.
Let us assume that $i\in \mathbb{N}$ and $F\colon G^{i} \to \mathbb{R}$.
Then, for any update function $\varphi\colon G \times [0,1]^s \to G$
of the transition kernel $K$ and any generator function $\psi \colon [0,1]^s \to G$
of $\nu$
the expectation of $F$ with respect to the joint distribution of
$X_1,\dots,X_{i}$ satisfies
\begin{equation} \label{eq: same_expect}
\begin{split}
& \mathbb{E}_{\nu,K}(F(X_1,\dots,X_{i})) \\
& =\int_{[0,1]^{is}}
F(\psi(u_0),\varphi_1(\psi(u_0),u_1),\dots,\varphi_{i-1}(\psi(u_0),u_1,\dots,u_{i-1}))\;\\
& \qquad\qquad \qquad \qquad \qquad \qquad \times \mathrm{d} u_0\, \mathrm{d} u_1\dots \mathrm{d} u_{i-1},
\end{split}
\end{equation}
whenever one of the integrals exist.
\end{lemma}
\begin{proof}
First, note that
\begin{align*}
& \mathbb{E}_{\nu,K}(F(X_1,\dots,X_{i})) \\
& = \underbrace{\int_G \dots \int_G}_{i\text{-times}}
F(x_1,\dots,x_{i})\, K(x_{i-1},\mathrm{d} x_{i}) \dots K(x_1,\mathrm{d} x_2)\, \nu(\mathrm{d} x_1).
\end{align*}
By the fact that $\psi$ is a generator function of $\nu$
we have
\begin{align*}
& \int_{[0,1]^{is}}
F(\psi(u_0),\varphi_1(\psi(u_0),u_1),\dots,\varphi_{i-1}(\psi(u_0),u_1,\dots,u_{i-1}))\,\mathrm{d} u_0\, \mathrm{d} u_1\dots \mathrm{d} u_{i-1}\\
& = \int_G \int_{[0,1]^{(i-1)s}}
F(x_1,\varphi_1(x_1,u_1),\dots,\varphi_{i-1}(x_1,u_1,\dots,u_{i-1}))\,
\mathrm{d} u_1\dots \mathrm{d} u_{i-1}\,\nu({\rm d}x_1),
\end{align*}
and by Lemma~\ref{lem: update_fct} we obtain
\begin{align*}
&\int_G \int_{[0,1]^{(i-1)s}}
F(x_1,\varphi_1(x_1,u_1),\dots,\varphi_{i-1}(x_1,u_1,\dots,u_{i-1}))\,
\mathrm{d} u_1\dots \mathrm{d} u_{i-1}\,\nu({\rm d}x_1)\\
& = \int_G \int_G \int_{[0,1]^{(i-2)s}}
F(x_1,x_2,\varphi_1(x_2,u_2),\dots,\varphi_{i-1}(x_2,u_2,\dots,u_{i-1}))\\
& \qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad \times
\mathrm{d} u_2\dots \mathrm{d} u_{i-1}\,K(x_1,{\rm d}x_2)\,\nu({\rm d}x_1).
\end{align*}
By iterating the application of Lemma~\ref{lem: update_fct}
the assertion is proven.
\end{proof}
Note that the right-hand-side of \eqref{eq: same_expect} is the expectation with respect
to the uniform distribution in $[0,1]^{is}$.
\begin{corollary}
Assume that the conditions of Lemma~\ref{lem: same_expect} are satisfied. Then, for $A\in \mathcal{B}(G)$, we have
\begin{equation} \label{eq: C_and_P}
\nu P^i(A)
= \lambda_{(i+1)s}(C_{i,\psi}(A)),
\end{equation}
and $\nu P^0(A) = \nu(A) = \lambda_s(C_{0,\psi}(A))$.
\end{corollary}
\begin{proof}
By Lemma~\ref{lem: same_expect} we have
\begin{align*} \label{eq: C_and_P}
\nu P^i(A)
& = \int_G K^i(x,A)\,\nu(\mathrm{d} x) \\
& = \int_G \underbrace{\int_G \dots \int_G}_{i\text{-times}}
1_{x_{i+1}\in A}\, K(x_{i},\mathrm{d} x_{i+1}) \dots K(x_1,\mathrm{d} x_2)\, \nu(\mathrm{d} x_1)\\
& = \int_{[0,1]^{(i+1)s}} 1_{\varphi_i(\psi(u_0),u_1,\dots,u_i)\in A} \, \mathrm{d} u_0 \, \mathrm{d} u_1 \dots \mathrm{d} u_i \\
& = \int_{[0,1]^{(i+1)s}} 1_{(u_0,u_1,\dots,u_i)\in C_{i,\psi}(A)} \, \mathrm{d} u_0 \, \mathrm{d} u_1 \dots \mathrm{d} u_i
= \lambda_{(i+1)s}(C_{i,\psi}(A)),
\end{align*}
which completes the proof.
\end{proof}
\section{On the pull-back discrepancy}\label{sec_discr}
Let $\mathscr{A} \subseteq \mathcal{B}(G)$ be a set of test sets.
Then the star-discrepancy of a point set
$S_{n} = \{x_1, \ldots, x_{n} \} \subseteq G$
with respect to the distribution $\pi$
is given by
\begin{equation*}
D^\ast_{\mathscr{A}, \pi}(S_n)
= \sup_{A \in \mathscr{A}} \left|\frac{1}{n} \sum_{i=1}^{n} 1_{x_i \in A} - \pi(A)\right|.
\end{equation*}
Assume that $u_0,u_1,\ldots, u_{n-1} \in [0,1]^s$ is a finite deterministic sequence.
We call this finite sequence \emph{driver sequence}.
Further, let $\varphi\colon G \times [0,1]^s \to G$ and $\psi \colon [0,1]^s \to G$
be measurable functions.
Then let
$S_n=\{x_1, \ldots, x_{n} \} \subseteq G$ be given by
\begin{equation} \label{eq: x_i_by_driver_seq}
x_{i+1} = x_{i+1}(x_1) = \varphi(x_{i};u_i) = \varphi_i(x_1;u_1,\dots,u_i), \quad i=1,\dots,n-1,
\end{equation}
where $x_1=\psi(u_0)$.
Note that $\psi$ might be considered a generator function
and $\varphi$ might be considered an update function.
We now define a discrepancy measure on the driver sequence.
We call it \emph{pull-back discrepancy}.
Below we show how this pull-back discrepancy
is related to the star-discrepancy of $S_n$.
\begin{definition}[Pull-back discrepancy]\label{def-pull-back-discrepancy}
Let $\;\mathcal{U}_{n} = \{u_0, u_1,\ldots, u_{n-1} \}
\subset [0,1]^{s}$ and
let $C_{i,\psi}(A)$ for $A\in\mathcal{B}(G)$ and $i\in \mathbb{N}\cup \{0\}$ be
defined as in \eqref{eq: C_i_psi}. Define the local discrepancy function by
\[
\Delta^{\text{\rm loc}}_{n,A,\psi,\varphi}(\mathcal{U}_n)
=\frac{1}{n} \sum_{i=0}^{n-1} \left[1_{(u_0,\ldots, u_i) \in C_{i,\psi}(A)} -
\lambda_{(i+1)s}(C_{i,\psi}(A)) \right].
\]
Let $\mathscr{A} \subseteq \mathcal{B}(G)$ be a set of test sets. Then we define the discrepancy of the driver sequence by
\begin{equation*}
D^\ast_{\mathscr{A},\psi,\varphi}(\mathcal{U}_n)
= \sup_{A \in \mathscr{A}} \left|\Delta^{\text{\rm loc}}_{n,A,\psi,\varphi}(\mathcal{U}_n) \right|.
\end{equation*}
We call $D^\ast_{\mathscr{A},\psi,\varphi}(\mathcal{U}_n)$ pull-back discrepancy of $\mathcal{U}_n$.
\end{definition}
The discrepancy of the driver sequence
$D^\ast_{\mathscr{A},\psi, \varphi}(\mathcal{U}_n)$
is a `pull-back discrepancy' since the test sets $C_{i,\psi}(A)$ are derived
from the test sets $A \in \mathscr{A}$ from the
star-discrepancy $D^\ast_{\mathscr{A}, \pi}(S_n)$
via inverting the update function and the generator.
The following theorem
provides a relation between the star-discrepancy of $S_n$ and the pull-back discrepancy
of $\mathcal{U}_n$, this is similar to \cite[Theorem~1]{DiRuZh13}.
\begin{theorem} \label{thm: est_discr}
Let $K$ be a transition kernel and $\nu$
be a distribution on $(G,\mathcal{B}(G))$.
Let $\mathscr{A} \subseteq \mathcal{B}(G)$ be a set of test sets.
Then, for any update function $\varphi\colon G \times [0,1]^s \to G$
of $K$ and any generator function $\psi \colon [0,1]^s \to G$ of $\nu$
we have, with driver sequence
$\mathcal{U}_n=\{u_0, u_1, \ldots, u_{n-1}\} \subset [0,1]^s$
and $S_n$ given by \eqref{eq: x_i_by_driver_seq}, that
\begin{align*}
\abs{D^\ast_{\mathscr{A}, \pi}(S_n) - D^\ast_{\mathscr{A},\psi,\varphi}(\mathcal{U}_n) }
& \le \sup_{A \in \mathscr{A}} \abs{ \frac{1}{n} \sum_{i=0}^{n-1} \nu P^i(A) - \pi(A) }.
\end{align*}
\end{theorem}
\begin{proof}
For any $A \in \mathscr{A}$
we have by \eqref{eq: C_and_P} that $\lambda_{(i+1)s}(C_{i,\psi}(A))=\nu P^i(A)$.
Thus
\begin{align*}
& \quad\, \left| \frac{1}{n} \sum_{i=1}^{n} 1_{x_i \in A} - \pi(A) \right| \\
&= \left| \frac{1}{n} \sum_{i=0}^{n-1} \left[ 1_{(u_0,\ldots, u_i) \in C_{i,\psi}(A)} - \nu P^i(A) + \nu P^i(A)- \pi(A) \right] \right| \\
& \leq \abs{ \frac{1}{n} \sum_{i=0}^{n-1} \left[1_{(u_0,\ldots, u_i) \in C_{i,\psi}(A)} - \lambda_{(i+1)s}(C_{i,\psi}(A))\right]}
+ \abs{ \frac{1}{n} \sum_{i=0}^{n-1} \nu P^i(A) - \pi(A) }.
\end{align*}
Hence
\[
D^\ast_{\mathscr{A}, \pi}(S_n)
\le D^\ast_{\mathscr{A},\psi, \varphi}(\mathcal{U}_n)
+ \sup_{A \in \mathscr{A}} \abs{ \frac{1}{n} \sum_{i=0}^{n-1} \nu P^i(A) - \pi(A) }.
\]
The inequality
\[
D^\ast_{\mathscr{A},\psi, \varphi}(\mathcal{U}_n)
\le D^\ast_{\mathscr{A}, \pi}(S_n) +
\sup_{A \in \mathscr{A}} \abs{ \frac{1}{n} \sum_{i=0}^{n-1} \nu P^i(A) - \pi(A) }
\]
follows by the same arguments.
\end{proof}
\begin{corollary} \label{coro: D_U_almost_D_P_spec}
Assume that the conditions of Theorem~\ref{thm: est_discr} are satisfied.
By $P$ denote the Markov operator of $K$. Further, let $K$ be reversible
with respect to $\pi$, let $P$ be variance bounding and let $\frac{d\nu}{d\pi} \in L_2$.
Then, for any update function $\varphi\colon G \times [0,1]^s \to G$
of $K$ and any generator function $\psi \colon [0,1]^s \to G$ of $\nu$
we have, with driver sequence
$\mathcal{U}_n=\{u_0, u_1, \ldots, u_{n-1}\} \subset [0,1]^s$
and $S_n$ given by \eqref{eq: x_i_by_driver_seq}, that
\begin{align*}
\left|D^\ast_{\mathscr{A}, \pi}(S_n) - D^\ast_{\mathscr{A},\psi,\varphi}(\mathcal{U}_n) \right| \le &
\frac{1-\Lambda_0^n}{n\cdot(1-\Lambda_0)} \norm{\frac{d\nu}{d\pi}-1}_{2},
\end{align*}
where $\Lambda_0=\max\{0,\Lambda\}$ and $\Lambda$ is defined in \eqref{eq: Lambda_variance_bounding}.
\end{corollary}
\begin{proof}
With $P_n= \frac{1}{n}\sum_{i=0}^{n-1} P^i$ we have
\[
\sup_{A \in \mathscr{A}} \abs{ \frac{1}{n} \sum_{i=0}^{n-1} \nu P^i(A) - \pi(A) }
\leq \norm{\nu P_n - \pi}_{{\rm tv}}.
\]
Thus, the assertion follows by Lemma~\ref{lem: tv_variance_bounding} and
Theorem~\ref{thm: est_discr}.
\end{proof}
\begin{remark} \label{rem: direc_simulation}
For the moment let us assume that we can sample with respect to $\pi$.
For any initial distribution $\nu$ with $\frac{d\nu}{d\pi}\in L_2$, for all $x\in G$ and $A\in \mathcal{B}(G)$
we set $K(x,A)=\pi(A)$, hence $\Lambda = 0$. Thus,
for any update function $\varphi$ of $K$ and generator function $\psi$ of $\nu$ we have
\begin{align*}
\left|D^\ast_{\mathscr{A}, \pi}(S_n) - D^\ast_{\mathscr{A},\psi,\varphi}(\mathcal{U}_n) \right| \le &
\frac{1}{n} \norm{\frac{d\nu}{d\pi}-1}_{2}.
\end{align*}
Note that the discrepancies do not coincide. The reason for this is that the initial state is taken
into account in the average computation.
However, if $\nu=\pi$,
then for any reversible transition kernel with respect to $\pi$
we obtain $D^\ast_{\mathscr{A}, \pi}(P_n) = D^\ast_{\mathscr{A},\psi,\varphi}(\mathcal{U}_n)$.
\end{remark}
\section{Monte Carlo rate of convergence}\label{sec: MC_rate}
In this section we show
for any update function $\varphi \colon G \times [0,1]^s \to G$ of a
variance bounding
transition kernel $K$
and any generator function $\psi \colon [0,1]^s \to G$ of a distribution $\nu$ the existence of finite sequences
$\mathcal{U}_n = \{u_0, u_1,\ldots, u_{n-1}\} \subset [0,1]^{s}$,
which define $S_n$ by \eqref{eq: x_i_by_driver_seq}, such that
\begin{align*}
D^\ast_{\mathscr{A},\psi,\varphi}(\mathcal{U}_n)
\quad \text{and} \quad
D^\ast_{\mathscr{A}, \pi}(S_n)
\end{align*}
converge to $0$ approximately with order $n^{-1/2}$.
The main result is proven for $D^\ast_{\mathscr{A}, \pi}(S_n)$.
The result with respect to $D^\ast_{\mathscr{A},\psi, \varphi}(\mathcal{U}_n)$ holds by Theorem~\ref{thm: est_discr}.\\
\subsection{Useful tools: delta-cover and Hoeffding inequality}
The concept of a $\delta$-cover will be useful (cf. \cite{Gn08} for a discussion of $\delta$-covers,
bracketing numbers and Vapnik-\v{C}ervonenkis dimension).
\begin{definition} \label{def: delta_cover}
Let $\mathscr{A} \subseteq \mathcal{B}(G)$ be a set of test sets.
A finite subset $\Gamma_\delta \subseteq \mathcal{B}(G)$ is called a $\delta$-cover
of $\mathscr{A}$ with respect to $\pi$
if for every $A \in \mathscr{A}$ there are sets $C, D \in \Gamma_\delta$ such that
\begin{equation*}
C \subseteq A \subseteq D
\end{equation*}
and
\begin{equation*}
\pi(D \setminus C) \le \delta.
\end{equation*}
We assume that $\emptyset \in \Gamma_\delta$.
\end{definition}
The following result is well known for the uniform distribution, see \cite[Section~2.1]{HeNoWaWo01} (see also \cite[Remark~3]{DiRuZh13} for the particular case below).
\begin{proposition}
Let $\mathscr{A} \subseteq \mathcal{B}(G)$ be a set of test sets.
Let $\Gamma_\delta$ be a $\delta$-cover of $\mathscr{A}$ with respect to $\pi$.
Then, for any point set $Z_n=\{z_1,\dots,z_n\} \subseteq G$, we have
\[
D^\ast_{\mathscr{A}, \pi}(Z_n)
\leq \max_{C\in \Gamma_\delta}
\left| \frac{1}{n} \sum_{i=1}^n 1_{z_i\in C}-\pi(C) \right| + \delta.
\]
\end{proposition}
Instead of considering the supremum
over the possibly infinite set of test sets $\mathscr{A}$ in the star-discrepancy
we use a finite set $\Gamma_\delta$ and take
the maximum over $C\in \Gamma_\delta$ by paying the price of adding $\delta$.
For variance bounding Markov chains on discrete state spaces
a Hoeffding inequality is proven in \cite{LePe04}.
In \cite{Mi12} this is extended to non-reversible Markov chains on general state spaces.
The following Hoeffding inequality for reversible, variance bounding Markov chains
follows by \cite[Theorem~3.3 and the remark after $(3.4)$]{Mi12}.
\begin{proposition}[Hoeffding inequality for Markov chains] \label{prop: Hoeffd}
Let $K$ be a reversible transition kernel with respect $\pi$ and let
$\nu$ be a distribution on $(G,\mathcal{B}(G))$ with $\frac{d\nu}{d\pi}\in L_2$. Let
us assume that the Markov operator of $K$ is variance bounding.
Further, let $(X_n)_{n\in \mathbb{N}}$ be a Markov chain with transition kernel $K$
and initial distribution $\nu$.
Then, for any $A\in\mathcal{B}(G)$ and $c>0$, we obtain
\begin{equation}
\P_{\nu,K} \left[ \abs{\frac{1}{n} \sum_{i=1}^{n} 1_{X_i\in A}-\pi(A)} \geq c \right]
\leq 2 \norm{\frac{d\nu}{d\pi}}_{2}\exp\left( -\frac{1-\Lambda_0}{1+\Lambda_0}\, c^2 n\right),
\end{equation}
with $\Lambda_0 = \max\{0,\Lambda\}$ and where $\Lambda$ is defined in \eqref{eq: Lambda_variance_bounding}.
\end{proposition}
We provide a lemma to state the Hoeffding inequality for Markov chains in terms
of the driver sequence.
To do so, let $\varphi \colon G \times [0,1]^s \to G$ and $\psi \colon [0,1]^s \to G$.
We need the following notation.
Let $\Delta_{n,A,\varphi,\psi} \colon [0,1]^{ns} \to [-1,1]$ be given by
\begin{align} \label{al: loc_disc_star_disc}
\Delta_{n,A,\varphi,\psi}(u_0,\dots,u_{n-1})
& =\frac{1}{n} \sum_{i=0}^{n-1} \left[1_{(u_0,\ldots, u_i) \in C_{i,\psi}(A)} - \pi(A) \right].
\end{align}
\begin{lemma} \label{lem: same_conc}
Let $K$ be a transition kernel and $\nu$ be a distribution on $(G,\mathcal{B}(G))$.
Let $(X_n)_{n\in\mathbb{N}}$ be a Markov chain with
transition kernel $K$ and initial distribution $\nu$.
Then, for any update function $\varphi \colon G \times [0,1]^s \to G$ of $K$,
any generator function $\psi \colon [0,1]^s \to G$, any $A\in\mathcal{B}(G)$ and $c>0$,
we have
\begin{equation}
\P[ \abs{\Delta_{n,A,\varphi,\psi}} \geq c ]
= \P_{\nu,K} \left[ \abs{\frac{1}{n} \sum_{i=1}^{n} 1_{X_i\in A}-\pi(A)} \geq c \right],
\end{equation}
where $\P$ denotes the uniform distribution in $[0,1]^{ns}$
and $\P_{\nu,K}$ denotes
the joint distribution of $X_1,\dots,X_{n}$.
\end{lemma}
\begin{proof}
Let
$
J(A,c) = \left \{ (z_1,\dots,z_{n})\in G^{n} \colon \abs{\frac{1}{n} \sum_{i=1}^{n} 1_{z_i\in A}-\pi(A)} \geq c \right\}
$
and let
\[
F(x_1,\dots,x_{n})
= 1_{(x_1,\dots,x_{n})\in J(A,c)}
= \begin{cases}
1 & \abs{\frac{1}{n} \sum_{i=1}^{n} 1_{x_i\in A}-\pi(A)} \geq c,\\
0 & \text{otherwise}.
\end{cases}
\]
By
$
\mathbb{E}_{\nu,K}(F(X_1,\dots,X_{n})) = \P_{\nu,K} (J(A,c)),
$
Lemma~\ref{lem: same_expect} and
\begin{align*}
& 1_{(\psi(u_0),\varphi_1(\psi(u_0),u_1),\dots,\varphi_{n-1}(\psi(u_0),u_1,\dots,u_{n-1}))\in J(A,c)}\\
&\quad = \begin{cases}
1 & \abs{\frac{1}{n} \sum_{i=0}^{n-1} \left[1_{(u_0,\ldots, u_i) \in C_{i,\psi}(A)} - \pi(A) \right]}\geq c,\\
0 & \text{otherwise},
\end{cases}
\end{align*}
the assertion follows.
\end{proof}
\subsection{Discrepancy bounds}\label{subsec_main}
We show that for any $s\in\mathbb{N}$,
for any update function $\varphi \colon G \times [0,1]^s \to G$
of the transition kernel $K$,
for any generator function $\psi \colon [0,1]^s \to G$ of
initial distribution $\nu$ with
$\frac{d \nu}{d \pi}\in L_2$
and every natural number $n$ there exists a finite sequence
$u_0, u_1, \ldots, u_{n-1} \in [0,1]^s$
such that the star-discrepancy of
$S_n$, given by \eqref{eq: x_i_by_driver_seq}, converges approximately with order $n^{-1/2}$.
The main idea to prove the existence result is to use probabilistic arguments.
We apply a Hoeffding inequality for variance bounding Markov chains and show that
for a fixed test set the probability
of point sets with small $\Delta_{n,A,\varphi,\psi}$, see \eqref{al: loc_disc_star_disc}, is large.
We then extend this result to all sets in the $\delta$-cover using the union bound and finally to all test sets.
The result shows that if the finite driver sequence is chosen at random from the uniform distribution, most choices satisfy the Monte Carlo rate
of convergence of the discrepancy for the induced point set $S_n$.
\begin{theorem}\label{thm_main}
Let $K$ be a reversible transition kernel with respect to $\pi$
and $\nu$ be a distribution on $(G,\mathcal{B}(G))$ with $\frac{d\nu}{d\pi}\in L_2$.
Assume that $P$, the Markov operator of $K$, is variance bounding.
Let $\mathscr{A}\subseteq \mathcal{B}(G)$ be a set of test sets
and for every $\delta > 0$ assume that there exists a set $\Gamma_\delta \subseteq \mathcal{B}(G)$
with $|\Gamma_\delta| < \infty$ such that $\Gamma_\delta$ is a $\delta$-cover of $\mathscr{A}$
with respect to $\pi$.
Then, for any update function $\varphi \colon G \times [0,1]^s \to G$ of $K$,
and any generator function $\psi \colon [0,1]^s \to G$ of $\nu$,
there exists a driver sequence $u_0, u_1, \ldots, u_{n-1} \in [0,1]^s$
such that $S_n=\{x_1,\dots,x_n\}$ given by $x_1=\psi(u_0)$ and
\begin{equation*}
x_{i+1} = x_i(x_1) = \varphi(x_{i};u_i) = \varphi_i(x_1; u_1,\dots,u_i), \quad i=1,\dots,n-1,
\end{equation*}
satisfies
\begin{align} \label{al: first_disc_bound}
D^\ast_{\mathscr{A}, \pi}(S_n)
& \le \sqrt{\frac{1 + \Lambda_0}{1-\Lambda_0}}\cdot\frac{\sqrt{ 2 \log( |\Gamma_\delta|^2 \norm{\frac{d \nu}{d \pi}}_{2} )}}{\sqrt{n}} + \delta,
\end{align}
with $\Lambda_0 = \max\{0,\Lambda\}$ and $\Lambda$ defined in \eqref{eq: Lambda_variance_bounding}.
\end{theorem}
\begin{remark}\label{rem_delta_cover}
In Lemma~\ref{lem_delta_cover_ex2} in Section~\ref{subsec_delta}
we show
for the set of test sets of axis parallel boxes that for any $\delta > 0$
there exists a $\delta$-cover with
$|\Gamma_\delta| = \mathcal{O}(\delta^{-d/ (1-\varepsilon)})$ for any $\varepsilon > 0$.
Hence, for instance, by choosing $\delta = n^{-3/4}$,
we obtain that $|\Gamma_{n^{-3/4}}| = \mathcal{O}(n^{d})$,
where we used $\varepsilon=1/4$.
\end{remark}
\begin{proof}
Let $A\in\mathcal{B}(G)$.
By Lemma~\ref{lem: same_conc} and Proposition~\ref{prop: Hoeffd} we have for any $c_n \geq 0$ that
\begin{equation} \label{eq: hoeffd_for_us}
\mathbb{P}\left[\abs{ \Delta_{n,A,\varphi,\psi}} \leq c_n \right]
\geq 1- 2 \norm{\frac{d \nu}{d \pi}}_{2}
\exp\left( -\frac{1-\Lambda_0}{1+\Lambda_0} c_n^2 n\right).
\end{equation}
Let
\begin{equation*}
\widehat{\Gamma}_{\delta}=\{D\setminus C : C \subseteq A \subseteq D, \mbox{ and } C,D \in\Gamma_\delta \}.
\end{equation*}
If for all $A\in \widehat{\Gamma}_{\delta}$ we have
\begin{equation} \label{eq: for_all_sets}
\mathbb{P}\left[\abs{ \Delta_{n,A,\varphi,\psi}} \leq c_n \right] > 1 - \frac{1}{|\widehat{\Gamma}_\delta |},
\end{equation}
then there exists a finite sequence $u_0,\dots,u_{n-1} \in [0,1]^s$ such that
\begin{equation} \label{eq: gamma_prime}
\max_{A \in \widehat{\Gamma}_\delta}
\abs{ \Delta_{n,A,\varphi,\psi}(u_0,\dots,u_{n-1})} \leq c_n.
\end{equation}
For
\[
c_n=\sqrt{\frac{1+\Lambda_0}{1-\Lambda_0}}\cdot\frac{\sqrt{2 \log(2\, |\widehat{\Gamma}_\delta | \, \norm{\frac{d \nu}{d \pi}}_{2} )}}{\sqrt{n}}
\]
we obtain by \eqref{eq: hoeffd_for_us}
that \eqref{eq: for_all_sets} holds and
that there exists a finite sequence $u_0,\dots,u_{n-1} \in [0,1]^s$ such that
\eqref{eq: gamma_prime} is satisfied.
Now we extend the result from $\widehat{\Gamma}_\delta$ to $\mathscr{A}$.
By the $\delta$-cover we have for $A\in\mathscr{A}$, that there are $C,D \in \Gamma_\delta$ such that $C \subseteq A \subseteq D$
and $\pi(D\setminus C)\leq \delta$.
Hence
\begin{align*}
& \left|\frac{1}{n} \sum_{i=0}^{n-1} \left[ 1_{(u_0,\ldots, u_i) \in C_{i,\psi}(A) } - \pi(A) \right] \right|\\
=& \left|\frac{1}{n} \sum_{i=0}^{n-1} \left[ 1_{(u_0,\ldots, u_i) \in C_{i,\psi}(D) } - \pi(D) \right]
-\frac{1}{n} \sum_{i=0}^{n-1} \left[ 1_{(u_0,\ldots, u_i) \in C_{i,\psi}(D\setminus A) } - \pi(D\setminus A) \right] \right|\\
\leq & \left|\frac{1}{n} \sum_{i=0}^{n-1} \left[ 1_{(u_0,\ldots, u_i) \in C_{i,\psi}(D) } - \pi(D) \right] \right|\\
& +\left|\frac{1}{n} \sum_{i=0}^{n-1} \left[ 1_{(u_0,\ldots, u_i) \in C_{i,\psi}(D\setminus A) } - \pi(D\setminus A) \right] \right|.
\end{align*}
Set
\[
I_1=\left|\frac{1}{n} \sum_{i=0}^{n-1} \left[ 1_{(u_0,\ldots, u_i) \in C_{i,\psi}(D) } - \pi(D) \right] \right|
\]
and
\[
I_2=\left|\frac{1}{n} \sum_{i=0}^{n-1}
\left[ 1_{(u_0,\ldots, u_i) \in C_{i,\psi}(D\setminus A) } - \pi(D\setminus A) \right] \right|.
\]
Since $\emptyset \in \Gamma_\delta$ we have $D = D\setminus \emptyset \in \widehat{\Gamma}_{\delta}$ and therefore
\begin{equation*}
I_1\leq\max_{A \in \widehat{\Gamma}_\delta} \left| \Delta_{n,A,\varphi,\psi} \right| \leq c_n.
\end{equation*}
Furthermore
\begin{align*}
I_2
&\;\, = \left|\frac{1}{n} \sum_{i=0}^{n-1} 1_{(u_0,\ldots, u_i) \in C_{i,\psi}(D\setminus A)} - \pi(D\setminus C) + \pi(D\setminus C) -\pi(D\setminus A) \right| \\
&\;\, \leq \left|\frac{1}{n} \sum_{i=0}^{n-1}
\left[ 1_{(u_0,\ldots, u_i) \in C_{i,\psi}(D\setminus C)} - \pi(D\setminus C)\right]\right|
+ \left| \pi(D\setminus C) - \pi(D \setminus A) \right| \\
&\;\, \leq c_n + \delta.
\end{align*}
The last inequality follows by the $\delta$-cover property, \eqref{eq: gamma_prime}
and the fact that $D\setminus C \in \widehat{\Gamma}_\delta$.
Finally note that $|\widehat{\Gamma}_{\delta} | \leq |\Gamma_\delta |^2/2$, which completes the proof.
\end{proof}
\begin{remark}
We did not impose any regularity conditions on the update functions.
In particular, for any transition kernel $K$ on $(G,\mathcal{B}(G))$ there exists
an update function $\varphi \colon G \times [0,1] \to G$, with $s=1$, see for example
\cite[Lemma~2.22, p.~34]{Ka02}.
Thus, there exists a driver sequence $U_n=\{u_0,\dots,u_{n-1}\} \subset [0,1]$
such that $S_n$ driven by $U_n$ satisfies \eqref{al: first_disc_bound}.
\end{remark}
\begin{remark}\label{rem4}
The proof of Theorem~\ref{thm_main} shows that with probability greater than $0$,
there is a driver sequence $u_0, u_1, \ldots, u_{n-1} \in [0,1]^s$
which yields a Markov chain quasi-Monte Carlo point set satisfying the discrepancy bound.
By increasing the constant in the discrepancy bound \eqref{al: first_disc_bound},
we can increase this probability to $> 1/2$.
Assume now we are given two different Markov chains
with different transition kernels
satisfying the assumptions of Theorem~\ref{thm_main}.
Since with probability $>1/2$ there is a driver sequence
for each
transition kernel
satisfying the conclusion of Theorem~\ref{thm_main},
it follows that there is a single driver sequence $u_0, u_1,\ldots, u_{n-1}$
such that the discrepancy bound \eqref{al: first_disc_bound} holds for both
Markov chain-quasi Monte Carlo point sets
simultaneously.
\end{remark}
By Corollary~\ref{coro: D_U_almost_D_P_spec} and Theorem~\ref{thm_main}
we can also state an upper bound on the pull-back discrepancy.
\begin{theorem} \label{thm: push_back_push_forward}
Let the assumptions of Theorem~\ref{thm_main} be satisfied.
Then,
for any update function $\varphi \colon G \times [0,1]^s \to G$ of $K$ and
for any generator function $\psi \colon [0,1]^s \to G$ of $\nu$
there exists a driver sequence
$\mathcal{U}_n= \{u_0, u_1, \ldots, u_{n-1} \} \subset [0,1]^s$
such that
\begin{equation*}
D^\ast_{\mathscr{A},\psi ,\varphi}(\mathcal{U}_n)
\le \sqrt{\frac{1 + \Lambda_0}{1-\Lambda_0}}\cdot\frac{\sqrt{2\log( |\Gamma_\delta|^2 \norm{\frac{d \nu}{d \pi}}_{2} )}}{\sqrt{n}}
+ \frac{1-\Lambda_0^n}{n\cdot(1-\Lambda_0)} \norm{\frac{d\nu}{d\pi}-1}_{2}
+ \delta,
\end{equation*}
with $\Lambda_0 = \max\{0,\Lambda\}$ and $\Lambda$ defined in \eqref{eq: Lambda_variance_bounding}.
\end{theorem}
We refer to Remark~\ref{rem_delta_cover} and Lemma~\ref{lem_delta_cover_ex2}
for a relation between $\delta$ and $|\Gamma_\delta|$.
Thus, we showed the existence of a driver sequence with small pull-back discrepancy.
Note that by using Corollary~\ref{coro: D_U_almost_D_P_spec}
one could also argue the other way around: If one
can construct a sequence with small pull-back discrepancy
then the star-discrepancy of $S_n$ is also small.
\begin{remark}
Let us consider a special case of
Theorem~\ref{thm_main} and Theorem~\ref{thm: push_back_push_forward}.
Namely, let us assume that we can sample with respect to $\pi$.
Thus, we set $\nu = \pi$ and $K(x,A) = \pi(A)$ for any $x\in G$, $A\in \mathcal{B}(G)$.
Then, for any update function $\varphi$
of $K$ and generator function $\psi$ of $\pi$ we have
\[
D^\ast_{\mathscr{A}, \pi}(S_n) =
D^\ast_{\mathscr{A},\psi ,\varphi}(\mathcal{U}_n)
\leq \frac{\sqrt{2 \log \abs{\Gamma_\delta}^2 }}{ \sqrt{n}} + \delta,
\]
since $\Lambda_0 = \Lambda = 0$.
This is essentially the same as Theorem~1 in \cite{HeNoWaWo01} in their setting.
However, it is not as elaborate as Theorem~4 in \cite{HeNoWaWo01},
which is based on results by Talagrand~\cite{Ta94} and Haussler~\cite{Ha95}.
We do not know a version of these results which apply to Markov chains
(such a result could yield an improvement of Theorems~\ref{thm_main} and \ref{thm: push_back_push_forward}).
\end{remark}
\subsection{Burn-in period} \label{subsec_burn_in}
For Markov chain Monte Carlo a burn-in period is used to
reduce the bias of the initial distribution. We show how a burn-in
changes the discrepancy bound of Theorem~\ref{thm: push_back_push_forward}.
Let us introduce the following notation.
Let $\varphi\colon G \times [0,1]^s \to G$ and $\psi \colon [0,1]^s \to G$
be measurable functions.
Let $n_0,n\in \mathbb{N}$, let
\[
\mathcal{U}_{n_0,n} = \{ u_0,\dots,u_{n_0},u_{n_0+1},\dots,u_{n_0+n-1} \}
\subset [0,1]^s
\]
and assume that
$S_{[n_0,n]} = \{ x_{n_0+1},\dots, x_{n_0+n} \} \subseteq G$
is given by \eqref{eq: x_i_by_driver_seq}, i.e.
\begin{equation*}
x_{i+1} = x_{i+1}(x_1)
= \varphi(x_{i};u_i) = \varphi_i(x_1;u_1,\dots,u_i),
\quad i=1,\dots,n_0+n-1,
\end{equation*}
where $x_1=\psi(u_0)$.
As before $\psi$ might be considered as a generator function
and $\varphi$ might be considered as an update function.
We now define a discrepancy measure on the driver sequence where the
burn-in period is taken into account.
We call it \emph{pull-back discrepancy with burn-in}.
\begin{definition}[Pull-back discrepancy with burn-in]
Let $C_{i,\psi}(A)$
for $A\in\mathcal{B}(G)$
and $i\in \mathbb{N}\cup \{0\}$ be
defined as in \eqref{eq: C_i_psi}.
Define the local discrepancy function with burn-in by
\[
\Delta^{\text{\rm loc}}_{n_0,n,A,\psi,\varphi}(\mathcal{U}_{n_0,n})
=\frac{1}{n} \sum_{i=n_0}^{n_0+n-1} \left[1_{(u_0,\ldots, u_i) \in C_{i,\psi}(A)} -
\lambda_{(i+1)s}(C_{i,\psi}(A)) \right].
\]
Let $\mathscr{A} \subseteq \mathcal{B}(G)$ be a set of test sets.
Then we define the discrepancy of the driver sequence by
\begin{equation*}
D^{\ast}_{n_0, \mathscr{A},\psi,\varphi}(\mathcal{U}_{n_0,n})
= \sup_{A \in \mathscr{A}} \left|\Delta^{\text{\rm loc}}_{n_0,n,A,\psi,\varphi}(\mathcal{U}_{n_0,n}) \right|.
\end{equation*}
We call $D^\ast_{n_0,\mathscr{A},\psi,\varphi}(\mathcal{U}_{n_0,n})$
pull-back discrepancy with burn-in of $\mathcal{U}_{n_0,n}$.
\end{definition}
By adapting Proposition~\ref{prop: Hoeffd}
and Lemma~\ref{lem: same_conc} to the setting with burn-in we obtain,
by the same steps as in the proof of Theorem~\ref{thm_main}, a bound
on the star-discrepancy for $S_{[n_0,n]}$.
Further, adapting Theorem~\ref{thm: est_discr} and
Corollary~\ref{coro: D_U_almost_D_P_spec} to the burn-in leads to
a bound on $D^{\ast}_{n_0, \mathscr{A},\psi,\varphi}(\mathcal{U}_{n_0,n})$
for a certain set $\mathcal{U}_{n_0,n}$.
\begin{theorem}
Let the assumptions of Theorem~\ref{thm_main} be satisfied.
Then, for any update function $\varphi \colon G \times [0,1]^s \to G$ of $K$ and any
generator function $\psi \colon [0,1]^s \to G$ of $\nu$
there exists a driver sequence
\[
\mathcal{U}_{n_0,n}= \{u_0, u_1, \ldots, u_{n_0+n-1} \} \subset [0,1]^s
\]
such that
\begin{align*}
D^\ast_{\mathscr{A}, \pi}(S_{[n_0,n]})
& \le \sqrt{\frac{1 + \Lambda_0}{1-\Lambda_0}}\cdot
\frac{\sqrt{2 \log\left( |\Gamma_\delta|^2 \norm{\frac{d (\nu P^{n_0})}{d \pi}}_{2} \right)}}{\sqrt{n}} + \delta,
\end{align*}
with $\Lambda_0 = \max\{0,\Lambda\}$ and $\Lambda$ defined in \eqref{eq: Lambda_variance_bounding}.
If the Markov operator $P$ has an absolute $L_2$-spectral gap we have
\begin{equation} \label{eq: burn_in_mixed_spectral_gap}
\begin{split}
D^\ast_{n_0,\mathscr{A},\psi ,\varphi}(\mathcal{U}_{n_0,n})
& \le \sqrt{\frac{1 + \Lambda_0}{1-\Lambda_0}}\cdot
\frac{\sqrt{2 \log(|\Gamma_\delta|^2 (1+ \beta^{n_0}\norm{\frac{d \nu}{d \pi}-1}_{2} )}}{\sqrt{n}} \\
& \qquad \qquad + \frac{(1-\Lambda_0^n)\beta^{n_0}}{n\cdot(1-\Lambda_0)} \norm{\frac{d\nu}{d\pi}-1}_{2}
+ \delta,
\end{split}
\end{equation}
with $\beta = \norm{P}_{L_2^0 \to L_2^0}$, see Definition~\ref{def: abs_spec_gap}.
In particular, by $\Lambda \leq \Lambda_0 \leq \beta < 1$ and $|\Lambda| \le \beta$, we deduce
\begin{align} \label{al: burn_in_spectral_gap}
D^\ast_{n_0,\mathscr{A},\psi ,\varphi}(\mathcal{U}_{n_0,n}) \leq
\frac{4 \sqrt{\log\left( |\Gamma_\delta|^2 (1+ \beta^{n_0}\norm{\frac{d \nu}{d \pi}-1}_{2} )\right)}}{\sqrt{n \cdot (1-\beta)}}
+ \frac{2 \beta^{n_0} \norm{\frac{d\nu}{d\pi}-1}_{2}}{n\cdot(1-\beta)}
+ \delta.
\end{align}
\end{theorem}
Equations \eqref{eq: burn_in_mixed_spectral_gap} and \eqref{al: burn_in_spectral_gap}
reveal that the burn-in $n_0$ can eliminate the influence of the initial state induced by $\psi$ under the assumption that there exists an absolute $L_2$-spectral gap.
A variance bounding transition kernel
is not enough, since it could be periodic and then $\nu P^{n_0}$ would not
converge to $\pi$ at all.
\section{Application}\label{sec_application}
We consider the set of test sets $\mathscr{B}$
which consists of all axis parallel boxes anchored at $-\infty$
restricted to $G\subseteq \mathbb{R}^d$, i.e.
\[
\mathscr{B}= \{ (-\infty,x)_ G \colon x\in \mathbb{R}^d \},
\]
with $(-\infty,x)_G=(-\infty,x)\cap G$ and $(-\infty,x)= \Pi_{i=1}^d (-\infty,x_i)$. In the following we study the size of $\delta$-covers with respect to such rectangular boxes.
We then focus on the application of Theorem~\ref{thm_main} and
state the relation between the discrepancy and the
error of the computation
of expectations.
The Metropolis algorithm with ball walk proposal
provides an example where one can see that
the existence result shows an error bound which depends
polynomially on the dimension $d$.
\subsection{Delta-cover with respect to distributions}\label{subsec_delta}
We now use an explicit version of a result due to Beck~\cite{Be84},
for a proof and
further details we refer to \cite[Theorem~1]{AiDi13}. We state it as a lemma.
\begin{lemma}\label{lem_Beck}
Let $([0,1]^d, \mathcal{B}([0,1]^d), \mu)$ be a probability space.
Let the set of test sets $\mathscr{A} = \{[0,y) \mid y \in [0,1]^d\}$, with
$[a,b)=\Pi_{j=1}^d [a_j,b_j)$ for
$a,b\in \mathbb{R}^d$,
be the set of anchored boxes. Let ${\rm supp }\mu$ be the closure of
\[
\{ x \in [0,1]^d: \forall \mbox{open neighborhoods } B \mbox{ of } x: \mu(B) > 0\}
\]
Then, for any $r\in\mathbb{N}$ there exists a set $Z_r=\{ z_1,\dots,z_r \}$
with $z_1,\dots,z_r \in {\rm supp }\mu$
such that
\begin{equation} \label{eq: disc_bound_AD}
D^\ast_{\mathscr{A}, \mu}(Z_r) \le 63 \sqrt{d}\, \frac{(2 + \log_2 r)^{(3d+1)/2}}{r}.
\end{equation}
\end{lemma}
Note that $\log_2$ denotes the dyadic and $\log$ the natural logarithm.
\begin{proof}
The assertion
follows by \cite[Theorem~3]{AiDi13}
with $\P=\mu$, $X=[0,1]^d\cap {\rm supp }\,\mu$, $\mathscr{C}= \{ [0,y)\cap {\rm supp }\,\mu \mid y\in \mathbb{Q}^d \} $.
This implies a version
of \cite[Corollary~1]{AiDi13}, thus a version of \cite[Theorem~1]{AiDi13},
with $x_1,\dots, x_N \in {\rm supp }\mu$.
\end{proof}
By a linear transformation we extend the result to general, bounded state spaces
$G \subset \mathbb{R}^d$.
\begin{corollary} \label{coro: boxes_disc_AD}
Let $G \subset \mathbb{R}^d$ be a bounded, measurable set
and let $(G, \mathcal{B}(G), \pi)$ be a probability space.
Let the set of test sets
$\mathscr{B} = \{(-\infty, x)_ G \mid x \in \mathbb{R}^d\}$.
Then, for any $r\in\mathbb{N}$ there exists a set
$S_r=\{ x_1,\dots,x_r \} \subseteq G$
such that
\begin{equation*}
D^\ast_{\mathscr{B}, \pi}(S_r) \le 63 \sqrt{d}\, \frac{(2 + \log_2 r)^{(3d+1)/2}}{r}.
\end{equation*}
\end{corollary}
\begin{proof}
Since $G$ is bounded there exist $a,b \in \mathbb{R}^d$ such that
$G\subseteq \prod_{j=1}^d [a_j, b_j]$.
There is a linear transformation $T\colon \prod_{j=1}^d [a_j, b_j] \to [0,1]^d$
which induces a probability measure $\mu$ on $([0,1]^d, \mathcal{B}([0,1]^d))$
with $\pi(A)=\mu(T(A))$ for $ A \in \mathcal{B}(G)$.
In particular, for $A \in \mathcal{B}([0,1]^d \setminus T(G))$ we have $\mu(A) = 0$.
By Lemma~\ref{lem_Beck} we have that there
exists a set $Z_r=\{z_1,\dots,z_r\} \subseteq {\rm supp}\, \mu$
such that \eqref{eq: disc_bound_AD} is satisfied.
Let $x_i = T^{-1}(z_i)$ for $i=1,\dots,r$
and for $z\in [0,1]^d$ let $x=T^{-1}(z)$.
Then
\begin{align*}
\frac{1}{r} \sum_{i=1}^r 1_{(-\infty,x)_G}(x_i) - \pi((-\infty,x)_G)
= & \frac{1}{r} \sum_{i=1}^r 1_{[0,z) \cap T(G)}(z_i) - \mu([0,z) \cap T(G) ).
\end{align*}
Since $z_1,\dots,z_r \in {\rm supp} \mu \subset T(G)$ and $\mu(A) = 0$ for
$A \in \mathcal{B}([0,1]^d \setminus T(G))$ we have
\begin{align*}
\frac{1}{r} \sum_{i=1}^r 1_{(-\infty,x)_G}(x_i) - \pi((-\infty,x)_ G)
= & \frac{1}{r} \sum_{i=1}^r 1_{[0,z)}(z_i) - \mu([0,z) ).
\end{align*}
By taking the supremum over the test sets on the right-hand side and using
\eqref{eq: disc_bound_AD} the assertion follows.
\end{proof}
As in \cite[Lemma~4]{DiRuZh13} a point set which satisfies a discrepancy bound
can be used to construct a $\delta$-cover.
The idea is to define
for each subset of the point set a
minimal and maximal set for the $\delta$-cover, see \cite[Lemma~4]{DiRuZh13}.
To simplify the bound of Corollary~\ref{coro: boxes_disc_AD},
for any $r \in \mathbb{N}$ and $0 < \varepsilon < 1$ we have
\begin{equation*}
\frac{ \left(2 + \log_2 r \right)^{(3d+1)/2}}{r} \le r^{\varepsilon-1} C_{\varepsilon,d},
\end{equation*}
where
\begin{equation}\label{eq_C}
C_{\varepsilon,d} = \max_{x \ge 1} \frac{(2+\log_2 x)^{(3d+1)/2}}{x^\varepsilon}
= 4^{\varepsilon} \left(\frac{3d+1}{2\mathrm{e} \varepsilon \log 2} \right)^{(3d+1)/2}.
\end{equation}
With this notation we obtain the following result.
\begin{lemma}\label{lem_delta_cover_ex2}
Let $G \subset \mathbb{R}^d$ be a bounded measurable set and
let $\pi$ be a probability measure on $(G,\mathcal{B}(G))$ which
is absolutely continuous with respect to the Lebesgue measure.
For the set
$\mathscr{B} = \{(-\infty,x)_G \mid x\in \mathbb{R}^d \}$,
any $0 < \delta \le 1$ and $0 < \varepsilon < 1$,
there is a $\delta$-cover $\Gamma_\delta$ of $\mathscr{B}$
with respect to $\pi$ with
\[
|\Gamma_\delta| \le \left(2 + \left\lceil (2 C_{\varepsilon,d} \delta^{-1})^{1/(1-\varepsilon)} \right \rceil \right)^d,
\]
where $C_{\varepsilon,d}$ is given by \eqref{eq_C}.
\end{lemma}
\begin{proof}
The proof of the assertion follows essentially by the same steps as the proof
of \cite[Lemma~4]{DiRuZh13}. The only difference is that we use the discrepancy
bound of Corollary~\ref{coro: boxes_disc_AD} instead of \cite[Theorem~4]{HeNoWaWo01}.
\end{proof}
The dependence of the size of the $\delta$-cover on $\delta$ is arbitrarily
close to order $\delta^{-d}$
in Lemma~\ref{lem_delta_cover_ex2}, whereas in \cite[Lemma~4]{DiRuZh13}
it is of order $\delta^{-2 d}$.
Furthermore, the constant in Lemma~\ref{lem_delta_cover_ex2}
is fully explicit (one can choose $0 < \varepsilon < 1$ to obtain
the best bound on the size of the $\delta$-cover).
By Theorem~\ref{thm_main} and Lemma~\ref{lem_delta_cover_ex2} we obtain the following result.
\begin{corollary} \label{coro_main}
Let $G\subset \mathbb{R}^d$ be a bounded set. Let $K$ be a reversible transition
kernel with respect to $\pi$ and $\nu$ be a distribution on $(G,\mathcal{B}(G))$
with $\frac{d\nu}{d\pi}\in L_2$.
Assume that $P$, the Markov operator of $K$,
is variance bounding.
Further, let $\mathscr{B}=\{ (-\infty,x)_G \mid x\in \mathbb{R}^d \}$
be the set of test sets.
Then,
for any update function $\varphi \colon G \times [0,1]^s \to G$ of $K$,
any generator function $\psi \colon [0,1]^s \to G$ of $\nu$ and
for all $n\geq16$, there exists a driver sequence $u_0,\dots,u_{n-1} \in [0,1]^s$ such that
$S_n=\{ x_1,\dots,x_n \}$ given by \eqref{eq: x_i_by_driver_seq} satisfies
\begin{equation}
D^*_{\mathscr{B},\pi}(S_n)
\leq \sqrt{\frac{1+\Lambda_0}{1-\Lambda_0}} \cdot
\frac{\sqrt{2}\,(\log\norm{\frac{d\nu}{d\pi}}_{2}+d \log n + 3d^2 \log(5d))^{1/2}}{\sqrt{n}}
+\frac{8}{n^{3/4}},
\end{equation}
with $\Lambda_0=\max\{ \Lambda,0\}$.
\end{corollary}
\begin{proof}
Let $\varepsilon=1/4$.
Thus $C_{1/4,d}=\sqrt{2}(\frac{6d+2}{\rm{e}\log2})^{(3d+1)/2}$ and
$\abs{\Gamma_\delta} \leq (16 \delta^{-4/3} (5d)^{3d})^d$.
By $\delta=8/n^{3/4}$ and Theorem~\ref{thm_main} the assertion follows.
\end{proof}
Let us discuss the result. The factor depending on $\Lambda_0$ is the penalty
for the convergence of the Markov chain. The term $\log\norm{\frac{d\nu}{d\pi}}_{2}$
shows the dependence on $\psi$ and the additional summand $\frac{8}{n^{3/4}}$
comes from the $\delta$-cover approximation. The rest is basically as in
\cite[Theorem~1]{HeNoWaWo01}.
\subsection{Integration error}
\label{subsec: int_err}
In this section we state a relation between a reproducing kernel Hilbert space and
the star-discrepancy.
As in \cite[Appendix~B]{DiRuZh13} we define a reproducing kernel $Q$ by
\begin{equation*}
Q(x,y) = 1 + \int_{\mathbb{R}^d} 1_{(-\infty, z)_G}(x)\, 1_{(-\infty, z)_G}(y) \, \rho(\mathrm{d} z),
\end{equation*}
where $\rho$ is a finite measure on $\mathbb{R}^d$,
i.e. $\int_{\mathbb{R}^d} \rho(\mathrm{d} z) < \infty$.
The function $Q$ uniquely defines a reproducing kernel Hilbert space $H_2 = H_2(Q)$ of functions defined on $\mathbb{R}^d$.
Reproducing kernel Hilbert spaces were studied in detail in \cite{Ar50}.
It is also known that the functions $f$ in $H_2$ permit the representation
\begin{equation}\label{eq_f_rep}
f(x) = f_0 + \int_{\mathbb{R}^d} 1_{(-\infty,z)_G}(x) \widetilde{f}(z) \,\rho(\mathrm{d} z),
\end{equation}
for some $f_0 \in \mathbb{C}$
and function $\widetilde{f} \in L_2( \mathbb{R}^d, \rho )$,
see for instance \cite[Theorem~4.21, p.~121]{StCh08} or follow the same arguments as in \cite[Appendix~A]{BrDi13}.
The inner product in $H_2$ is given by
\begin{equation*}
\langle f, g \rangle = f_0\, \overline{g_0}
+ \int_{\mathbb{R}^d} \widetilde{f}(z)\, \overline{\widetilde{g}(z)}\, \rho(\mathrm{d} z).
\end{equation*}
With these definitions we have the reproducing property
\begin{equation*}
\langle f, Q(\cdot, y)\rangle
= f_0 + \int_{\mathbb{R}^d} \widetilde{f}(z) 1_{(-\infty, z)_G}(y) \rho(\mathrm{d} z) = f(y).
\end{equation*}
For $1 \le q \le \infty$ we also define the space $H_q$ of
functions of the form \eqref{eq_f_rep}
for which $\widetilde{f} \in L_q(G, \rho)$, with finite norm
\begin{equation}\label{norm_H1}
\|f\|_{H_q} = \left(|f_0|^q + \int_{\mathbb{R}^d} |\widetilde{f}(z)|^q \rho(\mathrm{d} z) \right)^{1/q}.
\end{equation}
The following result concerning the integration error
in $H_q$ is proven in \cite[Theorem~3]{DiRuZh13}.
\begin{theorem}\label{thm_int_error}
Let $G \subseteq \mathbb{R}^d$ and $\pi$ be a probability measure on $G$.
Further let $\mathscr{B} = \{(-\infty, x)_G: x \in \mathbb{R}^d\}$.
We assume that $1 \leq p, q \leq \infty$ with $1/p + 1/q = 1$.
Then for $Z_n=\{z_1, z_2, \ldots, z_n \} \subseteq G$
and for all $f \in H_q$ we have
\begin{equation*}
\left|\int_G f(z) \pi(\mathrm{d} z) - \frac{1}{n} \sum_{i=1}^n f(z_i)\right|
\leq \|f\|_{H_q} D^\ast_{p, \mathscr{B}, \pi}(Z_n),
\end{equation*}
where
\begin{equation*}
D^\ast_{p, \mathscr{B}, \pi}(Z_n)
= \left( \int_{\mathbb{R}^d} \left| \int_G 1_{(-\infty, z)_G}(y) \pi(\mathrm{d} y)
- \frac{1}{n} \sum_{i=1}^n 1_{(-\infty, z)_G}(z_i) \right|^p \rho(\mathrm{d} z) \right)^{1/p},
\end{equation*}
and for $p=\infty$ let
\begin{equation*}
D^\ast_{\mathscr{B}, \pi}(Z_n)
:= D^\ast_{\infty, \mathscr{B}, \pi}(Z_n) = \sup_{z \in \mathbb{R}^d}
\left| \int_G 1_{(-\infty, z)_G}(y) \pi(\mathrm{d} y)
- \frac{1}{n} \sum_{i=1}^n 1_{(-\infty, z)_G}(z_i) \right|.
\end{equation*}
\end{theorem}
\begin{corollary}[Markov chain Koksma-Hlawka inequality]\label{cor_KH_inequality}
Assume that the conditions of Corollary~\ref{coro: D_U_almost_D_P_spec}
are satisfied.
Further let $\mathscr{B} = \{(-\infty, x)_G: x \in \mathbb{R}^d\}$.
Let $H_1$ denote the space of functions $f\colon\mathbb{R}^d \to \mathbb{C}$
with finite norm given by \eqref{norm_H1}.
Then, for any update function $\varphi\colon G \times [0,1]^s \to G$
of $K$ and any generator function $\psi \colon [0,1]^s \to G$ of $\nu$
we have, with driver sequence
$\mathcal{U}_n=\{u_0, u_1, \ldots, u_{n-1}\} \subset [0,1]^s$
and $S_n$ given by \eqref{eq: x_i_by_driver_seq}, that
\begin{align*}
&\left|\int_G f(x) \pi(\mathrm{d} z) - \frac{1}{n} \sum_{i=1}^n f(x_i) \right| \\
& \qquad \qquad\le
\left( D^\ast_{\mathscr{B}, \psi, \varphi}(\mathcal{U}_n)
+ \frac{1-\Lambda_0^n}{n \cdot(1-\Lambda_0)} \norm{\frac{d\nu}{d\pi}-1}_{2} \right) \|f\|_{H_1},
\end{align*}
with
$\Lambda_0 = \max\{ 0,\Lambda \}$, where $\Lambda$ is defined in \eqref{eq: Lambda_variance_bounding}.
\end{corollary}
In the spirit of
Remark~\ref{rem: direc_simulation} we obtain for $K(x,A)=\pi(A)$ that $\Lambda=0$.
Further, if $\nu=\pi$ we have the Koksma-Hlawka inequality (cf. \cite[p. 151, Theorem~5.5]{KN})
\begin{equation*}
\left|\int_G f(x) \pi(\mathrm{d} x) - \frac{1}{n} \sum_{i=1}^n f(x_i) \right|
\le D^\ast_{\mathscr{B}, \psi, \varphi}(\mathcal{U}_n)\, \|f\|_{H_1}.
\end{equation*}
\subsection{Metropolis algorithm with ball walk proposal}
The goal of this subsection is the application of the previously
developed theory
to an example. Let us assume that $G=\mathbb{B}_d$
is the Euclidean unit ball, i.e.
$\mathbb{B}_d=\{ x\in \mathbb{R}^d \mid \|x \|:= (\sum_{i=1}^d \abs{x_i}^2)^{1/2} \leq1 \}$.
Let $\rho \colon \mathbb{B}_d \to (0,\infty)$ be integrable with respect to the Lebesgue measure. We define the
distribution $\pi_\rho$ on $(\mathbb{B}_d,\mathcal{B}(\mathbb{B}_d))$ by
\[
\pi_\rho(A) = \frac{\int_A \rho(x)\, {\rm d} x}{\int_{\mathbb{B}_d} \rho(x)\, {\rm d} x}.
\]
The goal is to compute
\[
\mathbb{E}_{\pi_\rho}(f)
= \int_{\mathbb{B}_d} f(x)\, \pi_\rho({\rm d} x)
= \frac{\int_{\mathbb{B}_d} f(x) \rho(x)\, {\rm d} x}{\int_{\mathbb{B}_d} \rho(x)\, {\rm d} x},
\]
for functions $f\colon \mathbb{B}_d \to \mathbb{R}$ which are integrable with respect to $\pi_\rho$.
Note that for an approximation of $\mathbb{E}_{\pi_\rho}(f)$ the functions $f$ and $\rho$ are part of
the input of a possible approximation scheme. We
assume that sampling directly with respect to $\pi_\rho$ is not feasible.
We use the Metropolis algorithm with ball walk proposal
to sample approximately according to $\pi_\rho$.
Let $\gamma>0$, $x\in \mathbb{B}_d$
and $C\in \mathcal{B}(\mathbb{B}_d)$,
then the transition kernel of the $\gamma$ ball walk
is given by
\[
W_{\gamma}(x,C)
= \frac{{\lambda}_d(C\cap D_{\gamma}(x))}{ {\lambda}_d (D_\gamma (0))}
+ 1_{x\in A} \left[ 1 - \frac{{\lambda}_d(\mathbb{B}_d\cap D_{\gamma}(x))}{ {\lambda}_d (D_\gamma (0))}\right],
\]
where ${\lambda}_d$ denotes the $d$-dimensional Lebesgue measure and
$D_\gamma(x) = \{y \in \mathbb{R}^d \mid \|x-y\| \le \gamma \}$ denotes the Euclidean ball with radius
$\gamma$ around $x\in \mathbb{R}^d$.
The transition kernel of the Metropolis algorithm with ball walk proposal is
given by
\[
M_{\rho,\gamma}(x,C) = \int_{C} \theta(x,y)\, W_{\gamma}(x,{\rm d}y)
+ 1_{x\in A}\left[1 - \int_{\mathbb{B}_d} \theta(x,y)\,W_{\gamma}(x,{\rm d}y) \right],
\]
where $ \theta(x,y) = \min \{ 1 ,\rho(y)/\rho(x)\}$ is the so-called
acceptance probability.
The transition kernel $M_{\rho,\gamma}$ is reversible with respect
to $\pi_\rho$.
Now we provide update functions of the ball walk and the Metropolis algorithm with ball walk proposal.
Let $\mathbb{S}^{d-1} = \{ x\in \mathbb{R}^d \mid \|x\|=1 \}$ be the unit sphere in $\mathbb{R}^d$.
Let $\widetilde{\psi} \colon [0,1]^{d-1} \to \mathbb{S}^{d-1}$ be a generator
for the uniform distribution on the sphere, see for instance \cite{FaWa94}.
Then, $\psi_\gamma \colon [0,1]^d \to D_\gamma (0)$ given by
\begin{equation} \label{eq: psi_uniform}
\psi_\gamma (\bar{u}) = \gamma \, v_d^{1/d} \widetilde{\psi}(v_1,\dots, v_{d-1}),
\end{equation}
with $\bar{u}=(v_1,\dots,v_d)\in[0,1]^d$,
is a generator for the uniform distribution in $D_\gamma(0)$ (the Euclidean ball with radius $\gamma$ around $0$).
Thus, an update function
$\varphi_{W,\gamma}\colon \mathbb{B}_d \times [0,1]^d \to \mathbb{B}_d$
of the $\gamma$ ball walk, with $\bar{u}=(v_1,\dots,v_d)\in[0,1]^d$, is
\[
\varphi_{W,\gamma}(x,\bar{u}) = \begin{cases}
x+\psi_\gamma (\bar{u}) & x+\psi_\gamma (\bar{u}) \in \mathbb{B}_d \\
x & \mbox{otherwise}.
\end{cases}
\]
This leads to an update function $\varphi_{M,\gamma, \rho} \colon \mathbb{B}_d \times [0,1]^{d+1} \to \mathbb{B}_d$
of the Metropolis algorithm with ball walk proposal.
Let
\[
A(x;\bar{u}) = \min\{ 1, \rho(\varphi_{W,\gamma}(x,\bar{u}))/\rho(x) \},
\]
then an update function for the Metropolis algorithm with ball walk proposal is
\begin{equation} \label{eq: Metro_update}
\varphi_{M,\gamma, \rho}(x,u) = \begin{cases}
\varphi_{W,\gamma}(x,v_1,\dots,v_d) & v_{d+1} \leq A(x,v_1,\dots,v_d)\\
x & v_{d+1} > A(x,v_1,\dots,v_d),
\end{cases}
\end{equation}
where $u=(v_1,\dots,v_{d+1})\in[0,1]^{d+1}$ and $x\in \mathbb{B}_d$.
Thus, we have an update function of $W_\gamma$.
For the convenience of the reader we provide a transition
of the Metropolis algorithm with
ball walk proposal
from $x$ to $y$ with driving point $(v_1,\dots,v_{d+1}) \in [0,1]^{d+1}$
in algorithmic form:
\begin{algorithm}
Metropolis algorithm with ball walk proposal\\[1ex]
\begin{tabular}{ll}
Input: & driving point $(v_1,\dots,v_{d+1}) \in [0,1]^{d+1}$, and \\
& current state $x\in\mathbb{B}_d$; \\
Output:& next state $y\in\mathbb{B}_d$;
\end{tabular}
\begin{enumerate}
\item Compute $z:=\gamma\, v_d^{1/d}\, \widetilde \psi (v_1,\dots,v_{d-1})$
where $\widetilde \psi$ is a generator function for the uniform
distribution on $\mathbb{S}^{d-1}$.
\item
\subitem a) If $x+z \in \mathbb{B}_d$ and $v_{d+1} \le \min \left\{1,\rho(x+z)/\rho(x) \right\}$, then $y:= x + z$.
\subitem b) Otherwise $y := x$.
\item Return $y$.
\end{enumerate}
\end{algorithm}
We assume that the functions $f \colon \mathbb{B}_d \to \mathbb{R}$
and $\rho \colon \mathbb{B}_d \to (0,\infty)$
have some additional structure. Let $f\in H_1$ with $\|f\|_{H_{1}} \leq 1$,
where $H_{1}$ is defined in Subsection~\ref{subsec: int_err}.
For $\alpha > 0$ let $\rho \in \mathcal{R}_{\a,d}$
if the following conditions are satisfied:
\begin{enumerate}[(i)]
\item\label{it: log_conc} $\rho$ is log-concave, i.e. for all $\lambda \in (0,1)$ and for all $x,y \in \mathbb{B}_d$ holds
\[
\rho(\lambda x + (1-\lambda )y) \geq \rho(x)^\lambda \rho(y)^{1-\lambda}.
\]
\item\label{it: log_lip} $\rho$ is log-Lipschitz continuous with $\alpha$, i.e.
\[
\abs{\log \rho(x)-\log \rho(y)} \leq \alpha \| x-y \|.
\]
\end{enumerate}
Thus
\begin{equation} \label{eq: def_fct_class}
\mathcal{R}_{\a,d} = \{ \rho \colon \mathbb{B}_d \to (0,\infty) \mid
\rho\;\mbox{log-concave},
\abs{\log \rho(x)-\log \rho(y)} \leq \alpha \| x-y \|
\}.
\end{equation}
Next we provide a lower bound for $\Lambda_{\gamma,\rho}$, defined as in \eqref{eq: Lambda_variance_bounding}
for the transition kernel $M_{\gamma,\rho}$, where the density $\rho$ is log-concave and log-Lipschitz.
The result follows by \cite[Corollary~1, Lemma~13]{MaNo07}.
\begin{proposition} \label{prop: low_conduct}
Let us assume that $\rho\in \mathcal{R}_{\a,d}$. Further let
\[
\gamma^* = \min\{1/\sqrt{d+1},1/\alpha\}.
\]
Then
\begin{equation} \label{eq: lower_bd_var_bound}
1 - \Lambda_{\gamma^*,\rho} \geq \frac{3.125 \cdot 10^{-6} }{d+1} \min\left\{ \frac{1}{d+1},\frac{1}{\alpha} \right\}.
\end{equation}
\end{proposition}
The combination of Proposition~\ref{prop: low_conduct},
Theorem~\ref{thm_int_error}, Lemma~\ref{lem_delta_cover_ex2}
and Corollary~\ref{coro_main} lead to
the following error bound for the computation
of $\mathbb{E}_{\pi_\rho}(f)$ for $f\in H_1$ and $\rho\in \mathcal{R}_{\a,d}$.
\begin{theorem}
Let $\nu$ be the uniform distribution
on $(\mathbb{B}_d,\mathcal{B}(\mathbb{B}_d))$ with generator function $\psi_1$,
see \eqref{eq: psi_uniform}.
Let
\[
\gamma^* = \min\{1/\sqrt{d+1},1/\alpha\}
\]
and recall that $\varphi_{M,\gamma^*,\rho}$ is an update function
of the Metropolis algorithm with ball walk proposal,
see \eqref{eq: Metro_update}.
Then, for all $n\geq16$ and any $\rho\in\mathcal{R}_{\a,d}$
there exists a driver sequence
$u_0,u_1,\dots,u_{n-1} \in [0,1]^{d+1}$ such that $S_n=\{ x_1,\dots,x_n \}$
given by
\begin{align*}
x_1 & = \psi_1(\bar{u}_0)\\
x_{i+1} & = \varphi_{M,\gamma^*,\rho} (x_i;u_i), \qquad i=1,\dots,n-1,
\end{align*}
with $\bar{u}_0=(v_1,\dots,v_d)$ where $u_0=(v_1,\dots,v_d,v_{d+1})$, satisfies
\begin{align*}
& \sup_{f \in H_1, \Vert f \Vert_{H_1} \leq 1} \abs{ \mathbb{E}_{\pi_{\rho}}(f)-\frac{1}{n} \sum_{i=1}^n f(x_i) } \\
&\qquad \leq \frac{5000 \sqrt{d} \max\{ \sqrt{2d}, \sqrt{\alpha} \}
\left( \alpha+d \log n + 3d^2 \log(5d)\right)^{1/2}
}{\sqrt{n}} + \frac{8}{n^{3/4}}.
\end{align*}
\end{theorem}
\begin{proof}
By
\[
\frac{d\nu}{d \pi_{\rho}} (x) = \frac{\int_{\mathbb{B}_d} \rho(y)\, {\rm d}y}{\lambda_d(\mathbb{B}_d) \rho(x)},
\]
and by $\rho(x)/\rho(y) \leq \exp(2\alpha)$ for any $x,y\in \mathbb{B}_d$
we have $\|\frac{d\nu}{d\pi_\rho}\|_2 \leq \exp\alpha$.
Further, by Proposition~\ref{prop: low_conduct} we obtain
\[
1-\Lambda_0 \geq \frac{3.125\cdot 10^{-6}}{d+1} \min\left\{ \frac{1}{d+1}, \frac{1}{\alpha} \right \}.
\]
Thus by Corollary~\ref{coro_main} and Theorem~\ref{thm_int_error} the assertion follows.
\end{proof}
Let us emphasize that the theorem shows that for any $\rho \in \mathcal{R}_{\alpha,d}$
there exist a deterministic algorithm where the error depends only polynomially
on the dimension $d$ and the Log-Lipschitz constant $\alpha$.
\section{Beyond the Monte Carlo rate} \label{sec: beyound_MC}
In the previous sections we have seen that
there exist deterministic driver sequences
which yield almost the Monte Carlo rate of convergence of $n^{-1/2}$.
Roughly speaking, the proof of Theorem~\ref{thm_main} reveals that,
if the driver sequence is chosen at random from the uniform distribution
the discrepancy bound of \eqref{al: first_disc_bound}
is satisfied
with high probability.
In this section we use a stronger assumption
to achieve a better rate of convergence.
Again this result is an existence result.
We want to point out that
the proof of the
result
does not reveal any information on how to find driver sequences
which lead to good discrepancy bounds.
Its proof is based on an additional regularity condition of the
update function, the `anywhere-to-anywhere'
condition, and Corollary~\ref{coro: boxes_disc_AD}.
\begin{definition}
Let $\varphi:G \times [0,1]^s \to G$ be an update function of a
transition kernel $K$.
We say that $\varphi$ satisfies the
\emph{`anywhere-to-anywhere' condition} if for
all $x, y \in G$ there exists a $u \in [0,1]^s$ such that
\begin{equation*}
\varphi(x; u) = y.
\end{equation*}
\end{definition}
Now we use
the `anywhere-to-anywhere' condition to
reformulate Corollary~\ref{coro: boxes_disc_AD}.
We
obtain a bound on the
star-discrepancy for the Markov chain quasi-Monte Carlo construction.
\begin{corollary}\label{cor_higher_order}
Let $G \subset \mathbb{R}^d$ be a bounded,
measurable set and let $(G, \mathcal{B}(G), \pi)$ be a probability space.
Let the set of test sets
$\mathscr{B} = \{(-\infty, x) \cap G \mid x \in \mathbb{R}^d\}$
be the set of anchored boxes intersected with $G$.
Then, for any update function $\varphi \colon G \times [0,1]^s \to G$
of the transition kernel $K$ which satisfies
the `anywhere-to-anywhere' condition,
any surjective
function $\psi \colon [0,1]^s \to G$
and
for any $n\in\mathbb{N}$ there exists a driver sequence
$u_0, u_1, \dots u_{n-1} \in [0,1]^s$
such that $S_n=\{ x_1,\dots,x_n \}$
given by $x_1=\psi(u_0)$ and
\begin{equation*}
x_i = \varphi(x_{i-1}; u_i),
\qquad i=1,\dots,n-1,
\end{equation*}
satisfies
\begin{equation*}
D^\ast_{\mathscr{B}, \pi}(S_n) \le 63 \sqrt{d}\, \frac{(2 + \log_2 n)^{(3d+1)/2}}{n}.
\end{equation*}
\end{corollary}
The corollary states that if the `anywhere-to-anywhere'
condition is satisfied, in principle, we can get the same discrepancy
for the Markov chain quasi-Monte
Carlo construction as without using any Markov chain.
If the update function and underlying Markov operator $P$
satisfies the conditions of Corollary~\ref{coro: D_U_almost_D_P_spec},
then a similar discrepancy bound as in Corollary~\ref{cor_higher_order}
also holds for the driver sequence $\mathcal{U}_n=\{u_0, u_1, \ldots, u_{n-1}\}$.
Namely
\begin{equation*}
D^\ast_{\mathscr{B}, \psi,\varphi}(\mathcal{U}_n)
\le 63 \sqrt{d}\, \frac{(2 + \log_2 n)^{(3d+1)/2}}{n}
+ \frac{1 - \Lambda_0^n}{n \cdot (1- \Lambda_0)}
\left\|\frac{d \nu}{d \pi} - 1 \right\|_2.
\end{equation*}
\section{Concluding remarks}
Let us point out that
the discrepancy results of Subsection~\ref{subsec_main}
and Subsection~\ref{subsec_burn_in},
in particular, also hold for local Markov chains
which do not satisfy the `anywhere to anywhere'
condition
and the proof of this bound reveals that a uniformly i.i.d.
driver sequence satisfies
the discrepancy estimate with high probability.
In other words, there are many driver sequences which satisfy
the discrepancy bound of order $(\log n)^{1/2} n^{-1/2}$.
On the other hand, the choice of the driver sequence depends on the
initial distribution $\nu$ and the transition kernel.
It would be interesting to prove the existence of a universal
driver sequence, which yields Monte Carlo type behavior
for a class of initial distributions and transition kernels.
(For a finite set of initial distributions and
transition kernels such a result can be obtained
from our results since for any given initial distribution
and transition kernel we can show the existence of good driver
sequences with high probability, see Remark~\ref{rem4}.)
Further, the proven bounds on the discrepancy are based on a covering argument
with the Vapnik-\v{C}ervonenkis dimension. It is natural to ask
whether one can get better estimates with other covering arguments, for
example Dudley's entropy \cite{Du67} or its variants.
Another open problem is
the explicit construction of suitable driver sequences.
The results in this paper do not give any indication
how such a construction could be obtained.
However, as a step towards explicit constructions, we do obtain that the pull-back discrepancy
is the relevant criterion for constructing driver sequences.
|
\section{Introduction}
WASP-4b is one of the exoplanets most studied in the literature. Since
its discovery \citep{wilson08}, many observations of this target
have been made and several authors have determined the physical
properties of the host-star and the exoplanet \citep{gil09,winn09, nik12}. These works
reveal that the system is formed by a G7V star with a close-in hot
Jupiter ($M_{p}=1.28 M_{J}$, $R_{p}=1.39 R_{J}$) in a circular
orbit which transits the star every 1.33 days. WASP-4b is a highly
irradiated planet with a radius larger than the one predicted by
models \citep{for07}. One possibility is that the ongoing
orbital circularization provides the heat needed to inflate the
planet \citep{bee11}.
The transit timing
variations (TTVs) technique has become a very promising
method to estimate the mass of a non-transiting planet when it is not
possible to get radial velocity measurements \citep{hym05}.
Since the time between transits of a single planet should be
constant, variations in this time can be due to the gravitational
interaction with another planet in the system. If both planets show
transits, it is possible to estimate the radius and mass for each of
them, even without spectroscopic observations. In this way, it is
possible to determine the densities of planets orbiting late stars.
This is one of the key aspects of the TTVs technique.
Different authors carried out TTVs analysis looking for
another planetary-mass body in the WASP-4 system
without success. However, most of them employed mid-transit times
fitted with different models and error treatments. As it has
been shown \citep{sou12,nas13} the lack of
homogeneity in the analysis technique can lead to wrong conclusions
about TTVs.
In this work we present the light curves of 6 new transits
of WASP-4b obtained with telescopes located in Argentina, and
perform an homogeneous study of TTVs, analyzing 34 light curves spanning 6 years of
observations. For all these transits we employed the same
fitting procedure and error treatment to
obtain consistent photometric and physical parameters of the star and
the exoplanet.
In Section \S\ref{sec.obs} we present our observations and data reduction, in
Section \S\ref{sec.proc} we describe the procedure used to fit the light curves and the
parameters derived for the 34 transits. In Section \S\ref{sec.eph} we discuss the
new calculated ephemeris. In Section \S\ref{sec.comp} we compare the
results obtained with the fit provided by the Exoplanet Transit Database and, finally, in Section \S\ref{sec.conc} we present the conclusions.
\section{Observations and data reduction}\label{sec.obs}
We observed 6 transits of WASP-4b between October 2011 and
July 2013 employing two different telescopes: the Horacio
Ghielmetti Telescope (THG) located at the Complejo Astron\'omico El
Leoncito in San Juan (Argentina), and the 1.54 m telescope located at
the Estaci\'{o}n Astrof\'{i}sica
de Bosque Alegre (EABA, C\'{o}rdoba, Argentina). One of these transits
was observed with both telescopes simultaneously. In the analysis, we considered
these two measurements as independent. In Table \ref{tbl-1}
we show a log of the observations.
The THG is a remotely-operated 40-cm MEADE - RCX 400, with a focal ratio of
$f/8$. The instrument is currently equipped with an Apogee
Alta U16M camera with 4098 $\times$ 4098, 9 $\mu$m pixels, resulting in a
scale of 0.57"/pix and a 49'$\times$49' field of view.
At the EABA,
we used the 1.54 m telescope in the Newtonian focus, equipped with a 3070 $\times$ 2048, 9 $\mu$m pixels
Apogee Alta U9 camera. This camera provides a
scale of 0.25"/pix and a 8'$\times$12' field of view. For four transits, we employed the Johnson R filter available at both sites, while for the remaining two we made the observations without filter.
At the beginning of each observing night the computer clock was
automatically synchronized with the GPS. The central times
of the images were expressed in Heliocentric Julian Date based on
Coordinated Universal Time ($HJD_{UTC}$). Whenever possible, we observed 90 minutes
before and after each transit to obtain a large number of out-of-transit (OOT)
data-points to correct possible trends in the light
curves. We took 10 bias frames, 8 dark frames and
between 15 and 20 dome flat-fields. We averaged all the biases and median-combined
the bias-corrected darks. Finally, the bias- and
dark-corrected flats were median-combined to generate a master flat in
the corresponding band. All the images were processed using standard IRAF\footnote{IRAF is
distributed by the National Optical Astronomy Observatories, which
are operated by the Association of Universities for Research in
Astronomy, Inc., under cooperative agreement with the National
Science Foundation.} tasks.
To obtain instrumental magnitudes with aperture photometry, we
developed an algorithm called FOTOMCC. This is a quasi-automatic
pipeline developed for the IRAF environment using the “DAOPHOT”
package. Initially, FOTOMCC
employs a reference image, previously selected by the user, to
identify the centroids of the stars in all the images.
The optimal size for
the aperture is chosen through the growth-curves technique \citep{how89}. Specifically, we adopted the aperture size for which the star
magnitude was stable at the level of 0.001 mag.
The thickness of the sky-subtraction area was set to 5 pixels.
The magnitude errors were those provided by the
DAOPHOT task.
To carry out the
differential photometry, for every image we first subtracted from the
magnitude of the science star the one of each star
in the field. Then, we computed the standard deviation of all the
magnitude differences obtained in this way and we selected those stars
which gave the light curves with the lower sigma.
With the selected stars we built a master star
whose magnitude and error were the average magnitude and
error of all the chosen stars. The final light curve was built by
the subtraction of the magnitudes of the target and the
master star. For each photometric data, we estimated the formal error
as the quadrature sum of the errors of the target star and the master star.
Light curves present smooth trends mainly originated by differential
extinction and/or spectral type differences between the comparison and
the target star. To eliminate these slow variations we fitted
a Legendre polynomial to the OOT data-points and modified its order until the
dispersion of the residuals was minimum. In almost all cases we
used a second-order fit, although in some cases a lower dispersion was found
by fitting a straight line. Finally, for each light curve we removed
the fit from all the data (including transit points) and normalized
the OOT to unity. In Figure \ref{fig1} we present our six
light curves, and the best-fit to the data. Errorbars are also shown.
\subsection{Archival light curves}
To study TTVs and for the parameters determination we also
included all other transits publicly available.
We considered in particular 20 light curves found in the
literature: 1 from
\cite{wilson08}, 2 from \cite{winn09}, 1 from \cite{gil09}, 4 from \cite{soj11} and 12 observed by \cite{nik12}. We did not include the 4 transits from \cite{sou09}, since the authors reported failures in the
computer clock which make the mid-transit times unreliable
\citep{sou13}. We also included 8 transits observed by
amateurs and published in the Exoplanet Transit
Database (ETD\footnote{\textsf{http://var2.astro.cz/ETD.}}).
We only analyzed the complete transits with the
four contact points clearly visible.
\section{Light-curves fitting procedure}\label{sec.proc}
\subsection{Photometric parameters}\label{subsec.photpar}
Based on HARPS high signal-to-noise archival spectra of WASP-4, we derived stellar parameters: effective temperature $T_{eff}$, surface gravity $\log g$, metallicity $[Fe/H]$ and microturbulence $\xi$, using the FUNDPAR code (Saffe 2011).
The parameters obtained from
the analysis are: $T_{eff}=(5436 \pm 34)$ K, $\log g= (4.28 \pm 0.06) $ cm/s,
$\xi=(0.94 \pm 0.03)$ km/s, $[Fe/H]=(-0.05 \pm 0.04)$ dex (Jofr\'{e} et al. in
preparation). These agree with previously reported values,
except for $\log g$ which is slightly lower (e.g. \citealt{doy13}).
These stellar parameters were adopted as initial input for the program
JKTLD\footnote{\textsf{http://www.astro.keele.ac.uk/~jkt/codes/jktld.html.}},
which calculates theoretical limb-darkening coefficients
by bilinear interpolation of the effective temperature and surface
gravity using different tabulations. In particular, we employed the
tabulations provided by \cite{van93} and \cite{cla04}. For those
transits observed with no filter we used bolometric limb-darkening coefficients.
All the light curves were fitted using the JKTEBOP
code\footnote{\textsf{http://www.astro.keele.ac.uk/~jkt/codes/jktebop.html.}}. This
code models the light curve of a system of two components by
performing numerical integration over the
surface of concentric circles, under the assumption that the projection of each
component is a biaxial ellipsoid. It employs the Levenberg-Marquardt
optimization algorithm to get the best-fitting model. One of the
advantages of JKTEBOP over other fitting models is that it considers
small distortions from sphericity. Since WASP-4b is
a bloated planet, this program can give more realistic parameters from
the observed data.
For each transit, we ran JKTEBOP following the same fitting
procedure:
1) We assumed as free parameters: the inclination of the orbit ($i$), the sum of
the fractional radii\footnote{$r_{\star}=R_{\star}/a$ and $r_{P}=R_{P}/a$
are the ratios of the absolute radii (of the star and the exoplanet respectively) to the semimajor axis.}
($r_{\star} + r_{P}$), the ratio of the fractional
radii ($k=r_{\star}/r_{P}$) and the mid transit time
($T_{0}$). We fitted every light curve with the linear,
quadratic, logarithmic and square-root limb-darkening laws. For each
case, we tried with a) both coefficients fixed, b) the linear coefficient fitted
and the nonlinear fixed and c) both coefficients fitted. Finally, we
adopted as the best model for a given transit the one which minimizes the
$\chi^2$ of the fit and gives realistic parameters.
tbf{ 2) For a few transits, the convergence of some of the adjusted parameters was not achieved in 1). In these cases, assuming the limb-darkening law obtained in the first step, we iterated JKTEBOP taking as initial parameters of each iteration those obtained in
the previous one. This process was repeated until convergence.
3) For the solution achieved in 2), we first multiplied the photometric errors
by the square-root of the reduced chi-squared of the fit to
get $\chi_{r}^2=1$. Then, we ran the three algorithms available in JKTEBOP:
Bootstrapping and Monte Carlo simulations and
Residuals Permutation (RP), which takes red noise into account. For the
first two options we performed 1000 iterations. We conservatively adopted as
the final errors of the parameters the largest values given by these
algorithms.
We adopted as the final value for every parameter the median of those
obtained for every transit (except for $T_0$, see \S\ref{sec.eph}). We
adopted as the final error
the asymmetric uncertainties $\sigma_{+}$ and $\sigma_{-}$
of the selected
distribution, since they
are based on the empirical data and are more realistic than those
derived by a Gaussian distribution of the parameters.
\subsection{Physical parameters}\label{subsec.bozomath}
The physical parameters were determined using standard formulae
\citep{sousolo09} implemented in the JKTABSDIM
code\footnote{\textsf{http://www.astro.keele.ac.uk/~jkt/codes/jktabsdim.html.}}. This
code requires as input the measured quantities: $i$, $r_{\star}$,
$r_{P}$, the orbital period $P$, the velocity amplitudes of the star
and the exoplanet, $K_{\star}$ and $K_{P}$ respectively, the eccentricity $e$, $T_{eff}$, $[Fe/H]$ and their
errors. For each light curve, we employed the photometric parameters
($i$, $r_{\star}$, $r_{P}$)\footnote{The error considered as input was
the larger between $\sigma_{+}$ and $\sigma_{-}$.} obtained with the
program JKTEBOP, $P$ determined from the
ephemeris, and $T_{eff}$ and $[Fe/H]$ derived using
HARPS spectra. We used $e=0$ , and the $K_{\star}$ value given by
\cite{tri10}.
The procedure was the following: First, assuming $K_{P}=150\,km/s$ we calculated a stellar mass (see Eq. (5) of \cite{sousolo09}). By
linearly interpolating this stellar mass and the $[Fe/H]$ calculated in \S\ref{subsec.photpar} within tabulated theoretical model, we determined a predicted radius ($R^{(calc)}_{\star}$) and effective
temperature ($T^{(calc)}_{eff}$) for the star. Then, we evaluated the figure of merit:
\begin{equation}
fom = \Bigg[\frac{r^{(obs)}_{\star}-(R^{(calc)}_{\star}/a)}{\sigma(r^{(obs)}_{\star})}\Bigg]^2+\Bigg[\frac{T^{(obs)}_{eff}-T^{(calc)}_{eff})}{\sigma(T^{(obs)}_{eff})}\Bigg]^2
\end{equation}
We repeated this process until finding the value for $K_{P}$ which minimizes Eq. (1).
In order to avoid any dependence with the stellar-model, we performed this analysis
for 4 different sets of stellar models: $Y^2$ \citep{dem04},
Padova \citep{gir00}, Teramo \citep{pie04} and VRSS \citep{van06}.
We adopted as the final value for $K_{P}$ the average of
the amplitudes given by each model, and the standard deviation as the error of the velocity.
Finally, the solution for the system was determined using the JKTABSDIM
code.
From this procedure, we also estimated the age of the system considering
series of models bracketing the lifetime of the star in the main
sequence.
The resulting physical parameters of the star and
the planet obtained for each transit are listed in Table
\ref{tbl-2}. For the exoplanet, the surface gravity was calculated
with:
\begin{equation}
g_{P} = \frac{2\pi}{P}\frac{\sqrt{(1-e^2)} K_{\star}}{r^2_{P} sin(i)}
\end{equation}
\citep{sou07} and the modified equilibrium temperature as:
\begin{equation}
T'_{eq} = T_{eff}\sqrt{\frac{R_{\star}}{2a}}
\end{equation}
\noindent(Southworth 2010). Therefore, both $g_{P}$ and $T^{\prime}_{eq}$ are independent of the stellar
models. We performed a weighted average of all the measurements to
obtain the final value for each parameter, and the uncertainty
was determined as the standard deviation of the sample. Table \ref{tbl-3} shows
the final values and errors calculated for the photometric and
physical parameters of the star and the exoplanet. All these
are in good agreement with previous determinations, except for a
slightly lower inclination.
The presence of a perturber in the system could produce long-term
variations in these parameters (\citealt{sar99}, \citealt{car10}). Considering that our data comprises 6
years of observations, we studied the long-term behaviour of
$i$ and $R_{P}/R_{\star}$ (Fig \ref{fig2}a and \ref{fig2}b). We found
that these parameters remain constant within the $\pm 1\sigma$ error
of the weighted average, except for the outlier data-point in $i$
corresponding to the epoch 1307, which could have been caused by variable
observing conditions such as the presence of cirrus clouds during that night.
\section{Transit ephemeris and timing}\label{sec.eph}
We transformed the central times of all the observations to $BJD_{TDB}$ (Barycentric
Julian Date based on Barycentric Dynamical Time) with the \cite{eas10} online converter. For the amateur light curves,
we contacted the observers when extra information was needed.
For the mid-transit times we adopted the mean values obtained
in Section \S\ref{sec.proc}, and considered
the symmetric errors ($\pm\sigma$) given by the algorithm with the
largest uncertainty.
In most cases, the error
obtained with the RP method was the largest, indicating the presence of red
noise in the data \citep{pont06}. This implies that there are
correlations between adjacent data points in a light curve, reducing
the number of free parameters. The existence of red noise leads to an
underestimation of the errors in the adjusted parameters which, in
turn, might cause an inaccurate determination of the central time of
the transit. The red noise can be quantified with the factor
$\beta=\sigma_{r}/\sigma_{N}$, defined by \cite{winn08}. Here,
$\sigma_{r}$ is obtained by averaging the residuals into M bins of N
points and calculating the standard deviation of the binned residuals,
and $\sigma_{N}$ is the expected deviation, calculated by:
\begin{equation}
\sigma_{N} = \frac{\sigma_{1}}{\sqrt{N}}\sqrt{\frac{M}{M-1}}
\end{equation}
where $\sigma_{1}$ is the standard deviation of the unbinned
residuals. Considering that the duration of the ingress/egress of the
WASP-4b transits is about 20 minutes, we averaged the residuals in bins
of between 10 and 30 minutes and calculated the parameter $\beta$ for
each case. Finally, we used the median value as
the red noise factor corresponding to that light curve. In the
absence of red noise, we expect $\beta =1$. For these transits
$\beta$ ranges from 0.58 to 2.36.
The whole sample of
mid-transit times presents 2 big outliers corresponding to the epochs
298 and 1085. The first transit was obtained from the ETD.
In the latter case, we believe there was a failure in
the computer clock. We did not considered these points for further
analysis. Therefore, we determined the ephemeris in three different ways: a) considering all
the 32 remaining transits, b) excluding the incomplete transit (indicated in Fig \ref{fig1} as 2013-06-06 and observed at EABA), and c) only considering
those transits with $\beta \le 1.6$ (30 points). In the three cases
we fitted the data through weighted least-squares to obtain the best
period and the minimum reference time. We re-scaled the uncertainties
multiplying them by $\sqrt{\chi^2_{r}}$. The final values and errors for $P$ and $T_{0}$ obtained from
different sets are:
\begin{eqnarray*}
a)\,P=1.33823251(31)\,\textrm{days},\,T_{0}=2454697.797973(76)\,BJD_{TDB}\\
b)\,P=1.33823251(32)\,\textrm{days},\,T_{0}=2454697.797973(77)\,BJD_{TDB}\\
c)\,P=1.33823227(32)\,\textrm{days},\,T_{0}=2454697.797973(77)\,BJD_{TDB} \\
\end{eqnarray*}
\noindent {Since there are no differences in $T_{0}$, the inclusion of partial
transits, or those obtained with large red noise, does not affect the
calculation. We adopted the ephemeris given by the sample a)
including all the transits.
With the new ephemeris, we calculated the $O-C$ mid-transit times, which
are shown in Figure \ref{fig2}c. Except for the already mentioned outliers, all
differences are within the $\pm 1\sigma$
error. The RMS of the data is 54 seconds. We ran a Lomb-Scargle
periodogram \citep{hor86} to the data, excluding the 2 big
outliers, and no significant peak was found.
\section{Comparison between JKTEBOP and the fitting model in the ETD}\label{sec.comp}
For the light curves taken from the ETD, we
compared the mid-transit times obtained with JKTEBOP and those given by the ETD,
which provides an
automatic fit, modeling the photometric data with the
function (Poddany et al. 2010):
\begin{equation}
m(t_{i}) = A-2.5\log F(z[t_{i},t_{0},D,b], p, c_{1}) + B(t_{i}-t_{mean})+ C(t_{i}-t_{mean})^{2}
\label{eq.mod}
\end{equation}
where $m(t_{i})$ are the relative magnitudes taken at the times
$t_{i}$, $N$ is the number of data-points,
$t_{mean} =t_{i}/N$ is the mean time of the observations,
$z$ is the projected relative-separation of the planet from
the star, $p$ is the ratio of the planet to star radii and $F(z,p,
c_{1})$ is the $occultsmall$ routine of \cite{man02}, giving
the relative flux of the star as the planet transits.
This model assumes a linear limb-darkening law with the
coefficient $c_{1}$ fixed at an arbitrary value of $0.5$.
The user has the possibility to fit or maintain fixed
the mid-transit time, the duration and the depth parameters.
The coefficients of Eq. (\ref{eq.mod}) are calculated using the
Levenberg-Marquardt non-linear least-squares algorithm
from \cite{press92}. The optimal parameters are determined by
iterating the procedure until the difference between two
successive values of $\Delta\chi^2$ is negligible.
We fitted the three parameters simultaneously and converted the
resulting $HJD_{UTC}$ mid-transit times to $BJD_{TDB}$. In Figure \ref{fig2}d we
show the differences between the central times determined in both
ways. The errorbars are those derived using the ETD model.
The differences are as large as 1.5
minutes. We believe these disagreements are due to the very simple
limb-darkening law assumed in the ETD fit. In any case, these differences point out to
the need to derive the mid-transit times with an homogeneous method,
when searching for TTVs.
\section{Summary and conclusions}\label{sec.conc}
In this work we present 6 new observations of transits of WASP-4b, observed between
2011 and 2013. Using these observations together with another 28
transits previously reported (including 8 observed by amateurs), we
performed an homogeneous study of the system taking into
account the realistic possibility of distortions in its components.
The physical parameters of the star and the exoplanet are consistent
with previous determinations, except for the inclination which is slightly
lower, probably due to the fitting procedure.
In addition, we analyzed the long-term behaviour of different parameters.
Except for one outlier in $i$, and two
for the $O-C$
mid-transit times, all these parameters remain stable within the
$\pm 1\sigma$ error of the weighted averages. The RMS of the
mid-transit times is 54 seconds. Therefore, we
confirm previous
results, and found that the system does not show significant TTVs
attributable to the presence of a perturber, a conclusion
we expanded with two more years of observations, to a baseline of 6
years. The lack of temporal variations in the
rest of the parameters supports this conclusion.
Finally, we report differences as large as 1.5 minutes between the
mid-transit times modeled by the fitting programs provided by the
ETD and JKTEBOP. Therefore, we believe that the central times provided
by the ETD should be used with caution in TTV studies.
\acknowledgments
We are grateful to Pablo Perna and the CASLEO staff for technical support, and to CONICET for funding this research. We thank the anonymous referee for their useful comments, and
Phil Evans, Ivan Curtis, and T. G. Tan for kindly providing us
with information about their observations.
\bibliographystyle{apj}
|
\section{Introduction}
In this note we consider graphs which are finite, undirected, and
have no loops or multiple edges and multigraphs which may contain
multiple edges but no loops. Let $V(G)$ and $E(G)$ denote the sets
of vertices and edges of a multigraph $G$, respectively. The degree
of a vertex $v\in V(G)$ is denoted by $d_{G}(v)$, the maximum degree
of $G$ by $\Delta (G)$, and the chromatic index of $G$ by
$\chi^{\prime }\left(G\right)$. A multigraph $G$ is Eulerian if it
has a closed trail containing every edge of $G$. The terms and
concepts that we do not define can be found in \cite{West}.
A proper edge-coloring of a multigraph $G$ is a coloring of the
edges of $G$ such that no two adjacent edges receive the same color.
If $\alpha $ is a proper edge-coloring of $G$ and $v\in V(G)$, then
$S\left(v,\alpha \right)$ denotes the set of colors of edges
incident to $v$. A proper edge-coloring of a multigraph $G$ with
colors $1,\ldots,t$ is called an interval $t$-coloring if all colors
are used, and for any vertex $v$ of $G$, the set $S\left(v,\alpha
\right)$ is an interval of integers. A multigraph $G$ is interval
colorable if it has an interval $t$-coloring for some positive
integer $t$. The set of all interval colorable multigraphs is
denoted by $\mathfrak{N}$.
The concept of interval edge-coloring of multigraphs was introduced
by Asratian and Kamalian \cite{AsrKam1}. In \cite{AsrKam1,AsrKam2},
they proved the following result.\\
\noindent\textbf{Theorem 1.} If $G$ is a multigraph and $G\in
\mathfrak{N}$, then $\chi^{\prime }\left(G\right)=\Delta (G)$.
Moreover, if $G$ is a regular multigraph, then $G\in \mathfrak{N}$
if and only if $\chi^{\prime }\left(G\right)=\Delta (G)$.\\
Some results on interval edge-colorings of multigraphs were obtained
in \cite{Kam}. In \cite{PetKhach}, the authors described some
methods for constructing of interval non-edge-colorable bipartite
graphs and multigraphs.
In this note we show that all Eulerian multigraphs with an odd
number of edges have no interval coloring. We also give some methods
for constructing of interval non-edge-colorable Eulerian multigraphs.\\
\section{Results}
Let $G$ be a multigraph. For any $e\in E(G)$, by $G_{e}$ we denote
the multigraph obtained from $G$ by subdividing the edge $e$. For a
multigraph $G$, we define a multigraph $G^{\star}$ as follows:
\begin{center}
$V(G^{\star})= V(G)\cup \{u\}$, $u\notin V(G)$,\\
$V(G^{\star})= E(G)\cup \{uv:v\in V(G)~and~d_{G}(v)~is~odd\}$.
\end{center}
For a graph $G$, by $L(G)$ we denote the line graph of the graph
$G$.\\
We also need a classical result on Eulerian multigraphs.\\
\noindent\textbf{Euler's Theorem.} (\cite{Euler}) A connected
multigraph $G$ is Eulerian if and only if every vertex of $G$ has an even degree.\\
Now we can prove our result.\\
\noindent\textbf{Theorem 2.} If $G$ is an Eulerian multigraph and
$\vert E(G)\vert$ is odd, then $G\notin
\mathfrak{N}$.\\
\noindent\textbf{Proof} Suppose, to the contrary, that $G$ has an
interval $t$-coloring $\alpha$ for some $t$. Since $G$ is an
Eulerian multigraph, $G$ is connected and $d_{G}(v)$ is even for any
$v\in V(G)$, by Euler's Theorem. Since $\alpha$ is an interval
coloring and all degrees of vertices of $G$ are even, we have that
for any $v\in V(G)$, the set $S\left(v,\alpha\right)$ contains
exactly $\frac{d_{G}(v)}{2}$ even colors and $\frac{d_{G}(v)}{2}$
odd colors. Now let $m_{odd}$ be the number of odd colors in the
coloring $\alpha$. By Handshaking lemma, we obtain
$m_{odd}=\frac{1}{2}\sum\limits_{v\in
V(G)}\frac{d_{G}(v)}{2}=\frac{\vert E(G)\vert}{2}$. Thus $\vert
E(G)\vert$ is even, which is a contradiction. $\square$\\
\noindent\textbf{Corollary 1.} If $G$ is an Eulerian multigraph and
$G\in \mathfrak{N}$, then $\vert E(G)\vert$ is even.\\
Let us note that there are Eulerian graphs with an even number of
edges that have no interval coloring. For example, the complete
graph $K_{5}$ has no interval coloring. On the other hand, there are
many Eulerian graphs with an even number of edges that have an
interval coloring. In
\cite{Jaeger}, Jaeger proved the following result.\\
\noindent\textbf{Theorem 3.} If $G$ is a connected $r$-regular graph
($r\geq 2$), $\chi^{\prime }\left(G\right)=r$ and $\vert E(G)\vert$
is even, then
$\chi^{\prime }\left(L(G)\right)=2r-2$.\\
Since $G$ is a connected $r$-regular graph ($r\geq 2$) and $\vert
E(G)\vert$ is even, we have that $L(G)$ is a connected
$(2r-2)$-regular graph with an even number of edges. Moreover, by
Theorems 1 and 3 and Euler's Theorem,
we obtain the following\\
\noindent\textbf{Corollary 2.} If $G$ is a connected $r$-regular
($r\geq 2$) graph with an even number of edges and $G\in
\mathfrak{N}$, then
$L(G)$ is an Eulerian graph with an even number of edges and $L(G)\in \mathfrak{N}$.\\
Let us note that Theorem 2 also gives some methods for constructing
of interval non-edge-colorable Eulerian multigraphs from interval
colorable multigraphs.\\
\noindent\textbf{Corollary 3.} If $G$ is an Eulerian multigraph and
$G\in \mathfrak{N}$, then for each $e\in E(G)$, $G_{e}\notin
\mathfrak{N}$.\\
\noindent\textbf{Corollary 4.} If $G$ is a connected multigraph with
an odd number of edges and $G\in \mathfrak{N}$, then
$G^{\star}\notin \mathfrak{N}$.\\
|
\section{Introduction}
\label{sec:introduction}
The controlled production of emulsions composed of micron or submicron-sized drops with a
well defined mean size possesses uncountable applications in industry, medicine and
pharmacology~\citep{Basaranaiche,AnnuRevStone,BarreroAnnurev}. One of the most successful
methodologies for the production of micro or nano drops consists in generating a thin
thread of the fluid to be dispersed within a coflowing stream of the carrier one. Under
these conditions, the narrow jet subsequently breaks downstream of the injection tube due
to the growth of capillary disturbances. To avoid the clogging of these type of
flow-focusing~\citep*{Anna} or coflow devices~\citep*{SuryoBasaran,UtadaPRL,Marin}, the
diameter of the feeding tube is usually much larger than that of the produced jet which,
therefore, suffers a strong stretching in the downstream direction.
Recently,~\citet*{JFM12} reported experiments on the generation of concentrated monodisperse
emulsions composed of drops with \emph{uniform} sizes of even below 1 $\mu$m using a low
Reynolds number co-flow configuration inspired by the numerical experiments due
to~\cite{SuryoBasaran}. In these experiments a flow rate $Q_i$ of a fluid with a viscosity
$\mu_i$ discharges through a cylindrical tube of inner radius $R_i$ into an immiscible liquid
of viscosity $\mu_o=\mu_i/\lambda$ flowing in parallel with the axis of the injector at a
velocity $U_o$ with, from now on, $\sigma$ indicating the interfacial tension coefficient
between both immiscible liquids and $\lambda$ the inner-to-outer viscosity ratio. The
experimental observations of figures~\ref{fig1} and~\ref{fig2} reveal that the inner jet
highly stretches downstream provided that the flow-rate ratio $q=Q_i/(\pi\,R^2_iU_o)$ is
sufficiently small and the capillary number $\Ca=\mu_o\,U_o/\sigma$ is of order unity or
larger. The approximately cylindrical ligament emitted from the tip of the conical drop
breaks under the action of capillary forces, giving rise to a train of droplets with a low
size dispersion \emph{if} $\Ca$ is sufficiently large. Under these conditions, it was found
in~\citet{JFM12} that the drop diameter scales as $D_d \propto R_i q^{1/2}$. Therefore,
uniformly-sized drops with arbitrarily small diameters could be, in principle, obtained if
the control parameter $q$ is set to sufficiently small values and the capillary number
$\Ca$ is above a so far unknown threshold capillary number $\Ca^*$, whose determination
as a function of $q$ and $\lambda$, with $\lambda\ll 1$, is the main purpose of the present work.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{figure1}
\caption{For each of the values of the viscosity ratio considered in this study, namely,
(from top to bottom) $\lambda=10^{-2}$, $\lambda=5\times 10^{-3}$, $\lambda=3\times 10^{-3}$,
$\lambda=2\times 10^{-3}$ and $\lambda=10^{-3}$, there exists a critical value of the outer
velocity $U_o$ and, consequently, of the capillary number $\Ca=\mu_o\,U_o/\sigma$, above which
a steady jet issues from the apex of a conical drop and a train of uniform-sized drops is
produced (left column, figures 1a--e). However, if the capillary number is below its critical
value, the jet issued at the apex of the cone is unsteady and breaks unevenly, forming drops
of different sizes (right column, figures 1f--j). The values of $\Ca$ and $\lambda$ in each of
the experimental images depicted in this figure are the following:
(a) $\lambda=10^{-2}$, $\Ca=1.12$, (b) $\lambda=5\times10^{-3}$, $\Ca=1.15$,
(c) $\lambda=3\times 10^{-3}$, $\Ca=1.7$, (d) $\lambda=2\times 10^{-3}$, $\Ca=2.09$,
(e) $\lambda=10^{-3}$, $\Ca=2.69$, (f) $\lambda=10^{-2}$, $\Ca=0.9$,
(g) $\lambda=5\times10^{-3}$, $\Ca=0.98$, (h) $\lambda=3\times 10^{-3}$, $\Ca=1.37$,
(i) $\lambda=2\times 10^{-3}$, $\Ca=1.69$, (j) $\lambda=10^{-3}$, $\Ca=2.06$.
The outer diameter of the injection tube is $200$ $\mu$m.\label{fig1}}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\textwidth]{figure2}
\caption{The figure shows the effect of varying the flow rate ratio $q$, which increases
from left to right, on the drop formation processes for a fixed value of the viscosity
ratio, $\lambda=10^{-2}$, and three values of the capillary number, namely, $\Ca=0.9$
(top row), $\Ca=0.98$ (middle row) and $\Ca=1.12$ (bottom row). The analysis of the images
reveals that the value of the critical capillary number, which in this case is between
$\Ca=0.90$ and $\Ca=1.12$ is rather insensitive to changes in $q$. The same conclusion
can be extracted from the analysis of the experimental images corresponding to the rest
of viscosity ratios investigated. The values of $Q_i$ in each of the experimental images
are the following: a) $Q_i=2.1\times10^{-4}$ mm$^3\,$s$^{-1}$, b) $Q_i=3.8\times10^{-3}$
mm$^3\,$s$^{-1}$, c) $Q_i=8.6\times10^{-3}$ mm$^3\,$s$^{-1}$, d) $Q_i=7.3\times10^{-4}$
mm$^3\,$s$^{-1}$, e) $Q_i=4.2\times10^{-3}$ mm$^3\,$s$^{-1}$, f) $Q_i=7.4\times10^{-3}$
mm$^3\,$s$^{-1}$, g) $Q_i=1.3\times10^{-3}$ mm$^3\,$s$^{-1}$, h) $Q_i=2.9\times10^{-3}$
mm$^3\,$s$^{-1}$, i) $Q_i=1.4\times10^{-2}$ mm$^3\,$s$^{-1}$.
The outer diameter of the injection tube is $200$ $\mu$m \label{fig2}}
\end{figure}
To illustrate the effect of the capillary number on the distribution of drop sizes,
figure~\ref{fig1} as well as the three movies provided as supplementary material, show how water drops
are generated and dispersed within co-flowing
streams of different silicon oils with viscosities ranging from $\mu_o=100$ cp to
$\mu_o=1000$ cp at $25^o$$C$. Figures~\ref{fig1}a-e reveal that, for each value of the
viscosity ratio investigated, there exists a threshold outer velocity $U_o$ above which
the thin liquid thread ejected from the tip of a conical drop breaks regularly, forming
tiny droplets with very similar diameters. In the experimental situations depicted in
figures~\ref{fig1}f-j, the overall geometry of the drop attached at the injection tube
is similar to that showed in figures~\ref{fig1}a-e, but the jet issued at the apex of
the cone is unsteady and breaks unevenly, forming droplets with a broad size distribution.
It is interesting to note that the cone itself is steady even in cases in which the drop
formation process is not periodic.
The reason for the differences observed in the two series of experiments depicted in
figure~\ref{fig1} is that, in figures~\ref{fig1}a-e $\Ca>\Ca^*$, whereas $\Ca<\Ca^*$ in
figures~\ref{fig1}f-j, where $\Ca^*$ is defined as the minimum value of the capillary number
for which a train of uniform-sized drops is produced for fixed values of $\lambda$ and $q$.
It is also interesting to note that, in contrast with the noticeable dependence of $\Ca^*$
on $\lambda$ deduced from the analysis of the type of experimental images depicted in
figure~\ref{fig1}, the results in figure~\ref{fig2} reveal that the critical capillary number
is rather insensitive to changes in $q$.
Since many different applications such as drug delivery demand that the size distribution of
the drops composing the emulsion is as narrow as possible, it is desirable that the capillary
number at which the coflow device is operated is larger than $\Ca^*$. It is thus our purpose in
this paper to develop a theory to quantify and explain why drops with uniform sizes are produced
only for values of the capillary number above the critical value $\Ca^*$ deduced from the analysis of the experimental images in figures~\ref{fig1} and~\ref{fig2}. Notice that, according to the experimental
observations in figures~\ref{fig1} and~\ref{fig2}, figure~\ref{fig3} shows that
$\Ca^*$ is a decreasing function of $\lambda$ which does not appreciably vary with $q$.
As a first attempt to predict and quantify the dependence of the critical capillary number
on $\lambda$ depicted in figure~\ref{fig3}, let us recall that the non-regularity in the drop
formation process is associated with the fact that the thin liquid thread from which drops are
emitted, is itself unsteady. This experimental evidence suggests that the unstable capillary
perturbations that give rise to the generation of drops cannot be convected downstream at a
sufficiently large speed when $\Ca<\Ca^*$. Therefore, the two regimes delimited by the
experimental data points in figure~\ref{fig3} should correspond to those conditions for which
unstable perturbations are either convected downstream the cone tip (convectively unstable
flow), or are able to propagate upstream, thus inhibiting the formation of a steady jet
(absolutely unstable flow).
\begin{figure}
\centering
\includegraphics[width=1\textwidth]{figure3}
\caption{The critical capillary number, $\Ca^*$, as a function of the viscosity ratio $\lambda$.
Symbols show the experimental transition points separating the monodisperse ($\Ca>\Ca^*$,
figures 1a--e), and polydisperse ($\Ca<\Ca^*$, figures 1f--j) droplet generation regimes.
The line shows the critical capillary number for the transition from convective to absolute
instability computed using Tomotika's dispersion relation.
\label{fig3}}
\end{figure}
Previous studies~\citep{PoF05,UtadaPRL,UtadaPRL08,GuillotPRL} have demonstrated that the experimentally observed transition from jetting to dripping in cylindrical capillary jets can be predicted by means of a \emph{local} linear stability
analysis of parallel streams~\citep{AnnRevHuerre}. Thus, with the purpose of calculating the
boundary separating the two different drop formation processes depicted in figures~\ref{fig1}, \ref{fig2} and in the movies provided as supplementary material, we have determined the values of the capillary number for which the temporal growth rate of the perturbations with zero
group velocity is also equal to zero. The result obtained using Tomotika's dispersion
relation~\citep{Tomotika,PRLGoldstein,PoFPowers,PREInnombrable}, which describes the growth and
propagation of perturbations in a cylindrical jet immersed into another immiscible liquid,
is represented together with the experimental data in figure~\ref{fig3}. Although the
conditions of validity of Tomotika's analysis are fulfilled at large distances from the
injector, where the diameter of the liquid jet is nearly constant, figure~\ref{fig3} reveals
that the parallel flow stability analysis predicts larger values for the critical capillary
number than those measured experimentally. Consequently, the full jet, which strongly stretches
in the downstream direction, is more stable than the cylindrical portion located far downstream.
To improve the agreement with experimental results, in this study we have performed a
\emph{global stability analysis} that takes into account the real shape of the stretched jet
and describes the growth and propagation of perturbations without resorting to the parallel
flow simplification used in previous studies~\citep{Tomotika,PREInnombrable}. To avoid the
lengthy numerical computations associated with a global stability analysis in two or three
spatial directions~\citep[see e.g.][]{Estela1,Christodoulou,Estela2,Theofilis2011}, in the present work
we have developed a one-dimensional model based on slender-body theory that yields a single
partial differential equation governing the spatiotemporal evolution of the jet radius. In \cite{JFM12} it was shown that
the axial variations of both the steady shape of the jet and the outer flow field are very well
reproduced by our previous theory. Here we will demonstrate that our global stability analysis, that resorts to a new equation for the jet radius, faithfully reproduces Tomotika's dispersion relation in the long wave limit for the case of cylindrical jets of small viscosity ratio. Moreover, the experimentally measured values of the critical
capillary number are very well reproduced since our final equation retains the axial dependence of the flow field in the formulation. Let us point out that the global stability analysis, developed in \S 2.2, shares many
similarities with that by~\citet*{Rubio2013}, where the global stability of gravitationally
stretched jets is studied using a different one-dimensional model and excellent agreement with experiments
is found.
\section{Unsteady slender-body theory for the description of highly stretched jets in Stokes flow}
\label{sec:theory}
One of the possible ways to analyze the conditions under which drops are formed regularly from the tip of the conical drop, would be to solve the unsteady velocity field using the integral formulation due to \cite{Laz}, as was done in ~\citet*{JFM12}. However, following a strategy similar to that successfully applied to the case of liquid jets in air by \cite{GyC,EggersDupont,basaran04b,Rubio2013}, the dynamics of the low Reynolds number jets depicted in figures 1-2 can be described using a much simpler one-dimensional approach which reveals the physics underlying the transition observed in these figures in a neat way. The central idea behind our one-dimensional theory can be understood by noticing that the slender jets shown in figures 1-2 resemble the case of drops immersed in a purely straining flow \cite{Stone94}. The deformation of slender drops immersed in this type of flow field has been rigorously analyzed using asymptotic methods by \cite{Buckmaster,AcrivosyLo} and, from these studies, it is learned
that the outer flow field in the case of low viscous drops can be decomposed, in a first approximation, as the addition of the unperturbed flow field plus a distribution of sources located at the axis of symmetry. Since, in our case, the capillary that confines the outer fluid is ten times larger than the inner one, the dynamics of the jet is not affected by this confinement, as already demonstrated in \cite{JFM12}. Therefore, the conclusions derived from \cite{Buckmaster,AcrivosyLo} are fully applicable to the situation at hand, and figure 4 sketches the key idea under which our theory is built. This figure shows that the outer velocity field can be decomposed, in a first approximation, as the addition of two simpler flow fields: the unperturbed velocity field, which is the one that would exist in the coflowing device \emph{if the inner fluid was not injected} (see
figure \ref{exterior}a) and that induced on the outer stream by the presence of the jet (see \ref{exterior}b), which can be viewed as a perturbation to the former. From now on, $T$, $R$ and $Z$ will indicate time and the radial and axial coordinates respectively, with $Z=0$ located at the exit of the injection tube, whereas $R_j(Z,T)$ will denote the radius of the jet. In our approach, the unperturbed velocity field is calculated numerically, by solving the Navier-Stokes equations in the zero Reynolds number limit (Stokes equations) subjected to the appropriate boundary conditions, as it will be explained below. Regarding the perturbed flow, we will take advantage of the fact that the outer velocity field induced by a slender jet of a low viscosity fluid immersed in an outer axisymmetric flow, as is the case of the experiments shown in figures 1 and 2, can be approximated to that created by a line of sources located at the axis of symmetry \citep{Ashley,Taylor1964,Buckmaster}. The intensity of the source
distribution, $S(Z,T)$, must be such that the kinematic boundary condition is verified at the jet interface (see figure \ref{exterior}b).
The representation of the outer flow field as the addition of two simpler velocity fields greatly facilitates the theoretical analysis of the spatiotemporal evolution of the jet. Firstly, notice that the numerical computation of the unperturbed flow field sketched in figure \ref{exterior}a is straightforward since this solution does not depend on $Ca$, $\lambda$ or $q$: it is just a function of the geometry of the coflowing device and thus, it is unique for a given injector. But, what's more, notice that the unperturbed velocity field does not even need to be known everywhere in the flow domain; instead, for our purposes, an accurate representation of the flow field in the vicinity of the axis of symmetry is enough. This is due to the fact that our interest here is to deduce an equation for $R_j$, with $R_j/R_i\ll 1$ for $Z/R_i\gtrsim O(1)$. Since the jet radius is located very near the axis of symmetry in most of the flow domain, all the information needed from the numerics to build our theoretical approach
are just two functions: the axial component of the velocity at the axis, $U_x=U_n(Z,R=0)$, as well as $F(Z)=1/2\, \partial^2 U_n/\partial R^2 (Z,R=0)$, where the subscript $n$ will be used in the following to denote quantities related to the unperturbed problem. Indeed, given these two functions, the analytical expression of the axial component of the unperturbed velocity field in the near-axis region will be expressed in Taylor series as $U_n(Z,R)\simeq U_x(Z)+F(Z)\,R^2+O[(R_j/R_i)^4]$. Moreover, the corresponding approximate expressions of both the radial component of the unperturbed velocity field namely, $V_n(Z,R)$ and of the unperturbed pressure field, $P_n(Z,R)$, will be written in terms of $U_x$ and $F$ making use of the equations of continuity and momentum, as it will be shown below. Let us point out that the main difference between this work and those by \cite{Taylor1964,AcrivosyLo,HinchAcrivos,Sherwood,WendyPRL}, where the deformation of drops and bubbles immersed in a purely straining flow in the
limit of negligible inertial effects is studied, is that our theory retains the effect of the real outer flow field on the spatiotemporal evolution of the jet. More precisely, the effect of the injector geometry on the outer flow field will be represented in the equation to be deduced for $R_j(Z,T)$ through the two numerically determined functions $U_x(Z)$ and $F(Z)$ which, for the case of the specific geometry used in the experiments explained above, are represented graphically in figure \ref{Uaxis} (see also \cite{JFM12}). Let us point out that the theory to be developed in \S 2.1 can be extended to any other types of axisymmetric geometries provided that the functions $U_x$ and $F$ are straightforwardly determined from the numerical solution of the Stokes equations subjected to the no slip boundary conditions imposed by the geometry of the injector and to the corresponding inflow/outflow boundary conditions.
\begin{figure}
\begin{center}
\includegraphics[width=1\columnwidth]{figure4}
\end{center}
\caption{Sketch showing that the outer velocity field is decomposed
as the sum of two simpler velocity fields: (a) the one that does not take
into account the presence of the inner fluid and (b) the flow field
induced by the jet, which can be approximated as a line of
sources of intensity $S(Z,T)$ located at the axis of symmetry.}\label{exterior}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.45\columnwidth]{figure5a}
\includegraphics[width=0.45\columnwidth]{figure5b}
\end{center}
\caption{Downstream variation of the velocity at the axis of
symmetry, $u_x$ (a) and $f=1/2\,\partial^2\, u/\partial r^2 (r=0)$ (b)
corresponding to the situation sketched in figure \ref{exterior}a,
i.e, when the inner jet is not present. The dimensionless counterparts of $U_x$ and $F$, which are represented using lower case variables, have been non-dimensionalized using $U_o$ and $R_i$ as characteristic scales of velocity and length.}\label{Uaxis}
\end{figure}
\subsection{Theory}
The analysis presented in this section generalizes the work by~\citet{JFM12}, and aims to describe
the propagation and growth of capillary disturbances in unsteady coflowing streams with negligible
inertial effects. In our theory, the flow is represented as the superposition of the velocity field created by a distribution of sources with an unknown intensity $S(Z,T)$ plus the unperturbed flow field, which is the solution of the Stokes equations subjected to the the impermeability condition ($\mathbf{U_n}=0$) that needs to be satisfied at: i) the tube exit, namely, the circle defined by $Z=0$, $R\leq R_i$, ii) the outer surface of the injection tube, which is the cylinder of radius $R=R_i$ that extends along $Z\leq 0$ and iii) the inner surface of the outer tube, which is the cylinder of radius $R=R_{outer}$, where $R_{outer}=10 R_i$ is defined in figure 4. Moreover, Poiseuille velocity profiles are imposed at the upstream and downstream boundaries of the flow domain (see \citet{JFM12} for further details). Regarding the perturbed flow, it was demonstrated by \cite{Taylor1964,AcrivosyLo,HinchAcrivos} that, when a slender drop or bubble with a viscosity much smaller than that of the outer fluid ($\
lambda\ll 1$) is immersed within a straining flow, the first order modification to the unperturbed outer velocity field can be expressed as a superposition of sources located at the axis of symmetry. This is so because the contribution to the perturbed flow field associated with the addition of Stokeslets at $R=0$, is of the order of $\sim O[(R_i/L)^2]$ \citep{Buckmaster}. Here, $L$ denotes the characteristic axial length scale for which the radius of the thread suffers variations of its same order of magnitude and, in the case of slender flows, $R_i/L\ll 1$. The same idea can thus be applied to describe the experiments depicted in figures \ref{fig1}-\ref{fig2} since, in all the cases under study here, $\lambda\ll 1$ and the resulting jet is slender, namely, $\partial\,R_j/\partial\,Z\sim R_i/L\ll 1$. Therefore, the outer flow field can be approximately expressed as~\citep{JFM12},
\begin{equation}
{\bf U}\simeq U_n{\bf
e_z}+\left(V_n+\dfrac{S(Z,T)}{R}\right){\bf
e_r}\simeq\left[U_x(Z)+F(Z) R^2\right]{\bf e_z}+
\left[-\dfrac{R}{2} \dfrac{{\rm d}U_x}{{\rm d}Z}-
\dfrac{R^3}{4}\dfrac{{\rm d}F}{{\rm d}Z}+\dfrac{S(Z,T)}{R}\right]{\bf e_r}\,.\label{U1}
\end{equation}
In equation~\eqref{U1}, $\mathbf{e_r}$ and $\mathbf{e_z}$ are the
unit base vectors,
\begin{equation}
V_n=-\dfrac{R}{2} \dfrac{{\rm d}U_x}{{\rm d}Z}-
\dfrac{R^3}{4}\dfrac{{\rm d}F}{{\rm d}Z}\label{Vn}
\end{equation}
has been expressed as a function of $U_n$ making use of the continuity equation in cylindrical
coordinates, namely, $\partial U_n/\partial Z+R^{-1}\partial (RV_n)/\partial R=0$ and, as it was pointed out above, the axial component of the velocity field is expressed retaining just the first two terms in the in Taylor series expansion of $U_n$, what leads to
\begin{equation}
U_n(R,Z)=U_n(Z,R=0)+1/2\, \partial^2 U_n/\partial\,R^2 (Z,R=0)\,R^2+O[(R_j/R_i)^4]\simeq U_x(Z)+F(Z)\,R^2\, .\label{Un}
\end{equation}
Moreover, in equation (\ref{U1}), only the radial component of the perturbed velocity field created by a distribution of sources, $S(Z,T)/R\,\mathbf{e_r}\sim O(R_j/L)$, has been retained in the analysis due to the fact that the order of magnitude of the corresponding axial component is $\sim O[(R_j/L)^2]\ll 1$ (see, e.g. \cite{Ashley}).
Now notice that the unknown source intensity $S(Z,T)$ can be easily expressed as a function of the jet radius $R_j(Z,T)$ by means of the kinematic boundary condition at the interface,
\begin{equation}
\dfrac{\partial R_j}{\partial T}+\left[U_x(Z)+F(Z)R_j^2(Z,T)\right]
\dfrac{\partial R_j}{\partial Z}=V_n(R=R_j(Z,T))+\dfrac{S(Z,T)}{R_j(Z,T)}.\label{S0}
\end{equation}
Using the expression for $V_n$ given in equation (\ref{Vn}), equation~\eqref{S0} yields
\begin{equation}
\dfrac{S}{R_j}=\dfrac{\partial R_j}{\partial T}+\left(U_x+F R_j^2\right)
\dfrac{\partial R_j}{\partial Z}+\dfrac{R_j}{2}
\dfrac{{\rm d}U_x}{{\rm d}Z}+\dfrac{R_j^3}{4}\dfrac{{\rm d}F}{{\rm d}Z}\,. \label{S}
\end{equation}
Making use of equation (\ref{S}), the equation for $R_j$ can be easily deduced using the continuity equation
\begin{equation}
\dfrac{\partial R^2_j}{\partial T}+\dfrac{\partial}{\partial\,Z}
\left[\left(U_x+F R_j^2\right)R_j^2-\frac{R_j^4}{8\mu_i}
\frac{\partial P_i}{\partial Z}\right]=0\,,\label{Continuidad}
\end{equation}
where it has been taken into account that the flow in the jet is the addition of the plug flow induced by the outer axial velocity field at the radial position where the interface is located plus the Poiseuille flow created by the inner pressure gradient, $-\partial P_i/\partial Z$. It is of interest to point out that the flow within the jet in our one-dimensional theoretical approach may exhibit a recirculation bubble, in contrast with the type of one dimensional models which assume that the velocity profile within the jet is uniform (see, e.g. \cite{Basaran1D2D,basaran04b,Rubio2013} and references therein). Indeed, a recirculation bubble could exist in our case if the negative velocities of the parabolic velocity profile associated with an adverse inner pressure gradient ($\partial P_i/\partial Z>0$), were larger in magnitude than the corresponding positive velocity of the plug flow velocity profile.
In (\ref{Continuidad}), notice that the inner stream pressure $P_i$, which in the slender approximation is constant in the
radial direction, is related to the outer pressure $P_n$ through the normal stress jump
\begin{equation}
P_i=P_n+\sigma \nabla\cdot{\bf n}-
2\mu_o\left[\dfrac{\partial V_n}{\partial R}(R=R_j)-\dfrac{S}{R_j^2}\right]\,,\label{Pi}
\end{equation}
where the term corresponding to the normal viscous stresses associated with the inner stream has been neglected due to the fact that $\lambda\ll 1$. Moreover $\nabla\cdot{\bf n}$ is the interfacial curvature, given by
\begin{equation}
\nabla\cdot{\bf n}=\dfrac{1}{R_j(1+\dot{R}^2_j)^{1/2}}-
\dfrac{\ddot{R}_j}{(1+\dot{R}^2_j)^{3/2}}\,,\label{curvatura}
\end{equation}
with dots denoting derivatives with respect to $Z$. In equation~\eqref{Pi}, notice also that we have taken into account the fact that the velocity field created by the presence of the jet does not contribute to modify the outer pressure field. Indeed, the velocity field $\mathbf{U'}$ generated by a distribution of sources can be expressed in terms of the gradient of a velocity potential, namely, $\mathbf{U'}=\nabla\Phi$ due to the fact that a single three dimensional source and thus, a superposition of sources, are solutions of the Laplace equation. Therefore, since in the zero Reynolds number limit the momentum equation reads
\begin{equation}
-\nabla P'+\mu_{o}\nabla\,^{2}\textbf{U}'=-\nabla
P'-\mu_{o}\nabla\times\left(\nabla\times\textbf{U}'\right)=-\nabla
P'-\mu_{o}\nabla\times\left(\nabla\times\nabla\Phi\right)=0\rightarrow -\nabla
P'=0\, ,
\end{equation}
the only contribution to the outer pressure field comes from the unperturbed flow field, $P_n$. Consequently, the outer axial pressure gradient in the near-axis region, where the
jet surface is located, is calculated by means of the axial projection of the momentum equation,
\begin{equation}
-\dfrac{\partial P_n}{\partial Z}=-\mu_o\left[\frac{\partial^2 U_n}{\partial\,Z^2}+\frac{1}{R}\frac{\partial}{\partial\,R}\left(R\frac{\partial\,U_n}{\partial\,R}\right)\right]\rightarrow
-\dfrac{\partial P_n}{\partial Z}=
-\mu_o\left(\dfrac{{\rm d}^2 U_x}{{\rm d}Z^2}+R^2\dfrac{{\rm d}^2 F}{{\rm d}Z^2}+
4F\right)\,,\label{P0}
\end{equation}
where the expression for $U_n$, given by equation (\ref{Un}), has been used.
Taking $R_i$, $R_i/U_o$ and $\mu_oU_o/R_i$ as characteristic scales for length, time and pressure
respectively, substituting~\eqref{P0} and~\eqref{S} into the
$Z$-derivative of~\eqref{Pi}, and introducing the final expression into~\eqref{Continuidad},
yields the following dimensionless partial differential equation for the jet radius,
\begin{equation}
\begin{split}
&\dfrac{\partial r_j}{\partial t}\left[2r_j+\dfrac{1}{4\lambda}
\left(2r_j\left(\dfrac{\partial r_j}{\partial z}\right)^2+r^2_j\,
\dfrac{\partial^2 r_j}{\partial z^2}\right)\right]-\dfrac{1}{2\lambda}\,r^2_j\,
\dfrac{\partial r_j}{\partial z}\dfrac{\partial}{\partial z}
\left(\dfrac{\partial r_j}{\partial t}\right)-\dfrac{1}{4\lambda}\,r^3_j\,
\dfrac{\partial^2}{\partial z^2}\left(\dfrac{\partial r_j}{\partial t}\right)+\\
&+\dfrac{\partial}{\partial z}\left[\left(u_x+f r_j^2\right)r_j^2+
\dfrac{1}{8\lambda}\left(-3r^4_j\dfrac{{\rm d}^2 u_x}{{\rm d}z^2}-3\,r^6_j\,
\dfrac{{\rm d}^2 f}{{\rm d}z^2}-
4fr^4_j-\Ca^{-1}\,r^4_j\dfrac{\partial\mathcal{C}}{\partial z}-6r^5_j\,
\dfrac{\partial r_j}{\partial z}\dfrac{{\rm d}f}{{\rm d}z}-\right.\right.\\
&\left.\left.-2\left(\dfrac{\partial r_j}{\partial z}\right)^2\,fr^4_j+
2r^2_j\left(\dfrac{\partial r_j}{\partial z}\right)^2\,u_x-2r^3_j
\dfrac{\partial r_j}{\partial z}\dfrac{{\rm d}u_x}{{\rm d}z}-2 u_x r^3_j
\dfrac{\partial^2 r_j}{\partial z^2}-2 f r^5_j
\dfrac{\partial^2 r_j}{\partial z^2}\right)\right]=0\, ,\label{Continuidad1}
\end{split}
\end{equation}
with $u_x$ and $f$ in~\eqref{Continuidad1} given in figure \ref{Uaxis}.
>From now on, lower case variables like those appearing in
equation~\eqref{Continuidad1} indicate the dimensionless version of
their dimensional counterparts and $\mathcal{C}$ indicates the non dimensional expression for
the interfacial curvature given by equation~\eqref{curvatura}. Notice that the equation for
the steady jet shape $r_{j0}(z)$ deduced in~\cite{JFM12}, namely
\begin{eqnarray}
&&q=\left(u_x+f r_{j0}^2\right)r_{j0}^2+\dfrac{1}{8\lambda}\left(-3r^4_{j0}
\dfrac{{\rm d}^2 u_x}{{\rm d}z^2}-3\,r^6_{j0}\,\dfrac{{\rm d}^2 f}{{\rm d} z^2}-
4 f r^4_{j0}-\Ca^{-1}\,r^4_{j0}\dot{\mathcal{C}}_0-\right.\nonumber\\
&&\left.-6 r^5_{j0}\,\dfrac{{\rm d}r_{j0}}{{\rm d}z}
\dfrac{{\rm d}f}{{\rm d}z}-2\left(\dfrac{{\rm d}r_{j0}}{{\rm d}z}\right)^2\,f r^4_{j0}
+2 r^2_{j0}\left(\dfrac{{\rm d}r_{j0}}{{\rm d}z}\right)^2\,u_x-\right.\nonumber\\
&&\left.-2 r^3_{j0}\dfrac{{\rm d}r_{j0}}{{\rm d}z}\dfrac{{\rm d}u_x}{{\rm d}z}-
2 u_x r^3_{j0}\dfrac{{\rm d}^2 r_{j0}}{{\rm d}z^2}-2 f r^5_{j0}
\dfrac{{\rm d}^2 r_{j0}}{{\rm d}z^2}\right)\label{steady}
\end{eqnarray}
can be recovered by simply equating to zero the time derivatives in~\eqref{Continuidad1}.
Let us point out that, in our description, the continuity of tangential stresses across the interface is never invoked since $\lambda\ll 1$. In fact, this condition can be used to calculate the distribution of Stokeslets at the axis, what would introduce a higher order correction ($\sim O[(R_i/L)^2]$) to the velocity field created by the distribution of sources \citep{Buckmaster}. It is also interesting to note that, in our approach to deduce equation (\ref{Continuidad1}), terms of the order of $\sim O[(R_i/L)^2]$ have been neglected in the expression for the outer velocity field given by (\ref{U1}). However, the azimuthal curvature term $\sim \partial^2 R_j/\partial\,Z^2$, which is also of the order of $\sim O[(R_i/L)^2]$, has been retained in equation (\ref{curvatura}). This apparent inconsistency is justified since, as pointed out in \cite{GyC,Basaran1D2D,basaran04b,EggersVillermaux,Rubio2013} and references therein, to improve the agreement with experiments in one dimensional models such as the one
developed here, the exact expression of the interfacial curvature needs to be retained in the analysis of capillary dominated flows. Also, it can be appreciated in equation (\ref{P0}) that, in order for the contribution of $P_n$ to the inner pressure gradient to possess the same degree of accuracy as the contribution of the capillary pressure, the terms $\partial^2 U_n/\partial Z^2$ and $R^2 \partial^2 F/\partial Z^2$, which are $\sim O[(R_i/L)^2]$ when compared to $F$, have also been retained in the analysis.
\subsection{Global stability analysis}
\label{subsec:gla}
To find whether the solutions of equation~\eqref{steady} are stable or not, we substitute
the ansatz
\begin{equation}
r_j=r_{j0}(z)+r_{j1}(z,t)\label{rj0rj1}\,,
\end{equation}
into equation~\eqref{Continuidad1} and retain only linear terms in $r_{j1}(z,t)$ and its
derivatives, yielding the following linear partial differential equation for the perturbed
jet radius,
\begin{equation}
\begin{split}
&\dfrac{\partial r_{j1}}{\partial t}\left[2 r_{j0}+\dfrac{1}{4\lambda}
\left(2 r_{j0}\dot{r}_{j0}^2+r^2_{j0} \ddot{r}_{j0}\right)\right]-
\dfrac{1}{2\lambda}\,r^2_{j0}\dot{r}_{j0}\dfrac{\partial}{\partial z}
\left(\dfrac{\partial r_{j1}}{\partial t}\right)-\dfrac{1}{4\lambda}\,r^3_{j0}\,
\dfrac{\partial^2}{\partial z^2}\left(\dfrac{\partial r_{j1}}{\partial t}\right)+\\
&+r_{j1}B_0+\dfrac{\partial r_{j1}}{\partial z}B_1+
\dfrac{\partial^2 r_{j1}}{\partial z^2}B_2+\dfrac{\partial^3 r_{j1}}{\partial z^3}B_3+
\dfrac{\partial^4 r_{j1}}{\partial z^4}B_4=0\,,\label{Continuidadresumida}
\end{split}
\end{equation}
where the $B_j$ are five functions of $r_{j0}(z),f(z),u_x(z)$ and their $z$-derivatives,
whose expressions~\eqref{AB0}-\eqref{AB4} are provided in Appendix~A.
Equation~\eqref{Continuidadresumida} needs to be solved subjected to the boundary condition
$r_{j1}(z=0)=0$, and to the additional boundary condition expressing the fact that the
flow rate at $z=0$ is independent of time and equal to $q$, namely
\begin{equation}
\begin{split}
&\dfrac{\partial r_{j1}}{\partial z}\left[-\dfrac{3}{4\lambda}\,r^5_{j0}\dot{f}-
\dfrac{r^3_{j0}}{4\lambda}\dot{u}_x+\dot{r}_{j0}\left(\dfrac{1}{2\lambda}r_{j0}^2\,u_x-
\dfrac{1}{2\lambda}r^4_{j0}\,f\right)-\dfrac{1}{8\lambda \Ca}
\left(r^2_{j0}\dot{r}_{j0}^2d_0^{-3/2}-\right.\right.\\
&\left.\left.-r^2_{j0}d_0^{-1/2}+3r^3_{j0}\dot{r}^2_{j0}\ddot{r}_{j0}\,d_0^{-5/2}-r^3_{j0}
\ddot{r}_{j0}\,d_0^{-3/2}-15\dot{r}_{j0}^2\ddot{r}_{j0}^2\,r^4_{j0}\,d_0^{-7/2}+\right.\right.\\
&\left.\left.+3\ddot{r}_{j0}^2\,r_{j0}^4\,d_0^{-5/2}+3\,r^4_{j0}\dot{r}_{j0}\,r^{(3)}_{j0}\,
d_0^{-5/2}\right)\right]+
\dfrac{\partial^2 r_{j1}}{\partial z^2}\left[-\frac{u_x}{4\lambda}\,r^3_{j0}-\right.\\
&\left.-\dfrac{f}{4\lambda}r^5_{j0}-\dfrac{1}{8\lambda \Ca}\left(-r^3_{j0}\dot{r}_{j0}\,
d_0^{-3/2}+6r^4_{j0}\dot{r}_{j0}\ddot{r}_{j0}\,d_0^{-5/2}\right)\right]+
\dfrac{\partial^3 r_{j1}}{\partial z^3}\,\dfrac{1}{8\lambda \Ca}\,r_{j0}^4\,d_0^{-3/2}=0\,.
\label{CC1}
\end{split}
\end{equation}
Equation~\eqref{CC1} has been deduced retaining only linear terms in $r_{j1}$ and its
derivatives in the expression resulting from the substitution of the ansatz~\eqref{rj0rj1}
into equation~\eqref{steady}.
Writing the perturbed radius as
\begin{equation}
r_{j1}(z,t)=e^{\omega t}\,\bar{r}_{j1}(z)\,,\label{Omega}
\end{equation}
the different eigenvalues $\omega$ and their corresponding global eigenfunctions
$\bar{r}_{j1}(z)$ are found numerically through a Chebychev spectral collocation method applied
to the expression resulting from the substitution of the ansatz~\eqref{Omega}
into~\eqref{Continuidadresumida}. For that purpose, the physical space $0\leq z\leq l_{\infty}$,
with $l_{\infty}\gg 1$ the dimensionless jet length, is mapped into the Chebychev space
$-1\leq \xi\leq 1$ through the transformation
\begin{equation}
z=\dfrac{c\,(1-\xi)}{1-\xi^2+\frac{2c}{l_\infty}}\,.\label{Transformacion}
\end{equation}
To cluster as many points as possible in the cone-jet transition region, which is where the
interfacial curvature suffers the strongest variations, the constant $c$ in~\eqref{Transformacion}
is chosen to be twice the distance from the injection nozzle where the curvature of the steady
jet shape is minimum. Forcing the discrete version of equation~\eqref{Continuidadresumida} to be
satisfied at the $N-2$ Chebychev collocation points within the physical domain (all the
collocation points except those at the boundaries, namely $z=0$ and $z=l_\infty$) and imposing
at $z=0$ the boundary condition $\bar{r}_{j1}(z=0)=0$ as well as the one given by the discrete
version of the equation resulting from the substitution of~\eqref{Omega} into~\eqref{CC1},
yields the following linear system of equations for the $N$ eigenvalues and their corresponding
eigenfunctions,
\begin{equation}
\omega\mathcal{A}\cdot\mathbf{\bar{r}_{j1}}=\mathcal{B}\cdot\mathbf{\bar{r}_{j1}}\,,\label{Eceig}
\end{equation}
where the eigenvector $\mathbf{\bar{r}_{j1}}$ contains the values of the perturbed jet radius at
the $N$ collocation points. The $N\times N$ matrices $\mathcal{A},\mathcal{B}$ in
equation~\eqref{Eceig} are respectively given by
\begin{equation}
\mathcal{A}=\mathrm{diag}\left(-\frac{1}{4\lambda}\,r^3_{j0}\right)\mathcal{D}^2-
\mathrm{diag}\left(\frac{1}{2\lambda}\,r^2_{j0}\dot{r}_{j0}\right)\mathcal{D}+
\mathrm{diag}\left(2\,r_{j0}+\frac{1}{4\lambda}\left(2\,r_{j0}\dot{r}_{j0}^2+r^2_{j0}\,
\ddot{r}_{j0}\right)\right)\mathcal{I}\,,\label{A}
\end{equation}
and
\begin{equation}
\mathcal{B}=\mathrm{diag}\left(B_4\right)\mathcal{D}^4+\mathrm{diag}\left(B_3\right)\mathcal{D}^3+
\mathrm{diag}\left(B_2\right)\mathcal{D}^2+\mathrm{diag}\left(B_1\right)\mathcal{D}+
\mathrm{diag}\left(B_0\right)\,,
\end{equation}
where $\mathrm{diag}(\cdot)$ indicates the diagonal matrix whose elements are the values of the
$z$-dependent argument particularized at the $N$ collocation points, $\mathcal{I}$ indicates the
identity matrix, and the column vector
\begin{equation}
\mathcal{D}^n\cdot \mathbf{\bar{r}_{j1}}
\end{equation}
contains the value of the $n$-th derivative of $\bar{r}_{j1}$ in the physical space at the $N$
collocation points. Let us point out that our stability results were independent of
the numerical jet length, $l_{\infty}$, provided that $l_{\infty}\gg 1$.
The steady solution $r_{j0}(z)$, which is found solving the nonlinear equation~\eqref{steady}
by means of a Newton-Raphson method, will be globally stable when the real parts of \emph{all}
the eigenvalues are negative, whereas it will be globally unstable if the real part of at least
one eigenvalue, is positive. Thus, to find the function $\Ca^*(\lambda,q)$ we proceed as follows:
for given values of $\lambda$ and $q$, we first find the unperturbed jet shape $r_{j0}(z)$ by
solving equation~\eqref{steady} for a sufficiently large value of $\Ca$. The good agreement of
the resulting function $r_{j0}(z)$ with experiments shown in~\citet{JFM12}, constitutes the
first evidence that supports the accuracy of our generalized slender body theory. Once the
steady solution is found, the matrices $\mathcal{A},\mathcal{B}$ in equation~\eqref{Eceig} can
be computed since they only depend on $r_{j0}$, its derivatives and on the control parameters
$(\lambda,\Ca,q)$. Then, we proceed to find the $N$ eigenvalues and their corresponding
eigenvectors using standard Matlab functions (see figures~\ref{fig4} and~\ref{fig5}). If the
real parts of all eigenvalues are negative, as in the cases shown in figures~\ref{fig4}a
and~\ref{fig5}a, $\Ca$ is decreased and the full process is repeated again until the real part
of one of the eigenvalues crosses zero, as illustrated in figures~\ref{fig4}d and~\ref{fig5}d.
Note that, for given values of $\lambda$ and $q$, the value of $\Ca^*$ is determined by the
condition that the real part of the leading eigenvalue is zero.
\begin{figure}
\centering
\includegraphics[width=1\textwidth]{figure6}%
\caption{(a,d) Spectra, (b,e) unperturbed jet radius and (c,f) eigenfunctions associated with
the oscillatory global mode 1 for $\lambda=0.01$, $q=10^{-3}$ and $\Ca=1>\Ca^*$ (globally
stable flow, a--c), and $\Ca=0.8<\Ca^*$ (globally unstable flow, d--f). From
figure~\ref{fig4}d, notice that $\omega_i>0$ when the real part of the leading eigenvalue is
$\omega_r>0$. Also, from figure~\ref{fig4}f, notice that the associated eigenfunction
experiences spatial oscillations. Thus, the unstable global mode of type 1 represents a
perturbation that grows and propagates in both space and time. Crosses indicate the position of the eigenvalues, whereas the leading eigenvalue (the one corresponding to the largest growth rate), is encircled. \label{fig4}}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=1\textwidth]{figure7}%
\caption{(a,d) Spectra, (b,e) unperturbed jet shape and (c,f) eigenfunctions associated with
the non-oscillatory global mode 2 for $\lambda=0.1$, $q=10^{-3}$ and values of $\Ca=1>\Ca^*$
(stable flow, a--c), and $\Ca=0.8<\Ca^*$ (unstable flow, d--f). From figure~\ref{fig5}d,
notice that $\omega_i=0$ when the real part of the leading eigenvalue is $\omega_r>0$.
Also, from figure~\ref{fig5}f, notice that the associated eigenfunction does not experience
any kind of spatial oscillations. Thus, the unstable global mode of type 2 represents the
temporal growth of a perturbation which does not oscillate neither in time nor in space. Crosses indicate the position of the eigenvalues, whereas the leading eigenvalue (the one corresponding to the largest growth rate), is encircled.
\label{fig5}}
\end{figure}
We have identified two different types of eigenfunctions that represent either a spatio-temporally
oscillating mode (mode 1, see figure~\ref{fig4}) or a non-oscillatory mode (mode 2, see figure~\ref{fig5}). Both
families are also differentiated by the fact that, contrarily to the case of the non-oscillatory
mode, the amplitude of the oscillatory eigenfunction is virtually zero in the cone region, as it
can be deduced from figure~\ref{fig4}. At this point, let us anticipate that, within the ranges of
$\lambda$ and $q$ explored in this study, the stability analysis predicts that the oscillatory
mode dominates over the steady one, so it is reasonable to expect that the shape of the
eigenfunction depicted in figure~\ref{fig4}f is somehow related with our experimental
observations. This is indeed the case since, as revealed by figures~\ref{fig1} and~\ref{fig2},
the cone region is unperturbed even in the cases in which the drop formation process is not
periodic (see also the movies provided as supplementary material). Moreover, the maximum amplitude of the eigenfunction
associated with the oscillatory mode is located slightly downstream the cone tip, and this is
consistent with the experimental observation that the aperiodic formation process is associated
with an unsteady emission of drops from a region located very close to the cone apex. Our global
stability analysis also reveals that the value of the critical capillary number is hardly dependent
on $q$ for values of the viscosity ratio $\lambda\lesssim 0.1$ (see figure~\ref{fig6}). This
finding is also in agreement with the experimental evidence shown in figure~\ref{fig2}, where it
can be appreciated that $\Ca^*$ is rather insensitive to changes in $q$. Therefore, in view of the
slight variation of $\Ca^*$ with $q$ shown in figure~\ref{fig6}, the dependence of the critical
capillary number with $\lambda$ is calculated for a fixed value of $q=10^{-3}$. The result of this
calculation, depicted in figure~\ref{fig7}, compares favorably with experiments, a fact that
further supports our theory. Since our theoretical predictions are much closer to the experimental
observations than those obtained assuming that the jet is cylindrical, we can conclude that the
strong variation of the jet radius in the axial direction plays an essential role in stabilizing
the drop formation process.
We would like to emphasize that our analysis is able to predict the transition from the regular
drop formation process that takes place when a thin steady ligament is issued from the cone apex
(jetting regime) to the regime in which droplets detach aperiodically at the cone-jet region
(dripping regime), but not the reverse dripping to jetting transition. Note also that the
destabilization of the regular drop formation regime observed in our experiments shares some similarities with
the transition from jetting to dripping observed in liquid jets stretched by
gravity~\citep{clanet99,Basaran1D2D,basaran04b}, where the dripping regime manifests itself by the formation of drops right from or slightly downstream the exit of the injection tube. The main difference with the gravitational case is that, in the physical situation analyzed here, dripping occurs right from the tip of the conical drop, whose global shape is stable except very close to the microdrop emission region. The local oscillations appearing at the tip explain why the emission process is not periodic and leads to the formation of unevenly sized drops when the flow is globally unstable (see figures 1f-j, 6f and the movies provided as supplementary material). Let us point out that we also observe the formation of drops right from the exit tube, but this regime is found for values of the capillary number notably smaller than those investigated here \citep{SuryoBasaran}. This type of dripping regime, which is much more similar to that observed in the gravitational case, has not been reported here since the
diameters of the drops obtained are imposed by the geometry of the injector and, thus, are not of interest for applications due to the fact that their sizes are of the order of 100 microns or even larger.
\begin{figure}
\centering
\includegraphics[width=1\textwidth]{figure8}%
\caption{Dependence of $\Ca^*$ on $q$ for the oscillatory mode 1 (left plot), and the
non-oscillatory mode 2 (right plot).\label{fig6}}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=1\textwidth]{figure9}%
\caption{The critical capillary number, $\Ca^*$, as a function of the viscosity ratio
$\lambda$. Symbols show the experimental transition points separating the monodisperse
($\Ca>\Ca^*$, figures~\ref{fig1}a-e), and polydisperse ($\Ca<\Ca^*$, figures~\ref{fig1}f-j)
droplet generation processes. The solid and dashed lines represent the instability thresholds
of the global modes 1 (oscillatory) and 2 (non-oscillatory), respectively. The dash-dotted
and dotted lines correspond to the absolute/convective (A/C) transition, computed applying
the zero group velocity condition to equation~\eqref{reldisp} in the limit $z\to\infty$, and to
Tomotika's dispersion relation, respectively.\label{fig7}}
\end{figure}
\subsection{Comparison with Tomotika's stability analysis and elucidation of the physical mechanism leading to the higher stability exhibited by stretched jets}
There is still a remaining question that needs to be answered: why do our stretched microjets
exhibit a more stable behavior than their cylindrical counterparts? At first sight it could
be thought that, since ${\rm d}u_x/{\rm d}z>0$~ (see figure \ref{Uaxis}), the growth
rate of perturbations decreases thanks to the kinematic mechanism described in~\citet{Tomotika2}
and in~\citet{FrankelWeihs1985}, whereby the amplitude of the perturbations decreases while
their wavelenght increases due to the fact that fluid particles are elongated under the action
of the outer straining flow field. However, we will show below that, although the jet is indeed
stabilized thanks to this kinematic mechanism, the main reason for the higher stability exhibited
by the stretched microjets of figures~\ref{fig1} and~\ref{fig2} is in fact related with the
gradient of capillary pressure.
Indeed, let us adopt the slender approximation of assuming that the wavelength of the perturbation,
$2\pi/k$, with $k$ the wavenumber, is much smaller than the length along which $r_{j0}$ experiences
variations of its same order of magnitude. This condition, clearly verified for $z\gg 1$, permits
to look for traveling-wave solutions to equation~\eqref{Continuidadresumida}, of the form
$r_{j1}=e^{i\left(kz-\omega\,t\right)}\bar{\bar{r}}_{j1}$, with $k\in \mathcal{R}$ and
$\bar{\bar{r}}_{j1}$ slowly varying functions of $z$. At leading order, this approach provides the
following dispersion relation for $\omega$ as a function of $k$,
\begin{equation}
\begin{split}
&D(\omega,k)=-i\omega\left(2 r_{j0}+\dfrac{r^3_{j0}}{4\lambda}\,k^2\right)+
\left(\dfrac{b_{c0}}{\Ca}+2 r_{j0}\dot{u}_x\right)+\\
&+\left[2 r_{j0} u_x+r^3_{j0}\left(f\left(4-\dfrac{2}{\lambda}\right)-
\dfrac{7}{4\lambda}\ddot{u}_x\right)\right]ik+\dfrac{r^2_{j0}}{8\lambda\Ca}\,k^2+\\
&+\left(\dfrac{r^3_{j0}}{2 \lambda}\,u_x+\dfrac{r^5_{j0}}{4\lambda}\,f\right)ik^3+
\dfrac{r^4_{j0}}{8\lambda\Ca}\,k^4=0\,,\label{reldisp}
\end{split}
\end{equation}
where the term $b_{c0}\in \mathcal{R}$, provided in equation~\eqref{bc0} of Appendix~A, arises
when linear perturbations in the expression for the \emph{full curvature} of the jet are retained.
Neglecting all the $z$-derivatives of $r_{j0}$, what would correspond to the case of a purely
cylindrical jet or to the solution of~\eqref{steady} at $z\rightarrow \infty$, leads to $u_x=1$
and to $b_{c0}=\dot{u}_x=\ddot{u}_x=f=0$. The resulting dispersion relation ~\eqref{reldisp} in the limit $z\rightarrow \infty$ is identical to Tomotika's long wave limit dispersion relation deduced by ~\cite{PRLGoldstein,PoFPowers}, a fact constituting a further proof of our theory. Applying the zero group velocity condition
$\partial\omega/\partial k=0$, $\omega_i=0$ to equation~\eqref{reldisp} in the limit $z\rightarrow \infty$,
the boundary separating the absolutely and convectively unstable states can be calculated as
explained in~\cite{AnnRevHuerre,Gordillo01c,PoF05}. The resulting A/C transition curve,
shown in figure~\ref{fig7} (dash-dotted line), is very close to that obtained from the much more
involved calculation~\citep{PRLGoldstein,PoFPowers,PREInnombrable} that makes use of Tomotika's dispersion
relation (dotted line).
However, as stated above, to understand the higher stability of stretched jets we must retain
the \emph{non-parallel} terms in the dispersion relation~\eqref{reldisp} associated with axial
variations of $r_{j0}$ and $u_x$.
The main effect of non-parallel terms is to decrease the real part of $-i\omega$, namely, the
growth rate of perturbations. Indeed, $b_{c0}\Ca^{-1}+2\,r_{j0}\dot{u}_x>0$ since $\dot{u}_x>0$
(see figure \ref{Uaxis}) and $b_{c0}\Ca^{-1}>0$ (see figure~\ref{fig8}). It was already
pointed out above that positive axial velocity gradients stabilize the
jet~\citep{Tomotika2,FrankelWeihs1985,EggersVillermaux}, but figure~\ref{fig8} reveals that
$2r_{j0}\dot{u}_x<b_{c0}\Ca^{-1}$ near the tip of the conical drop, whose steady shape is depicted in figure 6b. The reason why the non-parallel
term $b_{c0}\Ca^{-1}$ contributes to the stabilization of the stretched jet can be easily explained
as follows: the gradients of capillary pressure in~\eqref{Continuidad} produce an axial variation
of the flow rate per unit length given by
$\Delta\,Q_{i,cap}=\sigma/(8\mu_i)\partial\left(R_j^4\,\partial(-\nabla\cdot{\bf n})/
\partial Z\right)/\partial Z$. Using the slender approximation for the curvature, namely,
$\nabla\cdot{\bf n}\sim 1/R_{j}$, linearizing the resulting expression around $R_{j0}$, and
using dimensionless variables, yields
$\Delta q_{i,cap}\sim(\dot{r}_{j0}^2+r_{j0}\ddot{r}_{j0})\,r_{j1}/(4\lambda \Ca)$ which, as can
be appreciated in figure~\ref{fig8}, is an excellent approximation to $b_{c0}\Ca^{-1}$ in the
cone-jet transition region. Therefore, since $\dot{r}_{j0}^2+r_{j0}\ddot{r}_{j0}>0$, perturbations
such that $r_{j1}>0$ (resp. $r_{j1}<0$), that tend to increase (resp. decrease) the jet radius,
cause $\Delta\,q_{i,cap}>0$ (resp. $\Delta\,q_{i,cap}<0$), leading to
$\partial r_{j1}/\partial t<0$ (resp. $\partial r_{j1}/\partial t>0$) by virtue of the mass
balance~\eqref{Continuidad}, thus explaining the stabilizing effect of the gradient of capillary
pressure existing inside stretched jets.
\begin{figure}
\centering
\includegraphics[width=0.8\textwidth]{figure10}%
\caption{This figure shows, for $\lambda=0.01$ and $Ca=1$ that the function $b_{c0}\Ca^{-1}$
in~\eqref{reldisp} (dashed line) can be well approximated as
$(\dot{r}_{j0}^2+r_{j0}\ddot{r}_{j0})/(4\lambda \Ca)$ (solid line), and that $2r_{j0}\dot{u}_x$
(dotted line), is smaller that $b_{c0}\Ca^{-1}$, meaning that the stabilization mechanism
associated with the action of capillary forces on highly curved interfaces, dominates over the
kinematic stabilization produced by the elongation of fluid
particles~\citep{Tomotika2,FrankelWeihs1985}.
\label{fig8}}
\end{figure}
\section{Conclusions}
\label{sec:conclusions}
In this paper we study the effect of the capillary number on the size distribution of the
drops composing the microemulsions obtained when two immiscible liquid streams with $\lambda\ll 1$ co-flow at
low Reynolds numbers and highly stretched liquid threads are
produced~\citep{SuryoBasaran,Marin}. Our experiments reveal that, when $\Ca\gtrsim O(1)$
and the flow rate ratio is such that $q\ll 1$, a liquid jet with a diameter substantially
smaller than that of the injection tube is issued from the tip of a conical drop pinned
at the exit of the injector. We find that this ligament is unsteady and breaks into unevenly
sized droplets for $\Ca<\Ca^*(\lambda)$ with $\Ca^*\sim O(1)$ the critical capillary number
and that, thanks to the fact that the growing capillary perturbations are convected downstream
the cone-jet transition region, the thin liquid jet is steady and breaks into uniformly
sized drops when $\Ca>\Ca^*(\lambda)$. With the purpose of predicting the dependence of the
critical capillary number on $\lambda$, we have first determined the values of $\Ca$ for which the group velocity of neutral waves is zero under the parallel flow assumption. The
results of this analysis using the dispersion relation which describes the growth of
capillary perturbations in cylindrical capillary jets~\citep{Tomotika}, reveal that the
boundary separating the absolutely unstable region from the convectively unstable
one~\citep{PRLGoldstein,PoFPowers,PREInnombrable} is well above the experimental values of the
critical capillary number, a fact indicating that the non-cylindrical liquid ligaments of our experiments are more
stable than their cylindrical counterparts. Thus, to improve the agreement between
experiments and theory, we have performed a global stability analysis retaining the highly
stretched shape of the liquid jet as well as realistic inner and outer velocity fields.
For that purpose, following the slender body theory developed in~\citet{JFM12}, we have
deduced the partial differential equation~\eqref{Continuidad1} that describes the
spatiotemporal evolution of the jet radius and whose steady limit accurately reproduces
the shape of the unperturbed liquid ligaments. Also, we have shown here that our equation~\eqref{Continuidad1} is able to reproduce Tomotika's stability analysis in the limit of cylindrical jets of a low viscosity fluid, a fact that further supports the validity of our theory. Moreover, the results of our stability analysis, which
are in fair agreement with experimental measurements, indicate that the regime in which
monodisperse emulsions are produced corresponds to those values of the capillary number for which
the cone-jet system is \emph{globally stable}. We can thus conclude that the
gradients of interfacial curvature and of axial velocity play an essential role in the
propagation of perturbations along stretched capillary jets. Notice that this idea has also been successfully applied to describe the jetting to dripping transition experienced by gravitationally stretched capillary jets using the one-dimensional equations for the jet radius and the axial velocity~\citep{Rubio2013}. Let us point out that the simplicity of the one dimensional approximation used here to study the stability of stretched jets contrasts with the difficulty of analyzing the stability of two dimensional or three dimensional free surface flows \citep[see e.g.][]{Christodoulou}.
One of our major findings here is the discovery of a new stabilization mechanism that
resorts to the existence of capillary pressure gradients inside low Reynolds number
stretched jets that differs from the one based on the elongation of fluid
particles~\citep{Tomotika2}. To conclude, we would like to emphasize that equation~\eqref{Continuidad1}
is rather general in the sense that it can be applied to describe the propagation and
growth of capillary perturbations along the jets produced, for instance,
in flow-focusing devices operated at low Reynolds numbers. Indeed, notice that the influence of each
particular geometry in the equation for the jet radius~\eqref{Continuidad1} comes through
the functions $u_x$ and $f$, which can be easily found numerically by solving Stokes
equations subjected to the appropriate boundary conditions when the inner fluid is removed from the flow domain. Currently, we are extending the idea of decomposing the flow field as the addition of two simpler velocity fields to those cases in which the Reynolds number of the outer stream is large. The same type of ideas based on the slender body approach could also be applied to describe the stability of electrohydrodynamic tip streaming regimes that, under appropriate conditions, take place when electrical stresses act on the interface of liquid threads with a cone-jet structure \citep{DelaMora,BarreroAnnurev,Collins1,Collins2}.
\begin{acknowledgements}
JMG and FCC thank financial support by the Spanish MINECO under Project DPI2011-28356-C03-01
and the Junta de Andaluc\'ia under Project P08-TEP-03997. AS thanks financial support by the
Spanish MINECO under Project DPI2011-28356-C03-02. These research projects have been partly
financed through European funds. The authors wish to express their gratitude to Elena de
Castro-Hen\'andez for providing them with the numerical values of functions $f$ and $u_x$
and for her careful revision of the equations in the text.
\end{acknowledgements}
|
\section{Introduction}
Quantum transport at the nanoscale \cite{DiCarlo2004,Zimbovskaya2011,Cunibertibook} is a
blooming field where the properties of matter can be explored in a realm
where quantum effects become crucial. In particular, the control of quantum
interference phenomena and their interplay with the electronic structure
offers a fascinating opportunity to overcome some of the usual constraints
of our macroscopic classical world. \cit
{Nature2001,Nature2012,Nitzan2003,Bustos2013,Rickhaus} However, at the
nanoscale, both quantum \textit{and} classical behavior can be expected.
This last emerges from the unavoidable environmental degrees of freedom.
\cite{Ratner2013} An exciting example of the competition among those
behaviors is electron-transfer in natural and artificial photosynthesis.
There, the interplay between localizing interferences and environmentally
induced decoherence seems to have a fundamental role in optimizing excitonic
transfer. \cite{Huelga2008,Lloyd2009} This phenomenon falls in line with
what is known in low dimensional conductors. Indeed, transport properties of
highly ordered 1-D systems is determined by the fast quantum diffusion of
local excitations, and thus become weakened by decoherence. On the other
hand, in disordered 1-D wires, quantum coherence allows the destructive
interferences that produce electronic localization. While these phenomena
are roughly described by introducing imaginary energies in the Kubo
formulation, it is at the cost of overlooking charge conservation. \cit
{Thouless-Kirkpatrick}
Landauer's picture has almost no rival in what concerns to electronic
coherent transport.\cite{Landauer1999} In its simplest form, conductance is
determined by the transmission probability (either quantum or classical)
among electrodes. Paradoxically, quantum transmittance is much simpler to
evaluate than its classical counterpart. Thus, the great majority of work
focus on the evaluation of the coherent transmittance setting aside
incoherent processes. An extension of this approach, developed by Markus
\"{u}ttiker,\cite{Buttiker1986} applies the Kirchhoff laws to a system
connected to multiple terminals. This allows to consider different voltage
probes as well as multiple current sources and drains. The self-consistent
non-equilibrium chemical potentials at the voltmeters must ensure current
cancellation. The resulting transport coefficients fulfill the Onsager's
reciprocity relations. Additionally, B\"{u}ttiker had the crucial insigh
\cite{Buttiker1986PRB} that a voltage probe implies a classical measurement
and thus it acts as a decoherence source. This concept was further
formulated by D'Amato and Pastawski introducing a Hamiltonian description
\cite{Damato-Pastawski} (henceforth the DP model). In this description, the
decoherent local probes can be assimilated to incoherent scattering by
delta-function potentials\cite{GLBE1,GLBE2}. This is founded in the Keldysh,
Kadanoff and Baym's quantum fields formalism\cite{Danielewicz1984} for the
non-equilibrium Green's functions. \cit
{Danielewicz1984,Kadanoff1989,Rammer1986} There, the integro-differential
equations are simplified by evaluating the currents and chemical potentials
in a linearized scheme that involves a matrix containing only transmittances
among different points in the sample. The DP model also provides a compact
solution for an arbitrary distribution of incoherent local scattering
processes. These lead to a momentum relaxing decoherence that produces
diffusion and a further increase in the resistance. The final set of linear
equations relate the local chemical potentials and the currents through a
transmittances matrix. \cite{Datta90} This results in the Generalized
Landauer-B\"{u}ttiker Equations\ (GLBE) that solve the DP model.
The original presentation of the DP model is constrained to two terminal
problems. Thus, in spite of the growing need to include the effects of
decoherent processes,\cite{Maassen2009, Horvat2013} its applications
remained mostly reduced to a few one-dimensional problems. \cit
{Zimbovskaya2002,Zwolak2002,Gagel1996,Nozaki2008,PAni-CBP,Nozaki2012,Anantram2013}
Besides, since the method deals with a great number of self-consistent local
chemical potentials, it often involves a cumbersome matrix inversion. Thus,
a general multi-terminal formulation of the DP model for decoherent
transport and an efficient computational strategy are still lacking.
In this paper we generalize the D'Amato-Pastawski model for multi-terminal
problems, presenting a decimation-based method for the calculation of the
decoherent conductance. In Sec. \ref{sec:BasicTools} we introduce the basic
tools, based on a decimation procedure that yields the parameters of an
effective Hamiltonian. In Sec. \ref{sec:DP} we overview the original DP
model. In Sec. \ref{sec:Computational}, we generalize the DP model for
multi-terminal setups. We also provide a recursive algorithm for the
calculation of Green's functions of general banded Hamiltonians. Then, we
show two application examples. In Sec. \ref{sec:example1} we consider a
simple model of a phonon-laser (SASER) based on the electron-phonon
interaction in a quantum dot \cite{Kent} where we asses the role of
decoherence in the SASER efficiency. In Sec. \ref{sec:example2}, we consider
the spin dependent electronic transport in a ferromagnetic wire where the
Giant Magnetoresistance (GMR) \cite{Fert08} shows up. We show that our
formulation describes the complete cross-over from a quantum transport to
the GMR semiclassical regime. In Sec. \ref{sec:conclusions} we summarize our
results and conclude that our formulation can handle decoherent transport in
a wide variety of problems beyond the typical two-terminal calculations.
\section{\label{sec:BasicTools}Decimation Procedures and Effective
Hamiltonians}
Even the simplest quantum devices involve a huge number of degrees of
freedom and thus their study can not be carried out without proper
simplifications. For example, a tight-binding Hamiltonian describing a
device or molecule with $N$ states (or orbitals) is, \cite{Pastawski-Medina}
\begin{equation}
\hat{H}_{S}=\sum\limits_{i=1}^{N}\left\{ E_{i}\hat{c}_{i}^{\dagger }\hat{c
_{i}^{{}}+\sum\limits_{\substack{ j=1 \\ (j\neq i)}}^{N}\left[ V_{i,j}\hat{
}_{i}^{\dagger }\hat{c}_{j}^{{}}+V_{j,i}\hat{c}_{j}^{\dagger }\hat{c
_{i}^{{}}\right] \right\} . \label{Hamil-sys}
\end{equation
Here, $\hat{c}_{i}^{\dagger }$ and $\hat{c}_{i}^{{}}$correspond to the
creation and anihilation fermionic operators acting on the vacuum
\left\vert 0\right\rangle $. \ Site energies are $E_{i}$ and hopping
amplitudes $V_{i,j}$ define the matrix Hamiltonian whose single particle
eigenstates are $\left\vert k\right\rangle =\sum_{i}u_{i,k}\hat{c
_{i}^{\dagger }\left\vert 0\right\rangle $ of energy $\varepsilon _{k}$
which are filled up to the Fermi energy, $\varepsilon _{F}$.
The decimation procedures, inspired in the renormalization group techniques
of statistical mechanics \cite{Kadanoff1983,Jose1982}, seek to recursively
reduce the number of degrees of freedom of a general $N\times N$ Hamiltonian
into another of lower rank, without altering the physical properties. The
basic idea can be captured by considering a system with $N=3$ states whose
secular equation is
\begin{equation}
\left[
\begin{array}{ccc}
\varepsilon -E_{1} & -V_{12} & -V_{13} \\
-V_{21} & \varepsilon -E_{2} & -V_{23} \\
-V_{31} & -V_{32} & \varepsilon -E_{3
\end{array
\right] \left(
\begin{array}{c}
u_{1} \\
u_{2} \\
u_{3
\end{array
\right) =\left[ \varepsilon \mathbb{I}-\mathbb{H}_{S}\right] \overrightarrow
u}\equiv \overrightarrow{0}.
\end{equation
Quite often we are interested in the transfer of an excitation from an
initial state to another one, say 1 and 2. Thus, instead of diagonalizing
the matrix, we could isolate $u_{3}$ from the third row and use it to
eliminate $u_{3}$ in the first and the second equations. In this way, we
obtain a new set of equations where $u_{3}$ is \textit{decimated}
\begin{equation}
\left[
\begin{array}{cc}
\varepsilon -\overline{E}_{1} & -\overline{V}_{12} \\
-\overline{V}_{21} & \varepsilon -\overline{E}_{2
\end{array
\right] \left(
\begin{array}{c}
u_{1} \\
u_{2
\end{array
\right) \label{H_S2eff} \\
=[\varepsilon \mathbb{I}-\mathbb{H}_{\mathrm{eff.}}]\vec{u}=0.
\end{equation
The renormalized coefficients hide their non-linear dependence on the energy
variable $\varepsilon :$
\begin{equation}
\begin{array}{c}
\overline{E}_{1}=E_{1}+\Sigma _{1}(\varepsilon )=E_{1}+V_{13}\dfrac{1}
\varepsilon -E_{3}}V_{31}, \\
\overline{E}_{2}=E_{2}+\Sigma _{2}(\varepsilon )=E_{2}+V_{23}\dfrac{1}
\varepsilon -E_{3}}V_{32}, \\
\overline{V}_{12}=V_{12}+V_{13}\dfrac{1}{\varepsilon -E_{3}}V_{32}
\end{array}
\label{eq:decim-ex}
\end{equation
In this case, the terms $\Sigma _{j}(\varepsilon );j=1,2$ are the real
self-energies accounting for the energy shifts due to the coupling with the
eliminated state. Notice that as long as one conserves the analytical
dependence on $\varepsilon $ of $\Sigma _{j}$, the actual secular equation
is still cubic in $\varepsilon $ and provides the exact spectrum of the
whole system. This procedure can be performed systematically in a
Hamiltonian of any size $N\times N$ to end up with an effective Hamiltonian
of size one desires, in particular a $2\times 2$ one. The effective
interaction parameter $\overline{V}_{12}$, together with the self-energies
\Sigma _{j}$, accounts for transport through the whole sample. Their
dependence on $\varepsilon $ provides all the needed information on the
steady state transport as well as on quantum dynamics. \cite{Levstein1990}
In practice, it is convenient to add an infinitesimal imaginary part, $
\mathrm{i}\eta $, to each energy $E_{j}\rightarrow E_{j}-\mathrm{i}\eta $.
Since a finite $\eta >0$ is equivalent to a decay process, it ensures that
one recovers the retarded time dependences of the observables through a well
defined Fourier transform.
The terminals connected to the system are described as semi-infinite leads
coupled to it. They are handled in a similar way as the system itself. The
idea is to eliminate all the internal degrees of freedom decimating them
progressively, renormalizing the states of the system which are directly
coupled to the external reservoirs. For further clarification we consider a
lead modeled as a semi-infinite one dimensional chain,
\begin{equation}
\hat{H}_{L}=\sum\limits_{i=0}^{-\infty }\left\{ E_{i}\hat{c}_{i}^{\dagger
\hat{c}_{i}^{{}}-V\left[ \hat{c}_{i}^{\dagger }\hat{c}_{i-1}^{{}}+\hat{c
_{i-1}^{\dagger }\hat{c}_{i}^{{}}\right] \right\} ,
\end{equation
that yields a tridiagonal matrix of infinite dimension. The elements $E_{i}
's and $V$'s are now the diagonal and off-diagonal terms of a tridiagonal
matrix $\mathbb{H}_{L}.$ This lead is connected at the left of the system,
say, with site $1$:
\begin{equation}
\hat{V}_{SL}=V_{L}\left[ \hat{c}_{1}^{\dagger }\hat{c}_{0}^{{}}+\hat{c
_{0}^{\dagger }\hat{c}_{1}^{{}}\right] .
\end{equation
Instead of dealing with the whole Hamiltonian
\begin{equation}
\hat{H}=\hat{H}_{S}+\hat{H}_{L}+\hat{V}_{SL},
\end{equation
we perform the decimation procedure. It becomes particularly simple because
of the chain structure of the lead. The energy of the $i$-th site, is \
\textquotedblleft shifted\textquotedblright\ by the elimination of $(i-1)
-th site, which itself is shifted by sites at its left \cit
{Pastawski-Medina}, with the self-energies resulting in a
continued-fraction:
\begin{eqnarray}
\Sigma _{i} &=&V_{i,i-1}\dfrac{1}{\varepsilon -E_{i-1}-\Sigma _{i-1}
V_{i-1,i} \\
(i &=&0,-1,-2,...-\infty ) \notag
\end{eqnarray
In a perfect propagating channel: $V_{i,i-1}\equiv V$ and $E_{i}=E_{0}$, and
thus, $\Sigma _{i}=\Sigma _{i-1}\equiv \Sigma $, we arrive to the
self-consistent solution:
\begin{eqnarray} \label{Dyson}
&&\Sigma (\varepsilon )=\dfrac{V^{2}}{\varepsilon -E_{0}-\Sigma }=\Delta
(\varepsilon )-\mathrm{i}\Gamma (\varepsilon ). \notag \label{sigma-leads}
\\
&=&\dfrac{\varepsilon -E_{0}+\mathrm{i}\eta }{2}-\text{sgn}(\varepsilon
-E_{0})\sqrt{\left( \dfrac{\varepsilon -E_{0}+\mathrm{i}\eta }{2}\right)
^{2}-V^{2}},
\end{eqnarray
where the generalized square root \cite{SquareRoot} in the limit $\eta
\rightarrow 0^{+},$ yields the imaginary component of the self-energy for
\varepsilon $ within the band of allowed energies. It becomes real otherwise.
Thus, once the states in the left lead are fully decimated the energy of the
first site becomes
\begin{eqnarray}
\widetilde{E}_{1}(\varepsilon ) &=&\overline{E}_{1}(\varepsilon )+\Sigma
_{L1}(\varepsilon ) \\
\text{with~~}\Sigma _{L1}(\varepsilon ) &=&\left( \frac{V_{L}}{V_{{}}
\right) ^{2}\Sigma (\varepsilon ) \\
&=&\Delta _{L1}(\varepsilon )-\mathrm{i}\Gamma _{L1}(\varepsilon ) \label{self_energy}
\end{eqnarray
As before, the real part $\Delta _{L1}(\varepsilon )$ indicates how the
unperturbed site energies are shifted by the leads. The important difference
with the simple decimation example discussed above is that, as a consequence
of the infinite nature of the lead, the self-energies may acquire a finite
imaginary component, $\Gamma _{L1}(\varepsilon ),$ even in the limit $\eta
\rightarrow 0^{+}$. It describes the rate at which coherent density
excitation in the system decays into the lead propagating states.
Note that, the imaginary part is roughly consistent with the exponential
decays of the survival probability predicted by the Fermi Golden Rule (FGR).
For instance, in a \textquotedblleft system\textquotedblright\ with a single
state $\left\vert 1\right\rangle $ interacting with a lead, the survival
probability at time $t$ after it has been placed in state $\left\vert
1\right\rangle $ is,
\begin{eqnarray}
\left\vert \left\langle 1\right\vert \exp [-\mathrm{i}\hat{H}~t/\hbar
]\left\vert 1\right\rangle \theta (t)\right\vert ^{2} &\equiv &\left\vert
\mathrm{i}\hbar G_{11}^{R}(t)\right\vert ^{2} \\
&\simeq &\exp [-2\Gamma _{L1}(E_{1})t/\hbar ],
\end{eqnarray
where we introduced the time dependent retarded Green's function,
G_{11}^{R}(t)$. However, we remember that the self-energies obtained above
have an explicit functional dependence on $\varepsilon .$ In consequence,
the actual decay can depart from this naive exponential approximation.
Indeed, a quantum decay should start quadratically as $1-\left( V_{L}t/\hbar
\right) ^{2}$ turning into an exponential at very short times. At very long
times the decay may even become a non-monotonous. \cite{Rufeil2006} In
practice, we will stay in the exponential approximation by neglecting the
dependence on $\varepsilon $ unless it is close to a band edge.
For the sake of simplicity, we may idealize the terminal leads as quasi 1-D
wires. As waveguides, they can be described in terms of open channels at the
Fermi energy or propagating modes. Thus, we chose a basis for the system's
Hamiltonian in which each independent propagation mode $l$ of a lead is
connected to a single system's state. This might require a unitary
transformation to choose a system's basis that matches the propagating modes
of leads (see Fig. \ref{Gr:leads-indep}). There is no restriction to the
converse: i.e. each \textquotedblleft site\textquotedblright\ can be coupled
to different quantum channels. Since the leads can be represented by
homogeneous infinite tight-binding chains, their decimation is just the
procedure implemented above with the appropriate $V$'s and $E$'s describing
each mode $l$.
\begin{figure*}[tbh]
\begin{center}
\includegraphics[width=5.5in]{fig1.eps}
\end{center}
\caption{Diagrammatic representation of an unitary transformation of the
system to a basis in which leads are independent. Here, dots represent
diagonal elements of the Hamiltonian in a site basis and lines non-diagonal
ones.}
\label{Gr:leads-indep}
\end{figure*}
The observation of DP was that any \textquotedblleft
local\textquotedblright\ electronic state weakly coupled to a huge number of
environmental degrees of freedom should decay from its initial decoupled
state according to the FGR. This would require a restitution or re-injection
of any escaping particle. Thus the DP model treats these decoherent
scattering channels sources as on-site fictitious voltage probes. Much as it
occurs with real voltmeters, local current conservation on each scattering
channels must be imposed. This ensures that each electron with definite
energy that escapes from a state towards a fictitious probe, is balanced by
an electron \textit{with the same energy} re-injected into the same state.
In the DP model, these decoherent channels are described by local
corrections to site energies of the sample, on the same footing as the real
channels:
\begin{equation}
\hat{\Sigma}_{\phi i}=-\mathrm{i}\Gamma _{\phi i}\hat{c}_{i}^{\dagger }\hat{
}_{i}^{{}}. \label{sigma-decoher}
\end{equation
Here, $\Gamma _{\phi i}$ represents an energy uncertainty associated with
the interaction process $\phi $ that mixes the local electron state $i$ with
environmental degrees of freedom. This introduces a decay of the state $i$
that can be described by the FGR. Notice that the state $i$ does not
necessarily represent a local basis, but it could be a channel mode or a
momentum basis state as well. The energy uncertainties due to decoherent
processes can be estimated for each specific process, \cite{PAni-CBP} and
may not necessarily be the same for every state $i$. Accordingly, each
\textquotedblleft site\textquotedblright\ $i$ may be subject to different
decay processes $\alpha $: those associated with real leads, $\alpha ={l,}$
and those related to decoherent processes (or fictitious probes), $\alpha =
\phi }$. The resulting effective Hamiltonian, $\hat{H}_{\mathrm{eff.}}$,
that includes the real and fictitious probes, is non-Hermitian \cit
{Rotter2009}:
\begin{equation}
\hat{H}_{\mathrm{eff.}}=(\hat{H}_{S}-\mathrm{i}\eta \hat{I
)+\sum\limits_{\alpha }\sum\limits_{i=1}^{N}\hat{\Sigma}_{\alpha i}.
\label{Hamil-efectivo}
\end{equation
Here, $\mathrm{Im}\hat{\Sigma}_{\alpha i}\neq 0$ only for those sites $i$
subject to decoherent processes ($\alpha =\phi $) or escapes to the leads (
\alpha =l$). Trivially, if the full imaginary part correction were
homogeneous (the same value for each state $i$), it just shifts the
eigenenergies into the complex plane. In contrast, inhomogeneous corrections
might produce spectral bifuctations that result in a quantum dynamical phase
transition. \cite{Dente2008}
In transport problems, most of the information on system dynamics is
distilled into the retarded and advanced Green functions. More practical
expressions are obtained using its Fourier transform into the energy
variable $\varepsilon $, from the effective Hamiltonian given by Eq. \re
{Hamil-efectivo}. In matrix representation:
\begin{equation}
\mathbb{G}^{R}(\varepsilon )=\left[ \varepsilon \mathbb{I}-\mathbb{H}_
\mathrm{eff}.}\right] ^{-1}=\mathbb{G}^{A\dagger }(\varepsilon )
\label{eq:green-def}
\end{equation
These Green's functions contain all the information of the quantum system
coupled to the leads and environment and constitute the kernel to move into
the non-equilibrium problem. Also, diagonal elements provide the
\textquotedblleft local\textquotedblright\ density of state
\begin{equation}
N_{i}(\varepsilon )=-\frac{1}{\pi }\mathrm{Im}G_{i,i}^{R}(\varepsilon )=
\frac{1}{2\pi \mathrm{i}}\left[ G_{i,i}^{R}(\varepsilon
)-G_{i,i}^{A}(\varepsilon )\right]
\end{equation
In particular, the transmission amplitudes of electronic excitations\
between the channels identified with process $\alpha $ at site $i$ and
process $\beta $ at site $j$ can be evaluated from the generalized form of
Fisher-Lee formula \cite{Pastawski-Medina}
\begin{equation}
t_{\alpha i,\beta j}(\varepsilon )=\mathrm{i}2~\sqrt{\Gamma _{\beta
j}^{{}}(\varepsilon )}~G_{j,i}^{R}(\varepsilon )~\sqrt{\Gamma _{\alpha
i}^{{}}(\varepsilon )}
\end{equation
and the transmission probabilities are given by
\begin{eqnarray}
T_{\alpha i,\beta j}(\varepsilon ) &=&\left\vert t_{\alpha i,\beta
j}(\varepsilon )\right\vert ^{2}\text{~~~~}(\alpha i\neq \beta j) \notag \\
&=&4\Gamma _{\beta j}(\varepsilon )G_{j,i}^{R}(\varepsilon )\Gamma _{\alpha
i}(\varepsilon )G_{i,j}^{A}(\varepsilon ) \label{Fisher-Lee-generalizada}
\end{eqnarray
where ${\Gamma }_{\alpha i}=\mathrm{i}(\Sigma _{\alpha ,i}^{R}-\Sigma
_{\alpha ,i}^{A})/2$ is proportional the escape rate at site $i$ due to a
process $\alpha $.
\section{\label{sec:DP}Two-terminal D'Amato-Pastawski Model.}
Retarded and advanced Green's functions and the transmission probabilities
associated with them contain the basic quantum dynamics. In order to
describe the non-equilibrium properties of a system, one has to evaluate the
density matrix or simply the diagonal terms of non-equilibrium density
functions,
\begin{equation}
G_{j,j}^{<}(\varepsilon )=\mathrm{i}2\pi N_{j}(\varepsilon )\mathrm{f
_{j}(\varepsilon ). \label{eq:Keldyshdensitydef}
\end{equation
These, in turn, are determined by the boundary conditions imposed by the
external reservoirs $\beta j$ that act as a source or drain of particles.
Their occupation is described by a non-equilibrium distribution function
approximated by a shifted Fermi distribution $\mathrm{f}_{\beta
j}(\varepsilon )=1/(\exp [\left( \varepsilon -\varepsilon _{F}-\delta \mu
_{\beta j}\right) /k_{B}T])$. In the Quantum Fields formalism, the $G_{\phi
j,\phi j}^{<}(\varepsilon )$ Green's functions result from the quantum
evolution in presence of the boundary conditions. In the time independent
case, energy is conserved, and the non-equilibrium density function takes
the form,
\begin{equation}
G_{j,k}^{<}(\varepsilon )=2\mathrm{i}\sum\limits_{\alpha
i}G_{j,i}^{R}(\varepsilon )\Gamma _{\alpha i}(\varepsilon )\mathrm{f
_{\alpha i}(\varepsilon )G_{i,k}^{A}(\varepsilon ),
\label{eq:Keldyshdensity-integral}
\end{equation
i.e. densities and correlations inside the system result from the
occupations $\mathrm{f}_{\beta i}(\varepsilon )$ imposed by the
experimentalist at the current terminals and the environment at the\
\textquotedblleft fictitious\textquotedblright\ probes. The equilibrium
density function $G_{j,j}^{(0)<}(\varepsilon )$ results when $\delta \mu
_{\beta j}\equiv 0$ for all $\beta j$. The actual observables are evaluated
from this non-equilibrium density function. The change respect to the
equilibrium in the local density can be expressed in terms of the above
boundary conditions as \cite{GLBE2}:
\begin{eqnarray}
\delta \rho _{j}^{{}} &=&-\frac{\mathrm{i}}{2\pi }\int \left[
G_{j,j}^{<}-G_{j,j}^{(0)<}\right] \mathrm{d}\varepsilon \\
&\simeq &N_{j}(\varepsilon _{F})\delta \mu _{j}, \notag
\end{eqnarray
while the currents between sites $i$ and $j$ are given b
\begin{equation}
I_{i,j}=\int \left[ V_{i,j}G_{j,i}^{<}-V_{j,i}G_{i,j}^{<}\right] \mathrm{d
\varepsilon .
\end{equation
These integral expressions of the observables, expressed in the linear
response approximation of small biases $e\mathtt{V}_{L}=\mu
_{Li}-\varepsilon _{F}\ll \varepsilon _{F}$, become the Generalized
Landauer-B\"{u}ttiker equations that describe the balance of electronic
current. These are no other than the Kirchhoff laws expressed in terms of
the generalized Landauer's conductances, given by the Fisher-Lee formulas of
Eq. \ref{Fisher-Lee-generalizada}. Because of the linear approximation these
transmittances are evaluated at the Fermi energy, and now become:
\begin{equation}
I_{\alpha i}=\frac{e}{h}\underset{\text{processes}}{\sum\limits_{\beta
=L,\phi }}\underset{}{}\underset{\text{sites}}{\sum\limits_{j=1(\alpha i\neq
\beta j)}^{N}}\left( T_{\alpha i,\beta j}\delta \mu _{\beta j}-T_{\beta
j,\alpha i}\delta \mu _{\alpha i}\right) \label{eq:Kirchhoff1}
\end{equation
where the quantities $\delta \mu _{\alpha i}=\mu _{\alpha i}-\varepsilon
_{F},$ are the chemical potentials of the electron reservoirs, at state $i$
for a process $\alpha $.
The requirement in the DP model that no net current flows through the
decoherent channels imposes
\begin{equation}
0\equiv I_{\phi i}. \label{eq:CurrentConservation}
\end{equation
These equations imply the self-consistent determination of the internal
non-equilibrium chemical potentials $\delta \mu _{\phi i}.$ Thus, we are
faced to a linear problem. Once again, its solution can be laid as a
decimation procedure, as we did to obtain the effective Hamiltonian.
Consider the case where two real leads are connected to the sites $1$ and $N$
of the system (thus identified as channels $\ell 1$ and $\ell N$), and a
\textit{single} decoherent process $\phi k$ is connected to the state $k$.
Thus, charge conservation implies:
\begin{equation}
0=T_{\phi k,\ell 1}\delta \mu _{\ell 1}+T_{\phi k,\ell N}\delta \mu _{\ell
N}-(T_{\ell 1,\phi k}+T_{\ell N,\phi k})\delta \mu _{\phi k},
\label{eq:Decim-T_first}
\end{equation
which can be rewritten as:
\begin{equation}
\delta \mu _{\phi k}=\frac{T_{\phi k,\ell N}}{(T_{\ell 1,\phi k}+T_{\ell
N,\phi k})}\delta \mu _{LN}+\frac{T_{\phi k,\ell 1}}{(T_{\ell 1,\phi
k}+T_{\ell N,\phi k})}\delta \mu _{\ell 1}
\end{equation
Using this relation for the current on real channels we obtain:
\begin{equation}
I_{\ell N}=-I_{\ell 1}=\frac{e}{h}\tilde{T}_{\ell N,\ell 1}(\delta \mu
_{\ell N}-\delta \mu _{\ell 1}), \label{eq:currentDP}
\end{equation
where $\tilde{T}_{\ell N,\ell 1}$ represent the \textquotedblleft
effective\textquotedblright\ transmission between leads $\ell 1$ and $\ell N$
after the decimation of the incoherent channel associated with $\phi k$,
given by:
\begin{equation}
\tilde{T}_{\ell N,\ell 1}=T_{\ell N,\ell 1}+T_{\ell N,\phi k}\frac{1}
(T_{\ell 1,\phi k}+T_{\ell N,\phi k})}T_{\phi k,\ell 1}. \label{eq:Decim-T}
\end{equation
Note that the zero current constrain at the decoherent channels allows us to
pile up (i.e. decimate) those processes into an incoherent contribution to
the total transmission. This is the reason why Eq. \re
{eq:CurrentConservation} is the key factor in the computation of the total
transmission. At this point one recognizes the analogy of the second term on
the right-hand side of Eq. \ref{eq:Decim-T} with the effective interaction
shown in Eq. \ref{eq:decim-ex}. This analogy will be used in the following
section to develop a simple matrix solution for the total decoherent
transmission in a multi-terminal setup. In the case of two current probes,
identifying the index label $L=\ell 1$ and $R=\ell N$ for the leads, and
\phi k=k$ for the decoherence probes, one has that the total transmission
probability is given by:\cite{Damato-Pastawski}
\begin{equation}
\tilde{T}_{L,R}=T_{L,R}+\sum\limits_{i,j}T_{R,i}\left[ \mathbb{W}^{-1}\right]
_{i,j}T_{j,L}. \label{eq:Teff_DP}
\end{equation
The elements of the matrix $\mathbb{W}$ are:
\begin{equation}
W_{ij}=-T_{ij}+\left( \sum\limits_{j=L,i,R}T_{ij}\right) \delta _{ij}.
\end{equation}
Eqs. \ref{eq:currentDP} and \ref{eq:Teff_DP} provide the decoherent current
and the effective transmission of DP model for two-terminal setups. However,
they need to be reformulated to deal with a multi-terminal setup as when
there are more than two externally controlled chemical potentials or when
one requires to discriminate among different processes that contribute to
the current.
\section{\label{sec:Computational}Multi-Terminal D'Amato-Pastawski Model}
The two-probe Landauer conductance requires the computation of a single
element of the Green's function matrix: that connecting sites where the
leads are attached. In a 1-D case, this is $G_{1N}$ (where $N$ is the number
of sites of the system) and can be calculated through a decimation procedure
\cite{Levstein1990} While this can be readily generalized to deal with
finite systems of any dimension, not all formulations result numerically
stable in presence of strong disorder or band gaps.\cite{PastawskiSlutzky}
We will present a particular algorithm that is stable in such conditions.
The method is applicable to block tridiagonal Hamiltonians. These are very
common in many physicaly relevant situations, specifically when interactions
are truncated, or when the Hamiltonian matrix presents some form of banded
structure.
The DP model requires the computation of the transmittances among all
possible pairs of fictitious and physical probes, roughly $M(M-1)/2$, where
M~(\leq N)$ is the number of phase-breaking scattering channels. Also the
computation of the effective transmission requires the inversion of $\mathbb
W}$, a $M\times M$ matrix, as expressed in Eq. \ref{eq:Teff_DP}. It is our
purpose to extend the scheme of the DP model to account for decoherence in
quantum transport problems that involves many terminals. We seek for a
decoherent transmission analogous to Eq. \ref{eq:Teff_DP} for each pair of
physical leads. Thus, the computational approach to the DP model would
require an efficient matrix inversion algorithm.
In the next subsection, we present a computational procedure that, being
based on decimation schemes, preserves the physical meaning of matrix
inversions. This may allow one to take advantage of system's symmetries as
they can usually be expressed as relations between $\mathbb{G}$'s elements.
\subsection{Green's Function and recursive algorithms.}
In order to obtain the Green's functions of Eq. \ref{eq:green-def}, a matrix
inversion is needed. The \textit{matrix continued fractions} \cit
{Butler1973,MCF-Pastawski} scheme offers a decimative approach well suited
to perform this task. This procedure can be constructed recalling the well
known $2\times 2$ block matrix inversion,
\begin{widetext}
\begin{equation}
\left[
\begin{array}{cc}
\mathbb{A} & \mathbb{B} \\
\mathbb{C} & \mathbb{D
\end{array
\right] ^{-1}=\left[
\begin{array}{cc}
(\mathbb{A}-\mathbb{BD}^{-1}\mathbb{C})^{-1} & -\mathbb{A}^{-1}\mathbb{B}
\mathbb{D}-\mathbb{CA}^{-1}\mathbb{B})^{-1} \\
-\mathbb{D}^{-1}\mathbb{C}(\mathbb{A}-\mathbb{BD}^{-1}\mathbb{C})^{-1} &
\mathbb{D}-\mathbb{CA}^{-1}\mathbb{B})^{-1
\end{array
\right] , \label{eq:BlockInvert}
\end{equation
\end{widetext}where $\mathbb{A}$, $\mathbb{B}$, $\mathbb{C}$ and $\mathbb{D}$
are arbitrary size subdivisions of the original matrix.
Let's assume that we have an effective Hamiltonian, $\hat{H}_{\mathrm{eff.}}$
which has block tridiagonal structure. We start \textquotedblleft
partitioning\textquotedblright\ the basis states in two portions: a cluster
labeled as $1$ that contains the first block, and the cluster of remaining
states of the system which we label as $B$. Thus, the Green's function
matrix in Eq. \ref{eq:green-def} is subdivided into four blocks,
(\varepsilon \mathbb{I}-\mathbb{E}_{1})$,$(\varepsilon \mathbb{I}-\mathbb{E
_{B})$,$-\mathbb{V}_{1B}$, and $-\mathbb{V}_{B1}$ of dimensions $N_{1}\times
N_{1}$, $N_{B}\times N_{B}$, $N_{1}\times N_{B}$ and $N_{B}\times N_{1}$
respectively. Thus,
\begin{equation}
\mathbb{G}(\varepsilon )=\left[
\begin{array}{cc}
\mathbb{G}_{11} & \mathbb{G}_{1B} \\
\mathbb{G}_{B1} & \mathbb{G}_{BB
\end{array
\right] =\left[
\begin{array}{cc}
\varepsilon \mathbb{I}-\mathbb{E}_{1} & -\mathbb{V}_{1B} \\
-\mathbb{V}_{B1} & \varepsilon \mathbb{I}-\mathbb{E}_{B
\end{array
\right] ^{-1}. \label{eq:Green-block}
\end{equation
Here, it is important to recall that the effective Hamiltonian $\hat{H
_{eff.}$ already includes all corrections due to fictitious and real probes,
by virtue of Eq. \ref{Hamil-efectivo}. In this way, the block with energies
and interactions, denoted here by $\mathbb{E}_{i}$, contain the
self-energies that account for the openness of the system, and may be
complex numbers. Combining Eq. \ref{eq:BlockInvert} and Eq. \re
{eq:Green-block} is easy to show that,
\begin{equation}
\begin{array}{c}
\mathbb{G}_{11}=\left( \varepsilon \mathbb{I}-{\mathbb{E}}_{1}-\mathbf
\Sigma }_{1}^{(B)}\right) ^{-1}=\left( \varepsilon \mathbb{I}-\tilde{\mathbb
E}}_{1}\right) ^{-1}, \\
\mathbb{G}_{BB}=\left( \varepsilon \mathbb{I}-{\mathbb{E}}_{B}-\mathbf
\Sigma }_{B}^{(1)}\right) ^{-1}=\left( \varepsilon \mathbb{I}-\tilde{\mathbb
E}}_{B}\right) ^{-1}, \\
\mathbb{G}_{1B}=\mathbb{G}_{11}\mathbb{V}_{1B}(\varepsilon \mathbb{I}
\mathbb{E}_{B})^{-1}=\mathbb{G}_{11}\left[ \mathbf{\Sigma }_{1}^{(B)}\mathbb
V}_{B1}^{-1}\right] ,\text{ and} \\
\mathbb{G}_{B1}=\mathbb{G}_{BB}\mathbb{V}_{B1}(\varepsilon \mathbb{I}
\mathbb{E}_{1})^{-1}=\mathbb{G}_{BB}\left[ \mathbf{\Sigma }_{B}^{(1)}\mathbb
V}_{1B}^{-1}\right]
\end{array}
\label{eq:Green-blocks2}
\end{equation
Here, the similarity with Eq. \ref{eq:decim-ex} allows us to define the
block self energies, $\mathbf{\Sigma }$'s, which in this simple $2\times 2$
block scheme, are given by:
\begin{equation}
\begin{array}{c}
\left[ \mathbf{\Sigma }_{1}^{(B)}\mathbb{V}_{B1}^{-1}\right] =\left[ \mathbb
V}_{1B}(\varepsilon \mathbb{I}-\mathbb{E}_{B})^{-1}\right] , \\
\left[ \mathbf{\Sigma }_{B}^{(1)}\mathbb{V}_{1B}^{-1}\right] =\left[ \mathbb
V}_{B1}(\varepsilon \mathbb{I}-\mathbb{E}_{1})^{-1}\right]
\end{array}
\label{eq:Sigmas1}
\end{equation
Notice, that in the expressions of Eqs. \ref{eq:Green-blocks2} and \re
{eq:Sigmas1}, the inverse of the hopping matrix must cancel with the hopping
that enters in the self-energies definition. Since the hoppings may be
non-square matrices, this definition is crucial to avoid its inversion.
Considering the bracket factors $\left[ \mathbf{\Sigma }\mathbb{V}_{{}}^{-1
\right] $ as a single object ensures stability of the recurrence procedure.
The decimation of the degrees of freedom associated with the portion $B$ of
the effective Hamiltonian is implied in Eq. \ref{eq:Green-blocks2}, where:
\begin{equation}
\tilde{\mathbb{E}}_{1}=\mathbb{E}_{1}+\mathbf{\Sigma }_{1}^{(B)}=\mathbb{E
_{1}+\left[ \mathbb{V}_{1B}(\varepsilon \mathbb{I}-\mathbb{E}_{B})^{-1
\right] \mathbb{V}_{B1}. \label{eq:decimation_block1+R}
\end{equation
Likewise, the decimation of block $1$ into $B$ gives the effective block:
\begin{equation}
\tilde{\mathbb{E}}_{B}=\mathbb{E}_{B}+\mathbf{\Sigma }_{B}^{(1)}=\mathbb{E
_{B}+\left[ \mathbb{V}_{B1}(\varepsilon \mathbb{I}-\mathbb{E}_{1})^{-1
\right] \mathbb{V}_{1B}. \label{eq:decimation_blockR+1}
\end{equation
Note that with the adopted notation for the self energies, $\Sigma
_{i}^{(j)} $ is the correction to block site $i$ when all block sites
between $i$ and $j $ (with $j$ included) are decimated. Therefore the
supra-index in parentheses indicate the subspace that has been decimated.
Since we are dealing with tridiagonal block matrices, we may resort to a
further partition for the matrix inversion involved in Eq. \re
{eq:decimation_block1+R}. i.e. the block $B$ describes states that can be
subdivided into two clusters where the first one, labeled 2, corresponds to
the first tridiagonal block from $(\varepsilon \mathbb{I}-\mathbb{E}_{B})$.
The other block $B^{\prime }$ now satisfies $\mathbb{V}_{1B^{\prime }}\equiv
\mathbb{O}$. Then, we hav
\begin{equation}
\mathbb{G}(\varepsilon )=\left[
\begin{array}{c|cc}
\varepsilon \mathbb{I}-\mathbb{E}_{1} & -\mathbb{V}_{12} & \mathbb{O} \\
\hline
-\mathbb{V}_{21} & \varepsilon \mathbb{I}-\mathbb{E}_{2} & -\mathbb{V
_{2B^{\prime }} \\
\mathbb{O} & -\mathbb{V}_{B^{\prime }2} & \varepsilon \mathbb{I}-\mathbb{E
_{B^{\prime }
\end{array
\right] ^{-1}.
\end{equation
Again, we can also decimate the degrees of freedom associated with block $2
, taking
\begin{equation}
\begin{array}{cc}
\tilde{\mathbb{E}}_{1}=\mathbb{E}_{1}+\mathbf{\Sigma }_{1}^{(2)}, & \tilde
\mathbb{E}}_{B^{\prime }}=\mathbb{E}_{B^{\prime }}+\mathbf{\Sigma
_{B^{\prime }}^{(2)} \\
\multicolumn{2}{c}{\tilde{\mathbb{V}}_{1B^{\prime }}=\mathbb{V
_{12}(\varepsilon \mathbb{I}-\mathbb{E}_{2})^{-1}\mathbb{V}_{2B^{\prime }}
\end{array
\end{equation
which leads to an effective equation analogous to Eq. \ref{eq:Green-block},
in terms of the new effective block sites:
\begin{equation}
\left[
\begin{array}{cc}
\mathbb{G}_{11} & \mathbb{G}_{1B^{\prime }} \\
\mathbb{G}_{B^{\prime }1} & \mathbb{G}_{B^{\prime }B^{\prime }
\end{array
\right] =\left[
\begin{array}{cc}
\varepsilon \mathbb{I}-\tilde{\mathbb{E}}_{1} & -\tilde{\mathbb{V}
_{1B^{\prime }} \\
-\tilde{\mathbb{V}}_{B^{\prime }1} & \varepsilon \mathbb{I}-\tilde{\mathbb{E
}_{B^{\prime }
\end{array
\right] ^{-1}
\end{equation
Therefore, an expression analogous to Eq. \ref{eq:Green-blocks2} is
obtained:
\begin{equation}
\begin{array}{c}
\mathbb{G}_{11}=\left( \varepsilon \mathbb{I}-\mathbb{E}_{1}-\mathbf{\Sigma
_{1}^{(B^{\prime })}\right) ^{-1} \\
\mathbb{G}_{B^{\prime }B^{\prime }}=\left( \varepsilon \mathbb{I}-\mathbb{E
_{B^{\prime }}-\mathbf{\Sigma }_{B^{\prime }}^{(1)}\right) ^{-1} \\
\mathbb{G}_{1B^{\prime }}=\mathbb{G}_{11}\tilde{\mathbb{V}}_{1B^{\prime
}}(\varepsilon \mathbb{I}-\tilde{\mathbb{E}}_{B^{\prime }})^{-1} \\
\mathbb{G}_{B^{\prime }1}=\mathbb{G}_{B^{\prime }B^{\prime }}\tilde{\mathbb{
}}_{B^{\prime }1}(\varepsilon \mathbb{I}-\tilde{\mathbb{E}}_{1})^{-1
\end{array}
\label{eq:Green-blocks3}
\end{equation
where the diagonal blocks of the Green's function matrix involv
\begin{equation}
\begin{array}{c}
\mathbf{\Sigma }_{1}^{(B^{\prime })}=\left[ \mathbb{V}_{12}(\varepsilon
\mathbb{I}-\mathbb{E}_{2}-\mathbf{\Sigma }_{2}^{(B^{\prime })})^{-1}\right]
\mathbb{V}_{12} \\
\mathbf{\Sigma }_{B^{\prime }}^{(1)}=\left[ \mathbb{V}_{B^{\prime
}2}(\varepsilon \mathbb{I}-\mathbb{E}_{2}-\mathbf{\Sigma }_{2}^{(1)})^{-1
\right] \mathbb{V}_{2B^{\prime }
\end{array}
\label{eq:Sigmas_border}
\end{equation
Note that in the self-energies of Eq. \ref{eq:Sigmas_border}, the decimated
space (denoted by the supra-index) always includes one of the border blocks
(in this case, $1$ or $B$). However, as shown hereafter, the non-diagonal
terms can also be written in terms of the block self-energies $\mathbf
\Sigma }^{(1)}$'s and $\mathbf{\Sigma }^{(B^{\prime })}$'s
\begin{equation}
\begin{array}{c}
\mathbb{G}_{1B^{\prime }}=\mathbb{G}_{11}[\mathbf{\Sigma }_{1}^{(B^{\prime
})}\mathbb{V}_{12}^{-1}][\mathbf{\Sigma }_{2}^{(B^{\prime })}\mathbb{V
_{B^{\prime }2}^{-1}], \\
\mathbb{G}_{B^{\prime }1}=\mathbb{G}_{B^{\prime }B^{\prime }}[\mathbf{\Sigma
}_{B^{\prime }}^{(1)}\mathbb{V}_{1B^{\prime }}^{-1}][\mathbf{\Sigma
_{2}^{(1)}\mathbb{V}_{12}^{-1}]
\end{array}
\label{eq:G1Bprime}
\end{equation
Both expressions are crucial to visualize the seed of our recursive
procedure.
The generalization by further partition into an arbitrary number of clusters
is straightforward. The Green's functions are expressed as a product of
non-singular self-energy blocks that are calculated recursively.
Independently of how the effective Hamiltonian is subdivided, if there are
N $ blocks of arbitrary size and the entire system is decimated into the $i
-th and $j$-th block, we have simply as matrix continued fractions: \cit
{MCF-Pastawski}
\begin{equation}
\begin{array}{c}
\mathbf{\Sigma }_{i}^{(j)}=\left[ \mathbb{V}_{i,i+1}\left( \varepsilon
\mathbb{I}-\mathbb{E}_{i+1}-\mathbf{\Sigma }_{i+1}^{(j)}\right) ^{-1}\right]
\mathbb{V}_{i+1,i} \\
\mathbf{\Sigma }_{j}^{(i)}=\left[ \mathbb{V}_{j,j-1}\left( \varepsilon
\mathbb{I}-\mathbb{E}_{j-1}-\mathbf{\Sigma }_{j-1}^{(i)}\right) ^{-1}\right]
\mathbb{V}_{j-1,j}^{{}} \\
\text{for}~~~~j>i
\end{array
\end{equation
provided that the final structure preserves a block three-diagonal. We
recall that matrix inversions are further stabilized by the presence of the
imaginary site energies imposed by the real and fictitious probes (Eq. \re
{Hamil-efectivo}). In this way, the decimation of the entire system into the
arbitrary \textquotedblleft block\textquotedblright\ sites $i$ and $j$,
leads to the effective quantities
\begin{equation}
\begin{array}{c}
\tilde{\mathbb{E}}_{i}=\mathbb{E}_{i}+\mathbf{\Sigma }_{i}^{(1)}+\mathbf
\Sigma }_{i}^{(j)} \\
\tilde{\mathbb{E}}_{j}=\mathbb{E}_{j}+\mathbf{\Sigma }_{j}^{(i)}+\mathbf
\Sigma }_{j}^{(N)} \\
\tilde{\mathbb{V}}_{i,j}=\tilde{\mathbb{V}}_{i,j-1}(\varepsilon \mathbb{I}
\mathbb{E}_{j}-\mathbf{\Sigma }_{j}^{(1)})^{-1}\mathbb{V}_{j-1,j
\end{array
\end{equation
\textit{which determine exactly} each $(i,j)$ element of the total Green's
function,
\begin{equation}
\left[
\begin{array}{cc}
\mathbb{G}_{ii} & \mathbb{G}_{ij} \\
\mathbb{G}_{ij} & \mathbb{G}_{jj
\end{array
\right] =\left[
\begin{array}{cc}
\varepsilon \mathbb{I}-\tilde{\mathbb{E}}_{i} & -\tilde{\mathbb{V}}_{ij} \\
-\tilde{\mathbb{V}}_{ji} & \varepsilon \mathbb{I}-\tilde{\mathbb{E}}_{j
\end{array
\right] ^{-1}.
\end{equation
The last expression is similar to Eq. \ref{eq:Green-block}, and therefore we
have,
\begin{equation}
\begin{array}{c}
\mathbb{G}_{ii}=\left[ (\varepsilon \mathbb{I}-{\mathbb{E}}_{i})-\mathbf
\Sigma }_{i}^{(1)}-\mathbf{\Sigma }_{i}^{(N)}\right] ^{-1}, \\
\mathbb{G}_{jj}=\left[ (\varepsilon \mathbb{I}-{\mathbb{E}}_{j})-\mathbf
\Sigma }_{j}^{(1)}-\mathbf{\Sigma }_{j}^{(N)}\right] ^{-1}, \\
\mathbb{G}_{ij}=\mathbb{G}_{ii}\left[ \tilde{\mathbb{V}}_{ij}(\varepsilon
\mathbb{I}-\tilde{\mathbb{E}}_{j})^{-1}\right] , \\
\mathbb{G}_{ji}=\mathbb{G}_{jj}\left[ \tilde{\mathbb{V}}_{ji}(\varepsilon
\mathbb{I}-\tilde{\mathbb{E}}_{i})^{-1}\right]
\end{array}
\label{eq:Green-blocksF}
\end{equation
This procedure is shown diagrammatically on Fig. \ref{Gr:Decimation}.
\begin{figure*}[tbph]
\begin{center}
\includegraphics[width=5.5in]{fig2.eps}
\end{center}
\caption{Decimation scheme for the calculation of the elements of the
Green's function matrix.}
\label{Gr:Decimation}
\end{figure*}
Note that the diagonal elements are easily calculated evaluating $\mathcal{O
}N)$ energy corrections of the form $\Sigma _{i}^{(1)}$ and $\Sigma
_{i}^{(N)}$, where all the sites have been decimated into site $i$. Also, in
order to compute all the non diagonal elements of the Green's function
matrix in Eq. \ref{eq:Green-blocks3} and \ref{eq:Green-blocksF} we would
need to evaluate $\sim N^{2}$ energy corrections $\mathbf{\Sigma }_{i}^{(j)}
's. However, following the insight given in Eq. \ref{eq:G1Bprime}, for
tridiagonal block Hamiltonians, the non-diagonal block matrix elements of
the Green function can be obtained in terms of the diagonal ones, avoiding
the need of the evaluation of $\mathcal{O(}N^{2})$ terms $\Sigma _{i}^{(j)}
's. In this case, if the Hamiltonian matrix is subdivided in $N$ arbitrary
blocks, we have
\begin{eqnarray}
&&\underset{\text{where}~i<j}{\mathbb{G}_{ij}=\mathbb{G}_{ii}\pro
\limits_{k=i}^{j-1}\left[ \mathbf{\Sigma }_{k}^{(N)}\mathbb{V}_{k+1,k}^{-1
\right] }, \label{eq:thoulessR} \\
&&\underset{\text{where}~i<j}{\mathbb{G}_{ji}=\mathbb{G}_{jj}\pro
\limits_{k=j}^{i+1}\left[ \mathbf{\Sigma }_{k}^{(1)}\mathbb{V}_{k-1,k}^{-1
\right] }. \label{eq:thoulessL}
\end{eqnarray
Note that now it is not necessary to evaluate any extra $\mathbf{\Sigma }$
in order to calculate $\mathbb{G}_{ij}$ for $i\neq j$, because those
self-energies have been already calculated for the diagonal Green's
Functions matrix blocks, $\mathbb{G}_{ii}$. This implies that only $\mathcal
O}\left( N\right) $ self-energies are required for the calculation of the
whole Green's function. These equations can help to take advantage of
possible symmetries of the $\mathbb{V}$ and $\mathbf{\Sigma }$ matrices to
speed up even more the calculation of Green's Functions.
Although Eqs. \ref{eq:thoulessR}-\ref{eq:thoulessL} have been formally
written in terms of hopping matrix inverses, $\mathbb{V}^{-1}$, these
expressions are accurate even when the hopping matrices are singular. This
is because the hopping matrix inverse cancels out with the hopping in the
\Sigma $ definition, as it can be seen, for example, in Eq. \ref{eq:Sigmas1
. In most cases, $\mathbb{G}_{ij}^{R}=\mathbb{G}_{ji}^{R}$, and therefore
Eqs. \ref{eq:thoulessR} and \ref{eq:thoulessL} are equivalent. However, both
equations are needed in some cases of quantum pumping \cite{LuisMPQP} or in
the presence of magnetic fields. The origin of the extraordinary stability
of Eqs. \ref{eq:thoulessR}-\ref{eq:thoulessL} can be easily grasped
analytically by considering a linear chain with three sites and expressing
the self-energies in terms of continued fractions before applying Eq. \re
{eq:G1Bprime}. Explicitly
\begin{widetext}
\begin{equation}
G_{1,3}(\varepsilon )=\cfrac{1}{\varepsilon -E_{1}-V_{12}\cfrac{1}
\varepsilon -E_{2}-V_{23}\cfrac{1}{\varepsilon -E_{3}}V_{32}}V_{21}}V_{12
\cfrac{1}{\varepsilon -E_{2}-V_{23}\cfrac{1}{\varepsilon -E_{3}}V_{32}}V_{23
\cfrac{1}{\varepsilon -E_{3}}. \label{Eq-G13}
\end{equation}
\end{widetext}Here, we clearly see that the divergences in the last factor
are exactly canceled by the zeros of the second one, while the singularities
in this one, are canceled by the zeros in the first factor. This equation
holds when the elements $E$'s and $V$'s are replaced by matrices, with
\mathbb{V}_{n,n+1}$'s mixing subspaces $\mathbb{E}_{n}$ and $\mathbb{E
_{n+1} $ of different dimensions. In general, the divergences in $\left[
\mathbf{\Sigma }_{k}^{(N)}\mathbb{V}_{k-1,k}^{-1}\right] $ are compensated
by the zeros of the previous term, i.e. $\left[ \mathbf{\Sigma }_{k-1}^{(N)
\mathbb{V}_{k,k-1}^{-1}\right] $. Furthermore, the regularization of poles
and divergencies imposed by decoherent processes (see below) ensure the
numerical precision of this cancellation.
\subsection{Physical Observables in the Multi-Terminal D'Amato-Pastawski
model}
The application of the DP model for multi-terminal devices requires a
generalization of Eq. \ref{eq:Teff_DP}. To obtain the total transmission on
each terminal, we can take advantage of the decimation procedures discussed
above. Eq. \ref{eq:Kirchhoff1} is easily rearranged in terms of the
transmissivity $(1-R_{\alpha i})$ from each channel $\alpha i$. \cit
{Pastawski-Medina} For process $\alpha $ at site $i$, one defines
\begin{eqnarray}
\left\vert t_{\alpha i,\alpha i}\right\vert ^{2}+(1-R_{\alpha i})
&=&\left\vert t_{\alpha i,\alpha i}\right\vert ^{2}\underset{\left( \beta
j\neq \alpha i\right) }{+\sum_{\beta ,j}}T_{\beta j,\alpha i} \label{Eq:1-R}
\\
&=&(1/g_{\alpha ,i})=4\pi N_{i}\Gamma _{\alpha i}, \label{Eq:g-T}
\end{eqnarray
where $N_{i}$ is the density of states at the site $i$. The Fisher-Lee
formula is extended by defining a \textquotedblleft
self-transmission\textquotedblright\ $\left\vert t_{\alpha i,\alpha
i}\right\vert ^{2}$ that is not a transmittance in the standard sense, and
certainly it is not the diagonal term $T_{\alpha i,\alpha i}\equiv R_{\alpha
i}-1$. However, it is required to obtain the sum of Eq. \ref{Eq:g-T} as the
product of the local density of states and the decay rate. It describes all
the electrons that, at a certain instant, are leaving the $\alpha i^{\mathrm
th}}$ reservoir to eventually return after wandering around. The inclusion
of this term is important because it contributes to define $(1/g_{\alpha
,i}) $, which plays a central role in a Keldysh perturbative expansion \cit
{GLBE1,Pastawski-Medina} and in a time dependent formulation of transport.
\cite{GLBE2}
Therefore, in a steady state calculation is enough to express eq. \re
{eq:Kirchhoff1} as:
\begin{equation}
I_{\alpha i}=\frac{\left\vert e\right\vert }{h}\left[ (R_{\alpha i}-1)\delta
\mu _{\alpha i}+\sum\limits_{\beta =L,\phi }\underset{\alpha i\neq \beta j}
\sum\limits_{j=1}^{N}}T_{\alpha i,\beta j}\delta \mu _{\beta ,j}\right] .
\label{Eq:Kirchhoff-current}
\end{equation
It can be arranged in a compact matrix notation, separating the processes
associated with the leads from the decoherent ones. The actual currents at
the leads are arranged in the vector $\overrightarrow{I}_{\lambda }$ while
the vanishing currents at the decoherent channels, in $\overrightarrow{I
_{\phi }\equiv \overrightarrow{0}$. Thus,
\begin{equation}
\left(
\begin{array}{c}
\overrightarrow{I}_{\lambda } \\
\overrightarrow{0
\end{array
\right) =\frac{\left\vert e\right\vert }{h}\mathbb{T}\left(
\begin{array}{c}
\delta \overrightarrow{\mu _{\lambda }} \\
\delta \overrightarrow{\mu _{\phi }
\end{array
\right) . \label{eq:CurrentsMatrix}
\end{equation
Here, the non-diagonal elements of $\mathbb{T}$ are transmission
probabilities and thus, they are definite positive. In contrast, the
diagonal elements are negative. Thus, a sum over any column or row cancels
out. This matrix can also be subdivided in the same block structure:
\begin{equation}
\mathbb{T}=\left[
\begin{array}{cc}
\mathbb{T}_{\lambda \lambda } & \mathbb{T}_{\lambda \phi } \\
\mathbb{T}_{\phi \lambda } & \mathbb{T}_{\phi \phi
\end{array
\right] .
\end{equation
This notation stress that $\mathbb{T}_{\lambda \lambda }$ only involves
terms that connect real leads, $\mathbb{T}_{\phi \phi }$ only involves
transmissions between decoherent channels and, finally, the blocks $\mathbb{
}_{\lambda \phi }$ and $\mathbb{T}_{\phi \lambda }$ connect leads with
decoherent processes. Thus, both $\lambda $ and $\phi $ subscripts may be
vectors themselves indicating processes (current leads $\ell $ or dephasing
processes $\phi $) and states in the system ($n=1,...N$). For instance, for
a system with a single resonant state identified as 1 coupled to two
terminals and a single decoherent process, $\lambda =(L1,R1)$ and $\phi
=\phi 1$. The fact that on-site chemical potentials at decoherent channels
ensure that no net current flows through them, allows us to evaluate
\overrightarrow{\delta \mu }_{\phi }$, from Eq. \ref{eq:CurrentsMatrix}:
\begin{equation}
\overrightarrow{\delta \mu }_{\phi }=\left[ -\mathbb{T}_{\phi \phi }\right]
^{-1}\mathbb{T}_{\phi \lambda }\overrightarrow{\delta \mu }_{\lambda }.
\end{equation
Here, $\overrightarrow{\delta \mu }_{\phi }$ provides the chemical potential
profile at the sites undergoing decoherence. Notice that, if used in a local
space representation, these chemical potentials do not distinguish left from
right going electrons. Thus they induce momentum relaxing decoherence. \cit
{GLBE1,Gasparian96,Datta2007}
The decimative procedure involves a simple algebraic relation between the
real channels of the system and the chemical potentials associated with
currents drains or sources. From Eq. \ref{eq:CurrentsMatrix}, it is
straightforward to isolate $\overrightarrow{I}_{\lambda }$, arriving to the
expression
\begin{equation}
\overrightarrow{I}_{\lambda }=\frac{e}{h}\widetilde{\mathbb{T}}_{\lambda
\lambda }\overrightarrow{\delta \mu }_{\lambda }, \label{eq:Idec_matrix}
\end{equation
and therefore, the adimensional effective conductances are the non-diagonal
elements of the matrix
\begin{equation}
\tilde{\mathbb{T}}_{\lambda \lambda }=\mathbb{T}_{\lambda \lambda }+\mathbb{
}_{\lambda \phi }[-\mathbb{T}_{\phi \phi }]^{-1}\mathbb{T}_{\phi \lambda },
\label{eq:Teff_matrix}
\end{equation
where the first term represents the coherent transmissions while the second
involves all the possible transmissions undergoing at least one decoherent
process. This last term, involves the inversion of a typically big $N\times
N $ matrix. Notice that the matrix in square brackets would correspond to
\mathbb{W}$ in the original D'Amato and Pastawski's paper, see Eq. \re
{eq:Teff_DP}. \cite{Damato-Pastawski} However, the matrix inversion can be
performed resorting to a recursive decimation of the ${N}$ dephasing
channels, taken one by one. Starting from the first one, at each stage of
decimation, all the remaining probes and dephasing channels become
renormalized according to the following recursive scheme for the matrix
elements of $\tilde{\mathbb{T}}$:
\begin{eqnarray}
\tilde{T}_{ij}^{\left[ 0\right] } &=&T_{ij}^{{}} \label{Eq:Trans_order_0} \\
\tilde{T}_{ij}^{\left[ k\right] } &=&\tilde{T}_{ij}^{\left[ k-1\right] }
\tilde{T}_{i,k}^{\left[ k-1\right] }\frac{-1}{\tilde{T}_{k,k}^{\left[ k-
\right] }}\tilde{T}_{k,j}^{\left[ k-1\right] }. \label{Eq:Trans_recursive}
\end{eqnarray
Here, $k$ runs over the dephasing channel index $\phi {1}...\phi {N}$ and
\tilde{T}_{ij}^{\left[ k\right] }$ stands for the matrix element ${i,j}$ \
(each of them take the values $\{\ell {1}...\ell {M,}\phi {1,}...,\phi {N
\}) $ of matrix $\mathbb{T},$ after the decimation of $k$ incoherent
channels. This recursion algorithm could become particularly useful when
only the effective transmission among a few external channels is needed.
Once that all of them were decimated, we have an effective transmission
matrix $\tilde{\mathbb{T}}\equiv \tilde{\mathbb{T}}^{(N)}$ given by:
\begin{equation}
\widetilde{\mathbb{T}}=\left[
\begin{array}{cccc}
\tilde{R}_{\ell 1}-1 & \tilde{T}_{\ell 1,L2} & \cdots & \tilde{T}_{\ell
1,\ell M} \\
\vdots & \vdots & \ddots & \vdots \\
\tilde{T}_{\ell M,\ell 1} & \tilde{T}_{\ell M,\ell 2} & \cdots & \tilde{R
_{\ell M}-
\end{array
\right] \label{eq:Weff-matrix}
\end{equation
which accounts for the overall (coherent plus incoherent) transmission
through the system between different current channels. This effective
transmission matrix relates real currents on each site of the sample with
the voltages associated with each electron reservoir. It should by noticed
that sums over rows or columns, both on the original $\mathbb{T}$ and on
\tilde{\mathbb{T}},$ must be zero, in accordance to the Kirchhoff law.
At this point there is a particular situation that should be discussed: a
unique voltage difference between two channel sets. This results in a single
chemical potential difference. For example, assuming that all the channels
associated with a current source in the \textquotedblleft
left\textquotedblright\ source $L$ have the same chemical potential, $\delta
\mu _{L}$ and all those in the current sink $R,$ have $\delta \mu _{R}$. We
can rewrite the net current as:
\begin{eqnarray}
\mathtt{I} &=&\sum\limits_{i}I_{i}=\frac{e}{h}\sum\limits_{j}^{M_{R}}\su
\limits_{i}^{M_{L}}\tilde{T}_{Rj,Li}(\delta \mu _{L}-\delta \mu _{R})
\label{eq:Curr_2Volt} \\
&=&\mathrm{Tr}\left[ 4\boldsymbol{\Gamma }_{R}\mathbb{G}_{N1}^{R}\mathbf
\Gamma }_{L}\mathbb{G}_{1N}^{A}\right] (\delta \mu _{L}-\delta \mu _{R})
\label{Eq. Tr(ImGImG)} \\
&=&\mathtt{GV} \label{eq:GV}
\end{eqnarray
where $\mathtt{G}$ is the effective conductance, $\mathtt{V}=(\delta \mu
_{L}-\delta \mu _{R})/e$ is the applied voltage. Notice that $\Gamma _{L}$
and $\Gamma _{R}$ are square matrices with dimensions $M_{L}\times M_{L}$
and $M_{R}\times M_{R}$ associated with the $M=M_{L}+M_{R}$ quantum channels
at the leads $L$ and $R$. Since the final expression is the trace of a
matrix product, the result does not depend on the chosen basis.
For the most general case of several chemical potentials, Eqs. \re
{eq:Curr_2Volt}-\ref{eq:GV} can not be used and one should rely on Eqs. \re
{eq:Idec_matrix} and \ref{eq:Teff_matrix} that are the general solution to
the multi-terminal DP model. These are the main results of this work
together with the algorithms for the Green's functions, Eqs. \re
{eq:thoulessL} and \ref{eq:thoulessR}, and for the effective transmittances,
Eqs. \ref{Eq:Trans_order_0} and \ref{Eq:Trans_recursive}. All of them will
be tested in physically relevant situations in the next two sections.
\section{\label{sec:example1} Application: Decoherence in a Model for a SASE
}
The explicit description of vibrational degrees of freedom in a transport
problem requires a multichannel formulation even in a two probe
configuration. This is because one must resort to a Fock-space
representation of the Hamiltonian describing electrons and phonons. This
situations occur in vibrational spectroscopy\cite{Stipe98,Park},
polaronic models,\cite{BoncaTrugman,BoncaTrugman97} photon-assisted tunneling
\cite{Stafford96,Jauho98} as well as in time-dependent classical electromagnetic fields
in Floquet representations.\cite{LuisAPL11}
We will analyze a simple model that represents this family of problems:
independent electrons tunneling through a resonance where they are strongly
coupled to a quantized vibrational mode. In particular, we describe the
optical phonon-assisted tunneling in a double barrier device. It manifests
as a satellite peak in the I-V curve. This mechanism led to one \cit
{Foa2001} of the various proposals for a phonon laser (SASER).\cite{Kent2010}
In such proposal, a substantial part of the electrons contributing to the
current emit an optical phonon. This constitute the basis for a coherent
ultrasound source. \cite{ChemPhys2002,Camps2001}. The efficiency of the
device depends on the contrast between the satellite peak and the valley,
which in turn is determined by specific quantum interferences among the
participating channels. Thus, we will explore if these interferences survive
the decoherence induced by the acoustic phonons.
\begin{figure}[tbph]
\begin{center}
\includegraphics[width=2.8in]{fig3.eps}
\end{center}
\caption{Fock-space representation of states $|j,n\rangle $. The middle row
represents local electronic states $j$ with $n$ phonons. Lower and upper
rows describe the same electronic tight-binding chain but with different
numbers of phonons. Vertical lines are local electron-phonon couplings
restricted to site 0th. }
\label{Gr:FockSpace}
\end{figure}
\textit{Model.} Consider a \textquotedblleft local\textquotedblright\
electronic resonant state labeled as $0$. There, the electron is coupled to
a single vibrational mode, with frecuency $\omega _{0}$, whose occupation is
associated with the bosonic number operator $\hat{b}^{\dagger }\hat{b}$.
This is represented by the electron-phonon Hamiltonian
\begin{equation}
\hat{H}_{S}=E_{0}\hat{c}_{0}^{+}\hat{c}_{0}^{{}}+\left( \hbar \omega _{0}
\tfrac{1}{2}\right) \hat{b}^{+}\hat{b}^{{}}+V_{g}(\hat{b}^{+}+\hat{b}^{{}}
\hat{c}_{0}^{+}\hat{c}_{0}^{{}}.
\end{equation
The eigenstates of this Hamiltonian are the polaron states,\cit
{BrazJP,Wingreen1988} whose eigenenergies are
\begin{equation}
E_{0,n}=E_{0}+\hbar \omega _{0}\left( n+\frac{1}{2}\right) -\frac{|V_{g}|^{2
}{\hbar \omega _{0}}.
\end{equation
The electrons can jump \textit{in} and \textit{out} the resonant state to
the left and right leads. They can also suffer decoherent processes with a
rate $2\Gamma _{\phi }/\hbar $ in a FGR approximation. The effective
Hamiltonian results:
\begin{equation}
\hat{H}_{\mathrm{eff}}=\hat{H}_{S}+\hat{\Sigma}_{L}+\hat{\Sigma}_{R}+\hat
\Sigma}_{\phi },
\end{equation
where $\hat{\Sigma}_{L}$ and $\hat{\Sigma}_{R}$ describe the escape to the
current leads and $\hat{\Sigma}_{\phi }$ the escape associated with
decoherence. They are,
\begin{equation}
\hat{\Sigma}_{L}+\hat{\Sigma}_{R}+\hat{\Sigma}_{\phi }=\left[ \Sigma
_{L}(\varepsilon )+\Sigma _{R}(\varepsilon )-\mathrm{i}\Gamma _{\phi }\right]
\hat{c}_{0}^{+}\hat{c}_{0}^{{}}.
\end{equation
Notice that, these self-energies must account for the high voltage
difference required by SASER operation as an offset in the band centers of
the left and right leads $E_{L}-E_{R}=e\mathtt{V}$. We have omitted a real
part of the decoherent process which is not relevant in the present case. As
discussed before \cite{ChemPhys2002}, the optical phonon absorption and
emission can be viewed as a \textquotedblleft vertical\textquotedblright\
processes in a two-dimensional network. Thus, transport in the Fock space is
computationally equivalent to a tight-binding model with an expanded
dimensionality, as shown in Fig. \ref{Gr:FockSpace}.\cit
{ChemPhys2002,BrazJP,BoncaTrugman}
\begin{figure*}[tbph]
\begin{center}
\includegraphics[width=5.5in]{fig4.eps}
\end{center}
\caption{Multichannel decoherent transmission for the polaron model, with
\hbar \protect\omega _{0}=0.2$ eV, $E_{0}=-1.5$ eV. (a) Local electronic
state without coupling with the phonons ($V_{g}=0$); (b) Transmission
probability for an electron leaving the sample without a change in the
phonon state ($V_{g}=0.1$); (c) Transmission probability for an electron
that leaves the sample emitting one phonon ($V_{g}=0.1$); (d) Total
decoherent transmission probability}
\label{Gr:all}
\end{figure*}
When an electron comes from the left side, it arrives at the resonant site
where it couples to the $n_{0}$ phonons present in the well. It can either
keep its original kinetic energy $\varepsilon -\left( n+\tfrac{1}{2}\right)
\hbar \omega _{0}$ or change it by emitting or absorbing $\Delta n$ phonons.
Thus, the transmission probabilities of each contribution are given by:
\begin{equation}
T_{R(n_{0}+\Delta n),Ln_{0}}^{{}}=2\Gamma _{R(n_{0}+\Delta
n)}G_{n_{0}+\Delta n,n_{0}}^{R}2\Gamma _{Ln_{0}}G_{n_{0},n_{0}+\Delta n}^{A}.
\label{eq:T_eph}
\end{equation
Notice that the subscripts represent channels in the Fock space. As a
consequence of the trivial energy shift, associated with the presence of
phonons,
\begin{equation}
\Gamma _{\alpha n}(\varepsilon )=\Gamma \left( \varepsilon -E_{\alpha
}-\left( n+\tfrac{1}{2}\right) \hbar \omega _{0}\right) ,
\end{equation
for $\alpha =L,R,$ as defined in Eqs. \ref{Dyson}-\ref{self_energy}. Voltages are accounted
by $E_{\alpha }$. Each of this processes contributes to the total coherent
transmission which is given by,
\begin{equation}
T_{RL}(\varepsilon )=\sum\limits_{\Delta n=-n_{0}}^{\infty }T_{n_{0}+\Delta
n,n_{0}}(\varepsilon ).
\end{equation
In an actual device, the current would be obtained integrating $\varepsilon $
with the appropriate Fermi functions. Here, we might recall that Ref. \cit
{Emberly2000} suggested that in the Fock space, \textquotedblleft
vertical\textquotedblright\ hoppings could be blocked by the presence of
other electrons arriving with different initial energies. However, when the
kinetic energy of the incoming electrons satisfies $E_{F}\leq \hbar \omega
_{0}\leq e\mathtt{V}$, the applied voltage always enables phonon emission
\cite{ChemPhys2002,BrazJP,BoncaTrugman} ruling out the eventual problem of
overflow \cite{Cattena2012} ensuring the physical significance of our model.
The decoherence is induced by the finite lifetime for the polaron states
through an imaginary correction in the self-energies of Eq. \re
{sigma-decoher}. The available \textquotedblleft direct\textquotedblright\
channels are associated with the transmission probabilities of Eq. \re
{eq:T_eph}. Because of the wide band approximation for the dephasing
channels, the energy uncertainty is independent of $\varepsilon $
\begin{equation}
\Gamma _{\phi n}(\varepsilon )\equiv \Gamma _{\phi }.
\end{equation
Optical phonon emission or absorption processes give rise to decoherent
processes, even when $\Gamma _{\phi }=0$. This leaves us with several
possible dephasing channels, whose transmittances ar
\begin{equation}
T_{\beta (n_{0}+\Delta n),\alpha n_{0}}^{{}}=2\Gamma _{\beta \left(
n_{0}+\Delta n\right) }(\varepsilon )|G_{n_{0}+\Delta n,n_{0}}^{R}(\epsilon
)|^{2}2\Gamma _{\alpha n_{0}}(\varepsilon ). \label{eq:T_eph_dec}
\end{equation
Here $\alpha $,$\beta $ are either$~R,L$ or $\phi $. From these
transmissions, and using Eqs. \ref{eq:Teff_matrix} and \ref{eq:Weff-matrix},
we obtain the effective transmissions through the available real channels.
Instead of using the SASER operation regime ($n_{0}\gg 1$), for pedagogical
reasons we will assume that injected electrons find $n_{0}=0$ phonons, a
situation that describes a vibrational spectroscopy experiments. Then the
total transmission is simply,
\begin{equation}
\tilde{T}_{LR}(\varepsilon )=\sum\limits_{n=0}^{N}\tilde{T
_{LR}^{(n)}(\varepsilon ),
\end{equation
where each $\tilde{T}_{LR}^{(n)}$ includes the decimation of the incoherent
channels as in Eq. \ref{eq:Decim-T}. In what follows we will analyze $\tilde
T}_{LR}(\varepsilon )$ which is also the relevant quantity to study the
non-linear response (see Eq. 134 in Ref. \cite{Pastawski-Medina}).
The total transmission as function of energy is shown in Fig. \ref{Gr:all}.
The Hamiltonian parameters are roughly representative of a double-well
resonant tunneling devices where electron-phonon interactions manifest as a
satellite peak in the conductance. \cite{Foa2001} There $E_{0}=-1.5$ eV,
V_{R}=V_{L}=-0.1$ eV, $\hbar \omega _{0}=0.2$ eV and $V_{g}\simeq -0.1$ eV.
We discriminate among different vertical processes contributing to the total
transmittance. When the coupling between the local electronic state and the
phonon mode is neglected, $V_{g}=0$, the problem becomes one dimensional
with a unique resonance, as shown in Fig. \ref{Gr:all}-a. The effect of the
environment, accounted with the DP model, is a broadening of the original
resonance. When the local electronic state is strongly coupled with the
phonon field, $\left\vert V_{g}\right\vert \gg 0$, there are extra available
paths for the conduction electrons in the Fock space. Different electron
pathways in the coherent picture can interfere destructively, e.g. those
that traverse the resonance straight away and those that previously emit and
absorb a virtual phonon. These give rise to anti-resonances in Figs. \re
{Gr:all}-b and \ref{Gr:all}-c. Since they are a coherent phenomena, they may
be destroyed when decoherent events are present. This is made evident in
Fig. \ref{Gr:all}-d where the total electron transmission probability in a
multi-phonon process is compared with the same configuration with added
decoherence, according to the multi-terminal DP model.
The energy uncertainty used is $\Gamma _{\phi }=0.026$ $e$\texttt{V} $\sim
k_{B}T_{R}$, where $k_{B}$ is the Boltzmann constant and $T_{R}$ stands for
room temperature of $300K$. Although one might evaluate $\Gamma _{\phi }$
from the electronic energy uncertainties obtained with the help of ab-initio
computations, the behavior of $\tilde{T}$ as a function of $\Gamma _{\phi }$
is smooth, provided that these local uncertainties are small compared with
typical tunneling rates from the local resonances, $\Gamma _{L(R)}\gg \Gamma
_{\phi }$. Therefore, small variations of the precise value of $\Gamma
_{\phi }$ do not change the general behavior of $\tilde{T}$. This is
illustrated in Fig. \ref{Gr:colormap} where a color map shows how $\Gamma
_{\phi }$ affects the total transmission probability in the range [$0$ eV,
0.025$ eV].
\begin{figure}[tbph]
\begin{center}
\includegraphics[width=3.4in]{fig5.eps}
\end{center}
\caption{Multichannel decoherent transmission for the polaron model in a
color map. The transmission probability is shown in a color scale, as a
function of the incident electron Fermi energy and the strength of the
imaginary energy shift $\Gamma _{\protect\phi }$. The behavior of $\tilde{T}$
is shown to be a smooth function of $\Gamma _{\protect\phi }$.}
\label{Gr:colormap}
\end{figure}
We confirmed the general trend that decoherence broadens and lowers the
resonance peaks and raise the tails. But more importantly, valleys are
shaped by multi-phonon coherent processes that produce anti-resonances.
These resulted very sensitive to decoherence. Thus, these effects should be
considered in assessing the efficiency of a SASER.
\section{\label{sec:example2} Application: Quantum to Classical transition
in a Model for Giant Magnetoresistance.}
Spintronics often requires to distinguish how each spin projection
contribute to the current and to identify the spin dependent voltage
profiles, i.e. the chemical potentials $\delta \mu $'s. These are absent
from the original solution of the DP model that just provides the total
current, $I_{LR}=(e/h)T_{eff}\delta \mu $ (see section \ref{sec:DP} ). This
limitation was overcomed by the previous sections, where a specific current
I_{j}$, at spin-channel $j$, can be readily calculated from eq. \re
{eq:CurrentsMatrix}, as $I_{j}=e/h\sum_{i}\left( \mathbb{T}\right)
_{ji}\delta \mu _{i}$.
Spin-dependent electron transport in ferromagnetic metals presents high
rates of scattering events that could make a fully coherent treatment
somewhat unrealistic. The standard approach is to use the semiclassical
Boltzmann equation \cite{Valet-Fert}. However, in these models quantum
mechanic effects are completely neglected from the very beginning. These
effects can become important and interesting to study. For instance, ref.
\cite{FernAlcPast13} shows that spin-dependent transmittances in nanowires
with a modulated magnetic field may present Rabi oscillations. In these
situations, a Hamiltonian model capable of reaching a semiclassical limit,
such as the DP, can be very useful.
In this section, we use the multi-terminal DP model to treat one of the
paradigmatic phenomena of the spintronics, the Giant Magnetoresistance
(GMR). We will show that one can go from a purely quantum regime, described
by a Hamiltonian, to the (semi)classical limit of GMR, just by varying a
single parameter: the `decoherent' scattering rate.
Giant Magnetoresistance may occur in systems composed of two layers of a
ferromagnetic metal where their relative magnetization can be switched. In
these materials, the rate of scattering depends on the electron spin. Thus,
the electrical resistance depends on the relative orientation between the
spin and the layer's magnetization. If the two layers have their
magnetization aligned, there is a spin orientation with low resistance that
dominates transport. On the other hand, when the magnetizations are
anti-aligned, both spin channels have high resistance. \cite{Fert08}
\textit{Model}. Let us consider a one-dimensional system composed of two
adjacent `layers' or portions of a single-domain ferromagnetic metal. We
choose the relative magnetization in a anti-aligned configuration (Fig. \re
{fig:ChemPot}-$a$). This system is connected to non-magnetic contacts at
each side, labeled by $L$ and $R$. Thus, the current flows perpendicular to
the magnetic interface. As usual, each spin is regarded as an independent
channel at the contacts. Thus, at the leads, each spin projection is
characterized by the chemical potentials $\mu _{L\uparrow }$, $\mu
_{L\downarrow }$, $\mu _{R\uparrow }$, and $\mu _{R\downarrow }$. Since we
are considering non-ferromagnetic contacts, the chemical potentials at the
leads are spin independent.
Inside the system, the electrons undergo scattering processes producing the
spin dependent resistance. Since its fair to neglect Anderson localization,
we can use the equivalence between delta function impurities and local
decoherent scattering processes. As in the Ohmic limit of the DP model \cit
{Damato-Pastawski,GLBE1} they can be characterized by the parameter $\Gamma
_{\phi }$. This is related with the mean free time, $\tau _{\sigma }$,
through $\Gamma _{\sigma }=\hbar /(2\tau _{\sigma })$. Then, the Ohmic
conductance is proportional to the mean free path $\ell _{\sigma }$ which
results $\ell _{\sigma }=v_{F}\tau _{\sigma }$. Note that the rate $\Gamma
_{\sigma }$ depends on the relative orientation between the spin and the
local magnetization. Thus, spin $\uparrow $ has a scattering rate $\Gamma
_{\phi 1}$, at the first layer, and $\Gamma _{\phi 2}$, at the second one.
The opposite spin has the complementary rates.
As in previous works,\cite{Gopar,FernAlcPast13} the system's Hamiltonian
\hat{H}_{S}$ is described in a tight-binding approach that includes local
spin-reversing interactions:
\begin{eqnarray}
\hat{H}_{S} &=&{\sum\limits_{i=-N}^{N}}\sum\limits_{\sigma =\uparrow
,\downarrow }[E_{i,\sigma }^{{}}\hat{c}_{i,\sigma }^{\dag }\hat{c}_{i,\sigma
}^{{~}}+V\left[ \hat{c}_{i\sigma }^{\dagger }\hat{c}_{i+1\sigma }^{{~}}
\mathrm{c.c.}\right] \notag \\
&&+{\sum\limits_{i=-N}^{N}}V_{\downarrow \uparrow }\left[ \hat{c
_{i\downarrow }^{\dagger }\hat{c}_{i\uparrow }^{{~}}+\mathrm{c.c.}\right] .
\label{HamGMR}
\end{eqnarray
The label $i$ indicates sites on a lattice with unit cell $a$, $E_{i,\sigma
}^{{}}$ is the energy at the site $i$ with spin $\sigma $, the operator
\hat{c}_{i,\sigma }^{\dag }$ ($\hat{c}_{i,\sigma }^{{~}}$) creates
(annihilates) a particle at the site $i$ with spin $\sigma $. The firsts two
terms of $\hat{H}$ accounts for the site energies and the spin-conserving
hopping, $V$, between adjacent sites. $V$ is chosen as the unit of energy.
In a graphical representation, each spin orientation is represented by a
chain of sites interconnected by $V$. Thus, two chains of sites are needed
to represent the spin-dependent transport along this ferromagnetic system
(Fig. \ref{fig:ChemPot}-$a)$). The last term of $\hat{H}$, models the
scattering processes that may change the spin projection, such as scattering
with magnetic impurities. Thus, $V_{\downarrow \uparrow }$ is the local
spin-reversing or spin-mixing hopping parameter. This is related to a
characteristic length scale identified as the spin diffusion length, $L_{sd}
, by
\begin{equation}
L_{sd}=\frac{\hslash v_{F}}{2\left\vert V_{\downarrow \uparrow }\right\vert
,
\end{equation
where $v_{F}$ is the Fermi velocity and $L_{sd}$ is the length scale at
which the spin-flipping processes relax the diffusing spin. Thus, within
this length, both spin orientations can be considered as independent.
L_{sd} $ is typically much larger than the mean free path. When the electron
gets into the ferromagnetic material, it undergoes an \textit{exchange
interaction} that can be regarded as a Zeeman interaction. Thus, the site
energy is $E_{i,\uparrow (\downarrow )}=E_{0}\pm \Delta E_{Z}$, where $i$ is
a site of the first layer.
As in Eq. \ref{Hamil-efectivo}, the effective Hamiltonian incorporates the
leads and the scattering processes through the appropriate self-energies.
Now, $\hat{\Sigma}_{L(R)}=\hat{\Sigma}_{L(R)\uparrow }+\hat{\Sigma
_{L(R)\downarrow }$ is the self-energy operator describing the escape to the
left (right) lead, given by Eq. \ref{Dyson}, where all hoppings are equal to
$V$. Decoherent channels accounting for resistive scattering are associated
to each site and included into $\hat{H}$ through the $\hat{\Sigma}_{\phi }$
operator. Thus, $\hat{\Sigma}_{\phi }$ is diagonal in a matrix
representation. In the wide band limit, their elements are purely imaginary,
i.e. $\left( \hat{\Sigma}_{\phi }\right) _{ii}=-\mathrm{i}\Gamma _{\phi i}$.
\begin{figure}[tbh]
\begin{center}
\includegraphics[width=3.4in]{magnetoresistance.eps}
\end{center}
\caption{-$a)$ On top is a scheme showing the layer's magnetization in the
two resistor model for GMR. Below is a tight binding representation
discriminating the spin projection. The coherence lengths of electrons in
the first layer are $\ell _{1}$ and $\ell _{2}$ for up and down spin
electrons respectively. Note that coherence lengths are inverted in the next
layer. $\ell _{1}/\ell _{2}=1/2$ in all cases. Fig. $b)$ to $d)$ Site
dependent chemical potentials with $\ell _{1}=15~a$ in Fig. $b)$, $\ell
_{1}=1500~a$ in Fig. $c)$, and $\ell _{1}=150~a$ in Fig. $d)$. The system
length is $1000a$ and $V_{\downarrow \uparrow }=0$ ($L_{sd}\rightarrow
\infty $), and the chemical potentials at the leads are $\protect\mu _{L}=
\mathrm{V}$ and $\protect\mu _{R}=0$. The Fermi wavelengths at the left side
are $\protect\lambda _{F}=45a$, for up spins, and $\protect\lambda _{F}=30a
, for down spins. The opposite holds at the right ferromagnet. The chosen
parameters do not represent a specific experimental set up.}
\label{fig:ChemPot}
\end{figure}
\textit{Classical regime of GMR: two resistors model (TRM)}. Here, the
system length is much shorter than $L_{sd}$, i.e. $V_{\downarrow \uparrow
}\approx 0$ in Eq. \ref{HamGMR}. Here, when electrons enters into a
ferromagnetic layer they undergo an electrical resistance $\delta R=I_{LR
\mathrm{V}$ (Ohm's law) that manifest in a linear drop in the
chemical-potential $\delta \mu $. Therefore, in the anti-aligned
configuration, there are two linear potential drops of $\delta \mu $ with
slopes proportional to the spin-dependent resistance of each layer. Then, it
is expected a splitting of the chemicals potentials that form a diamond like
figure. This is precisely what we obtain using the multi-terminal DP method
with mean free paths shorter that the system size. Fig. \ref{fig:ChemPot}-$b
$ to \ref{fig:ChemPot}-$d$ show this, through the site-dependent chemical
potential. In contrast, for the quantum limit of long mean free paths,
quantum interferences are evident. However, they are smoothed out by
increasing the scattering rate until they reach the expected classical
diamond like figure (Fig. \ref{fig:ChemPot}-$b$).
\begin{figure}[tbh]
\begin{center}
\includegraphics[width=2.5in]{diamond.eps}
\end{center}
\caption{Upper figure, site-dependent chemical potential $\protect\delta
\protect\mu _{i}$ profile for the semiclassical model of GMR with finite
spin diffusion length, $L_{sd}=100~a$. Lower figure, shows the local
currents $I_{i}$ for up and down spin electrons. System size is $1000a$,
\ell _{1}/\ell _{2}=1/2$, and $\ell _{1}=15a$.}
\label{fig:diamond}
\end{figure}
\textit{Semiclassical regime of GMR: Valet and Fert theory.} Considering
finite values for the spin diffusion length, $L_{sd}$, Valet and Fert \cit
{Valet-Fert} showed that the difference of the spin-dependent local chemical
potentials decays exponentially with the distance to the magnetic interface
with a length scale given by $L_{sd}$. They also showed that the
spin-dependent current is inverted in this length scale. In Fig \re
{fig:diamond} we show that the multi-terminal DP model is also capable to
reproduce these behaviors provided that we turn on the spin flip term in Eq.
\ref{HamGMR}. In the upper figure we show the spin and site dependent
chemical potentials. One can see that in regions far from the interface,
distances larger than $L_{sd}$, the chemical potentials are nearly the same.
In regions close to the interface, the chemical potential drop forms a
diamond-like figure that show the expected spin-dependent exponential
contributions summed up to the trivial mean linear drop. In the lower figure
we can observe how the inversion of the currents is produced in the length
scale $L_{sd}$. For longer distances, the currents reach a stationary value.
All these behaviors are in agreement with Ref. \cite{Valet-Fert}. This
situation reinforces the descriptive conceptual value of the DP model and
the versatility of the numerical algorithms developed in this paper.
\section{\label{sec:conclusions}Conclusion}
In this work, we first reviewed the original two-terminal DP model, which
accounts for decoherent effects in quantum transport. Then, we presented an
extension of this model which is capable to deal with multi-terminal setups.
Also, we introduced recursive algorithms that allows us to take advantage of
the problem symmetries, in particular in the case of general banded
Hamiltonians. The incorporation of a unified notation gives more
transparency to its potentialities. Using the specific Hamiltonian models
for phonon laser and giant magnetoresistance, we exemplified how to treat
multi-channel problems in presence of decoherence.
We made special emphasis on the role of decimation procedures in the context
of banded effective Hamiltonians, since they can be used as the basis for
efficient computational schemes. In particular, one of the keys is given by
Eqs. \ref{eq:thoulessR}- \ref{eq:thoulessL}. Note that, in the very common
situation of block tridiagonal (i.e. banded) matrix Hamiltonians, these
recursive equations provide an efficient decimation procedure that allows
one to obtain all the $N\times (N-1)$ non-diagonal blocks of the whole
Green's function matrix, $\mathbb{G}$, in terms of the $N$ diagonal blocks.
In turn, these last can be calculated as matrix continued fractions. \cit
{MCF-Pastawski} The idea here is to take advantage of particular system's
symmetries using these expressions to build an efficient computation
approach for the problem under study.
Profiting from a parallelism between the computation of $\mathbb{G}$ and the
decoherent transmitance $\tilde{\mathbb{T}}$ already hinted by the DP
solution \cite{GLBE1}, we also derived a compact matrix equation for $\tilde
\mathbb{T}}$ in a generalized multi-terminal scheme. This recursive
algorithm relies on decimation procedures.
As a first application, we added decoherent processes to the usual model for
phonon-assisted tunneling in the configuration used for a phonon laser or
SASER. As is well known, \cite{Foa2001} in the I-V curve of a SASER
configuration, the contrast between the valley (out of resonance) and the
satellite peak (corresponding to phonon emission) is enhanced by the effect
of antiresonances. These last result from the interference between different
paths in the Fock's space. \cite{ChemPhys2002} Besides of the expected
smoothing out of the resonances because of the decoherence, we found that it
leads to the degradation of the contrast mainly from the suppression of the
antiresonances. This could set up new bounds for the efficiency of SASER
operation. \cite{Camps2001}
We also solved a simple multi-terminal DP model representative of the giant
magnetoresistance (GMR) phenomenon. There, each spin orientation is a
different conduction channel. Thus, the spin-dependent transport is
intrinsically multi-terminal. We essentially showed that the main
characteristics of the GMR can be well reproduced with this simple model.
While preserving a Hamiltonian description, it is able to reach the expected
classical and semiclassical regimes by means of a single parameter, the
local decoherent rate $\Gamma _{\phi i}$. What is more important, as in Fig.
\ref{fig:ChemPot}-$c$ and $d$, it opens the possibility to profit from
situations where quantum interference becomes relevant. \cit
{FernAlcPast13,Richter12}
With increasing system's size, molecular electronics suffers a paradigm
shift on its dominant transport mechanism, from \textquotedblleft coherent
tunneling\textquotedblright\ to \textquotedblleft incoherent
hopping\textquotedblright. Within this context, the present work should
result specially helpful in providing a computational bridge between these
limiting situations, while maintaining a general, transparent, and efficient
approach to quantum transport.
\section{Acknowledgments}
We acknowledge L. E. F. Foa Torres for his comments and stimulating
discussions at an early stage of this work. We received financial support
from ANPCyT, CONICET, MiNCyT-Cor, and SeCyT-UNC.
|
\section{Introduction}
If $\Gamma$ is a~f\/inite subgroup of the unitary group ${\rm{U}}(2)$ which acts freely on the unit sphere around the
origin in $\mathbb{C}^2$, it is natural to ask existence of a~K\"ahler metric def\/ined on the minimal resolution of the
quotient space $\mathbb{C}^2/\Gamma$, which has `small curvature' and which is asymptotically locally Euclidean (ALE) at
inf\/inity.
If $\Gamma$ is a~f\/inite subgroup of ${\rm{SU}}(2)$, Kronheimer~\cite{Kr89a} constructed ALE Ricci-f\/lat K\"ahler
metrics on the minimal resolution by means of so called the hyperK\"ahler quotient, and further showed~\cite{Kr89b} that
the metrics are determined by the the cohomology classes of a~collection of K\"ahler forms associated to the
hyper-K\"ahler structure.
When $\Gamma$ is a~f\/inite cyclic subgroup of ${\rm{U}}(2)$ generated by scalar matrices, $\Gamma$ is not included in
${\rm{SU}}(2)$ unless $|\Gamma|=2$, and the minimal resolution of $\mathbb{C}^2/\Gamma$ is simply the total space of the
line bundle $\mathscr O(-n)$, where $n=|\Gamma|$, and the unique negative section is the exceptional locus of the
resolution.
Because the section intersects positively with the canonical class if $n>2$, there exists no Ricci-f\/lat K\"ahler
metric on $\mathscr O(-n)$ if $n>2$.
LeBrun~\cite{LB88} constructed on this complex surface a~scalar-f\/lat K\"ahler (SFK) metric which is also ALE.
The metric is invariant under a~natural ${\rm{U}}(2)$-action, and may be considered to be the natural K\"ahler metric on
$\mathscr O(-n)$.
Later, Calderbank--Singer~\cite{CS2004} pointed out that, for any cyclic subgroup $\Gamma\subset {\rm{U}}(2)$, the
minimal resolution of $\mathbb{C}^2/\Gamma$ admits an ALE SFK metric.
All these metrics are anti-self-dual (ASD) with respect to the complex orientation.
ALE spaces can be compactif\/ied to be an orbifold by adding a~point at inf\/inity, and after an appropriate conformal
change the metric can be extended to the compactif\/ication as an ASD metric on the orbifold.
Small deformations of ASD conformal structures are governed by a~deformation complex, and if the space is compact, the
index of the complex is expressed in terms of topological invariants of the space.
Viaclovsky~\cite{V} computed the index of the deformation complex for various compact orbifolds in explicit form, and
show in particular that if the space is the compactif\/ication $\widehat{\mathscr O(-n)}$ of $\mathscr O(-n)$, the index
is $12-4n$.
As the obstruction for the deformation complex vanishes, this means that LeBrun's metric has a~non-trivial deformation
as an ALE ASD metric (if $n>3$).
In~\cite{HonCMP2} we computed the ${\rm{U}}(2)$-action on the relevant cohomology group in concrete form, and computed
the dimension of the moduli space of ALE ASD metrics near the LeBrun metric.
Also we found that there exists a~real 1-parameter family of deformation of the LeBrun metric which preserves not only
ASD-ALE property but also a~K\"ahler representative.
In this article we investigate all small deformations of the LeBrun metric on $\mathscr O(-n)$ {\textit{as an ALE SFK metric}}.
If the complex structure is f\/ixed, the following rigidity is shown:
\begin{Proposition}[=~Proposition~\ref{prop:rigid}]
\label{prop:rigid0}
When we fix the complex structure on $\mathscr O(-n)$, LeBrun's metric on $\mathscr O(-n) $ cannot be deformed as an
ALE SFK metric by small deformations.
\end{Proposition}
In order to explain what happens when we allow the complex structure on $\mathscr O(-n)$ to vary in deformations, we
recall that the line bundle $\mathscr O(-n)$ is included as a~special member of an $(n-1)$-parameters family of
af\/f\/ine $\mathbb{C}$-bundles over $\mathbb{CP}^1$, whose transition law for f\/iber coordinates $\zeta_0$ over
$U_0=\mathbb{C}(u)\subset\mathbb{CP}^1$ and $\zeta_1$ over $U_1=\mathbb{C}(1/u)\subset\mathbb{CP}^1$ is concretely given
by
\begin{gather}
\label{af00}
\zeta_0 = \frac 1{u^n} \zeta_1 + \sum\limits_{l=1}^{n-1} \frac {t_l} {u^l},
\qquad
(t_1,\dots,t_{n-1})\in \mathbb{C}^{n-1},
\end{gather}
where $t_l$-s are parameters.
If $(t_1,\dots,t_{n-1})=(0,\dots,0)$, this gives the line bundle $\mathscr O(-n)$,
but if $(t_1,\dots,t_{n-1})\neq(0,\dots,0)$, the af\/f\/ine bundle~\eqref{af00} has no global section and it is just an af\/f\/ine bundle.
For $t=(t_1,\dots,t_{n-1})\in\mathbb{C}^{n-1}$ we denote by $A_t$ for the af\/f\/ine $\mathbb{C}$-bundle over
$\mathbb{CP}^1$ def\/ined by~\eqref{af00}.
Then we prove the following
\begin{Theorem}
\label{thm:0}
There exists a~neighborhood $B\subset\mathbb{C}^{n-1}$ of the origin for which LeBrun's ALE SFK metric on $\mathscr
O(-n)$ extends naturally to $A_t$ if $t\in B$, as ALE SFK metrics.
\end{Theorem}
This will be shown as Theorem~\ref{thm:main01}, and from the proof, this family of metrics can be regarded as the versal
family for the LeBrun metric on $\mathscr O(-n)$ as ALE SFK metrics.
The 1-parameter family of ALE SFK metrics on the 4-manifold $\mathscr O(-n)$ obtained in~\cite{HonCMP2} is exactly the
restriction of the family of ALE SFK metrics in Theorem~\ref{thm:0} to the f\/irst (or the last) coordinate axis.
Next for explaining an immediate consequence of Theorem~\ref{thm:0}, we recall that any af\/f\/ine $\mathbb{C}$-bundle
over $\mathbb{CP}^1$ can be naturally compactif\/ied to a~Hirzebruch surface by attaching a~section at inf\/inity.
We call the negative of the last self-intersection number as \textit{degree} of the af\/f\/ine bundle.
Then for any $t\in\mathbb{C}^{n-1}$ the af\/f\/ine bundle def\/ined by the transition law~\eqref{af00} is of degree~$n$.
Conversely, if $n>1$, any af\/f\/ine $\mathbb{C}$-bundle over $\mathbb{CP}^1$ of degree $n$ is of the form $A_t$ for
some $t\in\mathbb{C}^{n-1}$.
Now because the equation~\eqref{af00} is linear in the variables $\zeta_0$, $\zeta_1$, $t_1, t_2,\dots,t_{n-1}$, the af\/f\/ine
bundle $A_t$ and $A_{ct}$ is isomorphic for any $c\in \mathbb{C}^*$.
Therefore, Theorem~\ref{thm:0} implies the following
\begin{Corollary}[=~Corollary~\ref{cor:exaf}]
Any affine $\mathbb{C}$-bundle over $\mathbb{CP}^1$ of negative degree $($in the above sense$)$ admits an ALE SFK metric.
\end{Corollary}
Finally we explain some property of the family of ALE SFK metrics obtained in Theorem~\ref{thm:0}.
In contrast with the LeBrun metric for which the rigidity holds as in Proposition~\ref{prop:rigid0}, for the deformed
metrics, we have the following
\begin{Proposition}\label{prop:var}
Let $B\subset\mathbb{C}^{n-1}$ be as in Theorem~{\rm \ref{thm:0}}, and for $t\in B$ let $g_t$ be the ALE SFK metric on~$A_t$.
Then if $t\neq 0$ and $t$ is sufficiently close to the origin, there exists a~smooth arc $\gamma_t \subset B$
passing through the point $t$ which satisfies the following:
\begin{itemize}\itemsep=0pt
\item[$(i)$] the complex structure of $A_t$ is constant along the arc $\gamma_t$,
\item[$(ii)$] the conformal class of the ALE SFK metric $g_t$ varies when $t$ moves along $\gamma_t$.
\end{itemize}
\end{Proposition}
\section{Preliminary computations for Hirzebruch surfaces}
\label{s:Hirz}
\subsection{Notation and convention}
\label{ss:notation}
For an integer $n\ge 0$, $\mathbb{F}_n$ denotes the Hirzebruch surface of degree $n$; namely $\mathbb{F}_n=\mathbb
P(\mathscr O(-n)\oplus \mathscr O)$ over $\mathbb{CP}^1$.
We write $\pi:\mathbb{F}_n\to\mathbb{CP}^1$ for the projection, and $f$ for a~f\/iber (class) of $\pi$.
We denote $\Gamma_0$ for $(-n)$-section of $\pi$, which is unique when $n>0$.
We have $H^2(\mathbb{F}_n,\mathbb{Z})\simeq {\rm{Pic}}\mathbb{F}_n\simeq \mathbb{Z}[\Gamma_0]\oplus\mathbb{Z}[f]$, and
$-K_{\mathbb{F}_n}\simeq \mathscr O(2\Gamma_0 + (n+2)f)$ for the anticanonical class.
${\rm{Aut}}_0\mathbb{F}_n$ denotes the identity component of holomorphic transformation group of $\mathbb{F}_n$, and
for a~section~$L$ of $\mathbb{F}_n\to\mathbb{CP}^1$, ${\rm{Aut}}_0(\mathbb{F}_n,L)$ denotes the subgroup of
${\rm{Aut}}_0\mathbb{F}_n$ consisting of transformations which keep~$L$ invariant.
If $n>0$ and $L=\Gamma_0$, we have ${\rm{Aut}}_0\mathbb{F}_n= {\rm{Aut}}_0(\mathbb{F}_n,L)$.
Two pairs $(\mathbb{F}_n,L)$ and $(\mathbb{F}_n,L')$ are called isomorphic as a~pair if there is a~biholomorphic map
$\phi:\mathbb{F}_n \to \mathbb{F}_n$ which satisf\/ies $\phi(L) = L'$.
$\Gamma_{\infty}$ means a~section whose self-intersection number is $(+n)$.
${\rm{Aut}}_0\mathbb{F}_n$ acts transitively on the space of $(+n)$-sections, and $\Gamma_{\infty}$ may be
identif\/ied with the section $\mathbb P(\mathscr O(-n))$.
Thus the complement $\mathbb{F}_n\backslash \Gamma_{\infty}$ can be identif\/ied with the total space of the line bundle~$\mathscr O(-n)$.
We write the linear system to which a~section of $\mathbb{F}_n\to\mathbb{CP}^1$ belongs in the form $|\Gamma_0+kf|$, $k\ge 0$.
It is well-known that this system has an irreducible member only when $k=0$ or $k\ge n$.
So if~$L$ is a~section with positive self-intersection number, we have~$L\in |\Gamma_0 + (n+l)f|$ for some $l\ge 0$.
The letter $l$ is always used in this meaning throughout the article.
We have $\Gamma_{\infty}\in |\Gamma_0+nf|$.
The system $|\Gamma_0 + (n+l)f|$ is very ample if and only if $l>0$.
Moreover we have $h^0(\mathscr O_{\mathbb{F}_n}(\Gamma_0+(n+l)f))= n+2l+2$, where $h^0$ means $\dim H^0$.
Therefore the complement of any member of these systems is realized in an af\/f\/ine space $\mathbb{C}^{n+2l+1}$.
We will also use the following result regarding the dimension of the cohomology group $H^i(\Theta_{\mathbb{F}_n})$ of
the tangent sheaf.
Namely if $n>0$ we have \cite{MKbook}
\begin{gather}
\label{cohohir}
h^0(\Theta_{\mathbb{F}_n}) = n + 5,
\qquad
h^1(\Theta_{\mathbb{F}_n}) = n - 1,
\qquad
h^2(\Theta_{\mathbb{F}_n}) = 0.
\end{gather}
This will also be shown in the proof of Proposition~\ref{prop:l0}.
\subsection[Af\/f\/ine $\mathbb{C}$-bundles over $\mathbb{CP}^1$]{Af\/f\/ine
$\boldsymbol{\mathbb{C}}$-bundles over $\boldsymbol{\mathbb{CP}^1}$}
\label{ss:ab}
Let ${\rm Af}(\mathbb{C})$ be the group of complex af\/f\/ine transformations of $\mathbb{C}$; namely those of the form
$\zeta\mapsto a\zeta + b$ for $\zeta\in \mathbb{C}$, where $a\in \mathbb{C}^*$ and $b\in \mathbb{C}$.
Let $X$ be a~projective algebraic manifold.
By an af\/f\/ine $\mathbb{C}$-bundle over~$X$, we mean (as usual) a~$\mathbb{C}$-bundle $A\to X$ whose structure group
is ${\rm {Af}}(\mathbb{C})$.
In this subsection, according to~\cite{At55}, we brief\/ly explain a~classif\/ication of af\/f\/ine $\mathbb{C}$-bundles
over~$X$, and then apply it to a~concrete description of af\/f\/ine $\mathbb{C}$-bundles over~$\mathbb{CP}^1$.
As in the case of any f\/iber bundle with prescribed structure group, isomorphic classes of af\/f\/ine
$\mathbb{C}$-bundles over $X$ are naturally in 1-1 correspondence with the cohomology set $H^1(X, \mathscr {A}f)$, where
$\mathscr {A}f$ means the sheaf of germs of holomorphic maps from open sets in~$X$ to the group~${\rm{Af}}(\mathbb{C})$.
The set $H^1(X, \mathscr {A}f)$ is of course the inductive limit of $H^1(\mathscr U,\mathscr Af)$ with respect to open
covering $\mathscr U$-s of~$X$.
For each $\mathscr U$, there is a~natural map
\begin{gather*}
\rho_{\mathscr U}: \ H^1(\mathscr U,{\mathscr A}f)
\longrightarrow
H^1(\mathscr U,\mathscr O^*),
\end{gather*}
which is induced from the natural homomorphism $ {\rm {Af}}(\mathbb{C})\to \mathbb{C}^*$ that takes the coef\/f\/icient
of the linear part.
These naturally induce a~map $\rho:H^1(X, \mathscr Af)\to H^1(X,\mathscr O^*)$.
Therefore we have
\begin{gather}
\label{star1}
H^1(X,\mathscr Af) \simeq \bigsqcup_{\xi \in H^1(X,\mathscr O^*)}\rho^{-1}(\xi).
\end{gather}
Geometrically, for an (isomorphism class of) af\/f\/ine bundle $A\in H^1(X,\mathscr Af)$, the image $\rho(A)\in
H^1(X,\mathscr O^*)$ is exactly (the isomorphism class of) the dual line bundle of the normal bundle of the section
$\overline A \backslash A$ in $\overline A$, where $\overline A$ means the compactif\/ied $\mathbb{CP}^1$-bundle which
is obtained from $A\to X$ by the standard inclusion ${\rm Af}(\mathbb{C})\subset\mathrm{PGL}(2,\mathbb{C})$.
Returning to the \v Cech cohomology group, analogously to~\eqref{star1}, we clearly have, for each open covering
$\mathscr U$ of $X$,
\begin{gather*
H^1 (\mathscr U,\mathscr Af) = \bigsqcup_{\xi \in H^1(\mathscr U,\mathscr O^*)}\rho_{\mathscr U}^{-1}(\xi).
\end{gather*}
In order to describe the set $\rho_{\mathscr U}^{-1}(\xi)$, we write $\mathscr U = \{U_i\}$, and let $\xi\in
H^1(\mathscr U,\mathscr O^*)$ be represented by a~1-cocycle $\{a_{ij}\}$, so that $a_{ij}\in \mathscr O^*(U_{ij})$ where
$U_{ij}=U_i\cap U_j$.
We f\/ix a~collection $\{h_i\}$ of meromorphic functions, where $h_i$ is def\/ined on $U_i$, that satisfy $h_i = a_{ij}
h_j$ on $U_{ij}$ (this is possible from the projectivity assumption for $X$), and let $D$ be the divisor def\/ined by
$\{h_i = 0\}$.
Though this is not necessarily ef\/fective, it is `linear equivalent' to the line bundle $\xi$.
Under these f\/ixing of $\{a_{ij}\}$ and $\{h_i\}$, let $\{(a'_{ij},b'_{ij})\}\in H^1(\mathscr U,\mathscr Af)$, where
$a'_{ij}\in\mathscr O^*(U_{ij})$ and $b'_{ij}\in \mathscr O(U_{ij})$, be a~representative of an element of
$\rho_{\mathscr U}^{-1}(\xi)$.
If we choose any $\phi_i\in \mathscr O^*(U_i)$ and $\psi_i\in \mathscr O(U_i)$ for each~$i$ and apply a~f\/iber
coordinate change $\tilde{\zeta}_i= \phi_i\zeta_i + \psi_i$ on $U_i$, then the new 1-cocycle $\{(\tilde a_{ij},\tilde
b_{ij})\}$ associated to $\{\tilde{\zeta}_i\}$, which is another representative of the same element of $H^1(\mathscr
U,\mathscr Af)$, is readily seen to be given by
\begin{gather}
\label{cd88}
\tilde a_{ij} = \frac{\phi_i}{\phi_j}a'_{ij},
\qquad
\tilde b_{ij} = \phi_i b'_{ij} + \psi_i - \frac{\phi_i}{\phi_j}\psi_j a'_{ij}.
\end{gather}
The f\/irst equation of these shows that any element of $\rho_{\mathscr U}^{-1} (\xi)$ can be represented by a~cocycle
of the form $\{(a_{ij},b'_{ij})\}$ (namely, by using the original representative $\{a_{ij}\}$ for $\xi\in H^1(\mathscr
U,\mathscr O^*)$), and in the following, for any element of $\rho_{\mathscr U}^{-1} (\xi)$, we only consider such
representatives.
This means that we only consider f\/iber coordinate changes $\{(\phi_i,\psi_i)\}$ which satisfy $\phi_i = \phi_j$ on
$U_{ij}$, and hence we can write $\phi_i = t$ for all~$i$ for some constant $t\in \mathbb{C}^*$.
Then the second equation of~\eqref{cd88} becomes (after replacing $a'_{ij}$ by $a_{ij}$)
\begin{gather}
\label{cc89}
\tilde b_{ij} = t b'_{ij} + \psi_i -\psi_j a_{ij}.
\end{gather}
This is the transformation law for representatives, under coordinate changes that satisfy the above constraint.
We are still f\/ixing $\mathscr U=\{U_i\}$, $\xi \in H^1(\mathscr U,\mathscr O^*)$, a~representative $\{a_{ij}\}$ of~$\xi$,
and $\{h_i\}$ that satisf\/ies $h_i=a_{ij}h_j$ on $U_{ij}$.
If $\{(a_{ij},b_{ij})\}$ is a~1-cocycle that represents an element of $\rho^{-1}_{\mathscr U}(\xi)$, we def\/ine
\begin{gather*}
c_{ij}:= \frac{b_{ij}}{h_i}
\qquad
\text{on}
\quad
U_{ij}.
\end{gather*}
Then from the cocycle condition for $\{(a_{ij},b_{ij})\}$, it follows that $\{c_{ij}\}$ is a~1-cocycle whose value is in
$\mathscr O(D)$, where $\mathscr O(D)$ is the sheaf of holomorphic functions $f$ for which $fh_i$ is holomorphic for any~$i$.
Moreover if we apply f\/iber coordinate changes of the form $\{(\phi_i,\psi_i)\}=\{(t,\psi_i)\}$, it follows readily
from~\eqref{cc89} that the new 1-cocycle $\{\tilde c_{ij} = \tilde b_{ij}/h_i\}$ is cohomologous to the 1-cocycle
$\{tc_{ij}\}$.
Thus the assignment $\{(a_{ij},b_{ij})\} \mapsto \{c_{ij}=b_{ij}/h_i\}$ induces a~map $\rho_{\mathscr U}^{-1}(\xi)\to
H^1(\mathscr U,\mathscr O(D))/\mathbb{C}^*$, where $\mathbb{C}^*$ acts on $H^1(\mathscr U,\mathscr O(D))$ as the scalar
multiplication.
Conversely the assignment $\{c_{ij}\}\mapsto \{(a_{ij},h_ic_{ij})\}$ induces a~map $H^1(\mathscr U,\mathscr O(D))\to
\rho_{\mathscr U}^{-1}(\xi)$, which descends (by~\eqref{cc89}) to a~map from $H^1(\mathscr U,\mathscr
O(D))/\mathbb{C}^*$.
The last map is clearly the inverse of the above map $\rho_{\mathscr U}^{-1}(\xi)\to H^1(\mathscr U,\mathscr O(D))$.
Thus, under f\/ixing $\{a_{ij}\}$ for $\xi\in H^1(\mathscr U,\mathscr O^*)$ and $\{h_i\}$ satisfying $h_i = a_{ij}h_j$,
we obtained a~bijection
\begin{gather}
\label{af2.5}
\rho_{\mathscr U}^{-1}(\xi)
\; \stackrel{\sim}{\longrightarrow}
\;
H^1(\mathscr U,\mathscr O(D))/\mathbb{C}^*.
\end{gather}
Here the quotient space $H^1(\mathscr U,\mathscr O(D))/\mathbb{C}^*$ is of course a~single point if $H^1(\mathscr
U,\mathscr O(D))=0$ and otherwise a~single point plus a~projective space.
The single point corresponds to the line bundle~$\xi$ itself.
In~\cite{At55} it was proved that this map is independent of the choice of $\{a_{ij}\}$ and $\{h_i\}$.
So by taking the inductive limit in~\eqref{af2.5} with respect to open coverings, we obtain a~bijection
\begin{gather*
\rho^{-1}(\xi)
\;\stackrel{\sim}{\longrightarrow}
\;
H^1(X,\mathscr O(D))/\mathbb{C}^*.
\end{gather*}
Thus from~\eqref{star1} there is a~natural 1-1 correspondence
\begin{gather}
\label{af3}
H^1(X,\mathscr Af)
\;\stackrel{\sim}{\longrightarrow}
\;
\bigsqcup_{\xi\in H^1(X,\mathscr O^*)} H^1(X, \mathscr O(\xi)) /\mathbb{C}^*.
\end{gather}
When $X=\mathbb{CP}^1$, by natural isomorphisms $H^1(X,\mathscr O^*) \simeq H^2(\mathbb{CP}^1,\mathbb{Z})\simeq
\mathbb{Z}$,~\eqref{af3} can be rewritten as
\begin{gather}
\label{af4}
H^1(\mathbb{CP}^1,\mathscr Af)
\;\stackrel{\sim}{\longrightarrow}
\;
\bigsqcup_{n\in\mathbb{Z}} H^1(\mathbb{CP}^1, \mathscr O(n)) /\mathbb{C}^*.
\end{gather}
\begin{Definition}\label{def:degree}
The {\em degree} of an af\/f\/ine bundle $A\to\mathbb{CP}^1$ is the image of~$A\in H^1(\mathbb{CP}^1,\mathscr Af)$ by
the composition of the natural map and identif\/ications
\begin{gather*}
H^1(\mathbb{CP}^1,\mathscr Af)
\;\stackrel{\rho}{\longrightarrow}
\;
H^1(\mathbb{CP}^1,\mathscr O^*)\simeq
H^2(\mathbb{CP}^1,\mathbb{Z})\simeq \mathbb{Z}.
\end{gather*}
\end{Definition}
Evidently, if an af\/f\/ine bundle is a~line bundle, its degree coincides with the usual degree as a~line bundle.
Denoting $\overline A$ for the $\mathbb{CP}^1$-bundle naturally associated to~$A$ as before, the degree of~$A$ is
exactly the negative of the self-intersection number of $\overline A\backslash A$ in $\overline A$.
If $n\ge -1$ we have $H^1(\mathbb{CP}^1,\mathscr O(n))=0$ and $\rho^{-1}(n)$ consists of a~single point which is exactly
the line bundle $\mathscr O(n)$.
Hence if the degree of a~line bundle is more than $-2$, it cannot be deformed even as an af\/f\/ine bundle.
So we are mainly interested in the case where the degree is less than $-1$.
We write such line bundles in the form $\mathscr O(-n)$, so that $n\ge 2$.
Then as $h^1(\mathscr O(-n)) = n-1>0$, $\rho^{-1}(-n)$ consists of the single point (which corresponds to the line
bundle $\mathscr O(-n)$) and the projective space $\mathbb{CP}^{n-2}$ (which is also a~point if $n=2$).
For each $n\ge 2$ we now compute transition law for f\/iber coordinates on arbitrary af\/f\/ine bundles with degree $-n$
in a~concrete form.
For this we take the standard covering $\mathscr U_0:=\{U_0,U_1\}$ where $U_0 =\{(z:w)\in\mathbb{CP}^1\,|\, z\neq 0\}$
and $U_1 =\{(z:w)\in\mathbb{CP}^1\,|\, w\neq 0\}$, and put $u = w/z$, $v=1/u$.
Then we have $H^1(\mathscr U_0,\mathscr O^*)\simeq H^1(\mathbb{CP}^1, \mathscr O^*)$, and so for $\xi=\mathscr O(-n)$,
as $\{a_{ij}\}$ and $\{h_i\}$ we can take
\begin{gather*}
a_{01} = \frac 1{u^n},
\qquad
h_0 =1
\qquad
(\text{so that}
\quad
a_{10} = u^n,
\quad
h_1 = {u^n}).
\end{gather*}
Moreover as a~basis of $H^1(\mathscr U_0,\mathscr O(D)) \!=\!H^1(\mathscr U_0,\mathscr O(-n))\!\simeq\! \mathbb{C}^{n-1}$ we can
take $\{u^{-1},u^{-2},\dots,u^{1-n}\}$ where $u^{-l}\in H^0(U_{01},\mathscr O(D))$.
Thus for each element $b_{01}=t_1u^{-1}+t_2u^{-2} + \dots + t_{n-1} u^{1-n}\in H^1(\mathscr U_0,\mathscr O(D))$ we can
associate an af\/f\/ine $\mathbb{C}$-bundle $A\to\mathbb{CP}^1$ whose transition law for f\/iber coordinates is given by
\begin{gather}
\label{af5}
\zeta_0 = \frac 1{u^n} \zeta_1 + \sum\limits_{l=1}^{n-1} \frac {t_l} {u^l}
\qquad
\text{on}
\quad
U_{01},
\end{gather}
where $\zeta_0$ and $\zeta_1$ are f\/iber coordinates over $U_0=\mathbb{C}(u)$ and $U_1=\mathbb{C}(v)$ respectively.
This can be regarded as def\/ining a~holomorphic family of af\/f\/ine $\mathbb{C}$-bundles over $\mathbb{CP}^1$
parametrized by $\mathbb{C}^{n-1}$, and over the origin we have the line bundle $\mathscr O(-n)$.
We write the total space of this family by~$\mathscr A_n$, thereby obtaining a~holomorphic map
\begin{gather}
\label{af6}
\mathscr A_n\to\mathbb{C}^{n-1}.
\end{gather}
The equation~\eqref{af5} is linear in the variables $\zeta_0$, $\zeta_1$, $t_1,t_2,\dots,t_{n-1}$.
Hence the total space of the family~\eqref{af6} admits a~$\mathbb{C}^*$-action which is the multiplication to all these
variables by weight~$1$.
This $\mathbb{C}^*$-action clearly descends to the scalar multiplication on the base space $\mathbb{C}^{n-1}$, and
f\/ibers of the family~\eqref{af6} are mutually isomorphic along orbits of this $\mathbb{C}^*$-action on
$\mathbb{C}^{n-1}$.
This explains geometrically why two cocycles $\{(a_{ij},b_{ij})\}$ and $\{(a_{ij},tb_{ij})\}$ with values in $\mathscr
Af$ determines mutually isomorphic af\/f\/ine bundles.
As above, for any $t\in\mathbb{C}^*$, the two cocycles $\{(a_{ij}, b_{ij})\}$ and $\{(a_{ij}, tb_{ij})\}$ determines the
same (or isomorphic, more precisely) af\/f\/ine bundles.
This can also be seen directly by noticing that the equation~\eqref{af5} is linear in the variables
$\zeta_0$, $\zeta_1$, $t_1,t_2,\dots,t_{n-1}$.
Hence by identifying f\/ibers of~\eqref{af6} lying over the same linear 1-dimensional subspace, we have obtained
a~family of af\/f\/ine $\mathbb{C}$-bundles over $\mathbb{CP}^1$ which is parametrized by $H^1(\mathbb{CP}^1,\mathscr
O(-n))/\mathbb{C}^*$.
By varying $n$ in $\mathbb{Z}$, this gives a~concrete realization of the bijection~\eqref{af4}.
Strictly speaking, in the argument of the last paragraph, we need to show that the natural map $H^1(\mathscr
U_0,\mathscr Af)\to H^1(\mathbb{CP}^1,\mathscr Af)$ is bijective; especially we need to show that any af\/f\/ine bundle
over $\mathbb{CP}^1$ can be trivialized over the open sets $U_0$ and $U_1$ respecting the structure of af\/f\/ine
bundle.
But this can be proved by standard adjusting argument using coboundaries, and we omit the detail.
\subsection[Af\/f\/ine $\mathbb{C}$-bundles over $\mathbb{CP}^1$ and Hirzebruch surfaces]{Af\/f\/ine
$\boldsymbol{\mathbb{C}}$-bundles over $\boldsymbol{\mathbb{CP}^1}$ and Hirzebruch surfaces}
\label{ss:AH}
We are concerned with ALE SFK metrics on the total spaces of af\/f\/ine $\mathbb{C}$-bundles over $\mathbb{CP}^1$ whose
degree is negative.
We will investigate this through the natural compactif\/ication of the af\/f\/ine bundles to $\mathbb{CP}^1$ bundles.
The latter are of course Hirzebruch surfaces.
In this subsection we will brief\/ly explain relationship between af\/f\/ine $\mathbb{C}$-bundles over $\mathbb{CP}^1$
and the Hirzebruch surfaces.
First let $A\to \mathbb{CP}^1$ be an af\/f\/ine $\mathbb{C}$-bundle, and let $\overline A\to \mathbb{CP}^1$ be the
natural compactif\/ication to a~$\mathbb{CP}^1$-bundle induced by the inclusion ${\rm Af}(\mathbb{C})\subset
\mathrm{PGL}(2,\mathbb{C})$ as before.
We have $\overline A\simeq \mathbb{F}_n$ for some $n\ge 0$.
We write $L:= \overline A\backslash A$ for the added locus, which is of course a~section of the projection $\overline A\to \mathbb{CP}^1$.
Then $-L^2$ is exactly the degree of $A\to \mathbb{CP}^1$.
Hence any af\/f\/ine $\mathbb{C}$-bundle over $\mathbb{CP}^1$ of negative degree is naturally identif\/ied with the
complement of a~section of some $\mathbb{F}_n$ whose self-intersection number is positive.
We write by
\begin{gather}
\label{Kf}
\mathscr F_n\to \mathbb{C}^{n-1}
\end{gather}
for the family of Hirzebruch surfaces that is obtained as the simultaneous compactif\/ication for members of the family
$\mathscr A_n\to \mathbb{C}^{n-1}$ in~\eqref{af6}.
For this family, it is well-known that the Kodaira--Spencer map
\begin{gather}
\label{ks001}
T_0\mathbb{C}^{n-1}\to H^1(\mathbb{F}_n,\Theta)
\end{gather}
at the origin is isomorphic (see~\cite[pp.~309--312]{K}), and the family~\eqref{Kf} gives the Kuranishi family of the
Hirzebruch surface $\mathbb{F}_n$.
Thus the parameter space $\mathbb{C}^{n-1}$ of $\mathscr A_n\to\mathbb{C}^{n-1}$ and $\mathscr F_n\to\mathbb{C}^{n-1}$
may also be naturally identif\/ied with~$H^1(\Theta_{\mathbb{F}_n})$.
Although the transition law for each member of the family~\eqref{Kf} is concretely given as in~\eqref{af5}, it is not
easy to identify them with $\mathbb{F}_m$ for a~precise value of $m$.
This was intensively studied in~\cite[p.~143, Theorem]{Suwa}, where an explicit answer was given, but it is too
complicated to write the result here.
Some exceptions are identif\/ication for f\/ibers on the coordinate axes of~$\mathbb{C}^{n-1}$.
Namely letting $\mathbb{C}(t_l)$ to be the $l$-th coordinate axis of~$\mathbb{C}^{n-1}$, if we introduce new f\/iber
coordinates $\tilde{\zeta}_0$ and $\tilde{\zeta}_1$ by
\begin{gather}
\label{zeta'}
\tilde{\zeta}_0 = \frac{u^l\zeta_0-t_l}{t_l\zeta_0}
\qquad
\text{and}
\qquad
\tilde{\zeta}_1= \frac{\zeta_1}{t_lv^{n-l}\zeta_1 + t_l^2}
\end{gather}
on the open sets $U_0=\mathbb{C}(u)$ and $U_1=\mathbb{C}(v)$ respectively, then with the aid of~\eqref{af5}, we readily
obtain the relation $\tilde{\zeta}_0 = v^{n-2l}\tilde{\zeta}_1$, which means that the ruled surface over the axis
$\mathbb{C}(t_l)$ is isomorphic to $\mathbb{F}_{n-2l}$, except the central f\/iber.
Here we are allowing the case $n-2l<0$ and in that case $\mathbb{F}_{n-2l}$ means $\mathbb{F}_{2l-n}$.
We also note that, as a~natural extension of the $\mathbb{C}^*$-action on $\mathscr A_n$, the total space of $\mathscr
F_n\to\mathbb{C}^{n-1}$ has a~$\mathbb{C}^*$-action, and it also descends to the scalar multiplication on $\mathbb{C}^{n-1}$.
Conversely if~$L$ is a~section of $\pi:\mathbb{F}_m\to\mathbb{CP}^1$ for some $m\ge 0$, then the complement
$\mathbb{F}_m\backslash L$ is biholomorphic to an af\/f\/ine $\mathbb{C}$-bundle over $\mathbb{CP}^1$.
This can be seen in the following way.
Let~$\Gamma_0$ and~$\Gamma_{\infty}$ be sections satisfying $\Gamma_0^2 = -m$ and $\Gamma_{\infty}^2 = m$.
If $m=0$, we assume $\Gamma_0\neq \Gamma_{\infty}$.
Let $\mathscr U=\{U_i\}$ be an open covering of $\mathbb{CP}^1$ which satisf\/ies for any~$i$ at least one of $L\cap
\Gamma_0\cap \pi^{-1} (U_i)=\varnothing$ or $L\cap \Gamma_{\infty}\cap \pi^{-1} (U_i)=\varnothing$ holds for any~$i$.
Let $\zeta_i$ be any f\/iber coordinate over $U_i$ of the line bundle $\mathscr O(-m)\subset\mathbb{F}_m$ (so that
$\Gamma_0\cap\pi^{-1} (U_i)$ and $\Gamma_{\infty}\cap\pi^{-1} (U_i)$ are def\/ined by $\zeta_i=0$ and $\zeta_i=\infty$
respectively), and $f_i$ be a~meromorphic function on $U_i$ such that $L\cap\pi^{-1} (U_i)$ is def\/ined by $\zeta_i =
f_i$.
From the choice, $f_i$ does not have both a~zero and a~pole.
Then for any~$i$ such that $f_i$ does not have a~pole, we def\/ine a~new f\/iber coordinate $\tilde{\zeta}_i$ over $U_i$
as an af\/f\/ine bundle by setting
\begin{gather}
\label{af8}
\tilde{\zeta}_i = \frac 1{\zeta_i-f_i}.
\end{gather}
Then from the choice of~$i$, this may be used as a~f\/iber coordinate on $\mathbb{F}_m\to\mathbb{CP}^1$, and we have
$L\cap \pi^{-1}(U_i) = \{\tilde{\zeta}_i = \infty\}$.
For the remaining~$i$-s, $f_i$ does not have a~zero.
$L\cap \Gamma_0 \cap \pi^{-1} (U_i)=\varnothing$.
For these~$i$-s we put
\begin{gather}
\label{af9}
\tilde{\zeta}_i = \frac{f_i\zeta_i}{f_i-\zeta_i}.
\end{gather}
Then this can also be used as a~f\/iber coordinate over $U_i$, and we again have $L\cap \pi^{-1}(U_i) =
\{\tilde{\zeta}_i = \infty\}$.
From~\eqref{af8} and~\eqref{af9} we readily see that the transition law for the new coordinate system
$\{\tilde{\zeta}_i\}$ is included in the af\/f\/ine group ${\rm Af}(\mathbb{C})$.
Therefore $\mathbb{F}_m\backslash L$ is actually an af\/f\/ine bundle.
However, even if the equation for a~section~$L$ is given in a~concrete form, it is not immediate again to trivialize the
af\/f\/ine bundle $\mathbb{F}_m\backslash L\to\mathbb{CP}^1$ over $U_0=\mathbb{C}(u)$ and $U_1=\mathbb{C}(v)$ and write
down the transition function in the form~\eqref{af5}.
\subsection{Computations for Hirzebruch surfaces}
As we mentioned we will investigate ALE SFK metrics on the af\/f\/ine $\mathbb{C}$-bundles over $\mathbb{CP}^1$ through
the compactif\/ication to Hirzebruch surfaces.
More precisely the Hirzebruch surfaces are included in the twistor spaces of a~conformal compactif\/ication of the ALE
SFK metrics on the af\/f\/ine bundles, and the added section will be the twistor line over the added point at
inf\/inity, whose self-intersection number in the surface is positive.
For this purpose, in this subsection, we make computations for pairs $(\mathbb{F}_n,L)$ where~$L$ is a~section
satisfying $L^2>0$.
(So $\mathbb{F}_n\backslash L$ is an af\/f\/ine $\mathbb{C}$-bundle over $\mathbb{CP}^1$ of negative degree.) Especially
we compute the dimension $h^i(\Theta_{\mathbb{F}_n,L})$ for arbitrary pairs, where $\Theta_{\mathbb{F}_n,L}$ is the
sheaf of germs of holomorphic vector f\/ields on $\mathbb{F}_n$ which are tangent to~$L$.
If~$L$ is a~section of $\mathbb{F}_n\to\mathbb{CP}^1$ which satisf\/ies $L^2>0$, we have $L\in |\Gamma_0 + (n+l)f|$ for
some $l\ge 0$ (see Section~\ref{ss:notation} for notation). Of course the value of $h^i(\Theta_{\mathbb{F}_n,L})$
depends on the number $l$.
We begin with the case $l=0$.
In this case we have $n>0$ as we are supposing $L^2>0$.
Moreover if we identify $\mathbb{F}_n$ with $\overline A$ where $A=\mathscr O(-n)$, then~$L$ can be identif\/ied with
$\overline A\backslash A$, the section at inf\/inity.
Hence regarding $\mathscr O(-n)$ (or $\mathbb{F}_n$) as the minimal resolution of $\mathbb{C}^2/\mathbb{Z}_n$ (or
$\mathbb{CP}^2/\mathbb{Z}_n$) where $\mathbb{Z}_n\subset{\rm{GL}}(2,\mathbb{C})$ is a~cyclic subgroup of scalar matrices
of order $n$, the pair $(\mathbb{F}_n,L)$ has an ef\/fective action of ${\rm{GL}}(2,\mathbb{C})/\mathbb{Z}_n$.
In particular we have $h^0(\Theta_{\mathbb{F}_n,L}) \ge 4$.
\begin{Proposition}
\label{prop:l0}
Suppose $n>0$ and let~$L$ be any $(+n)$-section of $\mathbb{F}_n\to\mathbb{CP}^1$.
Then we have the following:
\begin{enumerate}\itemsep=0pt
\item[$(i)$] $h^0(\Theta_{\mathbb{F}_n, L})=4$ and $H^2(\Theta_{\mathbb{F}_n, L})=0$.
\item[$(ii)$] The natural homomorphism $H^1(\Theta_{\mathbb{F}_n, L}) \to H^1(\Theta_{\mathbb{F}_n})$ is isomorphic,
and these are $(n-1)$-dimensional vector spaces.
\item[$(iii)$] The complex structure of the pair $(\mathbb{F}_n,L)$ is independent of the choice of~$L$.
\item[$(iv)$] We have ${\rm{Aut}}_0(\mathbb{F}_n,L)\simeq {\rm{GL}}(2,\mathbb{C})/\mathbb{Z}_n$ $($see above$)$.
\end{enumerate}
\end{Proposition}
\begin{proof}
Let $\Theta_{\mathbb{F}_n/\mathbb{CP}^1}\subset\Theta_{\mathbb{F}_n}$ be the subsheaf consisting of germs of holomorphic
vector f\/ields which are tangent to f\/ibers of $\pi:\mathbb{F}_n\to\mathbb{CP}^1$.
Scalar matrices in ${\rm{GL}}(2,\mathbb{C})$ induce a~$\mathbb{C}^*$-action on $\mathbb{F}_n$ which preserves each
f\/iber of $\mathbb{F}_n\to\mathbb{CP}^1$, and it def\/ines a~vector f\/ield which is tangent to each f\/iber of $\pi$.
Hence we obtain a~section of $\Theta_{\mathbb{F}_n/\mathbb{CP}^1}$.
Moreover, as the vector f\/ield has simple zeros on $\Gamma_0\sqcup L$ and no other zeros, we obtain
$\Theta_{\mathbb{F}_n/\mathbb{CP}^1}\simeq \mathscr O_{\mathbb{F}_n}(\Gamma_0+ L )$.
With the aid of this isomorphism, we have the following commutative diagram of exact sequences of sheaves on
$\mathbb{F}_n$:
\begin{gather*}
\begin{CD}
& & 0 & & 0 & & 0 & &
\\
& & @VVV @VVV @VVV
\\
0 @>>> \mathscr O_{\mathbb{F}_n} (\Gamma_0) @>>> \Theta_{\mathbb{F}_n/\mathbb{CP}^1} @>>> N_{ L /\mathbb{F}_n} @>>> 0
\\
& & @VVV @VVV @|
\\
0 @>>> \Theta_{\mathbb{F}_n, L } @>>> \Theta_{\mathbb{F}_n} @>>> N_{ L /\mathbb{F}_n} @>>> 0
\\
& & @VVV @VVV @VVV
\\
0 @>>> \pi^*\Theta_{\mathbb{CP}^1} @= \pi^*\Theta_{\mathbb{CP}^1} @>>> 0
\\
& & @VVV @VVV
\\
&& 0 && 0
\end{CD}
\end{gather*}
Since the group ${\rm{GL}}(2,\mathbb{C})$ acts transitively on $\mathbb{CP}^1$, the natural map
$H^0(\Theta_{\mathbb{F}_n,L})\to H^0(\pi^*\Theta_{\mathbb{CP}^1})$ is surjective.
Hence from the f\/irst column of the diagram, since $h^0(\mathscr O_{\mathbb{F}_n}(\Gamma_0)) = 1$ as $n>0$ and
$h^0(\Theta_{\mathbb{CP}^1}) = 3$, we obtain $h^0(\Theta_{\mathbb{F}_n,L}) = 1 + 3 = 4$.
Also from the same column, as we readily have $h^1(\mathscr O_{\mathbb{F}_n}(\Gamma_0))=n-1$ and $h^2(\mathscr
O_{\mathbb{F}_n}(\Gamma_0))=0$, we obtain $h^1(\Theta_{\mathbb{F}_n,L}) = n-1$ and $h^2(\Theta_{\mathbb{F}_n,L}) = 0$.
From the isomorphism $\Theta_{\mathbb{F}_n/\mathbb{CP}^1}\simeq \mathscr O_{\mathbb{F}_n}(\Gamma_0+ L )$ we also obtain
$h^0(\Theta_{\mathbb{F}_n/\mathbb{CP}^1}) = h^0(\mathscr O_{\mathbb{F}_n}(\Gamma_0+ L )) = n + 2$.
Hence as $h^0(N_{ L /\mathbb{F}_n}) =h^0(\mathscr O_{\mathbb{CP}^1}(n)) = n+1$ and $h^0(\mathscr
O_{\mathbb{F}_n}(\Gamma_0)) = 1$, from the f\/irst row, we obtain that the map $H^0( \Theta_{\mathbb{F}_n/\mathbb{CP}^1}
) \to H^0(N_{ L /\mathbb{F}_n})$ is surjective.
Hence from the commutative diagram the map $H^0( \Theta_{\mathbb{F}_n} ) \to H^0(N_{ L /\mathbb{F}_n})$ is also
surjective.
Therefore from the middle row, as $h^0(\Theta_{\mathbb{F}_n,L}) = 4 $, we obtain $h^0(\Theta_{\mathbb{F}_n}) = 4 + (n+1)
= n+5$.
Moreover, as $h^i(N_{L/\mathbb{F}_n}) = h^i(\mathscr O(n))=0$ for $i\in\{1,2\}$ and $h^2(\Theta_{\mathbb{F}_n,L}) = 0$,
we obtain from the same row that the natural map $H^1(\Theta_{\mathbb{F}_n,L})\to H^1(\Theta_{\mathbb{F}_n})$ is
isomorphic and $h^2(\Theta_{\mathbb{F}_n})=0$.
The assertion (iii)
is clear since $\mathbb{F}_n\backslash L$ is isomorphic to the line bundle $\mathscr O(-n)$ for any
$(+n)$-section~$L$.
(iv) follows from the remark preceding to Proposition~\ref{prop:l0} and the assertion (i) which is already shown.
\end{proof}
Next we consider the case $l=1$, which requires some care.
\begin{Proposition}\label{prop:l1}
Suppose $n>0$.
If~$L$ is a~section of $\pi:\mathbb{F}_n\to\mathbb{CP}^1$ belonging to the system $|\Gamma_0 + (n+1)f|$, we have the following:
\begin{enumerate}\itemsep=0pt
\item[$(i)$] $ h^0(\Theta_{\mathbb{F}_n,L})=2$, $h^1(\Theta_{\mathbb{F}_n,L})= n-1$, and $h^2(\Theta_{\mathbb{F}_n,L})= 0$.
\item[$(ii)$] The natural map $H^1(\Theta_{\mathbb{F}_n,L})\to H^1(\Theta_{\mathbb{F}_n})$ is isomorphic.
\item[$(iii)$] The complex structure of the pair $(\mathbb{F}_n,L)$ is independent of the choice of the section~$L$.
\end{enumerate}
\end{Proposition}
\begin{proof}
Let $L\in |\Gamma_0 +(n+1)f|$ be a~section as in the proposition.
Then we readily have
\begin{gather}
\label{inter1n1}
(L,\Gamma_0)=1,
\qquad
(L,\Gamma_0+nf) = n+1.
\end{gather}
Note that we have not specif\/ied a~$(+n)$-section $\Gamma_{\infty}$ yet.
We write $p=L\cap\Gamma_0$ and let $q\in L$ be any point which is dif\/ferent from $p$.
Then as $\dim |\Gamma_0+nf| = n+1$, by dimension counting, there exists a~section $\Gamma_{\infty} \in |\Gamma_0 + nf|$
which touches~$L$ at the point $q$ by multiplicity $(n+1)$.
Let $T_{\mathbb{C}}\subset{\rm{Aut}}\mathbb{F}_n$ be the maximal torus which is determined by the property that it
preserves the two sections $\Gamma_0$, $\Gamma_{\infty}$ and f\/ixes the two points $p$, $q$.
The complement $\mathbb{F}_n\backslash \Gamma_{\infty}$ may be identif\/ied with the line bundle $\mathscr O(-n)$.
Let $u$ be an af\/f\/ine coordinate on $\mathbb{CP}^1\backslash \pi(q)$ (where $\pi$ is the projection
$\mathbb{F}_n\to\mathbb{CP}^1$ as before), and $\zeta$ a~f\/iber coordinate of the line bundle $\mathscr O(-n)$ over
$U_0=\mathbb{C}(u)$, so that
\begin{gather*}
p=(0,0),
\qquad
q=(\infty,\infty),
\qquad
\Gamma_{0} = \{\zeta=0\},
\qquad
\text{and}
\qquad
\Gamma_{\infty} = \{\zeta=\infty\}.
\end{gather*}
Then as~$L$ intersects $\Gamma_0$ transversally at $p$ by~\eqref{inter1n1}, a~def\/ining equation for~$L$ has to be of
the form, in the above coordinates,
\begin{gather*}
\zeta = uh(u),
\qquad
h(0)\neq 0,
\end{gather*}
where $h=h(u)$ is a~holomorphic function on $U_0 = \mathbb{C}(u)$.
In the coordinates $(v,\eta^{-1}):= (u^{-1},u^{-n}\zeta^{-1})$ around the point $q=(\infty,\infty)$, this can be
rewritten as $ \eta^{-1} = v^{n+1}/h(v^{-1})$.
Then since~$L$ touches $\Gamma_{\infty}$ at the point $q$ by multiplicity $(n+1)$, the function $h(v^{-1})$ cannot have
a~pole at $v=0$.
This means that $h(u)$ is a~constant.
Hence~$L$ is def\/ined by the equation $\zeta = cu$ for some $c\in \mathbb{C}^*$.
But we may assume $c=1$ by changing the f\/iber coordinate $\zeta$ to $c^{-1}\zeta$.
Thus~$L$ is def\/ined by $\zeta=u$ in the coordinates $(u,\zeta)$.
In particular this proves the assertion (iii).
Next in order to determine $h^0(\Theta_{\mathbb{F}_n,L})$, we recall that, in terms of the coordinates $(u,\zeta)$, any
vector f\/ield $\theta\in H^0(\Theta_{\mathbb{F}_n})$ is concretely written as (see~\cite[pp. 43--44]{MKbook})
\begin{gather}
\label{MK1}
\theta= g(u) \frac{\partial}{\partial u} + \left( f(u) \zeta^ 2 + c\zeta u \right) \frac{\partial}{\partial \zeta},
\end{gather}
where $g(u) = a_1 u^2 + a_2 z + a_3$
$(a_i\in\mathbb{C})$,
$f(u) = b_1 u^n + b_2 u^{n-1} + \dots + b_{n+1}$
$(b_i\in\mathbb{C})$ and $c\in \mathbb{C}$.
(So we have $3 + (n+1) + 1 = n + 5$ parameters in total, which agrees with $h^0(\Theta_{\mathbb{F}_n}) = n+5$.)
For later use, we let $l\ge 1$ and let the section~$L$ to be def\/ined by $\zeta = u^l$, and def\/ine $F(u,\zeta):=
\zeta - u^l$, so that $F$ is a~def\/ining equation of~$L$.
Then $\theta\in H^0(\Theta_{\mathbb{F}_n,L})$ if\/f the derivation $\theta F$ satisf\/ies $\theta F|_L = 0$.
By~\eqref{MK1} we have
\begin{gather*}
\theta F = f\zeta^2 + c u \zeta - lgu^{l-1}.
\end{gather*}
Hence by substituting $\zeta = u^l$, the restriction becomes
\begin{gather}
\theta F |_L = \big(b_1 u^n + b_2 u^{n-1} + \dots + b_{n+1}\big) u^{2l} + c u^{l+1} - l \big(a_1 u^2 + a_2 u + a_3\big) u^{l-1}
\nonumber
\\
\phantom{\theta F |_L}{}
= \big(b_1 u^{2l+n} + b_2 u^{2l+n-1} + \dots + b_{n+1}u^{2l}\big) + \big\{(c-la_1) u^{l+1} - la_2 u^l - la_3 u^{l-1}\big\}.
\label{der1}
\end{gather}
When $l=1$, we have $2l = l+1$, and we obtain
\begin{gather*}
\theta F|_L = \big(b_1 u^{n+2} + b_2 u^{n+1} + \dots + b_{n+1}u^{2}\big) + \big\{(c-a_1) u^{2} - a_2 u - a_3 \big\}
\\
\phantom{\theta F|_L}{}
= b_1 u^{n+2} + b_2 u^{n+1} + \dots + b_n u^3 + \{b_{n+1} + (c -a_1)\} u^2 - a_2 u - a_3.
\end{gather*}
Thus $\theta F|_L = 0$ if\/f{\samepage
\begin{gather*}
b_1=b_2=\dots=b_{n} = 0,
\qquad
b_{n+1} + c - a_1 = a_2 = a_3 = 0.
\end{gather*}
These imply $h^0(\Theta_{\mathbb{F}_n, L}) = 2$.}
It remains to compute $h^i(\Theta_{\mathbb{F}_n, L})$ for $i\in \{1,2\}$ and show the isomorphicity in~(ii).
But these follow readily from~\eqref{cohohir}, the standard exact sequence $ 0 \to \Theta_{\mathbb{F}_n, L} \to
\Theta_{\mathbb{F}_n} \to N_{L/\mathbb{F}_n}\to 0$ and $h^0(\Theta_{\mathbb{F}_n, L})=2$.
The assertion (iii) is already shown.
\end{proof}
By the proposition, the group ${\rm{Aut}}_0(\mathbb{F}_n,L)$ is 2-dimensional when $l=1$.
This group can also be readily determined in a~concrete form.
For this, as before let $p$ be the intersection point of~$L$ and $\Gamma_0$.
(By~\eqref{inter1n1}~$L$ and $\Gamma_0$ intersect transversally at a~unique point.)
\begin{Proposition}\label{prop:nonred}
Suppose $n>0$ and let~$L$ be any section of $\pi:\mathbb{F}_n\to \mathbb{CP}^1$ belonging to the system $|\Gamma_0+(n+1)f|$.
Then the $2$-dimensional group ${\rm{Aut}}_0(\mathbb{F}_n, L)$ can be naturally identified with
the group $\{g\in\mathrm{PGL}(2,\mathbb{C}) \,|\, g(\pi(p)) = \pi(p)\}$, which is isomorphic
to the affine transformation group ${\rm Af}(\mathbb{C})$.
\end{Proposition}
\begin{proof}
The kernel sheaf of the restriction of the natural surjection $\Theta_{\mathbb{F}_n}\to \pi^*\Theta_{\mathbb{CP}^1}$ to
the subsheaf~$\Theta_{\mathbb{F}_n,L}$ can be obtained in a~similar way to the f\/irst column of the commutative diagram
in the proof of Proposition~\ref{prop:l0}, and consequently we obtain the exact sequence
\begin{gather}
\label{ntses}
0
\longrightarrow
\mathscr O_{\mathbb{F}_n} (\Gamma_0 - f)
\longrightarrow
\Theta_{\mathbb{F}_n,L}
\longrightarrow
\pi^*\Theta_{\mathbb{CP}^1}
\longrightarrow
0.
\end{gather}
Clearly we have $H^0 (\mathscr O_{\mathbb{F}_n} (\Gamma_0 - f ) )=0$.
Hence we obtain that the natural homomorphism $H^0(\Theta_{\mathbb{F}_n,L} ) \to H^0( \pi^*\Theta_{\mathbb{CP}^1})\simeq
H^0(\Theta_{\mathbb{CP}^1})$ is injective.
Therefore, unlike ${\rm{Aut}}_0\mathbb{F}_n$, the subgroup ${\rm{Aut}}_0(\mathbb{F}_n,L)$ can be regarded as
a~subgroup of $\mathrm{PGL}(2,\mathbb{C})$.
Moreover any $g\in {\rm{Aut}}_0(\mathbb{F}_n,L)$ has to f\/ix the point $p = \Gamma_0\cap L$, since $\Gamma_0$ is
${\rm{Aut}}\mathbb{F}_n$-invariant as $n>0$.
Hence under the above inclusion ${\rm{Aut}}_0(\mathbb{F}_n,L)\subset\mathrm{PGL}(2,\mathbb{C})$,
${\rm{Aut}}_0(\mathbb{F}_n,L)$ is included in the subgroup of $\mathrm{PGL}(2,\mathbb{C})$ in the proposition.
But since we already know $\dim {\rm{Aut}}_0(\mathbb{F}_n,L)=2$ by Proposition~\ref{prop:l1}, we obtain the coincidence.
\end{proof}
Thus the computations for $h^i(\Theta_{\mathbb{F}_n,L})$ and ${\rm{Aut}}_0(\mathbb{F}_n,L)$ is over for arbitrary
sections when $l\in \{0,1\}$.
Next we consider the case $l>1$.
In this case, the situation is not completely homogeneous:
\begin{Proposition}\label{prop:l2}
Suppose $n>0$, $l>1$, and let~$L$ be any section of $\mathbb{F}_n\to \mathbb{CP}^1$ belonging to the system $|\Gamma_0+(n+l)f|$.
Then we have $H^2(\Theta_{\mathbb{F}_n,L})=0$.
Further one of the following holds:
\begin{enumerate}\itemsep=0pt
\item[$(i)$] $h^0(\Theta_{\mathbb{F}_n, L}) = 1$, $h^1(\Theta_{\mathbb{F}_n, L}) = n+2l-4$, and there is an exact
sequence
\begin{gather*}
0
\longrightarrow
\mathbb{C}^{2l-3}
\longrightarrow
H^1(\Theta_{\mathbb{F}_n,L})
\longrightarrow
H^1(\Theta_{\mathbb{F}_n})
\longrightarrow
0.
\end{gather*}
\item[$(ii)$] $h^0(\Theta_{\mathbb{F}_n, L}) = 0$, $h^1(\Theta_{\mathbb{F}_n, L}) = n+2l-5$, and there is an exact
sequence
\begin{gather*}
0
\longrightarrow
\mathbb{C}^{2l-4}
\longrightarrow
H^1(\Theta_{\mathbb{F}_n,L})
\longrightarrow
H^1(\Theta_{\mathbb{F}_n})
\longrightarrow
0.
\end{gather*}
\end{enumerate}
Furthermore, as long as the section~$L$ satisfies~$(i)$, the complex structure of the pair $(\mathbb{F}_n,L)$ is
independent of the choice of~$L$, and we have ${\rm{Aut}}_0(\mathbb{F}_n, L) \simeq \mathbb{C}^*$.
\end{Proposition}
\begin{proof}
The vanishing $H^2 (\Theta_{\mathbb{F}_n, L}) =0$ is an immediate consequence of the exact sequence
\begin{gather}
\label{ses:56}
0
\longrightarrow
\Theta_{\mathbb{F}_n,L}
\longrightarrow
\Theta_{\mathbb{F}_n}
\longrightarrow
N_{L/\mathbb{F}_n}
\longrightarrow
0
\end{gather}
since as $N_{L/\mathbb{F}_n}\simeq\mathscr O(n+2l)$ we have $H^1(N_{L/\mathbb{F}_n}) = 0$ and also
$H^2(\Theta_{\mathbb{F}_n}) =0$.
{\sloppy The ingredient is to show $h^0(\Theta_{\mathbb{F}_n, L}) \le 1$.
For this we f\/irst note that, in the same way to the f\/irst column of the commutative diagram in the proof of
Proposition~\ref{prop:l0} or the exact sequence~\eqref{ntses} in the case $l=1$, we have an exact sequence
\begin{gather*
0
\longrightarrow
\mathscr O_{\mathbb{F}_n} (\Gamma_0 - lf )
\longrightarrow
\Theta_{\mathbb{F}_n,L}
\longrightarrow
\pi^*\Theta_{\mathbb{CP}^1}
\longrightarrow
0.
\end{gather*}
This again means that the natural map $H^0(\Theta_{\mathbb{F}_n,L} ) \to H^0(\Theta_{\mathbb{CP}^1})$ is injective, and
hence ${\rm{Aut}}_0(\mathbb{F}_n,L)$ may be considered as a~subgroup of $\mathrm{PGL}(2,\mathbb{C})$.
Moreover, since $\Gamma_0$ is ${\rm{Aut}}\mathbb{F}_n$-invariant, ele\-ments of ${\rm{Aut}}_0(\mathbb{F}_n,L)$ f\/ix any
point of the intersection $\Gamma_0\cap L$.
Since we have $(L,\Gamma_0)=(\Gamma_0 + (n+l)f,\Gamma_0) = l\ge 2$, $L\cap \Gamma_0$ is non-empty.
If it consists of more than two points, then the image of ${\rm{Aut}}_0(\mathbb{F}_n,L)\to \mathrm{PGL}(2,\mathbb{C})$
is clearly identity, and so ${\rm{Aut}}_0(\mathbb{F}_n,L)$ is trivial, meaning $h^0(\Theta_{\mathbb{F}_n,L})=0$.
}
If $L\cap \Gamma_0$ consists of two points, the image of ${\rm{Aut}}_0(\mathbb{F}_n,L)\to \mathrm{PGL}(2,\mathbb{C})$ is
included in the $\mathbb{C}^*$-subgroup determined by the two points.
Suppose that the image is actually the $\mathbb{C}^*$-subgroup, and let $T_{\mathbb{C}}$ be the maximal torus of
${\rm{Aut}}_0(\mathbb{F}_n)$ which contains ${\rm{Aut}}_0(\mathbb{F}_n,L)(\simeq \mathbb{C}^*)$.
Then $T_{\mathbb{C}}$ determines on $\mathbb{F}_n$ a~structure of toric surface, and singles out a~$(+n)$-section
$\Gamma_{\infty}$ by $T_{\mathbb{C}}$-invariance.
Moreover~$L$ cannot intersect $\Gamma_{\infty}$ since~$L$ minus the two f\/ixed points $L\cap\Gamma_0$ forms an orbit of
the $\mathbb{C}^*$-subgroup of $T_{\mathbb{C}}$, and $\Gamma_{\infty}$ is disjoint from the unique 2-dimensional orbit
of the $T_{\mathbb{C}}$-action.
This contradicts $(L,\Gamma_{\infty})=n+l$ $(>0)$.
Therefore if $L\cap \Gamma_0$ consists of two points, ${\rm{Aut}}_0(\mathbb{F}_n,L)$ is trivial.
If $L\cap \Gamma_0$ consists of one point, since $(L,\Gamma_0)=l$, as in the same way to the proof of
Proposition~\ref{prop:l1}, we can f\/ind coordinates $(u,\zeta)$ on the line bundle $\mathscr O(-n)\subset\mathbb{F}_n$
such that the point $L\cap \Gamma_0$ corresponds to the origin and~$L$ is def\/ined by an equation $\zeta = u^l$.
In these coordinates the $(+n)$-section def\/ined by the equation $\zeta=\infty$ intersects~$L$ at the unique point
$(\infty,\infty)$ by the biggest multiplicity $(n+l)$.
Then using the computations in Proposition~\ref{prop:l1}, by writing a~vector f\/ield $\theta\in H^0(\Theta_{\mathbb{F}_n})$
as in~\eqref{MK1}, we have~\eqref{der1} for the derivative $\theta F|_L$ of the def\/ining equation $F=\zeta-u^l$ of~$L$.
Now as $l>1$ we have $2l>l+1$.
Hence looking the powers to $u$ in~\eqref{der1}, the vanishing $\theta F|_L = 0$ is equivalent to the equations
\begin{gather*}
b_1=b_2=\dots=b_{n+1} = 0,
\qquad
c-la_1 = a_2 = a_3 = 0.
\end{gather*}
From these we obtain $h^0(\Theta_{\mathbb{F}_n, L})=1$.
Thus we have seen that if $l>1$ we always have \mbox{$h^0(\Theta_{\mathbb{F}_n,L}) \le 1$} and the equality holds exactly
when~$L$ touches the section $\Gamma_0$ at a~point by the biggest multiplicity.
Once this is obtained, the assertions (i) and (ii) are readily obtained from the exact sequence~\eqref{ses:56}.
We omit the detail.
The f\/inal assertion is clear from the above argument since we have seen that if $h^0(\Theta_{\mathbb{F}_n,L}) = 1$,
equation for~$L$ can be taken as $\zeta=u^l$ in the coordinates $(u,\zeta)$ on the line bundle $\mathscr O(-n)$.
\end{proof}
As a~corollary to the results in this subsection, we obtain the following result on the existence and uniqueness up to
isomorphisms of $\mathbb{C}^*$-invariant pair $(\mathbb{F}_n,L)$ for each $n$ and $l$:
\begin{Corollary}\label{cor:unique}
For each integers $n\ge 0$ and $l\ge 0$, there exists a~section~$L$ of $\mathbb{F}_n\to\mathbb{CP}^1$ which satisfies
the following two properties:
\begin{enumerate}\itemsep=0pt
\item[$(i)$] $L\in |\Gamma_0+(n+l)f|$,
\item[$(ii)$] $L$ is $\mathbb{C}^*$-invariant, where $\mathbb{C}^*$ is a~subgroup of ${\rm{Aut}}\mathbb{F}_n$ which acts
non-trivially on~$L$.
\end{enumerate}
Moreover, for each $n$ and $l$ the complex structure of the pair $(\mathbb{F}_n,L)$ is independent of the choice of such
a~section~$L$.
\end{Corollary}
\begin{proof}
The assertion for the case $n>0$ follows from Propositions~\ref{prop:l0},~\ref{prop:l1} and~\ref{prop:l2}.
The assertion for the case $n=0$ is immediate to see.
\end{proof}
Finally in this subsection, we discuss variation of the complex structures of the af\/f\/ine bundles in the family
$\mathscr A_n\to\mathbb{C}^{n-1}$.
As in the beginning of this subsection, by identifying the total space of the line bundle $\mathscr O(-n)$ with the
minimal resolution of $\mathbb{C}^2/\mathbb{Z}_n$, the line bundle $\mathscr O(-n)$ admits
a~${\rm{GL}}(2,\mathbb{C})$-action.
This naturally gives rise to a~${\rm{GL}}(2,\mathbb{C})$-action on the cohomology group $H^1(\mathbb{CP}^1,\mathscr
O(-n))$.
Recalling that this cohomology group is exactly the base space of the family $\mathscr A_n\to\mathbb{C}^{n-1}$, f\/ibers
of the family are mutually biholomorphic if they are over the same orbit of the ${\rm{GL}}(2,\mathbb{C})$-action.
From the results in~\cite{HonCMP2}, the ${\rm{GL}}(2,\mathbb{C})$-action on the base space $\mathbb{C}^{n-1}$ is
identif\/ied with the tensor product
\begin{gather*}
S_1^{n-2}\mathbb{C}^2:=S^{n-2}\mathbb{C}^2\otimes\mathbb{C}_1,
\end{gather*}
where $S^{n-2}\mathbb{C}^2$ is the $(n-2)$-th symmetric product of the natural ${\rm{GL}}(2,\mathbb{C})$-action on
$\mathbb{C}^2$, and $\mathbb{C}_1$ is the 1-dimensional representation of ${\rm{GL}}(2,\mathbb{C})$ which is just the
multiplication of the determinant.
(See Section~\ref{explicit}, especially the isomorphisms~\eqref{isom49}.) It follows that if $n\in\{2,3\}$ the
${\rm{GL}}(2,\mathbb{C})$-action on the base space $\mathbb{C}^{n-1}\backslash\{0\}$ is transitive.
Therefore when $n\in\{2,3\}$, any member of the family $\mathscr A_n\to\mathbb{C}^{n-1}$ is biholomorphic except the
central f\/iber $\mathscr O(-n)$.
These can also be seen by just noting that, if $n\in\{2,3\}$, any f\/iber except the central f\/iber is identif\/ied
with $\mathbb{F}_{n-2}\backslash L$ with $L\in |\Gamma_0+(n-1)f|$, and also that the complex structure of the pair
$(\mathbb{F}_{n-2},L)$ is independent of the choice of a~non-singular member $L\in|\Gamma_0+(n-1)f|$ by
Proposition~\ref{prop:l1}.
On the other hand, if $n\ge 4$, the ${\rm{GL}}(2,\mathbb{C})$-action on the base space
$S_1^{n-2}\mathbb{C}^2=\mathbb{C}^{n-1}$ minus the origin is not transitive, and so the quotient space
$(\mathbb{C}^{n-1}\backslash\{0\})/{\rm{GL}}(2,\mathbb{C})$ consists of more than two elements.
As the total spaces of the af\/f\/ine bundles are not only open but also do not have compact holomorphic curves, it
seems dif\/f\/icult to determine when two af\/f\/ine surfaces lying over dif\/ferent ${\rm{GL}}(2,\mathbb{C})$-orbits
are mutually biholomorphic (if $n\ge 4$).
\subsection{Computations for surfaces of smooth normal crossing}
\label{ss:sncs}
In this subsection we f\/irst construct a~variety of smooth normal crossing from two copies of the Hirzebruch surface
$\mathbb{F}_n$ by identifying the same sections, and then compute cohomology groups for them.
In the next section these varieties will be included as a~subvariety in twistor spaces of the 4-dimensional orbifold
$\widehat{\mathscr O(-n)}$.
For this let $D$ be a~non-singular complex surface and $L\subset D$ a~non-singular rational curve.
Denoting $J$ for the complex structure on~$D$, we denote by $\overline D$ the complex surface obtained from~$D$ by
changing the complex structure~$J$ to $-J$.
Let $\id :D\to \overline D$ be the identity map.
This is an {\em anti-}holomorphic map.
Write $\overline L:=\id(L)\subset \overline D$.
Let $\tau:L\to L$ be an anti-holomorphic involutions of~$L\simeq\mathbb{CP}^1$, and we def\/ine a~map
$\phi:L\to\overline L$ by $\phi:=\id|_{L}\circ\tau$.
Since both~$\id|_{L}$ and~$\tau$ are anti-holomorphic, $\phi$ is a~holomorphic map.
Let $D\cup_{L,\tau}\overline D$ be the space obtained from the disjoint union $D\sqcup \overline D$ by identifying~$L$
and~$\overline L$ by~$\phi$.
By the holomorphicity of~$\phi$, $D\cup_{L,\tau}\overline D$ is naturally equipped with the structure of a~complex
variety which is smooth normal crossing.
Let $\sigma:D\sqcup \overline D\to D\sqcup \overline D$ be the map def\/ined by
\begin{gather*}
\sigma (p) =
\begin{cases}
\id (p) &\text{if}\quad p\in D,
\\
\id^{-1}(p) &\text{if}\quad p\in \overline D.
\end{cases}
\end{gather*}
This is clearly an involution which f\/lips the two components (as the map $\id$ f\/lips from the def\/inition),
and is an anti-holomorphic map since $\id$ and $\id^{-1}$ are.
If the two points $p\in L$ and $q\in \overline L$ satisfy $q = \phi(p)$, we have
\begin{gather*}
\phi(\sigma(q)) = \phi(\id^{-1}(q))=\id\circ\tau\circ\id^{-1}(q)=\id\circ\tau\circ\id^{-1} (\id\circ\tau(p)) = \id(p)=\sigma(p).
\end{gather*}
Namely we have $\phi(\sigma(q)) = \sigma(p)$.
Hence $\sigma$ descends to an endomorphism of $D\cup_{L,\tau}\overline D$.
We use the same letter $\sigma$ for this map.
This is an anti-holomorphic involution since the original $\sigma$ is.
Thus the variety $D\cup_{L,\tau}\overline D$ is naturally equipped with a~real structure.
The structure of $D\cup_{L,\tau}\overline D$ as a~complex variety with a~real structure depends not only on the
rational curve~$L$ but also on the involution $\tau$.
Further, if $p\in L$, we have
\begin{gather*}
\phi^{-1}(\sigma(p)) = (\id\circ\tau)^{-1}(\id (p)) = \tau^{-1}(p) = \tau (p).
\end{gather*}
This means that on the intersection $D\cap \overline D\subset D\cup_{L,\tau}\overline D$, the involution $\sigma$ may be
identif\/ied with the involution $\tau$ on~$L$.
We apply this construction to the pair $(D,L) = (\mathbb{F}_n,L)$, where $n > 0$ and~$L$ is a~section of
$\mathbb{F}_n\to \mathbb{CP}^1$ with a~positive self-intersection number and an anti-holomorphic involution $\tau:L\to L$ without a~f\/ixed point.
As above the structure of the resulting variety $\mathbb{F}_n\cup_{L,\tau}\overline{\mathbb{F}}_n$ depends on the choice
of the involution $\tau$.
But if the section~$L$ is supposed to be invariant under a~$\mathbb{C}^*$-action on $\mathbb{F}_n$ that acts on
non-trivially on~$L$, then the choice of $\tau$ is naturally constrained to be $\mathbb{C}^*$-equivariant, and
consequently if $p$ and $q$ denote the f\/ixed points of the $\mathbb{C}^*$-action on~$L$, we have $\tau(p)\in\{p,q\}$.
But since $\tau$ is supposed to have no f\/ixed point, we obtain $\tau(p) = q$.
This means that in an af\/f\/ine coordinate $u$ on~$L$ for which the $\mathbb{C}^*$-action is given by $u\mapsto tu$ for
$t\in\mathbb{C}^*$, we can write $\tau(u) = -a/\overline u$ for some $a>0$.
Therefore the ef\/fect of varying $\tau$ (namely varying the number $a>0$) is absorbed in the $\mathbb{C}^*$-action on
$\mathbb{F}_n$, and moreover by Corollary~\ref{cor:unique}, the complex structure of the pair $(\mathbb{F}_n,L)$ is
independent of the choice of such a~section~$L$.
Consequently the variety $\mathbb{F}_n\cup_{L,\tau}\overline{\mathbb{F}}_n$ makes a~unique sense.
Further the $\mathbb{C}^*$-actions on $(\mathbb{F}_n,L)$ and $(\overline{\mathbb{F}}_n,\overline L)$ are naturally glued
and the variety is equipped with a~$\mathbb{C}^*$-action.
As we are particularly interested in these varieties, we introduce notation for them:
\begin{Definition}
\label{def:scn}
For integers $n \ge 0$ and $l\ge 0$, let $L\in |\Gamma_0 + (n+l)f|$ be any $\mathbb{C}^*$-invariant section on
$\mathbb{F}_n$, and we denote by $\mathbb{F}_n\cup_l\overline{\mathbb{F}}_n$ for the variety of simple normal crossing
with $\mathbb{C}^*$-action, which is obtained from the two copies of the pair $(\mathbb{F}_n,L)$ by identifying two
$L$-s by an anti-holomorphic involution $\tau$ without a~f\/ixed point in the above way.
\end{Definition}
The notation $\mathbb{F}_n\cup_l\overline{\mathbb{F}}_n$ ref\/lects the independency from the choices of~$L$ and $\tau$.
Thus the complex structure of this variety is solely determined by two non-negative integers $n$ and $l$.
For these varieties we have the following.
\begin{Proposition}\label{prop:fd2}
Let $n>0$ and $l\ge 0$.
Then for the tangent sheaf $\Theta$ of the variety $\mathbb{F}_n \cup_l \mathbb{F}_n$ above, we have the following:
\begin{itemize}\itemsep=0pt
\item[$(i)$] If $l=0$, we have
\begin{gather*}
h^0(\mathbb{F}_n \cup_l \overline{\mathbb{F}}_n,\Theta) = 5,
\qquad
h^1(\mathbb{F}_n \cup_l \overline{\mathbb{F}}_n,\Theta) = 2(n-1),
\qquad
h^2(\mathbb{F}_n \cup_l \overline{\mathbb{F}}_n,\Theta) = 0.
\end{gather*}
\item[$(ii)$] If $l\ge 1$, we have
\begin{gather*}
h^0(\mathbb{F}_n \cup_l \overline{\mathbb{F}}_n,\Theta) = 1,
\qquad
h^1(\mathbb{F}_n \cup_l \overline{\mathbb{F}}_n,\Theta) = 2(n+2l-3),
\qquad
h^2(\mathbb{F}_n \cup_l \overline{\mathbb{F}}_n,\Theta) = 0.
\end{gather*}
\end{itemize}
\end{Proposition}
\begin{proof}
Though these can be shown in a~standard way by using Propositions~\ref{prop:l0},~\ref{prop:l1} and~\ref{prop:l2}, we
write a~proof as there is a~subtle point that relies on our construction of the variety
$\mathbb{F}_n\cup_l\overline{\mathbb{F}}_n$.
We have the standard exact sequence $0 \to \Theta_{\mathbb{F}_n\cup_l \overline{\mathbb{F}}_n} \to
\Theta_{\mathbb{F}_n,L}\oplus\Theta_{\overline{\mathbb{F}}_n,\overline L}\to \Theta_L\to 0$, where $L\in
|\Gamma_0+(n+l)f|$ is a~$\mathbb{C}^*$-invariant section identif\/ied by $\phi$.
For the case $l=0$, the natural map $H^0(\Theta_{\mathbb{F}_n,L}) \to H^0(\Theta_{L})$ is surjective from that of
${\rm{Aut}}_0(\mathbb{F}_n,L)\to \Aut L$.
Therefore from the above exact sequence and Proposition~\ref{prop:l0} (i), (ii) we obtain the required value for $h^i
(\mathbb{F}_n \cup_0 \overline{\mathbb{F}}_n,\Theta)$ for any~$i$ as well as natural isomorphisms {\samepage
\begin{gather}
H^1(\mathbb{F}_n \cup_0 \overline{\mathbb{F}}_n,\Theta) \simeq H^1(\mathbb{F}_n, \Theta_{\mathbb{F}_n,L}) \oplus H^1
(\overline{\mathbb{F}}_n,\Theta_{\overline{\mathbb{F}}_n,\overline L})
\label{isom001}
\\
\hphantom{H^1(\mathbb{F}_n \cup_0 \overline{\mathbb{F}}_n,\Theta)}{}
\simeq H^1(\mathbb{F}_n, \Theta_{\mathbb{F}_n}) \oplus H^1 (\overline{\mathbb{F}}_n,\Theta_{\overline{\mathbb{F}}_n}).
\label{isom002}
\end{gather}}
\noindent
Next for the case $l=1$, the natural homomorphism ${\rm{Aut}}_0(\mathbb{F}_n,L)\to \Aut L$ is not surjective and the
image is the af\/f\/ine transformation group as in Proposition~\ref{prop:nonred}.
Namely it consists of elements of $\Aut L$ which f\/ixes the point $p=\Gamma_0\cap L$.
As our involution $\tau$ is supposed to interchange the two f\/ixed points $p$ and $q$ of the $\mathbb{C}^*$-action, it
follows that the image of the natural map $H^0(\mathbb{F}_n, \Theta_{\mathbb{F}_n,L}) \oplus H^0
(\overline{\mathbb{F}}_n,\Theta_{\overline{\mathbb{F}}_n,\overline L})\to H^0(\Theta_L)$ is again surjective since the
two af\/f\/ine groups generate $\Aut L$.
Hence the cohomology exact sequence takes the same form as the case $l=0$, and by using Proposition~\ref{prop:l1} (i)
and (ii), we obtain the required value of $h^i (\mathbb{F}_n \cup_1 \overline{\mathbb{F}}_n,\Theta)$ as well as the
natural isomorphisms~\eqref{isom001} and~\eqref{isom002}.
Finally if $l>1$, by Proposition~\ref{prop:l2} the image of the natural injection ${\rm{Aut}}_0(\mathbb{F}_n,L)\to \Aut
L$ is the $\mathbb{C}^*$-subgroup that f\/ixes the two points $p$ and $q$.
Therefore from our choice of $\tau$, the image of the natural map $H^0(\mathbb{F}_n, \Theta_{\mathbb{F}_n,L}) \oplus H^0
(\overline{\mathbb{F}}_n,\Theta_{\overline{\mathbb{F}}_n,\overline L})\to H^0(\Theta_L)$ is 1-dimensional.
Therefore from the cohomology sequence we obtain $H^0(\Theta_{\mathbb{F}_n\cup_l\overline{\mathbb{F}}_n})
\simeq\mathbb{C}$, the exact sequence
\begin{gather*}
0
\longrightarrow
\mathbb{C}^2
\longrightarrow
H^1 (\mathbb{F}_n \cup_l \overline{\mathbb{F}}_n,\Theta)
\longrightarrow
H^1(\mathbb{F}_n, \Theta_{\mathbb{F}_n,L}) \oplus H^1
(\overline{\mathbb{F}}_n,\Theta_{\overline{\mathbb{F}}_n,\overline L})
\longrightarrow
0
\end{gather*}
and the isomorphism $H^2 (\mathbb{F}_n \cup_l \overline{\mathbb{F}}_n,\Theta) \simeq H^2(\mathbb{F}_n,
\Theta_{\mathbb{F}_n,L}) \oplus H^2 (\overline{\mathbb{F}}_n,\Theta_{\overline{\mathbb{F}}_n,\overline L})$.
From Proposition~\ref{prop:l2}, we f\/inish the proof of the assertion (ii).
\end{proof}
\section{Computations for twistor spaces}
\label{explicit}
In this section, based on the results in the previous section, we intensively study small deformations of the LeBrun
metric on $\mathscr O(-n)$ which preserve ALE SFK properties, and in particular show that any af\/f\/ine
$\mathbb{C}$-bundle over $\mathbb{CP}^1$ of negative degree admits an ALE SFK metric.
Next we investigate small deformations of the metrics on the af\/f\/ine bundles again as ALE SFK metrics, and in
particular show that even if we f\/ix the complex structure on the af\/f\/ine bundles, they admit a~1-parameter
deformation for which the conformal classes are not constant.
\subsection{Some generalities on twistor spaces of ALE SFK metrics}
Before starting actual computations, we brief\/ly recall basic properties of the twistor spaces of ALE SFK metrics,
including its natural compactif\/ication.
These will be used for investigating deformations of metrics which preserve ALE SFK property.
For more precise treatment on compactif\/ications of ALE ASD 4-manifolds, we refer the paper~\cite{V}.
Let $(X,J)$ be a~complex surface, $g$ an ASD Hermitian metric on it, $Z$ the twistor space of the ASD conformal class
$[g]$, and $F:=K_Z^{-1/2}$ the natural square root of the anticanonical line bundle of $Z$, which is available on any
twistor space.
Then the complex structure $J$ determines a~section of the twistor projection $Z\to X$ in a~tautological way, and its
image becomes a~non-singular divisor $D$ on $Z$.
$D$ is biholomorphic to $X$ by the projection $Z\to X$.
Let $\overline D$ be the divisor determined by the conjugate complex structure $-J$ on $X$.
We always have $D\cap \overline D=\varnothing$ as $J\neq -J$.
Then Pontecorvo's theorem~\cite{Pont92} means that the ASD Hermitian metric $g$ is K\"ahler with respect to $J$ if and
only if $D + \overline D\in |F|$.
If $X$ is non-compact with one end and the ASD metric $g$ is asymptotically Euclidean at inf\/inity, then $(X,[g])$ can
be compactif\/ied as an ASD manifold by adding a~point at inf\/inity.
Let $(\hat X,[\hat g])$ be the resulting compact ASD manifold, and $\hat Z$ the twistor space of $(\hat X, [\hat g])$,
which is smooth.
Then the closure ${\rm{Cl}}(D)$ of the above divisor $D\subset Z$ is a~divisor in $\hat Z$, and from the ALE SFK
property of the metric~\cite[proof of Proposition~6, p.~312]{LB92}, the divisor ${\rm{Cl}}(D)$ satisf\/ies the following
properties: (i)~${\rm{Cl}}(D)$ is still non-singular and ${\rm{Cl}}(D) = D\sqcup L$, where~$L$ is the twistor line over
the point at inf\/inity, (ii)~${\rm{Cl}}(D)\cap {\rm{Cl}}(\overline D)=L$, and the intersection is transverse, and (iii)~the normal bundle of~$L$ in ${\rm{Cl}}(D)$ (and also in ${\rm{Cl}}(\overline D)$) is of degree one.
Conversely a~divisor in $\hat Z$ satisfying these properties determines, up to overall constants, an ASD K\"ahler metric
on $(X,J)$ which is asymptotically Euclidean at inf\/inity.
When the SFK surface $(X,J,g)$ is ALE in a~strict sense (i.e.\ asymptotic to the f\/lat Euclidean orbifold
$\mathbb{C}^2/\Gamma$ at inf\/inity, where $\Gamma$ is a~non-trivial f\/inite subgroup of ${\rm{U}}(2)$ acting freely on
the unit sphere), the pair $(X,g)$ has a~natural compactif\/icaton $(\hat X, \hat g)$ as an ASD orbifold, which means
that~$\hat X$ is an orbifold of the form $X\cup\{\infty\}$ with $\infty$ being an orbifold point of~$\hat X$, and~$\hat
g$ is an ASD orbifold metric on~$\hat X$ whose conformal class on $X$ remains to be equal to~$g$.
Also the twistor space~$Z$ of~$(X,g)$ has a~natural compactif\/ication, for which we again denote by~$\hat Z$.
This is of course the twistor space of the ASD orbifold~$(\hat X,\hat g)$ in a~natural sense, and we again have~$\hat Z
= Z\sqcup L$, where~$L$ is the twistor line over the orbifold point~$\infty$.
We have $\Sing\hat Z\subset L$, and all singularities are quotient singularity by the group which is
orientation-reversing conjugate (namely conjugate after reversing the orientation; see~\cite[Def\/inition~1.4]{V} for the precise
def\/inition) to the above group~$\Gamma$.
Especially, denoting ${\rm{U}}(1)\subset{\rm{U}}(2)$ for the subgroup of consisting of scalar matrices, if~$\Gamma$ is
a~cyclic subgroup of ${\rm{U}}(1)$ with order $n\ge 2$, then $\hat Z$ has $A_{n-1}$-singularities along~$L$.
(This is particular to these subgroups, and for other subgroup $\Gamma\subset U(2)$, singular points of~$\hat Z$ are
isolated.) Moreover if $D\subset Z$ again denotes the divisor determined by the complex structure~$J$ on~$X$ and
${\rm{Cl}}(D)$ means its closure in $\hat Z$, then ${\rm{Cl}}(D)$ itself (and therefore ${\rm{Cl}}(\overline D)$ also)
is a~non-singular (but non-Cartier) divisor on $\hat Z$.
Moreover we have ${\rm{Cl}}(D)\cap {\rm{Cl}}(\overline D)=L$, and the normal bundle of~$L$ in ${\rm{Cl}}(D)$ (and also
in ${\rm{Cl}}(\overline D)$) is of degree $n$.
Furthermore the union ${\rm{Cl}}(D) \cup {\rm{Cl}}(\overline D)$ itself is smooth normal crossing.
We also note that in this situation the natural extension of the line bundle $F$ over $Z$ to $\hat Z$ is not just an
orbifold bundle but an ordinary line bundle; in other words the sum ${\rm{Cl}}(D) + {\rm{Cl}}(\overline D)$ is a~Cartier
divisor on $\hat Z$, while ${\rm{Cl}}(D)$ and ${\rm{Cl}}(\overline D)$ are not.
\looseness=-1
Because $\mathscr O(-n)$ is obtained as the minimal resolution of the quotient space $\mathbb{C}^2/\Gamma$ where
$\Gamma\subset{\rm{U}}(2)$ is the cyclic subgroup of scalar matrices of order $n$, ALE SFK metrics on $\mathscr O(-n)$
give rise to the last situation where the compactif\/ied twistor space $\hat Z$ has $A_{n-1}$-singularities along the
twistor line~$L$ at inf\/inity.
Here, we do not suppose that the complex structure on $\mathscr O(-n)$ is the natural one and we will also consider
complex structures which support the af\/f\/ine bundles in Section~\ref{ss:ab}.
Let $\widehat{\mathscr O(-n)}$ be the one-point compactif\/ication of the 4-manifold $\mathscr O(-n)$, and in the
following, instead of the letters $\hat Z$ and ${\rm{Cl}}(D)$, we use the letters $Z$ and $D$ respectively to mean the
twistor space of the conformal compactif\/ication of an ALE SFK metric on the 4-mani\-fold~$\mathscr O(-n)$ and the
(non-Cartier) divisor on $Z$ determined by the complex structure on the 4-mani\-fold~$\mathscr O(-n)$.
In this situation $D$ is biholomorphic to $\mathbb{F}_{n-2k}$ for some $k\ge 0$ satisfying $n-2k\ge 0$.
This is because $D$ contains the twistor line~$L$ at inf\/inity as a~$(+n)$-curve as above, which means the rationality
of $D$; further the decomposition $D=\mathscr O(-n)\sqcup L$ as a~smooth manifold means $b_2(D) = 2$, and hence
$D\simeq\mathbb{F}_m$ for some $m\ge 0$; but $\mathbb{F}_m\to\mathbb{CP}^1$ has a~$(+n)$-section if\/f $m=n-2k$ for some $k\ge0$.
Of course we have $k=0$ if the complex structure on $\mathscr O(-n)$ is the natural one.
We also remark that if $\tau$ is an anti-holomorphic involution of~$L$ without a~f\/ixed point, the union
$D\cup\overline D$, which is a~Cartier divisor in $Z$ as above, is isomorphic to the surface $D\cup_{L,\tau}\overline D$
constructed in the f\/irst half of Section~\ref{ss:sncs}, as a~complex variety with a~real structure.
Thus if $(Z,D)$ is a~pair of a~compact but singular twistor space and a~divisor determined by an ALE SFK metric on the
4-manifold $\mathscr O(-n)$, deformations of the metric preserving ALE SFK property are equivalent to locally trivial
deformations of the pair $(Z,D\cup\overline D)$ preserving the real structure.
For details on locally trivial deformations for complex spaces and pairs of a~complex space and a~complex subspace of
it, we refer a~book~\cite[Section~3.4]{SeBook}.
In particular, if we def\/ine the subsheaf $\Theta_{Z, D\cup\overline D}$ of the tangent sheaf $\Theta_Z$ by
\begin{gather*}
\Theta_{Z, D\cup\overline D}:=\big\{v\in \Theta_Z\,|\, v(f) \in \mathscr I_{D\cup\overline D}
\text{ if }
f\in \mathscr I_{D\cup\overline D}\big\},
\end{gather*}
then f\/irst order deformations of the pair $(Z,D\cup\overline D)$ which are locally trivial are in one to one
correspondence with the cohomology group $H^1(\Theta_{Z,D\cup\overline D})$, and obstructions are in
$H^2(\Theta_{Z,D\cup\overline D})$.
In particular if $H^2(\Theta_{Z,D\cup\overline D})=0$, the Kuranishi family for locally trivial deformations of the pair
$(Z, D\cup\overline D)$ is constructed over a~neighborhood of the origin in $H^1(\Theta_{Z,D\cup\overline D})$.
\subsection{Deformations of the LeBrun metric}
\label{ss:DLm}
Having recalled these basic materials, we start to investigate deformations of the LeBrun's ALE SFK metric on $\mathscr
O(-n)$ as an ALE SFK metrics, by investigating locally trivial deformations of the pair $(Z,D\cup\overline D)$ of
compactif\/ied singular twistor space and the divisor.
The following proposition provides basic information about such deformations.
\begin{Proposition}\label{prop:isom1}
Suppose $n\ge 3$ and let $Z$ be the twistor space on the orbifold $\widehat{\mathscr O(-n)}$, which is associated to the
conformal compactification of the LeBrun's ALE-SFK metric on $\mathscr O(-n)$ with negative mass.
Let $D$ be the divisor on $Z$ which is the closure of the section of the twistor fibration that is determined by the
complex structure of $\mathscr O(-n)$.
$(D$ is biholomorphic to $\mathbb{F}_n.)$ Then we have
\begin{gather}
\label{cohomdim1}
H^i( \Theta_Z(-D-\overline D) ) = 0,
\quad
i\neq 1,
\qquad
H^1( \Theta_Z(-D-\overline D) ) \simeq \mathbb{C},
\\
H^2(\Theta_{Z, D\cup\overline D}) = H^2 ( \Theta_{D\cup \overline D} ) = 0.
\label{van1}
\end{gather}
Moreover there is a~natural isomorphism
\begin{gather}
\label{basicisom1}
H^1 (\Theta_{Z, D \cup \overline D} ) \simeq H^1 (D, \Theta_{ D \cup \overline D} ),
\end{gather}
and these are $2(n-1)$-dimensional.
Furthermore the natural map
\begin{gather}
\label{forget}
H^1(\Theta_{Z,D\cup\overline D})
\longrightarrow
H^1(\Theta_Z)
\end{gather}
is injective, and if $n=3$, this is moreover surjective.
\end{Proposition}
The isomorphism~\eqref{basicisom1} will be of fundamental importance in the rest of this article.
\begin{proof}
The vanishing $H^2( \Theta_Z(-D-\overline D) ) = 0$ immediately follows from~\cite[Proposition~3.1]{HonCMP2} since $S$ in the
proposition is a~divisor in the system $|F|$ and hence $\Theta_Z(-D-\overline D)\simeq\Theta_Z(-S)$.
In order to compute $h^i(\Theta_Z(-D-\overline D) )$ for $i\in \{0,1,3\}$,
we use computations in the proof of the above proposition in~\cite{HonCMP2}.
Noting $\Theta_Z (-D-\overline D ) \simeq \Theta_Z\otimes F^{-1}$, the isomorphisms~(3.5),~(3.6) and~(3.8) in the proof
of~\cite[Proposition~3.1]{HonCMP2} are valid not only for $H^2$ but also for $H^i$ for any~$i$ because we have $H^i(\mathscr
O_{\mathbb{CP}^1} (-1))=H^i({\mathbb{CP}^1\times\mathbb{CP}^1},\mathscr O(-1,-1))=0$ for any~$i$.
Therefore in the notation of that proof, we have $H^i(\Theta_Z\otimes F^{-1}) \simeq H^i(X,\mathscr L')$ for any~$i$.
Further from the exact sequence~(3.9) there, we have $H^i(\mathscr L')\simeq H^i(\mathscr F')$ for any~$i$.
Furthermore from the exact sequence~(3.10) there, we obtain
\begin{gather*}
H^i(\mathscr F') = 0,
\quad
i\neq 1,
\qquad
H^1(\mathscr F')\simeq H^0(\Delta,\mathscr O) (\simeq\mathbb{C}).
\end{gather*}
These in particular imply~\eqref{cohomdim1}.
Next in order to deduce~\eqref{van1} and~\eqref{basicisom1} we consider the standard exact sequence
\begin{gather}
\label{ses:1}
0
\longrightarrow
\Theta_Z(-D-\overline D)
\longrightarrow
\Theta_{Z, D\cup\overline D}
\longrightarrow
\Theta_{D\cup \overline D}
\longrightarrow
0.
\end{gather}
Since the isometry group of the LeBrun metric on $\mathscr O(-n)$ is $U(2)/\mathbb{Z}_n$ \cite{LB88}, where
$\mathbb{Z}_n$ is the cyclic subgroup consisting of scalar matrices of order $n$, we have $ h^0 ( \Theta_{Z,
D\cup\overline D}) = 4.
$ On the other hand, as $D\cup \overline D\simeq \mathbb{F}_n\cup_0\overline{\mathbb{F}}_n$ biholomorphically, by
Proposition~\ref{prop:fd2}~(i) we have
\begin{gather}
\label{cohomdim2}
h^0 ( \Theta_{D\cup\overline D}) = 5,
\qquad
h^1 ( \Theta_{D\cup\overline D}) = 2(n-1),
\qquad
h^2 ( \Theta_{D\cup\overline D}) = 0.
\end{gather}
Therefore using~\eqref{cohomdim1} the cohomology exact sequence of~\eqref{ses:1} implies
\begin{gather}
\label{les:1}
0
\longrightarrow
H^0(\Theta_{Z, D\cup\overline D})
\big({\simeq}\,\mathbb{C}^4\big)
\longrightarrow
H^0(\Theta_{D\cup \overline D})
\big({\simeq}\, \mathbb{C}^5\big)
\nonumber
\\
\hphantom{0}{}
\longrightarrow
H^1(\Theta_Z(-D-\overline D))
(\simeq \mathbb{C})
\longrightarrow
H^1(\Theta_{Z, D\cup\overline D})
\longrightarrow
H^1(\Theta_{D\cup \overline D})
\big({\simeq}\, \mathbb{C}^{2(n-1)}\big)
\nonumber
\\
\hphantom{0}{}
\longrightarrow
0
\longrightarrow
H^2(\Theta_{Z, D\cup\overline D})
\longrightarrow
0.
\end{gather}
From this we obtain $H^2( \Theta_{Z, D\cup\overline D}) = 0$, an exact sequence
\begin{gather}
\label{LB:key}
0
\longrightarrow
H^0(\Theta_{Z,D\cup\overline D})
\big({\simeq}\, \mathbb{C}^4\big)
\longrightarrow
H^0(\Theta_{D\cup \overline D})
\big({\simeq}\, \mathbb{C}^5\big)
\longrightarrow
H^1( \Theta_Z(-D-\overline D))
(\simeq \mathbb{C})
\longrightarrow
0,\!\!\!\!
\end{gather}
and also the isomorphism~\eqref{basicisom1}.
From the last isomorphism we obtain $h^1(\Theta_{Z,D\cup\overline D}) = 2(n-1)$ by~\eqref{cohomdim2}.
Finally we show that the map~\eqref{forget} is injective.
For this let $N'$ be the cokernel sheaf of the natural injection $\Theta_{Z,D\cup\overline D}\to \Theta_Z$.
We have an exact sequence $0\to \Theta_{Z, D\cup \overline D}\to \Theta_Z\to N'\to 0$, and so for the injectivity it
suf\/f\/ices to show $H^0(N') = 0$.
Let $N:=\mathscr O_Z(D+\overline D)|_{D\cup\overline D}$ be the normal sheaf of the divisor $D\cup \overline D$ in~$Z$.
Since $D+\overline D\in |F|$ and $F$ is an ordinary line bundle on~$Z$, the sheaf~$N$ is an invertible $\mathscr
O_{D\cup\overline D}$-module, and isomorphic to $F|_{D+\overline D}$.
Then by computing local generators of the sheaves $\Theta_{Z}$ and $\Theta_{Z,D\cup\overline D}$ in coordinates, and
then comparing the resulting generators of the cokernel sheaf~$N'$ with local generators of~$N$, we obtain a~natural
isomorphism
\begin{gather}
\label{nottrivial}
N'\simeq N\otimes_{\mathscr O_{D\cup\overline D}}\mathscr I_{L},
\end{gather}
where $L=D\cap \overline D$ is the twistor line over the point at inf\/inity as before, and $\mathscr I_L$ is the ideal
sheaf of~$L$ in $D\cup \overline D$.
On the other hand, by the adjunction formula we have $K_{D\cup\overline D} \simeq K_Z + [D+\overline D]|_{D\cup\overline
D}\simeq -2F + F|_{D+\overline D}\simeq -F|_{D\cup\overline D}$.
Hence from~\eqref{nottrivial} we obtain $N'\simeq -K_{D\cup\overline D}\otimes \mathscr I_L$.
Further for the canonical sheaf of $D\cup\overline D$, as this itself is smooth normal crossing, we have
\begin{gather*
K_{D\cup\overline D}|_D \simeq K_D +[\overline D]|_D \simeq K_D+\mathscr O_D(L),
\end{gather*}
and similar for $K_{D\cup\overline D}|_{\overline D}$.
Hence by taking the inverse for these and taking a~tensor product with~$\mathscr O(-L)$, we obtain
\begin{gather}
\label{N''}
N'|_D\simeq -K_D-\mathscr O_D(2L),
\qquad
N'|_{\overline D}\simeq -K_{\overline D}-\mathscr O_{\overline D}(2L).
\end{gather}
Now as $D\simeq\mathbb{F}_n$ we have $-K_D\simeq 2\Gamma_0+(n+2)f$, and as~$L$ is a~$(+n)$-section
we have $\mathscr O_D(L)\simeq \Gamma_0+nf$.
Hence we have
\begin{gather}
\label{N'}
-K_D-\mathscr O_D(2L) \simeq 2\Gamma_0+(n+2)f -2(\Gamma_0 + n f) \simeq -(n-2)f.
\end{gather}
Thus as $n-2>0$ from the assumption $n>2$ we obtain $H^0(-K_D-\mathscr O_D(2L))=0$.
With reality, this means $H^0(D\cup \overline D,N')=0$.
Thus the injectivity of~\eqref{forget} follows.
If $n=3$, the map is also surjective since we have $h^1(\Theta_Z) = 4(n-2) = 4$ by~\cite[Proposition~2.1]{HonCMP2}, which
coincides with $2(n-1) = 4$.
\end{proof}
\begin{Remark}
The computations and the conclusions in the proposition are valid also for the case $n=2$ except the injectivity of the
map~\eqref{forget}.
For the case $n=2$, as in~\eqref{N'}, we have $N'|_D\simeq \mathscr O_D$.
With reality this means $N'\simeq \mathscr O_{D\cup\overline D}$, and hence we have $H^0(N')\simeq\mathbb{C}$.
Further from the cohomology exact sequence this is mapped to $H^1(\Theta_{Z,D\cup \overline D})$ injectively.
Thus the map~\eqref{forget} has a~1-dimensional kernel.
\end{Remark}
Next, letting $Z$ and $D$ be as in Proposition~\ref{prop:isom1}, we collect basic results on versal families of locally
trivial deformations of $Z$, $(Z,D\cup\overline D)$ and $D\cup \overline D$ and their relationship, which are readily
derived from Proposition~\ref{prop:isom1} and the results in Section~\ref{s:Hirz}.
First, for the the twistor space $Z$ of the LeBrun structure on $\widehat{\mathscr O(-n)}$, as showed
in~\cite[Proposition~2.1]{HonCMP2}, we have $H^2(\Theta_Z)=0$ and $h^1(\Theta_Z)=4n-8$.
Hence the parameter space of the Kuranishi family of locally trivial deformations of $Z$ may be identif\/ied with
a~neighborhood of the origin in $H^1(\Theta_Z)\simeq\mathbb{C}^{4n-8}$.
Versal family of~$Z$ as twistor spaces is obtained as the restriction of the Kuranishi family onto the real locus of the
neighborhood.
We denote the last real locus by~$\mathscr U_{\rm{ASD}}$, which is clearly smooth and real $(4n-8)$-dimensional.
As in~\cite{HonCMP2} we call the corresponding family of ASD conformal structures on $\widehat{\mathscr O(-n)}$
(parameterized by $\mathscr U_{\rm{ASD}}$) as the {\em versal family of ASD structures} for the LeBrun structure.
If $n>3$, not all these ASD structures preserve the K\"ahler representative.
From the construction we have a~canonical isomorphism
\begin{gather*
T_0\mathscr U_{\rm{ASD}}\simeq H^1(\Theta_Z)^{\sigma}
\end{gather*}
as real vector spaces, where the upper-script means the real subspace.
Second, for the pair $(Z,D\cup\overline D)$, in a~similar way to the above argument, since $ H^2 (\Theta_{Z, D \cup
\overline D} ) =0$ and $ h^1 (\Theta_{Z, D \cup \overline D} ) =2(n-1)$ as in Proposition~\ref{prop:isom1}, the
parameter space of the Kuranishi family for locally trivial deformations of the pair $(Z,D\cup\overline D)$ is
identif\/ied with a~neighborhood of the origin in $H^1 (\Theta_{Z, D \cup \overline D} ) \simeq\mathbb{C}^{2(n-1)}$.
Restricting this to the real locus, we obtain a~deformation of $Z$ preserving not only a~structure of twistor space but
also the K\"ahler representative in the conformal class.
Let $\mathscr K'\subset H^1(\Theta_{Z,D\cup\overline D})^{\sigma}$ be the parameter space of this family.
We have a~natural isomorphism $ T_0\mathscr K'\simeq H^1(\Theta_{Z,D\cup\overline D})^{\sigma} $.
For a~relationship between the families over $\mathscr U_{\rm{ASD}}$ (of twistor spaces) and $\mathscr K'$ (of pairs of
twistor spaces and Cartier divisors), by versailty, after a~possible shrinking of the domain, there is an induced map,
for which we denote by $\psi_1$, from $\mathscr K'$ to $\mathscr U_{\rm{ASD}}$, such that the pullback by~$\psi_1$ of
the family over $\mathscr U_{\rm{ASD}}$ is isomorphic to the $Z$-portion of the family of pairs over~$\mathscr K'$.
Though $\psi_1$ is not uniquely determined, the derivative $\psi'_1(0)$ is exactly the restriction of the
map~\eqref{forget} to the real locus.
By the proposition the last map is injective, and moreover isomorphism if $n=3$.
So if $n=3$ we may think $\mathscr K' \simeq \mathscr U_{\rm{ASD}}$ by $\psi_1$.
If $n>3$, since $h^1(\Theta_{Z,D\cup\overline D}) = 2(n-1)$ by Proposition~\ref{prop:isom1}, $\psi_1:\mathscr
K'\to\mathscr U_{\rm{ASD}}$ is an embedding as a~real submanifold of dimension $2(n-1)$ in~$\mathscr U_{\rm{ASD}}$ (and
$\dim \mathscr U_{\rm{ASD}} = 4(n-2)$ as above).
We call the image $\psi_1(\mathscr K')$ the {\em K\"ahler locus} in~$\mathscr U_{\rm ASD}$ and denote it by $\mathscr
K$.
If $n=3$, we may think $\mathscr K = \mathscr U_{\rm{ASD}}$ as above.
From the construction we have a~natural isomorphism
\begin{gather*}
\psi_1'(0)^{-1}: \ T_0\mathscr K
\;\stackrel{\sim}{\longrightarrow}
\;
H^1(\Theta_{Z,D\cup\overline D})^{\sigma}
\end{gather*}
as real vector spaces, where the upper-script means the real subspace.
Next for locally trivial deformations of the variety $D\cup\overline D$, since $H^2(\Theta_{D\cup\overline D}) = 0$ by
Proposition~\ref{prop:fd2} (i), the Kuranishi family is parameterized by a~neighborhood of the origin in
$H^1(\Theta_{D\cup\overline D})$.
Denote $\mathscr J\subset H^1(\Theta_{D\cup\overline D})^{\sigma}$ for the real locus of the neighborhood.
Then again by versality, after a~possible shrinking of the domain, there is an induced map, for which we denote by
$\psi_2$, from~$\mathscr K'$ to~$\mathscr J$ that induces an isomorphism between the two families.
Similarly to $\psi_1$, while~$\psi_2$ is not uniquely determined, the derivative $\psi'_2(0)$ is identif\/ied with the
real part of the natural map $H^1(\Theta_{Z,D\cup\overline D})\to H^1(\Theta_{D\cup\overline D})$.
The last map is an isomorphism by~\eqref{basicisom1}, and therefore $\psi_2$ is isomorphic in a~neighborhood of the
origin in~$\mathscr K'$.
Hence the composition $\psi_2\circ\psi_1^{-1}$ gives an isomorphism from the K\"ahler locus $\mathscr K\subset\mathscr
U_{\rm{ASD}}$ to $\mathscr J$, and the $D\cup\overline D$-portion of the families of pairs over~$\mathscr K'$ and the
family over $\mathscr J$ are isomorphic by~$\psi_2$.
The situation is summarized as in the following diagram:
\begin{gather}
\begin{CD}
H^1(\Theta_{D\cup\overline D})^{\sigma} @<{\sim}<< H^1(\Theta_{Z,D\cup\overline D})^{\sigma} @>{\rm{inj.}}>>
H^1(\Theta_{Z})^{\sigma}
\\
@A{\rm{incl.}}AA @A{\rm{incl.}}AA @ A{\rm{incl.}}AA
\\
\mathscr J @<{\sim}<{\psi_2}< \mathscr K' @>{\sim}>_{\psi_1}>\mathscr K\subset \mathscr U_{\rm{ASD}}
\label{diagram02}
\end{CD}
\end{gather}
(Note again that $\mathscr K = \mathscr U_{\rm{ASD}}$ when $n=3$.) Thus in order to understand the complex structures on
$\mathscr O(-n)$ determined by points on $\mathscr K$, it is enough to understand the complex structures on f\/ibers of
the family over $\mathscr J$.
For this purpose we recall from Sections~\ref{ss:ab} and~\ref{ss:AH} that the Kuranishi family $\mathscr
F_n\to\mathbb{C}^{n-1}$ of $\mathbb{F}_n$ is obtained from the family $\mathscr A_n\to\mathbb{C}^{n-1}$ of af\/f\/ine
bundles by taking a~simultaneous compactif\/ication.
Let $\mathscr L_n:=\mathscr F_n\backslash \mathscr A_n$ be the family of sections at inf\/inity.
We now apply the construction in Section~\ref{ss:sncs} to all f\/ibers of $(\mathscr F_n,\mathscr
L_n)\to\mathbb{C}^{n-1}$ simultaneously.
For this, we need to give an involution $\tau$ on each section to make the variety.
For this purpose we note that since all f\/ibers of $\mathscr F_n\to\mathbb{C}^{n-1}$ have a~common projection to
$\mathbb{CP}^1$ (equipped with the coordinates $u$ and $v$ as before), all f\/ibers of $\mathscr L_n\to\mathbb{C}^{n-1}$
are naturally identif\/ied each other.
Let $\tau_0:\mathbb{CP}^1\to\mathbb{CP}^1$ be an anti-holomorphic involution def\/ined by $\tau_0(u)=-1/\overline u$,
and through the identif\/ication we regard $\tau_0$ as an anti-holomorphic involution which is def\/ined on each f\/iber
of $\mathscr L_n\to\mathbb{C}^{n-1}$.
Then taking this $\tau_0$ as the involution $\tau$ in the construction of Section~\ref{ss:sncs} for any
$(t_1,\dots,t_{n-1}) \in\mathbb{C}^{n-1}$, we obtain a~family of smooth normal crossing surfaces, whose parameter space
is $\mathbb{C}^{n-1}$.
We write this family as
\begin{gather}
\label{rsl1}
\mathscr F_n\cup\overline{\mathscr F}_n\to\mathbb{C}^{n-1}.
\end{gather}
From the construction in Section~\ref{ss:sncs} each f\/iber of this family has a~canonical real structure that
interchanges the two components.
In the notation of Def\/inition~\ref{def:scn}, the f\/iber over the origin of this family is isomorphic to
$\mathbb{F}_n\cup_0\overline{\mathbb{F}}_n$ as a~complex variety with real structure, while on the $l$-th coordinate
axis $\mathbb{C}(t_l)$, f\/ibers are isomorphic to $\mathbb{F}_{n-2l}\cup_l\overline{\mathbb{F}}_{n-2l}$ except over the
origin.
The family~\eqref{rsl1} is in ef\/fect isomorphic to the (abstract) family over $\mathscr J$:
\begin{Lemma}
\label{lemma:isom4}
In a~sufficiently small neighborhood of the origin, the family~\eqref{rsl1} is isomorphic to the family of smooth
normal crossing surfaces over $\mathscr J$.
\end{Lemma}
\begin{proof}
By versality of the Kuranishi family for locally trivial deformations of $D\cup\overline D\simeq\mathbb{F}_n\cup_0
\overline{\mathbb{F}}_n$, we have an induced map, for which we denote by $\alpha$, from a~neighborhood of the origin of
the parameter space $\mathbb{C}^{n-1}$ of~\eqref{rsl1} to that of the last Kuranishi family, such that the pull-back by~$\alpha$ is isomorphic to the family~\eqref{rsl1}.
Though $\alpha$ is not uniquely determined, from naturality, the derivative~$\alpha'(0)$ is nothing but the
Kodaira--Spencer map for~\eqref{rsl1} at the origin.
On the other hand as in~\eqref{isom001} and~\eqref{isom002} we have natural isomorphisms
\begin{gather*}
H^1(\mathbb{F}_n \cup_0 \overline{\mathbb{F}}_n,\Theta) \simeq H^1(\mathbb{F}_n, \Theta_{\mathbb{F}_n,L}) \oplus H^1
(\overline{\mathbb{F}}_n,\Theta_{\overline{\mathbb{F}}_n,\overline L}) \simeq H^1(\mathbb{F}_n, \Theta_{\mathbb{F}_n})
\oplus H^1 (\overline{\mathbb{F}}_n,\Theta_{\overline{\mathbb{F}}_n}),
\end{gather*}
and the composition of the Kodaira--Spencer map $\alpha'(0)$ with these two isomorphisms is an injection onto the real
locus of the last direct sum, because from the construction of the family~\eqref{rsl1}, if we further take the
composition with the projection to the f\/irst factor $H^1(\mathbb{F}_n, \Theta_{\mathbb{F}_n})$ of the last direct sum,
we obviously obtain the Kodaira--Spencer map~\eqref{ks001}, which is an isomorphism.
This means that, in the neighborhood of the origin, the family~\eqref{rsl1} is isomorphic to the real locus of the
Kuranishi family of $\mathbb{F}_n\cup_0\overline{\mathbb{F}}_n$.
Since the isomorphism $D\cup\overline D \simeq \mathbb{F}_n\cup\overline{\mathbb{F}}_n$ respects the real structure, the
last real locus is exactly $\mathscr J$, as desired.
\end{proof}
Now we are able to prove our main result, concerning extendability of the LeBrun metric on $\mathscr O(-n)$ to all
nearby f\/ibers of the above family $\mathscr A_n\to \mathbb{C}^{n-1}$ as an ALE SFK metric:
\begin{Theorem}
\label{thm:main01}
The LeBrun metric on $\mathscr O(-n)$ extends smoothly to all nearby fibers of the family $\mathscr
A_n\to\mathbb{C}^{n-1}$ in~\eqref{af6} of affine $\mathbb{C}$-bundles, as an ALE SFK metric.
\end{Theorem}
\begin{proof}
As in Lemma~\ref{lemma:isom4}, via the induced map $\alpha$, the family $\mathscr F_n\cup\overline{\mathscr
F}_n\to\mathbb{C}^{n-1}$ is isomorphic to the family over $\mathscr J$.
Moreover, as we have already seen, the family over $\mathscr J$ is isomorphic to the $(D\cup\overline D)$-portion of the
deformation of the pair $(Z,D\cup\overline D)$ parameterized by $\mathscr K'$ via the induced map $\psi_2$.
Furthermore, the $Z$-portion of the family of pairs over $\mathscr K'$ is identif\/ied with the family of twistor spaces
over the K\"ahler locus $\mathscr K$ via the map $\psi_1$.
(See the diagram~\eqref{diagram02}.) By the theorem of Pontecorvo~\cite{Pont92}, for any point of $\mathscr K$, the
corresponding twistor space determines an SFK metric on the 4-manifold $\mathscr O(-n)$ up to overall constants.
These SFK metrics can be made to be ALE by multiplying overall constant for each metrics, because the af\/f\/ine bundles
we are considering have a~compactif\/ication by a~$(+n)$-curve.
Via the isomorphisms $\psi_1$, $\psi_2$ and $\alpha$, we conclude that all f\/ibers of the family $\mathscr
A_n\to\mathbb{C}^{n-1}$ admit ALE SFK metrics at least as long as the f\/ibers are suf\/f\/iciently close to the central f\/iber.
The smoothness for the variation of the metrics immediately follows from smoothness for $\psi_1$, $\psi_2$ and $\alpha$.
\end{proof}
We note that as in the above proof, the ALE SFK metrics on all nearby f\/ibers of the central f\/iber are uniquely
determined up to overall constants once we f\/ix the maps $\psi_1$, $\psi_2$ and $\alpha$.
From Theorem~\ref{thm:main01} it is immediate to prove the existence of an ALE SFK metric on any af\/f\/ine
$\mathbb{C}$-bundle over $\mathbb{CP}^1$ of negative degree (see Def\/inition~\ref{def:degree}).
For this, we recall that as we have explained in Section~\ref{ss:ab}, any af\/f\/ine $\mathbb{C}$-bundle over
$\mathbb{CP}^1$ of degree~$-n$ $({\le}-1)$ is a~member of the family $\mathscr A_{n}\to\mathbb{C}^{n-1}$.
Also, from the $\mathbb{C}^*$-action on the total space of $\mathscr A_n$ which is a~lift of the scalar multiplication
on $\mathbb{C}^{n-1}$, any f\/iber over the same line through the origin is mutually biholomorphic except over the
origin.
Thus for any sequence $\{U_{-n}\,|\, n\ge 1\}$ of neighborhoods of the origin in $\mathbb{C}^{n-1}$, the union
$\cup_{n\ge 1} U_{-n}$ contains arbitrary af\/f\/ine $\mathbb{C}$-bundles over $\mathbb{CP}^1$ of negative degree.
Hence by Theorem~\ref{thm:main01} we obtain
\begin{Corollary}\label{cor:exaf}
Any affine $\mathbb{C}$-bundle over $\mathbb{CP}^1$ of negative degree $($see Definition~{\rm \ref{def:degree})} admits an
ALE SFK metric.
\end{Corollary}
Also now it is easy to show the following rigidity result for the LeBrun metric on $\mathscr O(-n)$ when the complex
structure is f\/ixed:
\begin{Proposition}
\label{prop:rigid}
Let $k>0$ and $\Delta$ be a~unit disk in $\mathbb{R}^k$ around the origin, and let $\{g_t\,|\, t\in\Delta\}$ be
a~smooth family of ALE SFK metrics on the complex surface $\mathscr O(-n)$ equipped with the natural complex structure
as a~line bundle.
Assume that $g_0$ is isometric to the LeBrun metric.
Then there exists a~neighborhood $\Delta' \subset \Delta$ of the origin, such that $g_t$ is isometric to the LeBrun
metric up to overall constants for any $t\in\Delta'$.
\end{Proposition}
\begin{proof}
For each $t\in \Delta$ we take a~conformal compactif\/ication $\hat g_t$ of $g_t$ to $\widehat{\mathscr O(-n)}$.
Let $Z_t$ be the twistor space of $\hat g_t$ and $F_t$ be the line bundle $K_{Z_t}^{-1/2}$.
Then by the assumption for complex structure on $\mathscr O(-n)$, for any $t\in \Delta$, the twistor space $Z_t$ has
a~Cartier divisor $D_t\cup \overline D_t(\simeq \mathbb{F}_n\cup_0\overline{\mathbb{F}}_n)$ in the system $|F_t|$
which is biholomorphic to the divisor $D\cup \overline D$ in Proposition~\ref{prop:isom1}.
Hence the family $\{(Z_t,D_t\cup\overline D_t)\,|\, t\in\Delta\}$ gives a~locally trivial deformation of the pair
$(Z_0,D_0 \cup \overline D_0)$ for which the complex structure of $D_0\cup \overline D_0$ does not vary.
By versailty of the Kuranishi family for locally trivial deformations of the pair $(Z_0,D_0\cup \overline D_0)$, there
exist a~neighborhood $\Delta'\subset\Delta$ of the origin and a~smooth map $\varphi:\Delta'\to H^1(\Theta_{Z_0,D_0\cup
\overline D_0})$ which satisf\/ies $\varphi(0) = 0$ and whose pullback of the Kuranishi family is isomorphic to the
original family $\{(Z_t,D_t\cup \overline D_t)\,|\, t\in\Delta'\}$.
But because of the constancy $D_t\cup\overline D_t\simeq D_0\cup \overline D_0$ and the natural isomorphism $H^1 (
\Theta_{Z, D \cup \overline D} ) \simeq H^1 (\Theta_{ D \cup \overline D} )$ in~\eqref{basicisom1}, $\varphi$ has
to satisfy $\varphi(t)=0$ for any $t\in\Delta'$.
This means that the family $\{Z_t\,|\, t\in \Delta'\}$ itself is a~trivial family.
Hence the conformal classes $[\hat g_t]$ do not vary.
This means the required rigidity of the LeBrun's K\"ahler metric.
\end{proof}
Next we take group actions into account for the moduli problem.
Since LeBrun's metric on~$\mathscr O(-n)$ is ${\rm{U}}(2)$-invariant, its twistor space $Z$ admits
a~${\rm{U}}(2)$-action and the divisor $D\cup\overline D$ is ${\rm{U}}(2)$-invariant.
Hence the cohomology group $H^1(\Theta_Z)$ and $H^1(\Theta_{Z,D\cup\overline D})$ have natural ${\rm{U}}(2)$-actions.
The action on $H^1(\Theta_Z)$ was computed in~\cite{HonCMP2}, and if~$H^1(\Theta_Z)^{\sigma}$ denotes the relevant real
locus, we have, as a~real ${\rm{U}}(2)$-module,
\begin{gather*
H^1(\Theta_Z)^{\sigma}\simeq S^{n-2}_1\mathbb{C}^2 \oplus S^{n-4}_2\mathbb{C}^2.
\end{gather*}
Here, $S^m_k\mathbb{C}^2:= S^m\mathbb{C}^2\otimes_{\mathbb{C}}\mathbb{C}_k$, where $S^m\mathbb{C}^2$ denotes the $m$-th
symmetric product of the natural representation on $\mathbb{C}^2$, and $\mathbb{C}_k$ is the 1-dimensional
representation obtained by multiplying $(\det)^k$.
(If $m < 0$, $S^m\mathbb{C}^2$ means $0$, and $S^0\mathbb{C}^2$ means the trivial representation on $\mathbb{C}$.) For
the ${\rm{U}}(2)$-action on $H^1(\Theta_{Z,D+\overline D})$, it is immediate from Proposition~\ref{prop:isom1} to derive
the following
\begin{Proposition}
As a~real ${\rm{U}}(2)$-module, we have $H^1(\Theta_{Z,D\cup\overline D})^{\sigma} \simeq S^{n-2}_1\mathbb{C}^2$.
\end{Proposition}
\begin{proof}
By the injectivity of the natural map~\eqref{forget}, $H^1(\Theta_{Z,D\cup\overline D})^{\sigma}$ is naturally
a~subspace of $H^1(\Theta_{Z})^{\sigma}$ which is of course ${\rm{U}}(2)$-invariant.
Since both $S^{n-2}_1\mathbb{C}^2 $ and $S^{n-4}_2\mathbb{C}^2$ are irreducible ${\rm{U}}(2)$-modules,
$H^1(\Theta_{Z,D\cup\overline D})^{\sigma}$ has to coincide with one of these two spaces or the whole space
$S^{n-2}_1\mathbb{C}^2 \oplus S^{n-4}_2\mathbb{C}^2$.
But it has to be $S^{n-2}_1\mathbb{C}^2$ as $\dim H^1(\Theta_{Z,D\cup\overline D})^{\sigma}=2(n-1)$ from
Proposition~\ref{prop:isom1} while $\dim_{\mathbb{R}}S_2^{n-4}\mathbb{C}^2 = 2(n-3)\neq 2(n-1)$.
\end{proof}
Thus connecting the series of the natural isomorphisms
\begin{gather}
\label{isom49}
H^1(\Theta_{Z,D\cup\overline D})^{\sigma}
\; \stackrel{\eqref{basicisom1}}{\simeq}
\;
H^1(\Theta_{D\cup\overline D})^{\sigma}
\; \stackrel{\eqref{isom001}\&\eqref{isom002}}{\simeq}
\;
H^1(\Theta_D)
\;\stackrel{\eqref{ks001}}{\simeq}
\;
\mathbb{C}^{n-1},
\end{gather}
we obtain that the parameter space $\mathbb{C}^{n-1}$ of the family $\mathscr A_n\to\mathbb{C}^{n-1}$ may be
identif\/ied with $S^{n-2}\mathbb{C}^2$ as a~real ${\rm{U}}(2)$-module.
We also have the following result concerning ${\rm{U}}(2)$-action on another cohomology group.
Recall from Proposition~\ref{prop:isom1} that we have $H^1(\Theta_Z(-D-\overline D))\simeq\mathbb{C}$.
\begin{Proposition}
For the LeBrun metric, the natural ${\rm{U}}(2)$-action on $H^1(\Theta_Z(-D-\overline D))\simeq\mathbb{C}$ is trivial.
\end{Proposition}
\begin{proof}
As in~\eqref{LB:key} we have the exact sequence
\begin{gather}
\label{les:10}
0
\longrightarrow
H^0(\Theta_{Z, D\cup\overline D}) \big({\simeq}\, \mathbb{C}^4\big)
\;\stackrel{\iota}{\longrightarrow}
\;
H^0(\Theta_{D\cup\overline D})
\big({\simeq}\, \mathbb{C}^5\big)
\longrightarrow
H^1( \Theta_Z(-D-\overline D))
\longrightarrow
0,
\end{gather}
and hence the space $H^1(\Theta_Z(-D-\overline D))$ $(\simeq\mathbb{C})$ can be identif\/ied with the cokernel of the
injection~$\iota$.
In the space $H^0(\Theta_{D\cup \overline D})$ we have the 2-dimensional subspace generated by the scalar multiplication
on each of $D$ and $\overline D$, where we are viewing these as the compactif\/ication of the line bundle $\mathscr
O(-n)$ for the scalar multiplication.
As the scalar multiplications commute with any element of ${\rm{U}}(2)$, the group ${\rm{U}}(2)$ acts trivially on this
2-dimensional subspace.
Moreover the image of $\iota$ cannot contain this subspace since any real element of $H^0(\Theta_{Z,D\cup\overline D})$
is a~lift of a~conformal Killing f\/ield on $\widehat{\mathscr O(-n)}$ and hence on $D$ and $\overline D$ the vector
f\/ield cannot move independently each other.
This means that the 2-dimensional subspace of $H^0(\Theta_{D\cup\overline D})$ is mapped surjectively to the
1-dimensional space $H^1(\Theta_Z(-D-\overline D))$.
Since the sequence~\eqref{les:10} is ${\rm{U}}(2)$-equivariant, the assertion follows.
\end{proof}
Next we investigate the restrictions of the family of ALE SFK metrics in Theorem~\ref{thm:main01} to the coordinate axes
of $\mathbb{C}^{n-1}$, which provide ${\rm{U}}(1)$-equivariant deformations of the LeBrun metric.
Suppose $n\ge 3$ and for each $1\le l\le n-1$ let $\mathbb{C}(t_l)$ be the $l$-th coordinate axis of $\mathbb{C}^{n-1}$
as before, and let
\begin{gather}\label{1para}
\mathscr A_{n,l}\to \mathbb{C}(t_l)
\qquad
\text{and}
\qquad
\mathscr F_{n,l}\to \mathbb{C}(t_l)
\end{gather}
be the restrictions of the families $\mathscr A_n\to\mathbb{C}^{n-1}$ and $\mathscr F_n\to\mathbb{C}^{n-1}$ respectively
to the $t_l$-axis.
From the $\mathbb{C}^*$-action which is a~lift of the scalar multiplication on $\mathbb{C}^{n-1}$, all f\/ibers of
$\mathscr A_{n,l}\to \mathbb{C}(t_l)$ and $\mathscr F_{n,l}\to \mathbb{C}(t_l)$ are mutually isomorphic except the
central f\/iber for each, and as in~\eqref{zeta'}, f\/ibers of $\mathscr F_{n,l}\to \mathbb{C}(t_l)$ are isomorphic to
$\mathbb{F}_{n-2l}$ except the f\/iber over the origin.
Moreover recalling that~$\mathscr A_n$ is def\/ined by the equation
\begin{gather*
\zeta_0 = \frac 1{u^n} \zeta_1 + \sum\limits_{l=1}^{n-1} \frac {t_l} {u^l}
\qquad
\text{on}
\quad
U_{01},
\end{gather*}
as in~\eqref{af5}, we obtain that, as an enlargement of the above $\mathbb{C}^*$-action on $\mathscr A_n$ (and $\mathscr
F_n$), the total space of the family $\mathscr A_n$ (and $\mathscr F_n$) carries
a~$(\mathbb{C}^*\times\mathbb{C}^*)$-action def\/ined by
\begin{gather}
\label{taction}
(u,\zeta_0,t_l)
\;\stackrel{(s_1,s_2)}{\longmapsto}
\;
(s_1u,s_2\zeta_0,s_1^ls_2t_l),
\qquad
(s_1,s_2)\in\mathbb{C}^*\times\mathbb{C}^*.
\end{gather}
On the central f\/iber this gives a~$(\mathbb{C}^*\times\mathbb{C}^*$)-action of the toric structure on $\mathscr O(-n)$
or $\mathbb{F}_n$.
Putting $s_1=0$ in~\eqref{taction} gives the original $\mathbb{C}^*$-action on $\mathscr A_n$ and $\mathscr F_n$.
If we def\/ine a~$\mathbb{C}^*$-subgroup $G_l$ of $\mathbb{C}^*\times\mathbb{C}^*$ by
\begin{gather}
\label{G_l}
G_l:=\big\{(s_1,s_2)\in\mathbb{C}^*\times\mathbb{C}^*\,|\, s_1^ls_2=1\big\},
\end{gather}
which acts trivially on $\mathbb{C}(t_l)$, then the two families~\eqref{1para} may be regarded as $G_l$-equivariant
deformations of $\mathscr O(-n)$ and $\mathbb{F}_n$ respectively.
Then basically by restricting the family of ALE SFK metrics in Theorem~\ref{thm:main01} to the coordinate axis
$\mathbb{C}(t_l)$, we obtain the following result about existence of ${\rm{U}}(1)$-equivariant deformations of the
LeBrun metric:
\begin{Proposition}
\label{prop:invm}
Let $n\ge 3$ and $l\ge 1$ be integers satisfying $n-2l\ge 0$, and $\mathscr A_{n,l} \to\mathbb{C}(t_l)$ be the
$1$-parameter deformation of $\mathscr O(-n)$ to an affine $\mathbb{C}$-bundle as above.
Then the LeBrun metric on the central fiber $\mathscr O(-n)$ extends smoothly to any nearby fibers of $\mathscr
A_{n,l}\to\mathbb{C}(t_l)$ preserving not only the ALE SFK property but also a~${\rm{U}}(1)$-action, where ${\rm{U}}(1)$
is the compact torus of the stabilizer subgroup $G_l(\simeq\mathbb{C}^*)$ of the axis $\mathbb{C}(t_l)$ defined in~\eqref{G_l}.
\end{Proposition}
\begin{proof}
From the proof of Theorem~\ref{thm:main01}, we just need to show that the isomorphism between the K\"ahler locus
$\mathscr K$ and the parameter space of the family $\mathscr F_n\cup \overline{\mathscr F}_n\to\mathbb{C}^{n-1}$ around
the origin can be taken to be $T^2$-equivariant, where $T^2$ is the standard maximal torus of ${\rm{U}}(2)$ consisting
of diagonal matrices.
For this it is enough to see that the induced maps $\psi_1$, $\psi_2$ and $\alpha$, which were used to identify the
relevant families in Section~\ref{ss:DLm}, can be taken to be $T^2$-equivariant.
This holds for $\psi_1$ and $\psi_2$ since $Z$, $D\cup\overline D$ and the pair $(Z,D\cup\overline D)$ are
${\rm{U}}(2)$-invariant.
For $T^2$-equivariance of the remaining map $\alpha$, it is enough to see that the total space of the family $\mathscr
F_n\cup\overline{\mathscr F}_n\to\mathbb{C}^{n-1}$ has a~$T^2$-action whose restriction to the central f\/iber is
identif\/ied with the $T^2$-action on $D\cup\overline D$.
Each component of $\mathscr F_n\cup\overline{\mathscr F}_n$ has a~$T^2$-action which is the restriction of the
$(\mathbb{C}^*\times\mathbb{C}^*)$-action given in~\eqref{taction} to the maximal torus, and these are clearly
identif\/ied with the $T^2$-actions on $D$ and $\overline D$ respectively.
So to complete the proof we just need to see that the gluing map which was used for making the family $\mathscr
F_n\cup\overline{\mathscr F}_n$ is $T^2$-equivariant.
But this is immediate if we notice that the anti-podal map $\tau_0$ commutes with the $T^2$-actions.
\end{proof}
\begin{Remark}
The parameter space $\mathbb{C}(t_l)$ of the above family of the metric is of course real 2-dimensional, but the family
is in ef\/fect real 1-dimensional by the following reason.
As in the above proof, the axis $\mathbb{C}(t_l)$ is naturally identif\/ied (via the isomorphisms in~\eqref{isom49})
with a~$T^2$-invariant subspace of $H^1(\Theta_{Z,D\cup \overline D})^{\sigma}$ on which the subgroup $G_l$ acts
trivially.
This $T^2$-action has clearly real 1-dimensional orbits, and along each orbit the complex structure of the pairs of
twistor spaces and the divisors are constant.
We also note that although the group $T_{\mathbb{C}}$ (and $T_{\mathbb{C}}/G_l$) is acting on the axis
$\mathbb{C}(t_l)\subset\mathbb{C}^{n-1}$, the corresponding subspace is {\em not} $T_{\mathbb{C}}$- (nor
$T_{\mathbb{C}}/G_l$-) invariant, because the isomorphism in~\eqref{isom49} are not $T_{\mathbb{C}}$-equivariant and
just $T^2$-equivariant.
\end{Remark}
By Corollary~\ref{cor:unique} and Proposition~\ref{prop:invm}, we obtain the following
\begin{Proposition}
Let $n\ge 0$, $l\ge 0$, and let $L\in |\Gamma_0+(n+l)f|$ be any $\mathbb{C}^*$-invariant section of
$\mathbb{F}_n\to\mathbb{CP}^1$, where $\mathbb{C}^*$ acts non-trivially on~$L$.
Then the complement $\mathbb{F}_n\backslash L$ admits a~${\rm{U}}(1)$-invariant ALE SFK metric, where ${\rm{U}}(1)$ is
the compact torus of $\mathbb{C}^*$.
\end{Proposition}
\begin{proof}
By Corollary~\ref{cor:unique}, the complex structure of the pair $(\mathbb{F}_n,L)$ satisfying the properties in the
proposition is uniquely determined from $n$ and $l$.
If $l=0$, we have $\mathbb{F}_n\backslash L\simeq\mathscr O(-n)$ and the existence of the metric on it is guaranteed by
the original LeBrun metric on $\mathscr O(-n)$.
If $l>0$, $\mathbb{F}_n\backslash L$ is biholomorphic to general f\/ibers of the 1-parameter family $\mathscr
A_{n+2l,l}\to\mathbb{C}(t_l)$, and the existence of the metric is guaranteed by Proposition~\ref{prop:invm}, as long as
$n+2l\ge 3$.
The situation where $n+2l\ge 3$ does not hold is only the case $(n,l) = (0,1)$.
But the existence of an ALE SFK metric on $\mathbb{F}_0\backslash L$ ($L\in|\mathscr O(1,1)|$) is guaranteed by the
Eguchi--Hanson metric.
\end{proof}
When $l>1$, if~$L$ is the $\mathbb{C}^*$-invariant section as in the above proposition, then we have
${\rm{Aut}}_0(\mathbb{F}_n,L) \simeq\mathbb{C}^*$ by Proposition~\ref{prop:l2}.
But when $l=1$, we have ${\rm{Aut}}_0(\mathbb{F}_{n}, L)\simeq \rm{Af}(\mathbb{C})$ by Proposition~\ref{prop:nonred}.
Thus the af\/f\/ine surface $\mathbb{F}_n\backslash L$ admits an ALE SFK metric with an ef\/fective ${\rm{U}}(1)$-action
but ${\rm{Aut}}_0(\mathbb{F}_{n}, L)$ is not reductive.
But we do not know whether the holomorphic transformation group of the complex surface $\mathbb{F}_n\backslash L$ itself
is reductive, nor even whether it is of f\/inite dimensional.
In this respect, for the surface $\mathscr O(-n)$, the holomorphic automorphism group is known to be {\em not} of
f\/inite dimensional~\cite[Remark 2.20]{KPZ}.
\subsection{Deformations of the metrics on the af\/f\/ine bundles}
In the last subsection we have obtained ALE SFK metrics on af\/f\/ine $\mathbb{C}$-bundles over $\mathbb{CP}^1$ as small
deformations of the LeBrun metric on $\mathscr O(-n)$.
In this subsection we in turn investigate small deformations of these metrics on af\/f\/ine bundles, which again
preserve ALE SFK property.
So let $\mathscr A_n\to\mathbb{C}^{n-1}$ ($n\ge 3$) be the family~\eqref{af6} of af\/f\/ine bundles over $\mathbb{CP}^1$
as before, and for each $t\in\mathbb{C}^{n-1}$ write $A_t$ for the af\/f\/ine bundle lying over $t$.
By Theorem~\ref{thm:main01}, if $t$ is suf\/f\/iciently close to the origin, the af\/f\/ine bundle $A_t$ admits an ALE
SFK metric.
We write $g_t$ for this metric.
We recall that these metrics are uniquely determined up to overall constants by the (natural but non-unique) maps
$\psi_1$, $\psi_2$ and $\alpha$.
Then these metrics satisfy the following property:
\begin{Proposition}\label{prop:eg}
If $t\neq 0$, the metric $g_t$ on the affine bundle $A_t$ is a~member of a~non-trivial $($see below$)$, real $1$-parameter
family of ALE SFK metrics, in which the complex structure on $A_t$ does not deform.
\end{Proposition}
Here, `non-trivial family' means that the complex structures of the corresponding 1-parameter family of twistor spaces
actually vary as the parameter moves.
Thus the situation is in contrast with the LeBrun metric on the line bundle $\mathscr O(-n)$, for which the metric
cannot be deformed as an ALE SFK metric when the complex structure is f\/ixed (see Proposition~\ref{prop:rigid}).
\begin{proof}[Proof of Proposition~\ref{prop:eg}] Let $Z_t$ be the twistor space of a~conformal compactif\/ication of~$g_t$, and
let~$D_t$ be the divisor determined by the complex structure of~$A_t$.
The sum $D_t+\overline D_t$ is a~Cartier divisor belonging to $|K_{Z_t}^{-1/2}|$.
We f\/irst show that the natural map
\begin{gather}
\label{surj}
H^0(\Theta_{Z_t,D_t\cup\overline D_t})
\longrightarrow
H^0(\Theta_{D_t\cup\overline D_t})
\end{gather}
is surjective as long as $t\neq 0$ and $t$ is suf\/f\/iciently close to $0$.
(Note that if $t=0$ this is not surjective as in the sequence~\eqref{les:1}.) This is trivially satisf\/ied if
$H^0(\Theta_{D_t\cup\overline D_t}) = 0$.
If $H^0(\Theta_{D_t\cup\overline D_t}) \neq 0$, we clearly have $H^0(\Theta_{D_t,L_t}) \neq 0$, where
$L_t=D_t\cap\overline D_t$.
From Propositions~\ref{prop:l1},~\ref{prop:nonred} and~\ref{prop:l2}, this is the case only when $(D_t,L_t)$ is
a~$\mathbb{C}^*$-invariant pair.
Again by the same propositions, the complex structure of such a~pair is unique once the two integers $m\ge 0$ and $l\ge
1$ are f\/ixed for which $D_t\simeq\mathbb{F}_m$ and $L\in |\Gamma_0 + (m+l)f|$ hold.
This means that we have $D_t\cup \overline D_t\simeq \mathbb{F}_{n-2l}\cup_l\overline{\mathbb{F}}_{n-2l}$ for some $l\ge
1$ (as $t\neq 0$) in the notation of Def\/inition~\ref{def:scn}.
For the central f\/iber $(Z_0,D_0\cup\overline D_0)$, as in~\eqref{basicisom1}, we have a~natural isomorphism
\begin{gather}
\label{edisom}
H^1( \Theta_{Z_0,D_0\cup\overline D_0}) \simeq H^1(\Theta_{D_0\cup\overline D_0}),
\end{gather}
and this is ${\rm{U}}(2)$-equivariant.
Since $H^2( \Theta_{Z_0,D_0\cup\overline D_0}) = H^2(\Theta_{D_0\cup\overline D_0})=0$ as
in~\eqref{van1},~\eqref{edisom} means that, for any subgroup $G\subset {\rm{U}}(2)$, $G$-action on the surface
$D_0\cup\overline D_0$ extends to $D_t\cup\overline D_t$ for suf\/f\/iciently small $t$ if and only if the $G$-action on
the pair $(Z_0,D_0\cup\overline D_0)$ extends to $(Z_t,D_t\cup\overline D_t)$.
Moreover the restriction of the family $\mathscr F_n\cup\overline{\mathscr F}_n\to\mathbb{C}^{n-1}$ given
in~\eqref{rsl1} to the coordinate axis~$\mathbb{C}(t_l)$ actually connects $D_0\cup\overline D_0$ and $D_t\cup\overline
D_t$ in a~${\rm{U}}(1)$-equivariant way, where ${\rm{U}}(1)$ is the compact torus of~$G_l$ def\/ined in~\eqref{G_l}.
Thus the ${\rm{U}}(1)$-action on $(Z_0,D_0\cup\overline D_0)$ actually extends to $(Z_t,D_t\cup\overline D_t)$.
This means that the map~\eqref{surj} is surjective for suf\/f\/iciently small $t$.
By the upper semi-continuity and the invariance of the Euler characteristic under deformation, we have
\begin{gather*
H^i(Z_t,\Theta_{Z_t}(-D_t-\overline D_t)) =0
\qquad
\text{if}
\quad
i\neq 1,
\qquad
\text{and}
\qquad
H^1(Z_t,\Theta_{Z_t}(-D_t-\overline D_t)) \simeq\mathbb{C},
\end{gather*}
because these are true for the case of the LeBrun metric as in~\eqref{cohomdim1}.
Therefore from the exact sequence~\eqref{ses:1} (which remains obviously valid for $(Z_t,D_t\cup\overline D_t)$) and the
surjectivity of the map~\eqref{surj}, we obtain the short exact sequence
\begin{gather}
\label{ses:10}
0
\longrightarrow
H^1(\Theta_{Z_t}(-D_t-\overline D_t))
(\simeq\mathbb{C})
\;\stackrel{\alpha}{\longrightarrow}
\;
H^1(\Theta_{Z_t, D_t\cup\overline D_t})
\;\stackrel{\beta}{\longrightarrow}
\;
H^1(\Theta_{D_t\cup \overline D_t})
\longrightarrow
0.
\end{gather}
Then as we have $H^2(\Theta_{Z_t,D_t+\overline D_t}) = 0$ by~\eqref{van1} and the upper semi-continuity again, the
parameter space of the Kuranishi family for locally trivial deformations of the pair $(Z_t,D_t\cup\overline D_t)$ can be
regarded as a~small disk (for which we denote by $\Delta_1$) about the origin in $H^1(\Theta_{Z_t,D_t\cup\overline
D_t})$.
Similarly as we have $H^2(\Theta_{D_t\cup\overline D_t})=0$ from~\eqref{van1}, the Kuranishi family for $D_t\cup
\overline D_t$ may be regarded as a~small disk (for which we denote by $\Delta_2$) about the origin in
$H^1(\Theta_{D_t\cup\overline D_t})$.
If $\varphi:\Delta'_1\to\Delta_2$ denotes a~holomorphic map from a~possibly smaller disk $\Delta'_1\subset\Delta_1$
which is induced by the versality of the Kuranishi family for $D_t\cup\overline D_t$, then we naturally have
$\varphi'(0) = \beta$.
Hence from the surjectivity of $\beta$ in~\eqref{ses:10}, $\varphi$ is a~submersion at least in a~neighborhood of the
origin, and therefore $\varphi^{-1}(0)$ is non-singular and 1-dimensional near the origin, at which the tangent space is
exactly the line $\Image(\alpha)$.
Thus the Kuranishi family for locally trivial deformations of the pair $(Z_t,D_t\cup \overline D_t)$ contains
a~(complex) 1-parameter subfamily over which the complex structure of $D_t\cup\overline D_t$ is constant.
By restricting to the real locus of $\varphi^{-1}(0)$, we obtain the real 1-parameter family of deformation of the pair
$(Z_t,D_t\cup \overline D_t)$ for which the complex structure of $D_t\cup\overline D_t$ is constant.
By the same reason as in the proof of Theorem~\ref{thm:main01}, the ASD structure determined by $(Z_t,D_t\cup\overline
D_t)$ has a~K\"ahler representative which is ALE at inf\/inity.
It remains to show that the 1-parameter family of ALE SFK metrics on the 4-manifold $\mathscr O(-n)$ thus obtained is
non-trivial in the sense explained right after the proposition.
For this, it suf\/f\/ices to see that the Kodaira--Spencer class of the 1-parameter family of $Z_t$ is non-zero, because
this means that the complex structure of $Z_t$ actually deforms.
From the above argument the Kodaira--Spencer class of the deformation of the pair $(Z_t,D_t\cup\overline D_t)$ is
non-zero (belonging to $\Image(\alpha)$), and the genuine Kodaira--Spencer class (belonging to $H^1(\Theta_{Z_t})$) is
obtained from this by sending it under the natural map $H^1(\Theta_{Z_t,D_t\cup\overline D_t})\to H^1(\Theta_{Z_t})$.
We show that the last map is injective in the same way to the last part in the proof of Proposition~\ref{prop:isom1}.
Exactly by the same argument as to deduce~\eqref{N''}, for the cokernel sheaf $N'_t$ of the natural injection
$\Theta_{Z_t,D_t\cup\overline D_t} \to \Theta_{Z_t}$, we obtain
\begin{gather*}
N_t'|_{D_t}\simeq -K_{D_t}-\mathscr O_{D_t}(2L_t),
\qquad
N_t'|_{\overline D_t}\simeq -K_{\overline D_t}-\mathscr O_{\overline D_t}(2L_t).
\end{gather*}
Now as $t\neq 0$ we can write $D_t\simeq\mathbb{F}_{n-2l}$ for some $l$ satisfying $1\le l\le n-1$, and we have $L_t\in
|\Gamma_0+(n-l)f|$.
Hence by the same computation to deduce~\eqref{N'} we obtain
\begin{gather*
-K_{D_t}-\mathscr O_{D_t}(2L_t) \simeq 2\Gamma_0+(n-2l+2)f -2\{\Gamma_0 + (n-l)f\}
\simeq -(n-2)f.
\end{gather*}
Thus as $n\!>\!2$ we obtain $H^0(-K_{D_t}\!-\mathscr O_{D_t}(2L_t))\!=\!0$.
With reality, this means \mbox{$H^0(D_t\!\cup\! \overline D_t,N'_t)\!=\!0$}.
Hence the injectivity of the map $H^1(\Theta_{Z_t,D_t\cup\overline D_t})\to H^1(\Theta_{Z_t})$ follows, and we obtained
the assertion of the proposition.
\end{proof}
\begin{proof}[Proof of Proposition~\ref{prop:var}]
Let $\mathscr Z\to B$ be the family of twistor spaces associated to ALE SFK
metrics on nearby f\/ibers for the central f\/iber of the family $\mathscr A_n\to \mathbb{C}^{n-1}$ of af\/f\/ine
$\mathbb{C}$-bundle over~$\mathbb{CP}^1$.
For $t\in B$ we write $Z_t$ for the twistor space over the af\/f\/ine bundle $A_t$.
Since the family $\mathscr Z\to B$ is versal at the origin, it is also versal at $t\in B$ as long as $t$ is
suf\/f\/iciently close to the origin.
Therefore for a~real 1-parameter family of twistor spaces associated to the family of ALE SFK metrics obtained in
Proposition~\ref{prop:eg}, there exists an induced map from the parameter space of the last family to~$B$.
Then if we take the image of the map as the arc, it clearly satisf\/ies the property of the proposition.
\end{proof}
\subsection*{Acknowledgements}
I would like to thank Jef\/f Viaclovsky for numerous useful discussion from which this work has originated.
Also I would like to thank Akira Fujiki for invaluable comments on deformations of line bundles.
This work was supported by JSPS KAKENHI Grant Number 24540061.
\pdfbookmark[1]{References}{ref}
|
\subsection*{{\normalsize\b#1}}}
\renewcommand{\t}{^{\mbox{\tiny\sf T}}}
\newcommand{^{\mbox{\tiny\sf T}}}{^{\mbox{\tiny\sf T}}}
\newcommand{\ensuremath{\mathcal{B}_{\varepsilon}}}{\ensuremath{\mathcal{B}_{\varepsilon}}}
\newcommand{\ensuremath{\textup{int}\,}}{\ensuremath{\textup{int}\,}}
\newcommand{\tu}[1]{\,\,\textup{#1}\,\,}
\newcommand{\ensuremath{\textup{cl}\,}}{\ensuremath{\textup{cl}\,}}
\newcommand{\ensuremath{\textup{bd}\,}}{\ensuremath{\textup{bd}\,}}
\newcommand{\ensuremath{x_{0}}}{\ensuremath{x_{0}}}
\newcommand{\ensuremath{\operatorname{co}}}{\ensuremath{\operatorname{co}}}
\newcommand{\ensuremath{\operatorname{cone}}}{\ensuremath{\operatorname{cone}}}
\newcommand{\ensuremath{\textup{sgn}\,}}{\ensuremath{\textup{sgn}\,}}
\newcommand{\varepsilon}{\varepsilon}
\newcommand{\ensuremath{\operatorname{aff}}}{\ensuremath{\operatorname{aff}}}
\newcommand{\ensuremath{\operatorname{rank}}}{\ensuremath{\operatorname{rank}}}
\newcommand{\begin{enumerate}}{\begin{enumerate}}
\newcommand{\end{enumerate}}{\end{enumerate}}
\newcommand{\newcommand}{\newcommand}
\newcommand{\renewcommand}{\renewcommand}
\newcommand{\begin{verbatim}}{\begin{verbatim}}
\newcommand{\end{verbatim}}{\end{verbatim}}
\DeclareMathOperator{\fl}{\mathcal{F}}
\DeclareMathOperator{\sat}{Sat}
\DeclareMathOperator{\Co}{Co}
\newcommand{\textit}{\textit}
\newcommand{\llbracket}{\llbracket}
\newcommand{\rrbracket}{\rrbracket}
\newcommand{\Rightarrow}{\Rightarrow}
\newcommand{\Leftarrow}{\Leftarrow}
\newcommand{\bq}{\textbf}
\newcommand{\m}{\textrm}
\newcommand{\bb}{\mathbb}
\renewcommand{\bf}{\mathbf}
\newcommand{\til}{\texttildelow}
\newcommand{\be}{\begin{equation}}
\newcommand{\ee}{\end{equation}}
\newcommand{\dps}{\displaystyle}
\renewcommand{\l}{\left(}\renewcommand{\r}{\right)}
\newcommand{\lc}{\left\{}\newcommand{\rc}{\right\}}
\newcommand{\lb}{\left[}\newcommand{\rb}{\right]}
\newcommand{\ba}[1]{\begin{array}{#1}}
\newcommand{\ea}{\end{array}}
\newcommand{\ra}{\rightarrow}
\newcommand{\li}{\left |}
\newcommand{\ri}{\right |}
\newcommand{\pde}[2]{\frac{\partial #1}{\partial #2}}
\newcommand{\ode}[2]{\frac{d #1}{d #2}}
\newcommand{\odee}[3]{\frac{d^{#3} #1}{d #2^{#3}}}
\newcommand{\pdee}[3]{\frac{\partial^{#3} #1}{\partial #2^{#3}}}
\newcommand{\bn}{\begin{enumerate}}
\newcommand{\en}{\end{enumerate}}
\newcommand{\bt}{\begin{theorem}}
\newcommand{\et}{\end{theorem}}
\newcommand{\y}[1]{\lambda_{#1}}
\newcommand{\ir}{\mathbb{I}\mathbb{R}}
\newcommand{\closure}[2][3]{{}\mkern#1mu\overline{\mkern-#1mu#2}}
\newcommand{\ep}{\mathcal{E}_{P}}
\newcommand{\mr}{\mathcal{M}_{r}}
\newcommand{\mfa}{\mathcal{M}_{f,a}}
\newcommand{\mfp}{\mathcal{M}_{f,p}}
\newcommand{\mt}{\m{T}}
\newcommand{\F}{\mathbb{F}}
\setlength{\marginparwidth}{0.7in}
\newcommand{\todoef}[1]{\marginpar{\raggedright\color{blue}\tiny \setstretch{1} [EF]: #1}}
\newcommand{\rpnote}[1]{\marginpar{\raggedright\color{red}\tiny \setstretch{1} [RP]: #1}}
\begin{document}
\newtheorem{proposition}{Proposition}
\newtheorem{thm}{Theorem}
\newtheorem{rmk}{Remark}
\definecolor{cgreen}{rgb}{0,0.6,0}
\lstset{ %
language=C,
basicstyle=\footnotesize\ttfamily,
numbers=left,
numberstyle=\scriptsize,
stepnumber=1,
numbersep=4pt,
commentstyle=\color{cgreen},
backgroundcolor=\color{white},
showspaces=false,
showstringspaces=false,
showtabs=false,
frame=single,
tabsize=2,
captionpos=b,
breaklines=true,
breakatwhitespace=false,
escapeinside={\%}{)}
}
\pagestyle{plain}
\title{\textbf{Verifiable Control System Development for Gas Turbine Engines}}
\author{Mehrdad Pakmehr \footnote{Postdoctoral fellow at the School of Aerospace Engineering, Georgia Institute of Technology, Atlanta, GA 30332, {\small mehrdad.pakmehr@gatech.edu.}},
Timothy Wang \footnote{PhD Candidate at the School of Aerospace Engineering, Georgia Institute of Technology, Atlanta, Georgia 30332, {\small timothy.wang@gatech.edu.}},
Romain Jobredeaux \footnote{PhD Candidate at the School of Aerospace Engineering, Georgia Institute of Technology, Atlanta, Georgia 30332, {\small romain.jobredeaux@gatech.edu.}},
Martin Vivies \footnote{System Engineer at the Price Induction Inc., Marietta, Georgia, 30067, {\small martin.vivies@price-induction.com.}},
Eric Feron \footnote{Professor at the School of Aerospace Engineering, Georgia Institute of Technology, Atlanta, GA 30332, {\small feron@gatech.edu.}} }
\date{\null}
\maketitle
\pagestyle{plain}
\section*{Abstract}
A control software verification framework for gas turbine engines is developed. A stability proof is presented for gain scheduled closed-loop engine system based on global linearization and linear matrix inequality (LMI) techniques. Using convex optimization tools, a single quadratic Lyapunov function is computed for multiple linearizations near equilibrium points of the closed-loop system. With the computed stability matrices, ellipsoid invariant sets are constructed, which are used efficiently for DGEN turbofan engine control code stability analysis. Then a verifiable linear gain scheduled controller for DGEN engine is developed based on formal methods, and tested on the engine virtual test bench. Simulation results show that the developed verifiable gain scheduled controller is capable of regulating the engine in a stable fashion with proper tracking performance.
\section*{Nomenclature}
$NL$: Low Pressure Spool Speed\\
$NH$: High Pressure Spool Speed\\
$W_f$: Fuel Flow Control Input\\
PLA: Power Lever Angle\\
SFC: Specific Fuel Consumption\\
FADEC: Full Authority Digital Engine Control\\
WESTT: Whole Engine Simulator Turbine Technology\\
ECU: Engine Control Unit\\
$\alpha$: Scheduling Parameter\\
\emph{Co}: Convex Hull\\
superscript \emph{c}: Controller\\
superscript \emph{p}: Plant\\
superscript \emph{ol}: Open-loop\\
superscript \emph{cl}: Closed-loop\\
\section{Introduction}
Stability and control of gas turbine engines have been of interest to researchers and engineers from a variety of perspectives. Some of the literature related to engine control can be found in \cite{powerplantControl-sobey-1963, control-spang-1999, enginecontrol-jaw-2009, robustAeroengine-arriffin-1997, lqg-athans-1986, lqg-garg-1989, turbofanControl-fredrick-2000, AdvControl-richter-2012, PhDThesis-pakmehr-2013}. To facilitate the stability analysis of nonlinear systems, such as gas turbine engines, an efficient technique is to approximate them by a linear time-varying (LTV) system. One of the control design approaches, which perhaps is one of the most popular nonlinear control design approaches and has been widely and successfully applied in fields ranging from aerospace to process control \cite{research-rugh-2000, surveyGS-leith-2000}, is gain scheduling. Gas turbine engines are no exception, and research on gain scheduling of gas turbine engines is presented in \cite{kapasouris-gainsched-1985, turbofanSched-garg-1997, lpv-bruzelius-2002, lpv-balas-2002, lpv-gilbert-2010, approximate-zhao-2011, GainSchedStabConf-pakmehr-2013, GSstability-pakmehr-2013}.
When the operation of a control system is highly critical due to human safety factors or the high cost of failure in damaged capital or products, the software designers have to expend more effort to validate and verify their software before it can be released. In flight-critical operations, validation and verification are part of the flight certification process \cite{SoftwareTech-heck-2003}. Software system certification involves many challenges, including the necessity to certify the system at the level of functional requirements, code and binary levels, the need to detect run-time errors, and the need for proving timing properties of the eventual, compiled system \cite{CertifyCont-feron-2007, DO178C-RTCA-2011}.
Provable closed-loop stability constitutes an essential attribute of control systems, especially when human safety is involved, as in many aeronautical systems like gas turbine engines. Motivated by such applications, there exist many theorems to support system stability and performance under various assumptions \cite{ContSoftware-feron-2010}. Stability criteria apply to a class of dynamical systems for which a stability proof is established; and Lyapunov's stability theory plays a critical role in this regard. Control-system domain knowledge, in particular, Lyapunov-theoretic proofs of stability and performance, can be migrated toward computer-readable and verifiable certificates \cite{ContSoftware-feron-2010, autocoding-feron-2011}. Some of the recent research results on the control software verification using Lyapunov proof of stability can be reviewed in \cite{CertifyCont-feron-2007, ContSoftware-feron-2010, autocoding-feron-2011, ProofCode-Jobredeaux-2012, PhdThesis-roozbehani-2008, LyapSoftVerif-roozbehani-2013, GraphEnviron-wang-2011, DesToImpAutomated-wang-2013}.
Software verification process for aerospace systems is explained in ``RTCA /DO-178B: Software Considerations in Airborne Systems and Equipment Certification" \cite{DO178B-RTCA-1992}. Currently there is no detailed theoretical process for software verification; and the verification is mainly performed by running the software long enough to make sure that it works properly for the system at hand. Since the publication of DO-178B \cite{DO178B-RTCA-1992}, experience and scientific advances have been gained in the formal methods, their application, and tools. Formal methods are mathematically based techniques for specification, development, and verification of software aspects of digital systems \cite{DO333-RTCA-2011}. ``RTCA/DO-333: Formal Methods Supplement to DO-178C and DO-278A" \cite{DO333-RTCA-2011} provides guidance for applicants to facilitate the use of formal methods in aerospace systems.
In this paper, we aim at taking the first steps towards a more rigorous software verification process for gas turbine engine control systems, by developing stability proofs for the entire engine control architecture using the Lyapunov stability theory. This approach helps us in constructing an ellipsoid invariant set \cite{lmi-boyd-1994, ellipsoidalCalc-kurzhanski-1997} to be used as an efficient tool for control code stability analysis.
The rest of this paper is organized as follows. Section 2 describes the DGEN 380 turbofan engine and its virtual test bench. Section 3 presents the credible autocoding and verification process for the controller code. Section 4 presents the verifiable controller design and development for the DGEN engine. Section 5 presents the stability analysis for the closed-loop engine system with its gain scheduled controller. Section 6 presents some of the output verifiable code from the credible autocoding process. Section 7 presents simulation results of the verifiable controller code executed on the engine virtual test bench. Section 8 concludes the paper.
\section{Turbofan Engine}
The gas turbine engine model used in this study is a high fidelity model of DGEN 380 turbofan engine provided as a virtual test bench by Price Induction \cite{dgen380-link}. Brief descriptions of the DGEN engine and its virtual test bench are given below.
\subsection{Price Induction DGEN 380 Turbofan Engine}
Price Induction, a French aerospace company based out of Biarritz, has been designing for the past 10 years a family of engines designed specifically for general aviation. These DGEN engines are optimized for a cruise altitude ranging from 15,000 to 20,000 ft, at speed up to Mach 0.35 and with a flight ceiling limited to 25,000 ft. To be competitive with piston engines, they were also designed to have low specific fuel consumption (SFC), be lightweight and reasonably priced to allow the emergence of 4 to 5-seat light aircraft with a max weight ranging between 1,400 kg and 2,150 kg (3,417 lb and 5,622 lb) commonly called Personal Light Jets (PLJ).
The DGEN 380, the first engine of its family shown in Figure \ref{fig:dgen}, is a two-spool, high bypass ratio (7.6), unmixed flow turbofan engine. Its simple architecture yields up to 560 pounds of thrust in a compact and lightweight format (175 pounds and 4 feet long) while maintaining low noise and pollution levels. Beside its optimized performances, the engine innovates with its all-electric system: its starter-generator located directly on the high-pressure shaft, and oil and fuel pumps driven by electric motors are controlled by the Engine Control Unit (ECU), allowing for a really fine and optimized tuning of the DGEN control laws.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.5\textwidth]{./figs/dgen.jpg}
\caption{DGEN 380 lightweight turbofan engine (\copyright Price Induction)\cite{dgen380-link}}\label{fig:dgen}
\end{figure}
\subsection{Price Induction Engine Virtual Test Bench}
Based on its design expertise, Price Induction developed the WESTT (Whole Engine Simulator Turbine Technology) Solutions, which are educational and research tools based on the DGEN 380 turbofan engine and intended for high schools, universities, aeronautical maintenance training centers and research institutes.
The WESTT CS-BV, shown in Figure \ref{fig:westt}, is a product of this family dedicated to the study of the DGEN 380 turbofan and its control. With the DGEN 380 actual ECU hardware and its model running real-time and generating its sensors analog outputs, the CS-BV constitutes a great control Hardware-In-the-Loop (HIL) platform for the testing of engine control design. The MPC555 microcontroller which constitutes the core of the ECU can be easily programmed through the already existing code framework with different control logics and tested in real time with the use of the SIMMOT (engine real-time simulation). All engine outputs are displayed on screen and all data recorded for later performance analysis.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.7\textwidth]{./figs/westt.jpg}
\caption{ Price Induction WESTT CS-BV: DGEN 380 turbofan engine virtual test bench (\copyright Price Induction) \cite{dgen380-link}} \label{fig:westt}
\end{figure}
\section{Credible Autocoding and Verification Process}\label{autocodeProcess}
The Lyapunov matrices which are computed in the next sections can be used at the level of the code in order to formally verify its stability. Indeed the sublevel sets of the Lyapunov functions they represent are invariant sets in which the variables remain throughout the execution of the program. The concept of invariants is familiar to the computer scientists that seek them in computer programs. However, a general computer program lacks the nice mathematical structures that exist in control systems, which makes it hard to compute invariants for the general computer code. Instead, we target control systems by developing a control domain-specific framework of \emph{credible autocoding} and implementing it using a number of existing tools that we have extended. A high level view of the credible autocoding framework is given in Figure \ref{fig:verified_process}. It enables the control engineer to provide the control semantics such as the proof of Lyapunov stability directly on the Simulink diagram. This part of the framework of can be both manual or automated. The implementation of the framework is done by extending an Simulink to C translation tool named Gene-Auto \cite{nassimaformal09} to generate formal annotations in additional to the C code. These annotation describe the invariant sets in which the variables will remain. The sets are extracted automatically from the Lyapunov matrices. The formal annotation language used at the level of the C code is called ACSL (ANSI/ISO C Specification Language) \cite{acsl-link}. It can be read by various formal analysis tools such as Frama-C \cite{framaC-link}. In order to prove the validity of the generated annotations, we have extended these tools with domain-specific, automated routines that discharge the proof based on control-theoretic techniques.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.7\textwidth]{./figs/verified_process}
\caption{Visualization of autocoding and verification process} \label{fig:verified_process}
\end{figure}
\section{Verifiable Engine Controller Design}
The FADEC used for numerical analysis is a DGEN 380 turbofan engine controller model provided by Price Induction. The main novelty with this FADEC is that it is intended to be adaptable to different components such as a different compressor or turbine. The engine controller is developed using gain scheduling, and the scheduling parameter is the engine high pressure spool speed, i.e. $\alpha=NH$.
\subsection{Initial Model and Alterations to the Model}
The initial model requires extensive changes before it become compatible with our tool-chain. The majority of the changes required are not due to the Simulink blocks but rather because of the heavy presence of Matlab functions in the model. A large portion of modifications are made to the parts of the model that is not directly related to the controller. The model is consisted of three major subsystems. The first major subsystem is the throttle subsystem. This subsystem generates turbine commands NH and NL from the throttle input PLA. This subsystem is mostly written in Matlab and therefore need to be completely re-implemented in Simulink. The second major subsystem is the controller which contained transfer functions, and a polynomial curves-based lookup table. Both type of blocks are not supported by our tool-chain hence need to be converted into subsystems consisted of more basic blocks. The overall controller is consisted of dual PID subsystems with a basic anti-windup mechanism on both PID controllers. The last major subsystem is the Butee which contained safety features such as saturation operators and min/max switching strategies that are designed to prevent controller commands that could cause the turbine to exceed its maximum operational safety limits. This part of the model also requires extensive alteration as most of the Butee is written in Matlab.
\subsection{Analysis}
In this section, we give a description of the open-loop stability characteristics of the FADEC. These results are to be used to annotate the Simulink model and then transformed by our credible compilation tool-chain down to ACSL annotations for the C code.
From the Simulink model of FADEC provided by Price Induction, the semantics of the controller is extracted and then reformulated as a discrete-time state-space system. Let the controller states be denoted by the vector $x^c \in \Re^{11}$. The controller states are
\be
\dps x^c= \lb \ba{ccccccccccc} b_{0} & b_{1} & \epsilon_{0} & \epsilon_{1} & c_{0} & c_{1} & f_{0} & f_{1} & b_{2} & \epsilon_{2} & c_{2} \ea \rb^{\mathsf{T}}.
\label{controller_states}
\ee
The symbols $b_{n},n=0,1,2$ represent the integrator states, and the symbols $\epsilon_{n}, c_{n}, n=0,1,2 $ denote the states of the anti-windup mechanisms. The states $f_{n},n=0,1$ are the states of the first order filters used in the derivative portion of the two PID controllers.
We have $y_{1} \in \Re^{2}$ as one of the input to the controller
\be
\dps y_{1} =\lb \ba{c} \Delta NH \cr \Delta NL \ea\rb
\label{input:01}
\ee.
The symbols $\Delta NH$ and $\Delta NL$ are respectively the changes to the high and low pressure turbine spool speeds commanded by the throttle subsystem. The output from the controller, denoted as $u \in \Re$, is the input to the Butee subsystem. The Butee component is a safety limiter on the signal $u$. The output from the Butee, denoted as $\hat{u}$, is the input to the engine fuel pump. There is a feedback loop to the controller. This feedback loop to the controller contains two signals. One is the output from the Butee $\hat{u}$ and the other is $u_{2}$ which is a vector consisted of the anti-windup states $c_{n},n=1,2,3$. Both signals are delayed by one sample period in the feedback loop.
The symbols $NH$ and $NL$ denote the angular velocity of the high-pressure and low-pressure spools. The PID controller gains are computed using polynomial functions $p_{n},n=1,\ldots,4$ that map $NH$ to a set of gains
\be \ba{c}
\dps p_{1}: NH \ra K_{p} \cr
\dps p_{2}: NH \ra K_{i} \cr
\dps p_{3}: NH \ra K_{d} \cr
\dps p_{4}: NH \ra T_{d}.
\ea
\label{gains}
\ee
Let the parameter
\be
\dps \theta(\alpha) =\lb \ba{cccc} p_{1}(\alpha) & p_{2}(\alpha) & p_{3}(\alpha) & p_{4}(\alpha) \ea \rb
\label{parameter}
\ee
denote the set of PID gains for some $\alpha$. The state-space transition function and the output function of the FADEC can be defined using the following parameter-varying matrices.
\be \ba{llll}
\dps A^c(\theta) \in \Re^{11 \times 11}, & B^c(\theta) \in \Re^{11 \times 2}, & B^c_{u_{2}}(\theta) \in \Re^{11 \times 3}, & B^c_{\hat{u}}(\theta) \in \Re^{11 \times 1} \cr
\dps C^c_{1}(\theta) \in \Re^{1 \times 11}, & C^c_{2}(\theta) \in \Re^{3 \times 11}, & D^c_{1}(\theta) \in \Re^{1 \times 2 }
\ea
\label{abcd}
\ee
The matrices are varying in a nonlinear fashion with the parameter $\theta$. Let $B^c_{w}(\theta) =\lb \ba{ccc} B^c_{\hat{u}_{1}}(\theta) & B^c_{u_{2}}(\theta) \ea \rb$, $C^c(\theta)=\lb \ba{c} C^c_{1}(\theta) \cr C^c_{2}(\theta) \ea \rb$, $D^c(\theta)=\lb \ba{c} D^c_{1}(\theta) \cr 0 \ea \rb$, $u_{1}=C^c_{1}(\theta) x^c + D^c_{1}(\theta) y$, $u_{2}=C^c_{2}(\theta) x^c$, $\hat{u}_{1} =\sigma(u_{1})$, where $\sigma$ is the nonlinear causual operator representing the Butee, and finally let $w = \lb \ba{c} \hat{u}_{1} \cr u_{2} \ea \rb$. The discrete-time linear state-space model of the FADEC system is
\be
\ba{l}
\dps x^c_{+} = A^c(\theta) x^c + B^c (\theta) y + B^c_{w}(\theta) w_{-} \cr
\dps u =C^c(\theta) x^c + D^c(\theta) y \cr
\ea
\label{fadecss}
\ee
A diagram of the system in Simulink is given in Figure \ref{fig:fadec}.
\begin{figure}[!ht]
\centering
\includegraphics[scale=0.28]{./figs/fadecsys}
\caption{State-space model of the FADEC developed in Simulink}
\label{fig:fadec}
\end{figure}
For the sake of brevity, we have chosen not to display the closed-form expression of the parameter varying matrices in (\ref{fadecss}) as they are very large. We now seek to compute an ellipsoid invariant for (\ref{fadecss}). There were several issues with the FADEC model that presented a challenge towards finding a single ellipsoid invariant:
\begin{enumerate}
\item There is a sample delay on the input vector $w$. This precluded the usage of a simple quadratic function $V^c(x)=x^{\m{cT}} P^c x^c$ for the stability analysis. Instead, a Lyapunov-Krasovskii type functional $V^c(x^c,x^c_{-})=x^{\m{cT}} P^c x^c - x_{-}^{\m{cT}} P^c x^c_{-}$ is needed to handle the sample delay. However the resulting LMI cannot be solved by SeDuMi.
\item A convex hull of the system matrices is needed since the system matrices are not linearly parameter-varying.
\item The Butee component contained complex safety limiters in addition to the simple saturation operators.
\end{enumerate}
We discuss each of these issues in the ensuing sections.
\subsubsection{Sample Delay}
The problem with the sample delay was resolved simply by its removal. We shifted some of the dynamics in the controller
forward by one sample and removed the one sample delay on $\hat{u}_{1}$ so that the signal $w$ no longer need to be delayed by one sample. The changes are small enough that the performance of the controller is not noticeably affected as indicated by the simulations. Without the sample delay on $w$, the system in (\ref{fadecss}) becomes the following state-space system
\be
\ba{l}
\dps x^c_{+} = \hat{A}^c(\theta) x^c + \hat{B}^c (\theta) \hat{u} \cr
\dps u =C^c(\theta) x^c + D^c(\theta) y \cr
\ea
\label{fadecss2}
\ee
with a new state-transition matrix $\hat{A}^c$, and
$\hat{B}^c(\theta) =\lb \ba{cc} B^c(\theta) & B^c_{\hat{u}_{1}}(\theta) \ea \rb$, $\dps \hat{u}=\lb \ba{c} y \cr \hat{u}_{1} \ea \rb$.
\subsubsection{Butee Component}
There are two modes of operation to the Butee. The simple operational mode of the Butee is consisted of two identical saturation operator that restricts the input to the Butee and the output from the Butee to the interval $\lb 0.07, 0.098 \rb$. We can model the saturation nonlinearities using a sector-bound inequality. Let $\delta$ be the midpoint of the output range of the saturation operator i.e. $\delta = 0.5 (0.07 + 0.098)=0.84$. Let the output from the Butee to be denoted by $w$ and let $\tilde{w} = w - \delta$. With $m_{1}=1$ and $0<m_{2}<m_{1}$, we have the sector-bound inequality
\be
\dps (\tilde{y} - m_{1} C^c x^c - m_{1} D^c u + m_{1} \delta)^{\mathsf{T}} (\tilde{y} -m_{2} C^c x^c - m_{2} D^c u + m_{2} \delta) \leq 0.
\label{sector}
\ee
Let $\kappa_{1}=m_{1} m_{2}$, $\kappa_{2} \dps = \frac{1}{2}(m_{1} + m_{2})$.
The sector bound constraint in (\ref{sector}) is equivalent to the quadratic inequality $\forall x$, $\forall u$, $\forall y$,
\be
\lb \ba{c} x^c \cr u \cr \tilde{y} \cr 1 \ea\rb^{\mathsf{T}}
\lb \ba{cccc}
\kappa_{1} C^{c\mathsf{T}} C^c & \kappa_{1} C^{c\mathsf{T}} D^c & - \kappa_{2} C^{c\mathsf{T}} & \kappa_{2} C^{c\mathsf{T}} \cr
\kappa_{1} D^{c\mathsf{T}} C^c & \kappa_{1} D^{c\mathsf{T}} D^c & -\kappa_{2} D^{c\mathsf{T}} & \kappa_{2} D^{c\mathsf{T}} \cr
-\kappa_{2} C^c & -\kappa_{2} D^c & 1 & -1 \cr
\kappa_{2} C^c & \kappa_{2} D^c & -1 & 1\ea \rb
\lb \ba{c} x^c \cr u \cr \tilde{y} \cr 1 \ea\rb \leq 0
\label{sector02}
\ee
In the complex operational mode, the Butee employs a type of min/max switching component typically encountered in engine controllers to handle the performance limits. Although its input-output relations cannot be captured using a sector-bound inequality, however this component is sandwiched between the two saturation operators. We can assume a simple bound on the output of the Butee even when this mode is switched on. In fact, a simple bound on the Butee output $\|\hat{u}_{1}\|\leq 1$ produces better results numerically speaking than using the inequality from (\ref{sector02}).
\subsubsection{Convex Hull of the System Matrices}
With the removal of the sample delays, we have the following linear state-space model of the FADEC (the same as
in (\ref{fadecss2}))
\be
\ba{l}
\dps x^c_{+} = \hat{A}^c(\theta) x^c + \hat{B}^c (\theta) \hat{u} \cr
\dps u =C^c(\theta) x^c + D^c(\theta) y.
\ea
\label{fadecss3}
\ee
\begin{figure}[!ht]
\centering
\subfloat[]{\includegraphics[scale=0.46]{./figs/m11}
}
\subfloat[]{\includegraphics[scale=0.46]{./figs/m12}
}
\\
\subfloat[]{\includegraphics[scale=0.46]{./figs/m21}
}
\subfloat[]{\includegraphics[scale=0.46]{./figs/m22}
}
\\
\subfloat[]{\includegraphics[scale=0.46]{./figs/m31}
}
\subfloat[]{\includegraphics[scale=0.46]{./figs/m32}
}
\\
\subfloat[]{\includegraphics[scale=0.46]{./figs/m41}
}
\subfloat[]{\includegraphics[scale=0.46]{./figs/m42}
}
\caption{Varying entries and the convex hull}
\label{plot01}
\end{figure}
To compute a single ellipsoid invariant for (\ref{fadecss3}), we need to find a convex hull that contains all $\hat{A}^c (\theta)$ and $\hat{B}^c (\theta)$ for an assumed range of $\theta$. Since $\theta$ is a polynomial function of the scheduling parameter $\alpha=NH$, the state-transition matrix $\hat{A}$ and the input matrix $\hat{B}^c$ are smooth matrix functions of $NH$. To construct the convex hull, we first compute 4 corner matrices $\lb \ba{cc} \hat{A}^c & \hat{B}^c \ea \rb_{ij},i,j=1,2$ using the following formulas
\be
\ba{c}
\dps \lb \ba{cc} \hat{A}^c & \hat{B}^c \ea \rb_{11}=\lb \ba{cc} \hat{A}^c(\alpha_{\min}) & \hat{B}^c(\alpha_{\min}) \ea \rb + \Delta_{1} \cr
\dps \lb \ba{cc} \hat{A}^c & \hat{B}^c \ea\rb_{12}=\lb \ba{cc} \hat{A}^c(\alpha_{\min}) & \hat{B}^c(\alpha_{\min}) \ea \rb - \Delta_{2} \cr
\dps \lb \ba{cc} \hat{A}^c & \hat{B}^c \ea\rb_{21}=\lb \ba{cc} \hat{A}^c(\alpha_{\max}) & \hat{B}^c(\alpha_{\max}) \ea \rb + \Delta_{3} \cr
\dps \lb \ba{cc} \hat{A}^c & \hat{B}^c \ea\rb_{22}=\lb\ba{cc} \hat{A}^c(\alpha_{\max}) & \hat{B}^c(\alpha_{\max}) \ea \rb - \Delta_{4},
\ea
\label{corner}
\ee
with some estimated perturbation matrices $\Delta_{i},i=1,\ldots,4$. The minimum and maximum values of the scheduling parameter are $\alpha_{\min}=85\%$ and $\alpha_{\max}=106\%$. Next, we need to check if the convex hull of the 4 corner matrices $\Co \lc \lb \ba{cc} \hat{A}^c & \hat{B}^c \ea \rb_{ij}, i,j=1,2\rc$ contains all $\hat{A}^c(\theta)$ and $\hat{B}^c(\theta)$ for the range of $NH$ that we assumed to be valid. A graphical way to check this is to plot each of the entries from $\hat{A}^c$ and $\hat{B}^c$ that varies with NH as a function of NH, and then check if any of the resulting curves violate the convex hull formed by the $4$ corners. If there is a violation in any one entry, then we have to recalculate the 4 corner matrices with larger perturbation matrices until there is no violation in any of the entries.
It turns out that, of the 154 possible entries in the matrices $\hat{A}^c$ and $\hat{B}^c$, only 30 of them vary according to the parameter NH. The rest are either constant or zero. Figure \ref{plot01} has some example plots of those entries as a function of the parameter NH. The blue curves are the plots of the entries of $\hat{A}^c$ or $\hat{B}^c$ that vary with respect to the parameter NH. The red and green lines delineate the convex hull formed by the four corner matrices.
As you can see, all the curves are well within the convex hull formed by the 4 corners.
\subsubsection{Open-Loop Stability Result}
With the 4 corner matrices, we can now apply the following stability criterion to generate an invariant for the system in (\ref{fadecss3}).
\begin{proposition}
Assume the matrix $\lb \ba{cc} \hat{A}^c(\theta) & \hat{B}^c(\theta) \ea \rb \in \Co \lc \lb \ba{cc} \hat{A}^c & \hat{B}^c \ea \rb_{ij}, i,j=1,2\rc$ for some range of NH and assume that $ \| \hat{u} \| \leq 1$. If there exists a positive definite matrix $P^c$ and a scalar $\xi>0$ that satisfies
\be
\dps \lb \ba{cc}
\hat{A}_{ij}^{c\mathsf{T}} P^c \hat{A}^c_{ij} - P^c + \xi P^c & \hat{A}_{ij}^{c\mathsf{T}} P^c \hat{B}^c_{ij} \\[5pt]
\hat{B}_{ij}^{c\mathsf{T}} P^c \hat{A}^c_{ij} & \hat{B}_{ij}^{c\mathsf{T}} P^c \hat{B}^c_{ij} - \xi I
\ea \rb \prec 0,
\label{lmi05}
\ee
then the set $\lc x^c | x^{c\mathsf{T}} P^c x^c\leq 1\rc$ is an invariant set with respect to (\ref{fadecss3}).
\label{prop06}
\end{proposition}
Using proposition \ref{prop06}, for $\xi=0.02354$, a Lyapunov matrix $P^c \in \Re^{11 \times 11}$ is computed. The numerical value of this matrix is given in Appendix A.
\section{Engine Control Stability Analysis}
In order to fulfill the stability analysis of the engine with its control system, we needed to obtain the open-loop and closed-loop model of the whole system using engine, controller, and fuel pump dynamics. Figure \ref{fig:linear_modeling_process} visualizes this process. The engine linearization data (including plant, controller, open-loop systems, and closed-loop system matrices) for the four important engine equilibrium points are given in Appendix D.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.6\textwidth]{./figs/linear_modeling_process}
\caption{Block diagram visualization of the modeling process using plant, controller, and fuel pump models} \label{fig:linear_modeling_process}
\end{figure}
The states in the linear plant model are the spool speeds and spool accelerations (i.e., $x^p \in \Re^4$), and the input to the plant is the fuel flow rate in kg/hr. The controller has 11 states ($x^c \in \Re^{11}$), and its output is fuel pump RPM in percentage and nondimensional. The output of the controller ($W_c$) is multiplied by $4100/160$ to obtain $W_p=(4100/160) W_c$, then the output signal $W_p$ is transformed using $v=3883 W_p-244.06$; the fuel pump transfer function ($T_{pump}(z)=\frac{W_f(z)}{v(z)}$) is
\begin{equation}
T_{pump}(z)=\frac{0.21756}{z-0.8187},
\end{equation}
where its output, fuel flow rate ($W_f$), is the input to the plant in kg/hr.
Stability analysis of the overall system with controller is done using two different approaches. These methods are the Bounded Real Lemma \cite{lmi-boyd-1994, NonlinDynSysCont-haddad-2008} and the closed-loop system stability analysis approach developed in \cite{GainSchedStabConf-pakmehr-2013, GSstability-pakmehr-2013}. To obtain a numerical value of stability matrix $P$ for software verification process, there is a set of LMIs in each of these approaches, which can be solved for all $i=\{1,2,...,L\}$ using YALMIP \cite{YALMIP-lofberg-2004} and SeDuMi \cite{sedumi-Sturm-2001} packages in Matlab.
\subsection{Bounded Real Lemma}
Having in mind that $G(z)$ is bounded real if and only if $G(z)$ is asymptotically stable and $||G(z)||_{\infty} \leq \gamma$, the \emph{Bounded Real Lemma} is given as follows
\begin{thm}
Consider the dynamical system
\begin{equation}
\begin{array}{l}
\displaystyle G_i(z) \overset{\min}{\sim} \left[
\begin{array}{c|c}
A^{ol}_i & B^{ol}_i \\ \hline
C^{ol}_i & D^{ol}_i
\end{array}
\right], ~~~ i=1,2,...,L,
\end{array}
\end{equation}
with input $u(.) \in \mathcal U$ and output $y(.) \in \mathcal Y$. If $\gamma_i I_n-D^{ol^{\mathsf{T}}}_i D^{ol}_i-B^{ol^{\mathsf{T}}}_i P B^{ol^{\mathsf{T}}}_i > 0$ for all $i=\{1,2,...,L\}$, then $G^{ol}_i(z)$ is bounded real if and only if there exist a $P=P^{\mathsf{T}}>0$ for all $i=\{1,2,...,L\}$ such that
\begin{equation}\label{eqn_gs015}
\left[ \begin{array}{cc}
A^{ol^{\mathsf{T}}}_i P A^{ol}-P+C^{ol^{\mathsf{T}}}_i C^{ol}_i & (B^{ol^{\mathsf{T}}}_i P A^{ol}_i+D^{ol^{\mathsf{T}}}_i C^{ol}_i)^{\mathsf{T}} \\[5pt]
(B^{ol^{\mathsf{T}}}_i P A^{ol}_i+D^{ol^{\mathsf{T}}}_i C^{ol}_i) & -(\gamma_i I_n-D^{ol^{\mathsf{T}}}_i D^{ol}_i-B^{ol^{\mathsf{T}}}_i P B^{ol}_i)
\end{array} \right] \preccurlyeq 0.
\end{equation}
\end{thm}
In this case, the LMI (\ref{eqn_gs015}) is solved for the four main equilibrium points of the system, i.e. $i=4$. A Lyapunov matrix $P \in \Re^{16 \times 16}$ is computed, and the numerical value of this matrix is given in Appendix B.
\subsection{Closed-Loop Stability}
The discussions in this section about the engine closed-loop stability are extended from \cite{GainSchedStabConf-pakmehr-2013, GSstability-pakmehr-2013}.
\begin{thm}\label{thm2}
Consider the closed-loop system
\begin{equation}\label{eqn_gs112}
x_{+}=F(x,r),
\end{equation}
and assume there is a family of equilibrium points $(x_{eq},r_{eq})$ such that $F(x_{eq},r_{eq})=0$. Define $A^{cl} = \frac{\partial F(.)}{\partial x} \in \overline{S}, ~\forall x \in D_x $, where $\overline{S}$ is the set of linearizations of the system (\ref{eqn_gs112})
\begin{equation}\label{eqn_gs113}
\overline{S} := \{ A^{cl}, \forall x \in D_x \}.
\end{equation}
Assume there exist symmetric positive definite matrix $P$, such that
\begin{equation}\label{eqn_gs54}
A^{cl^{\mathsf{T}}} P A^{cl} - P \prec 0, ~~~ \forall A^{cl} \in \overline{S},
\end{equation}
then the system (\ref{eqn_gs112}) is stable. In other words, assuming the initial state is sufficiently close to some equilibrium, then the closed-loop system remains in a neighborhood of the equilibrium manifold for all $t \geq 0$.
\end{thm}
\begin{rmk}\label{rmk1}
In practice we can not obtain $\overline{S}$, instead, we can linearize system (\ref{eqn_gs112}) for a large number of points $x_i$, $i=1, \ldots, L$, which we claim is sufficient to cover the set of actual operating conditions, to show the stability of the closed-loop system. Define $S$ as a matrix polytope described by its vertices
\begin{equation}\label{eqn_gs114}
S:= \mathrm{Co}\{ A^{cl}_{1}, ..., A^{cl}_{L} \},
\end{equation}
where $A^{cl}_{i} = \left. \frac{\partial F(.)}{\partial x(t)} \right|_{x=x_i} \in S$, for all $i \in \{ 1,2, ..., L \}$. Note that $A^{cl}_{i}$ can be obtained by linearizing the nonlinear system (\ref{eqn_gs112}) at non-equilibrium points (transient condition), and also at equilibrium points (steady state condition). Then using convex optimization tools \cite{YALMIP-lofberg-2004, sedumi-Sturm-2001}, we compute a common symmetric positive definite matrix $P$, such that
\begin{equation}\label{eqn_gs115}
A^{cl^{\mathsf{T}}}_{i} P A^{cl}_{i} - P \prec 0, ~~~ \forall i \in \{ 1,2, ..., L \}.
\end{equation}
\end{rmk}
In this case, the LMI (\ref{eqn_gs115}) is solved for the four main equilibrium points of the system, i.e. $i=4$. A Lyapunov matrix $P \in \Re^{16 \times 16}$ is computed, and the numerical value of this matrix is given in Appendix C.
\section{Autocoded C with Proof Annotations}
First we give a brief introduction to the formal C annotation language ACSL. A main function of the ACSL is that it can be used to formally specify properties about the variables of the code using an annotation language that is similar in syntax and semantics as the C language.
The specified properties can be either assumptions or inductive invariants. The former case does not require a proof as it is an assumption made on the variable(s). For example, for a real-time system that interacts with the environment, we need to assume some bounds on inputs from the environment.
The latter does require a proof on the level of the code. As mentioned in section \ref{autocodeProcess}, a plethora of tools exist that can be used to analyze ACSL annotations and discharge any necessary proof obligations.
\begin{figure}[!ht]
\begin{lstlisting}
/*@
assume input<=1
ensure x<=1
*/
float x=input;
/*@
require x*x<=1
ensures x*x<=1
*/
while (1) {
x=0.99*x;
}
\end{lstlisting}
\caption{Simple ACSL example}
\label{loop01}
\end{figure}
Typically in ACSL, the assumptions are specified using the keyword \emph{assume}. For example, in the C code shown in Figure \ref{loop01}, the first line of the code assigns the value of the variable $input$ to the variable $x$. We want to assume that the variable $input$ is bounded by $1$ so we inserted an ACSL statement, which is encapsulated within the symbols /*@ and */, that specifies this property. The keywords \emph{require} and \emph{ensures} are used to specify the invariants. The former denotes the valid condition before the execution of the line of the code e.g. the \emph{pre-condition} while the latter denotes the valid condition afterwards e.g. the \emph{post-condition}. In Figure \ref{loop01}, we have the invariant $x*x<=1$ which holds true throughout the execution of the infinite loop. These type of invariants are expressed as both a pre- and a post-condition for the loop.
\begin{figure}[!ht]
\begin{lstlisting}
/*@
requires in_ellipsoidQ(QMat_0,vect_of_11_scalar(_state_->delay_aw0_memory,_state_->delay_aw1_memory,_state_->delay_E0_memory,_state_->delay_E1_memory,_state_->delay_D0_memory,_state_->delay_D1_memory,_state_->delay_x1_memory,_state_->delay_x2_memory,_state_->delay_aw2_memory,_state_->delay_E2_memory,_state_->delay_D2_memory));
requires \valid(_io_) && \valid(_state_);
ensures in_ellipsoidQ(QMat_1,vect_of_11_scalar(_state_->delay_aw0_memory,_state_->delay_aw1_memory,_state_->delay_E0_memory,_state_->delay_E1_memory,_state_->delay_D0_memory,_state_->delay_D1_memory,_state_->delay_x1_memory,_state_->delay_x2_memory,_state_->delay_aw2_memory,_state_->delay_E2_memory,_state_->delay_D2_memory));
*/
void pla_compute(t_pla_io *_io_, t_pla_state *_state_) {
REAL NL;
REAL NH;
REAL P3_KPa_;
REAL PLA;
\end{lstlisting}
\caption{Ellipsoid invariants for the generated FADEC code}
\label{fadfunc}
\end{figure}
For the credible autocoding of the FADEC, we inserted the open-loop stability proof into the Simulink diagram. The autocoding process generate two functions.
The first one is the initialization function and the other, called the $pla\rm{\_}compute$, is an amalgamation of the state-transition and output functions of the FADEC. The stability proof is inserted as an ellipsoid invariant on the function $pla\rm{\_}compute$ in two instances. They are specified using both \emph{require} and \emph{require} keywords in the ACSL comment shown in Figure \ref{fadfunc}. The function $in\rm{\_}ellipsoidQ$ defines the ellipsoid invariant using two arguments: the ellipsoid matrix in the form $Q=P^{-1}$ and vector of variables that is captured by the ellipsoid set.
\begin{figure}[!ht]
\begin{lstlisting}
/*@
behavior ellipsoid544_1:
requires in_ellipsoidQ(QMat_562,vect_of_13_scalar(_state_->delay_aw1_memory,_state_->delay_aw2_memory,Sum_of_Elements12_1,Sum_of_Elements12_2,_state_->delay_aw0_memory,_state_->delay_E0_memory,_state_->delay_E1_memory,_state_->delay_D0_memory,_state_->delay_D1_memory,_state_->delay_x1_memory,_state_->delay_x2_memory,_state_->delay_E2_memory,_state_->delay_D2_memory));
ensures in_ellipsoidQ(QMat_563,vect_of_11_scalar(_state_->delay_aw0_memory,_state_->delay_aw1_memory,_state_->delay_E0_memory,_state_->delay_E1_memory,_state_->delay_D0_memory,_state_->delay_D1_memory,_state_->delay_x1_memory,_state_->delay_x2_memory,_state_->delay_aw2_memory,_state_->delay_E2_memory,_state_->delay_D2_memory));
@ PROOF_TACTIC (use_strategy (AffineEllipsoid));
*/
{
_state_->delay_aw2_memory = Sum_of_Elements12_2;
}
}
\end{lstlisting}
\caption{Verification of the invariant using the generated post-condition}
\label{funcend}
\end{figure}
The reason for specifying the ellipsoid invariant twice, as pre and post-conditions, is because the function $pla\rm{\_}compute$ is executed in a loop just like the loop shown in Figure \ref{loop01}. Any invariant that is valid before the execution of $pla\rm{\_}compute$ also need to be valid after its execution.
This is a property that needs to be proven on the C code level and the backend tools mentioned in section \ref{autocodeProcess} have been equipped to handle this type of proof obligation. To enable the backend analyzer, the credible autocoder also generates additional ellipsoid invariants along with the proof strategy used for every line of code inside the function $pla\rm{\_}compute$. This is done until the last line of the function as shown in Figure \ref{funcend}, in which there is a post-condition generated by the autocoder This generated post-condition is used to check against the ellipsoid invariant that was inserted as pre- and post-conditions on the function $pla\rm{\_}compute$. For proof of correctness, one just need to show that the latter implies the former.
\section{Simulation Results}
\begin{figure}[!ht]
\centering
\subfloat[]{\includegraphics[scale=0.48]{./figs/simulation_engine_1} } \\
\subfloat[]{\includegraphics[scale=0.48]{./figs/simulation_engine_2} }
\caption{Snapshots of the verifiable engine controller implementation on the DGEN 380 turbofan engine virtual test bench}
\label{fig:simulation_engine}
\end{figure}
Here we present the simulation results to show that the commented code works like the uncommented one. The code verification process is developed to be sure that the controller is certifiable; it is not supposed to replace the simulation process. We transform the proof into an observer to detect the system's malfunctions and to indicate it is actually a \textit{health monitoring} software.
Figure \ref{fig:simulation_engine} shows two snapshots of the WESTT command screen for the case where the verifiable controller implemented on the DGEN 380 turbofan engine virtual test bench. These snapshots illustrate evolution of both engine spool speeds that closely follow their reference signals, and also engine fuel and oil pressure time histories.
A visualizations of the engine related avionics is also presented in these snapshots. The snapshot shows real-time measurement of pressures ($P_{amb}$, $P_2$, and $P_3$), temperatures ($T_{amb}$, and $EGT$), speeds ($NH$, and $NL$), thrust, fuel pump rating, oil pump rating, fuel consumption, fuel pressure, and oil pressure.
\section{Conclusions}
A stability proof was presented for the closed-loop DGEN 380 turbofan engine system with its gain scheduled controller based on the Lyapunov stability theory. Using convex optimization tools, various numerical values of the Lyapunov stability matrix, for the closed-loop system and the controller, were computed separately. With these stability matrices, ellipsoid invariant sets were constructed, which were used efficiently for DGEN turbofan engine control code stability analysis and verification. The verifiable engine controller code then implemented successfully on the engine virtual test bench (WESTT) to control a high fidelity DGEN 380 engine model. Simulation results are presented to illustrate the efficiency of the presented approach. The code verification framework presented here, hopefully, can be used for gas turbine engine control software certification in the future.
\section*{Acknowledgment}
This material is based upon the work supported by the Air Force Office of Scientific Research (AFOSR), the National Aeronautics and Space Administration (NASA), and the National Science Foundation (NSF).
\bibliographystyle{plain}
|
\section{Introduction}
Type theory and denotational semantics have been successfully used to
model, design, and reason about programming languages for almost half
a century. The application of such methods to quantum computing is
much more recent, going back only about 10 years {\cite{Selinger04}}.
An important problem in the semantics of quantum computing is how to
combine quantum computing with higher-order functions, or in other
words, how to design a functional quantum programming language. A
syntactic answer to this question was arguably given with the design
of the quantum lambda calculus
{\cite{valiron08phd,selinger06lambda}}. The quantum lambda calculus
has a well-defined syntax and operational semantics, with a strong
type system and a practical type inference algorithm. However, the
question of how to give a {\em denotational} semantics to the quantum
lambda calculus turned out to be difficult, and has remained open for
many years {\cite{Selinger04b,SV09}}. One reason that designing such a
semantics is difficult is that quantum computation is inherently
defined on {\em finite dimensional} Hilbert spaces, whereas the
semantics of higher-order functional programming languages, including
such features as infinite data types and recursion, is inherently
infinitary.
In recent years, a number of solutions have been proposed to the
problem of finding a denotational semantics of higher-order quantum
computation, with varying degrees of success. The first
approach~\cite{valiron06fully} was to restrict the language to strict
linearity, meaning that each function had to use each argument exactly
once, in the spirit of linear logic. In this way, all infinitary
concepts (such as infinite types and recursion) were eliminated from
the language. Not surprisingly, the resulting finitary language
permitted a fully abstract semantics in terms of finite dimensional
spaces; this was hardly an acceptable solution to the general problem.
The second approach~\cite{malherbe2010} was to construct a semantics
of higher-order quantum computation by methods from category theory;
specifically, by applying a presheaf construction to a model of
first-order quantum computation. This indeed succeeds in yielding a
model of the full quantum lambda calculus, albeit without
recursion. The main drawbacks of the presheaf model are the absence of
recursion, and the fact that such models are relatively difficult to
reason about. The third
approach~\cite{GoIquantum} was based on the Geometry of
Interaction. Starting from a traced monoidal category of basic quantum
operations, Hasuo and Hoshino applied a sequence of categorical
constructions, which eventually yielded a model of higher-order
quantum computation. The problem with this approach is that the tensor
product constructed from the geometry-of-interaction construction does
not coincide with the tensor product of the underlying physical data
types. Therefore, the model drops the possibility of entangled states,
and thereby fails to model one of the defining features of quantum
computation.
\noindent
{\bf Our contribution.}
\quad
In this paper, we give a novel denotational semantics of higher-order
quantum computation, based on methods from {\em quantitative
semantics}. Quantitative semantics refers to a family of semantics
of linear logic that interpret proofs as linear mappings between
vector spaces (or more generally, modules), and standard lambda terms
as power series. The original idea comes from Girard's normal functor
semantics \cite{Girard88c}. More recently, quantitative semantics has
been used to give a solid, denotational semantics for various
algebraic extensions of lambda calculus, such as probabilistic and
differential lambda calculi (e.g.\ \cite{danosehrhard}, \cite{finsp}).
One feature of our model is that it can represent {\em infinite
dimensional} structures, and is expressive enough to describe
recursive types, such as lists of qubits, and to model recursion. This
is achieved by providing an exponential structure {\em \`{a} la}
linear logic. Unlike the Hasuo-Hoshino model, our model permits
general entanglement. We interpret (a minor variant of) the quantum
lambda calculus in this model. Our main result is the adequacy of the
model with respect to the operational semantics.
The model is the juxtaposition of a simple, finite-dimensional model
of quantum computation together with a canonical completion yielding
the structures of linear logic. Our model demonstrates that the
quantum and the classical ``universes'' work well together, but also
-- surprisingly -- that they do not mix too much, even at higher order
types.
\noindent
{\bf Outline.}
\quad
In Section~\ref{sec:background}, we briefly review some background.
Section~\ref{sec:qlc} presents the version of the quantum lambda
calculus that we use in this paper, including its operational
semantics. Section~\ref{sec:sem} presents the denotational semantics
of the quantum lambda calculus, and Section~\ref{sec:adequacy} proves
the adequacy theorem. Section~\ref{subsect:discussion} concludes with
some properties of the representable elements.
\section{Background}\label{sec:background}
\subsection{Quantum computation in a nutshell}
Quantum computation is a computational paradigm based on the laws of
quantum physics. We briefly recall some basic notions; please see
{\cite{nielsen02quantum}} for a more complete treatment. The basic
unit of information in quantum computation is a {\em quantum bit} or
{\em qubit}, whose state is given by a normalized vector in the
two-dimensional Hilbert space $\C^2$. It is customary to write the
canonical basis of $\C^2$ as $\{\ket0,\ket1\}$, and to identify these
basis vectors with the booleans false and true, respectively. The
state of a qubit can therefore be thought of as a complex linear
combination $\alpha\ket0 + \beta\ket1$ of booleans, called a {\em
quantum superposition}. More generally, the state of $n$ qubits is
an element of the $n$-fold tensor product
$\C^2\otimes\ldots\otimes\C^2$.
There are three kinds of basic operations on quantum data:
initializations, unitary maps and measurements. Initialization
prepares a new qubit in state $\ket0$ or $\ket1$. A unitary map, or
\define{gate}, is an invertible linear map $U$ such that $U^*=U\inv$;
here $U^*$ denotes the complex conjugate transpose of $U$. Finally,
the operation of measurement consumes a qubit and returns a classical
bit. If $n$ qubits are in state $\alpha\ket0\tensor\phi_0 +
\beta\ket1\tensor\phi_1$, where $\phi_0$ and $\phi_1$ are normalized
states of $n-1$ qubits, then measuring the leftmost qubit yields
false with probability $|\alpha|^2$, leaving the remaining
qubits in state $\phi_0$, and true with probability $|\beta|^2$,
leaving the remaining qubits in state $\phi_1$.
\begin{example}\label{ex:cointoss}
A small algorithm is the simulation of an unbiased coin toss:
initialize one quantum bit to $\ket{0}$, apply the Hadamard gate
sending $\ket0$ to $\frac1{\sqrt2}(\ket0+\ket1)$ and $\ket1$ to
$\frac1{\sqrt2}(\ket0-\ket1)$, then measure. The result is
true with probability $\frac12$ and false with probability
$\frac12$.
\end{example}
\begin{figure}
\begin{center}
\scalebox{0.8}{$\xymatrix@=.7pc{
& & & & &\ar@{}|*\txt{qubit 1:~~~~} & \ket{\phi}\ar@{-}[r] &
*-={\bullet}\ar@{-}[r]\ar@{-}[dd] &
*+[F]{H}\ar@{-}[r] &
*+<10pt>{\;}\\
& & \ar@{}|*!/lu8pt/\txt{(i)} & & & &&
\ar@{}|*!<15pt,5pt>\txt{(ii)} & &
*+{M}\ar@{}
\ar@/^2.2pc/@{.>}[rrddddd]|!{[llllddd];
[rrrddd]}\hole^<(0.2){x,y}\\
\ar@{}|*\txt{qubit 2:~~~~}&\ket{0}\ar@{-}[r] & *+[F]{H}\ar@{-}[r] &
*-={\bullet}\ar@{-}[rrrr]\ar@{-}[dd]
& & & & *-={\oplus}\ar@{-}[rr] & &
*+<10pt>{\;}\\
& & & & & & & & & *+<10pt>{\;}\\
\ar@{}|*\txt{qubit 3:~~~~} & \ket{0}\ar@{-}[rr] & & *-={\oplus}
\ar@{-}`r[rd][rd]&&
\ar@3{.}[rrrrrrrr]_(.6)*+\txt{location
B}^(.4)*+\txt{location A}
&&&&&&&&\\
& & & &*-={}\ar@{-}`d[rd][rd]&& & & & & & & & & \\
& & & & & *-={}\ar@{-}[rrrrrr] & & & & & &
*+[F]{U_{xy}}\ar@{-}[r] \ar@{}|*!<10pt,15pt>\txt{(iii)}
& \ket{\phi}
\save
"1,10"."3,10"*[F--]\frm{}
\restore
\save
"3,3"."5,4"*!<3pt,-3ptpt>++[F.]\frm{}
\restore
\save
"1,8"."3,10"*!<2pt,-6pt>++[F.]\frm{}
\restore
}$}
\end{center}
\vspace{-10pt}
\caption{\footnotesize The quantum teleportation protocol.}\label{fig:telep}
\end{figure}
\begin{example}\label{ex:telep-presentation}
A slightly more involved algorithm is the {\em quantum teleportation
algorithm} (see {\cite{nielsen02quantum}} for details). The
procedure is summarized in Figure~\ref{fig:telep}. Wires represent
the path of quantum bits in the computation, and time flows from
left to right. The gate \scalebox{0.8}{$\xy(0,0)*+[F]{H}\endxy$}
stands for an application of the Hadamard gate, whereas the gate
\raisebox{1ex}{\scalebox{0.8}{$\xy(0,0)*{\bullet};
(0,-3)*{\oplus}**\dir{-}\endxy$}}
is a controlled-not: it negates the bottom qubit if the upper one is
in state $\ket1$. The box $M$ is a measurement. The unitaries
$U_{xy}$ are
\[
U_{00} = \left(\begin{smallmatrix}1&0\\0&1\end{smallmatrix}\right),
~ U_{01} =
\left(\begin{smallmatrix}0&1\\1&0\end{smallmatrix}\right), ~ U_{10}
=
\left(\begin{smallmatrix}1&0\\0&\textrm{-}1\end{smallmatrix}\right),
~ U_{11} =
\left(\begin{smallmatrix}0&1\\\textrm{-}1&0\end{smallmatrix}\right).
\]
The goal is to send a quantum bit in an unknown state $\ket\phi$
from Location A to Location B using two classical bits. The
procedure can be reversed to send two classical bits using a quantum
bit. In this case it is called the \define{dense coding algorithm}
{\cite{nielsen02quantum}}.
The algorithm consists of three parts. In (i), two quantum bits
(qubits 2 and 3) are entangled in state
$\frac1{\sqrt2}(\ket{00}+\ket{11})$. In (ii), the input qubit 1 in
state $\ket\phi$ is entangled with qubit 2, then both are
measured. The result is sent over location B, where in (iii) an
correction $U_{xy}$ is applied on qubit 3, setting it to state
$\ket\phi$.
\end{example}
\subsection{Density matrices and completely positive maps}
If we identify $\ket0$ and $\ket1$ with the standard basis vectors
$\zzmatrix{c}{1\\0}$ and $\zzmatrix{c}{0\\1}$, the
state of a qubit can be expressed as a two-dimensional vector
$v=\alpha\ket0 + \beta\ket1=\zzmatrix{c}{\alpha\\\beta}$. Similarly,
the state of an $n$-qubit system can be expressed as an
$2^n$-dimensional column vector.
Often, it is necessary to consider {\em probability distributions} on
quantum states; these are also known as {\em mixed states}. Consider a
quantum system that is in one of several states $v_1,\ldots,v_k$ with
probabilities $p_1,\ldots,p_k$, respectively. The {\em density matrix}
of this mixed state is defined to be
$ A = \sum_i p_i v_iv_i^*,
$
where $(-)^*$ denotes the adjoint operator.
By a theorem of Von Neumann, the density matrix is a good
representation of mixed states, in the following sense: two mixed
states are indistinguishable by any physical experiment if and only if
they have the same density matrix {\cite{nielsen02quantum}}. Note that
$\tr A = p_1+\ldots+p_k$. For our purposes, it is often convenient to
permit sub-probability distributions, so that $p_1+\ldots+p_k\leq 1$.
Let us write $\C^{n\times n}$ for the space of $n\times
n$-matrices. Recall that a matrix $A\in\C^{n\times n}$ is called {\em
positive} if $v^*Av\geq 0$ for all $v\in\C^n$. Given
$A,B\in\C^{n\times n}$, we write $A\sqleq B$ iff $B-A$ is positive;
this is the so-called {\em {\Lowner} partial order}. A linear map
$F:\C^{n\times n}\to\C^{m\times m}$ is called {\em positive} if
$A\sqgeq 0$ implies $F(A)\sqgeq 0$, and {\em completely positive} if
$F\otimes\id_k$ is positive for all $k$, where $\id_k$ is the identity
function on $\C^{k\times k}$. If $F$ moreover satisfies
$\tr(F(A))\leq\tr A$ for all positive $A$, then it is called a {\em
superoperator}. The density matrices are precisely the positive
matrices $A$ of trace $\leq 1$. Moreover, the superoperators
correspond precisely to those functions from mixed states to mixed
states that are physically possible {\cite{nielsen02quantum,Selinger04}}.
\subsection{The category \texorpdfstring{$\CPM$}{CPM}}
The category $\CPMs$ is defined as follows: the objects are natural
numbers, and a morphism $F:n\to m$ is a completely positive map
$F:\C^{n\times n}\to\C^{m\times m}$. Let $\CPM$ be the free completion
of $\CPMs$ under finite biproducts; specifically, the objects of
$\CPM$ are sequences $\vec n=(n_1,\ldots,n_k)$ of natural numbers,
and a morphism $F:\vec n\to\vec m$ is a matrix $(F_{ij})$ of morphisms
$F_{ij}:n_j\to m_i$ of $\CPMs$. The categories $\CPMs$ and $\CPM$ are
symmetric monoidal, and in fact, compact closed {\cite{Selinger04}}.
\subsection{Limitations of \texorpdfstring{$\CPM$}{CPM} as a model}
\label{sec:rec-type}
The category $\CPM$ can serve as a
fully abstract model for a simple, strictly linear, finitary quantum
lambda calculus {\cite{valiron06fully}}. For example, the type $\bit$
is interpreted as $(1,1)$, and the type $\qubit$ is interpreted as
$(2)$. Measurement, as a map from $\qubit$ to $\bit$, sends
$(\begin{smallmatrix}a&b\\c&d\end{smallmatrix})$ to $(a,d)$. The coin
toss is a map $(1)\to(1,1)$ sending $(p)$ to $(\frac{p}2,\frac{p}2)$.
Function spaces are interpreted via the compact closed structure.
As mentioned in the introduction, the semantics of
{\cite{valiron06fully}} is extremely limited, because it is completely
finitary. Thus recursion, infinite data types, and non-linear
functions (i.e., those that can use their argument more than once) had
to be completely removed from the language in order to fit the
model. For example, even the simple squaring function $f\mapsto
\lambda x.f(f\,x)$ is not representable in $\CPM$.
The purpose of the present paper is to remove all of these
restrictions. As an example, consider the following pseudo-code (in
ML-style):
\begin{Verbatim}[fontsize=\footnotesize,commandchars=\\\{\}]
val\textrm{\bf qlist} : qubit -> qubit list
let rec\textrm{\tt\bf qlist} q = if (cointoss) then [q]
else let (x,y) = entangle q in x::(\textrm{\bf{}qlist} y)
\end{Verbatim}
Here, {\tt cointoss} is a fair coin toss, and the function
{\tt entangle} sends $\alpha\ket0+\beta\ket1$ to $\alpha\ket{00} +
\beta\ket{11}$.
So if the function ${\bf qlist}$ is applied to a qubit
$\alpha\ket0+\beta\ket1$, the output is $\alpha\ket0+\beta\ket1$ with
probability $\frac12$, $\alpha\ket{00}+\beta\ket{11}$ with probability
$\frac14$, $\alpha\ket{000}+\beta\ket{111}$ with probability
$\frac18$, and so on. Its semantics should be of type $2 \to
(2,4,8,\ldots)$, mapping
\[
\left(\begin{smallmatrix}a&b\\c&d\end{smallmatrix}\right)\mapsto
\left(
\frac12\left(\begin{smallmatrix}a&b\\c&d\end{smallmatrix}\right),
\frac14\left(\begin{smallmatrix}a&0&0&b\\0&0&0&0\\0&0&0&0\\
c&0&0&d\end{smallmatrix}\right)
,\ldots
\right).
\]
The category $\CPM$ is ``almost'' capable of handling this case, but
not quite, because it cannot express infinite tuples of matrices. The
model we propose in this paper is essentially an extension of $\CPM$
to infinite biproducts, using methods developed
in~\cite{Girard99coherentbanach,MelliesTT09,LairdMM12,LairdMMP13}.
\section{A quantum lambda calculus}\label{sec:qlc}
\begin{table}
\[
\begin{array}{l}
{\it Terms}\quad M,N,P
\quad {:}{:}{=}
\\[1ex]
\qquad x\bor \lambda x^A.M \bor MN\bor
\punit\bor\letunitterm{M}{N}\bor\\[1ex]
\qquad \tensterm{M}{N}\bor
\lettensterm{x^A}{y^B}{M}{N}\bor\\[1ex]
\qquad \injl M\bor\injr M\bor\match P{x^A}M{y^B}N\bor\\[1ex]
\qquad \splitlist[\!A]\bor
\letrec{f^{A{\multimap} B}}{\!x}{M\!}{\!N}
\bor\meas \bor \new \bor U
\\[3ex]
{\it Values}\quad V,W
\quad {:}{:}{=}
\\[1ex]
\qquad x \bor c \bor \lambda x^A.M \bor
\tensterm{V}{W}\bor\injl{V}\bor\injr{W}
\\[3ex]
{\it Types}\quad A, B, C
\quad{:}{:}{=}
\\[1ex]
\qquad \qubit\bor A{\loli}B\bor \bang{(A\,{\loli}\, B)}\bor
\tunit\bor A\,{\tensor}\,B \bor A\,{\sumtype}\, B\bor\tlist{A}.
\end{array}
\]
\caption{\footnotesize Grammars of terms, values and types.
}\label{table:terms_grammar}
\end{table}
We define a variant of the typed quantum lambda calculus
of~\cite{SV09}. The main difference is that the language in this
present paper is a true extension of linear logic (see the type
assignment system of Table~\ref{fig:typing rules}). In particular, in
contrast with~\cite{SV09}, $\oc( A\otimes B)\multimap\oc A\otimes \oc
B$ is not provable and there is no need for a subtyping relation.
The operational semantics implements a call-by-value strategy. An
untyped call-by-name variant has been studied in~\cite{LagoMZ11}.
The classes of \emph{terms}, \emph{values} and \emph{types} are
defined in Table~\ref{table:terms_grammar}. The symbol $c$ ranges over
the set of term constants $\{\punit, {\tt split}^A,$ $\meas, \new,
U\}$. The constant $U$ ranges over a set of elementary unitary
transformations on quantum bits. In the examples below, we will be
using the Hadamard gate $H$ and the controlled-not gate $N_c$, defined
as follows {\cite{nielsen02quantum}}:
\begin{align}
H&=\frac{1}{\sqrt 2}
\left(\begin{smallmatrix}
1&1\\
1&-1
\end{smallmatrix}\right)
&
N_c&=
\left(\begin{smallmatrix}
1&0&0&0\\
0&1&0&0\\
0&0&0&1\\
0&0&1&0
\end{smallmatrix}
\right)\label{eq:h_and_nc}
\end{align}
Notice that bound variables
are given in Church style, i.e., with a type annotation. This enables
Proposition~\ref{prop:unicity_derivation}, and simplifies the semantic
interpretation of the typed terms. We omit such
annotations in the sequel if uninteresting or obvious.
We have two kinds of arrows: the linear arrow $A{\multimap}B$, and the
intuitionistic arrow $\bang{(A\multimap B)}$, which is obtained by the
call-by-value translation of the intuitionistic implication into
linear logic \cite{ll}. Intuitively, only the terms of type
$\bang{(A\multimap B)}$ represent functions that can be used
repeatedly, whereas terms of type $A{\multimap}B$ must be used exactly
once.
A type of the form $\oc{A}$ is called a \emph{$\oc$-type} or
\emph{non-linear} type, and all other types are called {\em linear}.
The distinction between linear and non-linear types is crucial for
allowing the type system to enforce the no-cloning property of quantum
physics.
By convention, $\multimap$ is associative to the right, while
application and tensor are associative to the left. We use the
notation $A^{\tensor n}$ for $A$ tensored $n$ times. The type
$\tlist{A}$ denotes finite lists of type $A$. When doing structural
induction on types, we assume that $\tlist A$ is greater than
$A^{\otimes n}$, for any $n\in\N$.
The set of terms and types is somewhat spartan; however it
can be easily extended by introducing syntactic sugar. Note that,
for technical convenience, we have only allowed types of the form
$\bang{A}$ when $A$ is an arrow type. However,
for an arbitrary type $A$, the type $\bang{A}$ can be simulated
by using $\bang(\tunit\loli A)$ instead.
\begin{notation}\label{notation:syntax}
We write $\bit=\tunit\oplus\tunit$,
$\ttrue=\injr\punit$,
$\ffalse=\injl\punit$,
$\nil=\injl\punit$ and
$\cons MN=\injr (M\otimes N)$. We
write $\lambda\punit.M$ for the term $\lambda
z^{\tunit}.(\letunitterm zM)$, where $z$ is a fresh variable, and
$\iftermx PMN$ for $\match P{x^\tunit}{N}{y^\tunit}{M}$.
\end{notation}
\begin{table*}
\scalebox{0.9}{
\begin{minipage}{1.1\textwidth}
\[
\infer[ax]{\bang{\Delta},x:A\entail x:A}{\text{$A$ linear}}
\qquad
\infer[axd]{\bang{\Delta},x:\bang (A\loli B)\entail x:A\loli B}{}
\qquad
\infer[p]{\bang\Delta\entail V:\bang{(A\loli B)}}{
\bang\Delta\entail V:A\loli B
&
V\textrm{ value}
}
\qquad
\infer[\tunit_I]{
\bang{\Delta}\entail \punit:\tunit
}{}
\]
\[
\infer[\loli_I]{\Delta\entail\lambda x^A.M:A\loli B}{
\Delta,x:A\entail M:B}
\quad
\infer[\loli_E]{\bang{\Delta},\Gamma,\Sigma\entail MN:B}{
\bang\Delta,\Gamma\entail M:A\loli B
&
\bang\Delta,\Sigma\entail N:A
}
\quad
\infer[\tunit_E]{
\bang\Delta,\Gamma,\Sigma\entail
\letunitterm{M}{N}:A
}{
\bang\Delta,\Gamma\entail M:\tunit
&
\bang\Delta,\Sigma\entail N:A
}
\]
\[
\infer[\tensor_I]{
\bang{\Delta},\Gamma,\Sigma\entail \tensterm{M}{N}:A\tensor B
}{
\bang\Delta,\Gamma\entail M:A
&
\bang\Delta,\Sigma\entail N:B
}
\quad
\infer[\tensor_E]{
\bang{\Delta},\Gamma,\Sigma\entail
\lettensterm{x^{A}}{y^{B}}{M}{N}:C
}{
\bang\Delta,\Gamma\entail M:A\tensor B
&
\bang\Delta,\Sigma,x:A,y:B\entail N:C
}
\]
\[
\infer[\oplus_{I}^\ell]{
\bang\Delta,\Gamma\entail
\injl{M}:A\sumtype B
}{
\bang\Delta,\Gamma\entail M:A
}
\quad
\infer[\oplus_{I}^r]{
\bang\Delta,\Gamma\entail
\injr{M}:A\sumtype B
}{
\bang\Delta,\Gamma\entail M:B
}
\quad
\infer[\oplus_E]{
\bang{\Delta},\Gamma,\Sigma\entail \match{P}{x^A}{M}{y^B}N:C
}{
\bang\Delta,\Gamma\entail P:A\sumtype B
&
\deduce{\bang\Delta,\Sigma,y:B\entail N:C}{\bang\Delta,\Sigma,x:A\entail M:C}
}
\]
\vspace{-1ex}
\[
\infer[\tlist{-\!}_{\!\!I}]{
\bang\Delta,\Gamma\entail
M:\tlist{A}
}{
\bang\Delta,\Gamma\entail M:\tunit\oplus(A{\otimes}\tlist A)
}
\quad
\infer[\splitlist]{
\bang\Delta\entail\splitlist[A] :
\tlist A {\multimap}\tunit\oplus(A{\otimes}\tlist A)
}{}
\quad
\infer[{\tt rec}]{
\bang{\Delta},\Gamma\entail \letrec{f^{A\multimap B}}{x}{M}{N}:C
}{
\bang\Delta,f:\bang{(A\multimap B)},x:A\entail M:B
&
\bang\Delta,\Gamma,f:\bang{(A\multimap B)}\entail N:C
}
\]
\[
\infer[\meas]{
\bang\Delta\entail
\meas:\qubit\multimap \bit
}{}
\quad
\infer[\new]{
\bang\Delta\entail
\new:\bit\multimap\qubit
}{}
\quad
\infer[U]{
\bang\Delta\entail
U:\qubit^{\otimes n}\multimap\qubit^{\otimes n}
}{U \text{ of arity $n$}}
\]
\end{minipage}
}
\caption{\footnotesize Typing rules. The contexts $\Gamma$ and
$\Sigma$ are assumed to be linear.}\label{fig:typing rules}
\end{table*}
A \emph{context} $\Delta$ is a function from a finite set of variables
to types. We denote the domain of $\Delta$ by $\supp\Delta$, and we
write $\Delta=x_1:A_1,\dots, x_n:A_n$ whenever
$\supp\Delta=\{x_1,\dots,x_n\}$ and $\Delta(x_i)=A_i$. We call
$\Delta$ \emph{exponential} (resp.\ \emph{linear}) whenever all $A_i$
are $\oc$-types (resp.\ no $A_i$ is a $\oc$-type). We write
$\oc\Delta$ for a context that is exponential. The notation $\Gamma,
\Sigma$ refers to the union of the two contexts $\Gamma$ and $\Sigma$
and assumes that $\supp{\Gamma}$ and $\supp\Sigma$ are disjoint.
A \emph{judgement} is a triple $\Gamma\vdash M:A$ of a context
$\Gamma$, a term $M$ and a type $A$. A judgement is called \emph{valid}
if it can be inferred from the typing rules in
Figure~\ref{fig:typing rules}, using the convention that the contexts
$\Gamma$ and $\Sigma$ are linear.
\begin{proposition}\label{prop:unicity_derivation}
There is at most one derivation inferring a given typing judgement
$\Gamma\vdash M:A$.
\qed
\end{proposition}
\begin{example}\label{ex:term_type}
In Section~\ref{sec:rec-type}, we wrote the informal program ${\bf
qlist}$. Our language is expressive enough to represent it. The
term $\mathtt{cointoss}$ can be defined as $\meas (H (\new\,
\ttrue))$, and it has type $\bit$. The term $\mathtt{entangle}$ is
$\lambda x^\qubit.N_c(x\otimes(\new\,\ffalse))$, which has type
$\qubit\multimap\qubit\otimes\qubit$. Then, ${\bf qlist}$ is
\begin{multline*}
\mathtt{letrec}\; f^{\qubit\multimap\qubit^\ell} q =\\
\mathtt{if}\; \mathtt{cointoss}\; \mathtt{then}\;
\cons q\nil\hspace{2.5cm}\\
\mathtt{else}\; \mathtt{let}\; x^\qubit\otimes y^\qubit
=\mathtt{entangle}\; q\; \mathtt{in} \;\cons {x}{fy}
\end{multline*}
which has type $\qubit\multimap\qubit^\ell$. In
Examples~\ref{ex:term_red} and~\ref{ex:term_sem} we discuss its
operational and denotational semantics, respectively.
\end{example}
\begin{example}\label{ex:telep-term}
In Example~\ref{ex:telep-presentation} and Figure~\ref{fig:telep},
we sketched the quantum teleportation algorithm. We said that the
algorithm can be decomposed into 3 parts. Each of these parts can be
described and typed in the quantum lambda calculus, yielding a
higher-order term. This is an adaptation of an example provided
in~\cite{selinger06lambda}.
\begin{itemize}
\item[(i)] generates an EPR pair of entangled quantum bits. Its type
is therefore $\tunit\loli\qubit\tensor\qubit$. The corresponding
term is
\[
{\bf EPR} = \lambda\punit. N_c\left((H(\new\,\ffalse))\tensor(\new\,
\ffalse)\right).
\]
\item[(ii)] performs a Bell measurement on two quantum bits and
outputs two classical bits $x,y$. Its type is thus
$\qubit\loli\qubit\loli\bit\tensor\bit$, and the term {\bf
BellMeasure} is defined as
\[
\lambda q_1.\lambda q_2.
\left(\begin{array}{l}
{\tt let}\ x\tensor y = N_c\,(q_1\tensor q_2)
\\
{\tt in}\ (\meas\,(H\,x))\tensor(\meas\,y)
\end{array}\right).
\]
\item[(iii)] performs a correction. It takes one quantum bit, two
classical bits, and outputs a quantum bit. It has a type of the form
$\qubit\loli\bit\tensor\bit\loli\qubit$. The term is
\[ \begin{array}{r}
{\bf U} = \lambda q.\lambda x\tensor y.
\mbox{\tt if\,$x$\,then\,$($if\,$y$\,then\,$U_{11}\,q$\,else\,$U_{10}\,q)$}
\\
\mbox{\tt
else\,$($if\,$y$\,then\,$U_{01}\,q$\,else\,$U_{00}\,q)$}.
\hspace{-.8ex}
\end{array}
\]
\end{itemize}
We can now write the term
\[
{\bf telep} = \begin{array}[t]{l}
\begin{array}[t]{l@{}l@{~}l}
\lambda\punit.&{\tt let}~{x\tensor y}&= {\bf EPR}~\punit~{\tt in}\\
&{\tt let}\ {f} &= {\bf BellMeasure}\ x~{\tt in}\\
&{\tt let}\ {g} &= {\bf U}\ y
\end{array}\\
\quad\qquad\hspace{1.3ex}{\tt in}\ {f\tensor g}.
\end{array}
\]
It can then be shown that
\[
\entail {\bf telep}:!(\tunit\loli (\qubit \loli
\bit\otimes\bit)\tensor(\bit\otimes\bit \loli \qubit))
\]
is a valid typing judgement.
In other words, the teleportation algorithm produces a pair of
entangled functions $f:\qubit\to\bit\tensor\bit$ and
$g:\bit\tensor\bit\to\qubit$. These functions have the property that
$g(f(\ket\phi))=\ket\phi$ for
all qubits $\ket\phi$, and $f(g(x\tensor y))=(x\tensor y)$ for all
booleans $x$ and $y$. These two functions are each other's inverse,
but because they contain an embedded qubit each, they can only be
used once. They can be said to form a ``single-use isomorphism''
between the (otherwise non-isomorphic) types $\qubit$ and
$\bit\tensor\bit$. However, the whole procedure is duplicable: one
can generate as many one-time-use isomorphism pairs as desired.
\end{example}
\subsection{Operational semantics}
\label{sec:op-semant}
The operational semantics is defined in terms of an abstract machine
simulating the behavior of Knill's QRAM model~\cite{knill}. It is
similar to the semantics given in \cite{SV09}.
\begin{definition}
\label{def:qclos}
A \emph{quantum closure} is a triple $\am{\qarray,\qlist,M}$ where
\begin{itemize}
\item $\qarray$ is a normalized vector of $\C^{2^n}$, for some
integer $n\geq 0$. The vector $\qarray$ is called the \emph{quantum
state};
\item $M$ is a term, not necessarily closed;
\item $\qlist$ is a one-to-one map from the set of free variables of
$M$ to the set $\{1,\ldots, n\}$. It is called the \emph{linking
function}.
\end{itemize}
We write $\supp{\qlist}$ for the domain of $\qlist$.
By abuse of language we may call a closure
$\am{\qarray,\qlist,V}$ a \emph{value} when the term $V$ is a value.
We denote the set of quantum closures by $\Cl$ and the set of
quantum closures that are values by $\Val$.
We write $\ell|_M$ for the linking function whose domain is
restricted to the set of free variables of $M$.
We say that the quantum closure $\am{\qarray,\qlist,M}$ is
\emph{total} when $|\qlist |$ has cardinality $n$, the size of the
quantum state. In that case, if $|\qlist |=\{x_1,\ldots, x_n\}$ and
$\qlist(x_i)=i$, we write $\qlist$ as $\ket{x_1,\ldots, x_n}$. A
quantum closure $\am{\qarray,\ket{x_1,\ldots, x_n},M}$ \emph{has a
type $A$}, whenever $x_1:\qubit,\dots,x_n:\qubit\vdash M:A$. In
case $\qlist=\ket{x_1,\ldots, x_n}$ we can also write $\qlist\vdash
M:A$.
\end{definition}
The purpose of a quantum closure is to provide a mechanism to talk
about terms with embedded quantum data. The idea is that a variable
$y\in\FV(M)$ is bound in the closure $\am{\qarray,\qlist,M}$ to qubit
number $\qlist(y)$ of the quantum state $\qarray$. So for example,
the quantum closure
$
\am{\frac1{\sqrt2}(\ket{00}+\ket{11}),\ket{x_1,x_2},\lambda y^A.yx_1x_2}
$
denotes a term $\lambda y^A.yx_1x_2$ with two embedded qubits $x_1$, $x_2$ in
the entangled state $\ket{x_1x_2}=\frac1{\sqrt2}(\ket{00}+\ket{11})$.
The notion of $\alpha$-equivalence extends naturally to quantum
closures, for instance, the states $\am{\qarray,\ket{x},\lambda
y^A.x}$ and $\am{\qarray,\ket{z},\lambda y^A.z}$ are
equivalent. From now on, we tacitly
identify quantum closures up to renaming of bound variables.
The evaluation of a term is defined as a probabilistic rewriting
procedure on quantum closures, using a call-by-value reduction
strategy. We use the notation $\am{\qarray,\qlist,
M}\redto[p]\am{\qarray',\qlist', M'}$ to mean that the left-hand side
closure reduces in one step to the right-hand side with probability
$p\in[0,1]$.
\begin{table*}
\footnotesize
\centering
\subfloat[Classical control.]{
\label{subtable:reduction_classical}
\centering
\parbox{.95\textwidth}{
\centering
\begin{align*}
\am{\qarray{},\qlist{},(\lambda x^A.M)\,V}
&{}\redto[1]
\am{\qarray{},\qlist{},M\{V/x\}}
&
\am{\qarray{},\qlist{},\lettensterm{x^A}{y^B}{V\tensor W}{N}}
&{}\redto[1]
\am{\qarray{},\qlist{},N\{V/x, W/y\}}
\\
\am{\qarray{},\qlist{},\letunitterm{\punit}{N}}
&{}\redto[1]
\am{\qarray{},\qlist{},N}
&
\am{\qarray{},\qlist{},\match{(\injl V)}{x^{\!A}}{M}{y^{\!B}}{N}}
&{}\redto[1] \am{\qarray{},\qlist{},M\{V/x\}}
\\
\am{\qarray{},\qlist{},\splitlist V}
&{}\redto[1] \am{\qarray{},\qlist{},V}
&
\am{\qarray{},\qlist{},\match{(\injr V)}{x^{\!A}}{M}{y^{\!B}}{N}}
&{}\redto[1] \am{\qarray{},\qlist{},N\{V/y\}}
\end{align*}
\vspace{-4ex}
\begin{align*}
\am{\qarray{},\qlist{},\letrec{f^{A\multimap B}}{x}{M}{N}}
&{}\redto[1] \am{\qarray{},\qlist{},N\{(\lambda
x^A.\letrec{f^{A\multimap B}}{x}{M}{M})/f\}}
\end{align*}
}
}
\subfloat[Quantum data. The variable $y$ is fresh. The decomposition
of the quantum array in the case of $\meas\, x$ is explained in
Definition~\ref{def:rw}.]{
\label{subtable:reduction_quantum}
\centering
\parbox{.95\textwidth}{
\centering
\begin{align*}
\am{\qarray{},\qlist, U(x_{1}\otimes\dots\otimes x_{k})}
&{}\redto[1]
\am{\qarray{}',\qlist,x_{1}\otimes\dots\otimes x_{k}}
\end{align*}
\vspace{-4ex}
\begin{align*}
\am{\qarray,\emptyset,\new~\ffalse}
&{}\redto[1]
\am{\qarray\otimes\ket0,\{y\mapsto n+1\}, y}
&
\am{\alpha\qarray_0+\beta\qarray_1,\{x\mapsto i\},\meas~x}
&{}\xredto[\abs{\beta}^2]
\am{\qarray'_1,\emptyset, \ttrue}
\\
\am{\qarray,\emptyset,\new~\ttrue}
&{}\redto[1]
\am{\qarray\otimes\ket{1},\{y\mapsto n+1\}, y}
&
\am{\alpha\qarray_0+\beta\qarray_1,\{x\mapsto i\},\meas~x}
&{}\xredto[\abs{\alpha}^2]
\am{\qarray'_0,\emptyset, \ffalse}
\end{align*}}
}
\\
\subfloat[Congruence rules, under the hypothesis that for some
$\ell_0$ we have $\ell=\ell_0 \uplus \ell|_M$, $\ell'=\ell_0 \uplus
\ell'|_{M'}$ and
$\am{\qarray,\qlist|_M,M}{\redto[p]}\am{\qarray',\qlist'|_{M'},M'}$.
]{\label{subtable:reduction_congruence}
\centering
\parbox{.95\textwidth}{
\centering
\begin{align*}
\am{\qarray,\qlist{},MN}&\redto[p]\am{\qarray',\qlist{}',M'N}
&
\am{\qarray,\qlist{},M\otimes N}&
\redto[p]\am{\qarray',\qlist{}',M'\otimes N}
&
\am{\qarray,\qlist{},\injl M}&
\redto[p]\am{\qarray',\qlist{}',\injl M'}
\\
\am{\qarray,\qlist{},VM}&\redto[p]\am{\qarray',\qlist{}',VM'}
&
\am{\qarray,\qlist{},V\otimes M}&
\redto[p]\am{\qarray',\qlist{}',V\otimes M'}
&
\am{\qarray,\qlist{},\injr M}&
\redto[p]\am{\qarray{}',\qlist{}',\injr M'}
\end{align*}
\vspace{-4ex}
\begin{align*}
\am{\qarray,\qlist{},\letunitterm{M}{N}}&
\redto[p]\am{\qarray',\qlist{}',\letunitterm{M'}{N}}
&
\am{\qarray,\qlist{},\lettensterm{x^A}{y^B}{M}{N}}&
\redto[p]\am{\qarray',\qlist{}',\lettensterm{x^A}{y^B}{M'}{N}}
\end{align*}
\vspace{-4ex}
\begin{align*}
\am{\qarray,\qlist{},\match
M{x^A}P{y^B}{N}}&\redto[p]\am{\qarray{}',
\qlist{}',\match{M'}{x^A}P{y^B}{N}}
\end{align*}
}
}
\caption{\footnotesize Reduction rules on closures.}
\label{table:reduction}
\end{table*}
\begin{definition}\label{def:rw}
The reduction rules are shown in Table~\ref{table:reduction}. The
rules split into three categories:
\subref{subtable:reduction_classical} rules handling the classical
part of the calculus; \subref{subtable:reduction_quantum} rules
dealing with quantum data; and
\subref{subtable:reduction_congruence} congruence rules for the
call-by-value strategy. Note that in the statement of the rules, $V$
and $W$ refer to values.
In the rules in
Table~\ref{table:reduction}\subref{subtable:reduction_quantum},
the quantum state $q$
has size $n$. The
quantum state $\qarray'$ in the first rule is obtained by applying the
$k$-ary unitary gate $U$ to the qubits
$\qlist(x_{1}),\dots,\qlist(x_{k})$. Precisely, $\qarray'=(\sigma\circ
(U\otimes\id)\circ\sigma^{-1})(\qarray)$, where $\sigma$ is the action
on $\C^{2^n}$ of any permutation over $\{1,\dots, n\}$ such that
$\sigma(i)=\qlist(x_i)$ whenever $i\leq k$. In the rules about
measurements, we assume that if $\qarray_0$ and $\qarray_1$ are
normalized quantum states of the form
\begin{equation}\label{eq:def-op-meas1}
{\textstyle\sum_j}\alpha_j\ket{\phi_j}\otimes\ket0\otimes\ket{\psi_j},
~~
{\textstyle\sum_j}\beta_j\ket{\phi_j}\otimes\ket1\otimes\ket{\psi_j},
\end{equation}
then $\qarray'_0$ and $\qarray'_1$ are respectively
\begin{equation}\label{eq:def-op-meas2}
{\textstyle\sum_j}\alpha_j\ket{\phi_j}\otimes\ket{\psi_j},
~~
{\textstyle\sum_j}\beta_j\ket{\phi_j}\otimes\ket{\psi_j},
\end{equation}
where the vectors $\phi_j$ have dimension $\qlist(x)-1$ (so
that the measured qubit is $\ell(x)$).
\end{definition}
In summary, the quantum state acts as a shared global store that is
updated destructively by the various quantum operations.
Note that the only probabilistic reduction step is the one
corresponding to measurement. Also, we underline that the hypothesis
associated with a congruence rule
$\am{\qarray,\qlist,C[M]}{\redto[p]}\am{\qarray',\qlist',C[M']}$ takes
into account the whole quantum states $\qarray$ and $\qarray'$. In
fact, because of the entanglement, the evaluation of
$\am{\qarray,\qlist|_M,M}$ may have a side-effect on the state of the
qubits pointed to by the variables occurring in the context $C[\,]$.
The rules assume that the involved
closures are well-defined. In particular, whenever
$\am{\qarray,\qlist, M}\redto[p]\am{\qarray,\qlist, M'}$, the two
terms $M$ and $M'$ have the same free variables. For example, the
closure $\am{\ket{00},\ket{yz},(\lambda x.y)z}$ cannot reduce and it
represents an error: it would reduce to the erroneous quantum closure
$\am{\ket{00},\ket{yz},z}$,
where the domain of the linking function is not the set of free
variables, as specified by Definition~\ref{def:qclos}.
The type system will prevent such an error as
proven in Proposition~\ref{prop:safety}.
\begin{example}\label{ex:term_red}
Recall Example~\ref{ex:term_type}. We have
$\am{\ket{},\ket{},\mathtt{cointoss}}\redto[1]\am{\ket{1},\ket{x},\meas(H
x)}\redto[1]\am{\frac1{\sqrt 2}(\ket{0}+\ket{1}),\ket{x},\meas\;
x}$, the latter reducing to either $\am{\ket{},\ket{},\ttrue}$ or
$\am{\ket{},\ket{},\ffalse}$, with equal probability $\frac12$. As
for $\mathtt{entangle}$, we have that
\begin{align*}
&\am{\alpha\ket0+\beta\ket{1},\ket x,\mathtt{entangle}\;
x}
\\
\redto[1]~~&
\am{\alpha\ket0+\beta\ket{1},\ket
x,N_c(x\otimes(\new\;\ffalse))}
\\
\redto[1]~~&
\am{\alpha\ket{00}+\beta\ket{10},\ket{xy},N_c(x\otimes
y)}
\\
\redto[1]~~&
\am{\alpha\ket{00}+\beta\ket{11},\ket{xy},x\otimes
y}.
\end{align*}
Similarly, one can check that
$\am{\alpha\ket0+\beta\ket1,\ket q,\mathtt{qlist}\, q}$ behaves as
described in Section~\ref{sec:rec-type}, reducing to
$\am{\alpha\ket0+\beta\ket1,\ket q,\cons q\nil}$ with probability
$\frac 12$, to
$\am{\alpha\ket{00}+\beta\ket{11},\ket{qq'},\cons{q'}{\cons q
\nil}}$ with probability $\frac 14$, etc. In particular, notice
that in any single reduction sequence the variable $q$ has not been
duplicated, as correctly asserted by the type of $\mathtt{qlist}$.
\end{example}
\begin{lemma}[Substitution]\label{lemma:substitution}
Suppose $\oc\Delta,\Gamma,x:A\entail M:B$ and $\oc\Delta,\Sigma\entail
V:A$, where $\Gamma$ and $\Sigma$ are linear contexts with disjoint
domain. Then $\oc\Delta,\Gamma,\Sigma\entail M\{V/x\}:B$.\qed
\end{lemma}
\begin{proposition}[Subject reduction]\label{prop:subject_reduction}
When $\am{\qarray,\ket{y_1\dots y_n},
M}\redto[p]\am{\qarray',\ket{x_1\dots x_{n'}}, M'}$ and
$y_1:\qubit,\dots, y_n:\qubit\entail M:A$, then $x_1:\qubit,\dots,
x_{n'}:\qubit\entail M':A$.\qed
\end{proposition}
\begin{proposition}[Type safety]\label{prop:safety}
If $\am{\qarray,\qlist, M}$ is typable then either $M$ is a value or
there is a closure $\am{\qarray',\qlist', M'}$ such that
$\am{\qarray,\qlist, M}\redto[p]\am{\qarray',\qlist', M'}$.
Moreover, if $M$ is not a value, the total probability of all
possible single-step
reductions from $\am{\qarray,\qlist, M}$ is $1$. \qed
\end{proposition}
\begin{lemma}[Totality]\label{lemma:totality}
If $\am{\qarray,\qlist, M}\redto[p]\am{\qarray',\qlist', M'}$
and $\am{\qarray,\qlist, M}$ is total, then $\am{\qarray',\qlist',
M'}$
is total too.
\end{lemma}
\begin{proof}
By induction on a derivation of $\am{\qarray,\qlist, M}
\redto[p]\am{\qarray',\qlist', M'}$, one proves that
$
\dim(\qarray')=\dim(\qarray)+\dim(\qlist')-\dim(\qlist)
$
where $\dim(\qarray)$ is the size of the quantum state $\qarray$
and $\dim(\qlist)$ is the cardinality of the domain set of the
linking function $\qlist$. Then, one gets the statement, since
$\am{\qarray,\qlist, M}$ is total iff $\dim(\qarray)=\dim(\qlist)$.
\end{proof}
\begin{notation}
The reduction relation $\redto$ defines the probability that a
closure reduces to another one in a single step. We extend this
relation to an arbitrary large (but finite) number of reduction
steps with the notation
$\Red^n_{\am{\qarray,\qlist,M},\am{q',\qlist',V}}$: it is the total
probability of $\am{\qarray,\qlist,M}$ reducing to a value
$\am{q',\qlist',V}$. It is defined as the sum of all
$\prod_{i=1}^mp_i$, where $
\am{\qarray,\qlist,M}\redto[p_1]\am{\qarray_1,\qlist_1,M_1}
\cdots\redto[p_m]\am{\qarray,\qlist',V}
$ is a finite reduction sequence of $m\leq n$ steps. We write
$\Red^\infty_{\am{\qarray,\qlist,M},\am{q',\qlist',V}}$ for the sup
over $n$ of $\Red^n_{\am{\qarray,\qlist,M},\am{q',\qlist',V}}$.
Finally, we define the \define{total probability}
$\Halt_{\am{\qarray,\qlist,M}}$ of $\am{\qarray,\qlist,M}$
converging to any value as
$
\sum_{\am{\qarray',\qlist',V}\in{\Val}}
\Red_{\am{\qarray,\qlist,M},\am{\qarray',\qlist',V}}^\infty.
$
\end{notation}
\section{Denotational semantics}\label{sec:sem}
We interpret the quantum lambda calculus in a suitable extension
$\freecat[\ccpms]$ of the category
$\CPM$ described in Section~\ref{sec:background}.
What $\CPM$ essentially misses is the linear logic exponential $\oc
A$, and our plan is to introduce it via the equation
\begin{equation}\label{eq:bang_biproduct}
\oc A\ass\bigoplus_{k=0}^\infty A^{\odot k},
\end{equation}
where $\bigoplus_{k=0}^\infty$ is the infinite biproduct of the family
$\{A^{\odot k}\}_k$, each $A^{\odot k}$ being the symmetric $k$-fold
tensor power of $A$, i.e., the equalizer of the $k!$ symmetries of the
$k$-ary tensor $A^{\otimes k}\ass A\otimes\dots\otimes A$.
The category $\CPM$ cannot express this equation
because it lacks both infinite biproducts and a convenient definition
of symmetric tensor
powers. The category $\freecat[\ccpms]$ is in some sense the minimal
extension of $\CPM$ having these two missing
ingredients.
The plan of the section is as follows.
Section~\ref{sect:preli_cpm} presents some preliminary material.
Section~\ref{sect:thecategory} defines
$\freecat[\ccpms]$ and Section~\ref{sect:lafont_ccpms} develops the categorical
structure allowing us to interpret the quantum
lambda calculus. Section~\ref{Sect:soundness} sketches the proof of
the soundness of the model with respect to the operational semantics.
Finally, Section~\ref{sect:examples}
discusses the denotations of the programs {\bf qlist} and {\bf
teleport}.
\subsection{Preliminaries: from \texorpdfstring{\CPM{}}{CPM} to
\texorpdfstring{$\ccpms$}{bar(CPMs)}}\label{sect:preli_cpm}
\paragraph{Permutation groups.} Let $\symgroup_n$ be the symmetric
group of degree $n$, i.e., the group of permutations of
$n=\{0,\dots,n-1\}$. Any permutation $g\in\symgroup_n$ gives rise to a
matrix $P_g\in\C^{n\times n}$, defined by $P_g(e_i)=e_{g(i)}$, where
$e_i$ is the $i$th standard basis vector. We define an action of $g$
on $\C^{n\times n}$ by $g\cdot M := P_gMP_g\inv$. Moreover, for a
subgroup $G\subseteq\symgroup_n$, we define
\begin{equation}\label{eq:action_permutation}
G\cdot M := \frac{1}{\#G}\sum_{g\in G}g\cdot M,
\end{equation}
where $\#G$ is the number of elements of $G$.
\begin{lemma}
Given a subgroup $G\subseteq\symgroup_n$, its action on $\C^{n\times
n}$ is idempotent (i.e., $G\cdot G\cdot M=G\cdot M$ for all $M$)
and completely positive.
\end{lemma}
\begin{proof}
For the idempotence, notice that for every $g\in G$, $gG=G$,
therefore: $G\cdot G\cdot M = \frac{1}{\#G}\sum_{g\in G} gG\cdot M =
G\cdot M$.
The complete positivity of $G$ is derived from the complete positivity
of each map $M\mapsto g\cdot M = P_gMP_g\inv$.
\end{proof}
In the sequel, we use the notation $G$ both for a subgroup of
$\symgroup_n$ and for the completely positive map defined by it. The
above Lemma allows us to define the set of completely positive maps
from $\C^{n\times n}$ to $\C^{m\times m}$ invariant under the actions
of two subgroups $G\subseteq\symgroup_n$, $H\subseteq\symgroup_m$ by
\[
\cpms(G,H)\ass\{f\in\CPM(n,m)\such G\pc f\pc H=f\},
\]
where $f;g$ is the diagrammatic composition $(f;g)(x) = g(f(x))$, and
$\CPM(n,m)$ is the set of completely positive maps from $\C^{n\times
n}$ to $\C^{m\times m}$.
\paragraph{Completion of the L\"owner positive cone.} The set
$\cpms(G,H)$ is a module over the semi-ring $\Rp$ of the non-negative
real numbers. The L\"owner order $\sqleq$ on completely positive
maps~\cite{Selinger04}{} endows this module with the structure of a
\emph{bounded} directed complete partial order (bdcpo), i.e., there is
a minimum element (the zero function $\mathbf 0$), and any directed
set $D$ that is bounded (i.e., such that there exists $f\in
\cpms(G,H)$ such that for all $g\in D$, $g\sqleq f$) has a least upper
bound $\bigvee D\in\cpms(G,H)$. However there exist unbounded directed
subsets in $\cpms(G,H)$. We therefore need to complete $\cpms(G,H)$ to
a dcpo.
The relevant construction is the {\em D-completion} of
{\cite{ZhaoFan2010}}, which we briefly recall. Given any poset $P$,
say that a subset $X$ is {\em Scott-closed} if it is down-closed and
for every directed $I\seq S$, if the least upper bound
$\bigvee I$ exists in $P$, then $\bigvee
I\in S$. We say that a monotone function between posets $f:P\to Q$ is
{\em Scott-continuous} if it preserves all {\em existing} least upper
bounds of directed subsets. Let $\Gamma(P)$ be the set of Scott-closed
subsets of $P$; this forms a dcpo under the subset ordering. The {\em
D-completion} $\cs(P)$ is defined to be the smallest sub-dcpo of
$\Gamma(P)$ containing all sets of the form $\down x$. Then $\cs(P)$
is a dcpo, and there is a canonical injective Scott-continuous map
$\iota:P\to\cs(P)$, defined by $\iota(x)=\down x$, which allows us to
regard $P$ as a subset of $\cs(P)$. The D-completion preserves all
existing least upper bounds of directed sets, is idempotent, and
satisfies the following universal property: given any other dcpo $E$
and Scott-continuous map $f:P\to E$, there exists a unique
Scott-continuous $g:\cs(P)\to E$ such that $f=\iota\pc g$. It follows
that the D-completion is functorial. Moreover, if $P$ is a bounded
directed complete partial order, then $P$ is an initial subset of
$\cs(P)$, i.e., the only new elements added by the completion are ``at
infinity''. We call these the {\em infinite} elements of $\cs(P)$.
The homset $\cpms(G,H)$ is then extended by D-completion, namely,
$\overline\cpms(G,H) :=\cs(\cpms(G,H))$. The categorical operations
are extended in the unique Scott-continuous way, using the universal
property of D-completion. This allows us to define indexed sums over
$\ccpms(G,H)$, as follows. If $\{f_i\}_{i\in
I}\subseteq\overline\cpms(G,H)$ is a (possibly infinite) indexed
family, $\sum_{i\in I}f_i$ is defined as $\dirsup
_{F\subseteq_{\mathrm{fin}}I}\bigl(\sum_{i\in F}f_i\bigr)$. Indeed,
the set $\{\sum_{i\in F}f_i\;;\;F\subseteq_{\mathrm{fin}}I\}$ is
always directed, so has a least upper bound in the order completion
$\overline\cpms(G,H)$ of $\cpms(G,H)$.
\subsection{The category
\texorpdfstring{$\freecat[\ccpms]$}{bar(CPMs)+}}
\label{sect:thecategory}
Given a set $A$ and $a,a'\in A$, define the \emph{Kronecker symbol}
$\delta_{a,a'}\in\N$ which takes value $1$ if $a=a'$ and $0$ if $a\neq
a'$.
\begin{description}
\item[Objects] are given by indexed families $\qfin A=\{(\arity{\qfin
A}_a, \symm{\qfin A}_a)\}_{a\in\web{\qfin A}}$, where the index
set $\web{\qfin A}$ is called the \emph{web} of $\qfin A$ and, for
every $a\in\web{\qfin A}$, $\arity{\qfin A}_a$ is a natural
non-negative integer, and $\symm{\qfin A}_a$ a subgroup of
permutations of degree~$\arity{\qfin A}_a$, called respectively the
\emph{dimension} and the \emph{permutation group} of $\qfin A_a$.
\item[Morphisms] from $\qfin A$ to $\qfin B$ are matrices $\phi$
indexed by $\web{\qfin A}\times\web{\qfin B}$ and such that
$\phi_{a,b}\in\ccpms(\symm{\qfin A}_a, \symm{\qfin B}_b)$.
\item[Composition] of $\phi\in\freecat[\ccpms](\qfin A,\qfin B)$ and
$\psi\in\freecat[\ccpms](\qfin B,\qfin C)$ is the matrix
$\phi\pc\psi$ defined by, for $a\in\web{\qfin A}$ and
$c\in\web{\qfin C}$, $ (\phi\pc\psi)_{a,c}\ass\sum_{b\in\web{\qfin
B}}\phi_{a,b}\pc\psi_{b,c}. $
\item[Identity] is the diagonal matrix built with the symmetries of
$\qfin A$, i.e., for $a,a'\in\web{\qfin A}$, $\id^\qfin
A_{a,a'}\ass\delta_{a,a'}\symm{\qfin A}_a$.
\end{description}
The description of the objects and the morphisms as indexed families
is crucial for inferring the structure of a compact closed Lafont
category (Section~\ref{sect:lafont_ccpms}). However, it is worthwhile
to notice that $\freecat[\ccpms]$ can also be presented as a concrete
category of modules and linear maps between modules. Let us sketch
such an alternative presentation.
Let $\qfin A$ be an object of $\freecat[\ccpms]$. We define a module
$\pmatr(\qfin A)$ over $\overline{\Rp}=\Rp\cup\{\infty\}$ as
follows. For every $a$ in $\web{\qfin A}$, let us write $\pmatr(a)$
for the cone of the positive matrices in $\symm{\qfin
A}_a(\C^{\arity{\qfin A}_a\times \arity{\qfin A}_a})$, this latter
being the subspace of the matrices in $\C^{\arity{\qfin A}_a\times
\arity{\qfin A}_a}$ invariant under $\symm{\qfin A}_a$. This
positive cone $\pmatr(a)$ is an $\Rp$-module. We then define:
\begin{equation}\label{eq:pos_obj}
\pmatr(\qfin A)\ass
\bigoplus_{a\in\web{\qfin A}}(\cs(\pmatr(a))\}).
\end{equation}
In fact, we have that $\pmatr(a)\simeq\cpms(\symgroup_1,\symm{\qfin
A}_a)$ and $\pmatr(\qfin A)\simeq\bigoplus_{a\in\web{\qfin
A}}\ccpms(\symgroup_1,\symm{\qfin A}_a)$. Hence, $\pmatr(\qfin A)$
is a continuous module over $\overline{\Rp}$: addition and scalar
multiplication are defined pointwise and are continuous operations
with respect to the L\"owner order.
Let $f:\pmatr(\qfin A)\to \pmatr(\qfin B)$ be a continuous
module homomorphism. We say that $f$ is \define{completely positive}
if all the module homomorphisms $f_{a,b} = \inj{a}\pc f\pc\proj b$ are
completely positive maps, for all $a\in\web{\qfin A}$ and
$b\in\web{\qfin B}$. (Indeed, since the positive matrices span the
complex vector space of square matrices (of corresponding size), one
can canonically extend the definition of complete positivity to module
homomorphisms $\pmatr(a)\to\pmatr(b)$).
\begin{proposition}\label{prop:concrete_cpms_pi}
There is an isomorphism between the homset
$\homof{\freecat[\ccpms]}{\qfin A,\qfin B}$ and the continuous
module homomorphisms from $\pmatr(\qfin A)$ to $\pmatr(\qfin B)$
that are completely positive. \qed
\end{proposition}
\subsection{\texorpdfstring{$\freecat[\ccpms]$}{bar(CPMs)+}
as a model of the quantum lambda calculus}
\label{sect:lafont_ccpms}
A compact closed category is a special case of symmetric monoidal
closed category. A symmetric monoidal closed category with finite
products, such that each object has a corresponding free commutative
comonoid, is called a \emph{Lafont category}, which is known to be a
model of intuitionistic linear logic
\cite{lafont:these,melliespanorama}. The category $\freecat[\ccpms]$
can be endowed with such a structure, as we will show in
Sections~\ref{subsubsect:biprod}--\ref{subsubsect:exp} below. We can
therefore interpret the quantum lambda calculus in $\freecat[\ccpms]$.
The denotation $\denot{A}$ of a type $A$ is an object of
$\freecat[\ccpms]$. In case $A$ is the ground type (i.e., $\tunit$,
$\qubit$), its denotation is:
\begin{align*}
\web{\denot{\qubit}}&\ass\{\star\},
&\arity{\denot{\qubit}}_\star&\ass 2,
&\symm{\denot{\qubit}}_\star\ass\{\id\},
\\
\web{\denot{\tunit}}&\ass\{\star\},
&\arity{\denot{\tunit}}_\star&\ass 1,
&\symm{\denot{\tunit}}_\star\ass\{\id\}.
\end{align*}
The denotation of the other types is given by structural induction,
following the compact closed Lafont structure of
$\freecat[\ccpms]$. We note in particular that the permutation groups
play a role only when interpreting $\oc$-formulas.
Let $\Gamma=x_1{:}A_1,\dots,x_n{:}A_n$. The denotation of a typing
judgement $\Gamma\entail M:A$ is a morphism $\denot{M}^{\Gamma\entail
A}:\denot{A_1\otimes\cdots\otimes A_n}\to\denot{A}$.
The definition is by structural induction on the unique type
derivation $\pi$ of $\Gamma\entail M:A$
(see Proposition~\ref{prop:unicity_derivation}).
\begin{table}
\[
\denot{\meas}^{\oc\Delta\vdash\qubit\multimap\bit}_{\vec m,(\ast,b)}=
(\begin{smallmatrix}
\alpha & \beta\\
\gamma & \delta
\end{smallmatrix})\mapsto
\begin{cases}
\alpha&\text{if $\vec m=\vec{[\,]}$ and $b=\ffalse$,}\\
\delta&\text{if $\vec m=\vec{[\,]}$ and $b=\ttrue$,}\\
0&\text{otherwise.}
\end{cases}
\]
\[
\denot{\new}^{\oc\Delta\vdash\bit\multimap\qubit}_{\vec m,(b,\ast)}=
\alpha\mapsto
\begin{cases}
(\begin{smallmatrix}
\alpha & 0\\
0 & 0
\end{smallmatrix})&\text{if $\vec m=\vec{[\,]}$ and $b=\ffalse$,}\\
(\begin{smallmatrix}
0 & 0\\
0 & \alpha
\end{smallmatrix})&\text{if $\vec m=\vec{[\,]}$ and $b=\ttrue$,}\\
\mathbf 0&\text{otherwise.}
\end{cases}
\]
\[
\denot{U}^{\oc\Delta\vdash\qubit^{\otimes n}
\multimap\qubit^{\otimes n}}_{\vec m,(\vec\ast,\vec\ast)}=
M\mapsto
\begin{cases}
U M U^{-1}&\text{if $\vec m=\vec{[\,]}$,}\\
\mathbf 0&\text{otherwise.}
\end{cases}
\]
\caption{\footnotesize Interpretation of the quantum constants.
The writing $\vec m$ stands for a sequence of multisets in
$\web{\denot{\oc\Delta}}$, the equality $\vec m=\vec{[\,]}$
meaning that each of these multisets is empty. $U$ and $M$ have
the same dimension $\C^{2^n\times 2^n}$, $U$ being unitary.}
\label{table:denotation_quantum}
\end{table}
\begin{table*}
\centering
\def0.84{0.84}
\centering
\scalebox{0.84}{
\centering
\subfloat[$\oc\Delta,x:A\vdash x:A$]{
\centering
\parbox{100pt}{
\centering
\[
\xymatrix@R=0pt@C=1pc{
\oc\Delta\otimes A\ar[r]^-{\weak\otimes\id}&
\tunit\otimes A\simeq A
}
\]
\vspace{-10pt}
}
}
\subfloat[$\oc\Delta,x:\oc A\vdash x:A$]{
\centering
\parbox{100pt}{
\centering
\[
\xymatrix@R=0pt@C=1pc{
\oc\Delta\otimes \oc A\ar[r]^-{\weak\otimes \der}&
\tunit\otimes A\simeq A
}
\]
\vspace{-10pt}
}
}
\subfloat[$\oc\Delta,\vdash V:\oc A$]{
\centering
\parbox{120pt}{
\centering
\[
\xymatrix@R=0pt@C=1pc{
\oc\Delta\ar[r]^{\dig}&\oc\oc\Delta\ar[r]^-{\bierman}&
\oc(\oc\Delta)\ar[r]^-{\oc\phi}&\oc A
}
\]
\vspace{-10pt}
}
}
\subfloat[$\oc\Delta,\Gamma\vdash \lambda x^A.M:A\multimap B$]{
\centering
\parbox{150pt}{
\centering
\[
\xymatrix@R=0pt@C=2pc{
\oc\Delta\otimes\Gamma\ar[r]^-{\cmatrix{\phi}}&A\multimap B
}
\]
\vspace{-10pt}
}
}
}
\vspace{-2ex}
\scalebox{0.84}{
\centering
\subfloat[$\oc\Delta,\Gamma,\Sigma\vdash MN:B$]{
\centering
\parbox{240pt}{
\centering
\[
\xymatrix@R=0pt@C=1pc{
\oc\Delta\otimes\Gamma\otimes\Sigma\ar[r]^-{\contr\otimes\id}&
\oc\Delta\otimes\Gamma\otimes\oc\Delta\otimes
\Sigma\ar[r]^-{\phi\otimes\psi}&
A\otimes A\multimap B\ar[r]^-{\eval}&
B
}
\]
\vspace{-10pt}
}
}
\subfloat[$\oc\Delta\vdash \punit:\tunit$]{
\centering
\parbox{60pt}{
\centering
\[
\xymatrix@R=0pt@C=4ex{
\oc\Delta\ar[r]^{\weak}&\tunit
}
\]
\vspace{-10pt}
}
}
\subfloat[$\oc\Delta,\Gamma,\Sigma\vdash \letunitterm{M}{N}:A$]{
\centering
\parbox{265pt}{
\[
\xymatrix@R=0pt@C=1pc{
\oc\Delta\otimes\Gamma\otimes\Sigma\ar[r]^-{\contr\otimes\id}&
\oc\Delta\otimes\Gamma\otimes\oc\Delta\otimes
\Sigma\ar[r]^-{\phi\otimes\id}&
\tunit\otimes\oc\Delta\otimes\Sigma\simeq
\oc\Delta\otimes\Sigma\ar[r]^-{\psi}&
A
}
\]
\vspace{-10pt}
}
}
}
\vspace{-2ex}
\scalebox{0.84}{
\subfloat[$\oc\Delta,\Gamma,\Sigma\vdash \tensterm MN:A\otimes B$]{
\centering
\parbox{200pt}{
\centering
\[
\xymatrix@R=0pt@C=.5pc{
\oc\Delta\otimes\Gamma\otimes\Sigma\ar[rr]^-{\contr\otimes\id}&&
\oc\Delta\otimes\Gamma\otimes\oc\Delta\otimes
\Sigma\ar[rr]^-{\phi\otimes\psi}&&
A\otimes B
}
\]
\vspace{-10pt}
}
}
\subfloat[$\oc\Delta,\Gamma,\Sigma\vdash \lettensterm{x^A}{y^B}{M}{N}:C$]{
\centering
\parbox{251pt}{
\centering
\[
\xymatrix@R=0pt@C=.5pc{
\oc\Delta\otimes\Gamma\otimes\Sigma
\ar[rr]^-{\contr\otimes\id}&&
\oc\Delta\otimes\Gamma\otimes\oc\Delta
\otimes\Sigma\ar[rr]^-{\phi\otimes\id}&&
A\otimes B\otimes\oc\Delta\otimes\Sigma\ar[r]^-{\psi}&
C
}
\]
\vspace{-10pt}
}
}
\subfloat[$\oc\Delta,\Gamma\vdash \injl M:A\oplus B$]{
\centering
\parbox{133pt}{
\centering
\[
\xymatrix@R=0pt@C=1pc{
\oc\Delta\otimes\Gamma\ar[r]^-{\phi}&
A\ar[r]^-{\inj \ell{}}&
A\oplus B
}
\]
\vspace{-10pt}
}
}
}
\vspace{-2ex}
\scalebox{0.84}{
\subfloat[$\oc\Delta,\Gamma\vdash \injr M:A\oplus B$]{
\centering
\parbox{140pt}{
\centering
\[
\xymatrix@R=0pt@C=1pc{
\oc\Delta\otimes\Gamma\ar[r]^-{\phi}&
B\ar[r]^-{\inj r{}}&
A\oplus B
}
\]
\vspace{-10pt}
}
}
\subfloat[$\oc\Delta,\Gamma,\Sigma\vdash \match{M}{x^A}{N}{y^B}{L}:C$
]{
\centering
\parbox{449pt}{
\centering
\[
\xymatrix@R=0pt@C=.5pc{
\oc\Delta\otimes\Gamma\otimes\Sigma\ar[rr]^-{\contr\otimes\id}&&
\oc\Delta\otimes\Gamma\otimes\oc\Delta
\otimes\Sigma\ar[rr]^-{\psi\otimes\id}&&
(A\oplus B)\otimes\oc\Delta\otimes\Sigma\ar[rr]^-{\pdistr}&&
(A\otimes\oc\Delta\otimes\Sigma)\oplus(B\otimes\oc\Delta\otimes\Sigma)
\ar[rr]^-{\phi_A\oplus\phi_B}&&
C
}
\]
\vspace{-10pt}
}
}
}
\vspace{-2ex}
\scalebox{0.84}{
\subfloat[$\oc\Delta,\Gamma\vdash M:\tlist A$]{
\centering
\parbox{240pt}{
\centering
\[\xymatrixcolsep{1pc}
\xymatrix{
\oc\Delta\otimes\Gamma\ar[r]^-{\phi}&
\tunit\oplus(A\otimes\tlist A)\ar[rr]^-{\id\oplus\pdistr}&&
\tunit\oplus(\bigoplus_{n=1}^\infty A^{\otimes n})=\tlist A
}
\]
\vspace{-15pt}
}
}
\subfloat[$\bang\Delta,\Gamma\entail\letrec{f}{x}{M}{N}:C$]{
\centering
\parbox{350pt}{
\centering
\[
\oc\Delta\tensor\Gamma
\xrightarrow{\contr}
\oc\Delta\tensor\Gamma\tensor\oc\Delta
\xrightarrow{\id\tensor\fixpoint(\dig;\bierman;\oc{(\Lambda\phi)})}
\oc\Delta\tensor\Gamma\tensor\oc{(A\loli B)}
\xrightarrow{\psi}
C
\]
\vspace{-10pt}
}}
}
\caption{\footnotesize Sketch of the interpretation of the typing
judgements, using the Lafont structure of $\freecat[\ccpms]$
defined in Section~\ref{sect:lafont_ccpms}. The
morphisms $\phi$, $\psi$, $\phi_A$, $\phi_B$ refer to the denotation
of the premises of the unique derivation concluding a typing
judgement. In (c) and (n), the morphism $\bierman$ stands for
$\bierman^\unit$ or the suitable sequence of $\bierman^\otimes$,
depending on the context $\oc\oc\Delta$.}\label{table:denotation_rule}
\end{table*}
The denotations of the constants $\meas$, $\new$ and the unitary
transformations are given in
Table~\ref{table:denotation_quantum}. Table~\ref{table:denotation_rule}
briefly recalls the denotation of the usual linear logic rules. Here,
the morphisms $\phi, \psi, \phi_A, \phi_B$ refer to the denotation of
the premises of the last rule of $\pi$, which are uniquely defined
given $\Gamma\entail M:A$.
In the interpretation of the $\mathtt{letrec}$ constructor, the fixed
point operator $\fixpoint$ is defined as follows. Let $\phi$ be a
morphism in the set $\homof{\freecat[\ccpms]}{\oc{C}\tensor\oc{A},\oc A}$. By
induction on $n$, we define the morphism
$\phi^n\in\homof{\freecat[\ccpms]}{\oc{C},\oc A}$: $
\phi^0\ass\oc{C}\xrightarrow{\weak;\oc{\bf 0}}\bang{A}$,
$\phi^{n+1}\ass\oc{C}\xrightarrow{\contr}\bang{C}\tensor\bang{C}
\xrightarrow{\id\tensor
\phi^n} \bang{C}\tensor\bang{A}\xrightarrow{\phi}\bang{A}$. Since
$\phi$ can be regarded as a continuous module homomorphism (in
particular it is monotone), the set $\{\phi^n\}$ is directed
complete. We define $\fixpoint(\phi)$ as its least upper bound.
\subsubsection{Biproduct (\texorpdfstring{$\qfin A\oplus\qfin B$}{A + B})}
\label{subsubsect:biprod}
Let $I$ be a (possibly infinite) set of indexes. The biproduct
$\bigoplus_{i\in I}\qfin A_i$ of a family $\{\qfin A_i\}_{i\in I}$ of
objects in $\freecat[\ccpms]$ is defined by
\[
\web{\bigoplus_{i\in I}\qfin A_i}
\ass\!\bigcup_{i\in I}\{i\}\times\web{\qfin A_i},\;\;
\arity{\bigoplus_{i\in I}\qfin A_i}_{(j,a)}\!\!\ass\arity{\qfin A_j}_a,
\;\;
\symm{\bigoplus_{i\in I}\qfin A_i}_{(j,a)}\!\!\ass\symm{\qfin A_j}_a.
\]
The corresponding projections and injections are denoted respectively by
$\proj j$ and $\inj j$ and defined as:
\begin{align*}
\proj j_{(i,a),a'}&\ass\inj j_{a',(i,a)}\ass
\delta_{j,i}\delta_{a,a'}\symm{\qfin A_i}_a.
\end{align*}
The tupling $\fprod{\phi_i}{i\in I}$ (resp.\ (co)-tupling
$\fcoprod{\psi_i}{i\in I}$) of a family of morphisms
$\phi_i$ elements of $\homof{\freecat[\ccpms]}{\qfin A,\qfin B_i}$ (resp.\
$\psi_i$ elements of $\homof{\freecat[\ccpms]}{\qfin A_i,\qfin B}$)
is defined by
$(\fprod{\phi_i}{i\in
I})_{a,(j,b)}\ass(\phi_{j})_{a,b}$ (resp. $(\fcoprod{\psi_i}{i\in
I})_{(j,a),b}\ass(\psi_{j})_{a,b}$).
\begin{example}\label{ex:bool_sem}
Recall that in Notation~\ref{notation:syntax}, the type $\bit$ is
interpreted as the biproduct $\denot{\tunit}\oplus\denot{\tunit}$,
which is the two-element family
$\{(1,\{\id\})_{\ttrue},(1,\{\id\})_{\ffalse}\}$. The positive
cones associated with $\tunit$ and $\bit$ are:
$\pmatr(\denot{\tunit})=\overline{\Rp}$ and
$\pmatr(\denot{\bit})=\overline{\Rp}^2$.
The typing judgement $\vdash \ttrue:\bit$ is interpreted as the right
injection, which can be seen both as a family of two completely
positive maps from $\C$ to $\C$ (i.e., $\denot{\ttrue}^{\vdash
\bit}_{\star,\ttrue}=p\mapsto p$ and $\denot{\ttrue}^{\vdash
\bit}_{\star,\ffalse}=p\mapsto 0$) and as a quantum compatible and
completely positive map sending $p\in\overline{\Rp}$ to
$(0,p)\in\overline{\Rp}^2$. Symmetrically,
$\denot{\ffalse}^{\vdash\bit}$ is the map $p\mapsto (p,0)$.
As an example of a term with free variables, consider
$\texttt{Neg}_x\ass\iftermx{x}{\ffalse}{\ttrue}$. The denotation of
$x:\bit\vdash \texttt{Neg}_x:\bit$ can be seen both as a family of
four constant maps $\denot{\texttt{Neg}_x}^{\bit\vdash\bit}_{b,b'}$
from $\C$ to $\C$ of value $1$ if $b\neq b'$ and $0$ otherwise, and as
a single map from $\overline\Rp^2$ to $\overline\Rp^2$ sending
$(p,p')$ to $(p',p)$.
\end{example}
\subsubsection{Symmetric monoidal structure (\texorpdfstring{$\qfin
A\otimes \qfin B$}{A tensor B},
\texorpdfstring{$\unit$}{1} and
\texorpdfstring{$\tlist{\qfin A}$}{list(A)})}
\label{subsubsect:smc}
The bifunctor
$\tensor:\freecat[\ccpms]\times\freecat[\ccpms]\to\freecat[\ccpms]$ is
defined on objects $\qfin A, \qfin B$ by:
\[
\web{\qfin A\tensor\qfin B}\ass\web{\qfin A}\times\web{\qfin B},
\qquad
\arity{\qfin A\tensor\qfin B}_{(a,b)}\ass\arity{\qfin A}_a
\times\arity{\qfin B}_b,\]
\[
\symm{\qfin A\tensor\qfin B}_{(a,b)}\ass\{(g,h)\;;\;g\in
\symm{\qfin A}_a, h\in\symm{\qfin B}_b\},
\]
\noindent where $\arity{\qfin A}_a\times\arity{\qfin B}_b$ is the
multiplication of the two numbers $\arity{\qfin A}_a$ and
$\arity{\qfin B}_b$, which can be seen as the lexicographically
ordered set of pairs $(i,j)$, for $i<\arity{\qfin A}_a$,
$j<\arity{\qfin B}_b$. Hence, the action of a permutation
$(g,h)\in\symm{\qfin A\tensor\qfin B}_{(a,b)}$ on $\arity{\qfin
A}_a\times\arity{\qfin B}_b$ can be described as
$(i,j)\mapsto(g(i),h(j))$.
The bifunctor $\tensor$ on morphisms is defined componentwise, using
the standard tensor of the category $\CPM$ extended to the infinite
elements by the universal property of the D-completion
(Section~\ref{sect:preli_cpm}).
The tensor unit is the object $\denot{\tunit}$ interpreting the unit type.
The associativity, unit, and symmetry isomorphisms are defined
componentwise from the corresponding isomorphisms in $\CPM$, composed
with the actions of the groups of the objects. E.g., the symmetry is $
\sigma^{\qfin A,\qfin
B}_{(a,b),(b',a')}\!\ass\delta_{a,a'}\delta_{b,b'}\symm{\qfin
A\otimes\qfin B}_{(a,b)}\pc\swap^{\arity{\qfin A}_a,\arity{\qfin
B}_b}
$,
where $\swap^{\arity{\qfin A}_a,\arity{\qfin B}_b}$ is the symmetry in
\CPM{} between $\C^{\arity{\qfin A}_a\times\arity{\qfin
A}_a}\otimes\C^{\arity{\qfin B}_b\times\arity{\qfin B}_b}$ and
$\C^{\arity{\qfin B}_b\times\arity{\qfin B}_b}\otimes\C^{\arity{\qfin
A}_a\times\arity{\qfin A}_a}$. Notice that it is sufficient to
pre-compose $\swap^{\arity{\qfin A}_a,\arity{\qfin B}_b}$ with
$\symm{\qfin A\otimes\qfin B}_{(a,b)}$ (or, symmetrically,
post-compose with $\symm{\qfin B\otimes\qfin A}_{(b,a)}$), in order to
have a map invariant under both the permutation groups $\symm{\qfin
A\otimes\qfin B}_{(a,b)}$ and $\symm{\qfin B\otimes\qfin
A}_{(b,a)}$. This is because $\symm{\qfin A\otimes\qfin
B}_{(a,b)}\pc\swap^{\arity{\qfin A}_a,\arity{\qfin B}_b}=\symm{\qfin
A}_{a}\otimes\symm{\qfin B}_{b}\pc\swap^{\arity{\qfin
A}_a,\arity{\qfin B}_b}=\swap^{\arity{\qfin A}_a,\arity{\qfin
B}_b}\pc \symm{\qfin B}_{b}\otimes\symm{\qfin
A}_{a}=\swap^{\arity{\qfin A}_a,\arity{\qfin B}_b}\pc\symm{\qfin
B\otimes\qfin A}_{(b,a)}$. Similar simplifications will be done
henceforth without explicitly mentioning it.
\begin{example}\label{ex:ifthen_sem}
The denotation of $\qubit\otimes\qubit$ is the singleton web family
$\{(4,\{\id\})_\star\}$. This object is associated with the cone of
positive matrices of dimension $4\times 4$ plus the infinite
elements needed to complete the L\"owner order. The
denotation of $\bit\otimes\bit$ instead has a web of cardinality
$4$, i.e.,
$\{(\ffalse,\ffalse),(\ffalse,\ttrue),(\ttrue,\ffalse),(\ffalse,\ffalse)\}$,
and, for each index $b\in\web{\denot{\bit\otimes\bit}}$, we have
$\arity{\denot{\bit\otimes\bit}}_b=1$ and
$\symm{\denot{\bit\otimes\bit}}_b=\{\id\}$. This object is
associated with the biproduct
$\overline\Rp\oplus\overline\Rp\oplus\overline\Rp\oplus\overline\Rp$.
\end{example}
Notice that in the above example the tensor product distributes over
the biproducts:
$\denot{\bit\otimes\bit}=\denot{(\tunit\oplus\tunit)
\otimes(\tunit\oplus\tunit)}=\denot{\tunit\oplus\tunit\oplus
\tunit\oplus\tunit}$. This
is true in general: the isomorphism between $\qfin
A\otimes(\bigoplus_{i\in I}\qfin B_i)$ and $\bigoplus_{i\in I}(\qfin
A\otimes\qfin B_i)$ is
\[
\pdistr_{(a,(i,b)),(i',(a',b'))}\ass\delta_{i,i'}\delta_{a,a'}
\delta_{b,b'}\symm{\qfin A\otimes\qfin B_i}_{(a,b)}.
\]
This isomorphism allows us to define the list constructor as the
infinite biproduct of tensor powers
$
\tlist{\qfin A}\ass\bigoplus_{n=0}^\infty\qfin A^{\otimes n}.
$
In fact, we have $\tlist{\qfin A}\simeq\unit\oplus(\qfin A\otimes
\tlist{\qfin A})$.
\begin{example}\label{ex:list}
The denotation of the unit type list is:
$\web{\denot{\tlist\unit}}=\N$ and, for every $n\in\N$,
$\arity{\denot{\tlist\unit}}_n=1$,
$\symm{\denot{\tlist\unit}}_n=\{\id\}$. This object can be
associated with the module $\overline{\Rp}^\N$ and is suitable for
denoting the numerals in unary notation. Indeed, writing $\underline
n$ for the list $\cons{\punit}{\dots\cons{\punit}{\nil}}$ of length
$n$, we have $\denot{\underline
n}^{\vdash\tlist\unit}=p\mapsto(\underbrace{0,\dots,0}_{n-1\text{
times}},p,0,\dots)$.
\end{example}
\subsubsection{Compact closure (\texorpdfstring{$\qfin A^\perp, \qfin
A\multimap B$}{dual(A) and A -o B})}
\label{subsubsect:cc}
Dual objects coincide: we have
$
\dual{\qfin A}\ass\qfin A.
$
The unit $\eta^{\qfin A}\in\freecat[\ccpms](\unit,\dual{\qfin
A}\otimes \qfin A)$ and co-unit $\epsilon^{\qfin
A}\in\freecat[\ccpms](\qfin A\otimes\dual{\qfin A},\unit)$ are
defined componentwise composing the unit and co-unit of $\CPM$ with
the correspondent permutation group. Writing $E_{i,j}$ for the matrix
that has $0$ everywhere except $1$ at $(i,j)$, we have:
\begin{align*}
\eta^{\qfin A}_{\star,(a,a')}&
\ass 1\mapsto \sum_{i,j<\arity{\qfin A}}\symm{\qfin
A}_a(E_{i,j})\otimes
\symm{\qfin A}_a(E_{i,j})\\
\epsilon^{\qfin A}_{(a,a'),\star}
&\ass(E_{i,j}\otimes E_{i',j'})
\mapsto\!\!\!\!\sum_{g,g'\in \symm{\qfin A}_a}\frac{1}{
\#\symm{\qfin A}_a}\delta_{g(i),g'(i')}\delta_{g(j),g'(j')}.
\end{align*}
Compact closed categories are monoidal closed. Let us recall the
monoidal closure structure, which is needed to model the abstraction
and the application of the quantum lambda calculus.
The internal hom object is defined as
$
\qfin A\multimap\qfin B\ass(\dual{\qfin A}\otimes \qfin B)=
\qfin A\otimes \qfin B.
$
The evaluation morphism $\Eval^{\qfin A,\qfin
B}:\homof{\freecat[\ccpms]}{(\qfin A\multimap\qfin B)\otimes\qfin A,
\qfin B}$ and the currying isomorphism $\Lambda(\--)$ from
$\homof{\freecat[\ccpms]}{\qfin C\otimes\qfin A,\qfin B}$ to
$\homof{\freecat[\ccpms]}{\qfin C,\qfin A\multimap\qfin B}$ are,
\[
\Eval^{\qfin A,\qfin
B}\!\!\!\ass\sigma\pc\alpha\pc(\epsilon\tensor\id)\pc
\lambda,~~\cmatrix{\phi}\ass\lambda^{\!\!-1}\!\pc(\eta\otimes\id)\pc
\alpha^{\!\!-1}\!\pc(\id\otimes(\sigma;\phi)),
\]
\noindent where $\alpha$, $\lambda$, and $\sigma$ are the associative,
left unit and symmetric isomorphisms associated with $\otimes$.
\begin{example}\label{ex:lambda_sem}
Let us consider the abstraction $\lambda x.\texttt{Neg}_x$ of the
term $\texttt{Neg}_x$ discussed in Example~\ref{ex:bool_sem}. The
denotation $\denot{\lambda x.\texttt{Neg}_x}^{\vdash
\bit\multimap\bit}$ is obtained from
$\denot{\texttt{Neg}_x}^{x:\bit\vdash\bit}$ just by shifting the
matrix indexes: $\denot{\lambda x.\texttt{Neg}_x}^{\vdash
\bit\multimap\bit}_{\star,(b,b')}=
\denot{\texttt{Neg}_x}^{x:\bit\vdash\bit}_{b,b'}$.
Looking at this matrix as a module homomorphism, the map
$\denot{\lambda x.\texttt{Neg}_x}^{\vdash \bit\multimap\bit}$ is
$p\mapsto (0,p,p,0)$, which is a map from $\overline\Rp$ to
$\overline\Rp_{(\ffalse,\ffalse)}\oplus
\overline\Rp_{(\ffalse,\ttrue)}\oplus
\overline\Rp_{(\ttrue,\ffalse)}\oplus\overline\Rp_{(\ttrue,\ttrue)}$,
where we make explicit the correspondence between the web elements
of $\denot{\bit\multimap\bit}$ and the components of the biproduct
associated with.
Application corresponds basically to matrix multiplication. For example,
$\denot{(\lambda x.\texttt{Neg}_x)(\meas\,
y)}^{y:\qubit\vdash\bit}_{\star,b}$ is the function defined as
$
\sum_{b'\in\{\ttrue,\ffalse\}}\denot{\lambda
x.\texttt{Neg}_x}_{\star,(b',b)}^{\vdash\bit\multimap\bit}\denot{\meas\,
y}_{\star,b'}^{y:\qubit\vdash\bit}$,
which is sending $
(\begin{smallmatrix}
\alpha & \beta\\
\gamma & \delta
\end{smallmatrix})$
to $\delta$ if $b=\ffalse$, $\alpha$ if $b=\ttrue$, and $0$ otherwise.
\end{example}
\subsubsection{Free commutative comonoids (\texorpdfstring{$\qfin
A^{\odot k}$}{symmetric powers}, \texorpdfstring{$\oc \qfin A$}{!A})}
\label{subsubsect:exp}
Let us now focus on the crucial structure modeling the linear logic
modality $\oc$. We first define the notion of $k$-th symmetric power
of an object and then we show how the biproduct of all such symmetric
powers yields an exponential structure.
\begin{notation}\label{notation:multisets}
Given a set $X$, a \emph{multiset} $\mu$ over $X$ is a function
$X\mapsto\N$. The \emph{support} of $\mu$ is the set
$\supp\mu=\{a\mid\mu(a)\neq 0\}\subseteq X$, the \emph{disjoint
union} is $(\mu\uplus\nu)(a)=\mu(a)+\nu(a)$, and the \emph{empty
multiset} is the zero constant function. The \emph{cardinality} of
$\mu$ is $\sum_{a\in X}\mu(a)\in\N\cup\{\infty\}$. A multiset is
finite if it has finite cardinality. $\M_k(X)$ (resp.\ $\M_f(X)$) is
the set of the multisets over $X$ with cardinality $k$ (resp.\
finite). Finite multisets can be denoted by listing the occurrences
of their elements between square brackets, i.e., $\mu=[a,a,b]$ is
$\mu(a)=2$, $\mu(b)=1$ and zero on the other elements, and $[\,]$ is
the empty multiset.
\end{notation}
In a symmetric monoidal category, given a natural number $k$, the
\emph{$k$-th symmetric power} of an object $\qfin A$ is a pair $(\qfin
A^{\odot k},\eq{\qfin A^{\odot k}})$ of an object $\qfin A^{\odot k}$
and a morphism $\eq{\qfin A^{\odot k}}$ from $\qfin A^{\odot k}$ to
$\qfin A^{\otimes k}$, which is an equalizer of the $k!$ symmetries of
the $k$-ary tensor $\qfin A^{\otimes k}$. Such equalizers do not exist
in general, but they do exist in $\freecat[\ccpms]$ and can be
concretely represented using the multisets notation, as follows:
\begin{gather*}
\web{\qfin A^{\odot k}}\ass\M_k(\web{\qfin A}),\quad \arity{\qfin
A^{\odot k}}_\mu\ass\prod_{a\in\supp\mu}(\arity{\qfin
A}_a)^{\mu(a)},
\\
\symm{\qfin A^{\odot
k}}_\mu\ass\{(h_a,g_a^1,\dots,g_a^{\mu(a)})_{a\in\supp\mu}\;;\;
h_a\in\symgroup_{\mu(a)}, g_a^i\in\symm{\qfin A}_a\},
\end{gather*}
where $(h_a,g_a^1,\dots,g_a^{\mu(a)})_{a\in\supp\mu}$ is a
$\supp\mu$-indexed family of sequences of permutations and
$\symm{\qfin A^{\odot k}}_\mu$ is a group (composition being defined
componentwise) whose action on $\C^{\arity{\qfin A^{\odot
k}}_\mu\times\arity{\qfin A^{\odot k}}_\mu}$ can be described by
seeing $\arity{\qfin A^{\odot k}}_\mu$ as the set of families of
sequences of the form $(i_a^1,\dots,i_a^{\mu(a)})_{a\in\supp\mu}$,
with $i_a^j<\arity{\qfin A}_a$ for every $j\leq\mu(a)$. Then, the
action of $(h_a,g_a^1,\dots,g_a^{\mu(a)})_{a\in\supp\mu}$ on such
families is:
\[
(i_a^1,\dots,i_a^{\mu(a)})_{a\in\supp\mu}\mapsto
(g_a^1(i_a^{h_a(1)}),\dots,g_a^{\mu(a)}(i_a^{h_a(\mu(a))}))_{a\in\supp\mu}.
\]
The morphism $\eq{\qfin A^{\odot k}}$ is given by
\[
\eq{\qfin A^{\odot k}}_{\mu,(a_1,\dots, a_k)}\ass
\begin{cases}
\symm{\qfin A^{\odot k}}_\mu&\text{if $\mu=[a_1,\dots,a_k]$,}\\
\mathbf 0&\text{otherwise.}
\end{cases}
\]
\begin{remark}
The object $\denot{A}^{\odot k}$ describes $k$ {\em unordered} uses
of an element of type $A$. The fact that our model uses the
symmetric tensor power $\qfin A^{\odot k}$ instead of the $k$-fold tensor
$\qfin A^{\otimes k}$ means operationally that the behavior of a program
calling its input $k$ times does not depend on the order of the calls.
\end{remark}
\begin{example}\label{ex:symmetric_two_power}
In Example~\ref{ex:ifthen_sem}, we have seen that
$\denot{\qubit}^{\otimes 2}=\{(4,\{\id\})_\star\}$. The symmetric
$2$-power $\denot{\qubit}^{\odot 2}$ is instead the singleton web
family $\{(4,\{\id,\swap\})_\star\}$, where $4$ is represented as
the lexicographically ordered set $\{(0,0),(0,1),(1,0),(1,1)\}$ and
the permutation $\swap$ acts on it by $(b,b')\mapsto(b',b)$. The
group of permutations $\{\id,\swap\}$ shrinks the set of possible
morphisms to or from $\denot{\qubit}^{\odot 2}$. For example, the
matrix $N_c$ associated with the controlled-not gate
(Equation~\eqref{eq:h_and_nc}) defines a complete positive endo-map
of $\C^{4\times 4}$, which is an endo-morphism of
$\denot{\qubit}^{\otimes 2}$ but not of $\denot{\qubit}^{\odot 2}$,
because $N_c$ is not invariant under the action of $\{\id,\swap\}$:
\[
\{\id,\swap\}(N_c)=\frac12(\id(N_c)+\swap(N_c))=
\frac12
\left(\begin{smallmatrix}
2&0&0&0\\
0&1&0&1\\
0&0&1&1\\
0&1&1&0
\end{smallmatrix}
\right)\neq N_c.
\]
Concerning the module associated with symmetric tensor powers,
$\pmatr(\denot{\qubit}^{\odot 2})$ is the D-completion of
\[\left\{
\left(
\begin{smallmatrix}
\alpha_1&\alpha_2&\alpha_2&\alpha_3\\
\alpha_4&\alpha_5&\alpha_6&\alpha_7\\
\alpha_4&\alpha_6&\alpha_5&\alpha_7\\
\alpha_8&\alpha_9&\alpha_9&\alpha_{10}
\end{smallmatrix}
\right)\text{ positive}\;;\;\forall i, \alpha_i\in\C
\right\}
\]
which is a subcone of the positive cone of $\C^{4\times 4}$ of
dimension $10$.
Concerning biproducts, the denotation of $\qubit\oplus\qubit$ is given
by $\{(2,\{\id\})_{\ttrue}, (2,\{\id\})_{\ffalse}\}$, while its
symmetric tensor power $\denot{\qubit\oplus\qubit}^{\odot 2}$ is given
by the three-element family
$\{(4,\{\id,\sigma\})_{[\ttrue,\ttrue]},(4,\{\id\})_{[\ttrue,
\ffalse]},(4,\{\id,\sigma\})_{[\ffalse,\ffalse]}\}$. Notice the
difference between the pair $(4,\{\id\})$ associated with $[\ttrue,
\ffalse]$ and the pair $(4,\{\id,\sigma\})$ associated with the two
multisets of singleton support.
\end{example}
The biproduct $\oc\qfin A\ass\bigoplus_{k=0}^\infty\qfin A^{\odot k}$
of all symmetric powers of $\qfin A$ can be defined as
\begin{align*}
\web{\oc\qfin A}&=\M_f(\web{\qfin A}),&
\arity{\oc\qfin A}_\mu&=\arity{\qfin A^{\odot k}}_\mu\!\!,&
\symm{\oc\qfin A}_\mu&=\symm{\qfin A^{\odot k}}_\mu&
\!\!\!\!\text{($\mu\in\M_k(\web{\qfin A})$)}
\end{align*}
This object yields a concrete representation of the free commutative
comonoid generated by $\qfin A$. The counit (also called
\emph{weakening}) $\weak\in\homof{\freecat[\ccpms]}{\oc\qfin A,\unit}$
and the comultiplication (or \emph{contraction})
$\contr\in\homof{\freecat[\ccpms]}{\oc\qfin A,\oc\qfin
A\otimes\oc\qfin A}$ are:
\begin{align*}
\weak_{\mu,\star}&\ass\delta_{\mu,[\,]}\symm{\oc\qfin A}_{[]},&
\contr_{\mu,(\mu',\mu'')}&\ass\delta_{\mu,\mu'+\mu''}\symm{\oc\qfin A}_\mu.
\end{align*}
The freeness of the comonoid gives the structure of exponential
comonad. The functorial promotion maps an object $\qfin A$ to
$\oc\qfin A$ and a morphism $\phi\in\homof{\freecat[\ccpms]}{\qfin
A,\qfin B}$ to $\oc\phi\in\homof{\freecat[\ccpms]}{\oc\qfin
A,\oc\qfin B}$ defined by, for $\mu\in\M_f(\web{\qfin A})$ and
$\nu=[b_1,\dots,b_k]\in\M_f(\web{\qfin B})$,
\[
\oc\phi_{\mu,\nu}\ass\sum_{
\substack{
(a_1,\dots,a_k),\text{ st}\\
[a_1,\dots,a_k]=\mu
}
}
\symm{\oc\qfin{A}}_{\mu}\pc\bigotimes_{i=1}^k\phi_{a_i,b_i}
\pc\symm{\oc\qfin
B}_\nu.
\]
The counit of the comonad (or \emph{dereliction})
$\der\in\homof{\freecat[\ccpms]}{\oc\qfin A,\qfin A}$ and the
comultiplication (or \emph{digging})
$\dig\in\homof{\freecat[\ccpms]}{\oc\qfin A,\oc\oc\qfin A}$ are
\begin{align*}
\der_{\mu,a}&\ass\delta_{\mu,[a]}\symm{\qfin A}_a,&
\dig_{\mu,M}&\ass\delta_{\mu,\sum M}\symm{\qfin \oc\qfin A}_\mu,
\end{align*}
where $M\in\web{\oc\oc\qfin A}$ is a multiset of multisets $\nu$ over
$\web{\qfin A}$ and $\sum M\in\web{\oc\qfin A}$ is the multiset union
of such $\nu$'s, i.e., for every $a\in\web{\qfin A}$, $\sum
M(a)=\sum_{\nu\in\supp M}\nu(a)^{M(\nu)}$.
Finally, the last two morphisms that are essential to interpret our
calculus are Bierman's
$\bierman^\otimes\in\homof{\freecat[\ccpms]}{\oc\qfin A\otimes\oc\qfin
B,\oc(\qfin A\otimes\qfin B)}$ and
$\bierman^\unit\in\homof{\freecat[\ccpms]}{\unit,\oc\unit}$, given by
$
\bierman^\otimes_{(\mu,\nu),\eta}\ass\delta_{\eta,\mu\times\nu}\symm{\oc(\qfin
A\otimes\qfin B)}_\eta$ and $
\bierman^\unit_{\star,\mu}\ass\delta_{\mu,[\star]}\symm{\unit}_\mu$,
where $\mu\times\nu$ is the multiset in $\web{\oc{(\qfin A\otimes\qfin
B)}}$ defined by, $\mu\times\nu(a,b)\ass\mu(a)\nu(b)$.
\begin{example}\label{ex:bang_sem}
Using the isomorphism between $\M_f(\{\star\})$ and the set $\N$, and
between $\M_f(\{\ttrue, \ffalse\})$ and $\N\times\N$, the free
commutative comonoids associated with $\denot{\tunit}$ and
$\denot{\bit}$ are $\oc\denot{\tunit}=\{(1,\{\id\})_{n}\}_{n\in\N}$,
and $\oc\denot{\bit}=\{(1,\{\id\})_{(n,m)}\}_{n,m\in\N}$.
In general, notice that all constructions of the Lafont category
preserve the underlying pair $(1,\{\id\})$ and act only at the level
of webs. For more involved examples, one should look for objects with
larger dimension, like $\denot{\qubit}$. For example,
$
\oc\denot{\qubit}=\{(2^n,\symgroup_n)_n\}_{n\in\N}.
$
Notice that $\oc\tunit$, $\oc\bit$ and $\oc\qubit$ are not allowed
by our type grammar. In fact, $\oc\qubit$ is meaningless because of
the no-cloning constraint on quantum bits. However, such spaces
should exist in the model since they are isomorphic to the
denotations of legal types, like $\oc(\tunit\multimap\tunit)$,
$\oc(\tunit\multimap\bit)$ and $\oc(\tunit\multimap\qubit)$.
\end{example}
\subsection{The soundness theorem}\label{Sect:soundness}
The soundness of $\freecat[\ccpms]$ with respect to the operational
semantics given in Figure~\ref{table:reduction} is an easy consequence
of the fact that the category gives a (dcpo-enriched) model of linear
logic. In fact, the operational semantics is a trivial extension of a
head-reduction strategy of linear logic cut-elimination.
\begin{proposition}\label{prop:lafont}
The category $\freecat[\ccpms]$ is a dcpo-enriched compact closed
Lafont category, hence $\freecat[\ccpms]$ is a model of linear
logic.
\end{proposition}
\begin{proof}[Proof (Sketch)]
This basically amounts to showing that $\freecat[\ccpms]$ is the
result of a categorical construction applied to $\CPMs$ which is
known to give, under certain circumstances, a dcpo-enriched Lafont
category and to preserve the compact closed structure of
$\CPMs$. This construction was sketched in
\cite{Girard99coherentbanach} and detailed
in~\cite{MelliesTT09,LairdMM12,LairdMMP13}. It consists in moving:
(i) from $\CPMs$ to a category $\cpms$ with symmetric tensors, which
is actually a full sub-category of the Karoubi envelope of $\CPMs$;
(ii) to a dcpo-enriched category $\ccpms$ using the D-completion
defined in {\cite{ZhaoFan2010,Keimel2009}}; and, finally, (iii)
constructing the free biproduct completion $\freecat[\ccpms]$ of
$\ccpms$ and applying Equation~\eqref{eq:bang_biproduct}.
\end{proof}
Given a linking $\ell=\ket{y_1,\dots,y_m}$, we write $\ell\vdash M:A$
for the judgement $y_1:\qubit,\dots,y_m:\qubit\vdash M:A$.
\begin{proposition}[Invariance of the interpretation]
\label{prop:semantic_invariance}
Let $\ell$ be the linking $\ket{y_1,\dots,y_m}$, and assume
$\ell\vdash M:A$. If $M$ is not a value, then for all quantum states
$\qarray\in{\C^{2^m}}$,
\begin{equation}\label{eq:soundness}
\denot{M}^{\ell\vdash A}(\qarray\qarray^\ast)=\sum_{
\begin{subarray}{c}
\am{\qarray,\ell,M}\redto[p]\am{\qarray',\ell',N}
\end{subarray}
}p\cdot\denot{N}^{\ell'\vdash A}(\qarray'{\qarray'}^\ast).
\end{equation}
\end{proposition}
\begin{proof}
By hypothesis, $\am{\qarray,\qlist,M}$ is a typable total closure,
and so, by Proposition~\ref{prop:subject_reduction} and
Lemma~\ref{lemma:totality}, all of its reducts
$\am{\qarray',\qlist',N}$ are typable total closures, so that
$\denot{N}^{\ell'\vdash A}(\qarray'{\qarray'}^\ast)$ is
well-defined.
Equation~\ref{eq:soundness} is proven by cases, depending on the
rule applied to $\am{\qarray,\ell,M}$. The cases of
Table~\ref{table:reduction}\subref{subtable:reduction_classical}
follows from the fact that $\freecat[\ccpms]$ is a dcpo-enriched
model of linear logic. The quantum rules
(Table~\ref{table:reduction}\subref{subtable:reduction_quantum}) are
trivial consequences of Table~\ref{table:denotation_quantum}, and
the congruence rules of
Table~\ref{table:reduction}\subref{subtable:reduction_congruence}
are done by induction on $M$, using the fact that the category
$\freecat[\ccpms]$ is linear.
\end{proof}
\begin{corollary}\label{cor:invariance_interpretation}
We have
$\denot{M}^{\vdash\tunit}_{\ast}\geq\Halt_{\am{\ket{\,},\ket{\,},M}}$.
\end{corollary}
\begin{proof}
By induction on $n$ and using Proposition~\ref{eq:soundness} we can show that
$
\denot{M}^{\ell\vdash\tunit}_\ast(\qarray\qarray^\ast)
$
is greater or equal to
$
\sum_{\am{\qarray'\!,\ell'\!,V}}\Red^{n}_{\am{\ell,\qarray,M},
\am{\qarray'\!,\ell'\!,V}}$.
Then $\denot{M}^{\ell\vdash\tunit}_{\ast}(\qarray\qarray^\ast)
\geq\Halt_{\am{\qarray,\ell,M}}$ follows by taking the limit as
$n\to\infty$,and
invoking the monotonicity of $\{\Red^{n}\}_n$.
\end{proof}
\subsection{The denotations of {\bf qlist} and {\bf teleport}}
\label{sect:examples}
\begin{example}\label{ex:term_sem}
Recall the terms of Example~\ref{ex:term_type}. The web of
$\denot{\tlist\qubit}$ is $\N$, while
$\denot{\tlist\qubit}_n{=}(2^n,\{\id\})$. Note that
$\pmatr(\denot{\tlist\qubit})$ is equivalent to
the D-completion of $\bigoplus_nP(\C^{2^n\times 2^n})$ where
the set
$P(\C^{2^n\times 2^n})$ is
the cone of $2^n\times2^n$ positive matrices. The denotation of the
term {\bf qlist} is a morphism in
$\freecat[\ccpms](\qubit,\tlist\qubit)$, that is, a map sending a
$2\times 2$ positive matrix onto $\bigoplus_nP(\C^{2^n\times 2^n})$.
The program {\bf qlist} is defined using recursion: its semantics is
the limit of the morphisms $f_n$ sending
$(\begin{smallmatrix}a&b\\c&d\end{smallmatrix})$ to the
infinite sequence $ ({\bf
0},\frac12e_1,\ldots\!,\frac1{2^n}e_n,{\bf 0},{\bf 0},\ldots)$
where $e_i$ is the $2^i{\times}2^i$ positive matrix
\begin{center}
\vspace{-1ex}
\scalebox{0.6}{$\begin{pmatrix}
a&0&\cdots&0&b
\\
0&0&\cdots&0&0
\\[-1ex]
\vdots&\vdots&\ddots&\vdots&\vdots
\\[0.4ex]
0&0&\cdots&0&0
\\
c&0&\cdots&0&d
\end{pmatrix}$}.
\end{center}
\vspace{-1ex}
This limit is the map sending
$(\begin{smallmatrix}a&b\\c&d\end{smallmatrix})$ to the sequence of
infinitely increasing matrices
$
({\bf 0},\frac12e_1,\ldots,\frac1{2^n}e_n,\ldots)
$.
Note that the first element of the sequence is ${\bf 0}$, as the
program {\bf qlist} never return the empty list. Also note that all
the positive matrices in the sequence represent {\em entangled states
of arbitrary sizes}. Our semantics is the first one to be able to
account for such a case: in~\cite{GoIquantum}, only fixed sizes
were allowed for entangled states.
\end{example}
\begin{example}\label{ex:telep-denot}
We claim in the introduction that the model is expressive enough to
describe entanglement at higher-order types. As we discuss in
Example~\ref{ex:telep-term}, the encoding of the
quantum teleportation algorithm produces two entangled, mutually inverse
functions: $f:\qubit\loli\bit\tensor\bit$ and
$g:\bit\tensor\bit\loli\qubit$.
The term $({\bf teleport}\,\punit)$ of type $(\qubit \loli
\bit\otimes\bit)\tensor(\bit\otimes\bit \loli \qubit)$ is one
instance of such a pair of functions. Its denotation is a finite sequence
of $16$ square matrices of size $4\times 4$. Using a lexicographic
convention, we can lay them out as in Fig.~\ref{tab:Axyzt}.
Because of the convention, morally each row corresponds to an
element of type $\bit\otimes\bit \loli \qubit$ whereas each column
corresponds to an element of type $\qubit\loli
\bit\otimes\bit$. Picking a row, i.e., a choice of two left-sided
booleans, amounts to choosing the two booleans that will be passed to
the function $g$. Picking a column, i.e., a choice of two right-sided
booleans, amounts to deciding on the probabilistic result we get
from the function $f$. The intersection of a column and a row is
therefore the representation of a map $\qubit\loli\qubit$. This map
is a description of a possible path in the control flow of the
algorithm.
The matrices on the diagonal correspond to a run of the algorithm
as it was intended: applying $g$ to the result of $f$. Since they
are supposed to be the identity on $\qubit$, we can therefore deduce
that the matrices $A_{00,00}$, $A_{01,01}$, $A_{10,10}$ and
$A_{11,11}$ are all equal to $\left(\begin{smallmatrix}
1&0&0&1\\
0&0&0&0\\
0&0&0&0\\
1&0&0&1
\end{smallmatrix}\right).
$
Since this matrix cannot be written as the tensor of two $2\times2$
matrices, we conclude that the denotation $A$ of $({\bf
teleport}\,\punit)$ is indeed entangled.
We can compute the other matrices $A_{xy,zt}$ using the same
argument: in general, $A_{xy,zt}$ is a composition of $f$ and $g$,
except that instead of applying $g$ to $(x,y)$, we apply it to
$(z,t)$. We therefore get a function $\qubit\to\qubit$ constructed
out of the $U_{--}$ that might (if $xy=zt$) or might not be the
identity. In general, the matrix $A_{xy,zt}$ is the denotation of
the unitary $U_{zt}U_{xy}^*$. The denotation $A$ is given in full
detail in Table~\ref{tab:Axyzt}.
\end{example}
\begin{remark}
Example~\ref{ex:telep-denot} is a good illustration of what we
claimed in the introduction: the model reflects the
juxtaposition of quantum and classical structures, even at
higher-order types. Here, the control-flow is handled by
the biproduct structure, and the quantum part of the algorithm is
split across the list of $4{\times}4$~matrices.
\end{remark}
\begin{table*}
\[
\begin{array}{
r@{{}={}}c@{\quad}c@{{}={}}c@{\quad}c@{{}={}}c@{\quad}c@{{}={}}cc}
A \quad=\quad\frac14\bigg(\quad
A_{00,00}&
\left(\begin{smallmatrix}
1&0&0&1\\
0&0&0&0\\
0&0&0&0\\
1&0&0&1
\end{smallmatrix}\right),
&
A_{00,01}&
\left(\begin{smallmatrix}
0&0&0&0\\
0&1&1&0\\
0&1&1&0\\
0&0&0&0
\end{smallmatrix}\right),
&
A_{00,10}&
\left(\begin{smallmatrix}
1&0&0&\textrm{-}1\\
0&0&0&0\\
0&0&0&0\\
\textrm{-}1&0&0&1
\end{smallmatrix}\right),
&
A_{00,11}&
\left(\begin{smallmatrix}
0&0&0&0\\
0&1&\textrm{-}1&0\\
0&\textrm{-}1&1&0\\
0&0&0&0
\end{smallmatrix}\right),
\\
A_{01,00}&
\left(\begin{smallmatrix}
0&0&0&0\\
0&1&1&0\\
0&1&1&0\\
0&0&0&0
\end{smallmatrix}\right),
&
A_{01,01}&
\left(\begin{smallmatrix}
1&0&0&1\\
0&0&0&0\\
0&0&0&0\\
1&0&0&1
\end{smallmatrix}\right),
&
A_{01,10}&
\left(\begin{smallmatrix}
0&0&0&0\\
0&1&\textrm{-}1&0\\
0&\textrm{-}1&1&0\\
0&0&0&0
\end{smallmatrix}\right),
&
A_{01,11}&
\left(\begin{smallmatrix}
1&0&0&\textrm{-}1\\
0&0&0&0\\
0&0&0&0\\
\textrm{-}1&0&0&1
\end{smallmatrix}\right),
\\
A_{10,00}&
\left(\begin{smallmatrix}
1&0&0&\textrm{-}1\\
0&0&0&0\\
0&0&0&0\\
\textrm{-}1&0&0&1
\end{smallmatrix}\right),
&
A_{10,01}&
\left(\begin{smallmatrix}
0&0&0&0\\
0&1&\textrm{-}1&0\\
0&\textrm{-}1&1&0\\
0&0&0&0
\end{smallmatrix}\right),
&
A_{10,10}&
\left(\begin{smallmatrix}
1&0&0&1\\
0&0&0&0\\
0&0&0&0\\
1&0&0&1
\end{smallmatrix}\right),
&
A_{10,11}&
\left(\begin{smallmatrix}
0&0&0&0\\
0&1&1&0\\
0&1&1&0\\
0&0&0&0
\end{smallmatrix}\right),
\\
A_{11,00}&
\left(\begin{smallmatrix}
0&0&0&0\\
0&1&\textrm{-}1&0\\
0&\textrm{-}1&1&0\\
0&0&0&0
\end{smallmatrix}\right),
&
A_{11,01}&
\left(\begin{smallmatrix}
1&0&0&\textrm{-}1\\
0&0&0&0\\
0&0&0&0\\
\textrm{-}1&0&0&1
\end{smallmatrix}\right),
&
A_{11,10}&
\left(\begin{smallmatrix}
0&0&0&0\\
0&1&1&0\\
0&1&1&0\\
0&0&0&0
\end{smallmatrix}\right),
&
A_{11,11}&
\left(\begin{smallmatrix}
1&0&0&1\\
0&0&0&0\\
0&0&0&0\\
1&0&0&1
\end{smallmatrix}\right)&\quad\bigg).
\end{array}
\]
\caption{\footnotesize The denotation of the quantum
teleportation algorithm.}
\label{tab:Axyzt}
\end{table*}
\section{Adequacy}\label{sec:adequacy}
In the following, we prove the adequacy of $\freecat[\ccpms]$
(Theorem~\ref{th:adequacy}). This amounts to achieving the converse
inequality of Corollary~\ref{cor:invariance_interpretation}. The proof
uses a syntactic approach, following~\cite{GoIquantum}. We introduce a
bounded {\tt letrec}$^n$, which can be unfolded at most $n$ times. On
the one hand, the language allowing only bounded {\tt letrec} is
strongly normalizing (Lemma~\ref{lem:SN-ext}), hence the adequacy for
it can be easily achieved by induction on the longest reduction
sequence of a term (Corollary~\ref{cor:finit_adequacy}). On the other
hand, the unbounded {\tt letrec} can be expressed as the supremum of
its bounded approximants, both semantically
(Lemma~\ref{lem:finitary0}) and syntactically
(Lemma~\ref{lem:finitary2}). We then conclude the adequacy for the
whole quantum lambda calculus by continuity.
\begin{definition}\label{def:letrecn}
Let us extend the grammar of terms (Table~\ref{table:terms_grammar})
by adding: (i) a new term $\Omega^A$; (ii) a family of new term
constructs $\letrecn{n}{f^{A\loli B}}{x}{M}{N}$ indexed by natural
numbers $n\geq 0$.
The typing rules for these new constructs are
\[
\infer{
\bang\Delta\entail\Omega^A:A
}{}
\qquad
\infer{
\bang\Delta,\Gamma\entail\letrecn{n}{f^{A\loli B}}{x}{M}{N}:C
}{
\begin{array}{l}
\bang\Delta,f:\bang{(A\loli B)},x:A
\entail
M:B
\\
\bang\Delta,\Gamma,f:\bang{(A\loli B)}
\entail
N:C
\end{array}
}
\]
Their denotations are given, respectively, by the map $\bf 0$ and
the family of maps
\[
\oc\Delta\tensor\Gamma
\xrightarrow{\contr}
\oc\Delta\tensor\Gamma\tensor\oc\Delta
\xrightarrow{\!\!\id\tensor(\dig;\bierman;\oc{(\Lambda\phi)})^n\!\!}
\oc\Delta\tensor\Gamma\tensor\oc{(A\loli B)}
\xrightarrow{\psi}
C,
\]
where $\phi\in\freecat[\ccpms](\oc\Delta\otimes\oc(A\multimap
B)\otimes A,B)$ and
$\psi\in\freecat[\ccpms](\oc\Delta\otimes\Gamma\otimes\oc(A\multimap
B),C)$ are the denotations of the premises and
$(\dig;\bierman;\oc{(\Lambda\phi)})^n
\in\freecat[\ccpms](\oc\Delta,\oc(A\multimap
B))$ is defined in a similar fashion as in
Table~\ref{table:denotation_rule}.
The reduction rules are updated as follows.
\[
\begin{array}{@{}l@{}}
\am{q,\ell,\letrecn{0}{f^{A\loli B}}{x}{M}{N}} \xredto[1]
\am{q,\ell,N\{(\lambda x^A.\Omega^B)/f\}}
\\[1ex]
\am{q,\ell,\letrecn{n+1}{f^{A\loli B}}{x}{M}{N}} \\
~\qquad~\xredto[1]
\am{q,\ell,N\{(\lambda x^A.\letrecn{n}{f^{A\loli B}}{x}{M}{M})/f\}}.
\end{array}
\]
\end{definition}
The additions to the language do not modify the properties of the
language: subject reduction (Proposition~\ref{prop:subject_reduction})
and totality (Lemma~\ref{lemma:totality}) hold as they are stated,
while type safety (Proposition~\ref{prop:safety}) and soundness
(Proposition~\ref{prop:semantic_invariance}) are satisfied, with the
proviso of considering the set of normal forms to consist of the set
of values \emph{and} the set of terms containing $\Omega$ in
evaluating position.
\begin{definition}\label{def:finitary}
A term is called \define{finitary} when it does not contain any
occurrence of the un-indexed {\tt letrec} construct. It can however
contain $\Omega$ and any of the indexed {\tt letrec}${}^n$. We call a
closure \define{finitary} when its term is finitary.
\end{definition}
\begin{lemma}[Strong normalization]\label{lem:SN-ext}
If $\am{\qarray_1,\qlist_1,M_1}$ is finitary and typable, then every
reduction sequence of the form
$
\am{\qarray_1,\qlist_1,M_1}\xredto[p_1]
\am{\qarray_2,\qlist_2,M_2}\xredto[p_2]
\am{\qarray_3,\qlist_3,M_3}\xredto[p_3]
\cdots
$
is finite.
\end{lemma}
\begin{proof}[Proof (Sketch)]
We reduce the finitary quantum lambda calculus to a simply typed
non-deterministic language without quantum states, for which a
standard proof technique can be used. The terms of this language are
the terms of the extended quantum lambda calculus, minus the {\tt
letrec} construct. The operational semantics is obtained from
Table~\ref{table:reduction} and the rules for {\tt letrec}$^n$ by
replacing closures with the respective terms and the rules of
Table~\ref{subtable:reduction_quantum} by dummy reduction rules:
like $ U(\bullet\otimes\dots\otimes\bullet) \redto
\bullet\otimes\dots\otimes\bullet$, or $ \new~\ffalse \redto
\bullet$. The symbol $\bullet$ denotes a distinct term variable,
which, by convention, it is never bound by an abstraction. Clearly,
the strong normalization of this language implies that of the
finitary quantum lambda calculus.
\end{proof}
\begin{corollary}[Finitary adequacy]\label{cor:finit_adequacy}
Let $M$ be a closed finitary term of unit type. Then $
\denot{M}^{\vdash\tunit}_\ast=\Halt_{\am{\ket{\,},\ket{\,},M}}. $
\end{corollary}
\begin{proof}[Proof (Sketch)]
We prove that, for any total finitary quantum closure of unit type
$\am{\qarray,\qlist,M}$ we have $\denot{M}^{\qlist\vdash
1}(\qarray\qarray^\ast)=\Halt_{\am{\qarray,\qlist,M}}$. In fact,
by Lemma~\ref{lem:SN-ext}, there exists $m\in\N$ such that
$\Halt_{\am{\qarray,\qlist,M}}= \sum_{\am{\qarray',\qlist',V}}
\Red^m_{\am{\qarray,\!\qlist,\!M},\am{\qarray'\!,\!\qlist'\!,\!V}}$. We
conclude by induction on $m$.
\end{proof}
\begin{definition}\label{def:approx}
Let $\apprv$ be a relation between finitary terms and general terms
defined as the smallest congruence relation on terms satisfying, for
every $M\apprv M'$ and $N\apprv N'$:
\begin{align*}
N\{(\lambda x^A.\Omega^B)/f\}&\apprv(\letrec{f}{x}{M'}{N'}),\\
(\letrecn{n}{f}{x}{M}{N})&\apprv(\letrec{f}{x}{M'}{N'}).
\end{align*}
\end{definition}
\begin{lemma}\label{lem:finitary0}
If $\Gamma\vdash M:A$, then $
\denot{M}^{\Gamma\vdash A}\!\!\!\!=\dirsup_{\!\!
\substack{
M'\apprv M\\
M'\text{ finitary}\!\!\!\!\!
}
}\denot{M'}^{\Gamma\vdash A}$.\qed
\end{lemma}
\begin{lemma}\label{lem:finitary2}
If $M\apprv M'$, then
$\Halt_{\am{\qarray,\qlist,M}}\leq\Halt_{\am{\qarray,\qlist,M'}}$.
\end{lemma}
\begin{proof}[Proof (Sketch)]
By induction on $n$, one proves the inequality:
$\sum_{\am{q',\ell',V}}\Red^n_{\am{q,\ell,M},\am{q',\ell',V}} \leq
\sum_{\am{q',\ell',V}}\Red^n_{\am{q,\ell,M'},\am{q',\ell',V}}$, from
which the statement follows trivially.
\end{proof}
\begin{theorem}\label{th:adequacy}
Let $M$ be a program, i.e., a closed term of unit type. Then
$
\denot{M}^{\vdash\tunit}_\ast=\Halt_{\am{\ket{\,},\ket{\,},M}}.\
$
\end{theorem}
\begin{proof}
By Corollary~\ref{cor:invariance_interpretation} we have
$\denot{M}^{\vdash\tunit}_\ast\geq\Halt_{\am{\ket{\,},\ket{\,},M}}$.
Conversely,
by Lemma~\ref{lem:finitary0},
$\denot{M}^{\vdash\tunit}_\ast=\dirsup_{M'\apprv
M}\denot{M'}^{\vdash\tunit}_\ast$, which is equal to
$\dirsup_{M'\apprv M}\Halt_{\am{\ket{\,},\ket{\,},M'}}$ by Corollary
\ref{cor:finit_adequacy}, which is less or equal to
$\Halt_{\am{\ket{\,},\ket{\,},M}}$ by Lemma~\ref{lem:finitary2}.
\end{proof}
\section{Structure of the sets of representable elements}
\label{subsect:discussion}
We conclude this paper with an analysis of some of the properties of
the denotation of terms.
Recall that a morphism in $\freecat[\ccpms]$ is an indexed family
of either
completely positive maps, or infinite elements added during D-completion.
We show that (1) all types have a non-zero inhabitant; (2) provided
that the term constant $U$ ranges over arbitrary unitary
matrices, the representable elements of a given homset form a convex
set including ${\bf 0}$; and (3) infinite elements are not part of
any representable map.
We first need two auxiliary definitions.
\newcommand{\consume}[1][]{\overline\omega_{#1}}
\newcommand{\generate}[1][]{\omega_{#1}}
\begin{table*}
\[
\begin{array}{@{}l@{{}={}}ll@{{}={}}l@{}}
\generate[\qubit]
&
\lambda\punit.\new\,\ffalse
&
\consume[\qubit]
&
\lambda x^\qubit.\iftermx{\meas\,x}{\punit}{\punit}
\\
\generate[A\loli B]
&
\lambda\punit.\lambda x^A.\letunitterm{(\consume[A]\,x)}{
(\generate[B]\,\punit)}
&
\consume[A\loli B]
&
\lambda f^{A\loli B}.\consume[B]\,(f\,(\generate[A]\,\punit))
\\
\generate[\oc{(A\loli B)}]
&
\lambda\punit.\lambda x^A.(\generate[A\loli B]\,\punit)\,x
&
\consume[\oc{(A\loli B)}]
&
\mu g f^{\oc{(A\loli B)}}.\iftermx{{\bf c}}{\punit}{
(\consume[A\loli B]\,f);(g\,f)}
\\
\generate[\tunit]
&
\lambda\punit.\punit
&
\consume[\tunit]
&
\lambda\punit.\punit
\\
\generate[A\tensor B]
&
\lambda\punit.(\generate[A]\,\punit)\tensor(\generate[B]\,\punit)
&
\consume[A\tensor B]
&
\lambda x^{A\tensor B}.\lettensterm{z_1}{z_2}{x}{
\letunitterm{(\consume[A]\,z_1)}{(\consume[B]\,z_2)}}
\\
\generate[A\oplus B]
&
\lambda\punit.\iftermx{{\bf c}}{(\generate[A]\,\punit)}{
(\generate[B]\,\punit)}
&
\consume[A\oplus B]
&
\lambda x^{A\oplus B}.\match{x}{z_1^A}{\consume[A]\,z_1}{z_2^B}{
\consume[B]\,z_2}
\\
\generate[\tlist{A}]
&
\mu f
\punit.\iftermx{{\bf c}}{(\punit)}{\cons{(\generate[A]\,\punit)}{
(f\,\punit)}}
&
\consume[\tlist{A}]
&\begin{array}[t]{@{}l@{}}
\mu f
x^{\tlist{A}}.{\tt match}\,\splitlist\,x\,{\tt with}~
\\
\hspace{-.43in}(~z_1^\tunit~:~z_1\bor z_2^{A\tensor\tlist{A}}~:~
{\tt let}~{y_1}\tensor{y_2}={z_2}~{\tt in}~{
\letunitterm{(\consume[A]\,y_1)}{(f\,y_2)}})
\end{array}
\end{array}
\]
\caption{\footnotesize Two mutually recursive families of terms}
\label{tab:consume-generate}
\end{table*}
\begin{definition}
We define two type-indexed families of terms $\consume[A]$ and
$\generate[A]$ by mutual induction in
Table~\ref{tab:consume-generate}. The term ${\bf c}$ represents the
fair coin toss $\meas\,(H\,(\new\,\ffalse))$ (recall
Example~\ref{ex:cointoss}) and the notation $\mu f x.M$ stands for
$\letrec{f}{x}{M}{f}$.
\end{definition}
\begin{lemma}\label{lem:non-zero-image}
For all types $A$, we have $\entail\generate[A]:\tunit\loli A$ and
$\entail\consume[A]:A\loli\tunit$. Moreover, the morphisms
$\denot{\generate[A]}^{\entail\tunit\loli A}$ and
$\denot{\consume[A]}^{\entail A\loli\tunit}$, seen as indexed families,
do not contain the zero map.\qed
\end{lemma}
\begin{corollary}
All types are inhabited by at least one closed value of non-null
denotation.
\end{corollary}
\begin{proof}
Immediate with Lemma~\ref{lem:non-zero-image}: for a given type $A$,
choose the term $(\generate[A]\,\punit)$.
\end{proof}
\begin{proposition}
Given a type $A$ and a context $\Gamma$, the denotations
$\denot{M}^{\Gamma\entail A}$ of valid typing judgements $\Gamma\entail
M:A$ form a convex set including ${\bf 0}$.
\end{proposition}
\begin{proof}
Suppose that $\Gamma$ is $x_1:A_1,\ldots,x_n:A_n$.
A term $M$ mapping to $\bf 0$ is
$
(\consume[A_1] x_1;\ldots;\consume[A_n] x_n;{\bf\Omega})
$
where the term ${\bf\Omega}$ is a shortcut for
$\letrec{f}{x}{f\,x}{f\,\punit}$, of denotation
$\bf 0$.
Now, suppose that $f=\denot{M_1}^{\Gamma\entail A}$ and
$g=\denot{M_2}^{\Gamma\entail A}$, and choose two non-negative real
numbers $\rho_1$, $\rho_2$ such that $\rho_1+\rho_2 = 1$.
There exists an angle $\phi$ such that $(\cos \phi)^2=\rho_1$ and
that $(\sin \phi)^2=\rho_2$.
As the term constants $U$ range over arbitrary unitaries,
the unitary matrix $V_\phi=
(\begin{smallmatrix}\cos\phi&-\sin\phi\\\sin\phi&\cos
\phi\end{smallmatrix})$
is representable in the quantum lambda calculus. The term
$
{\bf c'} = \meas\,(V_\phi\,(\new\,\ffalse))
$
has denotation $(\rho_1,\rho_2)$. We then conclude that
the term
$\iftermx{{\bf c'}}{M_1}{M_2}$
has denotation $\rho_1f+\rho_2g$.
\end{proof}
\begin{proposition}
If $\Gamma\entail M:A$ is valid, then no infinite element is part of
the denotation $\denot{M}^{\Gamma\entail A}$ of $M$.
\end{proposition}
\begin{proof}
Suppose that one of the infinite elements of the D-comple\-tion were
to be found in the interpretation of $x_1:A_1,\ldots,x_n:A_n\entail
M:A$. Then the closed term
\[
(\lambda x_1\ldots x_n.\consume[A]\,M)
(\generate[A_1]\punit)\ldots(\generate[A_n]\punit)
\]
of type $\tunit$ has infinite denotation, contradicting
Theorem~\ref{th:adequacy}.
\end{proof}
This last proposition indicates that infinite elements introduced
during the D-completion are really an
artifact only needed for the categorical construction. The
representable elements in the model are only built out of families of
completely positive maps.
\section{Conclusion}
We presented a higher-order lambda calculus for quantum computation
featuring classical and quantum data, duplication, recursion, and an
infinite parametric type for lists. We then answered a long-standing
open question: the description of a model for the full quantum lambda
calculus. The model we propose is a free construction based on the
known model of completely positive maps, but nevertheless has a
concrete presentation.
One thing that this model explains and illustrates is the distinction
between the quantum and classical parts of the language. The quantum
part is described by completely positive maps (finite dimension),
whereas the classical
control is given by the Lafont category (i.e., linear logic). The
model demonstrates that the two ``universes'' work well together, but
also -- surprisingly -- that they do not mix too much, even at higher
order types (we always have an {\em infinite} list of {\em finite}
dimensional CPMs). The control flow is completely handled by the
biproduct completion, and not by the CPM structure.
The adequacy result, moreover, validates that the model is a ``good''
representation of the language.
One should also note that the product and the coproduct coincide in
our model. For example, the model has morphisms that correspond to a
program returning true with probability $1$ and false with probability
$1$. We would like to point out that our interpretation is not
{\em surjective}. For example, there are also morphisms in the model
corresponding to ``probability 2''. (Incidentally, adding terms with
such behavior makes it possible to build a term whose denotation is
$\infty$ -- so the fact that this provably does not happen somehow
captures the sanity of the model). Interpretations in denotational
models are often not surjective. In fact, it is an open problem to
give a non-syntactic characterization of the image of our
interpretation. Similarly, the problem of full-abstraction is still
open.
\bibliographystyle{abbrv}
|
\section{Introduction}
The growth of Internet raised the problem of illegal redistribution as a
major concern in digital content industry, because copying such material
is easy and no information is lost in the process. To protect digital copies,
however, is a complicated task. Methods like cryptography do not resolve this
problem, since the information must be decrypted at one point to be able to
use it. The goal of digital fingerprinting is to discourage people from
illegally redistributing their legally purchased copy. In this scenario, the
distributor embeds into the digital content, using a watermark algorithm, a
unique piece of information({\it{fingerprint}}) for each user. If an illegal copy is
found, the distributor can extract the fingerprint from it to identity the
dishonest user({\it{pirate}}). Because the pirate may try to damage the fingerprint
before redistribution, the watermarking algorithm must ensure robustness to the distributor.
Nevertheless, the most dangerous attack against digital fingerprinting is the
collusion attack introduced in \cite{bon}. The contents delivered to different
users are, since their
fingerprints differ, essentially different. Two or more pirates may compare
their copies and reveal the locations of part of fingerprint. With deleting or
modifying those locations, pirates can generate a new copy of content in order
not to be traced. This collusion attack could not only violate
pirate-identifying but frame an innocent user in some cases. We are interested
in designing a set of fingerprints({\it{fingerprinting code}}) with which the
distributor can always identity at least one colluder from a forged fingerprint
with a small error probability. In particular separating code, IPP code and TA code are
most important fingerprinting codes with different
collusion-secure properties for generic digital data.
We will denote the $i$th component of any tuple $x$ by $x_i$ and the Hamming
distance between two tuples $x, y$ by $d(x,y)$. Let $n,w,w_1$ and $w_2$ be positive
integers such that $n,w,w_1 \geq 2$ and $w_1 \geq w_2$. Suppose $C$ is a code of length
$n$ over $\F_q$.
\begin{itemize}
\item We define \textit{descendant set} of an arbitrary nonempty subset $U$ of $C$ by
\begin{equation*}
\textnormal{desc}U:=\{x \in \F_q^n \mid \forall i, \exists y \in U : x_i=y_i \}
\end{equation*}
\item $C$ is a $(w_1,w_2)$-\textit{separating code} provided that, if $U_1,U_2$ are
disjoint subsets of $C$ such that $1 \leq |U_1| \leq w_1$ and $1 \leq |U_2| \leq w_2$,
then their descendant sets are also disjoint.
\item $C$ is called a $w$-\textit{identifiable parent property code}(IPP code)
provided that for all $x \in \F_q^n$, the set $\textrm{IPP}_w(x):=\{U \subset C \mid
x \in \textrm{desc}U, 1 \leq |U| \leq w \}$ is empty or
$\bigcap \limits_{U \subset \textrm{IPP}_w(x)}U \neq \phi$.
\item $C$ is called a $w$-\textit{traceability code}(TA code) provided that if
$U \subset C, 1 \leq |U| \leq w$ and $x \in \textrm{desc}U$, there exist at
least one codeword $y \in U$ such that $d(x,y)<d(x,z)$ for all $z \in C \backslash U$.
\end{itemize}
The code classes defined above are known to satisfy the following relationships.
\begin{prop}
\textnormal{(see \cite{ssw01})} Let $d$ be the minimum distance of a code $C$ of
length $n$. Then for $C$, \\
\centerline{
$d>n(1-1/w^2) \Rightarrow w$-TA $\Rightarrow w$-IPP $\Rightarrow (w,w)$-separating}
\end{prop}
\begin{prop}
\textnormal{(see \cite{c})} Let $d$ be the minimum distance of a code $C$ of
length $n$. If $d>n(1-1/(w_1w_2))$, then $C$ is a $(w_1,w_2)$-separating code.
\end{prop}
Let $1 \leq k \leq q-1$ be an integer. The \textit{Reed-Solomon code} $\RS_k(q)$
of dimension $k$ over $\F_q$ is defined by $\RS_k(q):=\{ev(f) \mid f \in \F_q[x],
\textrm{deg}f<k\}$, where $ev:f \in F_q[x] \mapsto (f(\alpha^0),f(\alpha^1),
\cdots,f(\alpha^{q-2})) \in \F_q^{q-1}$ and $\alpha$ is a primitive element in $\F_q$.
It is well known that $\RS_k(q)$ is a $[q-1,k,q-k]$-linear code.
\par
Reed-Solomon code is one of the most famous error-correcting codes and
it also has an application in digital fingerprinting. A. Silverberg, et al.
\cite{ssw03} dealt with applying list decoding method to tracing algorithms
of fingerprinting codes. In their work, the collusion-secure properties of
Reed-Solomon codes and other algebraic geometry codes were studied, and
the following question was left as an open problem.
\begin{que}
Is it the case that $d>n-n/w^2$ for all $w$-IPP Reed-Solomon codes of length $n$
and minimum distance $d$?
\end{que}
Thus, Silverberg's question is a problem of the equivalence of IPP and traceability
for Reed-Solomon code family. \\
\indent The problem was studied in \cite{fcsd} and \cite{mfs}. In \cite{fcsd}, they
restated the separation property of Reed-Solomon codes algebraically, as a system of
equations, to get the following result.
\begin{thr} \textnormal{(see \cite{fcsd})}
Suppose $k-1$ divides $q-1$. If $\RS_k(q)$ is a $(w_1,w_2)$-separating code,
then $d>n-n/(w_1w_2)$ where $d$ is minimum distance.
\end{thr}
In \cite{mfs}, they presented the similar result as follows by establishing an
additive homomorphism over finite field.
\begin{thr} \textnormal{(see \cite{mfs})}
Suppose $w^2>q$ or $w$ divides $q$. If $\RS_k(q)$ is a $(w,w)$-separating code,
then $d>n-n/w^2$ where $d$ is minimum distance.
\end{thr}
As you can see, the previous works claimed the stronger fact than Silverberg's original
problem in certain cases. In this context, we naturally raise the following question,
which turns out to be the main topic of this article.
\begin{que}
Is it the case that $d>n-n/(w_1w_2)$ for all $(w_1,w_2)$-separating Reed-Solomon
codes of length $n$ and minimum distance $d$?
\end{que}
The rest of the paper is organized as follows: In Section 2, we will present a
sufficient condition for non-separation of linear codes, and prove that the previous
works can be derived from that condition. Some more parameter setups providing
positive answer about Question 2 will be obtained in Section 3. In Section 4, the
application of subspace subcodes of Reed-Solomon codes will be unveiled. We conclude
the paper in Section 5 after presenting experimental results to show the extension
of our work.
\par Throughout the remaining, $\F_q$ is Galois field with order $q=p^m$ and
characteristic $p$. Let $r_i=[\textnormal{log}_pw_i], i=1,2$. For any polynomial
$f$ over $\F_q$, let Im$f=f(\F_q)$. For an arbitrary word $x \in \F_q^n$, Im$x$
is the set of all its components, i.e. Im$x=\{x_i \mid 1 \leq i \leq n\}$.
For given two sets $E,F \subset \F_q$, we define $EF:=\{ab \mid a \in E,
b \in F\}$ and $E+F:=\{a+b \mid a \in E, b \in F\}$. We will denote the set of
all polynomials over $\F_q$ of degree less than $k$ by $P_k$. $n,w,w_1$ and
$w_2$ are positive integers satisfying $n,w,w_1 \geq 2$ and $w_1 \geq w_2$.
\section{Restatement of the Previous Works}
In this section we propose a sufficient condition for non-separation of linear
codes, which will integrate the former results in \cite{fcsd} and \cite{mfs}.
The idea was motivated by \cite{mfs}, where an additive homomorphism was
established such that its image set has a special property. Before presenting
the major result, we will formally define such "special property" of a set. \par
\indent Let $U$ be a subset of $\F_q$. $U$ is called \textit{additively}
(\textit{multiplicatively}) $(w_1,w_2)$- \textit{separable} and written by
$U=(E,F)_{w_1,w_2}$ provided that
there exist two subsets $E,F \subset U$ with $1 \leq |E| \leq w_1$ and
$1 \leq |F| \leq w_2$ such that $U \subset E+F$ ($U \subset EF$). \par
\indent The following theorem is the main result of this section. Note that
it is not just for Reed-Solomon codes, but for linear codes.
\begin{thr}
Let $C$ be $[n,k]_q$-linear code containing \textnormal{\textbf{1}}$=(1,1,\cdots,1)$.
Suppose there exists a codeword $c=(c_1,c_2,\cdots,c_n) \in C$ such that
$|\textnormal{Im}c| \geq 2$ and \textnormal{Im}$c$ is $(w_1,w_2)$-separable additively
or multiplicatively. Then, $C$ is not $(w_1,w_2)$-separating.
\end{thr}
\textit{Proof}. We will only prove when Im$c$ is additively $(w_1,w_2)$-separable,
since the other case can be proven in similar way. Let Im$c=(E,F)_{w_1,w_2}$.
Define $U:=\{\beta \cdot \textbf{1} \mid \beta \in E\}$ and
$V:=\{c-\gamma \cdot \textbf{1} \mid \gamma \in F\}$. Then $U,V \subset C$ since
$c,\textbf{1} \in C$ and $C$ is a linear code. Further, $U$ and $V$ are disjoint because
$|\textnormal{Im}c| \geq 2$.
For all $i \in \overline{1,n}$, there exist $\beta_i \in E$ and $\gamma_i \in F$ such that
$c_i=\beta_i+\gamma_i$. If we set $x:=(\beta_1,\beta_2,\cdots,\beta_n)$, it is clear
that $x \in \textrm{desc}U \cap \textrm{desc}V$ which implies non-separation. $\boxdot$ \par
\bigskip
\noindent The following corollary is the Reed-Solomon code version of Theorem 2.1.
\begin{cor}
Let $1 \leq k \leq q-1$ be an integer. If there exists a non-constant polynomial $f$
in $P_k$ such that \textnormal{Im}$f$ is $(w_1,w_2)$-separable additively or
multiplicatively, then the code $\RS_k(q)$ is not $(w_1,w_2)$-separating.
\end{cor}
\noindent By definition, $\RS_k(q) \subset \RS_{k+1}(q)$, therefore the code $\RS_{k+1}(q)$
is not $(w_1,w_2)$-separating if $\RS_k(q)$ is not $(w_1,w_2)$-separating. Meanwhile,
the inequality $d>n-n/(w_1w_2)$ is equivalent with $k-1<(q-1)/(w_1w_2)$ for
Reed-Solomon codes. Thus, it sufices to consider the case
$k=\lceil (q-1)/(w_1w_2) \rceil + 1$ when we study Question 2.
In other words, if $\RS_k(q)$ is not $(w_1,w_2)$-separating where
$k=\lceil (q-1)/(w_1w_2) \rceil + 1$ for given $q,w_1,w_2$, Question 2 has the positive
answer. (see \cite{mfs}) \par
\indent In this context, we will reprove the previous results done on Silverberg's
open problem more simply using Corollary 2.1. \par \bigskip
\indent \textit{Proof of Theorem 1.1 :} Suppose $d \leq n(1-1/(w_1w_2))$, i.e.
$k-1 \geq (q-1)/(w_1w_2)$. Set $f(x):=x^{k-1}$. Since $k-1 \mid q-1$,
the polynomial $f$ is a multiplicative homomorphism mapping $\F_q^{*}$ to $\F_q^{*}$.
So Im$f$ is a multiplicative subgroup of $\F_q^*$ with order
$|\textrm{Im}f|=|\F_q^*|/|\textrm{Ker}f|=(q-1)/(k-1) \leq w_1w_2$. For $\F_q^*$ is cyclic,
Im$f$ is also cyclic, thus, it has a generator $\gamma$. Set
$E:=\{\gamma^{iw_2} \mid 0 \leq i \leq w_1-1 \}$ and $F:=\{\gamma^j \mid 0 \leq j \leq w_2-1 \}$.
Then it is easy to check that Im$f=(E,F)_{w_1,w_2}$, which implies non-separation by
Corollary 2.1.
Therefore, if $\RS_k(q)$ is $(w_1,w_2)$-separating, then $k-1<(q-1)/(w_1w_2)$. $\boxdot$
\par
\bigskip
\indent \textit{Proof of Theorem 1.2 :} As we mentioned above, it sufices to consider
the case $k=\lceil (q-1)/w^2 \rceil + 1$ only. Assume $w^2>q$. Then $k=2$, thus
$k-1 | q-1$, which makes the condition of Theorem 1.1. Now let's assume that $w|q$.
The polynomial $f(x):=x^{q/w^2}-x$ is an additive homomorphism over $\F_q$ and
$|\textrm{Im}f|=w^2$. By finite group theory, there exist subgroups
$E,F<\textrm{Im}f$ with $w$ elements, respectively, such that Im$f=E+F$.
Further, $f \in P_k$. Thus, from Corollary 2.1 the code RS$_k(q)$ is not
$(w_1,w_2)$-separating. $\boxdot$ \par
\bigskip
From the preceding proofs, we claim that the results in
\cite{fcsd} and \cite{mfs} can be integrated into a simpler scheme.
We conclude this section with the following proposition that resembles
Theorem 1.2 without proof.
\begin{prop}
Suppose $w_1w_2$ divides $q$. If $\RS_k(q)$ is a $(w_1,w_2)$-separating code,
then $d>n-n/(w_1w_2)$ where $d$ is minimum distance.
\end{prop}
\section{New Parameter Setups}
$w_1,w_2$ and $q$ are the parameters specifying Reed-Solomon code and its
separation property. The aim of this section is to propose new configurations of
them that provide Question 2 with positive answer. The underlying principle is again
Theorem 2.1 or Corollary 2.1. \\
\indent In this section, we suppose that $k-1 \mid q$ where
$k=\lceil (q-1)/(w_1w_2) \rceil + 1$. Since $q$ is a prime power, there
exists an integer $s$ such that $q/(k-1)=p^s$. One can easily check that
$p^s$ is the largest power of $p$ which is equal or less than $w_1w_2$.
Therefore, $s=r_1+r_2$ or $s=r_1+r_2+1$. \\
\indent The main idea is to set $f(x):=x^{k-1}-x$, prove that Im$f$ is
additively or multiplicatively $(w_1,w_2)$-separable, and refer to Corollary 2.1.
It is obvious that $f \in P_k$ is an additive homomorphism over $\F_q$ and therefore
Im$f$ is an additive group with $p^s$ elements. Thus,
the problem is to find the setups such that Im$f$ is $(w_1,w_2)$-separable.
The first setup is $s=r_1+r_2$.
\begin{prop}
If $s=r_1+r_2$, then $\textnormal{Im}f$ is additively $(w_1,w_2)$-separable.
\end{prop}
$Proof.$ For $|\textnormal{Im}f|=p^{r_1+r_2}$, there exist two additive subgroups $E$
and $F$ with Im$f=E+F$ such that $|E|=p^{r_1}$ and $|F|=p^{r_2}$. Therefore,
Im$f=(E,F)_{w_1,w_2}$. $\boxdot$ \par
\bigskip
\noindent The second setup is $[w_1/p^{r_1}] \cdot [w_2/p^{r_2}] \geq p$.
\begin{prop}
If $[w_1/p^{r_1}] \cdot [w_2/p^{r_2}] \geq p$, \textnormal{Im}$f$ is
additively $(w_1,w_2)$-separable.
\end{prop}
$Proof.$ If $s=r_1+r_2$, Im$f$ is additively $(w_1,w_2)$-separating by Proposition 3.1.
Assume $s=r_1+r_2+1$. Then there exist three additive subgroups $E,F$ and $P$ of Im$f$ with
Im$f=E+F+P$ such that $|E|=p^{r_1},|F|=p^{r_2}$ and $|P|=p$. Moreover, $P$ is cyclic
since $p$ is a prime number. Let $\alpha$ be its generator. If we set
$P_1:=\{(i \cdot [w_2/p^{r_2}])\gamma \mid 0 \leq i \leq [w_1/p^{r_1}]-1\},
P_2:=\{j\gamma \mid 0 \leq j \leq [w_2/p^{r_2}]-1\}$ and $E'=E+P_1,F'=F+P_2$, then
we get $P=P_1+P_2$ and Im$f=E'+F'$ since $[w_1/p^{r_1}] \cdot [w_2/p^{r_2}] \geq p$.
Therefore, Im$f$ is additively $(w_1,w_2)$-separable. $\boxdot$ \par
\bigskip
\noindent The results of this section can be integrated into the following theorem.
\begin{thr}
Suppose $k-1$ divides $q$ with $q/(k-1)=p^s$,
and $s \leq r_1+r_2$ or $[w_1/p^{r_1}] \cdot [w_2/p^{r_2}] \geq p$.
If $\RS_k(q)$ is a $(w_1,w_2)$-separating code,
then $d>n-n/(w_1w_2)$ where $d$ is minimum distance.
\end{thr}
Now we state the following lemma which will be useful
for the next section.
\begin{lem}
The finite field $\F_{p^s}$ is $(w_1,w_2)$-separable if at least one of the followings
hold :
\begin{itemize}
\item $s \leq r_1+r_2$
\item $[w_1/p^{r_1}] \cdot [w_2/p^{r_2}] \geq p$
\item $w_1w_2-w_2 \geq p^s$
\end{itemize}
\end{lem}
$Proof.$ We can prove in the first and second cases similarly with the propositions above
since Im$f$ is additively isomorphism with $\F_{p^s}$. So we will only consider the third case.
It is well known that $\F_{p^s}^{*}=\F_{p^s} \backslash \{0\}$ is a multiplicative cyclic group.
Denote by $\alpha$ its generator. Set $E:=\{\alpha^{i(w_2-1)} \mid 0 \leq i \leq w_1-1\}$ and
$F:=\{\alpha^j \mid 0 \leq j \leq w_2-2\}$. Then $EF=\{\alpha^i \mid 0 \leq i \leq w_1w_2-w_2\}$ and
$\F_{p^s}^{*}=EF$ since $w_1w_2-w_2 \geq p^s$. Thus, if we set $F'=F \cup \{0\}$,
then $\F_{p^s}=EF'$ which implies $\F_{p^s}=(E,F')_{w_1,w_2}$. $\boxdot$ \par
\section{Application of Subspace Subcodes}
In a linear code, there are some codewords all of whose components belong to a certain
subset of $\F_q$. Collecting such codewords is a method of constructing a new code
from an existing code, and it was studied in \cite{d}, \cite{j} and \cite{hms}.
Subfield subcode in \cite{d} is a set of codewords whose components all lie in a subfield.
Subgroup subcodes, or subspace subcodes were introduced in \cite{j} and \cite{hms}, where
their dimensions were estimated. Let $S$ be a $v$-dimensional subspace of $\F_q$ where
$0 \leq v \leq m$. \textit{Subspace subcode} of Reed-Solomon code $C=\RS_k(q)$ with $S$ is
defined to be the set of codewords from $C$ whose components all lie in $S$, and is
denoted by $\SSRS_S(C)$.
In this section, we will study application of subspace subcodes of Reed-Solomon
codes to Question 2 in case $p=2$ and $q=2^m$.
It is related to the dimensions of $\SSRS_S(C)$. \\
\indent $\SSRS_S(C)$ is an $\F_2$-linear space. In \cite{hms}, the explicit formula
to calculate the binary dimension of $\SSRS_S(C)$ denoted by $K(C,S)$ was proposed as
follows :
\begin{equation*}
K(C,S)=\sum\limits_{j \in I_n}d_j(a_j-r_j)
\end{equation*}
\noindent where $I_n$ is the set consisting of the smallest integers in each modulo
$n=2^m-1$ cyclotomic coset, $d_j$ is the cardinality of the coset containing
$j$ denoted by $\Omega_j$, $e_j$ is the number of elements from $\Omega_j$
lying in the set $J=\{1,2,\cdots,k\}$, $a_j=me_j/d_j$ and $r_j$'s are the
ranks of certain $(m-v) \times a_j$ matrices called cyclotomic matrices. \par
\indent As well as the explicit formula, they presented the following lower bound
for the binary dimension :
\begin{equation*}
K(C,S) \geq L(k,v)=\sum\limits_{j \in I_n} \textrm{max} \{d_j(a_j-(m-v)),0\}
\end{equation*}
We will call subspace subcode
$\SSRS_S(C)$ \textit{trivial}, provided that $K(C,S) \leq v$. Then the following
lemma is immediately obtained.
\begin{lem}
Suppose that there exists a $v$-dimensional subspace of $\F_q$
denoted by $S$ such that the subspace subcode of $C=\RS_k(q)$ with $S$ is non-trivial.
Then, $C$ is not $(w_1,w_2)$-separating, provided that $S$ is $(w_1,w_2)$-separable.
\end{lem}
\textit{Proof.} There exists a codword $c \in \SSRS_S(C)$ with $|$Im$c| \geq 2$ because of
non-triviality. Therefore, by Theorem 2.1, $C$ is not $(w_1,w_2)$-separating. $\boxdot$ \par
\bigskip
By using the lower bound $L(k,v)$, we can get the more practical result about Question 2.
$L(k,v)$ depends on the dimension of the parent code $k$ and the dimension of the subspace
$s$, not the subspace $S$ itself. So we can restrict to $S=\F_{2^v}$. Suppose $\SSRS_S(C)$
is trivial, then $\SSRS_T(C)$ is also trivial where $T$ is a subspace of $S$. Therefore
it sufices to consider the largest power $2^v$ equal or less than $w_1w_2$.
\begin{thr}
Let $2^v$ be the largest power equal or less than $w_1w_2$, which satisfies
at least one of the following conditions hold, and suppose $L(k,v)>v$.
\begin{itemize}
\item $v \leq r_1+r_2$
\item $[w_1/2^{r_1}] \cdot [w_2/2^{r_2}] \geq 2$
\item $w_1w_2-w_2 \geq 2^v$
\end{itemize}
If $\RS_k(q)$ is a $(w_1,w_2)$-separating code, then $d>n-n/(w_1w_2)$ where $d$ is minimum
distance.
\end{thr}
\textit{Proof.} $\SSRS_S(C)$ is non-trivial since $K(C,S) \geq L(k,v)>v$ where $S=\F_{2^v}$.
Plus, $\F_{2^v}$ is $(w_1,w_2)$-separable by Lemma 3.1. Therefore, applying Lemma 4.1
implies the conclusion. $\boxdot$
\section{Examples and Conclusions}
In this article, we presented an algebraic statement for the generalized version of
Silverberg's open problem, and exploited it to integrate the former results.
Besides the previous results, we could procure some new parameter setups ensuring the
equivalence of separation and traceability properties for Reed-Solomon codes. Finally
using the concept of subspace subcode introduced in error-correcting code theory,
we proposed a new result when the characteristic of finite field is 2. \par
\indent Table-1 illustrates the contributions of our work to Silverberg's open problem
for some parameters. For each $w$ and $q$, we set $k=\lceil (q-1)/w^2 \rceil + 1$ and
check $(w,w)$-separation property of $\RS_k(q)$ by the existing results. In each cell,
the source of the work is written if $\RS_k(q)$ is not $(w,w)$-separating. For example, "\cite{fcsd}"
means that non-separation is proven by Theorem 1.1, and "3.1" represents that it is
followed by Theorem 3.1 of our paper. The symbol "*" denotes the trivial cases
$w^2 \geq q$, and "-" stands for pending cases. \par \bigskip
\indent \textbf{Example 1 :} Let $w=15$ and $q=256$. Then
$k=\lceil (q-1)/w^2 \rceil + 1 = 3$. So $k-1 \mid q$. Moreover, since
$(15/2^3)^2>2$, the condition of Theorem 3.1 holds. Therefore $\RS_k(q)$ is
not $(w,w)$-separating. Now let $w=10$ and $q=128$. Then $(10/2^3)^2<2$.
However, $q/(k-1)=64=2^6$, so the condition of Theorem 3.1 holds and
$\RS_k(q)$ is not $(w,w)$-separating. \par
\bigskip
\indent \textbf{Example 2 :} Let $w=12$ and $q=2048$. Then $k=16$. Theorem 3.1
cannot be applied in this case, since $k-1=15$ divides neither $q$ nor $q-1$.
The largest power of 2 equal or less than $w^2=144$ satisfying at least one of
the conditions in Lemma 4.1 is $2^7=128$, for $w^2-w=132>128$.
The modulo $n=2047$ cyclotomic coset containing 1 is
$\Omega_1=\{1,2,4,8,16,32,64,128,256,512,1024\}$, therefore $d_j=|\Omega_1|=11$
and $a_j=e_j=|\Omega_1 \cap J|=5$ where $J=\{1,2,\cdots,16\}$. So
$K(C,S) \geq L(k,v) \geq \textnormal{max}\{d_1(a_1-(m-v)),0\}=11>7$, which implies
$\RS_k(q)$ is not $(w_1,w_2)$-separating by Theorem 4.1. \par
\bigskip
\centerline{
\begin{tabular}
{|r|c|c|c|c|c|c|c|c|c|c|c|c|} \hline
$q$ & 16 & 32 & 64 & 81 & 125 & 128 & 243 & 256 & 512 & 1024 & 2048 & 2187 \\ \hline
$w=2$ & \cite{mfs} & \cite{mfs} & \cite{mfs} & \cite{fcsd} & \cite{fcsd} & \cite{mfs}
& - & \cite{mfs} & \cite{mfs} & \cite{mfs} & - & \cite{mfs}\\ \hline
$w=3$ & - & - & \cite{fcsd} & \cite{mfs} & - & -
& \cite{mfs} & - & - & - & - & \cite{mfs}\\ \hline
$w=4$ & \cite{fcsd} & \cite{mfs} & \cite{mfs} & \cite{fcsd} & - & \cite{mfs}
& - & \cite{mfs} & \cite{mfs} & \cite{mfs} & \cite{mfs} & -\\ \hline
$w=5$ & * & 3.1 & \cite{fcsd} & \cite{fcsd} & \cite{mfs} & -
& - & - & - & - & - & -\\ \hline
$w=7$ & * & * & 3.1 & 3.1 & - & 4.1
& - & - & - & - & - & -\\ \hline
$w=9$ & * & * & * & \cite{fcsd} & \cite{fcsd} & 3.1
& \cite{mfs} & 3.1 & \cite{fcsd} & - & - & \cite{mfs}\\ \hline
$w=10$ & * & * & * & \cite{fcsd} & \cite{fcsd} & 3.1
& 3.1 & \cite{fcsd} & - & \cite{fcsd} & - & -\\ \hline
$w=12$ & * & * & * & * & * & *
& \cite{fcsd} & 3.1 & 3.1 & 3.1 & 4.1 & -\\ \hline
$w=13$ & * & * & * & * & * & *
& \cite{fcsd} & 3.1 & 3.1 & 4.1 & - & -\\ \hline
$w=14-15$ & * & * & * & * & * & *
& \cite{fcsd} & 3.1 & 4.1 & - & - & -\\ \hline
$w=17$ & * & * & * & * & * & *
& * & * & 3.1 & 3.1 & 3.1 & -\\ \hline
$w=18$ & * & * & * & * & * & *
& * & * & 3.1 & 3.1 & 4.1 & -\\ \hline
$w=19-22$ & * & * & * & * & * & *
& * & * & 3.1 & \cite{fcsd} & - & -\\ \hline
$w=24$ & * & * & * & * & * & *
& * & * & * & 3.1 & 3.1 & -\\ \hline
$w=28-31$ & * & * & * & * & * & *
& * & * & * & 3.1 & 4.1 & 3.1\\ \hline
$w=34-40$ & * & * & * & * & * & *
& * & * & * & * & 3.1 & \cite{fcsd}\\ \hline
\end{tabular}
} \par
\bigskip
\centerline{
Table-1. Contributions to Silverberg's Problem for Some Parameters
}
\bigskip
\noindent Thus, for a large family of Reed-Solomon codes with $2 \leq w \leq 40$ and
$16 \leq q \leq 2187$, the separation and traceability properties are equivalent.
|
\section{Introduction}
Let $\Omega\subset\mathbb{R}^d$ $(d\geq 1)$ be an open and bounded set,
and consider a function $u\in L^\infty(\Omega)$.
The Neighborhood filter operator is defined by
\begin{equation}
\label{def.NF}
\NF^h u (\mathbf{x})=\frac{1}{C(\mathbf{x})}\int_\Omega \textrm{e}^{-\frac{\abs{u(\mathbf{x})-u(\mathbf{y})}^2 }{h^2}} u(\mathbf{y})d\mathbf{y},
\end{equation}
where $h$ is a positive constant, and \[C(\mathbf{x})=\int_\Omega \exp\left(-\abs{u(\mathbf{x})-u(\mathbf{y})}^2) h^{-2}\right) d\mathbf{y}\] is a normalization factor, intended to allow constants to be fixed points of $\NF^h$.
The Neighborhood filter (NF) is the simplest particular case of other related filters involving local terms,
notably the Yaroslavsky filter \cite{Yaroslavsky1985,Yaroslavsky2003}, the SUSAN filter \cite{Smith1997} introduced by Smith and Brady, the Bilateral filter \cite{Tomasi1998} of Tomasi and Manduchi, and the Nonlocal Means filter (NLM) \cite{Buades2005} by Buades, Coll and Morel.
These methods have been introduced in the last decades
as efficient alternatives to local methods such as those expressed in terms
of nonlinear diffusion partial differential equations (PDE's), among which the pioneering approaches of Perona and Malik \cite{Perona1990}, \'Alvarez, Lions and Morel \cite{Alvarez1992} and Rudin, Osher and Fatemi \cite{Rudin1992} are fundamental.
We refer the reader to \cite{Buades2010} for a review and comparison of these methods.
Among all these filters, the Neighborhood filter is the simplest, but yet useful,
method due to its compromise between denoising quality and computational speed.
Indeed, although it creates shocks and staircasing effects \cite{Buades2006a}, the computational cost
is by far lower than those of other integral kernel filters or PDE's based methods.
Since, usually, a single denoising step of the nonlocal filters is not enough,
an iteration is performed according to several choices of
the iteration actualization, see \fer{def.GFF} and \fer{def.GFV} for two of such
strategies.
In this context, the Neighborhood filter and its variants have been analyzed
from different points of view. For instance, Barash \cite{Barash2002}, Elad \cite{Elad2002}, Barash et al. \cite{Barash2004}, and Buades et al. \cite{Buades2006}
investigate the asymptotic relationship between the Yaroslavsky filter and the Perona-Malik PDE. Gilboa et al. \cite{Gilboa2008} study certain applications of
nonlocal operators to image processing. In \cite{Peyre2008}, Peyr\'e establishes a relationship
between the non-iterative nonlocal filtering schemes and thresholding in adapted
orthogonal basis.
In a more recent paper, Singer et al. \cite{Singer2009}
interpret the Neighborhood filter as a stochastic diffusion process,
explaining in this way the attenuation of high frequencies in the processed images.
In this article, we reformulate the Neighborhood filter in terms of the \emph{decreasing rearrangement} of the initial image, $u$,
which is defined as the inverse of the \emph{distribution function}
$q\in\mathbb{R}\to m_u(q) = \abs{\{\mathbf{x} \in\Omega : u(\mathbf{x}) >q\}}$,
see Section 2 for the precise definition.
Realizing that the structure of level sets of $u$ is invariant through the Neighborhood filter operation as well as through the decreasing rearrangement of
$u$ allows us to rewrite \fer{def.NF} in terms of a one-dimensional integral
expression, see Theorem~\ref{th.equivalence1}.
Although from expression \fer{def.NF} is readily seen that only computation on level lines is needed to perform the filtering, the alternative expression in terms of the decreasing rearrangement offers room for further analysis of the iterative scheme.
Perhaps, the most important consequence of the rearrangement is, apart from
the dimensional reduction, the reinterpretation of the NF as a
\emph{local} algorithm. Thanks to this, we may prove some properties of the Neighborhood filter nonlinear iterative scheme, see \fer{def.GFV}, among which
\begin{itemize}
\item The asymptotic behavior of the NF as a shock filter of the type introduced by \'Alvarez et al. \cite{Alvarez1994}, combined with a contrast loss effect, see Theorem~\ref{th.pde}.
\item The contrast change character of the NF, i.e. the existence of a continuous and
increasing function $g:\mathbb{R}\to\mathbb{R}$ such that $\NF^h (u(\mathbf{x}))=g(u(\mathbf{x}))$, see Corollary~\ref{cor.contrast}.
\end{itemize}
As mentioned above, the most salient advantage of the Neighborhood filter in its rearranged version is speed. If $N$ denotes the image size in pixels, the complexity of other
nonlocal filters such as the \emph{classical} NF, the Bilateral or the NLM filters are of the order $C\times N$,
where $C$ depends on different parameters (window sizes) used in those models. A typical value of $C$ may be of the order $10^5$, for the NLM. However, the complexity of the rearranged version of the NF depends on the number of the initial image intensity levels, which we assume quantized in $Q$ levels, and on some small constant related to the window size, $h$, resulting
in a complexity of the order $c\times Q$, being a typical value of $c$ of the order $10^2$.
Thus, the complexity of the NF is independent of the image size, once the level lines of the
initial image have been identified.
However, denoising quality of the NF is, by far, poorer. The staircaising
effect, always present in algorithms reducing the Total Variation of the initial image, is especially strong for the NF.
In fact, after several iterations, and depending on the window size $h$, the NF output image concentrates most of its pixels mass on few level sets, producing a segmentation-like effect on the initial image.
A quick explanation of the difference between the NF and the Bilateral and NLM filters is that the NF does not retain any local information of the image, diffusing the
intensity values just according to the mass of their corresponding level lines. Thus, a pixel belonging to a level line with large mass will retain its value even if it is isolated in a component of the image with different intensity value.
Due to this, the Neighborhood filter and morphological filters deduced from the topographic map of an image, see for instance the monograph by Caselles and Monasse \cite{Caselles2010}, such as the Grain and the Killer filters, are also different since the latter use the local geometry of level sets connected components in a fundamental manner.
However, they do share some similitudes in their level set based framework, and this could be used for combining both algorithms. For instance, the isolated small regions belonging to a level line with large mass which remain in the filtered image after
the NF application could be removed by morphological procedures, such as
opening and closing operators.
The segmentation-like behavior of the NF may be explained in terms of the relationship between
the inflexion points of the decreasing rearrangement and the local extreme points of the image histogram, $h(q)=-m'_u(q)$. Thus, since the NF behaves asymptotically as a shock filter, and shock filters accumulate mass in inflexion points, we deduce that the NF works as an histogram based segmentation algorithm.
Since for our analysis the Gaussian form of the integral kernel is not important, we consider the following generalization of the Neighborhood filter
\begin{equation}
\label{def.GF}
\NF^h u (\mathbf{x})=\frac{1}{C(\mathbf{x})}\int_\Omega \mathcal{K}_h(u(\mathbf{x})-u(\mathbf{y})) u(\mathbf{y})d\mathbf{y},
\end{equation}
with $C(\mathbf{x})=\int_\Omega \mathcal{K}_h(u(\mathbf{x})-u(\mathbf{y})) d\mathbf{y}$, and $\mathcal{K}_h(\xi)= \mathcal{K}(\xi/h)$. For the moment, we only assume
$\mathcal{K} \in L^1_{loc}(\mathbb{R})$ and $\mathcal{K}\geq0$ to have \fer{def.GF} well defined, although more meaningful conditions will be
stated later.
We consider the following iterative schemes,
for $n\in\mathbb{N}$ (including $n=0$), and $u^{(0)}=u$:
\begin{enumerate}
\item Iteration with fixed kernel (linear operator),
\begin{equation}
\label{def.GFF}
u^{(n+1)} (\mathbf{x})=\frac{1}{C_0(\mathbf{x})}\int_\Omega \mathcal{K}_h(u^{(0)}(\mathbf{x})-u^{(0)}(\mathbf{y})) u^{(n)}(\mathbf{y})d\mathbf{y},
\end{equation}
with $C_0(\mathbf{x})=\int_\Omega \mathcal{K}_h(u^{(0)}(\mathbf{x})-u^{(0)}(\mathbf{y})) d\mathbf{y}$.
\item Iteration with varying kernel (nonlinear operator),
\begin{equation}
\label{def.GFV}
u^{(n+1)} (\mathbf{x})=\frac{1}{C_n(\mathbf{x})}\int_\Omega \mathcal{K}_h(u^{(n)}(\mathbf{x})-u^{(n)}(\mathbf{y})) u^{(n)}(\mathbf{y})d\mathbf{y},
\end{equation}
with $C_n(\mathbf{x})=\int_\Omega \mathcal{K}_h(u^{(n)}(\mathbf{x})-u^{(n)}(\mathbf{y})) d\mathbf{y}$.
\end{enumerate}
Observe that, for both schemes, we have
\begin{equation}
\label{bound.iter}
\nor{u^{(n+1)}}_{L^\infty(\Omega)}\leq \nor{u^{(n)}}_{L^\infty(\Omega)},
\end{equation}
for all $n\in\mathbb{N}$, and therefore the iterations are well defined.
The plan of the article is the following. In Section~2 we introduce the notion
of decreasing rearrangement and establish the equivalence between
\fer{def.GFF} and \fer{def.GFV}
and its corresponding versions under this transformation. In addition, we
show some examples on the relationship between an image and its relative rearrangement and histogram.
In Section~3, we prove some qualitative properties of the nonlinear iterative scheme,
among which its behavior as a shock filter. In Section~4, we describe the
discretization of the continuous model and demonstrate with examples the denoising
capabilities of the NF in comparison with the Bilateral and NLM filters, and
its interpretation as a segmentation-like algorithm. In Section~5 we give our conclusions.
Finally, let us emphasize that the reformulation of the Neighborhood filter
we propose produces the same filtered image than that obtained through the classical
NF. However, there are two important advantages in our approach:
(i) the filtering process is performed through a one-dimensional
integral, reducing largely the algorithm complexity, and (ii) the rearranged version exposes the NF functioning to a deeper mathematical analysis.
\section{Neighborhood filters in terms of the decreasing rearrangement}
Let us denote by $\abs{E}$ the Lebesgue measure of any measurable set $E$.
For a Lebesgue measurable function $u:\Omega\to\mathbb{R}$, the function
$q\in\mathbb{R}\to m_u(q) = \abs{\{\mathbf{x} \in\Omega : u(\mathbf{x}) >q\}}$ is called the \emph{distribution function} corresponding to $u$.
Function $m_u$ is non-increasing and therefore admits a unique generalized inverse, called the
\emph{decreasing rearrangement}. This inverse takes the usual pointwise meaning when
the function $u$ has not flat regions, i.e. when $\abs{\{\mathbf{x} \in\Omega : u(\mathbf{x}) =q\}} =0$ for any $q\in\mathbb{R}$. In general,
the decreasing rearrangement $u_*:[0,\abs{\Omega}]\to\mathbb{R}$ is given by:
\begin{equation*}
u_*(s) =\left\{
\begin{array}{ll}
{\rm ess}\sup \{u(\mathbf{x}): \mathbf{x} \in \Omega \} & \qtext{if }s=0,\\
\inf \{q \in \mathbb{R} : m_u(q) \leq s \}& \qtext{if } s\in (0,\abs{\Omega}),\\
{\rm ess}\inf \{u(\mathbf{x}): \mathbf{x} \in \Omega \} & \qtext{if }s=\abs{\Omega}.
\end{array}\right.
\end{equation*}
Notice that since $u_*$ is non-increasing in $[0,\abs{\Omega}]$, it is continuous but at most a countable subset of
$[0,\abs{\Omega}]$. In particular, it is right-continuous for all $t\in (0,\abs{\Omega}]$.
The notion of rearrangement of a function is classical. We refer the reader to the textbook
\cite{Lieb2001} for the basic definitions and to the monograph \cite{Rakotoson2008} for a deeper insight into the subject.
The following equi-measurability property holds \cite[Corollary 1.1.1]{Rakotoson2008}.
Let $F:\mathbb{R}\to\mathbb{R}_+$ be a Borel function. Then
\begin{equation}
\label{prop.1}
\int_\Omega F(u(\mathbf{y}))d\mathbf{y} = \int_0^{\abs{\Omega}} F(u_* (s))ds.
\end{equation}
In particular, $\nor{u}_{L^p(\Omega)}=\nor{u_*}_{L^p(0,\abs{\Omega})}$ for all $p\in[0,\infty]$.
We shall use the following notation for the level sets of $u$:
\begin{equation}
\label{def.levelsets}
L_t(u)=\{\mathbf{y} \in \Omega : u(\mathbf{y})=u_*(t) \},\qtext{for }t \in [0,\abs{\Omega}].
\end{equation}
The following theorem asserts that we may compute the Neighborhood filters in
the one-dimensional space $[0,\abs{\Omega}]$.
\begin{theorem}
\label{th.equivalence1}
Let $\Omega\subset\mathbb{R}^d$ be an open and bounded set, $d\geq 1$, and $\mathcal{K}:\mathbb{R}\to\mathbb{R}_+$ be a Borel function. Let $u^{(0)}\in L^\infty(\Omega)$ be,
without loss of generality, non-negative and set $v_0 =u^{(0)}_*$.
Then, for $m=0$ (resp. $m=n$), the iterative scheme \fer{def.GFF} (resp \fer{def.GFV})
may be computed as, for $\mathbf{x} \in L_t(u^{(0)})$ and $t\in [0,\abs{\Omega}]$, $u^{(n+1)}(\mathbf{x})= v_{n+1}(t)$, with
\begin{equation}
\label{def.NFstar}
v_{n+1}(t) = \frac{1}{c_m(t)} \int_0^{\abs{\Omega}} \mathcal{K}_h(v_{m}(t)-v_{m}(s))v_{n}(s)ds,
\end{equation}
and $c_m(t)=\int_0^{\abs{\Omega}} \mathcal{K}_h(v_{m}(t)-v_{m}(s))ds$.
In addition, for $n\in \mathbb{N}$,
\begin{equation}
\label{bound.iterstar}
\nor{v_{n+1}}_{L^\infty(0,\abs{\Omega})}\leq \nor{v_n}_{L^\infty(0,\abs{\Omega})}.
\end{equation}
\end{theorem}
\noindent\emph{Proof. }
We start considering the one-step Neighborhood filter defined in \fer{def.NF}.
For each $\mathbf{x}\in \Omega$ such that $u(\mathbf{x}) <\infty$ (i.e. all $\mathbf{x}\in\Omega$ but a subset of zero measure), we consider the Borel function $F:\mathbb{R}\to\mathbb{R}_+$ given by
\begin{equation}
\label{def.F}
F(w)=\mathcal{K}_h(u(\mathbf{x})-w)w\qtext{if }w\geq 0,\quad F(w)=0\qtext{if }w<0.
\end{equation}
Then, \fer{prop.1} implies
\begin{equation*}
\int_\Omega \mathcal{K}_h(u(\mathbf{x})-u(\mathbf{y}))u(\mathbf{y})d\mathbf{y} = \int_0^{\abs{\Omega}} \mathcal{K}_h(u(\mathbf{x})-u_*(s))u_* (s))ds,
\end{equation*}
and
\begin{equation*}
C(\mathbf{x})= \int_0^{\abs{\Omega}} \mathcal{K}_h(u(\mathbf{x})-u_*(s))ds.
\end{equation*}
Since $\mathbf{x} \in L_t(u^{(0)})$,
for some $t\in [0,\abs{\Omega}]$, we have
\begin{align*}
\int_0^{\abs{\Omega}} \mathcal{K}_h(u(\mathbf{x})- & u_*(s)) u_* (s))ds
=\int_0^{\abs{\Omega}} \mathcal{K}_h(u_*(t)-u_*(s))u_* (s))ds,
\end{align*}
and similarly for $C(\mathbf{x})$.
We thus introduce the equivalent formulation to \fer{def.NF} given by, for $t\in [0,\abs{\Omega}]$
\begin{equation}
\label{def.onestep}
\NF^h u(\mathbf{x})=\NF^h u_* (t) \qtext{if}\quad \mathbf{x} \in L_t(u) ,
\end{equation}
with
\begin{equation*}
\NF^h u_* (t) = \frac{1}{c(t)} \int_0^{\abs{\Omega}} \mathcal{K}_h(u_*(t)-u_*(s))u_* (s)ds,
\end{equation*}
and $c(t)=\int_0^{\abs{\Omega}} \mathcal{K}_h(u_*(t)-u_*(s))ds$.
The extension of this transformation to the iterative schemes \fer{def.GFF} and \fer{def.GFV} ($m=0$ or $m=n$, respectively)
is straightforward
due to the invariance of the level sets structure with respect to the Neighborhood filtering, see \fer{def.onestep}.
For both schemes, the step $n=1$ gives
\begin{equation*}
u^{(1)}(\mathbf{x})= v_1(t):=\NF^h u_*^{(0)} (t) ,
\end{equation*}
for $\mathbf{x} \in L_t(u^{(0)})$.
In the case of the scheme with variable kernel \fer{def.GFV}, we may use the same
function $F$ defined in \fer{def.F} to deduce that $u^{(2)}$ is constant in
$L_t(u^{(0)})$ for each $t\in [0,\abs{\Omega}]$. An induction argument allows us to define
\begin{equation*}
u^{(n+1)}(\mathbf{x})= v_{n+1}(t):=\NF^h v_n (t) ,
\end{equation*}
for $n\in\mathbb{N}$, and $\mathbf{x} \in L_t(u^{(0)})$, which is \fer{def.NFstar} with $m=n$.
For the fixed kernel scheme \fer{def.GFF} ($m=0$) we can not use $F$ in the step $n=1$ because
the values of $u^{(0)}$ and $u^{(1)}$ inside the integral may be different on the level set $L_t(u^{(0)})$.
However, these functions are still constant in each level set,
implying the existence of a measurable function $f_1$ such that $f_1(u^{(1)}(\mathbf{x}))=u^{(0)}(\mathbf{x})$
for $\mathbf{x}\in L_t(u^{(0)})$, for $t\in [0,\abs{\Omega}]$. Therefore, we may repeat the above argument with function $F$ replaced
by $F_1(w)=\mathcal{K}_h(u^{(0)}(\mathbf{x})-f_1(w))w$ if $w\geq0$ and $F_1(w)=0$ if $w<0$, thus obtaining
\begin{align*}
u^{(2)}(\mathbf{x})= & \NF^h u^{(1)}(\mathbf{x})
= \frac{1}{C_1(\mathbf{x})}
\int_0^{\abs{\Omega}} \mathcal{K}_h(u^{(0)}(\mathbf{x})-f_1(v_1(s))v_1(s)ds \\
= & \frac{1}{C_1(\mathbf{x})}
\int_0^{\abs{\Omega}} \mathcal{K}_h(u^{(0)}(\mathbf{x})-u^{(0)}_*(s))v_1(s)ds.
\end{align*}
Reasoning in a similar way for $C_1(\mathbf{x})$, and recalling the definition of $v_0$ we obtain for
$\mathbf{x} \in L_t(u^{(0)})$,
\begin{align*}
u^{(2)}(\mathbf{x})=v_2(t):= \frac{1}{c_1(t)}
\int_0^{\abs{\Omega}} \mathcal{K}_h(v_0(t)-v_0(s))v_1(s)ds.
\end{align*}
Then \fer{def.NFstar} for $m=0$. follows from an inductive argument.
Finally, using \fer{bound.iter} and the equi-measurability property \fer{prop.1} we obtain
\fer{bound.iterstar} for all $n\in\mathbb{N}$.
$\Box$
\begin{figure*}[ht]
\centering
\subfigure%
{\includegraphics[width=4cm,height=4cm]{figure1_boat_clean.png}}
\subfigure%
{\includegraphics[width=4cm,height=4cm]{figure1_boat_DR.png}}
\subfigure%
{\includegraphics[width=4cm,height=4cm]{figure1_boat_hist.png}} \\
\subfigure%
{\includegraphics[width=4cm,height=4cm]{figure1_hexagonal_clean.png}}
\subfigure%
{\includegraphics[width=4cm,height=4cm]{figure1_hexagonal_DR.png}}
\subfigure%
{\includegraphics[width=4cm,height=4cm]{figure1_hexagonal_hist.png}} \\
\subfigure%
{\includegraphics[width=4cm,height=4cm]{figure1_sintetica_clean.png}}
\subfigure%
{\includegraphics[width=4cm,height=4cm]{figure1_sintetica_DR.png}}
\subfigure%
{\includegraphics[width=4cm,height=4cm]{figure1_sintetica_hist.png}}
\caption{Test images \emph{Boat}, \emph{Texture} and \emph{Squares}.
Second and third columns show the decreasing rearrangement and the histogram, respectively, of the original images (clean) and their counterparts obtained after the addition of a Gaussian noise with SNR$=10$.}
\label{fig1}
\end{figure*}
\subsection{Examples}
We show three examples of the decreasing rearrangement for clean and noisy images, all of them
quantized in the usual interval $[0,255]$.
We have chosen as test images a natural image, \emph{Boat},
a texture image, \emph{Texture}, and a synthetic image, \emph{Squares}.
The first two images are taken from the data base of the
Signal and Image Processing Institute, University of Southern California
({\tt boat.512.tiff} of Vol.~3 and {\tt 1.5.02.tiff} of Vol.~1,
respectively), see \cite{Collins1998,Kwan1999}, while the third is a synthetic image constructed with four gray levels
($0,~85,~170$ and $255$) and such that its four level sets have the same measure.
This choice is motivated by the bad (resp. good) performance of the NF for
uniform (resp. extreme) distribution of gray levels mass. These distributions are plotted in Fig.~\ref{fig1}.
A gray levels mass uniformly distributed image has a straight line as decreasing rearrangement, or a constant, as histogram. In Fig.~\ref{fig1}, we observe that the decreasing rearrangement of the Boat, is closer to a straight line than that of the Texture, which is still \emph{continuous}. The choice of the synthetic image is motivated by its extreme behavior: a piece-wise constant decreasing rearrangement.
We have added a Gaussian white noise of $\text{SNR}=10$ to the test images
according to the noise measure $\text{SNR}=\sigma(u)/\sigma(\nu)$,
where $\sigma$ is the empirical standard deviation, $u$ is the original image, and $\nu$ is the noise.
We may observe in Fig.~\ref{fig1} that the main consequence of noise addition on the decreasing rearrangement is its smoothening towards a straight line. Of course, the effect is stronger in
far from uniformly distributed images, like the Squares.
Finally, observe the connection between points of local maximum or minimum
for the histogram and inflexion points for the decreasing rearrangement.
This is the base for justifying the use of the NF as a segmentation algorithm.
\section{Properties of the nonlinear varying-kernel iterative scheme}
For the differential analysis we carry out in this section we have to assume regularity properties on the decreasing rearrangement of the given image.
In general, if $u\in W^{1,q}(\Omega)$, with $q>d$, and $\Omega$ is open, bounded and connected, then $u_* \in W^{1,p}(0,\abs{\Omega})$ for any $p<2$, see
\cite[Th. 3.3.2]{Rakotoson2008}.
However, not much more than this is expected, and even for
$C^\infty (\Omega)$ functions, their decreasing rearrangement may be non globally
differentiable, see Fig.~\ref{rem.fig}.
The following theorem asserts that the monotonicity property of the decreasing rearrangement of the initial image
is conserved along all the iterations.
In addition, for some type of kernels among which the Gaussian is included,
the measure of flat regions does not change in the iterative procedure. We also obtain a
condition in terms of the window size, $h$, ensuring that the steady state is a constant.
\begin{theorem}
\label{th.derivada}
Let $v_0= u^{(0)}_* \in W^{1,p}(0,\abs{\Omega})$ for some $p\geq 1$.
Let $\mathcal{K}\in W^{1,\infty}_{\text{loc}}(\mathbb{R})$, with $\mathcal{K} (\xi)\geq 0$ for all $\xi\in\mathbb{R}$,
and $\mathcal{K}(\xi) > 0$ for all $\xi$ in a neighborhood of $0$.
Let $v_{n+1}$ be given by the iterative scheme \fer{def.NFstar}, for $n\in\mathbb{N}$ and $m=n$.
Then we have, for all $n\in\mathbb{N}$,
\begin{enumerate}
\item[(i)] $v_{n+1}\in W^{1,p}(0,\abs{\Omega})$, and if $v'_0(t)=0$ then $v'_{n+1}(t)=0$.
\item[(ii)] Let $R_1:=(\xi_1-\xi_2)\big(\mathcal{K}'(\xi -\xi_1)\mathcal{K}(\xi -\xi_2)-\mathcal{K}'(\xi -\xi_2)\mathcal{K}(\xi -\xi_1)\big)$ and
assume
\begin{equation}
\label{cond.k}
R_1\geq 0 \text{ for all }\xi,\xi_1,\xi_2 \in\mathbb{R}.
\end{equation}
Then $v'_{n+1} \leq 0$ a.e. in $(0,\abs{\Omega})$. In addition, if $v'_0(t)<0$ and the inequality in
\fer{cond.k} is strict then $v'_{n+1}(t)<0$.
\item[(iii)] Let $\phi:\mathbb{R}_+\to\mathbb{R}_+$ be a continuous function
of $h$ such that $\lim_{h\to \infty} \phi(h)=0$.
Assume that $\mathcal{K} >0$ in $\mathbb{R}$, and let
\begin{equation*}
\label{def.R2}
R_2= (\xi_1-\xi_2)\Big(\frac{\mathcal{K}'(\xi -\xi_1)}{\mathcal{K}(\xi -\xi_1)}-\frac{\mathcal{K}'(\xi -\xi_2)}{\mathcal{K}(\xi -\xi_2)}\Big).
\end{equation*}
Assume
\begin{equation}
\label{cond.k2}
R_2 \leq \phi(h)(\xi_1-\xi_2)^q \quad \forall h\geq0,\quad \forall \xi,\xi_1,\xi_2 \in\mathbb{R},
\end{equation}
for some $q\geq 1$.
Then, if $h$ is large enough the sequence $v_n$ converges uniformly to a constant.
\end{enumerate}
\end{theorem}
\begin{figure}[t]
\centering
\subfigure
{\includegraphics[width=5cm,height=4cm]{ejemplo_reord_regularidad1.png}}
\hspace{1cm}
\subfigure
{\includegraphics[width=5cm,height=4cm]{ejemplo_reord_regularidad2.png}}
\caption{{\small Graphic of function $f(x)=x\sin(10x)$ (top panel) and that of its decreasing rearrangement, $f_*$ (bottom panel).
Although $f\in C^\infty (-\pi,\pi)$, $f_*$ is not even once continuously differentiable in $(0,2\pi)$. }}
\label{rem.fig}
\end{figure}
\begin{remark}
With the additional assumption of
$\mathcal{K}$ being symmetric,
condition \fer{cond.k} imposes $\mathcal{K}(\xi_2) \geq \mathcal{K}(\xi_1)$
if $\abs{\xi_1}\geq \abs{\xi_2}$, that is, it must be a decaying kernel.
Observe that the Gaussian kernel $\mathcal{K}(s)=\text{e}^{-s^2}$ satisfies all the assumptions
of Theorem~\ref{th.derivada}. In particular, condition \fer{cond.k2} is satisfied for $\phi(h)=1/h^2$ and $q=2$.
Functions $R_1$ and $R_2$ are related in the following way: $R_1=\mathcal{K}(\xi-\xi_1)\mathcal{K}(\xi-\xi_2)R_2$. Thus, if
$\mathcal{K}>0$ in $\mathbb{R}$ then condition \fer{cond.k} is equivalent to $R_2 \geq 0$.
\end{remark}
\emph{Proof of Theorem~\ref{th.derivada}.}
Differentiating \fer{def.NFstar}, for $m=n$, with respect to $t$ we obtain
\begin{align}
v'_{n+1}(t) = & \label{deri}
\frac{1}{c_n^2(t)}v'_{n} (t)
\Big(c_n(t) \int_0^{\abs{\Omega}}\mathcal{K}_h'(v_{n}(t)-v_{n}(s)) v_{n}(s) ds \\
& - \int_0^{\abs{\Omega}} \mathcal{K}_h(v_{n}(t)-v_{n}(s))v_{n}(s)ds
\int_0^{\abs{\Omega}}\mathcal{K}_h'(v_{n}(t)-v_{n}(s)) ds \Big). \nonumber
\end{align}
Let us consider the case $n=0$. By assumption, $v_0 \in W^{1,p}(0,\abs{\Omega})$.
On one hand, since $\mathcal{K}$ is continuous and positive in a neighborhood of $0$, and $v_0'$ is bounded in $L^1(0,\abs{\Omega})$, we deduce
$c_0(t) \geq \hat c \int_{\mathbb{R}} \mathcal{K}_h(\xi) d\xi$, for some positive constant $\hat c$.
On the other hand, the regularity assumed on $v_0$ also implies $v_0\in L^\infty(0,\abs{\Omega})$. Using these properties together with
the Lipschitz continuity of $\mathcal{K}$ we obtain that also $v_1\in W^{1,p}(0,\abs{\Omega})$. Then we proceed recursively to deduce $v_{n+1}\in W^{1,p}(0,\abs{\Omega})$
for all $n\in\mathbb{N}$, and thus (i) follows.
To check the sign of $v'_{n+1}$, let us simplify the notation introducing, for fixed $t\in(0,\abs{\Omega})$,
\begin{align*}
& f(s)=\mathcal{K}_h'(v_{n}(t)-v_{n}(s)),\quad
g(s)=v_{n}(s),\\
& h(s)=\mathcal{K}_h(v_{n}(t)-v_{n}(s)).
\end{align*}
Then $v'_{n+1} \leq 0$ follows from \fer{deri} if we prove
\begin{align*}
\int_0^{\abs{\Omega}} f(s)ds &\int_0^{\abs{\Omega}} g(z)h(z)dz \leq
\int_0^{\abs{\Omega}} h(s)ds \int_0^{\abs{\Omega}} f(z)g(z)dz,
\end{align*}
which is equivalent to $\alpha\geq 0$, with
\begin{equation}
\label{def.alpha}
\alpha=\int_0^{\abs{\Omega}}\int_0^{\abs{\Omega}}g(z)\big(f(z)h(s)-f(s)h(z)\big)dsdz.
\end{equation}
Interchanging the dummy variables, we get after addition of the corresponding
identities
\begin{equation*}
\label{alpha.2}
2 \alpha=\int_0^{\abs{\Omega}}\int_0^{\abs{\Omega}}(g(z)-g(s))\big(f(z)h(s)-f(s)h(z)\big)dsdz.
\end{equation*}
We then use assumption \fer{cond.k} to deduce $\alpha\geq 0$.
The second part of (ii) follows from similar arguments.
We, finally, prove (iii). Using \fer{deri} and the definition \fer{def.alpha} we have
\begin{equation*}
v_{n+1}'(t)=\frac{\alpha}{c_n^2(t)} v_{n}'(t).
\end{equation*}
Since $\mathcal{K}>0$, we may rewrite $\alpha$ as
\begin{equation*}
\alpha =\int_0^{\abs{\Omega}}\int_0^{\abs{\Omega}}g(z)h(s)h(z) \Big(\frac{f(z)}{h(z)}-\frac{f(s)}{h(s)}\Big)dsdz.
\end{equation*}
Using again the interchange of dummy variables and condition \fer{cond.k2} we obtain
\begin{equation*}
\frac{\alpha}{c_n^2(t)} \leq \phi(h) \nor{v_n}_{L^\infty}^q.
\end{equation*}
Therefore, for $h$ large enough (depending on the shape of function $\phi$) we have $\abs{v_{n+1}'(t)}\leq c \abs{v_{n}'(t)}$, with
$c<1$. The result follows. $\Box$
\begin{corollary}
\label{cor.reg}
Let $v_0= u^{(0)}_* \in C^m([0,\abs{\Omega}])$ and $\mathcal{K}\in C^m(\mathbb{R})$, with $\mathcal{K} >0 $. Then
$v_{n+1} \in C^m([0,\abs{\Omega}])$ for all $n\in\mathbb{N}$.
\end{corollary}
\noindent\emph{Proof. } From formula \fer{deri} we see that the $m-$order derivative of $v_{n+1}$
is given in terms of a quotient with a non-vanishing denominator and a numerator expressed as a composition of continuous functions, given in terms of the derivatives of order up to $m$ of $v_n$ and $\mathcal{K}$. $\Box$
Although an easy consequence of Theorems~\ref{th.equivalence1} and \ref{th.derivada}, the following result neatly exposes the functioning of the Neighborhood filter. We recall that
the mapping $g:\mathbb{R}\to\mathbb{R}$ is a \emph{contrast change} if $g$ is strictly increasing and continuous.
\begin{corollary}
\label{cor.contrast}
Assume the conditions of Theorem~\ref{th.derivada}, with the exception of condition (iii),
and suppose that $v_0'(t)<0$ for all $t\in(0,\abs{\Omega})$.
Then, there exists a contrast change, $g$ such that $u^{(n+1)}=g(u^{(0)})$ in $\Omega$, where $u^{(n+1)}$ is the $(n+1)-$th iteration of the nonlinear iterated Neighborhood filter \fer{def.GFV}.
\end{corollary}
\noindent\emph{Proof. }
Since, by assumption, $v_0\in W^{1,1}(0,\abs{\Omega})\subset C([0,\abs{\Omega}])$, and $v_0'(t)<0$ for $t\in(0,\abs{\Omega})$, the corresponding distribution function of $v_0$, $m_{v_0}$
is continuous and invertible in $(0,\abs{\Omega})$ and, actually, it is the inverse of $v_0$.
Notice that $m_{v_0}$ coincides with the distribution function of
$ u^{(0)}_*$, since $v_0= u^{(0)}_*$.
We define the function
\[
g(q)=v_{n+1}(m_{v_0}(q)), \qtext{for } q\in [\min{v_0},\max{v_0}].
\]
By Theorem~\ref{th.derivada}, $v_{n+1}\in W^{1,1}(0,\abs{\Omega})\subset C([0,\abs{\Omega}])$, and $v_{n+1}'(t)<0$, implying that $g$ is continuous and increasing (composition of two continuous and decreasing functions).
According to Theorem~\ref{th.equivalence1}, the structure of level sets is invariant under the NF, i.e. the level lines of $u^{(n+1)}$ are the same as those
of $u^{(0)}$. In addition, the values of $u^{(n+1)}$ on the level lines are given by \fer{def.NFstar}. Therefore, for
$\mathbf{x} \in L_s(u^{(0)})$, for all $s\in[0,\abs{\Omega}]$, we have
\[
g(u^{(0)}(\mathbf{x}))=g(v_0(s))=v_{n+1}(s)=u^{(n+1)}(\mathbf{x}).
\]
$\Box$
In the following theorem we establish a correspondence between the nonlocal diffusion
scheme \fer{def.NFstar} and local diffusion. This is a fundamental ingredient for the properties deduced later, which can not be
directly deduced from the N-dimensional model.
Indeed, since $v_n$ is non-increasing for all $n\in\mathbb{N}$ (Theorem~\ref{th.derivada}), we have that the values selected by the nonlocal kernel $\mathcal{K}_h(v_n(t)-v_n(s))$ are related to the independent variables through, for instance, Taylor's expansion.
For example, if $u_*$ is smooth, we may approximate
\begin{equation*}
\label{approxi}
\int_\Omega \text{e}^{-\frac{(v_n(t)-v_n(s))^2}{h^2}}v_n(s)ds \approx
\int_\Omega \text{e}^{-\frac{((t- s)v_n'(t))^2}{h^2}}v_n(s)ds.
\end{equation*}
Therefore, the size of the support of the cut-off function approximated by the Gaussian is related to
the size of $v_n'$. If $v_n'$ is large, the size is small and the diffusion only occurs in a small interval
around $v_n(t)$. The opposite effect holds if $v_n'$ is small.
Although more general assumptions on $\mathcal{K}$ may be prescribed, see Remark~\ref{remth.pde}, we estate this result for the Gaussian kernel, for clarity. We also ask for further regularity on $u_*$.
\begin{theorem}
\label{th.pde}
Let $v_0=u_* \in C^3([0,\abs{\Omega}])$ be such that $v_0'<0$ in $[0,\abs{\Omega}]$. Let
$ \mathcal{K}(\xi)=\text{e}^{-\xi^2}$.
Then, for all $t\in(0,\abs{\Omega})$, there exist positive constants $\alpha_1,\alpha_2$ independent of $h$ such that
\begin{align}
\label{app.NFstar}
v_{n+1}(t) = v_{n}(t)&+\alpha_1 \tilde k_h(t)v_{n} ' (t) \big( h+O(h^{3/2}) \big)
- \alpha_2 \frac{v_{n} '' (t)}{(v_{n} ' (t))^2} h^2 + O(h^{5/2}),
\end{align}
with
\begin{equation}
\label{def.ktilde}
\tilde k_h(t)= \frac{\mathcal{K}_h(v_n(t)-v_n(\abs{\Omega}))}{v_n'(\abs{\Omega})}-
\frac{\mathcal{K}_h(v_n(t)-v_n(0))}{v_n'(0)},
\end{equation}
and with $\alpha_1\approx 1/\sqrt{\pi}$, and $\alpha_2\approx 1$.
\end{theorem}
There are two interesting effects captured by \fer{app.NFstar}:
\begin{enumerate}
\item The border effect (loss of contrast). Function $\tilde k_h$ is \emph{active} only when $t$ is close to the boundaries, $t\approx 0$ and $t\approx \abs{\Omega}$. For
$t \approx 0$ the term $\tilde k_h(t)v_{n} ' (t) <0 $ contributes to the decrease of
the largest values of $v_{n+1}$ while for $t\approx \abs{\Omega}$ the opposite effect takes place. Therefore,
this term tends to flatten $v_{n+1}$, i.e. induces a loss of contrast.
\item The term $-\frac{h^2}{2} \frac{v_{n} '' (t)}{(v_{n} ' (t))^2}$ is anti-diffusive, inducing large
gradients of the iterated functions in a neighborhood of the inflexion points. In this sense, the scheme \fer{app.NFstar}
is related to the shock filter introduced by Alvarez and Mazorra \cite{Alvarez1994}
\begin{equation}
\label{alvarez}
u_t+F(G_t u_{xx},G_t u_{x})u_x=0,
\end{equation}
where $G_t$ is a smoothing kernel and function $F$ satisfies $F(p,q)pq\geq 0$ for any $p,q\in\mathbb{R}$. Indeed, neglecting the
border and the lower order terms, and defining $F(p,q)=p/q^3$, we obtain from \fer{app.NFstar}
\begin{equation*}
v_{n+1}(t) - v_{n}(t)+ \alpha_2 h^2 F(v_{n} '' (t),v_{n} ' (t)) v_{n} ' (t)=0 ,
\end{equation*}
which may be regarded as a time discretization of \fer{alvarez}.
\end{enumerate}
\emph{Proof of Theorem~\ref{th.pde}. }
We may rewrite the iterative scheme \fer{def.NFstar}, for $n\in\mathbb{N}$ and $m=n$
as
\begin{align}
\label{th2.1}
v_{n+1}&(t)-v_n(t)= \\
& \frac{1}{c_n(t)}\int_0^{\abs{\Omega}} \mathcal{K}_h(v_n(t)-v_n(s))(v_n(s)-v_n(t))ds.\nonumber
\end{align}
Due to (3) of Theorem~\ref{th.derivada} we have $v_{n}'<0$ in $(0,\abs{\Omega})$, and due to \fer{bound.iterstar}, $v_n(0,\abs{\Omega})\subset v_0(0,\abs{\Omega})$.
Let us denote the inverse of $v_{n}$ by $v_n^{-1}$.
Using the change of variable $s=v_n^{-1}(q)$ and writing
$t=v_n^{-1}(z)$, we obtain from \fer{th2.1}
\begin{equation}
\label{th2.3}
v_{n+1}(t)-v_n(t)=\frac{I_1(z)}{I_2(z)},
\end{equation}
with
\begin{align*}
& I_1(z)= \int_{v_n(\abs{\Omega})}^{v_n(0)} \mathcal{K}_h(z-q)(q-z)\frac{dq}{v_{n} ' (v_n^{-1}(q))},\\
& I_2(z)=\int_{v_n(\abs{\Omega})}^{v_n(0)} \mathcal{K}_h(z-q)\frac{dq}{v_{n} ' (v_n^{-1}(q))}.
\end{align*}
Using the explicit form of $\mathcal{K}$ and integrating by parts, we obtain
\begin{align}
\label{th2.i1}
I_1(z)=& \frac{h^2}{2}\Big( \tilde k_h (v_n^{-1}(z))
-\int_{v_n(\abs{\Omega})}^{v_n(0)} \mathcal{K}_h(z-q)\frac{v_{n} '' (v_n^{-1}(q))}{(v_{n} ' (v_n^{-1}(q)))^3}dq\Big),
\end{align}
with $\tilde k_h$ given by \fer{def.ktilde}.
By assumption, functions
\begin{equation*}
f(q)=\frac{v_{n} '' (v_n^{-1}(q))}{(v_{n} ' (v_n^{-1}(q)))^3} \qtext{and}\quad g(q)=\frac{1}{v_{n} ' (v_n^{-1}(q))}
\end{equation*}
are bounded in $[v_n(\abs{\Omega}),v_n(0)]$ and by Corollary~\ref{cor.reg} they are also continuously differentiable in $(v_n(\abs{\Omega}),v_n(0))$.
Consider the interval $J_h=\{q: \abs{z-q}<\sqrt{h}\}$.
By well known properties of the Gaussian kernel, we have
\begin{equation}
\label{gauss.1}
\kappa(h): = \int_{J_h} \mathcal{K}_h(z-q) dq < \int_\mathbb{R} \mathcal{K}_h(q) dq = h\sqrt{\pi},
\end{equation}
and
\begin{equation}
\label{gauss.2}
\mathcal{K}_h(z-q)\leq \text{e}^{-1/h} \quad\text{if}\quad q\in J_h^C=\{q:\abs{z-q}\geq\sqrt{h}\}.
\end{equation}
In particular, from \fer{gauss.2} we get
\begin{equation}
\label{th2.4}
\left| \int_{J_h^C} \mathcal{K}_h(z-q)f(q)dq \right| < O(h^\alpha)\qtext{for any }\alpha >0.
\end{equation}
Taylor's formula implies
\begin{align*}
\int_{v_n(\abs{\Omega})}^{v_n(0)} \mathcal{K}_h& (z-q) f(q)dq
= \int_{J_h} \mathcal{K}_h(z-q) (f(z) +O(\sqrt{h})) dq
+ \int_{J_h^C} \mathcal{K}_h(z-q)f(q)dq .
\end{align*}
Therefore, from \fer{th2.i1}, \fer{gauss.1} and \fer{th2.4} we deduce
\begin{align*}
I_1(z)=\frac{h^2}{2} \Big( \tilde k(v_n^{-1}(z)) -
\frac{v_{n} '' (v_n^{-1}(z))}{(v_{n} ' (v_n^{-1}(z)))^3} \kappa(h) + O(h^{3/2})\Big).
\end{align*}
Similarly,
\begin{align*}
I_2(z)& =\int_{v_n(\abs{\Omega})}^{v_n(0)} \mathcal{K}_h(z-q)g(q)dq
= \int_{J_h} \mathcal{K}_h(z-q) (g(z)+O(\sqrt{h})) dq
+ \int_{J_h^C} \mathcal{K}_h(z-q)g(q)dq \\
& =\frac{1}{v_{n} ' (v_n^{-1}(z))} \kappa(h) + O(h^{3/2}).
\end{align*}
Then,
the result follows from \fer{th2.3} substituting $z$ by $v_n(t)$. $\Box$
\begin{remark}
\label{remth.pde}
Theorem~\ref{th.pde} may be extended to Lipschitz continuous decaying kernels satisfying
the growth condition
\begin{equation}
\label{ext.kernel}
K(s)\leq \frac{k_0}{1+\abs{s}^p},\qtext{for some }p>1.
\end{equation}
In such case, the higher order terms in formula \fer{app.NFstar} must be replaced by
$O(h^{\alpha+1})$ and $O(h^{\alpha +2})$, respectively, with $\alpha=(p-1)/(p+1)$, and
$\alpha_1,~\alpha_2$ are just some positive constants. In addition, function $\tilde k_h$
given by \fer{def.ktilde} is replaced by some continuous function related to a primitive of $sK_h(s)$, and still inducing the border effect commented after the statement of the theorem. This primitive plays the same role as $-h^2K_h(s)/2$ (the primitive when $K$ is a Gaussian kernel) in formulas \fer{th2.i1},
\fer{gauss.1} and \fer{gauss.2}, from where condition \fer{ext.kernel} arises.
\end{remark}
\section{Discretization and numerical examples}
For computing the iterated Neighborhood filter \fer{def.GFV} of a function through its
decreasing rearrangement version \fer{def.NFstar}, we assume that the initial image, $u^{(0)}$ is quantized in some range, e. g. $[0,255]$, and compute its decreasing rearrangement $v_0=u^{(0)}_*$ as the inverse of the distribution function $m_{u^{(0)}}$.
Then, the iterations are performed by computing the integrals involved in the filter by a simple middle point formula, i.e. by assuming a constant-wise interpolation of the discrete image.
Only two parameters must be fixed in advance, the \emph{length} of the kernel window, $h$, and a tolerance for the stopping criterium or, alternatively, the number of filtering iterations.
When the iterations are stopped at iteration, say, $n+1$, we recover the output image by using the formula provided in Theorem~\ref{th.equivalence1},
\[
u^{(n+1)}(\mathbf{x})= v_{n+1}(t) \qtext{for }\mathbf{x} \in L_t(u^{(0)})\qtext{and } t\in [0,\abs{\Omega}],
\]
where $L_t(u^{(0)})$ stands for the level sets of $u^{(0)}$, see \fer{def.levelsets}.
Recall that the level sets structure of $u^{(n)}$, for $n=0,1,\ldots,$ is invariant.
We used a stopping criterium based on the variational approach of the NF given by
Kindermann et al. \cite{Kindermann2005}. In particular, the authors formally show that the
critical points of the functional
\begin{equation*}
J(u)=\int_{\Omega\times\Omega} g \Big(\frac{(u(\mathbf{x})-u(\mathbf{y}))^2}{h^2}\Big)d\mathbf{x} d\mathbf{y},
\end{equation*}
for $g(s)=\int_0^s\mathcal{K}_h(\sqrt{t}) dt$ , coincide with the fixed points of the Neighborhood filter.
The gradient descent scheme associated to the minimization of $J$ is just
the iterated Neighborhood filter \fer{def.GFV}, and thus the relative difference of
the decreasing sequence $J(u^{(n+1)})$ between successive iterations may be used as a stopping criterium. In fact, using
the equi-measurability property \fer{prop.1} we readily deduce
\begin{equation*}
J(u)=J_*(u_*):=\int_0^{\abs{\Omega}}\int_0^{\abs{\Omega}} g \Big(\frac{(u_*(s)-u_*(t))^2}{h^2}\Big)ds dt,
\end{equation*}
which is the actual form of the functional we use for the stopping criterium.
Let us mention that in \cite{Kindermann2005} the authors show that the functional
$J$ is not convex, in general, and therefore the existence and uniqueness of a global minimum for $J$ may not be deduced from the standard theory.
Finally, let us stress that in discrete computations the analytical results obtained in
Section~3 are not always observed.
The reason is, of course, that some of the assumptions are not fulfilled
in the discrete framework. Importantly, those referring to the unbounded
support of the kernel $\mathcal{K}$, or to the regularity of $u_*$.
For example, for the Gaussian kernel used in our numerical experiments,
it is proven in Theorem~\ref{th.derivada} that if $v_0'(t)<0$ then $v_n'(t)<0$ for all $n$. However, as it may be seen, for instance, in Fig.~\ref{fig2} (fourth row, second column), this property is
violated in the discrete framework. Nevertheless, the weaker result $v_n'(t)\leq 0$ is always observed in the experiments
\subsection{ Numerical examples for denoising}
In the first set of experiments we used the Neighborhood filter for denoising porpouses,
and compare it with other related filters: the Bilateral filter,
\begin{equation*}
\BF^{h,\rho} u (\mathbf{x})=\frac{1}{C(\mathbf{x})}\int_\Omega \textrm{e}^{-\frac{\abs{u(\mathbf{x})-u(\mathbf{y})}^2 }{h^2}}\textrm{e}^{-\frac{\abs{\mathbf{x}-\mathbf{y}}^2 }{\rho^2}} u(\mathbf{y})d\mathbf{y},
\end{equation*}
where $h$ and $\rho$ are positive constants, and \[C(\mathbf{x})=\int_\Omega \exp\left(-\abs{u(\mathbf{x})-u(\mathbf{y})}^2) h^{-2}\right)\exp\left(-\abs{\mathbf{x}-\mathbf{y}}^2\rho^{-2}\right) d\mathbf{y}, \]
and the Nonlocal Means filter,
\begin{equation*}
\NL^{h,\rho} u (\mathbf{x})=\frac{1}{C(\mathbf{x})}\int_\Omega \textrm{e}^{-\frac{G_\rho * \abs{u(\mathbf{x}+\cdot)-u(\mathbf{y}+\cdot)}^2 (0)}{h^2}} u(\mathbf{y})d\mathbf{y},
\end{equation*}
where $h>0$, $G_\rho$ is a Gaussian kernel of standard deviation $\rho>0$ and \[C(\mathbf{x})=\int_\Omega \exp\left(-G_\rho * \abs{u(\mathbf{x}+\cdot)-u(\mathbf{y}+\cdot)}^2 (0)) h^{-2}\right) d\mathbf{y}.\]
Since the usual version of the NF, given by \fer{def.GFV}, and the version introduced in this article, expressed through the decreasing rearrangement by \fer{def.NFstar}, are equivalent, there is no need of comparison between them.
The denoising properties of these three filters are well known, and a thoroughfull comparison among them (and among other filters) is given in \cite{Buades2010}. Here, we are not so interested in deciding which is the best performing denoising algorithm than in analyzing their behavior with respect to the histogram and the decreasing rearrangement redistributions.
We applied the filters on the test images given in the Introduction, see Fig.~\ref{fig1},
corrupted with an additive Gaussian white noise of $\text{SNR}=10$,
according to the noise measure $\text{SNR}=\sigma(u)/\sigma(\nu)$,
where $\sigma$ is the empirical standard deviation, $u$ is the original image, and $\nu$ is the noise.
In Figs.~\ref{fig2} to \ref{fig4} we show the results of applying these filters to the Boat, the Texture and the Squares images. The columns correspond to: noisy image, Neighborhood filter, Nonlocal means filter, and Bilateral filter. The rows correspond to: image, detail of the image, intensity histograms of noisy and denoised images, decreasing rearrangements of noisy and denoised images, level curves of image details showed in row 2.
Although the Bilateral and the Nonlocal Means filters are applied only once, their execution time is always much larger than that of the iterated Neighborhood filter,
for which we used the stopping criterium
\[
\frac{\abs{J_*(v_{n+1}) - J_*(v_{n})}}{\abs{J_*(v_{n})}}<10^{-5},
\]
producing between eight iterations, for the Squares image, and twenty iterations, for
the Texture image. We used the same parameter values for $h$ and $\rho$ in all the experiments.
As expected, the best visual result for the natural image is obtained with the Nonlocal Means filter: smoother and with a lower staircaising effect than the others. It is interesting to notice how the absence of local information in the Neighborhood filter
produces regions with rapid intensity value changes, for instance in the clouds of
the image. The smoothing effect of the local terms in the Bilateral and the NLM filters prevent the formation of this artifact.
A partial explanation of the worse behavior of the NF may be found in the corresponding plots for the histograms and the decreasing rearrangements. While the Bilateral and the NLM filters keep almost unchanged the gray intensity structure of the pixel mass, the NF concentrates most of the mass in few and disconnected values which, in general, is an undesired effect in natural images.
Finally, observe that all the filters produce a level lines shortening, notably the NLM filter.
For the Texture image, similar conclusions may be deduced. In this case, the level lines shortening is specially intense for the NF. The area between the circles is
\emph{cleaned} to one single intensity value, around 225. We may check in the corresponding histograms the large difference between the mass assigned to this value
in the different filters. This is a first clue in the consideration of the NF as a
segmentation-like filter.
For the synthetic image Squares, the result of applying the NF is almost a perfect image recovery, while the Bilateral and the NLM filters keep always some noise due to
the local diffusion. The spatial smoothening effect of the latter work against
denoising, for this image.
Let us finally point out to the border effects mentioned after Theorem~\ref{th.pde}, involving formula \fer{app.NFstar}, and related to the contrast loss induced by the NF. They
are clearly visualized in the plots of the decreasing rearrangement of these images.
Also the anti-diffusive behavior of the algorithm, captured by the second order term of formula
\fer{app.NFstar} is observed: concave
regions induce increase on the iterate while convex regions induce decrease. The result is a steeper slope around the inflexion points at each iteration.
\subsection{ Numerical examples for segmentation}
The intensity histogram, $h_u$, of an image $u$ is defined by the measure of its level sets
\begin{equation*}
h_u(q)=\abs{\{\mathbf{y} \in \Omega : u(\mathbf{y})=q \}}, \qtext{for }q\in [\min u, \max u].
\end{equation*}
We therefore have the following relationship between the histogram and the distribution function of $u$,
\begin{equation*}
m_u(q)=\int_q^{\max u} h(s)ds.
\end{equation*}
In particular, under regularity assumptions, critical points of the histogram coincides with inflexion points of the distribution function and, hence, of the decreasing rearrangement.
Observe that histogram critical points detection is the base for some segmentation algorithms, see for instance
\cite{Tobias2002,Chang2006,Nath2011,Qin2011}. Due to the discrete nature of
computations, finding the maxima of the histogram from where initiating
a segmentation procedure is a challenging task.
At this respect, the NF may be seen
as a way of detecting histogram maxima, i.e.
inflexion points of the decreasing rearrangement, which
produces an \emph{automatic} segmentation with the only tunning
of the the window size controlled by $h$.
To demonstrate this capability, we applied the NF to MRI brain segmentation.
We used a phantom brain from the Simulated Brain Database \cite{brainweb}
with a $9\%$ of additive Riccian noise.
In Fig.~\ref{fig5} we show an axial slice of the volume (initial image) an the
corresponding segmentation in four, three and two regions reached by setting $h=17,~20,~50$, respectively. The contour lines and the decreasing rearrangement
are shown too.
In Fig.~\ref{fig6} we show the masks of the segmented regions corresponding to $h=17,~20$.
In Fig.~\ref{fig7} we show the grey-white matter segmentation performed with the NF and with other
standard packages: Freesurfer \cite{freesurfer}, FSL \cite{fsl} and SPM8 \cite{spm8}.
The Dice coincidence coefficient is computed for all the algorithms, see Table~\ref{tab:label}, showing a good
performance of the NF in relation to the more sophisticated algorithms implemented in
the mentioned packages. The Dice coefficient is one if a perfect match to the ground truth
is attained. Zero, on the contrary.
Although in Fig.~\ref{fig7} we have shown the results for one slice, the NF is applied directly to the whole volume, meaning that the dimension reduction is from a three
dimensional space (the space of voxels) to a one dimensional space (the space of level lines measures). Thus, the time execution of the NF is several orders of magnitude lower than the others (a standard volume takes few seconds in a standard laptop).
However, this is no more than a toy example, from where general conclusions can not be inferred.
\begin{table}[!ht]
\centering
\caption{
\bf{Comparison among several algorithms}}
\begin{tabular}{|c||c|c|c|c||c|c|c|c|}
\hline\hline
& \multicolumn{4}{|c||}{Dice coefficient}
\\%\tabularnewline
\hline
& Freesurfer & FSL & SPM & NF \\\hline
white & 0.9490 & 0.9435 & 0.9468 & 0.9563 \\
grey & 0.8509 & 0.8599 & 0.8835 & 0.8797 \\
\hline\hline
\end{tabular}
\label{tab:label}
\end{table}
\newgeometry{textwidth=18cm}
\begin{figure}[ht]
\centering
{\includegraphics[width=4cm,height=4cm]{exp1_boat_noise_NF.png}}
{\includegraphics[width=4cm,height=4cm]{exp1_boat_final_NF.png}}
{\includegraphics[width=4cm,height=4cm]{exp1_boat_final_NLM.png}}
{\includegraphics[width=4cm,height=4cm]{exp1_boat_final_B.png}}\\
{\includegraphics[width=4cm,height=4cm]{exp1_boat_detail_noise_NF.png}}
{\includegraphics[width=4cm,height=4cm]{exp1_boat_detail_final_NF.png}}
{\includegraphics[width=4cm,height=4cm]{exp1_boat_detail_final_NLM.png}}
{\includegraphics[width=4cm,height=4cm]{exp1_boat_detail_final_B.png}}\\
\hspace{4cm}
{\includegraphics[width=4cm,height=4cm]{exp1_boat_hist_NF.png}}
{\includegraphics[width=4cm,height=4cm]{exp1_boat_hist_NLM.png}}
{\includegraphics[width=4cm,height=4cm]{exp1_boat_hist_B.png}}\\
\hspace{4cm}
{\includegraphics[width=4cm,height=4cm]{exp1_boat_DR_NF.png}}
{\includegraphics[width=4cm,height=4cm]{exp1_boat_DR_NLM.png}}
{\includegraphics[width=4cm,height=4cm]{exp1_boat_DR_B.png}} \\
{\includegraphics[width=4cm,height=4cm]{exp1_boat_contour_noise_NF.png}}
{\includegraphics[width=4cm,height=4cm]{exp1_boat_contour_final_NF.png}}
{\includegraphics[width=4cm,height=4cm]{exp1_boat_contour_final_NLM.png}}
{\includegraphics[width=4cm,height=4cm]{exp1_boat_contour_final_B.png}} \\
\caption{{ Denoising experiment. Columns: noisy image, Neighborhood, NLM, and Bilateral filters.
Rows: noisy image, detail of the image, histograms, decreasing rearrangements, and
level curves of image details shown in row 2. }}
\label{fig2}
\end{figure}
\begin{figure*}[ht]
\centering
{\includegraphics[width=4cm,height=4cm]{exp1_hexagonal_noise_NF.png}}
{\includegraphics[width=4cm,height=4cm]{exp1_hexagonal_final_NF.png}}
{\includegraphics[width=4cm,height=4cm]{exp1_hexagonal_final_NLM.png}}
{\includegraphics[width=4cm,height=4cm]{exp1_hexagonal_final_B.png}}\\
{\includegraphics[width=4cm,height=4cm]{exp1_hexagonal_detail_noise_NF.png}}
{\includegraphics[width=4cm,height=4cm]{exp1_hexagonal_detail_final_NF.png}}
{\includegraphics[width=4cm,height=4cm]{exp1_hexagonal_detail_final_NLM.png}}
{\includegraphics[width=4cm,height=4cm]{exp1_hexagonal_detail_final_B.png}}\\
\hspace{4cm}
{\includegraphics[width=4cm,height=4cm]{exp1_hexagonal_hist_NF.png}}
{\includegraphics[width=4cm,height=4cm]{exp1_hexagonal_hist_NLM.png}}
{\includegraphics[width=4cm,height=4cm]{exp1_hexagonal_hist_B.png}}\\
\hspace{4cm}
{\includegraphics[width=4cm,height=4cm]{exp1_hexagonal_DR_NF.png}}
{\includegraphics[width=4cm,height=4cm]{exp1_hexagonal_DR_NLM.png}}
{\includegraphics[width=4cm,height=4cm]{exp1_hexagonal_DR_B.png}} \\
{\includegraphics[width=4cm,height=4cm]{exp1_hexagonal_contour_noise_NF.png}}
{\includegraphics[width=4cm,height=4cm]{exp1_hexagonal_contour_final_NF.png}}
{\includegraphics[width=4cm,height=4cm]{exp1_hexagonal_contour_final_NLM.png}}
{\includegraphics[width=4cm,height=4cm]{exp1_hexagonal_contour_final_B.png}} \\
\caption{{Denoising experiment. Columns: noisy image, Neighborhood, NLM, and Bilateral filters.
Rows: noisy image, detail of the image, histograms, decreasing rearrangements, and
level curves of image details shown in row 2.}}
\label{fig3}
\end{figure*}
\begin{figure*}[ht]
\centering
{\includegraphics[width=4cm,height=4cm]{exp1_sintetica_noise_NF.png}}
{\includegraphics[width=4cm,height=4cm]{exp1_sintetica_final_NF.png}}
{\includegraphics[width=4cm,height=4cm]{exp1_sintetica_final_NLM.png}}
{\includegraphics[width=4cm,height=4cm]{exp1_sintetica_final_B.png}}\\
{\includegraphics[width=4cm,height=4cm]{exp1_sintetica_detail_noise_NF.png}}
{\includegraphics[width=4cm,height=4cm]{exp1_sintetica_detail_final_NF.png}}
{\includegraphics[width=4cm,height=4cm]{exp1_sintetica_detail_final_NLM.png}}
{\includegraphics[width=4cm,height=4cm]{exp1_sintetica_detail_final_B.png}}\\
\hspace{4cm}
{\includegraphics[width=4cm,height=4cm]{exp1_sintetica_hist_NF.png}}
{\includegraphics[width=4cm,height=4cm]{exp1_sintetica_hist_NLM.png}}
{\includegraphics[width=4cm,height=4cm]{exp1_sintetica_hist_B.png}}\\
\hspace{4cm}
{\includegraphics[width=4cm,height=4cm]{exp1_sintetica_DR_NF.png}}
{\includegraphics[width=4cm,height=4cm]{exp1_sintetica_DR_NLM.png}}
{\includegraphics[width=4cm,height=4cm]{exp1_sintetica_DR_B.png}} \\
{\includegraphics[width=4cm,height=4cm]{exp1_sintetica_contour_noise_NF.png}}
{\includegraphics[width=4cm,height=4cm]{exp1_sintetica_contour_final_NF.png}}
{\includegraphics[width=4cm,height=4cm]{exp1_sintetica_contour_final_NLM.png}}
{\includegraphics[width=4cm,height=4cm]{exp1_sintetica_contour_final_B.png}} \\
\caption{{ Denoising experiment. Columns: noisy image, Neighborhood, NLM, and Bilateral filters.
Rows: noisy image, detail of the image, histograms, decreasing rearrangements, and
level curves of image details shown in row 2.}}
\label{fig4}
\end{figure*}
\begin{figure*}[ht]
\centering
{\includegraphics[width=4cm,height=4cm]{exp2_image_noise.png}}
{\includegraphics[width=4cm,height=4cm]{exp2_h17_final.png}}
{\includegraphics[width=4cm,height=4cm]{exp2_h20_final.png}}
{\includegraphics[width=4cm,height=4cm]{exp2_h50_final.png}}\\
{\includegraphics[width=4cm,height=4cm]{exp2_image_noise_contour.png}}
{\includegraphics[width=4cm,height=4cm]{exp2_h17_contour.png}}
{\includegraphics[width=4cm,height=4cm]{exp2_h20_contour.png}}
{\includegraphics[width=4cm,height=4cm]{exp2_h50_contour.png}}\\
\hspace{4cm}
{\includegraphics[width=4cm,height=4cm]{exp2_h17_DR.png}}
{\includegraphics[width=4cm,height=4cm]{exp2_h20_DR.png}}
{\includegraphics[width=4cm,height=4cm]{exp2_h50_DR.png}}\\
\caption{{Segmentation experiment. Results of applying the Neighborhood filter with
several values of the window size $h$. Rows: Image, level curves and decreasing rearrangement, showing the number of segmented regions (flat regions). Columns: Image,
results of applying the NF with $h=17$, $h=20$, and $h=50$, respectively.}}
\label{fig5}
\end{figure*}
\begin{figure*}[ht]
\centering
\fbox{{\includegraphics[width=3.9cm,height=4cm]{exp2_h17_1.png}}}
\fbox{{\includegraphics[width=3.9cm,height=4cm]{exp2_h17_2.png}}}
\fbox{{\includegraphics[width=3.9cm,height=4cm]{exp2_h17_3.png}}}
\fbox{{\includegraphics[width=3.9cm,height=4cm]{exp2_h17_4.png}}}\\
\hspace*{-4.35cm}\fbox{{\includegraphics[width=3.9cm,height=4cm]{exp2_h20_1.png}}}
\fbox{{\includegraphics[width=3.9cm,height=4cm]{exp2_h20_2.png}}}
\fbox{{\includegraphics[width=3.9cm,height=4cm]{exp2_h20_3.png}}}\\
\caption{{Segmentation experiment. Masks of the segmented regions. First row: $h=17$. The NF produces four regions, corresponding to background, dura-mater and ventricles, grey matter and white matter. Second row: $h=20$. The NF produces three regions, corresponding to background, duramatter and ventricles, grey plus white matter.}}
\label{fig6}
\end{figure*}
\begin{figure*}[ht]
\centering
{\includegraphics[width=5cm,height=4.5cm]{fig_truth_normal_9.png}}
{\includegraphics[width=5cm,height=4.5cm]{fig_FS_normal_9}}
{\includegraphics[width=5cm,height=4.5cm]{fig_FSL_normal_9}}\\
{\includegraphics[width=5cm,height=4.5cm]{fig_SPM_normal_9}}
{\includegraphics[width=5cm,height=4.5cm]{fig_NF_normal_9}}
\caption{{Segmentation experiment. Details of the segmentation produced with several
MRI standard packages and with the Neighborhood filter. See Table~\ref{tab:label} for details. }}
\label{fig7}
\end{figure*}
\section{Summary}
In this paper we introduced the use of the decreasing rearrangement
to express nonlinear and nonlocal filters in terms of integral operators
in the one-dimensional space $[0,\abs{\Omega}]$.
We have proved properties related to the Neighborhood filter nonlinear iterative scheme. In particular, geometric properties like the invariance of level sets and the performance of the filter as a contrast change.
We have also proven a detailed qualitative behavior of the iterations as
a power expansion in terms of the window size, and with coefficients
which depend on up to second order derivatives of the iterations. This allowed us
to distinguish two kind of effects of the filtering process: an anti-diffusive effect of shock-filter type, and a contrast loss effect.
Motivated by the possible piece-wise constant steady state of the discrete problem,
we have illustrated the interpretation of the filter as a segmentation algorithm, indeed
connected to other techniques involving the histogram thresholding.
The main conclusion of our work is that, for certain kind of images, among which those
having concentrated their pixel mass around few intensity levels, the NF is appropriate both
as a denoising and as a histogram-maxima based segmentation algorithm. The execution time
of its rearranged version clearly out-performs those of other algorithms. However, for
other kind of images, specially those with a relatively flat histogram, the results of the NF are poor.
\section{Acknowledgments}
The authors thank to the anonymous reviewers for their interesting comments and suggestions, which have notably contributed to the improvement of our work.
The authors are partially supported by the Spanish DGI Project MTM2010-18427.
|
\section{Introduction}
Symmetry seems to play an important role in the classification and interactions of the elementary particles. The Standard Model (SM) based on the gauge symmetry $SU(3)_C \times SU(2)_L \times U_Y(1)$ has been extremely successful in describing all experimental results so far to a precision less than one percent.
The final ingredient of the SM, namely the Higgs boson, has finally been observed at the LHC \cite{ATLAS}. However, SM is unable to explain why the charges of the elementary particle are quantized because of the presence of $U(1)_Y$. This was remedied by enlarging the $SU(3)_C$ symmetry to $SU(4)_C$ with the lepton number as the fourth color,(or grand unifying all three interaction in SM in $SU(5)$ \cite{GG}
or $SO(10)$ \cite{gfm}).
SM also has no candidate for the dark matter whose existence is now well established experimentally \cite{dm}. Many extensions of the SM models, such as models with weakly interacting massive particles (WIMP) can explain the dark matter \cite{dm}. The most poplar examples are the lightest stable particles in supersymmetry \cite{dm}, or the lightest Kaluza-Klein partcle in extra dimensions \cite{kkdm}. Of course, axion \cite{ww} is also a good candidate for dark matter. Several experiments are ongoing to detect signals of dark matter in the laboratory. However, it is possible that the dark matter is just the analogue of ordinary matter belonging to a parallel universe. Such a parallel universe naturally appears in the superstring theory with the $E_8 \times E'_8$ gauge symmetry before compactification \cite{chsw}. Parallel universe in which the gauge symmetry is just the replication of our ordinary universe, i,e the gauge symmetry in the parallel universe being $SU(3)'\times SU(2)'\times U(1)'$ has also been considered \cite{fv}. If the particles analogous to the proton and neutron in the parallel universe is about five times heavier than the proton and neutron of our universe, then that will naturally explain why the dark matter of the universe is about five times the ordinary matter. This can be easily arranged by assuming strong coupling constant square$/4\pi, \alpha'_s$ is about five times larger than the QCD $\alpha_s$. Thus, in this work, we assume that the two universe where the electroweak sector is exactly symmetric, whereas the corresponding couplings in the strong sector are different, explaining why the dark matter is larger than the ordinary matter. Also, we assume that both universes are described by non-abelian gauge symmetry so that the kinetic mixing between the photon ($\gamma$) and the parallel photon ($\gamma '$) is forbidden.
We also assume that post-inflationary reheating in the two worlds are different, and the the parallel universe is colder than our universe \cite {mohapatra}. This makes it possible to maintain the successful prediction of the big bang nucleosynthesis, though the number of degrees of freedom is increased from the usual SM of 10.75 at the time of nucleosynthesis due the extra light degrees of freedom (due to the $\gamma ', e'$ and three $\nu '$s).
In this work, we explore the LHC implications of this scenario due to the mixing among the Higgs bosons in the two electroweak sectors. Such a mixing, which is allowed by the gauge symmetry, will mix the lightest Higgs bosons of our universe ($h_1$) and the lightest Higgs boson of the parallel universe ($h_2$), which we will call the dark Higgs. One of the corresponding mass eigenstates, $h_{SM}$ we identify with the observed Higgs boson with mass of $125$ GeV. The other mass eigenstate, which we denote by $h_{DS}$, the dark Higgs, will also have a mass in the electroweak scale. Due to the mixing effects, both Higgs will decay to the kinematically allowed modes in our universe and as well as to the modes of the dark universe. One particularly interesting scenario is when the two Higgs bosons are very close in mass, say within 4 GeV so that the LHC can not resolve it \cite {ATLASHgaga}
. However, this scenario will lead to the invisible decay modes\cite{ATLASINV} . The existence of such invisible decay modes can be established at the LHC when sufficient data accumulates. (The current upper limit on the invisible decay branching ratio of the observed Higgs at the LHC is $0.65$). At the proposed future International Linear Collider (ILC) \cite{ILC}, the existence of such invisible modes can be easily established, and the model can be tested in much more detail.
\section{Model and the Formalism}
The gauge symmetry we propose for our work is $ SU(4)_C \times SU(2)_L \times SU(2)_R $ for our universe, and $ SU(4)'_C \times SU(2)'_L \times SU(2)'_R $ for the parallel universe. Note that we choose this non-abelian symmetry not only to explain charge quantization (as in Pati-Salam model \cite{Pati:1974yy}), but also to avoid the kinetic mixing of $\gamma$ and $\gamma '$ as would be allowed in the Standard Model. All the elementary particles belong to the representations of this symmetry group and their interactions are governed by this symmetry. The 21 gauge bosons belong to the adjoint representations $(15,1,1)$, $(1,3,1)$, $(1,1,3)$. $(15,1,1)$ contain the 8 usual colored gluons, 6 lepto-quark gauge bosons $(X, \bar{X})$, and one $(B-L)$ gauge boson \cite{mm}. $(1,3,1)$ contain the 3 left handed weak gauge bosons, while $(1,1,3)$ contain the 3 right handed weak gauge bosons. The parallel universe contains the corresponding parallel gauge bosons. However, so far as the gauge interactions are concerned, we do not assume that the coupling for $SU(4)$ and $SU(4)'$ interactions are the same, but strong coupling in the parallel universe is larger in order to account for the $p'$ (proton of the parallel universe) mass to be about five times larger than the proton. For the electroweak sector, we assume the exact symmetry between our universe and the parallel universe.
The fermions belong to the fundamental representations $( 4, 2,1) + (4,1,2)$. The 4 represent three color of quarks and the lepton number as the 4th color, $(2,1)$ and $(1,2)$ represent the left and right handed doublets. The forty eight Weyl fermions belonging to three generations may be represented by the matrix
\begin{equation}
{\begin{pmatrix}
{\begin{pmatrix} u \\ d \end{pmatrix}}_1 &
{\begin{pmatrix} u \\ d \end{pmatrix}}_2 & {\begin{pmatrix} u \\ d \end{pmatrix}}_3
& {\begin{pmatrix} \nu_e \\ e \end{pmatrix}}_4\\
{\begin{pmatrix} c \\ s \end{pmatrix}}_1 &
{\begin{pmatrix} c \\ s \end{pmatrix}}_2 & {\begin{pmatrix} c \\ s \end{pmatrix}}_3
& {\begin{pmatrix} \nu_{\mu} \\ \mu \end{pmatrix}}_4\\
{\begin{pmatrix} t \\ b \end{pmatrix}}_1 &
{\begin{pmatrix} t \\ b \end{pmatrix}}_2 & {\begin{pmatrix} t \\ b \end{pmatrix}}_3
& {\begin{pmatrix} \nu_{\tau} \\ \tau \end{pmatrix}}_4\\
\end{pmatrix}}_{L, R}.
\end{equation}
We have similar fermion representations for the parallel universe, denoted by primes.
The model has 3 gauge coupling constants: $g_4$ for $SU(4)$ color which we will identify with the strong coupling constant of our universe, $g'_4$ for $SU(4)'$ color of the parallel universe, and $g$ for $SU(2)_L$ and $SU(2)_R$, and corresponding electroweak couplings for the parallel universe ($g_L = g_R = g'_L = g'_R = g$ (we assume that the gauge couplings of the electroweak sectors of the two universe are the same).
\subsection{Symmetry breaking}
$SU(4)$ color symmetry is spontaneously broken to $SU(3)_C \times U(1)_{B-L}$ in the usual Pati-Salam way using the Higgs fields $(15, 1,1)$ at a scale $V_c$. The most stringent limit on the scale of this symmetry breaking comes from the upper limit of the rare decay mode $K_L \rightarrow \mu e$ \cite{KL}.
$SU(2)_L \times SU(2)_R \times U(1)_{B-L}$ can be broken to the SM using the Higgs representations $(1,2,1)$ and $1,1,2)$ at a scale $V_{LR}$. Alternatively, one can use the Higgs multiplets $(1.3,1)$ and $(1,1,3)$ if we want to generate the light neutrino masses at the observed scale. Finally the remaining symmetry is broken to the $U(1)_{EM}$ using the Higgs bi-doublet $(1,2,2)$ as in the left-right model. The $(15, 2, 2)$ Higgs multiplet could also be added to eliminate unwanted mass relations among the charged fermions. Similar Higgs representations are used to break the symmetry in the parallel universe to $U'(1)_{EM}$. A study of the Higgs potential shows that there exist a parameter space where only one neutral Higgs in the bi-doublet remains light, and becomes very similar to the SM Higgs in our universe \cite{Senjanovic}. All other Higgs fields become very heavy compared to the EW scale. Similar is true in the parallel universe. The symmetry of the Higgs fields in the EW sector between our universe and the parallel universe will make the two electroweak VEV's the same. Thus the mixing terms between the two bi-doublets (one in our universe and one in the parallel universe) then leads to mixing between the two remaining SM like Higgs fields. The resulting mass terms for the remaining two light Higgs fields can be written as
$m_{VS}^2 h_{1}^2 + m_{DS}^2 h_{2}^2 + 2 \lambda v_{VS} v_{DS}h_1 h_2$, (where $v_{VS}$ and $v_{DS}$ are the electroweak symmetric breaking scale in the visible sector and dark sector respectively) from which the two mass eigenstates and the mixing can be calculated. The implications for this is when the two light Higgses are very close in mass (within about 4 GeV, which LHC can not resolve) leads to the invisible decay of the observed Higgs boson. Below we discuss the phenomenological implications for this scenario at the LHC, and briefly at the proposed ILC \cite{ILC}.
\section{Phenomenological Implications}
In the framework of this model, interaction between fermions and/or gauge bosons of dark sector and visible sector (the SM particles) are forbidden by the gauge symmetry. However, quartic Higgs interactions of the form $\lambda (H_{VS}^{\dag}H_{VS})(H_{DS}^{\dag}H_{DS})$ (where $H_{VS}$ and $H_{DS}$ symbols denote the Higgs fields in the visible sector and dark sector respectively) are allowed by the gauge symmetry and gives rise to mixing between the Higgses of dark and visible sector. The mixing between the lightest Higgses of dark sector and visible sector gives rise to interesting phenomenological implications at the collider experiments. In this section, we will discuss the phenomenological implications of the lightest dark and visible neutral Higgs mixing ($h_1$ and $h_2$). As discussed in the previous section, the bi-linear terms involving the lightest visible sector (denoted by $h_{1}$) and dark sector (denoted by $h_2$) Higgses in the scalar potential are given by,
\begin{equation}
{\cal L}_{Scalar} \supset m_{VS}^2h_1^2 + m_{DS}^2h_2^2 + 2 \lambda v_{VS} v_{DS} h_1 h_2
\end{equation}
where, $v_{VS}$ and $v_{DS}$ are the electroweak symmetric breaking scale in the visible sector and dark sector respectively. In our analysis, we have assumed the both $v_{VS}$ and $v_{DS}$ are equal to the SM electroweak symmetry breaking scale $v_{SM} \sim 250$ GeV. $m_{VS}$, $m_{DS}$ and $\lambda$ are the free parameters in the theory and the masses ($m_{h_1^{(p)}}$ and $m_{h_2^{(p)}}$) and mixing between physical light Higgs states (denoted by $h_1^{(p)}$ and $h_{2}^{(p)}$) are determined by these parameters:
\begin{eqnarray}
h_{1}^{(p)}&=&{\rm cos}\theta ~h_1 + {\rm sin}\theta ~h_2, \nonumber\\
h_{2}^{(p)}&=&-{\rm sin}\theta ~h_1 + {\rm cos}\theta ~h_2,
\end{eqnarray}
where the masses and the mixing angle of these physical states are given by,
\begin{eqnarray}
m_{h_1^{(p)},h_2^{(p)}}^2&=&\frac{1}{2}[(m_{VS}^2+m_{DS}^2)\mp\sqrt{(m_{VS}^2-m_{DS}^2)^2+4\lambda^2v_{VS}^2v_{DS}^2}]\nonumber\\
{\rm tan}2\theta &=&\frac{2\lambda ~v_{VS}~v_{DS}}{m_{DS}^2-m_{VS}^2}.
\end{eqnarray}
In the framework of this model, we have two light physical neutral Higgs ($h_1^{(p)}$ and $h_{2}^{(p)}$) states. Out of these two Higgs states, we define the SM like Higgs $h_{SM}$ is the state which is dominantly $h_1$-like, i.e., if ${\rm cos}\theta > {\rm sin}\theta$ then $h_{SM}=h_{1}^{(p)}$ and vice versa. The other Higgs is denoted as dark Higgs ($h_{DS}$). Since ATLAS and CMS collaborations have already detected a SM like Higgs boson with mass about 125 GeV, we only studied the scenario where the mass of $h_{SM}$ is between 123 to 127 GeV. Before going into the details of collider implication of visible sector and dark sector Higgs mixing, it is important to understand the correlation between the mixing and mass of the dark Higgs ($m_{h_{DS}}$). To understand the correlation, for few fixed values of $\lambda$, we have scanned the $m_{VS}-m_{DS}$ parameter space.
We have only considered the points which gives rise to a $h_{SM}$ in the mass range between 123 GeV to 127 GeV. For these points, the resulting dark Higgs masses ($m_{h_{DS}}$) and mixing ($\theta$) are plotted in Fig.~\ref{mixing}.
The scatter plot in Fig.~\ref{mixing} shows that large mixing in the visible and dark sector is possible only when the dark Higgs mass is near 125 GeV i.e., near the mass of SM like Higgs boson. It is important to note that the LHC is a proton-proton collider, i.e., LHC collides the visible sector particles only. Therefore, the production cross-section of dark Higgs at the LHC is proportional to the square of the visible sector Higgs component in $h_{DS}$. Therefore, in order to detect the signature of dark Higgs at the collider experiments, we must have significant mixing between the visible and dark sector Higgses. And Fig.~\ref{mixing} shows that significant mixing arises only when dark Higgs and SM like Higgs are nearly degenerate in mass. Therefore, in this article, we studied the phenomenology of two nearly degenerate Higgs bosons with mass about 125 GeV.
\begin{figure}
\begin{center}
\includegraphics[width=10.5 cm,height=8.5cm]{mH_mixing.pdf}
\end{center}
\caption{Scatter plot of dark Higgs mass vs mixing angle for different values of $\lambda$. The SM-like Higgs mass is kept fixed in the range between 123 to 127 GeV denoted by the shaded region in the plot.}
\label{mixing}
\end{figure}
\subsection{Interactions and Decays of light Higgses}
In the present model, two light Higgs physical states ($h_1^{(p)}$ and $h_2^{(p)}$) result from the mixing of visible sector and dark sector light Higgs weak eigenstate $h_1$ and $h_2$ respectively. Visible sector light Higgs weak eigenstates, $h_1$ interacts only with the visible sector fermions ($f$) via Yukawa interactions and gauge bosons ($V$) via gauge interactions. Whereas the dark sector light Higgs weak eigenstate interacts only with the dark fermions $f_D$ and dark gauge bosons $V_D$. However, as a result of mixing, the physical light Higgses interact with both the visible particles and dark particles and thus, they can be produced at the Large Hadron Collider(LHC) experiment. The coupling of the physical states $h_1^{(p)}$ and $h_2^{(p)}$ with the visible as well as dark fermions and gauge bosons can be written as a product of corresponding SM coupling and sine or cosine of the mixing angle. As a result the production cross sections of $h_1^{(p)}$ and $h_2^{(p)}$ and decay widths into visible as well as dark particles can be computed in terms of the SM Higgs production cross-sections/decay widths and the mixing angle. For example, total $h_1^{(p)}$ production cross section at the LHC is given by $\sigma_{SM}{\rm cos}^2\theta$, where $\sigma_{SM}$ is the production cross-section of the SM Higgs with equal mass. Similarly, the decay widths of $h_{1}^{(p)} (h_{2}^{(p)})$ into visible and dark sector fermions are given by $\Gamma_{SM}^{H\to f\bar f} {\rm cos}^2\theta$ ($\Gamma_{SM}^{H\to f\bar f} {\rm sin}^2\theta$) and $\Gamma_{SM}^{H\to f\bar f} {\rm sin}^2\theta$ ($\Gamma_{SM}^{H\to f\bar f} {\rm sin}^2\theta$) respectively, where $\Gamma_{SM}^{H\to f\bar f}$ is the decay width of the SM Higgs into fermions. It is important to note that since the QCD coupling in the dark sector is about $5$ times
larger than the QCD coupling in the visible sector, the Higgs coupling with dark gluon in this model is enhanced by a factor about $5$.\\
In this analysis we are considering both the higgs states in the mass range between $123-127$ GeV. Here we present the expressions for $\mu= \sigma/\sigma_{SM}$ and total $\sigma\times BR_{invible}$ for present model,
\begin{eqnarray}
\mu &=&\frac{(\sigma_{h1}{cos}^4\theta BR_{h1}/(1 + 24 BR_{h1}^{gg}{sin}^2\theta)) + (\sigma_{h2}{sin}^4\theta BR_{h2}/(1 + 24 BR_{h2}^{gg} {cos}^2\theta)) }{\sigma_{SM}*BR}\nonumber\\
\sigma \times BR_{inv} &=&\frac{\sigma_{h1}{cos}^2\theta {sin}^2\theta (BR_{h1}^{inv} + 25BR_{h1}^{gg})}{1+ 24BR_{h1}^{gg} {sin}^2\theta} + \frac{\sigma_{h2}{cos}^2\theta {sin}^2\theta (BR_{h2}^{inv} + 25BR_{h2}^{gg})}{1+ 24BR_{h2}^{gg} {cos}^2\theta}\
\end{eqnarray}
where $\sigma_{h1}$ corresponds to Standard Model Higgs production cross-section at mass of $h_{1}^{(p)}$ and $\sigma_{h2}$ corresponds to Standard Model production cross-section at mass of $h_{2}^{(p)}$ (see Table \ref{production cross section}) and $BR_{h1}$ and $BR_{h2}$ corresponds to Branching ratios of Higgs boson at mass $h_{1}^{(p)}$ and $h_{2}^{(p)}$ respectively(see Table \ref{Decay Branching Ratio1}). For calculating the $\mu$ values in present model we have used Branching Ratios of $H \rightarrow WW \rightarrow l \nu l \nu$ and $H \rightarrow \gamma \gamma$ channels(see Table \ref{Decay Branching Ratio}).
\subsection{Data used in Collider Analysis}
In this section, we discuss the collider phenomenology of invisible Higgs Decays. Before going into
the details of the collider prediction, we first need to study the constraints on the parameter space
coming from present Standard Model predictions and experimental data. The Higgs mass eigenstates of $h_{SM}$ and $h_{DS}$ will be produced in Colliders through the top loop as top quark has Standard Model couplings to the $h_{SM}$ mass eigen state.
The Higgs, which comprises of both $h_1$ and $h_2$ eigen states, will then decay in both the Standard Model decay modes along with Dark sector decay modes. We will perceive these dark sector decay modes as enhancement in the invisible Branching Fraction of the Higgs.\
We first discuss the different constraints on the mixing angle $\theta$ between the two eigenstates coming from experimental data of $H \rightarrow WW \rightarrow l \nu l \nu$ and $H \rightarrow \gamma \gamma$ channels. Along with these experimental data in Higgs decays in different modes, we have also taken into account constraints on the mixing angle parameter space coming from the ATLAS search for the invisible decays of a $125$ GeV Higgs Boson produced in association with a Z boson \cite{ATLASINV}. \
\begin{table}
\begin{center}
\begin{tabular}{|c|c|c|c|c|}
\hline
Mass of Higgs(GeV) & $\sigma_{ggf}$ & $\sigma_{ttH}$ &$\sigma_{VBF}$ &$\sigma_{Vh}$ \\
\hline
\hline
123 & 20.15 & 1.608 & 1.15 & 0.1366 \\
\hline
124 & 19.83 & 1.595 & 1.12 & 0.1334 \\
\hline
125 & 19.52 & 1.578 & 1.09 & 0.1302 \\
\hline
126 & 19.22 & 1.568 & 1.06 & 0.1271 \\
\hline
127 & 18.92 & 1.552 & 1.03 & 0.1241 \\
\hline
\end {tabular}
\end{center}
\vspace{0.2cm}
\caption{ Standard Model production cross section (pb) in different channels for $E_{CM}$ = 8 TeV.}
\label{production cross section}
\end{table}
\begin{table}
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|}
\hline
Mass of Higgs(GeV) & BR($H {\rightarrow}WW$) & BR($H {\rightarrow}ZZ$) & BR($H {\rightarrow}{\gamma\gamma}$) & BR($H {\rightarrow} gg $)& BR($H {\rightarrow}ff$) \\
\hline
\hline
123 & 0.183 & $2.18\times 10^{-2}$ & $2.27 \times 10^{-3}$ & $8.71\times 10^{-2}$ & 0.687 \\
\hline
124 & 0.199 & $2.41\times 10^{-2}$ & $2.27\times 10^{-3}$ & $8.65\times 10^{-2}$ & 0.687 \\
\hline
125 & 0.215 & $2.64\times 10^{-2}$ & $2.28\times 10^{-3}$ & $8.57\times 10^{-2}$ & 0.670 \\
\hline
126 & 0.231 & $2.89\times 10^{-2}$ & $2.28\times 10^{-3}$ & $8.48\times 10^{-2}$ & 0.651 \\
\hline
127 & 0.248 & $3.15\times 10^{-2}$ & $2.27\times 10^{-3}$ & $8.37\times 10^{-2}$ & 0.633 \\
\hline
\end {tabular}
\end{center}
\vspace{0.2cm}
\caption{ Standard Model Decay Branching Ratio in different channels.}
\label{Decay Branching Ratio1}
\end{table}
\begin{table}
\begin{center}
\begin{tabular}{|c|c|c|}
\hline
Channels for Higgs Decay & $\mu$ value by ATLAS & $\mu$ value by CMS \\
\hline
\hline
$H \rightarrow WW \rightarrow l \nu l \nu$ & $1.01 \pm 0.31$ & $0.76 \pm 0.21$ \\
\hline
$H \rightarrow \gamma \gamma$ & $1.65 \pm 0.24(stat) ^ {+0.25}_{-0.18}(syst) $ & $0.78 \pm 0.27$ \\
\hline
\end {tabular}
\end{center}
\vspace{0.2cm}
\caption{ Experimental values of best fit signal strength $\mu = \sigma/\sigma_{SM}$ at $E_{CM}$ = 8 TeV.}
\label{Decay Branching Ratio}
\end{table}
The Standard Model production cross-sections in different channels (such as gluon-gluon fusion, ttH, vector boson fusion and vector boson (both W boson and Z boson) in association with a Higgs boson) at $E_{CM}$ = 8 TeV and Decay Branching ratios in different channels (such as $H {\rightarrow}WW$, $H {\rightarrow}ZZ$,$H {\rightarrow}{\gamma\gamma}$,$H {\rightarrow} gg$,$H {\rightarrow}ff$) has been given by ATLAS collaboration in reference \cite{SM production} \cite{SM decay}. We have used these cross-sections and branching ratios in different channels in our analysis.
The relevant cross-sections and branching ratios used for our analysis are presented in Table \ref{production cross section} and Table \ref{Decay Branching Ratio1} respectively. We have taken the mass range between $123 - 127$ GeV which is the interesting parameter space for our analysis.\
In Table \ref{Decay Branching Ratio} we present the results of the different experimental searches in the $H \rightarrow WW \rightarrow l \nu l \nu$ channel by ATLAS collaborations \cite{ATLASWW} and CMS collaboration \cite{CMSWW} and in $H \rightarrow \gamma \gamma$ channel by ATLAS collaborations\cite{ATLASGG} CMS collaborations\cite{CMSGG} .
\subsection{Bounds on Mixing Angle}
In this section we use the data that we presented in the previous section to constrain the mixing angle parameter space.
In Fig \ref{fig:decayrate}, we present the total invisible decay rate i.e $\sigma \times BR$ in the invisible channel vs the mixing angle $\theta$ for
$ m_{h1}^{(p)} = 123$ GeV and $ m_{h2}^{(p)} = 127$ GeV ($m_{h1}^{(p)} = 124 GeV$ and $m_{h2}^{(p)} = 126 GeV)$ . ATLAS collaboration has searched for the invisible decay of higgs boson in Z H production channel at $E_{CM} = 8 TeV$. In absence of any significant deviation of data from the Standard Model background prediction, ATLAS collaboration has set an upper limit of $65\%$ on the invisible decay branching of a SM higgs boson of mass $125$ GeV \cite{ATLASINV}. Assuming $\sigma_{total}$ = $22.32$ pb Higgs cross-section at $125$ GeV (see Table \ref{production cross section}), $65\%$ upper limit on invisible decay branching ratio corresponds to $14.5$ pb upper limit on the invisible Higgs decay rate. This limit is shown in the shaded green region in Fig \ref{fig:decayrate}. It can be seen from the plot that present model is consistent with ATLAS experimental data for $\theta < 33^{o} $ and $\theta > 58^{o}$ in the parameter space region.
\begin{figure}
\begin{center}
\vspace*{-2.0cm}
\includegraphics[width=10.5cm,height=8.5cm]{sig_br.pdf}
\caption{Decay rate in invisible channels in present model as a function of mixing angle $\theta$. The shaded regions correspond to SM allowed values for $\sigma \times BR_{inv}$.}
\label{fig:decayrate}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\vspace*{-2.0cm}
\includegraphics[width=10.5cm,height=8.5cm]{H_gaga.pdf}
\caption{Higgs decaying into diphoton rate in present model as a function of mixing angle $\theta$. The shaded regions again correspond to ATLAS and CMS allowed $\mu = \sigma/\sigma_{SM}$ values.}
\label{fig:H_gaga}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\vspace*{-2.0cm}
\includegraphics[width=10.5cm,height=8.5cm]{H_WW.pdf}
\caption{$H \rightarrow WW \rightarrow l \nu l \nu$ rate in present model as a function of mixing angle $\theta$. The shaded regions correspond to ATLAS and CMS allowed $\mu = \sigma/\sigma_{SM}$ values.}
\label{fig:H_WW}
\end{center}
\end{figure}
In Fig. \ref{fig:H_gaga} we have presented a plot of $\mu = \sigma/\sigma_{SM}$ in the $H \rightarrow \gamma \gamma$ channel as a function of the mixing angle $\theta$. The plot shows prediction in present model for $m_{h1}^{(p)} = 123$ GeV and $ m_{h2}^{(p)} = 127$ GeV $(m_{h1}^{(p)} = 124$ GeV and $m_{h2}^{(p)} = 126$ GeV) mass values.The yellow shaded region corresponds for allowed region by CMS collaboration and green shaded region is allowed region for ATLAS collaboration in this channel. It can be seen from the plot that CMS allowed region is consistent for all $\theta$'s for the present model,but present model is not consistent with ATLAS allowed region for any values of $\theta$. We point out that $H \rightarrow \gamma \gamma$ data for ATLAS, is well above the SM expectation. If the present model is realized by Nature, with the accumulation of more data with higher luminosities at the Large Hadron Collider(LHC) the $H \rightarrow \gamma \gamma$ branching ratio measured by ATLAS experiment should should come down significantly from present experimental value of $1.65 \pm 0.24(stat) ^ {+0.25}_{-0.18}(syst) $. Our model is consistent with the lower $\mu$ value of $0.78 \pm 0.27$ for $H \rightarrow \gamma \gamma$ as measured by the CMS experiment for the whole parameter of the parameter space.
In Fig.~\ref{fig:H_WW} we present a plot of $\mu = \sigma/\sigma_{SM}$ in the $H \rightarrow WW \rightarrow l \nu l \nu$ channel with mixing angle $\theta$. Two curves for $ m_{h1}^{(p)} = 123$ GeV and $ m_{h2}^{(p)} = 127$ GeV $(m_{h1}^{(p)} = 124$ GeV and $m_{h2}^{(p)} = 126$ GeV) present the prediction for present model. The yellow shaded region corresponds for allowed region by CMS collaboration and green shaded region is for allowed region by ATLAS collaboration in this channel. It can be seen from the plot that ATLAS allowed region is consistent with present model for $\theta < 13 (16)^{o} $ and $\theta > 70 (71)^{o} $ region in the parameter space. It can also be seen that present model is also consistent with CMS allowed region for $\theta < 20 (23)^{o}$ and $\theta > 65 (66)^{o}$ parameter space. It is interesting to note that the prediction curves for the present model with mass values of $ m_{h1}^{(p)} = 123$ GeV and $ m_{h2}^{(p)} = 127$ GeV $(m_{h1}^{(p)} = 124$ GeV and $m_{h2}^{(p)} = 126$ GeV) are not symmetric. It can be understood by taking into the fact that in low $\theta$ region $ m_{h1}^{(p)}$ is SM like. As $ m_{h1}^{(p)}$ is lower than $ m_{h2}^{(p)}$ for both curves, the cross-section $\times$ Branching ratio is smaller in lower $\theta$ region. Whereas for high $\theta$ region $ m_{h2}^{(p)}$ is SM like and as it is heavier than $ m_{h1}^{(p)}$ for both curves the cross section $\times$ Branching Ratio is higher in this region,which makes the curves non-symmetric.\\
This present analysis in the $H \rightarrow WW \rightarrow l \nu l \nu$ channel gives the most stringent constraint of $\theta < 13 (16)^{o} $ and $\theta > 70 (71)^{o} $ on the parameter space for the mixing angle $\theta$ taking into account all the constraints coming from analysis in $\sigma \times BR_{invisible}$, $H \rightarrow \gamma \gamma$ and $H \rightarrow WW \rightarrow l \nu l \nu$ channels. From this analysis in different channels it is certain that there is still plenty of parameter space available for the present model taking into account all the known experimental constraints at the LHC.\\
We would also like to comment that in a linear collider like the proposed International Linear Collider(ILC) this analysis can be done without any ambiguity about the resolution of the two Higgs in the close range of $4GeV$. In a $e+e-$ collider the Higgs will be produced in association with a Z boson and from the mass recoil of the Z boson the peak resolution of the Higgs boson can be measured in the limit of $40$ MeV \cite{ILC}. So from linear colliders we will be able to tell for sure if there are two Higgs bosons in the comparable mass range between ($123-127$GeV), which is not possible in this precision from Hadron Collider like LHC.
\section{Summary and Conclusions}
Motivated by the fact that the dark matter is about five times the ordinary matter, we have proposed that the dark matter can just be like the ordinary matter in a parallel universe with the corresponding strong coupling constant, $\alpha^{'}_s$ about five times the strong coupling, $\alpha_s$ of our universe. The parallel universe needs to be much colder than our universe to keep the successful prediction for the big bang nucleosynthesis. We have used the non-abelian Pati-Salam gauge symmetry for both universe to have the charge quantization, as well as, to avoid any kinetic mixing between the photon of our universe and the parallel universe. However, the two universes will be connected via the electroweak Higgs bosons of the two universes. If the electroweak sector of the two universes are symmetric, the lightest Higgs bosons of the two universes will mix. In particular, if these two Higgses mix significantly, and their masses are close (say within 4 GeV), LHC will not be able to resolve if it is observing one Higgs or two Higgses. However, each Higgs will decay to the particles of our universe as well as to the corresponding particles of the the parallel universe. This leads to the invisible decays of the observed Higgs boson (or bosons). We have used all the available experimental data at the LHC to set constraint on this mixing angle, and find that in can be as large as $16^{o}$. If the mixing angle is not very small, LHC will be able to infer the existence of such invisible decays when sufficient data accumulates. (The current limit on the invisible branching ratio from the LHC data is
$<65\%$). We also find that the cross section times the branching ratio for Higgs to $\gamma \gamma$ channel is fully consistent with our model as measured by the CMS collaboration, but not by the ATLAS collaboration. The results by the ATLAS collaboration for this channel has to come down if our model is realized by nature. Our proposal of two Higgses around $125$ GeV , and significant invisible decay fraction can easily be tested in the proposed ILC where peak resolution of the Higgs boson can be measured to about $40$ MeV.
{\bf Acknowledgment:} We thank K. S. Babu for some important comments and him and S. Gottlieb for useful discussions. The work of SC, KG and SN was supported in part by the U.S. Department of Energy
Grant Number DE-SC0010108.
|
\section{Introduction}\label{sec:intro}
Consider the linear system
\begin{equation}
(H - zS) u = b,
\label{eqn:multishift}
\end{equation}
where $H$ and $S$ are $N\times N$ Hermitian matrices, \response{$S$ is positive definite}, $u$ and $b$ are vectors of
length $N$, and $z\in \mathbb{C}$ with $\Re z\le 0$.
Often~\eqref{eqn:multishift} needs to be solved a large number of
times, both for a large number of distinct parameters $z$, and for
multiple right hand sides $b$. This type of calculation arises in a
number of applications, such as time-independent density functional
perturbation theory
(DFPT)~\cite{BaroniGiannozziTesta1987,GonzeAllanTeter1992,BaroniGironcoliDalEtAl2001},
many body perturbation theory using the GW
method~\cite{Hedin1965,AryasetiawanGunnarsson1998,FriedrichSchindlmayr2006,UmariStenuitBaroni2010,PingRoccaGalli2013,GiustinoCohenLouie2010},
and the random phase approximation (RPA) of the electron correlation
energy~\cite{LangrethPerdew1975,LangrethPerdew1977,Furche2001,NguyenGironcoli2009}.
The connection between~\eqref{eqn:multishift} and these
applications will be illustrated in Section~\ref{sec:connection}.
\subsection{\response{Previous work}}
From an algorithmic point of view, solving linear systems with multiple
shifts has been widely explored in the
literature~\cite{Frommer2003,FrommerGlassner1998,SimonciniSzyld2007,DattaSaad1991,Meerbergen2003,GallivanGrimmeVan1996,BaiFreund2001,FeldmannFreund1995,SABAKI2012}.
These methods are often based on the Lanczos method. The basic idea here is
that when $S=I$ the Krylov subspace constructed in the Lanczos procedure
is invariant to the shifts $z$. Therefore the same set of Lanczos
vectors can be used simultaneously to solve~\eqref{eqn:multishift}
with multiple $z$. When $S$ is not the identity matrix, $S$ should be factorized,
such as by the Cholesky factorization, and \response{triangular solves with the Cholesky factors must be
computed} during every iteration, which can significantly increase the
computational cost.
In its simplest form the Lanczos method as described in Section
\ref{sec:lanczos} precludes the use of a preconditioner. However, there
are strategies, some of which are discussed in the aforementioned
references, for using a small number of preconditioners and perhaps multiple
Krylov spaces to find solutions for all of the desired parameters. Even
when using preconditioners it may be unclear how to pick which Krylov
spaces to use, or what the shifts for the preconditioners should be, see
\textit{e.g.}, \cite{Meerbergen2003}. Another specific class of methods are the so called recycled Krylov methods, see, \textit{e.g.}, \cite{Parks2006}. In such methods, the work done to solve one problem, potentially with a preconditioner, is reused to try and accelerate the convergence of related problems.
Another class of popular methods, especially in the context of model
reduction, are based on the Pad\'{e} approximation, which does not target the
computation of all the entries of $u(z)=(H-zS)^{-1} b$ as in~\eqref{eqn:multishift}, but rather a linear functional of $u(z)$ in the form
of $f(z)=l^{T}u(z)$. $f(z)$ is a scalar function of $z$,
which can be stably
expanded in the Pad\'{e} approximation via the Lanczos
procedure~\cite{FeldmannFreund1995}. However, it is not obvious how to
obtain a Pad\'{e} approximation for all entries of $u(z)$ directly, and with an
approximation that has a uniformly bounded error for all $\Re z\le 0$.
\subsection{\response{Contribution}}
In this paper we develop a new approach for solving para\-metrized
linear systems using a pole expansion. The pole expansion used here is a rational
approximation to all entries of $u(z)$ simultaneously. The main idea behind this
pole expansion comes directly from the work of Hale, Higham and
Trefethen~\cite{HHT}, which finds a nearly optimal quadrature rule of
the contour integral representation for $u(z)$ in the complex plane.
The idea of the work in~\cite{HHT} has also been adapted in the context of
approximating the spectral projection operator and the Fermi-Dirac
operator (a ``smeared'' spectral projection operator) for ground state electronic structure
calculation~\cite{LinLuYingE2009}.
Such a scheme directly expresses $u(z),$ and thus the solution to \eqref{eqn:multishift} for any $z,$
as the linear combination of solutions to a small set of linear systems, each of which has a fixed parameter.
Each linear system with a fixed parameter may be solved either with a preconditioned iterative
method or a direct method. Furthermore, the construction of the pole expansion depends only on the largest and smallest positive generalized eigenvalues of the matrix pencil and the number of poles. The number of poles used in the expansion yields control over the accuracy of the scheme. In fact, the pole expansion converges exponentially
with respect to the number of approximating terms across the range
$\Re z \le 0.$ Finally, the pole expansion allows the same
treatment for general $S$ as in the case of $S=I,$ without directly
using its Cholesky factor or $S^{-1}$ at every iteration. As a result,
the pole expansion addresses the disadvantages of both the Lanczos method
and the Pad\'{e} approximation.
\subsection{Notation}
In this paper we use the following notation. Let
$\{\lambda_{i}\},\{\psi_{i}\}$ be the generalized eigenvalues and eigenvectors of
the matrix \response{pencil} $(H,S)$ which satisfy
\begin{equation}
\label{eqn:GE}
H\psi_{i} = S \psi_{i} \lambda_{i},\quad i=1,\ldots,N,
\end{equation}
with $\{\lambda_{i}\}$ labeled in non-increasing order. Such an
eigen decomposition is only used for deriving the pole expansion, and is
not performed in practical calculations.
We initially assume that all the eigenvalues $\lambda_{i}$ are positive, and denote by
$E_{g}\equiv \lambda_N>0$. However, we later relax this assumption and allow for a small number of negative eigenvalues. We also denote by $\Delta E$ the spectrum
width of the $(H,S)$ \response{pencil} \textit{i.e.}\ $\Delta E = \lambda_{1}-\lambda_{N}$.
We refer to Fig.~\ref{fig:pole1} for an illustration of the relative position
of the spectrum and the range of the parameter $z$ in the complex plane.
\begin{figure}
\centering
\includegraphics[width = 0.6\textwidth]{pole1.pdf}
\caption{A schematic view of the range of the spectrum of the
$(H,S)$ \response{pencil} (thick red line on the positive real axis),
and the range of the parameter $z$ (light gray area), separated by a
positive distance $E_{g}$.}
\label{fig:pole1}
\end{figure}
Furthermore, $A^*$ and $A^{T}$
denote the conjugate transpose and the transpose of a matrix $A$,
respectively. We let $R$ be an upper triangular matrix that denotes the
Cholesky factorization of $S,$ \textit{i.e.},
\begin{equation*}
S = R^* R.
\end{equation*}
We use the notation
\begin{equation*}
\mathcal{K}_{n}(A,b) = \text{span}\left(b,Ab,\ldots,A^{n-1}b \right)
\end{equation*}
to denote the $n^{th}$ Krylov subspace associated with the matrix $A$ and
the vector $b.$ Finally, we let $N_z$ be the number of distinct complex
shifts $z$ for which we are interested in solving~\eqref{eqn:multishift}.
\subsection{Outline}
The rest of the manuscript is organized as follows. We first discuss
the standard Lanczos method for solving linear systems with multiple
shifts in Section~\ref{sec:lanczos}. We introduce the pole expansion method
for solving~\eqref{eqn:multishift}, and analyze
the accuracy and complexity of the approach in Section~\ref{sec:pole}. In
general the matrix \response{pencil} $(H,S)$ may not always have all positive eigenvalues,
and the case where there are positive and negative eigenvalues is discussed in
Section~\ref{sec:indefinite}.
The connection between~\eqref{eqn:multishift} and electronic
structure calculations is given in Section~\ref{sec:connection}.
The numerical results with applications
to density functional perturbation theory calculations are given in
Section~\ref{sec:numer}.
We conclude in Section~\ref{sec:conclusion}.
\section{Lanczos method for parametrized linear systems}\label{sec:lanczos}
We briefly describe \response{a basic variant of the} Lanczos method for solving parametrized linear systems of the form \response{\eqref{eqn:multishift}}
where the matrix \response{pencil} $(H,S)$ satisfies the conditions given in the
introduction.
Using the Cholesky factorization of $S$ we may transform~\eqref{eqn:multishift} in a manner such that we instead solve
\begin{equation}
\label{eqn:chol_eq}
(R^{-*}HR^{-1}-zI)\tilde{u} = \tilde{b},
\end{equation}
where $u = R^{-1}\tilde{u}$ and $\tilde{b} = R^{-*}b.$ Since $R^{-*}HR^{-1}$ is Hermitian positive definite we now briefly describe the Lanczos method for parametrized systems under the assumption that our equation is of the form
\begin{equation*}
(A-zI)x = \tilde{b},
\end{equation*}
for some Hermitian positive definite matrix $A,$ right hand side $\tilde{b}$ and complex shift $z$ such that $\Re z \leq 0.$
Given these systems, we note that
\begin{equation*}
\mathcal{K}_k(A,\tilde{b}) = \mathcal{K}_k(A-zI,\tilde{b})
\end{equation*}
for any complex scalar $z.$ Based on this observation, the well known Lanczos process, see, \textit{e.g.}, \cite{GVL} for details \response{and \cite{Lanczos1950} for a historical perspective}, may be slightly modified such that at each iteration approximate solutions, each of which satisfy the Conjugate Gradient (CG) error criteria, are produced for each shift. \response{However, in this basic formulation using the invariance of the Krylov subspaces precludes the use of preconditioners for the problems.}
\begin{remark}
We note that the restriction $\Re z \leq 0$ may be relaxed if we
instead use a method analogous to MINRES to simultaneously solve the
shifted systems, for details of MINRES see, \textit{e.g.}, \cite{PASA1975}. Such
an algorithm is only slightly more expensive than the CG style
algorithm given here and scales asymptotically in the same manner.
\end{remark}
We observe that the computational cost of using this method breaks down
into two distinct components. First, there is the cost of the Lanczos
procedure which only has to be computed once regardless of the number of
shifts. The dominant cost of this procedure is a single matrix vector
multiplication at each iteration. In the case where $S\neq I$ the method
actually requires a single application of $R^{-*}HR^{-1},$ which is
accomplished via two triangular solves and one matrix vector product
with $H.$ Second, for each shift $z$ a $k\times k$ tridiagonal system, denoted $T_k,$
must be solved. However, this may be done very cheaply by maintaining a
$LDL^*$ factorization of $T_k$ for each shift (see, \textit{e.g.}, \cite{GVL}
for a detailed description of the case without a shift). This means that the
computational costs at each iteration for solving the set of
sub-problems scales linearly with respect to both the
number of shifts and the problem size. Thus, the dominant factor in the
computation is often the cost of the Lanczos process.
Furthermore, properties of the Lanczos process and the structure of the $LDL^*$
factorization imply that if we neglect the cost of storing $A$ the memory
costs of the algorithm scale linearly in both the number of shifts and
the problem size. At the core of this memory scaling is the fact that
not all of the Lanczos vectors must be stored to update the solution, see, \textit{e.g.}, \cite{GVL}. In fact, for
large problems that take many iterations storing all of the vectors
would be infeasible. However, this does impact the use of the algorithm
for multiple shifts. Specifically, the set of shifts for which solutions are desired must be decided
upon before the algorithm is run so that all of the necessary
factorizations of $T_k$ may be built and updated at each iteration. If a
solution is needed for a new shift the algorithm would have to be run
again from the start unless all of the Lanczos vectors were stored.
\section{Pole expansion for parametrized linear systems}\label{sec:pole}
We now describe a method based on the work of Hale, Higham and Trefethen \cite{HHT} to simultaneously solve systems of the form
\begin{equation}
(H-z_lS)u_l = b
\label{eqn:pole_shift}
\end{equation}
\response{where the subscript $l$ explicitly denotes the set of $N_z$ distinct shifts. Here, the matrix pencil $(H,S)$ and the shifts $z_l$ satisfy the same properties as in the introduction.}
\subsection{Constructing a pole expansion}
First, we provide a brief overview of the method for computing functions
of matrices presented in \cite{HHT}. Given an analytic function $f$,
a Hermitian matrix $A,$ and a closed contour $\Gamma$ that
only encloses the analytic region of $f$ and winds once around the spectrum of
$A$ in a counter clockwise manner, then $f(A)$ may be represented via contour
integration as \begin{equation}
\label{eqn:contour}
f(A) = \frac{1}{2\pi i}\int_{\Gamma}f(\xi)(\xi I-A)^{-1}d\xi.
\end{equation}
The authors in \cite{HHT} provide a method via a sequence of conformal mappings to generate an efficient quadrature scheme for~\eqref{eqn:contour} based on the trapezoidal rule. Specifically, in \cite{HHT} a map from the region $S = [-K,K]\times [0,K']$ to the upper half plane of ${\Omega = \mathbb{C} \backslash ((-\infty,0]\cup [\lambda_{\min} (A), \lambda_{\max} (A)])}$ is constructed via
\begin{equation*}
\begin{split}
z &= \sqrt{\lambda_{\min} (A) \lambda_{\max} (A)}\left(\frac{k^{-1}+u}{k^{-1}-u} \right),\\
u &= sn(t) = sn(t|k),\\
k &= \frac{\sqrt{\lambda_{\max} (A) / \lambda_{\min} (A)}-1}{\sqrt{\lambda_{\max} (A) / \lambda_{\min} (A)}+1},
\end{split}
\end{equation*}
where $t\in S$ and $z$ is in the upper half plane of $\Omega.$ The constants $K$ and $K'$ are complete elliptic integrals and $sn(t)$ is one of the Jacobi elliptic functions. Application of the trapezoidal rule in $S$ using the $P$ equally spaced points
\begin{equation*}
t_j = -K + \frac{iK'}{2}+2\frac{(j-\frac{1}{2})K}{P}, \;\; 1\leq j \leq P,
\end{equation*}
yields a quadrature rule for computing~\eqref{eqn:contour}. In \cite{HHT} the assumption made is that the only non-analytic region of $f$ lies on the negative real axis.
Here, we instead consider the case where the non-analytic region of $f$
may be anywhere in the negative half plane. Therefore,
for our purposes a modification of the transform in \cite{HHT} must be
used. Specifically, we use the construction of the quadrature presented
in Section 2.1 of \cite{LinLuYingE2009}, where now an additional
transform of the form $\xi = \sqrt{z}$ is used to get the quadrature
nodes. We note that there is a slight difference between the contour
used here and the one is \cite{LinLuYingE2009}. Because we are assuming
the matrix is Hermitian positive definite we only need to consider a
single branch of the square root function in defining the nodes $\xi_j,$
in this case the positive one.
The procedure outlined may be used to generate a $P$ term pole expansion for a function $f(A)$ denoted
\begin{equation}
\label{eqn:Pole_expand}
f_P(A) \approx \sum_{k=1}^P w_k f(\xi_k)\left(\xi_k I - A\right)^{-1},
\end{equation}
where $\xi_k$ and $w_k,$ $k=1,\ldots,P$ are the quadrature nodes and weights respectively. Both the nodes and the weights depend on $P,$ $\lambda_{\max} (A),$ and $\lambda_{\min} (A)$. In \cite{HHT} asymptotic results are given for the error $\|f(A)-f_P(A)\|.$ For a Hermitian positive definite matrix $A$ the asymptotic error in the expansion of the form~\eqref{eqn:Pole_expand}, see \cite{HHT} Theorem 2.1 and Section 2 of \cite{LinLuYingE2009}, behaves as
\begin{equation}
\label{eqn:HHTerror}
\|f(A)-f_P(A)\| = \mathcal{O} \left( e^{-C P / \log\left(\lambda_{max}(A)/\lambda_{min}(A)\right)} \right),
\end{equation}
where $C$ is a constant independent of $P,\lambda_{max}(A)$ and $\lambda_{min}(A).$
\begin{remark}
We make special note of the fact that in the framework where we
have all positive generalized eigenvalues the constant in the exponent
does not depend on \response{where the function $f$ has poles in the left half plane.} Specifically, the
initial use of the $z = \xi^2$ mapping maps the poles on the imaginary
axis to the negative real axis, see \cite{LinLuYingE2009} for details.
All of the poles \response{with strictly negative real part} get mapped to a distinct sheet
of the Riemann surface from the one that contains the positive real axis
and the poles that were initially on the imaginary axis. As noted
previously, the only sheet we need to consider building a contour on is
the one containing the spectrum of $A$, and thus $C$ does not depend on the \response{locations of the poles.}
\end{remark}
Fig.~\ref{fig:pole2} illustrates the type of contours computed by this scheme and used throughout the remainder of the paper. To facilitate the computation of certain quantities necessary in the generation of these contours we used a Schwarz-Christoffel toolbox \cite{DriscollSC1,DriscollSC2}.
\begin{figure}
\centering
\includegraphics[width = 0.6\textwidth]{Pole2.pdf}
\caption{A schematic view of the range of the spectrum of the
$(H,S)$ \response{pencil} (thick red line on the positive real axis),
and the range of the parameter $z$ (light gray area), separated by a
positive distance $E_{g}$, together of the Cauchy contour (yellow
line) surrounding the positive part of the spectrum and the
discretized contour integration points (poles, black dots).}
\label{fig:pole2}
\end{figure}
\subsection{\response{Solving shifted linear systems}}
We now address the use of a pole expansion to solve problems of the form~\eqref{eqn:pole_shift}.
Denote by $\Psi = [\psi_{1},\cdots,\psi_{N}]$, and $\Lambda =
\mathrm{diag}[\lambda_{1},\cdots,\lambda_{N}]$, from
the generalized eigenvalue problem ~\eqref{eqn:GE} we then have
\begin{equation*}
\Psi^{*} H \Psi = \Lambda, \quad \Psi^{*} S \Psi = I.
\end{equation*}
We emphasize that the eigen decomposition is only used in the
derivation of the method and is not performed in practical calculations.
For the moment we assume that $\lambda_1,\ldots,\lambda_N$ are positive and ordered from largest to smallest. From~\eqref{eqn:multishift} we have for each shift $z_l$
\begin{equation*}
\begin{split}
u_l &= (H - z_l S)^{-1} b \\
&= \left(\Psi^{-*} \Lambda \Psi^{-1} - z_l \Psi^{-*}
\Psi^{-1}\right)^{-1} b\\
&= \Psi \left( \Lambda - z_l \right)^{-1} \Psi^{*} b.
\end{split}
\end{equation*}
Since $\Re z\le 0,$ we may now use a pole expansion for $f^l(A) = 1/(A-z_lI)$ generated by the procedure in \cite{LinLuYingE2009} to write
\begin{equation}
\begin{split}
u_l^P & = \Psi \sum_{k=1}^{P} \frac{\omega_{k}}{\xi_{k}-z_l}
\left(\Lambda - \xi_{k}\right)^{-1}
\Psi^{*} b\\
& = \sum_{k=1}^{P} \frac{\omega_{k}}{\xi_{k}-z_l}
\left(\Psi^{-*}\Lambda\Psi^{-1} - \xi_{k}\Psi^{-*}\Psi^{-1}\right)^{-1} b \\
&= \sum_{k=1}^{P} \frac{\omega_{k}}{\xi_{k}-z_l}
\left(H - \xi_{k}S\right)^{-1} b,
\end{split}
\label{eqn:pole_approx}
\end{equation}
which yields an approximate solution, $u_l^P,$ to the true solution $u_l.$
\begin{remark}
Using the pole expansion method to solve the parametrized systems~\eqref{eqn:multishift} motivated our choice of $f.$ However, the use of the pole
expansion only places mild requirements on $f$, so the method presented
here may potentially be used to solve systems with different types of
parametrization.
\end{remark}
To simplify the notation let us define $h_k$ as the solutions to the set of problems
\begin{equation}
\label{eqn:pole_sub}
\left(H - \xi_{k}S\right)h_k = b,
\end{equation}
such that our approximate solution is now simply formed as
\begin{equation}
\label{eqn:solution}
u_l^P = \sum_{k=1}^{P} \frac{\omega_{k}}{\xi_{k}-z_l} h_k.
\end{equation}
If we assume that $h_k$ is computed exactly the solutions to~\eqref{eqn:pole_shift} formed via~\eqref{eqn:solution} satisfy the asymptotic, $P \rightarrow \infty,$ error bound:
\begin{equation*}
\frac{\|u_l - u_l^P\|_S}{\|\Psi^* b\|_2} \leq \mathcal{O} \left( e^{-C P / \log\left(\lambda_1/\lambda_N\right)} \right).
\end{equation*}
Given that the $h_k$ are computed exactly we may write
\begin{equation}
\begin{split}
\|u_l - u_l^P\|_S &= \|\Psi f^l(\Lambda) \Psi^* b - \Psi f^l_P(\Lambda) \Psi^* b\|_S \\
&= \|f^l(\Lambda) \Psi^* b - f^l_P(\Lambda) \Psi^* b\|_2 \\
&\leq \|f^l(\Lambda) - f^l_P(\Lambda) \|_2 \|\Psi^* b\|_2.
\end{split}
\label{eqn:errS}
\end{equation}
Finally, using the error approximation~\eqref{eqn:HHTerror} in conjunction with~\eqref{eqn:errS} yields the desired result. Furthermore, in the special case where $S=I$ the error bound reduces to
\begin{equation*}
\frac{\|u_l - u_l^P\|_2}{\|b\|_2} \leq \mathcal{O} \left( e^{-C P / \log\left(\lambda_1/\lambda_N\right)} \right).
\end{equation*}
These bounds show that asymptotically the error decreases exponentially with respect to the number of poles used in the expansion for $f^l(A).$ However, there is additional error introduced since $h_k$ is computed inexactly. Thus, the overall error will often be dominated by the error in the computation of $h_k.$ More specifically, let us define
\begin{equation}
\tilde{u}_l^P = \sum_{k=1}^{P} \frac{\omega_{k}}{\xi_{k}-z_l} \tilde{h}_k
\label{eqn:approx_solution}
\end{equation}
where $\tilde{h}_k$ represents an approximation of $h_k.$ We assume that $\tilde{h}_k$ satisfies the error bound
\begin{equation}
\label{eqn:pole_sub_error}
\frac{\|\Psi^*(b - (H-\xi_kS)\tilde{h}_k)\|_2}{\|\Psi^*b\|_2} \leq \epsilon
\end{equation}
and define $r_k = b - (H-\xi_kS)\tilde{h}_k.$ In this case the solutions to~\eqref{eqn:pole_shift} formed via~\eqref{eqn:approx_solution} satisfy the asymptotic, $P \rightarrow \infty,$ error bound:
\begin{equation*}
\frac{\|u_l - \tilde{u}_l^P\|_S}{\|\Psi^*b\|_2} \leq \mathcal{O} \left( e^{-C P / \log\left(\lambda_1/\lambda_N\right)} \right) + \epsilon \left( \max_{i=1,\ldots,N}\frac{1}{\vert \lambda_i-z_l \vert} \right).
\end{equation*}
Under the assumptions about the inexact solutions made here we may write
\begin{equation*}
\begin{split}
\tilde{u}_l^P &= \sum_{k=1}^{P} \frac{\omega_{k}}{\xi_{k}-z_l} \tilde{h}_k \\
&= \sum_{k=1}^{P} \frac{\omega_{k}}{\xi_{k}-z_l} \left( h_k - (H-\xi_kS)^{-1}r_k \right) \\
&= u_l^P - \sum_{k=1}^{P} \frac{\omega_{k}}{\xi_{k}-z_l} (H-\xi_kS)^{-1}r_k.
\end{split}
\end{equation*}
In conjunction with the error bound~\eqref{eqn:pole_sub_error} this implies that
\begin{equation}
\label{eqn:utilde_error}
\begin{split}
\frac{\|u_l - \tilde{u}_l^P\|_S}{\|\Psi^*b\|_2} &= \frac{\|u_l - u_l^P + \Psi f_P^l(\Lambda) \Psi^*r_k\|_S}{\|\Psi^*b\|_2} \\
&\leq \frac{\|u_l - u_l^P\|_S}{\|\Psi^*b\|_2} + \frac{\|f_P^l(\Lambda) \Psi^*r_k\|_2}{\|\Psi^*b\|_2} \\
&\leq \frac{\|u_l - u_l^P\|_S}{\|\Psi^*b\|_2} + \epsilon \|f_P^l(\Lambda)\|_2\\
&\leq \frac{\|u_l - u_l^P\|_S}{\|\Psi^*b\|_2} + \epsilon \left( \|f_P^l(\Lambda) - f^l(\Lambda)\|_2 + \|f^l(\Lambda)\|_2\right)\\
&\leq \frac{\|u_l - u_l^P\|_S}{\|\Psi^*b\|_2} + \epsilon\|f_P^l(\Lambda) - f^l(\Lambda)\|_2 + \epsilon \left( \max_{i=1,\ldots,N}\frac{1}{\vert \lambda_i-z_l \vert} \right).
\end{split}
\end{equation}
Finally, using the estimate~\eqref{eqn:HHTerror} along with~\eqref{eqn:utilde_error} yields the desired result. Once again, in the case where $S=I$ the error bound simplifies to
\begin{equation*}
\frac{\|u_l - \tilde{u}_l^P\|_2}{\|b\|_2} \leq \mathcal{O} \left( e^{-C P / \log\left(\lambda_1/\lambda_N\right)} \right) + \epsilon \left( \max_{i=1,\ldots,N}\frac{1}{\vert \lambda_i-z_l\vert} \right).
\end{equation*}
The error bound shows us that the error may be dominated by either the error in the pole expansion or the error in the solutions of~\eqref{eqn:pole_sub}. Since the error in the pole expansion decays exponentially, it is often best to control the overall error via the relative error requested when solving~\eqref{eqn:pole_sub}.
Given that we are interested in solving systems of the form~\eqref{eqn:pole_shift} for a large number of shifts, the key observation
in~\eqref{eqn:pole_approx} is that the vectors $h_k$ are independent of
the shifts $z_l$ because the $\xi_k$ are independent of $z_l.$ Therefore, this method parametrizes the solutions to $N_z$ linear systems of
the form~\eqref{eqn:pole_shift} on the solutions of $P$ independent
linear systems of the form~\eqref{eqn:pole_shift}. In fact, once this parametrization has been done the method is completely flexible and any method may be used to solve the sub-problems.
\begin{remark}
Because the $\xi_k$ appear in complex conjugate pairs, if $H,S$ and $b$ are real then so do the solutions to the systems~\eqref{eqn:pole_sub}. Therefore, in this situation we only have to solve $P/2$ systems, where for simplicity we assume $P$ is even.
\end{remark}
The bulk of the computational cost in this method is the
necessity of solving $P$ systems of the form~\eqref{eqn:pole_sub} after
which the vectors $h_k$ may be combined with different weights to yield
approximation solutions for as many distinct $z_l$ as desired. In fact, because the systems are completely independent this method
can be easily parallelized with up to $P$ (or $P/2$ depending on
whether symmetry is used) machines. Once the sub-problems have been solved, computing a solution for all $N_z$ shifts costs $O(PN_zN).$
Furthermore, as long as the $h_k$ are saved, solutions for new shifts
may be computed as needed with negligible computational cost. This is in
contrast to the Lanczos method where, as discussed in Section~\ref{sec:lanczos}, computing a solution for a new shift generally
requires running the algorithm again from the start unless all
$\{v_i\}$ are stored.
In some situations, the set of systems~\eqref{eqn:pole_sub} may even be
simultaneously solved using existing Krylov based methods for
simultaneously solving shifted systems, see, \textit{e.g.},
\cite{Frommer2003,FrommerGlassner1998,SimonciniSzyld2007,DattaSaad1991,Meerbergen2003,GallivanGrimmeVan1996,BaiFreund2001,FeldmannFreund1995,SABAKI2012}.
Perhaps the simplest example would be to use the MINRES variation of
the method outlined in Section~\ref{sec:lanczos}. In this situation
the same methods may be applicable to the original systems, however,
the pole expansion reduces the number of shifts that have to be solved
for from $N_z$ to $P.$ Similarly, if a good preconditioner is known
for each of the $P$ distinct systems then each system may be
independently solved via an iterative method.
If the systems are amenable to the use of a direct method, \textit{e.g.}, LU
factorization or factorizations as in \cite{MARROK05}, then the $N_z$ shifted systems may be solved by computing
factorizations of the matrices $(H-\xi_kS)$ and using those
factorizations to solve the sub-problems. If a direct method is used the
procedure also allows for efficiently solving shifted systems with
multiple right hand sides. The specific solution methodologies we used
for the sub-problems of the pole expansion method will be discussed in
Section~\ref{sec:numer}.
\section{Indefinite Systems}\label{sec:indefinite}
Up until this point we have considered Hermitian matrix \response{pencils} $(H,S)$ for which the generalized eigenvalues are all positive. Motivated by the applications we discuss in Section~\ref{sec:connection} we now discuss the case where there are both positive and negative generalized eigenvalues and we seek a solution of a specific form. We still require that $H$ and $S$ are Hermitian and that $S$ is positive definite. We assume that the generalized eigenvalues of $(H,S)$ are ordered such that
\begin{equation*}
\lambda_1 \geq \dots \geq \lambda_M > 0 > \lambda_{M+1} \geq \dots \geq \lambda_N.
\end{equation*}
Let $\Psi_{+} = [\psi_{1},\cdots,\psi_{M}]$ and $\Psi_{-} = [\psi_{M+1},\cdots,\psi_{N}].$ Similarly let $\Lambda_{+} =
\mathrm{diag}[\lambda_{1},\cdots,\lambda_{M}]$ and $\Lambda_{-} = \mathrm{diag}[\lambda_{M+1},\cdots,\lambda_{N}].$
While the notation used here mirrors that earlier in the paper, the generalized eigenvectors and eigenvalues here are distinct from the rest of the paper.
\subsection{Lanczos Method}
For the Lanczos based method used here there are two considerations that
have to be made with respect to indefinite systems. For the purposes of
this section we make the simplification, as in
Section~\ref{sec:lanczos}, that we first transform the problem in a
manner such that $S=I.$ Under this assumption we are interested in
solving systems for which the right hand side $\tilde{b}$ satisfies
$\Psi_{-}^*\tilde{b} = 0$ and the solution satisfies $\Psi_{-}^*x^l =
0.$ We note that in exact arithmetic $\Psi_{-}^*\tilde{b} = 0$ implies
that $\Psi_{-}^*x^l = 0$. However, we must ensure that the numerical
method used to solve these systems maintains this property.
Specifically, we need to ensure that $\Psi_{-}^*x^l_k \approx 0,$ and,
if we are using a CG style method, we must ensure that the sub-problems
remain non-singular. Both of these conditions are reliant upon the
mutual orthogonality between the Lanczos vectors and the columns of
$\Psi_{-}.$
In practice, we have observed that there is no excessive build up of
components in the $\Psi_{-}$ directions amongst the Lanczos vectors and
thus good orthogonality is maintained between the solutions we compute
using the Lanczos based methods and the negative generalized eigenspace.
Furthermore, in practice if we assume that $\Psi_{-}$ is known, then at
each step of the Lanczos process we may project out any components of
the Lanczos vectors that lie in the negative generalized eigenspace.
Such a procedure enforces orthogonality, up to the numerical error in
the projection operation, between the computed solution $x_k$ and
$\Psi_{-}.$
\subsection{\response{Pole expansion}}
The expansion we used in~\eqref{eqn:pole_approx} is valid for
Hermitian positive definite matrices. For
the case where the diagonal matrix of generalized eigenvalues $\Lambda$
has both positive and negative entries,
the pole expansion does not directly give accurate results.
However, the pole expansion is
still applicable in the case where we wish to solve the systems
projected onto the positive generalized eigenspace. Section~\ref{sec:connection} provides the motivation for considering such
problems. Similar to before, we are interested in solving systems for
which the right hand side $b$ satisfies $\Psi_{-}^*b = 0$ and the
solution satisfies $\Psi_{-}^*Su_l = 0.$ To accomplish this we use a
pole expansion using $\lambda_1$ and $\lambda_M$ as the bounds of the
spectrum. We note that in exact arithmetic $\Psi_{-}^*b = 0$ implies
that $\Psi_{-}^*Su_l = 0.$ However we must ensure that our numerical
methods retain this property.
Earlier in this section we discussed
the impact of this generalization on the Lanczos based solver.
Here we restrict our discussion to the impact of solving an indefinite
system on the pole expansion method and provide an argument for why we do not observe difficulty in this regime.
If we once again assume that $\tilde{h}_k$ satisfies~\eqref{eqn:pole_sub_error} we may conclude that
\begin{equation*}
\frac{\|u_l - \tilde{u}_l^P\|_S}{\|\Psi_{+}^* b\|_2} \leq \mathcal{O} \left( e^{-C P / \log\left(\lambda_1/\lambda_M\right)} \right) + \mathcal{O} \left( \epsilon \right) \left( \max_{i=1,\ldots,M}\frac{1}{\vert \lambda_i-z_l \vert} \right).
\end{equation*}
To argue such a bound, in a manner similar to before, we may write
\begin{equation*}
\begin{split}
\frac{\|u_l - \tilde{u}_l^P\|_S}{\|\Psi_{+}^*b\|_2} &= \frac{\|u_l - u_l^P + \Psi f_P^l(\Lambda) \Psi^*r_k\|_S}{\|\Psi^*b\|_2} \\
&\leq \frac{\|u_l - u_l^P\|_S}{\|\Psi^*b\|_2} + \frac{\|f_P^l(\Lambda) \Psi^*r_k\|_2}{\|\Psi_{+}^*b\|_2} \\
&\leq \frac{\|u_l - u_l^P\|_S}{\|\Psi^*b\|_2} + \frac{\| f_P^l(\Lambda_{+})\Psi_{+}^*r_k\|_2 + \| f_P^l(\Lambda_{-})\Psi_{-}^*r_k\|_2 }{\|\Psi_{+}^*b\|_2} \\
&\leq \frac{\|u_l - u_l^P\|_S}{\|\Psi^*b\|_2} + \epsilon \|f_P^l(\Lambda_{+})\|_2 + \frac{\| f_P^l(\Lambda_{-})\Psi_{-}^*r_k\|_2}{\|\Psi_{+}^*b\|_2}\\
&\leq \frac{\|u_l - u_l^P\|_S}{\|\Psi^*b\|_2} + \epsilon \left( \|f_P^l(\Lambda_{+}) - f^l(\Lambda_{+})\|_2 + \|f^l(\Lambda_{+})\|_2\right) + \epsilon \| f_P^l(\Lambda_{-}) \|_2.
\end{split}
\end{equation*}
Here we observe that the rational function $f_P^l$ is well behaved on the negative real axis in a manner dependent on the closest poles, which are of order $\lambda_M$ away. Furthermore, if the residuals $r_k$ are orthogonal to the negative general eigenspace this term vanishes. Also the construction of $f_P^l$ once again implies a dependence on the distance between the shifts and the generalized eigenvalues of the system. When using the pole expansion method on an indefinite system as long as the sub-problems are appropriately solved the use of the expansion basically maintains the accuracy of the overall solution. Furthermore, as long as the overall solution is computed accurately enough, and as long as $S$ is reasonably conditioned we cannot have large components of the computed solution in the negative generalized eigenspace. Furthermore, if necessary we may simply project out the components of $\tilde{h}_k$ in the negative generalized eigenspace to ensure that the overall solution error is not impacted by the sub-problem solution method.
\section{Connection with electronic structure calculation}\label{sec:connection}
In this section we discuss the connection between the problem with
multiple shifts in~\eqref{eqn:multishift} and several aspects
of the electronic structure theory, which are based on perturbative
treatment of Kohn-Sham density functional
theory~\cite{HohenbergKohn1964,KohnSham1965} (KSDFT).
KSDFT is the most
widely used electronic structure theory for describing the ground state
electronic properties of molecules, solids and other nano structures.
To simplify our discussion, we assume the computational domain is
$\Omega=[0,L]^3$ with periodic boundary conditions.
We use linear algebra notation, and we do not distinguish integral
operators from their kernels. For example, we may simply denote
$\hat{f}(\bvec{r})=\hat{A}[\hat{g}](\bvec{r})\equiv \int \hat{A}(\bvec{r},\bvec{r}')
\hat{g}(\bvec{r}')\,\mathrm{d} \bvec{r}'$ by $\hat{f}=\hat{A}\hat{g}$, and represent the
operator $\hat{A}$ by its kernel $\hat{A}(\bvec{r},\bvec{r}')$. All
quantities represented in the real space are given in the form such as
$\hat{H}(\bvec{r},\bvec{r}')$ and $\hat{f}(\bvec{r})$, and the corresponding matrix or
vector coefficients represented in a finite dimensional basis set is
given in the form such as $H$ and $f$.
The Kohn-Sham equation defines a nonlinear eigenvalue problem
\begin{equation}
\begin{split}
&\hat{H}[\hat{\rho}]\hat{\psi}_{i} = \varepsilon_{i} \hat{\psi}_{i},\\
&\hat{\rho}(\bvec{r}) = \sum_{i=1}^{N_e} \abs{\hat{\psi}_{i}(\bvec{r})}^2, \quad \int
\hat{\psi}^{*}_{i}(\bvec{r}) \hat{\psi}_{j}(\bvec{r}) \,\mathrm{d} \bvec{r} = \delta_{ij},
\end{split}
\label{eqn:KS}
\end{equation}
where $N_e$ is the number of electrons (spin degeneracy is omitted
here for simplicity). The eigenvalues
$\{\varepsilon_{i}\}$ are ordered non-decreasingly. The lowest
$N_{e}$ eigenvalues $\{\varepsilon_{i}\}_{i=1}^{N_{e}}$ are called the occupied state energies, and
$\{\varepsilon_{i}\}_{j>N_{e}}$ are called the unoccupied state
energies. We assume
$\varepsilon_{N_{e}+1}-\varepsilon_{N_{e}}>0$, \textit{i.e.}\ the system is an
insulating system~\cite{Martin2004}.
The eigenfunctions $\{\hat{\psi}_{i}\}_{i=1}^{N_e}$ define the
electron density $\hat{\rho}(\bvec{r})$, which in turn defines the Kohn-Sham
Hamiltonian
\begin{equation}
\hat{H}[\hat{\rho}] = -\frac12 \Delta + \hat{V}_{c}[\hat{\rho}] + \hat{V}_{\mathrm{xc}}[\hat{\rho}] + \hat{V}_{\mathrm{ion}}.
\label{eqn:ksham}
\end{equation}
Here $\Delta$ is the Laplacian operator for characterizing the kinetic
energy of electrons.
\begin{equation*}
\hat{V}_{c}[\hat{\rho}](\bvec{r}) \equiv
\int \frac{\hat\rho(\bvec{r}')}{\abs{\bvec{r}-\bvec{r}'}} \,\mathrm{d} \bvec{r}'
\label{}
\end{equation*}
is the Coulomb potential which is linear with respect to the electron
density $\hat{\rho}$.
$\hat{V}_{\mathrm{xc}}[\hat{\rho}]$ is a nonlinear
functional of $\hat{\rho}$, characterizing the many body
exchange and correlation effect. $\hat{V}_{\mathrm{ion}}$ is the electron-ion interaction
potential and is independent of $\hat{\rho}$. Because the eigenvalue problem
(\ref{eqn:KS}) is nonlinear, it is often solved iteratively by a class
of algorithms called self-consistent field iterations
(SCF)~\cite{Martin2004}, until~\eqref{eqn:ksham} reaches self-consistency.
When the self-consistent solution of the Kohn-Sham equation is obtained,
one may perform post Kohn-Sham calculations for properties within and
beyond the ground state properties of the system. Examples of such
calculations include the density functional perturbation theory
(DFPT)~\cite{BaroniGiannozziTesta1987,GonzeAllanTeter1992,BaroniGironcoliDalEtAl2001},
the GW
theory~\cite{Hedin1965,AryasetiawanGunnarsson1998,FriedrichSchindlmayr2006,UmariStenuitBaroni2010,PingRoccaGalli2013,GiustinoCohenLouie2010}
and the random phase approximation
(RPA) of the electron correlation energy~\cite{LangrethPerdew1975,LangrethPerdew1977,Furche2001,NguyenGironcoli2009}.
In these theories, a key quantity is the so called independent particle
polarizability matrix, often denoted by $\hat{\chi}_{0}(\omega)$. The
independent particle polarizability matrix characterizes the first order
\textit{non-self-consistent} response of the electron density $\delta
\hat{\rho}_{0}(\omega)$ with respect to the time dependent external
perturbation potential $\delta \hat{V}(\omega)$, where $\omega$ is the
frequency of the time dependent perturbation potential. $\omega$ can be
chosen to be $0$, characterizing the static linear response of the
electron density with respect to the static external perturbation
potential.
In the density functional perturbation theory, the first order
\textit{self-consistent} static response (\textit{i.e.}\ the physical response) of
the electron density $\delta \hat{\rho}(\bvec{r})$ with respect to the static
external perturbation potential $\delta \hat{V}$ can be computed as
\begin{equation*}
\delta \hat{\rho} = \hat{\chi}(0) \delta \hat{V},
\label{}
\end{equation*}
where the operator $\hat{\chi}(0)$ is directly related
to $\hat{\chi}_{0}(0)$ as
\begin{equation*}
\hat{\chi}(0) = \left[ I - \hat{\chi}_{0}(0) \left( \hat{V}_{c} + \frac{\delta
\hat{V}_{\mathrm{xc}}}{\delta \hat{\rho}} \right)
\right]^{-1} \hat{\chi}_{0}(0).
\label{}
\end{equation*}
Many body perturbation theories such as the GW
theory computes
the quasi-particle energy which
characterizes the excited state energy spectrum of the system.
The key step for calculating the quasi-particle energy is the
computation of the screened Coulomb operator, which is defined as
\begin{equation*}
\hat{W}(i\omega) = \left( I - \hat{V}_{c} \hat{\chi}_{0}(i\omega)
\right)^{-1} \hat{V}_{c}.
\label{}
\end{equation*}
Here $\chi_{0}(i\omega)$ should be computed on a large set of
frequencies on the imaginary axis $i\omega$.
The random phase approximation (RPA) of the electron correlation energy
improves the accuracy of many existing exchange-correlation functional in
ground state electronic structure calculation.
Using the adiabatic connection
formula~\cite{LangrethPerdew1975,LangrethPerdew1977}, the correlation
energy can be expressed as
\begin{equation*}
E_{c} = -\frac{1}{2\pi} \int_{0}^{1} \int_{0}^{\infty}
\mathrm{Tr}\left\{ \hat{V}_{c} [\hat{\chi}_{\lambda}(i\omega) -
\hat{\chi}_{0}(i\omega)] \right\}
\,\mathrm{d}\omega \,\mathrm{d} \lambda,
\label{}
\end{equation*}
and $\hat{\chi}_{\lambda}(i\omega)$ can be computed from
$\hat{\chi}_{0}(i\omega)$ as
\begin{equation*}
\hat{\chi}_{\lambda}(i\omega) = \hat{\chi}_{0}(i\omega) + \lambda
\hat{\chi}_{0}(i\omega)
\hat{V}_{c} \hat{\chi}_{\lambda}(i\omega).
\label{}
\end{equation*}
In all the examples above, computing $\hat{\chi}_{0}(i\omega)$ is
usually the bottleneck. For simplicity we consider the case
when $\hat{H}$ is real and therefore the eigenfunctions
$\hat{\psi}_{i}(\bvec{r})$ are also real. In such case the kernel
$\hat{\chi}_{0}(i\omega)(\bvec{r},\bvec{r}')$ can be computed from the Adler-Wiser
formula~\cite{Adler1962,Wiser1963} as
\begin{equation*}
\hat{\chi}_{0}(\bvec{r},\bvec{r}',i\omega) = 2 \Re \sum_{i \le
N_{e}, j > N_{e}}
\frac{\hat{\psi}_{i}(\bvec{r})\hat{\psi}_{j}(\bvec{r})\hat{\psi}_{j}(\bvec{r}')\hat{\psi}_{i}(\bvec{r}')}
{\varepsilon_{i}-\varepsilon_{j}+i\omega} .
\end{equation*}
The summation $\sum_{j > N_{e}}$ requires the computation of a
large number of eigenstates which is usually prohibitively expansive.
Recent
techniques~\cite{UmariStenuitBaroni2010,PingRoccaGalli2013,GiustinoCohenLouie2010}
have allowed to avoid the direct computation of
$\{\hat{\psi}_{j}\}_{j>N_{e}}$ when $\hat{\chi}_{0}(i\omega)$ is
multiplied to an arbitrary vector $\hat{g}$ as
\begin{equation}
\begin{split}
\hat{\chi}_{0}(i\omega)[\hat{g}](\bvec{r}) &= 2 \Re \sum_{i \le N_{e}}
\hat{\psi}_{i}(\bvec{r}) \int
\sum_{j > N_{e}}
\frac{\hat{\psi}_{j}(\bvec{r})\hat{\psi}_{j}(\bvec{r}')}{\varepsilon_{i}-\varepsilon_{j}
+ i\omega }
\hat{\psi}_{i}(\bvec{r}') \hat{g}(\bvec{r}')\,\mathrm{d} \bvec{r}'\\
& \equiv 2 \Re \sum_{i \le N_{e}}
\hat{\psi}_{i}(\bvec{r}) \hat{u}_{i}(\bvec{r}).
\end{split}
\label{eqn:chi0}
\end{equation}
Here $\hat{u}_{i}(\bvec{r})$ can be solved through the equation
\begin{equation}
\hat{Q} \left[\hat{H} - (\varepsilon_{i}+i\omega)\right] \hat{Q}
\hat{u}_{i} =
-\hat{Q}[\hat{\psi}_{i}\odot \hat{g}].
\label{eqn:multishiftOriginal}
\end{equation}
The operator $\hat{Q}(\bvec{r},\bvec{r}')=\delta(\bvec{r},\bvec{r}') - \sum_{i\le
N_{e}}\hat{\psi}_{i}(\bvec{r})\hat{\psi}^{*}_{i}(\bvec{r}')$ is a projection
operator onto the occupied states (noting
$\hat{\psi}^{*}_{i}(\bvec{r}')=\hat{\psi}_{i}(\bvec{r}')$ is real), and
$[\hat{\psi}_{i} \odot \hat{g}](\bvec{r}) \equiv
\hat{\psi}_{i}(\bvec{r})\hat{g}(\bvec{r})$ is the element wise product between two
vectors.
Without loss of generality we may set the largest occupied state energy $\varepsilon_{N_{e}}=0$.
Equation~\eqref{eqn:multishiftOriginal} can be
reduced to the form
\begin{equation}
(\hat{H} - z) \hat{u} = \hat{f},
\label{eqn:multishiftFine}
\end{equation}
for multiple shifts $z = \varepsilon_{i} + i\omega$ ($\Re z \le 0$) and multiple right
hand sides $\hat{f}$. Furthermore, $\hat{u},\hat{f} \in \text{Ran}(
\hat{Q})$ where $\text{Ran}$ is the range of the operator $\hat{Q}$.
Equation~\eqref{eqn:multishiftFine} can be solved in practice using a finite
dimensional basis set, such as finite element, plane waves, or
more complicated basis functions such as numerical atomic
orbitals~\cite{SolerArtachoGaleEtAl2002}.
We denote the basis set by a collection of column vectors as
$\hat{\Phi}=[\hat{\varphi}_{1}(\bvec{r}),\cdots,\hat{\varphi}_{N}(\bvec{r})]$.
The overlap matrix associated with the basis set $\hat{\Phi}$ is
\begin{equation*}
S = \hat{\Phi}^{*} \hat{\Phi},
\label{}
\end{equation*}
and the projected Hamiltonian matrix in the basis $\hat{\Phi}$ is
\begin{equation*}
H = \hat{\Phi}^{*} \hat{H} \hat{\Phi}.
\label{}
\end{equation*}
Using the ansatz that both the solution and the right hand side can be
represented using the basis set $\hat{\Phi}$ as
\begin{equation}
\hat{u} = \hat{\Phi} u,\quad \hat{f} = \hat{\Phi} f,
\label{eqn:ufbasis}
\end{equation}
\eqref{eqn:multishiftFine} becomes
\begin{equation*}
(H - z S) u = \hat{\Phi}^* \hat{f} = S f,
\label{}
\end{equation*}
which is~\eqref{eqn:multishift} with $b=S f$. Using the
eigen decomposition of the matrix \response{pencil} $(H,S)$ as in~\eqref{eqn:GE}, each eigenfunction in the real space
$\hat{\psi}_{i}(\bvec{r})$ is given using the basis set
$\hat{\Phi}$ as
\begin{equation}
\hat{\psi}_{i} = \hat{\Phi}\psi_{i}.
\label{eqn:psibasis}
\end{equation}
Combining Eqs.~\eqref{eqn:psibasis} and~\eqref{eqn:ufbasis}, the
condition $\hat{u}\in \text{Ran}(\hat{Q})$ becomes
\begin{equation}
\hat{\psi}_{i}^{*} \hat{u} = \psi_{i}^{*} S u = 0,\quad \forall i\le
N_{e},
\label{eqn:ucond}
\end{equation}
and similarly $\hat{f}\in \text{Ran}(\hat{Q})$ becomes
\begin{equation}
\psi_{i}^{*} S f = 0,\quad \forall i\le N_{e},
\label{eqn:fcond}
\end{equation}
In practice the conditions~\eqref{eqn:ucond} and~\eqref{eqn:fcond} can
be satisfied by a projection procedure as described in
Section~\ref{sec:indefinite}.
\section{Numerical results}\label{sec:numer}
First we illustrate the accuracy and the scaling of the pole expansion
method.
We then present two distinct
numerical examples based on the general method presented here. In one
example we consider the case where $S=I$ and an iterative method is used
to solve the sub-problems, and in the second example we consider the
case where $S \neq I$ and a direct method is used to solve the
sub-problems. In each case the method presented here is compared with
the Lanczos style method described in Section~\ref{sec:lanczos}.
All of the numerical experiments were run in MATLAB on a Linux machine
with four 2.0GHz eight core CPUs and 256GB of RAM.
\subsection{Accuracy of the pole expansion}
Before presenting the examples motivated by the preceding section, we
first demonstrate the behavior of the pole expansion method for
approximating
\begin{equation*}
f(x;z) = \frac{1}{x-z}
\end{equation*}
on some interval of the positive real axis.
Fig.~\ref{fig:contour_ex}
shows an example contour as computed via the method in Section~\ref{sec:pole} when the region of interest is $[1,100]$.
\begin{figure}
\centering
\begin{subfigure}[b]{.49\textwidth}
\includegraphics[width = 1\textwidth]{contour_ex.pdf}
\caption{View of the entire contour.}
\end{subfigure}
\begin{subfigure}[b]{.49\textwidth}
\includegraphics[width = 1\textwidth]{contour_ex_zoom.pdf}
\caption{View near the origin.}
\end{subfigure}
\caption{Example of quadrature nodes (blue $+$'s) used when the spectrum of $A$ (red line) lies in the interval $[1,100].$}
\label{fig:contour_ex}
\end{figure}
First we consider approximating $f(x;z)$ for a fixed value $z=i$ on the
interval $[1,1000].$
Fig.~\ref{fig:pole_err} shows the
$\|\cdot \|_{\infty}$ error of approximating $f(x;i)$ by $f_P(x;i),$
computed via sampling at $10000$ equally spaced points in $[1,1000]$
along the $x$ direction, as
the number of poles $P$ increases. We observe the
exponential decay in error as the number of poles increases, which is aligned with the error analysis presented earlier.
Next we illustrate the number of poles required to reach fixed accuracy
as $x$ approaches $0$, \textit{i.e.}\ as $E_{g}\to 0$.
We consider approximating $f(x;i)$ via $f_P(x;i)$ on the interval $[\sigma,10]$
for 20
values of $\sigma \in [10^{-4},1]$ that are equally spaced on the
logarithmic scale.
Fig.~\ref{fig:numpoles} shows the number of poles required such that
$\|f_P(x;i)-f(x;i)\|_{\infty}$ is less than $10^{-8}$ over the interval
$[\sigma,10].$
\begin{figure}
\centering
\begin{subfigure}[b]{.49\textwidth}
\includegraphics[width = 1\textwidth]{pole_err.pdf}
\caption{}
\label{fig:pole_err}
\end{subfigure}
\begin{subfigure}[b]{.49\textwidth}
\includegraphics[width = 1\textwidth]{numpoles.pdf}
\caption{}
\label{fig:numpoles}
\end{subfigure}
\caption{Behavior of the pole expansion. (a) $\|f_P(x;i)-f(x;i)\|_{\infty},$ on a $\log_{10}$ scale, for the interval $x\in [1,1000]$ as the number of poles, $P,$ is increased. (b) Number of poles required to achieve $\|f_P(x;i)-f(x;i)\|_{\infty} \leq 10^{-8}$ for $x\in [\sigma,10].$}
\end{figure}
Finally we consider the approximation of $f(x;z)$ for a wide range
of $z$ with $\Re z \leq 0$, and demonstrate how $\| f_P(x;z) - f(x;z)
\|_{\infty}$ varies as $z$ changes using the same pole expansion. We
fixed the number of poles used in the approximation to be 60. Fig.~
\ref{fig:pole_range} shows that high accuracy is maintained for all
$z$ in the left half plane. Here we kept the region of interest as $x\in [1,1000]$ and
consider the accuracy at 5000 distinct $z$ evenly distributed with $\Re z \in [-50,0]$ and $\Im z \in
[-50,50].$
\begin{figure}
\centering
\includegraphics[width = .75\textwidth]{pole_range.pdf}
\caption{$\|f_P(x;z)-f(x;z)\|_{\infty}$ on a $\log_{10}$ scale for the interval $x\in [1,1000]$ as $z$ is varied.}
\label{fig:pole_range}
\end{figure}
\subsection{Orthogonal basis functions: $S=I$}
First, we consider the case where orthogonal basis functions are used
and thus in the notation here $S=I.$ The Hamiltonian matrix $H$ takes
the form
\begin{equation*}
H = -\frac{1}{2}\Delta + V,
\end{equation*}
and the operator $V$ is obtained by solving the Kohn-Sham density
functional theory problem for a benzene molecule using the KSSOLV
package~\cite{YangMezaLeeEtAl2009}, which is a MATLAB toolbox for
solving Kohn-Sham equations for small molecules and solids implemented
entirely in MATLAB m-files. The benzene molecule has $30$ electrons and
$15$ occupied states (spin degeneracy of $2$ is counted here).
The atomic configuration of the benzene molecule is given in
Fig.~\ref{fig:benzene}.
The computational domain is $[0,20) \times [0,20)
\times [0,10)$ along the $x,y,z$ directions, respectively. The
computational domain is discretized into $64\times 64 \times 32 =
131072$ points. The Laplacian operator is discretized using the plane
wave basis set, and this set of grid corresponds to the usage of the kinetic
energy cutoff at $50.5$ Hartree. The Laplacian operator is applied using
spectral method, and is done efficiently using the Fast Fourier
Transform (FFT). The $15$ negative eigenvalues (occupied states) and the
corresponding eigenvectors are
computed using
the locally optimal block preconditioned conjugate gradient (LOBPCG)
method~\cite{Knyazev2001},
with a preconditioner of the form $(-\frac{1}{2}\Delta+.001)^{-1}.$
We only use these computed eigenvectors to ensure that the
right hand side for the set of equations we solve is orthogonal to the
negative eigenspace of the $(H,S)$ \response{pencil}, as required in
Section~\ref{sec:indefinite} for using the pole expansion for indefinite
systems.
We now solve the set of problems
\begin{equation}
\label{eqn:n1}
(H-i\eta_lI)x_l = b,
\end{equation}
for $101$ equispaced $\eta_l$ in the interval $[-10,10].$
Because $H$ and $b$ are real, some systems are in essence solved
redundantly. However, here we are interested in the performance given the
number of shifts and not the specific solutions.
We monitor the behavior of our method and the Lanczos method as the
condition number of $H$ increases. To this end we sequentially refine
the number of discretization points in each direction by a factor of $2$
and generate a potential function $V$ via Fourier interpolation. The
grid size is denoted by $N\times N\times N/2$, and the largest problem
we consider is discretized on a $256\times 256 \times 128$ grid. For
these large problems the eigenvectors associated with negative
eigenvalues are approximated by Fourier interpolates of the computed
eigenvectors for the smallest problem.
To compare the methods we solved the set of
problems~\eqref{eqn:n1} via the pole expansion method using a number of
poles that depended on the size of the problem. To combat the slight
loss of accuracy that occurs for a fixed number of poles as the
condition number of the matrix grows, we increased the number of poles as
the problem size grew. During this step preconditioned
GMRES \cite{Saad1986} in MATLAB
was used to solve the sub-problems associated with the pole expansion.
The preconditioner used \response{is} of the form
\begin{equation*}
\left( -\frac{1}{2}\Delta + \xi_k\right)^{-1},
\end{equation*}
and similarly to $H$ it was efficiently applied via the FFT. The
requested accuracy of the GMRES routine is that the relative residual
is less than $10^{-7}.$ Since $H$ and $b$ are real only $P/2$
sub-problems had
to be solved. We remark that though the solution for different poles can
be straightforwardly parallelized, here we performed the calculations sequentially
in order to compare with the sequential
implementation of the Lanczos method. Let $\tilde{x}_l^P$ denote the
approximate solutions computed using the pole expansion. The relative
error metric
\begin{equation*}
r_l^P = \frac{ \| b - (H - i\eta_l I )\tilde{x}_l^P \|_2 }{\| b \|_2}
\end{equation*}
is then computed for each shift after the approximate solutions have been computed. Finally, the Lanczos procedure described in Section~\ref{sec:lanczos} is called with a requested error tolerance of
\begin{equation}
\label{eqn:lanczos_stop}
\frac{ \| b - (H - i\eta_l I )\hat{x}_l\|_2 }{\| b \|_2} \leq \max_{l} r_l^P,
\end{equation}
where $\hat{x}_l$ denotes the approximate solution computed via the Lanczos method. Since the Lanczos method reveals the residual for each shift at each iteration this stopping criteria is cheap computationally. Thus, the Lanczos method was run until all of the approximate solutions met the stopping criteria~\eqref{eqn:lanczos_stop}. The Lanczos method used here takes advantage of the cheap updates briefly described in Section~\ref{sec:lanczos} and to further save on computational time, once a solution for a given $\eta_l$ was accurate enough the implementation stopped updating that solution. For comparison purposes we also solved the problem with a version of the Lanczos algorithm that uses Householder reflectors to maintain orthogonality amongst the Lanczos vectors \cite{GVL}. For the smallest size problem used here, and without any shifts, the version from Section~\ref{sec:lanczos} took 460 iterations to converge to $10^{-8}$ accuracy while the version that maintained full orthogonality took 456 iterations to converge to the same accuracy. However, the method that maintained full orthogonality took 65 times longer to run and would be prohibitively expensive for the larger problems given the increased problem size and iteration count so for all the comparisons here we use the CG style method outlined in Section~\ref{sec:lanczos}.
The spectrum of the operator $H$ grows as $\mathcal{O}(N^2)$, and we
observe that the number of iterations
required for the Lanczos method to converge grow roughly as $\mathcal{O}(N)$. We
do not expect to see such growth of the number of
iterations in the pole expansion method, given the preconditioner used
in solving the sub-problems. Furthermore, even though the cost per
iteration of solving for multiple shifts scales linearly in the Lanczos
method, the time required is dependent on the number of iterations
required to converge. In fact, in this case where $H$ may be applied
very efficiently for a very large number of shifts, the cost at each
iteration may be dominated by the additional computational cost
associated with each shift. In contrast, once the sub-problems have been
solved, the pole expansion method has a fixed cost for computing the
solutions for all the $\eta_l$, which only depends on the number of
shifts and the number of poles. If $\vert \eta_l \vert$ is large the
Lanczos method will converge very quickly for this specific problem, so
in the case where only shifts with large magnitude are considered the
Lanczos method may perform better. However, Section~\ref{sec:connection}
motivates our use of shifts spaced out along a portion of the imaginary
axis that includes 0.
Table~\ref{tab:numer1} reports the results of the pole
expansion method and the Lanczos method for the problem
described above. For the pole expansion method, the number of iterations
reported is the total number of iterations required to solve all of the
sub-problems. In both cases the error reported is the maximum computed
relative residual over the shifts. Finally, the total time taken to
solve the problems is reported for each method.
We observe that as expected the Lanczos method performs better for
the smallest sized problem. However once we reach the mid sized problem
the methods perform comparably, and the pole expansion method is
actually a bit faster even though it takes a few more iterations
overall. What is important to notice is that the overall iteration count
of solving all the sub-problems associated with the pole expansion
method remains relatively constant even with the increased number of
poles. By the time we reach the largest
problem the pole expansion method outperforms the Lanczos
method, taking about half the time to solve the set of problems.
\begin{remark}
Because the GMRES method used here uses Householder transforms to
maintain orthogonality amongst the Krylov basis it is not very memory
efficient. However, we also ran this example using preconditioned TFQMR \cite{TFQMR}
in place of GMRES, and while the iteration count more than doubled for
the pole expansion method it was actually faster since the applications
of the Householder matrices in GMRES is very expensive. This comparison
actually helps demonstrate the flexibility of the method with respect to
the solver used. If a fast method was not available for applying $H$
then GMRES may be preferable, however in the case where $H$ may be
applied very efficiently such as TFQMR may be better suited to
the problem.
\end{remark}
\begin{table}
\centering
\begin{tabular}{c | c | c | c | c | c | c | c | }
\cline{2-8}
& \multicolumn{3}{|c|}{Lanczos} & \multicolumn{4}{|c|}{Pole Expansion} \\ \hline
\multicolumn{1}{|c|}{Problem Size} & Iter. & Time(s) & Error & P & Iter. & Time(s) & Error \\ \hline
\multicolumn{1}{|c|}{$64\times 64 \times 32$} & 297 & 97.23 & 9.83 $\times 10^{-7}$ & 70 & 1005 & 181.99 & 9.85 $\times 10^{-7}$ \\ \hline
\multicolumn{1}{|c|}{$128\times 128 \times 64$} & 911 & 2073.89 & 6.91 $\times 10^{-6}$ & 80 & 969 & 1920.75 & 6.93 $\times 10^{-6}$ \\ \hline
\multicolumn{1}{|c|}{$256\times 256 \times 128$} & 1746 & 35129.79 & 6.83 $\times 10^{-6}$ & 90 & 959 & 17127.93 & 6.85 $\times 10^{-6}$ \\ \hline
\end{tabular}
\caption{Comparison of the pole expansion method and the Lanczos method for solving $(H-i\eta_lI)x_l = b.$}
\label{tab:numer1}
\end{table}
To further demonstrate the scaling of the methods with respect to the
number of shifts we ran the problem at a fixed size and varied the
number of shifts. As before, preconditioned GMRES was used in the pole
expansion method and no parallelism was used when solving the
sub-problems. The same strategy as above was used to ensure that the
Lanczos method stopped once it had solved all the problems as accurately
as the least accurate solution found using the pole expansion. Fig.~
\ref{fig:omega_small} shows the time taken to solve the problems of size
$64\times 64 \times 32$ for a varying number of $\eta_l$ equispaced in
$[-10,10].$ In all cases the largest relative residual was on the order
of $7\times 10^{-7}.$ Similarly, Fig.~\ref{fig:omega_large} shows the
time taken to solve the problems of size $128\times 128 \times 64$ for a
varying number of $\eta_l$ equispaced in $[-10,10].$ In all cases the
largest relative residual was on the order of $7\times 10^{-6}.$ Here we
observe that in both cases the pole expansion method scales very well as
the number of shifts increases, especially in comparison to the Lanczos
method. For example, even in the case where the problem is $64\times
64\times 32$ and the Lanczos method takes around a third of the
iterations of the pole expansion method, if you take the number of
shifts to be large enough the pole expansion method becomes considerably
faster.
\begin{figure}
\centering
\begin{subfigure}[b]{.49\textwidth}
\includegraphics[width = 1\textwidth]{benzene_omega_small.pdf}
\caption{$64\times 64 \times 32$}
\label{fig:omega_small}
\end{subfigure}
\begin{subfigure}[b]{.49\textwidth}
\includegraphics[width = 1\textwidth]{benzene_omega_large.pdf}
\caption{$128\times 128 \times 64$}
\label{fig:omega_large}
\end{subfigure}
\caption{Time taken to solve problems for a varying number of $\eta_l$ equispaced in $[-10,10]$}
\end{figure}
\subsection{Computation of $\hat{\chi}_{0}(i\omega)[\hat{g}](\bvec{r})$}
Motivated by our discussion in Section~\ref{sec:connection} we used our
technique to compute $\hat{\chi}_{0}(i\omega)[\hat{g}](\bvec{r})$ as defined
in~\eqref{eqn:chi0} where we only need $\hat{\psi}_i(\bvec{r})$ for $i\leq
N_e$. In order to compute $\hat{\chi}_{0}(i\omega)[\hat{g}](\bvec{r})$ for a
large number of $i\omega$ we had to compute $\hat{u}_i(\bvec{r})$ via
\eqref{eqn:multishiftOriginal}.
Physically this corresponds to the calculation of the response of the
electron density with respect to external perturbation potential
$\hat{g}(\bvec{r})$.
We constructed $\hat{g}$ as a Gaussian centered at one of the carbon
atoms in the benzene molecule. Fig.~\ref{fig:g} shows a slice along
the $z$ direction of the
function $\hat{g},$ and~\ref{fig:V} shows the corresponding slice of potential function
$V$ for the benzene
molecule. The pole expansion for computing the required quantities in
\eqref{eqn:multishiftOriginal} took 4400 seconds. This time encompasses
solving 15 sets of systems each with 200 shifts of the form
$\varepsilon_{i}+i\eta_{l}$, and each $\varepsilon_{i}$ corresponds to a distinct right
hand side. Computing the 15 smallest eigenvectors took 46.86 seconds.
Conversely, to use the alternative formula in
\eqref{eqn:multishiftOriginal} for computing
$\hat{\chi}_{0}(i\omega)[\hat{g}](\bvec{r})$ requires computing a large
number of additional eigenvectors. Even just computing the first 1000
eigenvectors took 6563 seconds using LOBPCG. Computing the first
2000 eigenvectors took 82448 seconds. Fig.~\ref{fig:slice} shows an
example of the solution in a slice of the domain for $\omega = 0.$
\begin{figure}
\centering
\begin{subfigure}[b]{.3\textwidth}
\includegraphics[width = 1\textwidth]{g_slice16.pdf}
\caption{Slice of $\hat{g}.$}
\label{fig:g}
\end{subfigure}
\begin{subfigure}[b]{.3\textwidth}
\includegraphics[width = 1\textwidth]{V_slice16.pdf}
\caption{Slice of $V.$}
\label{fig:V}
\end{subfigure}
\begin{subfigure}[b]{.3\textwidth}
\includegraphics[width = 1\textwidth]{w0_slice16.pdf}
\caption{Slice of $\hat{\chi}_{0}(0)[\hat{g}](\bvec{r}).$}
\label{fig:slice}
\end{subfigure}
\caption{Slices $\hat{g}$ and $\hat{\chi}_{0}(0)[\hat{g}](\bvec{r})$ at the vertical midpoint of domain}
\end{figure}
\subsection{Non-orthogonal basis functions: $S\ne I$}
We now consider the case where non-orthogonal basis functions are used
in the formulation of the problems as described in Section~\ref{sec:connection}. The matrices $H$ and $S$ are obtained by solving the Kohn-Sham density functional
theory problem for a DNA molecule with $715$ atoms using the SIESTA
package~\cite{SolerArtachoGaleEtAl2002} using the atomic orbital basis.
The atomic configuration of the DNA molecule is given in Fig.~\ref{fig:dna}.
The number of electrons is $2442$ and the number of
occupied states $N_{e}=1221$ (spin degeneracy of $2$ is counted here). The $1221$ negative eigenvalues and the corresponding eigenvectors were directly computed and were only used to ensure that the right hand sides for the set of equations we solve are orthogonal to the negative eigenspace of the $(H,S)$ \response{pencil}.
In this case the basis functions are non-orthogonal and therefore $S\ne I$. This means that we are now interested in solving problems of the form
\begin{equation}
\label{eqn:exeq}
(H - i\eta_l S)x_l = b,
\end{equation}
for a large number of equispaced $\eta_l$ in the interval $[-10,10].$ Furthermore, we are interested in solving the system for the same set of $\eta_l$ for multiple right hand sides $b.$ In this case both $H$ and $S$ are sparse and of size $7752 \times 7752.$ Therefore, the problem lends itself to the use of a direct method when solving the sub-problems for the pole expansion method as discussed in~\ref{sec:pole}. Since once again the problem is real, this means that we have to compute $P/2$ LU factorizations of matrices of the form
\begin{equation}
\label{eqn:exsub}
(H - \xi_k S).
\end{equation}
After this step has been completed the pole expansion method may be used to quickly solve problems of the form~\eqref{eqn:exeq}. First, we may compute the set of weights for each $\eta_l.$ Then, we may use the LU factorizations of~\eqref{eqn:exsub} to find the vectors $\tilde{h}_k.$ Finally, we may compute the solutions to the set of problems~\eqref{eqn:exeq} for a fixed $b$ using the computed weights. For each additional right had side we just need to compute a new set of $\tilde{h}_k$ and then combine them using the same weights as before. The marginal cost for each additional right had side is $P /2$ forward and backward substitutions plus the cost of combining $P$ vectors with $N_z$ distinct sets of weights. Comparatively, using the Lanczos method to solve for multiple right hand sides requires starting over because the Krylov subspace is dependent on $b.$ It is important to note that if the number of shifts is less than the number of poles used it would be more efficient to just factor the $N_z$ shifted systems. However, we are interested in the case where there are more shifts than poles even though some cases in the example do not reflect this situation.
\begin{figure}
\centering
\begin{subfigure}[b]{.49\textwidth}
\begin{center}
\includegraphics[width=.7\textwidth]{benzene.png}
\caption{}
\label{fig:benzene}
\end{center}
\end{subfigure}
\begin{subfigure}[b]{.49\textwidth}
\begin{center}
\includegraphics[width=1\textwidth]{ADNA_715.png}
\caption{}
\label{fig:dna}
\end{center}
\end{subfigure}
\caption{(a) The atomic configuration of a benzene molecule with $12$ atoms. (b) The atomic configuration of a DNA molecule with $715$ atoms.}
\end{figure}
This solution strategy is not dependent on computing LU factorizations
of~\eqref{eqn:exsub}. Any direct method may be used that allows for
rapid computation of solutions given a new right hand side. Therefore,
the pole expansion algorithm used in this manner has two distinct parts.
There is the factorization step where $P$ factorizations are computed,
and there is the solving step, where the factorizations are used to
solve $P$ sub-problems for each right hand side and then the solutions
are combined for all the desired $\eta_l.$ For a small number of shifts
and for very few right hand sides we still expect the Lanczos method to
potentially outperform the pole expansion method. However, as soon as we
have a large number of shifts, or there are enough right hand sides to
make the use of the direct methods favorable to an iterative method we
expect the pole expansion method to take much less time than the Lanczos
method.
Based on the splitting of the work for the pole expansion between a
factorization step and a solve step, we present the results for this
example slightly differently than before. The problems were solved for a
varying number of $\eta_l,$ denoted $N_z.$ For each set of $\eta_l$ the
problem was solved using both the pole expansion method and the Lanczos
method for $N_{\text{rhs}}$ distinct right hand sides via the
transformation in~\eqref{eqn:chol_eq} and we report the average time to
solve the problem for a single right hand side along with the time for
computing the Cholesky factorization of $S,$ denoted $T_c.$ Table
\ref{tab:lanczos} shows the results of using the Lanczos method for
solving the set of problems. We observe that the number of iterations is
consistent regardless of the number of shifts, and because the problem
is small the method scales well as the number of shifts grows. However,
for each right hand side the Lanczos method must start from scratch.
Table~\ref{tab:pe} shows the time taken to factor the $P /
2$ sub-problems of the pole expansion method and then the cost for
computing a solution for all the $\eta_l$ per right hand side. Similar to before, we did not take advantage of the parallelism in the method and simply computed the factorizations sequentially. Here we
observe that the bulk of the computation time is in the factorization
step, which is expected since this now behaves like a direct method.
However, once the factorizations have been computed the marginal cost of
forming the solutions for all of the desired shifts with a new right
hand side is minimal.
\begin{table}
\centering
\begin{tabular}{c c | c | c | c | c | }
\cline{3-6}
& & \multicolumn{4}{|c|}{Lanczos} \\ \hline
\multicolumn{1}{|c|}{ $N_z$ } & \multicolumn{1}{|c|}{ $N_{\text{rhs}}$ } & Avg. Iter. & $T_c(s)$ & Avg. Solve Time(s) per $b$ & Max Error \\ \hline
\multicolumn{1}{|c|}{3} & \multicolumn{1}{|c|}{20} & 171 & 1.33 & 102 & 3.88 $\times 10^{-9}$ \\ \hline
\multicolumn{1}{|c|}{11} & \multicolumn{1}{|c|}{20} & 173 & 1.33 & 103.8 & 4.03 $\times 10^{-9}$ \\ \hline
\multicolumn{1}{|c|}{101} & \multicolumn{1}{|c|}{20} & 173 & 1.33 & 114.8 & 4.25 $\times 10^{-9}$ \\ \hline
\multicolumn{1}{|c|}{1001} & \multicolumn{1}{|c|}{20} & 171 & 1.33 & 223.6 & 4.94 $\times 10^{-9}$ \\ \hline
\end{tabular}
\caption{Lanczos method for solving $(H-i\eta_lS)x_l = b.$}
\label{tab:lanczos}
\end{table}
\begin{table}
\centering
\begin{tabular}{c c | c | c | c | c | }
\cline{3-6}
& & \multicolumn{4}{|c|}{Pole Expansion} \\ \hline
\multicolumn{1}{|c|}{ $N_z$ } &\multicolumn{1}{|c|}{ $N_{\text{rhs}}$ } & P & Factor Time(s) & Avg. Solve Time(s) per $b$ & Max Error \\ \hline
\multicolumn{1}{|c|}{3} & \multicolumn{1}{|c|}{20} & 60 & 149.4 & 3.95 & 5.52 $\times 10^{-10}$ \\ \hline
\multicolumn{1}{|c|}{11} & \multicolumn{1}{|c|}{20} & 60 & 149.4 & 3.98 & 5.46 $\times 10^{-10}$ \\ \hline
\multicolumn{1}{|c|}{101} & \multicolumn{1}{|c|}{20} & 60 & 149.4 & 5.14 & 5.45 $\times 10^{-10}$ \\ \hline
\multicolumn{1}{|c|}{1001} & \multicolumn{1}{|c|}{20} & 60 & 149.4 & 16.2 & 5.56 $\times 10^{-10}$ \\ \hline
\end{tabular}
\caption{Pole expansion method for solving $(H-i\eta_lS)x_l = b.$}
\label{tab:pe}
\end{table}
When a direct method is an option for solving problems of the form
\eqref{eqn:exeq} and there are a large number of shifts, the direct
method may be combined with the pole expansion method to essentially
parametrize the factorizations of $N_z$ distinct matrices on the
factorizations of $P$ distinct matrices. If the number of shifts is much
larger than the number of poles required for the desired accuracy, this
reduction in the number of required factorizations turns out to be very
beneficial computationally. Furthermore, if memory is an issue, it is
possible to only ever store one factorization at a time. Specifically,
once a factorization is computed for a given pole, the vectors
$\tilde{h}_k$ may be computed for each right hand side and then the
factorization may be discarded. Overall, the combined use of the pole
expansion and an efficient direct method appears to be a very efficient
method for solving sets of parametrized linear systems for multiple
right hand sides, or, in some cases, even for a single right hand side.
\section{Conclusion}\label{sec:conclusion}
We have presented a new method for efficiently solving a type of
parametrized linear systems, motivated from electronic structure
calculations. By building a quadrature scheme based on the ideas in
\cite{HHT} we are able to represent the solutions of the parametrized
shifted systems as weighted linear combinations of solutions to a set of
fixed problems, where the weights vary based on the parameter of the
system. This method scales well as the number of distinct parameters for
which we want to solve~\eqref{eqn:multishift} grows. Furthermore,
because the solutions to the parametrized equations are based on a fixed
set of sub-problems there is flexibility in how the sub-problems are
solved. We presented examples using both iterative and direct solvers
within the framework for solving the shifted systems. The method
presented here can be more favorable compared to a Lanczos based method,
especially when solutions to a large number of parameters or a large
number of right hand sides are required.
\section*{Acknowledgments}
A. D. is currently supported by NSF Fellowship DGE-1147470 and was partially supported by the National Science Foundation grant DMS-0846501.
L. L. was partially supported by the Laboratory Directed Research and
Development Program of Lawrence Berkeley National Laboratory under the
U.S. Department of Energy contract number DE-AC02-05CH11231, and by
Scientific Discovery through Advanced Computing (SciDAC) program funded
by U.S. Department of Energy, Office of Science, Advanced Scientific
Computing Research and Basic Energy Sciences. L. Y. was partially
supported by National Science Foundation under award DMS-0846501 and by
the Mathematical Multifaceted Integrated Capability Centers (MMICCs)
effort within the Applied Mathematics activity of the U.S. Department of
Energy's Advanced Scientific Computing Research program, under Award
Number(s) DE-SC0009409. We would like to thank Lenya Ryzhik for providing computing resources. We are grateful to Alberto Garcia and Georg Huhs for providing the atomic configuration for the DNA molecule. L. L. thanks Dario Rocca and Chao Yang for helpful discussions, and thanks the
hospitality of Stanford University where the idea of this paper started.
\bibliographystyle{siam}
|
\section{Introduction}
\label{intro}
Recent advances in mobile technologies have paved the way for a novel paradigm for achieving large-scale city-wide sensing known as Participatory Sensing \cite{Burke}. In Participatory sensing, the key idea is to recruit ordinary people to contribute in sensor data collection using their mobile phones. This revolutionary paradigm has been operationally used to crowdsource information ranging from personal health \cite{DietSense} and prices of consumer goods \cite{PetrolWatch} to environment monitoring \cite{EarPhone}.\\
As a crowdsensing platform, a key challenge in the success of participatory sensing is the recruitment of sufficient participants. Typically, participatory sensing campaigns rely on voluntary contributors without any explicit incentives for participation. The lack of adequate motivation may result in few participants which in turn, reduces the data reliability. Another challenge is the suitability of participants particularly for those tasks which require domain-specific knowledge or expertise \cite{reddyrecruitment}.\\
To address the aforementioned challenges, one proposed idea is to employ online social networks as the underlying substrate for recruiting well-suited contributors \cite{integrate, integratemain}. This marriage of participatory sensing and online social networks, referred to as \emph{social participatory sensing}, offers the following advantages. First, the identification of suitable participants can be done easily through the public profile information such as interests, expertise and education. Second, social ties can act as an effective motivation to contribute to tasks created by friends, since people normally like to be helpful to their friends. Third, it is possible to offer incentives in the form of e-coins \cite{ecash} or reputation points which can be published in participants' profile and seen by others. A real-world instantiation of social participatory sensing was recently presented in \cite{twitter}, wherein, Twitter was used as the underlying social network substrate. The proposed system was tested in the context of two applications: weather radar and noise mapping. Their experiment resulted in a considerable smartphone-based participation from Twitter members even without an incentive structure. This clearly demonstrated the suitability of online social networks as a publish-subscribe infrastructure for tasking/utilizing smartphones and pave the way for ubiquitous crowd-sourced sensing and social collaboration applications.
The open nature of participatory sensing which allows everyone to contribute, while valuable for encouraging participants, facilitates erroneous and untrusted data preparation. When combined with social network, new trust issues arise. For instance, following the devastation incurred due to Hurricane Sandy in the US in October 2012, social media was flooded with misinformation and fake photos \footnote{http://news.yahoo.com/10-fake-photos-hurricane-sandy-075500934.html}. While some of these were easy to identify as fake data (e.g., photoshopped images of sharks swimming in New York streets), several other fake pictures and reports were initially thought to be true.
In fact, the widespread use of social networks, along with fast and easy-to-use dissemination facilities such as re-sharing (a fake photo) or re-tweeting (a false event) make it difficult to identify the origin of the data and investigate its credibility. This clearly highlights the need for a trust system which is responsible for performing necessary validations both from the perspective of data trustworthiness and also the reliability of data contributors. In other words, it is important to know who and with what level of social trustworthiness produces the data and how much of the data can be trusted. While there exist works that address the issue of data trustworthiness in participatory sensing (see Section \ref{rel}), they do not provide means to include social trust and as such cannot be readily adopted for social participatory sensing.\\
In this paper, we present an application agnostic framework to evaluate trust in social participatory sensing systems. Our system independently assesses the quality of the data and the trustworthiness of the participants and combines these metrics using fuzzy logic to arrive at a comprehensive trust rating for each contribution. These trust ratings are then used to calculate and update the reputation score of participants. By adopting a fuzzy approach, our proposed system is able to concretely quantify uncertain and imprecise information, such as trust, which is normally expressed by linguistic terms rather than numerical values. We undertake extensive simulations to demonstrate the effectiveness of our reputation framework and benchmark against the state-of-the-art. The results demonstrate that considering social relations makes trust evaluation more realistic, as it resembles human behaviour in establishing trustful social communications. We also show that our framework is able to quickly adapt to rapid changes in the participant's behaviour by prompt and correct detection and revocation of unreliable contributions and accurate update of participant's reputation score. Moreover, we find that leveraging fuzzy logic provides considerable flexibility in combining the underlying components which leads to better assessment of the trustworthiness of contributions. Our framework results in a considerable increase in the overall trust over a method which solely associates trust based on the quality of contribution.
The rest of the paper is organised as follows. Related work is discussed in Section \ref{rel}. We present the details of our framework in Section \ref{pro}. Simulation results are discussed in Section \ref{sim}. Finally, Section \ref{con} concludes the paper.
\begin{figure}
\centering
\includegraphics[width=12cm]{framework.eps}
\caption{Reputation framework architecture}
\label{fig:frm}
\end{figure}
\section{Related Work}
\label{rel}
To the best of our knowledge, the issue of trust in social participatory sensing hasn't been addressed in prior work. As such, we discuss about related research focussing on trust issues in participatory sensing.\\
In a participatory sensing system, trustworthiness can be viewed as the quality of the sensed data. In order to ascertain the data trustworthiness, it is highly desirable to ascertain that the sensor data has been captured from the said location and at the said time. \cite{Lenders} has proposed a secure service which allows participants to tag their content with a spatial timestamp indicating its physical location, which is later used by a co-located infrastructure for verification. A similar approach has been proposed in \cite{LocProof}, in the form of a small piece of metadata issued by a wireless infrastructure which offers a timestamped signed location proof. Since these works rely on external infrastructure, they have limited scalability. Moreover, neither approach will work in situations where the infrastructure is not installed. In our proposed framework, we assume that sensor data is tagged with GPS coordinates/system time before being stored in phone memory, which is then used by trust server for verification. Data trustworthiness has been investigated from another point of view which tries to confirm that uploaded data preserves the characteristics of the original sensed data and has not been changed unintentionally or maliciously. In particular, there are several works which make use of Trusted Platform Module (TPM)\cite{TPM}, which is a micro-controller embedded in the mobile device and provides it with hardware-based cryptography as well as secure storage for sensitive credentials. In \cite{Dua}, each device has a trusted hardware element that implements cryptographic algorithms for content protection. \cite{Imasensor} presents two TPM-based design alternatives: the first architecture relies on a piece of trusted code and the second design incorporates trusted computing primitives into sensors to enable them sign their readings. YouProve \cite{YouProve} is another TPM-based system that allows client applications to directly control the fidelity of data they upload and services to verify that the meaning of source data is preserved.
However, TPM chips are yet to be widely adopted in mobile devices. There is also recent work that does not require TPM. \cite{RFSN} proposes a reputation-based framework which makes use of Beta reputation \cite{beta} to assign a reputation score to each sensor node in a wireless sensor network. Beta reputation has simple updating rules as well as facilitates easy integration of ageing. However, it is less aggressive in penalizing users with poor quality contributions. A reputation framework for participatory sensing was proposed in \cite{Brian}. A watchdog module computes a cooperative rating for each device according to its short-term behaviour which acts as input to the reputation module which utilizes Gompertz function \cite{gompertz} to build a long-term reputation score. Their results show an improvement over the non-trust aggregation based approaches and Beta reputation system. However, the parameters related to the participants' social accountability have not been considered. As such, their system cannot be readily used in our context.
\section{Fuzzy Trust Framework}
\label{pro}
In this section, we explain the proposed framework for evaluating trust and reputation in social participatory sensing systems. An overview of the architecture is presented in Section \ref{arc} followed by a detailed discussion of each component in Section \ref{comp}.
\subsection{Framework Architecture}
\label{arc}
Since our framework attempts to mimic how human's perceive trust, we first present a simple illustrative example. Suppose John is a member of an online social network (e.g., Facebook). He has made a profile and has friended several people. John is a vegetarian and is the member of several vegetarian social communities. He is also on a budget and is keen to spend the least possible amount for his weekly groceries. He decides to leverage his social circle to find out the cheapest stores where he can buy vegetarian products. Specifically, he asks his friends or community members to capture geotagged photos of price labels of vegetarian food items when they are out shopping and send these back to him. One of his friends, Alex decides to help out and provides him with several photos of price labels. In order to decide whether to rely on Alex's contributions, John would naturally take into account two aspects: (i) his personal trust perception of Alex, which would depend on various aspects such as the nature of friendship (close vs. distant), Alex's awareness of vegetarian foods, Alex's location, etc and (ii) the quality of Alex's data which would depend on the quality of the pictures, relevance of products, etc. In other words, John in his mind computes a trust rating for Alex's contribution based on these two aspects. Our proposed trust framework provides a means to obtain such trust ratings by mimicking an approach similar to John's perception of trustworthiness in a scalable and automated manner. This trust rating helps John to select trustable contributions and accordingly plan for his weekend shopping.
Moreover, the trust server provides a reputation score for each of the participating friends, according to the trustworthiness of their successive contributions.
\noindent Fig. \ref{fig:frm} illustrates the architecture of the proposed reputation framework. The social network serves as the underlying publish-subscribe substrate for recruiting friends as participants. In fact, the basic participatory sensing procedures (i.e., task distribution and uploading contributions) are performed by utilizing the social network communication primitives. A person wishing to start a participatory sensing campaign acts as a requester and disseminates the tasks to his friends via email, private message or by writing as a post on their profiles (e.g., Facebook wall). Friends transfer their contributions via email or in the form of a message. We can also benefit from group construction facilities in Facebook or community memberships in Google Plus. The contributions received in response to a campaign are transferred (e.g., by using Facebook Graph API\footnote{http://developers.facebook.com/docs/reference/api/}) to a third party trust server, which incorporates the proposed fuzzy inference system and arrives at an objective trust rating for each contribution. This trust rating is used as a criterion to accept the contribution or revoke it, by comparing against a predefined threshold.
At the end of each campaign, a cumulative objective trust rating, referred to as $Trust_{RP}$ is automatically updated for each participant, which denotes the trustworthiness degree of Requester upon the Participant. $Trust_{RP}$ is dependent on the trustworthiness of the contribution that the participant has prepared for the requester.
For certain campaigns, depending on the nature of task, the requester may desire to add a subjective evaluation in order to indicate how much the contribution is compatible with his needs and expectations. In such a case, this subjective rating is combined with the system-computed rating to update $Trust_{RP}$.
At regular intervals, a reputation score is also calculated for each participant, which is a combination of the trust ratings that requesters have assigned to him. This reputation score is further used as a weight for participant's evaluations, ratings or reviews. More details about trust update, subjective rating and reputation calculation are presented in Section \ref{repsec}.
\subsection{Framework Components}
\label{comp}
This section provides a detailed explanation of the framework components. In particular we focus on the trust sever, fuzzy inference system and reputation module.
\subsubsection{Trust Server}
\label{server}
The trust server is responsible for maintaining and evaluating a comprehensive trust rating for each contribution and calculating a reputation score for each participant. As discussed in Section \ref{intro}, there are two aspects that need to be considered: (1) Quality of Contribution (QoC) and (2) Trust of Participant (ToP). The server maintains a trust database, which contains the required information about participants and the history of their past contributions. When a contribution is received by the trust server, the effective parameters that contribute to the two aforementioned components are evaluated by the Evaluator and then combined to arrive at a single quantitative value for each. The two measures serve as inputs for the fuzzy inference system, which computes the trustworthiness of contribution. In the following, we present a brief discussion about the underlying parameters and the evaluation methods.\\
\noindent\underline{Quality of Contribution (QoC)}\\
In participatory sensing, contributions can be of any modality such as images or sounds. The quality of the data is affected not only by fidelity of the embedded sensor but also the sensing action initiated by the participant. The in-built sensors in mobile devices can vary significantly in precision. Moreover, they may not be correctly calibrated or even worse not functioning correctly, thus providing erroneous data. Participants may also use the sensors improperly while collecting data,(e.g., not focussing on the target when capturing images). Moreover, human-as-sensor applications such as weather radar in \cite{twitter} are exposed to variability in the data quality due to subjectivity. For example, what is hot for one person may be comfortable for another. In order to quantify QoC, a group of parameters must be evaluated such as: relevance to the campaign (e.g., groceries in the above example), ability in determining a particular feature (e.g., price tag), fulfilment of task requirements (e.g., specified diet restrictions), etc. There already exists research that has proposed methods for evaluating the quality of data in participatory sensing. Examples include image processing algorithms proposed in \cite{DietSense} and outlier detection \cite{outlier} for sound-based sensing tasks. Rather than reinventing the wheel, our system relies on these state-of-the-art methods for determining the QoC.\\
\noindent\underline{Trust of Participant (ToP)}\\
\begin{figure}
\centering
\includegraphics[width=12cm]{gompertz.eps}
\caption{(a) Gompertz function for Friendship score \,\,\,\,\, (b) Inverse Gompertz function for time gap score}
\label{fig:gompertz}
\end{figure}
ToP is a combination of personal and social factors. Personal factors consist of the following parameters:\\
\emph{Expertise(E):}
It is defined as the measure of a participant's knowledge and is particularly important in tasks that require domain expertise. Greater credence is placed in contributions made by a participant who has expertise in the campaign. We propose to use expert finding systems for evaluating expertise. These systems employ social networks analysis and natural language processing (text mining, text classification, and semantic text similarity methods) to analyse explicit information such as public profile data and group memberships as well as implicit information such as textual posts to extract user interests and fields of expertise \cite{expert}. In particular, Dmoz\footnote{http://www.dmoz.org} open directory project is used for expertise classification. Expertise evaluation is done by incorporating text similarity analysis to find a match between the task keywords (e.g., vegetarian) and participant's expertise.
We assume that the set \emph{TE} contains the Task's required Expertise and \emph{PE} is the set of Participant's Expertises. In this case, the expertise score of each participant is defined as Eq. \ref{eq:eqE}:
\begin{equation}
\label{eq:eqE}
E= \frac{n(TE\cap PE)}{n(TE)}
\end{equation}
where \emph{n(A)} is the number of elements in set A.\\
\emph{Timeliness(T):}
Timeliness measures how promptly a participant performs prescribed tasks. It depends on the contribution response time \emph{(t)} and the task deadline \emph{(d)}. To evaluate this parameter, inverse Gompertz function defined as \begin{math} T(t)= 1 - e^{-be^{-ct}} \end{math} is used because of its compatibility with timeliness evolution: timeliness score is highest when the contribution is received immediately after the task release time. The score begins to decrease as the response time increases, reaching the minimum value when the response is received just before the deadline. In the original inverse Gompertz function, the lower asymptote is zero; it means that the curve approaches to zero in infinity. In our case, timeliness rate will only be zero if contribution is received after the deadline; otherwise, a value between \emph{x} and 1 is assigned to it. It means that the lowest timeliness rating will be \emph{x} if contribution is received before the deadline, and is zero if received after the deadline. So, we modify the function as Eq. \ref{eq:eq1} to calculate the timeliness (T):
\begin{equation}
\label{eq:eq1}
T(t)= \begin{cases}
1 - [(1 - x)e^{-be^{-ct}}] & \text{ if } t<d \\
0& \text{otherwise}
\end{cases}
\end{equation}
\noindent\emph{Locality(L):}
Another significant parameter is locality, which is a measure of the participant's familiarity with the region where the task is to be performed. We argue that contributions received from people with high locality to the tasking region are more trustable than those received from participants who are not local, since the first group is more acquainted with and has better understanding of that region. According to the experimental results presented in \cite{shirazi}, people tend to perform tasks that are near to their home or work place (places that they are considered `local' to them). This implies that if we log the location of participants' contributions, we can estimate their locality. A participant's locality would be highest at locations from where they make maximum number of contributions.
In order to evaluate locality, we assume that the sensing area has been divided to \emph{n} regions, and a vector \emph{V} with the length equal to \emph{n} is defined for each participant, where, \emph{V(i)} is number of samples collected in region \emph{i}. In this case, locality of a participant to region \emph{i} is calculated by Eq. \ref{eq:eq2}:
\begin{equation}
\label{eq:eq2}
L(i)= \frac{V(i)}{\sum_{i=0}^{n-1} V(i)}
\end{equation}
\begin{table*}
\centering
\caption{ Fuzzy rule base for defining ToC according to QoC and ToP}\label{tab:rule}
\begin{tabular}{|c|c|c|c||c|c|c|c|}\hline
\scriptsize{Rule no.} & \scriptsize{if QoC} & \scriptsize{and ToP} & \scriptsize{Then ToC} & \scriptsize{Rule no.} & \scriptsize{if QoC} & \scriptsize{and ToP} & \scriptsize{Then ToC} \\\hline
\scriptsize{1} & \scriptsize{Low} & \scriptsize{Low} & \scriptsize{VL} & \scriptsize{9} & \scriptsize{Med2} & \scriptsize{Low} & \scriptsize{M} \\\hline
\scriptsize{2} & \scriptsize{Low} & \scriptsize{Med1} & \scriptsize{L} & \scriptsize{10} & \scriptsize{Med2} & \scriptsize{Med1} & \scriptsize{H} \\\hline
\scriptsize{3} & \scriptsize{Low} & \scriptsize{Med2} & \scriptsize{L} & \scriptsize{11} & \scriptsize{Med2} & \scriptsize{Med2} & \scriptsize{H}\\\hline
\scriptsize{4} & \scriptsize{Low} & \scriptsize{High} & \scriptsize{M} & \scriptsize{12} & \scriptsize{Med2} & \scriptsize{High} & \scriptsize{H}\\\hline
\scriptsize{5} & \scriptsize{Med1} & \scriptsize{Low} & \scriptsize{L} & \scriptsize{14} & \scriptsize{High} & \scriptsize{Low}& \scriptsize{H}\\\hline
\scriptsize{6} & \scriptsize{Med1} & \scriptsize{Med1} & \scriptsize{L} & \scriptsize{14} & \scriptsize{High} & \scriptsize{Med1}& \scriptsize{H}\\\hline
\scriptsize{7} & \scriptsize{Med1} & \scriptsize{Med2} & \scriptsize{M} & \scriptsize{15} & \scriptsize{High} & \scriptsize{Med2}& \scriptsize{VH}\\\hline
\scriptsize{8} & \scriptsize{Med1} & \scriptsize{High} & \scriptsize{M} & \scriptsize{16} & \scriptsize{High} & \scriptsize{High}& \scriptsize{VH}\\\hline
\end{tabular}
\end{table*}
Next, we explain the social factors that affect ToP:\\
\emph{Friendship duration(F):}
In real as well as virtual communications, long lasting friendship relations normally translate to greater trust between two friends. So, friendship duration which is an estimation of friendship length is a prominent parameter in trust development. We use the Gompertz function depicted in Fig. \ref{fig:gompertz}(a) to quantify friendship duration, since its shape is a perfect match for how friendships evolve. Slow growth at start resembles the friendship gestation stage. This is followed by a period of accumulation where the relationship strengthens culminating in a steady stage. As such, the friendship duration is evaluated according to Eq. \ref{eq:eq3}, in which, \emph{b} and \emph{c} are system-defined constants and \emph{t} is the time in years.
\begin{equation}
\label{eq:eq3}
F(t)= e^{-be^{-ct}}
\vspace{-1mm}
\end{equation}
\noindent\emph{Interaction time gap(I):}
In every friendship relation, interactions happen in form of sending requests and receiving responses. Interaction time gap, measures the time between the consequent interactions and is a good indicator of the strength of friendship ties. If two individuals interact frequently, then it implies that they share a strong relationship, which translates to greater trust.
We propose to use the inverse Gompertz function depicted in Fig. \ref{fig:gompertz}(b) to quantify the interaction time gap, since a smaller time gap indicates stronger relationship, which leads to high social trust and vice-versa. So, the interaction time gap is evaluated according to Eq. \ref{eq:eq4}, in which, \emph{b} and \emph{c} are system-defined constants and \emph{t} is the gap (in days) between the current time and the Latest Interaction(LI) time.
\begin{equation}
\label{eq:eq4}
I(t)= 1-e^{-be^{-ct}}
\end{equation}
The aforementioned parameters are combined by the Evaluator to arrive at a single value for ToP, as depicted in Eq. \ref{eq:top},
\begin{equation}
\label{eq:top}
ToP= w_1\times E + w_2 \times T + w_3\times L + w_4 \times F + w_5 \times I
\end{equation}
where, \emph{$w_{i}$} is the weight of each parameter, and $\sum_{i=1}^{5}(w_i)$ equals to 1. The adjustment of the weights depends on the nature of the task. For example, in location-based tasks, $w_3$ is set to be considerably high to give more impression to Locality parameter. Similarly, for tasks where real-time information is important, a higher weight may be associated with Timeliness ($w_2$).
\subsubsection{Fuzzy inference system}
\label{fuzzy}
Our proposed framework employs fuzzy logic to calculate a comprehensive trust rating for each contribution, referred to as the Trust of Contribution (ToC). We cover all possible combinations of trust aspects and address them by leveraging fuzzy logic in mimicking the human decision-making process. The inputs to the fuzzy inference system are the crisp values of QoC and ToP. In the following, we describe the fuzzy inference system components.
\noindent\emph{Fuzzifier:}
The fuzzifier converts the crisp values of input parameters into a linguistic variable according to their membership functions. In other words, it determines the degree to which these inputs belong to each of the corresponding fuzzy sets. The fuzzy sets for QoC, ToP and ToC are defined as: \\ T(QoC)=T(ToP)=\{Low, Med1, Med2, High\}\\
T(ToC)= \{ VL, L, M, H, VH\}.\\
For any set $X$, a membership function on $X$ is any function from $X$ to the real unit interval [0,1]. The membership function which represents a fuzzy set $A$ is usually denoted by $\mu_{A}$. The membership degree $\mu_{A}(x)$ quantifies the grade of membership of the element $x$ to the fuzzy set $A$. The value 0 means that $x$ is not a member of the fuzzy set; the value 1 means that $x$ is fully a member of the fuzzy set. The values between 0 and 1 characterize fuzzy members, which belong to the fuzzy set only partially.\\
Fig.\ref{fig:mfinput} represents the membership function of QoC and ToP and Fig.\ref{fig:mfoutput} depicts the ToC membership function. We used trapezoidal shaped membership functions since they provide adequate representation of the expert knowledge, and at the same time, significantly simplify the process of computation.
\begin{figure}[!h]
\centering
\subfigure[Membership function for QoC and ToP]
{
\includegraphics[height=1in, width=2.5in]{mf1.eps}
\label{fig:mfinput}
}
\hspace{0.1cm}
\subfigure[Membership function for ToC]
{
\includegraphics[height=1in,width=2.5in]{mf2.eps}
\label{fig:mfoutput}
}
\caption{Membership functions of input and output linguistic variables}
\label{fig:mf}
\end{figure}
\noindent\emph{Inference Engine:}
The role of inference engine is to convert fuzzy inputs (QoC and ToP) to the fuzzy output (ToC) by leveraging If-Then type fuzzy rules. The combination of the above mentioned fuzzy sets create 4*4=16 different states which have been addressed by 16 fuzzy rules as shown in Table \ref{tab:rule}. Fuzzy rules help in describing how we balance the various trust aspects. The rule based design is based on the experience and beliefs on how the system should work. To define the output zone, we used \emph{max-min} composition method as: \begin{math} \mu_{T(ToC)}(ToC)= max[\underset{\substack{X \in T(ToP),\\Y\in T(QoC)}}{min}(\mu_{X}(ToP), \mu_{Y}(QoC))] \end{math}.
The result of the inference engine is the ToC which is a linguistic fuzzy value.
\noindent\emph{Defuzzifier:}
A defuzzifier converts the ToC fuzzy value to a crisp value in the range of [0, 1]. We employed the Centre of Gravity (COG) \cite{cog} defuzzification method, which computes the center of gravity of the area under ToC membership function. COG is perhaps the most commonly used and popular defuzzification technique with the advantage of quick and highly accurate computations.
\subsubsection{Reputation Module}
\label{repsec}
Once the ToC is defined for a contribution, the corresponding requester-participant mutual trust is updated, which is then used to calculate/update the participant's reputation score. In the following, we describe these steps in details:\\
As mentioned before, for some tasks, it is desirable for the requester to assign a subjective rating to participants' contributions. This is particularly relevant for campaigns where it is difficult for the requester to express his real needs, desires or restrictions via task definition. Subjective rating is also useful when the requester does not have enough knowledge about the task and needs an expert review to confirm the validity of the contributions. For example, assume a requester with a strict gluten-free diet who asks his friends to take photos from the price tag and ingredients of gluten-free products. The term gluten-free is generally used to indicate a supposedly harmless level of gluten rather than a complete absence. For those with serious celiac disease, the maximum safe level of gluten in a finished product is even lower than the amount that exists in normal gluten-free products. So, a double check with product ingredients is essential to be performed either by the requester himself or by a nutritionist to assure that it is safe to be consumed.
To be brief, although the objective rating assigned to a contribution is perfect for many tasks, sometimes, a subjective rating is added to reassure the conformance of contribution to the specific needs of requester. In such a case, the need for such subjective evaluation is defined by the requester in the task definition step.
We denote the subjective rating as Requester Evaluation (RE) which implies the trustworthiness of contribution from the requester's point of view. Although RE value can be in any range, in our simulation in Section \ref{sim}, we assume that RE has a value in the range of $(ToC-\mu , ToC+\mu)$, where $\mu$ = 1- $\rho_{Req}$ and $\rho_{Req}$ is the requester's reputation score. For a requester with high reputation score, the value of $\mu$ is small, resulting in RE close to ToC. It means that a requester with high reputation score is likely to assign a rating, which is close to the system-computed rating.
In the absence of subjective ratings, the requester simply relies on the objective ratings assigned by the trust server. In this case, $\mu$ is simply set to zero, resulting in RE=ToC.
Based on the ToC assigned to each contribution, the trust of requester upon the corresponding participant ($Trust_{RP}$) is updated. In fact, we adopt a reward/penalty policy for this update. A participant with ToC values greater than a predefined threshold1$(Th_1)$ is rewarded, and the amount of $\left |ToC-\rho_{Req}*RE \right |$ is added to $Trust_{RP}$. Similarly, a participant with ToC less than a predefined threshold2$(Th_2)$ is penalized, and the amount of $\left |ToC-\rho_{Req}*RE \right |$ is reduced from $Trust_{RP}$. This can be summarized in Eq. \ref{eq:tr}. In our simulations in Section \ref{sim}, we set $(Th_1)=0.7$ and $(Th_2)=0.3$.
\begin{equation}
\label{eq:tr}
\small{Trust_{RP} =\left\{\begin{matrix}
Trust_{RP}+ \left |ToC-\rho_{Req}*RE \right | & \; if\, ToC>Th_1\\
Trust_{RP}- \left |ToC-\rho_{Req}*RE \right | & \; if\, ToC<Th_2
\end{matrix}\right.}
\end{equation}
Note that in the this equation, we use the requester's reputation score $(\rho_{Req})$ as a weight for his evaluation($RE$), since we believe an evaluation from a requester with high reputation score is more trustworthy than an evaluation from a low reputable requester.
\begin{figure}
\centering
\includegraphics[width=5cm]{graph.eps}
\caption{A sample social graph of 4 members with mutual trust ratings}
\label{fig:graph}
\end{figure}
This process is repeated for all participants at the end of each sensing campaign, and $Trust_{RP}$ is updated for all of them.
After every $n$ campaigns, $Trust_{RP}$ values upon each active participant act as inputs for reputation module, which updates the participant's reputation score accordingly.
While there are already different crowdsourcing applications of online reputation systems \cite{survey} such as eBay \footnote{http://www.ebay.com/}, Epinions \footnote{http://www.epinions.com/} and Amazon \footnote{http://www.amazon.com/}, we use Web Page ranking algorithms as the basis for computing reputation scores. We draw parallels between the rank of a page in a set of web pages and the reputation score of a member in a social network. Moreover, the weights of links from different pages to a specific page are considered to be equivalent to the trust ratings of one member as determined by the other members of the social network.
Having a set of objects, a ranking algorithm calculates a relative importance of all objects in the set and makes an ordered list according to the importance. Web page ranking algorithms such as PageRank \cite{PageRank} calculate and assign a rank to a web page by analysing the web graph. Roughly speaking, PageRank ranks a page according to how many other pages are pointing at it. This can be described as a reputation system, because the collection of hyperlinks to a given page can be seen as public information that can be combined to derive a reputation score. A single hyperlink to a given web page can be seen as a trust rating of that web page.
In PageRank, the rank of page $P$, denoted by $\rho(P)$ is defined as:
\begin{math} \rho(P)=\frac{\sum\limits_{P_i\rightarrow P}(\rho(P_i))}{L(P_i)} \end{math}
in which, $P_i$ is the set of all pages which have an outgoing link to page $P$, and $L(P_i)$ is the number of outgoing links from page $P_i$.\\
In the original PageRank algorithm, it is assumed that all the outgoing links have equal weights. This is not always true, since not all outgoing links from a web page are equally important. So, we adopted the extension offered in \cite{kaltix} which modifies the above equation as Eq. \ref{eq:pr},
\begin{equation}
\label{eq:pr}
\rho(P)=\sum\limits_{P_i\rightarrow P}{\frac{w_i}{\sum\limits_{P_i\rightarrow P_j}{w_j}} \rho(P_i)}
\end{equation}
in which, $w_i$ is the weight of the outgoing link, and the sum of weights of outgoing links is equal to 1.\\
We explain this further by presenting an illustrative example. Consider the graph in Fig. \ref{fig:graph} in which, $P_1$, $P_2$, $P_3$ and $P_4$ are the social network members. Links represent friendship relations with weights equal to the mutual trust between the pairs. In this case, according to Eq. \ref{eq:pr}:
\\
\begin{math} \rho(P_1)= T_{21}\times \rho(P_2)\end{math}\\
\begin{math} \rho(P_2)= T_{32}\times \rho(P_3)\end{math}\\
\begin{math} \rho(P_3)= T_{13}\times \rho(P_1)\end{math}\\
\begin{math} \rho(P_4)= T_{14}\times \rho(P_1) + T_{24}\times \rho(P_2) + T_{34}\times \rho(P_3)\end{math}\\
As can be seen in the above expressions, reputation calculation is an iterative process and continues until convergence is obtained. In our simulation in Section \ref{sim}, we assume that the convergence occurs when $\left |\rho_k(P_i)-\rho_{k-1}(P_i) \right| \leq 10^{-10}$ for all $P_i$. \\
To summarize, once a campaign is launched, participants begin to send a series of contributions. For each contribution, the Evaluator computes a value for QoC and ToP. These values are fed to fuzzy inference engine which calculates ToC for that contribution. The trust of requester upon each participant($Trust_{RP}$) is updated according to his ToC. The server utilizes $Trust_{RP}$ and $\rho_{Req}$ to update the reputation score of each participant.
\section{Experimental Evaluation}
\label{sim}
This section presents simulation-based evaluation of the proposed trust system. The simulation setup is outlined in Section \ref{setup} and the results are in Section \ref{res}.
\subsection{Simulation Setup}
\label{setup}
To undertake the preliminary evaluations outlined herein, we chose to conduct simulations, since real experiments in social participatory sensing are difficult to organise. Simulations afford a controlled environment where we can carefully vary certain parameters and observe the impact on the system performance. We developed a custom Java simulator for this purpose.
We simulate an online social network where 100 members participate in 5000 campaigns, producing one contribution for each. We assume that each member is connected to all others, similar to a social group; So, there are totally 10000 friendship relations. All members can serve both as requesters who launch sensing campaigns and as participants who contribute data to these sensing campaigns.
In our previous work \cite{Mobi}, we assumed of categorizing participants according to the trade-offs between ToP and QoC. We wanted to observe how accurate the system assigns trust ratings to contributions in case of different ToP and QoC levels. Moreover, we artificially created scenarios where participants begin producing contributions with low QoC, which results in a decrease in ToC. We wanted to see if the system is able to quickly detect this transition and revoke low trustable contributions in an accurate and robust manner.
In this paper, instead of categorizing the participants according to ToP and QoC, we designed the categories according to the trade-offs between personal factors and social factors inside ToP, and simply assumed that QoC has a value in the range of $(ToP-\mu , ToP+\mu)$. In fact, we are going to observe how the system reacts to behavioural changes of participants and how much it is successful to update the reputation scores in case of such fluctuations. As mentioned in Section \ref{comp}, ToP parameters can be divided into two groups: social factors which include Friendship duration and Interaction time gap, and personal factors which include Timeliness, Expertise and Locality. In the real-world, there are often situations where a friend with a high rating of social factors (i.e., a very close friend with whom one has repeated interactions) has a low rating for personal factors for a period of time (i.e., does not have related expertise or does not produce timely contributions). It other words, we may have participants who have high social trust, but low personal trust, and vice versa. We have thus 4 different states based on the combination of different levels of personal and social trusts.
Specifically, we assumed that 60 members (out of 100) belong to Category A whereas the remaining 40 belong to Category B, adding the assumption that category A members have high personal trust, while category B members have low personal trust. We also assume that for each member $P_A$ in category A, all other members score $P_A$ with high social trust, and for each member $P_B$ in category B, all other members score $P_B$ with low social trust.
When $P_A$ serves as requester, other members form two subcategories:\\
A-1: which includes 59 members from category A, excluding $P_A$. They have high personal trust and score $P_A$ with high social trust.\\
A-2: which includes 40 members from category B. They have low personal trust and score $P_A$ with high social trust.
Similarly, when $P_B$ serves as requester, other members form two subcategories:\\
B-1: which includes 60 members from category A. They have high personal trust and score $P_B$ with low social trust.\\
B-2: which includes 39 workers from category B, excluding $P_B$. They have low personal trust and score $P_B$ with low social trust.
It is but natural that not all friends in a social network would contribute data to sensing campaigns. As such, we assume that 10\% of the members in category A and 50\% of the members in category B do not upload any data. The rationale for assuming unequal percentages is that the first group are close friends and hence a higher percentage would be willing to contribute, whereas the second group are not so and have less willingness to contribute.
Whenever a task is launched, one of the participants is selected to be the requester. Without loss of generality we assume that tasks are launched in sequential order by the social network members, i.e., member 1 launches the first campaign, member 2 launches the second campaign and so on.\\
\noindent\underline{ToP Parameter Settings}
\begin{table}
\centering
\scriptsize{ \caption{ ToP parameter settings}\label{tab:setting}
\begin{tabular}{|c|c|c|c|}\hline
\multicolumn{2}{|c}{category A}&\multicolumn{2}{|c|}{category B}\\\hline
\cline{1-4}
param&value¶m&value\\\hline
n(PE)&4&n(PE)&2\\\hline
rt&$\left\{\begin{matrix}
(0,1]&prob=0.4\\
(1,7/2]&prob=0.65\\
(7/2,7]&prob=0.9
\end{matrix}\right.$&rt&$\left\{\begin{matrix}
(0,1]&prob=0.1\\
(1,7/2]&prob=0.3\\
(7/2,7]&prob=0.5
\end{matrix}\right.$\\\hline
$N_1$&random(0,1)&$N_1$&random(0,0.5)\\\hline
$N_2$&random(0,0.9)&$N_2$&0\\\hline
$N_3$&random(0,0.8)&$N_3$&0\\\hline
t&rand[4,5]&t&rand[0,1]\\\hline
LI&$\left\{\begin{matrix}
(0,d]&prob=0.8\\
0&prob=0.2
\end{matrix}\right.$&LI&$\left\{\begin{matrix}
(0,d]&prob=0.2\\
0&prob=0.8
\end{matrix}\right.$\\\hline
\end{tabular}
}
\end{table}
In the following, we will discuss the initialisation of the various parameters introduced in Section \ref{comp}.
In order to set the Expertise value for a participant, we assume that there are a total of 6 expertise areas defined and that each task needs at most 3 expertise areas $(n(TE)=3)$. To calculate the Expertise score for each participant, we assign a value to $n(PE)$ based on his category, as shown in Table \ref{tab:setting}. The expertise score $E$ is then calculated using Eq. \ref{eq:eqE}.
For Timeliness, we first set the response time \emph{(rt)} for each participant. Considering the task deadline to be 7 days (d=7), \emph{rt} is assigned a value according to the participant's personal category, as depicted in Table \ref{tab:setting}. For example, for a participant $P_A$ belonging to category A, with probability of 0.4, \emph{rt} is at most one day, with the probability=0.65, \emph{rt} is at most half of a week, and with the probability of 0.9, \emph{rt} is at most one week (Note that the greatest probability is 0.9, since with the probability of 0.1 (10\%), $P_A$ does not attend in sensing campaign). \emph{rt} then acts as the input value for Eq. \ref{eq:eq1} which results in Timeliness score $T$ for participant. Other input parameters for Eq. \ref{eq:eq1} have been set as \emph{x}=0.3, \emph{b}= 6, \emph{c}=0.6, and \emph{d}=7 days.
For Locality, we assume that there are a total of 25 regions and that each participant is local to 3 regions (i.e., Locality score $L$ for these 3 regions is 1). We also assume that when a participant has the maximum Locality score to a region, he has a relatively high locality to its surrounding regions. So, Locality score $L$ is assigned to the surrounding regions up to 3 levels of neighborhood,i.e., $N_1$, $N_2$ and $N_3$, based on participant's category, as shown in Table \ref{tab:setting}.
For Friendship duration, as mentioned in Section \ref{comp}, the input parameter $(t)$ is the time (in years) passed from the beginning of friendship establishment. The initial value of \emph{t} is set according to the participant's category, as shown in Table \ref{tab:setting} and a constant value of 0.02 is added to \emph{t} upon each participation.
$t$ is then serves as the input value for Eq. \ref{eq:eq3} which computes the Friendship duration score $F$ for the participant.
Other input parameters for Eq. \ref{eq:eq3} have been set as \emph{b}=5 and \emph{c}=1.
Finally, for the Interaction time gap, as mentioned in Section \ref{comp}, the input parameter $t$ is the gap (in days) between the current time and the Latest Interaction($LI$) time. We set $LI$ based on the category of each participant, as shown in Table \ref{tab:setting}, and calculate $t$ accordingly. $t$ is then fed to Eq. \ref{eq:eq4} which calculates the Interaction time gap score $I$ for the participant. Other input parameters for Eq. \ref{eq:eq4} have been set as \emph{b}=10 and \emph{c}=0.2.
Once all of the aforementioned parameters are computed, ToP is calculated by simply averaging them. In other words, we simply assume that $w_i=1/5$ in Eq. \ref{eq:top}. QoC is then assigned a value in a range of $(ToP-\mu , ToP+\mu)$ with $\mu=0.1$.
ToC is then calculated and $Trust_{RP}$ is updated according to Eq. \ref{eq:tr}. At intervals, reputation score is also updated for participants. We set the reputation interval to be after every 100 campaigns ($n$=100).
In the first scenario, we assume that ToPs follow the category settings throughout the entire simulation. In the second scenario, we assume that ToP parameters change for a group of participants which results in a transition from one category to another (details in Section \ref{res}).
As mentioned in Section \ref{arc}, a ToC rating is calculated for each contribution and those with ToC lower than a predefined threshold are revoked from further calculations. The ToCs for the non-revoked contributions are then combined to form an overall trust for that campaign. In other words, \begin{math} Overall Trust = \frac{\sum_{i=1}^{n} ToC}{n} \end{math} in which, \emph{n} is the number of non-revoked contributions. The revocation threshold is set to 0.5. We consider the overall trust as the evaluation metric. The greater the overall trust the better the ability of the system to revoke untrusted contributions. Overall trust has a value in the range of [0, 1]. We also calculate the reputation scores for all participants to see whether they reflect the behaviour of participants in normal and transition settings. Reputation score value is a number in the range of [0, 1] with initial value of 0.5 for each participant.
We compare the performance of our framework against the following systems: (1) Baseline-Rep: which follows the approach in \cite{Brian} by calculating a reputation score for each participant according to the QoC of his successive contributions. This reputation score is used as a weight for QoC. In other words, \begin{math} ToC= \sqrt{Rep* QoC}\end{math} (2) Average: which includes ToP but computes the ToC simply as an average of ToP and QoC (3) Fuzzy: our proposed framework.
\subsection{Simulation Results}
\label{res}
We first present the simulation results for the first scenario. Fig. \ref{fig:ch1} depicts the evolution of the average overall trust as a function of the number of campaigns. As shown in the figure, our fuzzy trust method outperforms the other methods. This confirms its success in mimicking the human trust establishing process by correct settings of fuzzy rules. In particular, we have set the rules in a way that results in early detection and severe punishment of untrusted contributions and also put greater emphasis on highly trusted contributions. The former has been done by assigning a very low (VL) value to ToC in case of low ToP and QoC (i.e., Rule no. 1 in Table \ref{tab:rule}), whereas the latter has been obtained through assigning very high(VH) value to ToC in case of high QoC and above average ToP (i.e., Rule no. 15 and 16 in Table \ref{tab:rule}).
\begin{figure
\centering
\includegraphics[width=8cm]{avgoverall.eps}
\caption{Evolution of average overall trust for all methods, Scenario 1}
\label{fig:ch1}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=8cm]{otrust1.eps}
\caption{Evolution of overall trust, Fuzzy method, Scenario 1}
\label{fig:ch2}
\end{figure}
\begin{figure
\centering
\includegraphics[width=8cm]{rep.eps}
\caption{Reputation score for all members, Fuzzy method, Scenario 1}
\label{fig:ch3}
\end{figure}
Figure \ref{fig:ch2} depicts the evolution of overall trust for 1000 contributions with Fuzzy method. As can be seen in this figure, at each interval containing 100 contributions, two different levels of overall trust are achieved. Remembering the order of requesters which is equal to members' order, higher level of overall trust is obtained when the requester is from category A. So, participants are located either in subcategory A-1 or A-2. This will result in either high ToC values (when participants are from category A-1) or medium ToC values (when participants are from category A-2), which in turn, results in high overall trust. Similarly, lower level of overall trust is obtained when the requester is from category B. So workers are located either in category B-1 or B-2. This will lead to either medium ToC values (when participants are from category B-1) or low ToC values (when participants are from category B-2), which results in low overall trust. This variation is repeated regularly at each interval of 100 contributions.
Fig. \ref{fig:ch3} presents the reputation of 100 participants after attending in 5000 sensing campaigns. As mentioned before, the initial value of reputation score for all participants is 0.5. Category A participants who have high ToPs, produce contributions with high ToC and hence, they get rewarded. This reward results in $trust_{RP}$ increase upon them, which in turn, increases their reputation score. On the contrary, for category B participants with low ToPs, ToCs will also be low, and hence, they are penalized, which results in the reduction of their reputation score. To summarize, our system continually tracks the contributions made over a series of campaigns and detects participants' behaviour, which is accurately reflected in the evolution of the reputation scores.\\
Next, we present results for the second scenario, wherein, the behaviour of participants change for a period of time, which results in a transition from one category to another. This scenario allows us to observe the performance of the schemes in the presence of noise. For example, consider a participant $P_A$ who is in category A, changes his behaviour for a period of time and behaves in a different manner which results in decrease of his personal and (hence) social trust. For example $P_A$ no longer provides timely contributions or does not care enough about the requirements of the task. This behavioural change results in a decrease in his personal trust, and consequently, others score him low with social trust. In other words, a participant may encounter a transition from category A to category B. In this scenario, we assume that 10 from 60 participants of category A transition to category B (e.g., a reduction in their personal and social factor values is created) in the period between 1000\textsuperscript{th} and 4000\textsuperscript{th} campaigns.\\
Fig. \ref{fig:ch4} shows the reputation score of 100 participants at the end of transition period (i.e., after attending in 4000 campaigns). As can be seen in this figure, the reputation of first ten participants who encounter such transition has a considerable decrease in comparison with others not encountering such transition. This again demonstrate the ability of our reputation module to adjust the reputation scores as a reflection of behavioural changes of participants.\\
Finally, Fig. \ref{fig:ch5} shows the reputation score evolution of participant no. 9 encountering such transition between 10\textsuperscript{th} and 40\textsuperscript{th} reputation intervals (between 1000\textsuperscript{th} and 4000\textsuperscript{th} campaigns). As can be observed, our proposed method shows an explicit and considerable reaction to this behavioural change, as compared with other methods. There is a decrease in reputation score due to dishonest behaviour during the transition period. At the end of transition period, transition encountered participant resumes his normal behaviour which results in a considerable increase in his reputation score.
\begin{figure
\centering
\includegraphics[width=8cm]{repnoise.eps}
\caption{Reputation score for all members at Campaign 4000th, Fuzzy method, Scenario 2}
\label{fig:ch4}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=8cm]{repuser.eps}
\caption{Evolution of Reputation score for participant no.9 in all methods, Scenario 2}
\label{fig:ch5}
\end{figure}
\section{Conclusions}
\label{con}
In this paper, we proposed an application agnostic reputation framework for social participatory sensing system. Our system independently assesses the quality of the data and the trustworthiness of the participants and combines these metrics using fuzzy inference engine to arrive at a comprehensive trust rating for each contribution. The system is then assigns a reputation score to participants by leveraging the concepts utilised in PageRank algorithm. Simulations demonstrated that our scheme increases the overall trust by over 15\% as compared to other methods, and assigns reputation scores to participants in a robust and reliable manner.
\bibliographystyle{plain}
|
\section{Acknowledgements}\label{sec:acknowledgement}
The authors would like to thank Mahesh Gupta for his valuable contributions in performing the experiments.
\section{Conclusion}\label{sec:conclude}
We introduce \system{GQBE}, a system that queries
knowledge graphs by example entity tuples.
As an initial step toward better usability of graph query systems,
\system{GQBE} saves users the burden of forming explicit query graphs.
To the best of our knowledge, there has been no such proposal in the past.
Its query graph discovery component derives a hidden query graph based on example tuples.
The query lattice based on this hidden graph may contain a large number of query graphs.
\system{GQBE}'s query algorithm only partially evaluates query graphs for
obtaining the top-\emph{k}\ answers.
Experiments on Freebase and DBpedia datasets show that \system{GQBE}
outperforms the state-of-the-art system \system{NESS} on both accuracy and efficiency.
\section{Experiments}\label{sec:exp}
This section presents our experiment results on the accuracy and
efficiency of \system{GQBE}. The experiments were
conducted on a double quad-core $24$ GB Memory $2.0$ GHz Xeon server.
\begin{table} [t]
\centering
\scriptsize
\begin{tabular}{|p{6mm}|l|l|}
\hline
{\bf Query} & {\bf Query Tuple} & {\bf Table Size} \\ \hline\hline
F$_1$ & $\langle$Donald Knuth, Stanford University, Turing Award$\rangle$ & 18 \\
F$_2$ & $\langle$Ford Motor, Lincoln, Lincoln MKS$\rangle$ & 25 \\
F$_3$ & $\langle$Nike, Tiger Woods$\rangle$ & 20 \\
F$_4$ & $\langle$Michael Phelps, Sportsman of the Year$\rangle$ & 55 \\
F$_5$ & $\langle$Gautam Buddha, Buddhism$\rangle$ & 621 \\
F$_6$ & $\langle$Manchester United, Malcolm Glazer$\rangle$ & 40 \\
F$_7$ & $\langle$Boeing, Boeing C-22$\rangle$ & 89 \\
F$_8$ & $\langle$David Beckham, A. C. Milan$\rangle$ & 94 \\
F$_9$ & $\langle$Beijing, 2008 Summer Olympics$\rangle$ & 41 \\
F$_{10}$ & $\langle$Microsoft, Microsoft Office$\rangle$ & 200 \\
F$_{11}$ & $\langle$Jack Kirby, Ironman$\rangle$ & 25 \\
F$_{12}$ & $\langle$Apple Inc, Sequoia Capital$\rangle$ & 300 \\
F$_{13}$ & $\langle$Beethoven, Symphony No. 5$\rangle$ & 600 \\
F$_{14}$ & $\langle$Uranium, Uranium-238$\rangle$ & 26 \\
F$_{15}$ & $\langle$Microsoft Office, C++$\rangle$ & 300 \\
F$_{16}$ & $\langle$Dennis Ritchie, C$\rangle$ & 163 \\
F$_{17}$ & $\langle$Steven Spielberg, Minority Report$\rangle$ & 40 \\
F$_{18}$ & $\langle$Jerry Yang, Yahoo!$\rangle$ & 8349 \\
F$_{19}$ & $\langle$C$\rangle$ & 1240 \\
F$_{20}$ & $\langle$TomKat$\rangle$ & 16 \\
D$_1$ & $\langle$Alan Turing, Computer Scientist$\rangle$ & 52 \\
D$_2$ & $\langle$David Beckham, Manchester United$\rangle$ & 273 \\
D$_3$ & $\langle$Microsoft, Microsoft Excel$\rangle$ & 300 \\
D$_4$ & $\langle$Steven Spielberg, Catch Me If You Can$\rangle$ & 37 \\
D$_5$ & $\langle$Boeing C-40 Clipper, Boeing$\rangle$ & 118 \\
D$_6$ & $\langle$Arnold Palmer, Sportsman of the year$\rangle$ & 251 \\
D$_7$ & $\langle$Manchester City FC, Mansour bin Zayed Al Nahyan$\rangle$ & 40 \\
D$_8$ & $\langle$Bjarne Stroustrup, C++$\rangle$ & 964 \\
\hline
\end{tabular}
\vspace{1mm}
\caption{Queries and Ground Truth Table Size}
\label{tab:groundTruth}
\vspace{-2mm}
\end{table}
\vspace{-1mm}
{\flushleft \textbf{Datasets}}\hspace{2mm} The experiments were conducted over two
large real-world knowledge graphs, the Freebase~\cite{Bollacker+08freebase}
and DBpedia~\cite{AuerBK+07} datasets.
We preprocessed the graphs so that the kept nodes are all named
entities (e.g.,\xspace \entity{Stanford University}) and abstract concepts (e.g.,\xspace
\entity{Jewish people}). The resulting Freebase graph contains $28$M nodes,
$47$M edges, and $5,428$ distinct edge labels. The DBpedia graph contains
$759$K nodes, $2.6$M edges and $9,110$ distinct edge labels.
\reminder{say how we removed symmetric edges, data cleaning steps}
\vspace{-1mm}
{\flushleft \textbf{Methods Compared}}\hspace{2mm}
\system{GQBE} was compared with a \system{Baseline} and \system{NESS}~\cite{ness}.
\system{NESS} is a graph querying framework that finds approximate matches of query
graphs with unlabeled nodes which correspond to query entity nodes in MQG.
Note that, like other systems, \system{NESS} must take a query graph (instead of a query tuple)
as input. Hence, we feed the MQG discovered by \system{GQBE} as the query
graph to \system{NESS}. For each node $v$ in the query graph, a set of
candidate nodes in the data graph are identified. Since, \system{NESS} does not consider
edge-labeled graphs, we adapted it by requiring each candidate node $v'$ of
$v$ to have at least one incident edge in the data graph bearing the
same label of an edge incident on $v$ in the query graph. The score of
a candidate $v'$ is the similarity between the neighborhoods of $v$ and $v'$,
represented in the form of vectors, and further refined using an iterative
process. Finally, one unlabeled query
node is chosen as the pivot $p$. The top-$k$ candidates for multiple unlabeled
query nodes are put together to form answer tuples, if they are within the neighborhood
of $p$'s top-$k$ candidates. Similar to the best-first method (Sec.\ref{sec:processing}),
\system{Baseline} explores a query lattice in a bottom-up manner
and prunes ancestors of null nodes. However, differently, it evaluates
the lattice by breadth-first traversal instead of in the order of
upper-bound scores. There is no early-termination by top-\emph{k}\ scores,
as \system{Baseline} terminates when every node is either evaluated or pruned.
We implemented \system{GQBE} and \system{Baseline} in Java and we obtained
the source code of \system{NESS} from the authors.
\reminder{say difference between NESS and GQBE}
\vspace{-1mm}
{\flushleft \textbf{Queries and Ground Truth}}\hspace{2mm}
Two groups of queries are used on the two datasets, respectively.
The Freebase queries F$_1$-- F$_{20}$ are based on Freebase tables
such as \textsf{\scriptsize http:// www.freebase.com/view/computer/programming\_language\_designer?instances},
except F$_1$ and F$_6$ which are from Wikipedia tables such as
\textsf{\scriptsize http://en.wikipedia.org/wiki/List\_of\_English\_football\_club\_owners}.
The DBpedia queries D$_1$-- D$_8$ are based on DBpedia tables such
as the values for property \textsf{\scriptsize is dbpedia-owl:author of}
on page \textsf{\scriptsize http://dbpedia.org/page/Microsoft}.
Each such table is a collection of tuples, in which each tuple consists
of one, two, or three entities. For each table, we used one or more tuples as
query tuples and the remaining tuples as the ground truth for query answers.
All the $28$ queries and their corresponding table sizes are summarized in Table~\ref{tab:groundTruth}.
They cover diverse domains, including people, companies, movies,
sports, awards, religions, universities and automobiles.
\begin{table} [t]
\centering
\scriptsize
\begin{tabular}{|@{\hspace{0.5mm}}c@{\hspace{0.5mm}}|@{\hspace{0.5mm}}c@{\hspace{0.5mm}}|}
\hline
{\bf Query Tuple} & {\bf Top-$3$ Answer Tuples} \\
\hline \hline
& $\langle$D. Knuth, Stanford, V. Neumann Medal$\rangle$ \\
$\langle$Donald Knuth, Stanford, Turing Award$\rangle$ & $\langle$J. McCarthy, Stanford, Turing Award$\rangle$ \\
& $\langle$N. Wirth, Stanford, Turing Award$\rangle$ \\\hline
& $\langle$David Filo, Yahoo!$\rangle$ \\
$\langle$Jerry Yang, Yahoo!$\rangle$ & $\langle$Bill Gates, Microsoft$\rangle$ \\
& $\langle$Steve Wozniak, Apple Inc.$\rangle$ \\\hline
& $\langle$Java$\rangle$ \\
$\langle$C$\rangle$ & $\langle$C++$\rangle$ \\
& $\langle$C Sharp$\rangle$ \\
\hline
\end{tabular}
\vspace{1mm}
\caption{Case Study: Top-$3$ Results for Selected Queries}
\label{tab:queryResults}
\vspace{-2mm}
\end{table}
\vspace{-2mm}
{\flushleft \textbf{Sample Answers}}\hspace{2mm} Table \ref{tab:queryResults} only
lists the top-$3$ results found by \system{GQBE} for $3$ queries (F$_1$, F$_{18}$, F$_{19}$),
due to space limitations.
\spara{(A) Accuracy Based on Ground Truth}
We measured the accuracy of \system{GQBE} and \system{NESS} by
comparing their query results with the ground truth.
The accuracy on a set of queries is the average of accuracy on individual queries.
The accuracy on a single query is captured by three widely-used measures~\cite{Manning08}, as follows.
\begin{list}{$\bullet$}
{ \setlength{\leftmargin}{0.5em} }
\item Precision-at-$k$ (P@$k$): the percentage of the top-$k$ results
that belong to the ground truth.
\item Mean Average Precision (MAP): The average precision of the top-$k$
results is AvgP$=$$\frac{\sum_{i=1}^k{\text{P}@i}\ \times\ rel_i}{\text{size of ground truth}}$,
where $rel_i$ equals $1$ if the result at rank $i$ is in the ground truth
and $0$ otherwise. MAP is the mean of AvgP for a set of queries.
\item Normalized Discounted Cumulative Gain (nDCG):
The cumulative gain of the top-$k$ results is
DCG$_k$$=$$rel_1$$+$$\sum_{i=2}^{k}{\frac{rel_i}{\log_2(i)}}$.
It penalizes the results if a ground truth result is ranked low.
DCG$_k$ is normalized by IDCG$_k$, the cumulative gain for an ideal ranking
of the top-$k$ results. Thus nDCG$_k$$=$$\frac{\text{DCG}_k}{\text{IDCG}_k}$.
\end{list}
\begin{figure}[t]
\centering
\hspace{-5mm}
\subfigure[P@$k$]{
\includegraphics[scale=0.184, angle=270]{figures/precision-k.eps}
\label{fig:precision}
}
\hspace{-6mm}
\subfigure[MAP]{
\includegraphics[scale=0.184, angle=270]{figures/map.eps}
\label{fig:map}
}
\hspace{-6mm}
\subfigure[nDCG]{
\includegraphics[scale=0.184, angle=270]{figures/ndcg.eps}
\label{fig:ndcg}
}
\vspace{-1mm}
\caption{Accuracy of \system{GQBE} and \system{NESS} on Freebase Queries}
\label{fig:accuracy}
\end{figure}
\begin{table} [t]
\centering
\scriptsize
\begin{tabular}{|c|p{4mm}|c|c||c|p{4mm}|c|c|}
\hline
{\bf Query} & {\bf P@$k$} & {\bf nDCG} & {\bf AvgP} & {\bf Query} & {\bf P@$k$} & {\bf nDCG} & {\bf AvgP}\\
\hline \hline
D$_1$ & 1.00 & 1.00 & 0.20 & D$_2$ & 1.00 & 1.00 & 0.04 \\
D$_3$ & 1.00 & 1.00 & 0.03 & D$_4$ & 0.80 & 0.94 & 0.19 \\
D$_5$ & 0.90 & 1.00 & 0.08 & D$_6$ & 1.00 & 1.00 & 0.04 \\
D$_7$ & 0.90 & 0.98 & 0.22 & D$_8$ & 1.00 & 1.00 & 0.01 \\
\hline
\end{tabular}
\vspace{1mm}
\caption{Accuracy of \system{GQBE} on DBpedia Queries, $k$=$10$}
\label{tab:dbpediaResults}
\vspace{-2mm}
\end{table}
Fig.\ref{fig:accuracy} shows these measures for different values
of $k$ on the Freebase queries. \system{GQBE} has high accuracy.
For instance, its P@$25$ is over $0.8$. The absolute value of MAP is not high, merely because
Fig.\ref{fig:map} only shows the MAP for at most top-$25$
results, while the ground truth size (i.e., the denominator in
calculating MAP) for many queries is much larger.
Moreover, \system{GQBE} outperforms \system{NESS} substantially, as its
accuracy in all three measures is almost always twice as better.
This is because \system{GQBE} gives priority to query entities and
important edges in MQG, while \system{NESS} gives equal
importance to all nodes and edges except the pivot.
Furthermore, the way \system{NESS} handles edge labels
does not explicitly require answer entities to be connected by
the same paths between query entities.
Table~\ref{tab:dbpediaResults} further shows the accuracy of \system{GQBE}
on individual DBpedia queries at $k$=$10$. It exhibits high accuracy on all queries,
including perfect precision in several cases.
\begin{table} [t]
\centering
\scriptsize
\begin{tabular}{|@{\hspace{1.5mm}}c@{\hspace{1.5mm}}|@{\hspace{1.5mm}}c@{\hspace{1.5mm}}||@{\hspace{1.5mm}}c@{\hspace{1.5mm}}|@{\hspace{1.5mm}}c@{\hspace{1.5mm}}||@{\hspace{1.5mm}}c@{\hspace{1.5mm}}|@{\hspace{1.5mm}}c@{\hspace{1.5mm}}||@{\hspace{1.5mm}}c@{\hspace{1.5mm}}|@{\hspace{1.5mm}}c@{\hspace{1.5mm}}|}
\hline
{\bf Query} & {\bf PCC} & {\bf Query} & {\bf PCC} & {\bf Query} & {\bf PCC} & {\bf Query} & {\bf PCC} \\
\hline
F$_1$ & 0.79 & F$_2$ & 0.78 & F$_3$ & 0.60 & F$_4$ & 0.80 \\
F$_5$ & 0.34 & F$_6$ & 0.27 & F$_7$ & 0.06 & F$_8$ & 0.26 \\
F$_9$ & 0.33 & F$_{10}$ & 0.77 & F$_{11}$ & 0.58 & F$_{12}$ & undefined \\
F$_{13}$ & undefined & F$_{14}$ & 0.62 & F$_{15}$ & 0.43 & F$_{16}$ & 0.29 \\
F$_{17}$ & 0.64 & F$_{18}$ & 0.30 & F$_{19}$ & 0.40 & F$_{20}$ & 0.65 \\
\hline
\end{tabular}
\vspace{1mm}
\caption{Pearson Correlation Coefficient (PCC) between \system{GQBE} and Amazon MTurk Workers, $k$=$30$}
\label{tab:pcc}
\vspace{-2mm}
\end{table}
\spara{(B) Accuracy Based on User Study}
We conducted an extensive user study through Amazon Mechanical Turk
(MTurk, \textsf{\scriptsize https://www.mturk.com/mturk/}) to evaluate \system{GQBE}'s
accuracy on Freebase queries, measured by Pearson Correlation Coefficient
(PCC). For each of the $20$ queries, we obtained the top-$30$
answers from \system{GQBE} and generated $50$ random pairs of these answers.
We presented each pair to $20$ MTurk workers and asked for their preference
between the two answers in the pair. Hence, in total, $20,000$ opinions were obtained.
We then constructed two value lists per query, $X$ and $Y$, which represent
\system{GQBE} and MTurk workers' opinions, respectively. Each list has $50$
values, for the $50$ pairs. For each pair, the value in $X$ is the difference between
the two answers' ranks given by \system{GQBE}, and the value in $Y$ is the difference
between the numbers of workers favoring the two answers. The PCC value for a query
is $(\text{E}(XY)-\text{E}(X)\text{E}(Y))/(\sqrt{\text{E}(X^2)-(\text{E}(X))^2}\sqrt{\text{E}(Y^2)-(\text{E}(Y))^2})$.
The value indicates the degree of correlation between the pairwise ranking orders
produced by \system{GQBE} and the pairwise preferences given by MTurk workers.
The value range is from $-1$ to $1$. A PCC value in the ranges of [$0.5$,$1.0$],
[$0.3$,$0.5$) and [$0.1$,$0.3$) indicates a strong, medium and small positive correlation,
respectively~\cite{pcc_cohen}.
PCC is undefined, by definition, when $X$ and/or $Y$ contain all equal values.
Table \ref{tab:pcc} shows the PCC values for F$_{1}$-F$_{20}$. Out of the $20$ queries,
\system{GQBE} attained strong, medium and small positive correlation
with MTurk workers on $9$, $5$ and $3$ queries, respectively. Only query F$_{7}$ shows no correlation.
Note that PCC is undefined for F$_{12}$ and F$_{13}$, because all the top-$30$ answer tuples
have the same score and thus the same rank, resulting in all zero values in $X$, i.e., \system{GQBE}'s list.
\begin{table*} [t]
\centering
\scriptsize
\begin{tabular}{|p{5mm}|p{4mm}|p{5mm}|p{5mm}|p{4mm}|p{5mm}|p{5mm}|p{4mm}|p{5mm}|p{5mm}|p{4mm}|p{5mm}|p{5mm}|p{4mm}|p{5mm}|p{5mm}|}
\hline
{\bf Query} & \multicolumn{3}{|c|}{{\bf Tuple1}} & \multicolumn{3}{|c|}{{\bf Tuple2}} & \multicolumn{3}{|c|}{{\bf Combined (1,2)}} & \multicolumn{3}{|c|}{{\bf Tuple3}} & \multicolumn{3}{|c|}{{\bf Combined (1,2,3)}}\\\hline
\hline
& P@$k$ & nDCG & AvgP & P@$k$ & nDCG & AvgP & P@$k$ & nDCG & AvgP & P@$k$ & nDCG & AvgP & P@$k$ & nDCG & AvgP\\\hline
F$_1$ & {\bf 0.36} & 0.76 & 0.32 & {\bf 0.36} & {\bf 1.00} & {\bf 0.50} & 0.12 & 0.38 & 0.02 & {\bf 0.36} & 0.73 & 0.22 & 0.12 & 0.49 & 0.02 \\
F$_2$ & 0.76 & {\bf 1.00} & 0.79 & 0.00 & 0.00 & 0.00 & {\bf 0.80} & {\bf 1.00} & 0.80 & 0.12 & 0.70 & 0.05 & {\bf 0.80} & {\bf 1.00} & {\bf 0.91} \\
F$_4$ & 0.32 & 0.73 & 0.09 & 0.40 & 0.65 & 0.08 & {\bf 1.00} & {\bf 1.00} & {\bf 0.45} & N/A & N/A & N/A & N/A & N/A & N/A \\
F$_6$ & 0.24 & 0.89 & 0.16 & 0.28 & 0.89 & 0.18 & {\bf 0.40} & 0.87 & 0.16 & 0.36 & 0.98 & {\bf 0.22} & 0.12 & {\bf 0.94} & 0.07 \\
F$_8$ & 0.92 & 0.79 & 0.20 & {\bf 1.00} & {\bf 1.00} & {\bf 0.27} & 0.96 & 0.98 & 0.24 & 0.48 & 0.86 & 0.08 & {\bf 1.00} & {\bf 1.00} & {\bf 0.27} \\
F$_9$ & 0.68 & 0.72 & 0.23 & 0.56 & 0.66 & 0.17 & 0.80 & 0.86 & 0.35 & {\bf 1.00} & {\bf 1.00} & 0.62 & {\bf 1.00} & {\bf 1.00} & {\bf 0.66} \\
F$_{17}$ & 0.32 & {\bf 1.00} & 0.33 & 0.64 & 0.83 & 0.25 & 0.32 & {\bf 1.00} & 0.32 & 0.56 & 0.84 & 0.23 & {\bf 0.68} & {\bf 1.00} & {\bf 0.46} \\
\hline
\end{tabular}
\vspace{1mm}
\caption{Accuracy of \system{GQBE} on Multi-tuple Queries, $k$=$25$}
\label{tab:multiTup}
\vspace{-3mm}
\end{table*}
\begin{figure*}[htb]
\begin{minipage}[b]{0.5\linewidth}
\centering
\includegraphics[width = 0.90\linewidth, keepaspectratio = true, scale=0.6]{figures/singleQueryProcessingTimeAllQueries.eps}
\caption{Query Processing Time}
\label{fig:run-time}
\end{minipage}\vspace{-2mm}
\begin{minipage}[b]{0.5\linewidth}
\centering
\includegraphics[width = 0.90\linewidth, keepaspectratio = true, scale=0.6]{figures/latticeNodesTraversedAllQueries.eps
\caption{Lattice Nodes Evaluated}
\label{fig:nodes-traversed}
\end{minipage}\vspace{-2mm}
\end{figure*}
\spara{(C) Accuracy on Multi-tuple Queries}
We investigated the effectiveness of the multi-tuple querying approach
(Sec.\ref{sec:multituple}). In aforementioned single-tuple query
experiment (A), \system{GQBE} attained perfect P@$25$ for $13$ of the
$20$ Freebase queries. We thus focused on the remaining $7$ queries.
For each query, Tuple1 refers to the query tuple in
Table~\ref{tab:groundTruth}, while Tuple2 and Tuple3 are two tuples
from its ground truth. Table~\ref{tab:multiTup} shows
the accuracy of top-$25$ \system{GQBE} answers for the three
tuples individually, as well as for the first two and three tuples
together by merged MQGs, which are denoted Combined(1,2) and
Combined(1,2,3), respectively. F$_4$ attained perfect precision after
Tuple2 was included. Therefore, Tuple3 was not used for F$_4$. The
results show that, in most cases, Combined(1,2) had better accuracy
than individual tuples and Combined(1,2,3) further improved accuracy.
\spara{(D) Efficiency Results}
We compared the efficiency of \system{GQBE}, \system{NESS} and
\system{Baseline} on Freebase queries. The total run time for a query
tuple is spent on two components---query graph discovery and query
processing. Fig.\ref{fig:run-time} compares the three
methods' query processing time, in logarithmic scale.
The figure shows the query processing time for each of the $20$ Freebase queries, and the edge cardinality of the MQG for each of those is shown below the corresponding query id.
The query cost does not appear to increase by edge
cardinality, regardless of the query method.
For \system{GQBE} and \system{Baseline}, this is because
query graphs are evaluated by joins and join selectivity plays a more
significant role in evaluation cost than number of edges.
\system{NESS} finds answers by intersecting postings lists on feature
vectors. Hence, in evaluation cost, intersection size matters more
than edge cardinality.
\system{GQBE} outperformed \system{NESS} on $17$ of the $20$ queries
and was more than $3$ times faster in $10$ of them. It finished within $10$
seconds on $17$ queries. However, it performed very poorly
on F$4$ and F$19$, which have $10$ and $7$ edges respectively.
This indicates that the edges in the two MQGs lead to poor join selectivity.
\system{Baseline} clearly suffered, due to its inferior pruning
power compared to the best-first exploration employed by \system{GQBE}.
This is evident in Fig.\ref{fig:nodes-traversed} which shows the numbers
of lattice nodes evaluated for each query.
\system{GQBE} evaluated considerably less nodes in most cases and
at least $2$ times less on $11$ of the $20$ queries.
MQG discovery precedes the query processing step and is shared by all
three methods. Column MQG$_1$ in Table~\ref{tab:multiTuple-mqg} lists
the time spent on discovering MQG for each Freebase query. This time
component varies across individual queries, depending on the sizes of
query tuples' neighborhood graphs. Compared to the values shown in
Fig.\ref{fig:run-time}, the time taken to discover an MQG in average
is comparable to the time spent by \system{GQBE} in evaluating it.
\begin{table} [t]
\centering
\scriptsize
\begin{tabular}{|@{\hspace{1.5mm}}c@{\hspace{1.5mm}}|@{\hspace{1.5mm}}c@{\hspace{1.5mm}}|@{\hspace{1.5mm}}c@{\hspace{1.5mm}}|@{\hspace{1.5mm}}c@{\hspace{1.5mm}}||@{\hspace{1.5mm}}c@{\hspace{1.5mm}}|@{\hspace{1.5mm}}c@{\hspace{1.5mm}}|@{\hspace{1.5mm}}c@{\hspace{1.5mm}}|@{\hspace{1.5mm}}c@{\hspace{1.5mm}}|}
\hline
{\bf Query} & {\bf MQG$_1$} & {\bf MQG$_2$} & {\bf Merge} & {\bf Query} & {\bf MQG$_1$} & {\bf MQG$_2$} & {\bf Merge} \\
\hline
F$_1$ & 73.141 & 73.676 & 0.034 & F$_2$ & 0.049 & 0.029 & 0.006 \\
F$_3$ & 12.566 & 4.414 & 0.024 & F$_4$ & 5.731 & 7.083 & 0.024 \\
F$_5$ & 9.982 & 2.522 & 0.079 & F$_6$ & 6.082 & 4.654 & 0.039 \\
F$_7$ & 0.152 & 0.107 & 0.007 & F$_8$ & 10.272 & 2.689 & 0.032 \\
F$_9$ & 62.285 & 2.384 & 0.041 & F$_{10}$ & 2.910 & 5.933 & 0.030 \\
F$_{11}$ & 59.541 & 65.863 & 0.032 & F$_{12}$ & 1.977 & 0.021 & 0.006 \\
F$_{13}$ & 9.481 & 5.624 & 0.034 & F$_{14}$ & 0.038 & 0.015 & 0.004 \\
F$_{15}$ & 0.154 & 5.143 & 0.021 & F$_{16}$ & 54.870 & 6.928 & 0.057 \\
F$_{17}$ & 60.582 & 69.961 & 0.041 & F$_{18}$ & 58.807 & 75.128 & 0.053 \\
F$_{19}$ & 0.224 & 0.076 & 0.003 & F$_{20}$ & 0.025 & 0.017 & 0.002 \\
\hline
\end{tabular}
\vspace{1mm}
\caption{Time for Discovering and Merging MQGs (secs.)}
\label{tab:multiTuple-mqg}
\vspace{-2mm}
\end{table}
\begin{figure}[t]
\centering
\includegraphics[width = 0.5\linewidth, keepaspectratio = true, scale=0.4]{figures/multiTupleBoxPlotData.eps}\vspace{-3mm}
\caption{Query Processing Time of $2$-tuple Queries}
\label{fig:multiTuple-exec-time}
\end{figure}
Fig.\ref{fig:multiTuple-exec-time} shows the distribution of the \system{GQBE}'s query
processing time, in logarithmic scale, on the merged
MQGs of 2-tuple queries in Table~\ref{tab:multiTup}, denoted by
Combined(1,2). It also shows the distribution of the total time for evaluating the two
tuples' MQGs individually, denoted Tuple1+Tuple2.
Combined(1,2) processes $10$ of the $20$ queries in less than a second while the fastest query for Tuple1+Tuple2 takes a second.
This suggests that the merged MQGs gave higher weights to more
selective edges, resulting in faster lattice evaluation.
Meanwhile, these selective edges are also more important edges
common to the two query tuples, leading to improved answer accuracy
as shown in Table~\ref{tab:multiTup}. Table~\ref{tab:multiTuple-mqg}
further shows the time taken to discover MQG$_1$ and MQG$_2$
for the two tuples, along with the time for merging them.
The latter is negligible compared to the former.
\section{Introduction}\label{sec:intro}
\reminder{mention Facebook entity graph, semantic graph, SPARQL, we ignore numeric values, some QBE-related work on RDF...}
\reminder{related work
SWiPE: Searching Wikipedia by Example
RDF-QBE: a Semantic Web Building Block
Semantic Query-by-Example for RDF data - Sogang University
}
There is an unprecedented proliferation of \emph{knowledge graphs} that record millions of entities (e.g., persons, products, organizations) and their relationships. Fig.\ref{fig:example-graph} is an excerpt of a knowledge graph, in which the edge labeled \edge{founded} between nodes \entity{Jerry Yang} and \entity{Yahoo!} captures the fact that the person is a founder of the company. Examples of real-world knowledge graphs include DBpedia~\cite{AuerBK+07}, YAGO~\cite{SuchanekKW07}, Freebase~\cite{Bollacker+08freebase} and Probase~\cite{probase}.
Users and developers are tapping into knowledge graphs for numerous applications, including search, recommendation and business intelligence.
Both users and application developers are often overwhelmed by the daunting task of understanding and using knowledge graphs. This largely has to do with the sheer size and complexity of such data. As of March 2012, the Linking Open Data community had interlinked over 52 billion RDF triples spanning over several hundred datasets. More specifically, the challenges lie in the gap between complex data and non-expert users.
Knowledge graphs are often stored in relational databases, graph databases and triplestores (cf.~\cite{KhanWY12} for a tutorial).
\begin{figure}[htb]
\centering
\includegraphics[width = 0.99\linewidth, keepaspectratio = true, scale=0.4]{figures/erGraph.eps}\vspace{-4mm}
\caption{An Excerpt of a Knowledge Graph}
\label{fig:example-graph}
\end{figure
\begin{figure}[tb!]
\centering
\includegraphics[width = 0.9\linewidth, keepaspectratio = true, scale=0.4]{figures/searchScreenWithoutAC.eps}\vspace{-3mm}
\caption{Query Interface of \system{GQBE}}
\label{fig:searchScreen}\vspace{1mm}
\end{figure}
In retrieving data from these databases, the norm is often to use structured query languages such as SQL, SPARQL, and those alike. However, writing structured queries requires extensive experiences in query language and data model and good understanding of particular datasets~\cite{usability}.
Graph data is not ``easier'' than relational data in either query language or data model. The fact it is schema-less makes it even more intangible to understand and query. \emph{If querying ``simple'' tables is difficult, aren't complex graphs harder to query?} \reminder{CL: Explain why it is difficult to write a structured query on data graph.}
\begin{figure*}[htb]
\centering
\includegraphics[keepaspectratio = true, scale=0.47]{figures/simple_architecture.eps
\caption{The Architecture and Components of \system{GQBE}}
\label{fig:architecture
\end{figure*}
Motivated by the aforementioned usability challenge, we build \system{GQBE}~\footnote{A description of \system{GQBE}'s user interface and demonstration scenarios can be found in~\cite{gqbedemo}. An accompanying demonstration video is at {\small \url{http://www.youtube.com/watch?v=4QfcV-OrGmQ}}.} (Graph Query by Example), a system that queries knowledge graphs by example entity tuples instead of graph queries. Given a data graph and a query tuple consisting of entities, \system{GQBE} finds similar answer tuples. Consider the data graph in Fig.\ref{fig:example-graph} and an
scenario where a Silicon Valley business analyst is interested in finding entrepreneurs who have founded technology companies head-quartered in California.
Suppose she knows an example 2-entity query tuple such as \etuple{Jerry Yang}{Yahoo!} that satisfies her query intent.
As the query interface in Fig.~\ref{fig:searchScreen} shows, entering such an example tuple to \system{GQBE} is simple, especially with the help of user interface
tools such as auto-completion in identifying the exact entities in the data graph.
The answer tuples can be \etuple{Steve Wozniak}{Apple Inc.} and \etuple{Sergey Brin}{Google}, which are founder-company pairs.
If the query tuple consists of 3 or more entities (e.g., $\langle \entity{Jerry Yang},\entity{Yahoo!},$ $\entity{Sunnyvale}\rangle$), the answers will be similar tuples of the same cardinality (e.g., $\langle \entity{Steve Wozniak},\entity{Apple Inc.},\entity{Cupertino}\rangle$).
Our work is the first to query knowledge graphs by example entity tuples. The paradigm of \emph{query-by-example} (QBE) has a long history in relational databases~\cite{qbe}. The idea is to express queries by filling example tables with constants and shared variables in multiple tables, which correspond to selection and join conditions, respectively. Its simplicity and improved user productivity make QBE an influential database query language. By proposing to query knowledge graphs by example tuples, our premise is that the QBE paradigm will enjoy similar advantages on graph data. The technical challenges and approaches are vastly different, due to the fundamentally different data models.
Substantial progress has been made on query mechanisms that help users construct query graphs or even do not require explicit query graphs. Such mechanisms include keyword search (e.g., \cite{KA11}), keyword-based query formulation~\cite{PoundIW10,YCHH12}, natural language questions~\cite{Yahya+12}, interactive and form-based query formulation~\cite{Demidova+12,Jarrar+12}, and visual interface for query graph construction~\cite{GRAPHITE, GBLENDER}. Little has been done on comparison across these graph query mechanisms. While a usability comparison of these mechanisms and \system{GQBE} is beyond the scope of this paper, we note that they all have
pros and cons and thus complement each other.
Particularly, QBE and keyword-based methods are adequate for different usage scenarios.
Using keyword-based methods, a user has to articulate query keywords, e.g., ``technology companies head-quartered in California and their founders'' for the aforementioned analyst.
Not only a user may find it challenging to clearly articulate a query, but also a query system might not return accurate answers, since it is non-trivial to precisely separate these keywords and correctly match them with entities, entity types and relationships.
This has been verified through our own experience on a keyword-based system adapted from SPARK~\cite{spark}.
In contrast, a \system{GQBE} user only needs to know the names of some entities in example tuples, without being required to specify how exactly the entities are related.
On the other hand, keyword-based querying is more adequate when a user does not know a few sample answers with respect to her query.
In the literature on graph query, the input to a query system in most cases is a
structured query, which is often graphically presented as a query graph or
pattern. Such is not what we refer to as query-by-example, because the query graphs and patterns are formed by using structured query
languages or the aforementioned query mechanisms. For instance, \system{PathSim}~\cite{Sun+11} finds the top-\emph{k}\ similar entities that are connected to a query entity, based on a user-defined meta-path semantics in a heterogeneous network. In~\cite{YSZH12}, given a query graph as input, the system finds structurally isomorphic answer graphs with semantically similar entity nodes.
In contrast, \system{GQBE} only requires a user to provide an entity tuple, without knowing the underlying schema.
There are several challenges in building \system{GQBE}. Below we provide a brief overview of our approach in tackling these challenges. The ensuing discussion refers to the system architecture and components of \system{GQBE}, as shown in Fig.~\ref{fig:architecture}.
(1) With regard to \emph{query semantics}, since the input to \system{GQBE} is a query tuple instead of an explicit query graph, the system must derive a hidden query graph based on the query tuple, to capture user's query intent. The \emph{query graph discovery} component (Sec.\ref{sec:qgraph}) of \system{GQBE} fulfills this requirement and the derived graph is termed a \emph{maximal query graph} (MQG).
The edges in MQG, weighted by several frequency-based and distance-based heuristics, represent important ``features'' of the query tuple to be matched in answer tuples. More concretely, they capture how entities in the query tuple (i.e., nodes in a data graph) and their neighboring entities are related to each other. Answer graphs matching the MQG are projected to answer tuples, which consist of answer entities corresponding to the query tuple entities.
\system{GQBE} further supports multiple query tuples as input which collectively better capture the user intent.
(2) With regard to \emph{answer space modeling} (Sec.\ref{sec:modeling}), there can be a large space of approximate answer graphs (tuples), since it is unlikely to find answer graphs exactly matching the MQG.
\system{GQBE} models the space of answer tuples by a \emph{query lattice} formed by the subsumption relation between all possible query graphs. Each query graph is a subgraph of the MQG and contains all query entities. Its answer graphs are also subgraphs of the data graph and are isomorphic to the query graph. Given an answer graph, its entities corresponding to the query tuple entities form an answer tuple.
Thus the answer tuples are essentially approximate answers to the MQG. For ranking answer tuples, their scores are calculated based on the edge weights in their query graphs and the match between nodes in the query and answer graphs.
(3) The query lattice can be large. To obtain top-\emph{k}\ ranked answer tuples, the brute-force approach of evaluating all query graphs in the lattice can be prohibitively expensive.
For \emph{efficient query processing} (Sec.\ref{sec:processing}), \system{GQBE} employs a top-$k$ lattice exploration algorithm that only partially evaluates the lattice nodes in the order of their corresponding query graphs' upper-bound scores
We summarize the contributions of this paper as follows
\begin{list}{$\bullet$}
{ \setlength{\leftmargin}{0.5em} \setlength{\itemsep}{0pt}}
\item For better usability of knowledge graph querying systems, we propose a novel approach of querying by example entity tuples, which saves users the burden of forming explicit query graphs. To the best of our knowledge, there was no such proposal in the past.
\item The query graph discovery component of \system{GQBE} derives a hidden maximal query graph (MQG) based on input query tuples, to capture users' query intent. \system{GQBE} models the space of query graphs (and thus answer tuples) by a query lattice based on the MQG.
\item \system{GQBE}'s efficient query processing algorithm only partially evaluates the query lattice to obtain the top-\emph{k}\ answer tuples ranked by how well they approximately match the MQG.
\item We conducted extensive experiments and user study on the large Freebase and DBpedia datasets to evaluate \system{GQBE}'s accuracy and efficiency (Sec.\ref{sec:exp}). The comparison with a state-of-the-art graph querying framework \system{NESS}\cite{ness} (using MQG as input) shows that \system{GQBE} is twice as accurate as \system{NESS} and outperforms \system{NESS} on efficiency in most of the queries.
\end{list}
\section{Answer Space Modeling}
\label{sec:modeling}
Given the maximal query graph $MQG_t$ for a tuple $t$, we model the
space of possible query graphs by a lattice. We further discuss the
scoring of answer graphs by how well they match query graphs.
\subsection{Query Lattice}
\begin{definition}\label{def:lattice
The \textbf{\em query lattice} $\mathcal{L}$ is a partially ordered set (poset)
($\mathcal{QG}_t$, $\prec$), where $\prec$ represents the
subgraph-supergraph subsumption relation and $\mathcal{QG}_t$ is the subset of query
graphs (Def.\ref{def:qgraph}) that are subgraphs of $MQG_t$, i.e.,
$\mathcal{QG}_t$$=$$\{Q|Q\in \mathcal{Q}_t \text{ and } Q \preceq MQG_t\}$.
The top element (root) of the poset is thus $MQG_t$. When represented by a
Hasse diagram, the poset is a directed acyclic graph, in which each node
corresponds to a distinct query graph in $\mathcal{QG}_t$. Thus we shall use the terms
{\em lattice node} and {\em query graph} interchangeably.
The {\em children} ({\em parents}) of a lattice node $Q$ are its subgraphs
(supergraphs) with one less (more) edge, as defined below.
\begin{align}
&\textsf{Children}(Q) = \{Q' | Q' \in \mathcal{QG}_t, Q' \prec Q, |E(Q)|\!-\!|E(Q')|\!=\!1 \}& \nonumber\\%\vspace{-2mm}\\
&\textsf{Parents}(Q) = \{Q' | Q' \in \mathcal{QG}_t, Q \prec Q', |E(Q')|\!-\!|E(Q)|\!=\!1\}& \nonumbe
\end{align}
\end{definition}
The leaf nodes of $\mathcal{L}$ constitute of the \emph{minimal query trees}, which
are those query graphs that cannot be made any simpler and yet still keep all the
query entities connected
\begin{definition
A query graph $Q$ is a \textbf{\em minimal query tree} if none of its subgraphs is also a query graph.
In other words, removing any edge from $Q$ will disqualify it from being a query graph---the resulting graph
either is not weakly connected or does not contain all the query entities. Note that such a $Q$ must be
a tree
\end{definition}
\begin{figure}[tb!]
\centering
\includegraphics[width = 1.0\linewidth, keepaspectratio = true, scale=0.4]{figures/lattice_realEntity.eps
\caption{Maximal Query Graph and Query Lattice}
\label{fig:lattice
\end{figure}
\begin{example}[Query Lattice and Minimal Query Tree]
Fig.\ref{fig:lattice}(a) shows a maximal query graph $MQG_t$, which
contains two query entities in shaded circles and five edges $F,G,H,L,$
and $P$. Its corresponding query lattice $\mathcal{L}$ is in
Fig.\ref{fig:lattice}(b). The root node of $\mathcal{L}$, denoted $FGHLP$,
represents $MQG_t$ itself. The two bottom nodes, $F$ and
$HL$, are the two minimal query trees. Each lattice node is a distinct
subgraph of $MQG_t$. For example, the node $FLP$ represents a query graph
with only edges $F, L$ and $P$. Note that there is no lattice node for
$GLP$, which is not a valid query graph since it is not connected
\end{example}
The construction of the query lattice, i.e., the generation of query graphs
corresponding to its nodes, is integrated with its exploration. In other
words, the lattice is built in a ``lazy'' manner---a lattice node is not
generated until the query algorithm (Sec.\ref{sec:processing}) must
evaluate it. The lattice nodes are generated in a bottom-up way. A node
is generated by adding exactly one appropriate edge to the query graph
for one of its children. The generation of bottom nodes, i.e., the minimal query trees,
is described below.
By definition, a minimal query tree can only contain edges on undirected paths
between query entities. Hence, it must be a subgraph of the weakly connected
component $M_s$ found from the core graph described in Sec.\ref{sec:mqg}.
To generate all minimal query trees, our method enumerates all distinct spanning
trees of $M_s$ by the technique in~\cite{GabowM78} and then trim them.
Specifically, given one such spanning tree, all non-query entities (nodes) of degree one
along with their edges are deleted. The deletion is performed iteratively until there is no
such node. The result is a minimal query tree. Only distinct minimal query trees are kept.
Enumerating all spanning trees in a large graph is expensive.
However, in our experiments on the Freebase dataset, the $MQG_t$ discovered by the
approach in Sec.\ref{sec:qgraph} mostly contains less than $15$ edges.
Hence, the $M_s$ from the core graph is also empirically small, for which the cost
of enumerating all spanning trees is negligible.
\subsection{Answer Graph Scoring Function}\label{sec:agraphscore}
The score of an answer graph $A$ (${\textsf{score}_{Q}}(A)$) captures
$A$'s similarity to the query graph $Q$. It is defined below and
is to be plugged into Eq.~(\ref{eq:ranking_function}) for defining
answer tuple score
\begin{align}
\begin{aligned}
\label{eq:scoreFunction}
{\textsf{score}_{Q}}(A) & = {\textsf{s\_score}}(Q) + {\textsf{c\_score}_{Q}}(A)\\%\vspace{-2mm}
{\textsf{s\_score}}(Q) & = \;\;\;\;\; \sum_{e \in E(Q)}{\textsf{w}(e)}\\%\vspace{-2mm}
{\textsf{c\_score}_{Q}}(A) & = \!\!\!\!\!\! \sum_{\substack{ e=(u,v) \in E(Q) \\ e'=(f(u), f(v)) \in E(A)}}\!\!\!\!\!\!\!\!\!\!\!\!\textsf{match}(e, e'
\end{aligned}
\end{align
In Eq.~(\ref{eq:scoreFunction}), ${\textsf{score}_{Q}}(A)$ sums up
two components---the \emph{structure score} of $Q$ (${\textsf{s\_score}}(Q)$)
and the \emph{content score} for $A$ matching $Q$ (${\textsf{c\_score}_{Q}}(A)$).
${\textsf{s\_score}}(Q)$ is the total edge weight of $Q$.
It measures the important structure in $MQG_t$ that is captured by $Q$ and thus by $A$.
${\textsf{c\_score}_{Q}}(A)$ is the total extra credit for identical nodes among
the matching nodes in $A$ and $Q$ given by $f$---the bijection between
$V(Q)$ and $V(A)$ as in Def.\ref{def:ansGraph}.
For instance, among the $6$ pairs of matching nodes between
Fig.\ref{fig:query-graph}(a) and Fig.\ref{fig:answer-graph}(a),
the identical matching nodes are \entity{USA}, \entity{San Jose}
and \entity{California}. The rationale for the extra credit is that although
node matching is not mandatory, the more nodes are matched, the more
similar $A$ and $Q$ are.
The extra credit is defined by the following function $\textsf{match}(e,e')$.
Note that it does not award an identical matching node excessively.
Instead, only a fraction of $\textsf{w}(e)$ is awarded, where the
denominator is either $|E(u)|$ or $|E(v)|$. ($E(u)$ are the edges
incident on $u$ in $MQG_t$.) This heuristic is based on that,
when $u$ and $f(u)$ are identical, many of their neighbors can be also
identical matching nodes.
\begin{multline}
\label{eq:match}
\textsf{match}(e,e')\text{=}
\begin{cases}
\frac{\textsf{w}(e)}{\lvert{E(u)}\lvert} & \text{if } u\text{=}f(u)\\
\frac{\textsf{w}(e)}{\lvert{E(v)}\lvert} & \text{if } v\text{=}f(v)\\
\frac{\textsf{w}(e)}{min(\lvert{E(u)}\lvert,\lvert{E(v)}\lvert)} & \text{if } u\text{=}f(u), v\text{=}f(v)\\
0 & \text{otherwise}
\end{cases
\end{multline}
In discovering $MQG_t$ from $H_t$ by Alg.\ref{alg:mqg}, the weights
of edges in $H_t$ are defined by Eq.~(\ref{eq:edge_wt_function}) which
does not consider an edge's distance from the query tuple.
The rationale behind the design is to obtain a balanced $MQG_t$ which
includes not only edges incident on query entities but also those
in the larger neighborhood. For scoring answers by
Eq.~(\ref{eq:scoreFunction}) and Eq.~(\ref{eq:match}), however,
our empirical observations show it is imperative to differentiate the
importance of edges in $MQG_t$ with respect to query entities,
in order to capture how well an answer graph matches $MQG_t$.
Edges closer to query entities convey more meaningful relationships
than those farther away. Hence, we define edge depth ($\textsf{d}(e)$)
as follows. The larger $\textsf{d}(e)$ is, the less important $e$ is.
\spara{Edge Depth}\hspace{2mm}
The depth $\textsf{d}(e)$ of an edge $e$=$(u,v)$ is its smallest
distance to any query entity $v_i \in t$, \i
\begin{align}
\textsf{d}(e) = \min_{v_i \in t} \min_{u,v} \{\textsf{dist}(u,v_i),\textsf{dist}(v,v_i)\
\end{align}
Here, $\textsf{dist}(.,.)$ is the shortest length of all undirected paths
in $MQG_t$ between the two nodes.
In summary, \system{GQBE} uses Eq.~(\ref{eq:edge_wt_function}) as the definition
of $\textsf{w}(e)$ in weighting edges in $H_t$. After $MQG_t$ is discovered from $H_t$
by Alg.\ref{alg:mqg}, it uses the following Eq.~(\ref{eq:mqg_weighting_function}) as
the definition of $\textsf{w}(e)$ in weighting edges in $MQG_t$.
Eq.~(\ref{eq:mqg_weighting_function}) incorporates $\textsf{d}(e)$ into Eq.~(\ref{eq:edge_wt_function}).
The answer graph scoring functions Eq.~(\ref{eq:scoreFunction}) and Eq.~(\ref{eq:match}) are based
on Eq.~(\ref{eq:mqg_weighting_function}).
\begin{align}
\label{eq:mqg_weighting_function}
\textsf{w}(e) = \textsf{ief}(e)\;/\;(\textsf{p}(e) \times \textsf{d}^2(e))
\end{align}
\subsection{Multi-tuple Queries}
\label{sec:multituple}
The query graph discovery component of \system{GQBE} essentially derives
a user's query intent from input query tuples. For that, a single query
tuple might not be sufficient. While the experiment results in
Sec.\ref{sec:exp} show that a single-tuple query obtains excellent
accuracy in many cases, the results also exhibit that allowing multiple
query tuples often helps improve query answer accuracy.
This is because important relationships commonly associated with multiple
query tuples express the user intent more precisely. For instance,
suppose a user has provided two query tuples together---\etuple{Jerry Yang}{Yahoo!}
and \etuple{Steve Wozniak}{Apple Inc.}. The query entities in both tuples
share common properties such as \edge{places\_lived} in \entity{San Jose}
and \edge{headquartered\_in} a city in \entity{California}, as shown in
Fig.\ref{fig:example-graph}. This might indicate that the user is
interested in finding people from San Jose who founded technology companies
in California.
Given a set of tuples $T$, \system{GQBE} aims at finding top-$k$ answer
tuples similar to $T$ collectively. To accomplish this, one approach is to
discover and evaluate the maximal query graphs (MQGs) of individual query
tuples. The scores of a common answer tuple for multiple query tuples
can then be aggregated. This has two potential drawbacks: (1) Our concern
of not being able to well capture user intent still remains.
If $k$ is not large enough, a good answer tuple may not appear in
enough individual top-$k$ answer lists, resulting in poor aggregated score.
(2) It can become expensive to evaluate multiple MQGs.
\begin{figure}[tb!]
\centering
\includegraphics[width = 0.85\linewidth, keepaspectratio = true, scale=0.25]{figures/multiTuple.eps
\caption{Merging Maximal Query Graphs}
\label{fig:multiple-tuple}
\end{figure}
We approach this problem by producing a merged and re-weighted MQG that
captures the importance of edges with respect to their presence across
multiple MQGs. The merged MQG is then processed by the same method for single-tuple
queries. \system{GQBE} employs a simple strategy to
merge multiple MQGs. The individual MQG for a query tuple
$t_i$=$\langle v_1^i, v_2^i, \ldots, v_n^i \rangle$$\in$$T$
is denoted $M_{t_i}$. A virtual MQG $M_{t_i}'$ is created for every
$M_{t_i}$ by replacing the query entities $v_1^i, v_2^i, \ldots, v_n^i$
in $M_{t_i}$ with corresponding virtual entities $w_1, w_2, \ldots, w_n$ in
$M_{t_i}'$. Formally, there exists a bijective function $g$$:$$V(M_{t_i})$$\rightarrow$$V(M_{t_i}')$
such that (1) $g(v_j^i)$=$w_j$ and $g(v)$=$v$ if $v$$\notin$$t_i$, and
(2) $\forall e$=$(u,v)$$\in$$E(M_{t_i})$, there exists an edge $e'$=$(g(u),g(v))$ $\in$$E(M_{t_i}')$
such that $label(e)$=$label(e')$; $\forall e'$=$(u',v')$$\in$$E(M_{t_i}')$,
$\exists$$e$ =$(g^{-1}(u'),g^{-1}(v'))$$\in$$E(M_{t_i})$ such that $label(e)$=$label(e')$.
The merged MQG is denoted $MQG_T$. It is produced by including vertices
and edges in all $M_{t_i}'$, merging identical virtual and regular vertices,
and merging identical edges that bear the same label and the same vertices
on both ends.
Formally
\begin{align}
\textstyle V(MQG_T)=\bigcup\limits_{t_i \in T} V(M_{t_i}') \text{ and } E(MQG_T)=\bigcup\limits_{t_i \in T} E(M_{t_i}').\nonumbe
\end{align}
The edge cardinality of $MQG_T$ might be larger than the target size $r$.
Thus Alg.\ref{alg:mqg} proposed in Sec.\ref{sec:mqg} is
also used to trim $MQG_T$ to a size close to $r$.
In $MQG_T$, the weight of an edge $e$ is given by $c*\textsf{w}_{max}(e)$,
where $c$ is the number of $M_{t_i}'$ containing $e$ and $\textsf{w}_{max}(e)$
is its maximal weight among all such $M_{t_i}'$
\begin{example}[Merging Maximal Query Graphs]
Let Figs.~\ref{fig:multiple-tuple} (a) and (b) be the $M_{t_i}$
for query tuples \etuple{Steve Wozniak}{Apple Inc.} and
\etuple{Jerry Yang}{Yahoo!}, respectively.
Fig.\ref{fig:multiple-tuple}(c) is the merged $MQG_T$.
Note that entities \entity{Steve Wozniak} and \entity{Jerry Yang} are
mapped to $w_1$ in their respective $M_{t_i}'$ (not shown,
for its mapping from $M_{t_i}$ is simple) and are merged into
\entity{$w_1$} in $MQG_T$. Similarly, entities \entity{Apple Inc.} and
\entity{Yahoo!} are mapped and merged into $w_2$.
The two \edge{founded} edges, appearing in both individual $M_{t_i}$ and
sharing identical vertices on both ends ($w_1$ and $w_2$) in the
corresponding $M_{t_i}'$, are merged in $MQG_T$. Similarly the two
\edge{places\_lived} edges are merged. However, the two
\edge{headquartered\_in} edges are not merged, since they share only one
end ($w_2$) in $M_{t_i}'$. The edges \edge{nationality} and
\edge{education}, which appear in only one $M_{t_i}$, are also present in
$MQG_T$. The number next to each edge is its weight
\end{example}
In comparison to evaluating a single-tuple query, the extra overhead
in handling a multi-tuple query includes creating multiple MQGs, which
is $\lvert T \lvert$ times the average cost of discovering an individual
MQG, and merging them, which is linear in the total edge cardinality of all MQGs.
\section{Problem Formulation}
\label{sec:prelim}
\system{GQBE} runs queries on knowledge data graphs.
A \emph{\textbf{data graph}} is a directed multi-graph $G$ with node set $V(G)$ and edge set $E(G)$.
Each node $v$$\in$$V(G)$ represents an entity and has a unique identifier $id(v)$.~\footnote{Without loss
of generality, we use an entity's name as its identifier in presenting
examples, assuming entity names are unique.}
Each edge $e$=$(v_i,v_j)$$\in$$E(G)$ denotes a directed relationship from
entity $v_i$ to entity $v_j$. It has a label, denoted as $label(e)$.
Multiple edges can have the same label.
The user input and output of \system{GQBE} are both entity tuples,
called \emph{\textbf{query tuples}} and \emph{\textbf{answer tuples}}, respectively.
A tuple $t$=$\langle v_1, \ldots, v_n \rangle$ is an ordered list of
entities (i.e., nodes) in $G$. The constituting entities of query (answer)
tuples are called \emph{query (answer) entities}. Given a data graph $G$
and a query tuple $t$, our goal is to find the top-\emph{k}\ answer tuples $t'$
with the highest similarity scores $\textsf{score}_t(t')$.
\begin{figure}[t]
\centering
\includegraphics[width = 0.88\linewidth, keepaspectratio = true, scale=0.40]{figures/maxQueryGraph.eps
\caption{Neighborhood Graph for \etuple{Jerry Yang}{Yahoo!}}
\label{fig:neighbor-graph}
\end{figure}
\begin{figure*}[htb]
\begin{minipage}[b]{0.33\linewidth}
\centering
\includegraphics[width = 1.0\linewidth, keepaspectratio = true, scale=0.6]{figures/queryGraph.eps
\caption{Two Query Graphs in Fig.\ref{fig:neighbor-graph}}
\label{fig:query-graph}
\end{minipage
\hspace{0.1cm}
\begin{minipage}[b]{0.33\linewidth}
\centering
\includegraphics[width = 1.0\linewidth, keepaspectratio = true, scale=0.6]{figures/answerGraph.eps
\caption{Two Answer Graphs for Fig.\ref{fig:query-graph}(a)}
\label{fig:answer-graph}
\end{minipage
\hspace{0.1cm}
\begin{minipage}[b]{0.33\linewidth}
\centering
\includegraphics[width = 1.0\linewidth, keepaspectratio = true, scale=0.4]{figures/answerGraph1.eps
\caption{Two Answer Graphs for Fig.\ref{fig:query-graph}(b)}
\label{fig:answer-graph1}
\end{minipage
\end{figure*}
We define $\textsf{score}_t(t')$ by matching the inter-entity relationships
of $t$ and that of $t'$, which entails matching two graphs constructed
from $t$ and $t'$, respectively. To this end, we define the
{\em neighborhood graph} for a tuple, which is based on the
concept of undirected path.
An \emph{undirected path} is a path whose edges are not necessarily
oriented in the same direction. Unless otherwise stated, we will
refer to undirected path simply as ``path''. We consider undirected
path because an edge incident on a node can represent an important
relationship with another node, regardless of its direction.
More formally, a path $p$ is a sequence of edges $e_1,\ldots,e_n$ and
we say each edge $e_i\in p$. The path connects two nodes $v_0$ and $v_n$ through
intermediate nodes $v_1,\ldots,v_{n-1}$, where either
$e_i$=$(v_{i-1},v_i)$ or $e_i$=$(v_i,v_{i-1})$, for all
$1$$\leq$$i$$\leq$$n$.
The length of the path, $len(p)$, is $n$ and the endpoints of the path,
$ends(p)$, are $\{v_0, v_n\}$. Note that there is no undirected cycle
in a path, i.e., the entities $v_0,\ldots,v_n$ are all distinct.
\begin{definition
\label{def:ng}
The \textbf{\em neighborhood graph} of query tuple $t$, denoted $H_t$, is the \emph{weakly connected
subgraph}\footnote{A directed graph is \emph{weakly connected}
if there exists an undirected path between every pair of vertices.}
of data graph $G$ that consists of all nodes reachable from at least one query entity by an
undirected path of $d$ or less edges (including query entities themselves)
and the edges on all such paths.
The \emph{path length threshold}, $d$, is an
input parameter. More formally, the nodes and edges in $H_t$
are defined as follows:
$V(H_t) = \{v \lvert v \in V(G)$ and $\exists p$ s.t. $ends(p)$=$\{v_i,v\}$ where $v_i \in t \text{, } len(p) \leq d\}$;
$E(H_t) = \{e \lvert e \in E(G)$ and $\exists p$ s.t. $ends(p)$=$\{v_i,v\}$ where $v_i \in t, len(p) \leq d$, and $e \in p\}$
\end{definition}
\begin{example}[Neighborhood Graph]
Given the data graph in Fig.\ref{fig:example-graph},
Fig.\ref{fig:neighbor-graph} shows the neighborhood graph for query tuple
\etuple{Jerry Yang}{Yahoo!} with path length threshold $d$=2.
The nodes in dark color are the query entities
\end{example}
Intuitively, the neighborhood graph, by capturing how query entities and
other entities in their neighborhood are related to each other,
represents ``features'' of the query tuple that are to be matched in
query answers. It can thus be viewed as a hidden query graph derived
for capturing user's query intent.
We are unlikely to find query answers that exactly match the
neighborhood graph. It is however possible to find exact matches to its subgraphs.
Such subgraphs are all query graphs and
their exact matches are approximate answers that match the neighborhood graph
to different extents
\begin{definition
\label{def:qgraph}
A \textbf{\em query graph} $Q$
is a weakly connected subgraph of $H_t$ that contains all the query entities.
We use $\mathcal{Q}_t$ to denote the set of all query graphs for $t$, i.e.,
$\mathcal{Q}_t$=\{$Q| Q$ is a weakly connected subgraph of $H_t$ s.t.
$\forall v \in t$, $v \in V(Q)$\}
\end{definition}
Continuing the running example, Fig.\ref{fig:query-graph} shows two query graphs
for the neighborhood graph in Fig.\ref{fig:neighbor-graph}.
Echoing the intuition behind neighborhood graph, the definitions of
answer graph and answer tuple are based on the idea that an answer
tuple is similar to the query tuple if their entities participate
in similar relationships in their neighborhoods
\begin{definition
\label{def:ansGraph}
An \textbf{\em answer graph} $A$ to a query graph $Q$
is a weakly connected subgraph of $G$ that is isomorphic to $Q$.
Formally, there exists a bijection $f$$:$$V(Q)$$\rightarrow$$V(A)$ such that:
\begin{list}{$\bullet$}
{ \setlength{\leftmargin}{1em} \setlength{\itemsep}{-1pt} }
\item For every edge $e = (v_i,v_j) \in E(Q)$, there exists an edge $e'=(f(v_i),f(v_j)) \in E(A)$ such that $label(e) = label(e')$;
\item For every edge $e' = (u_i,u_j) \in E(A)$, there exists $e = (f^{-1}(u_i),f^{-1}(u_j)) \in E(Q)$ such that $label(e) = label(e')$.
\end{list
For a query tuple $t$=$\langle v_1, \ldots, v_n \rangle$,
the \textbf{\em answer tuple} in $A$ is
$t_A$=$\langle f(v_1), \ldots, f(v_n)\rangle$.
We also call $t_A$ the \emph{projection} of $A$.
We use $\mathcal{A}_Q$ to denote the set of all answer graphs of $Q$.
We note that a query graph (tuple) trivially matches itself, therefore is not
considered an answer graph (tuple)
\end{definition}
\begin{example}[Answer Graph and Answer Tuple]
Fig.\ref{fig:answer-graph} and Fig.\ref{fig:answer-graph1} each show two
answer graphs for query graphs Fig.\ref{fig:query-graph}(a) and
Fig.\ref{fig:query-graph}(b), respectively.
The answer tuples in
Fig.\ref{fig:answer-graph} are $\langle \entity{Steve Wozniak}$, $\entity{Apple Inc.} \rangle$ and \etuple{Sergey Brin}{Google}.
The answer tuples in Fig.\ref{fig:answer-graph1} are
\etuple{Bill Gates}{Microsoft} and \etuple{Sergey Brin}{Google}
\end{example}
\begin{definition
\label{def:scoringAnsTuple}
The set of answer tuples for query tuple $t$ are
$\{ t_A | A$$\in$$\mathcal{A}_Q, Q$$\in$$\mathcal{Q}_t\}$.
The \emph{\textbf{score of an answer $t'$}} is given b
\begin{align}
\label{eq:ranking_function}
\textsf{score}_t(t')= \max_{A \in \mathcal{A}_Q, Q\in \mathcal{Q}_t} \{{\textsf{score}_Q}(A) | t'=t_A \
\end{align}
The score of an answer graph $A$ (${\textsf{score}_{Q}}(A)$) captures
$A$'s similarity to query graph $Q$. Its equation is given in
Sec.\ref{sec:agraphscore}
\end{definition}
The same answer tuple $t'$ may be projected from multiple answer graphs,
which can match different query graphs. For instance,
Figs.~\ref{fig:answer-graph}(b) and \ref{fig:answer-graph1}(b),
which are answers to different query graphs, have the same
projection---$\langle \entity{Sergey Brin}$, $\entity{Google}\rangle$.
By Eq.~(\ref{eq:ranking_function}), the highest score attained by the
answer graphs is assigned as the score of $t'$, capturing how well $t'$ matches $t$.
\section{Query Processing}
\label{sec:processing}
The query processing component of \system{GQBE} takes the maximal query
graph $MQG_t$ (Sec.\ref{sec:qgraph}) and the query lattice $\mathcal{L}$
(Sec.\ref{sec:modeling}) and finds answer graphs matching the query
graphs in $\mathcal{L}$. Before we discuss how $\mathcal{L}$ is
evaluated (Sec.\ref{sec:bestfirst}), we introduce the storage model and query plan for
processing one query graph (Sec.\ref{sec:onequery}).
\subsection{Processing One Query Graph}\label{sec:onequery}
The abstract data model of knowledge graph can be represented
by the Resource Description Framework (RDF)---the standard Semantic Web
data model. In RDF, a data graph is parsed into a set of triples, each
representing an edge $e$=$(u,v)$. A triple has the form
(subject, property, object), corresponding to ($u, label(e), v$).
Among different schemes of RDF data management, one important approach
is to use relational database techniques to store and query RDF graphs.
To store a data graph, we adopt this approach and, particularly, the
vertical partitioning method~\cite{abadi07}.
This method partitions a data graph into multiple
two-column tables. Each table is for a distinct edge label and stores
all edges bearing that label. The two columns are ($subj, obj$), for
the edges' source and destination nodes, respectively. For efficient query
processing, two in-memory search structures (specifically, hash tables)
are created on the table, using $subj$ and $obj$ as the hash keys, respectively.
The whole data graph is hashed in memory by this way, before any query comes in.
Given the above storage scheme, to evaluate a query graph is to process a
multi-way join query.
For instance, the query graph in Fig.\ref{fig:lattice}(a) corresponds to
{\footnotesize \textsf{SELECT F.subj, F.obj FROM F,G,H,L,P WHERE F.subj=G.sbj
AND F.obj=H.subj AND F.subj=L.subj AND F.obj=P.subj AND H.obj=L.obj.}}
We use right-deep hash-joins to process such a query.
Consider the topmost join operator in a join tree for
query graph $Q$. Its left operand is the \emph{build relation}
which is one of the two in-memory hash tables for an edge $e$.
Its right operand is the \emph{probe relation} which is a
hash table for another edge or a join subtree for
$Q'$=$Q$$-$$e$ (i.e., the resulting graph of removing $e$ from $Q$).
For instance, one possible join tree for the aforementioned
query is $G$$\bowtie$$(F$$\bowtie$$(P$$\bowtie$$(H$$\bowtie$$L)))$.
With regard to its topmost join operator, the left operand
is $G$'s hash table that uses $G.sbj$ as the hash key, and the right
operand is $(F$$\bowtie$$(P$$\bowtie$$(H$$\bowtie$$L)))$.
The hash-join operator iterates through tuples from the
probe relation, finds matching tuples from the build relation,
and joins them to form answer tuples.
\subsection{Best-first Exploration of Query Lattice}\label{sec:bestfirst}
Given a query lattice, a brute-force approach is to evaluate all lattice
nodes (query graphs) to find all answer tuples. Its exhaustive nature
leads to clear inefficiency, since we only seek top-\emph{k}\ answers.
Moreover, the potentially many queries are evaluated separately, without
sharing of computation. Suppose query graph $Q$ is evaluated by the
aforementioned hash-join between the build relation for $e$ and the
probe relation for $Q'$. By definition, $Q'$ is also a query graph in
the lattice, if $Q'$ is weakly connected and contains all query entities.
In other words, in processing $Q$, we would have processed one of its
children query graph $Q'$ in the lattice.
We propose Alg.\ref{alg:hierarchical}, which allows sharing of computation.
It explores the query lattice in a \emph{bottom-up} way, starting with the
minimal query trees, i.e., the bottom nodes. After a query graph is processed,
its answers are materialized in files.
To process a query $Q$, at least one of its children $Q'$=$Q$$-$$e$ must have been
processed. The materialized results for $Q'$ form the probe relation and
a hash table on $e$ is the build relation.
While any topological order would work for the bottom-up exploration,
Alg.\ref{alg:hierarchical} employs a {\em best-first} strategy
that always chooses to evaluate the most promising lattice node
$Q_{best}$ from a set of candidate nodes. The gist is to process the
lattice nodes in the order of their upper-bound scores and $Q_{best}$
is the candidate with the highest upper-bound score (Line~\ref{ln:qbest}).
If processing $Q_{best}$ does not yield any answer graph, $Q_{best}$
and all its ancestors are pruned (Line~\ref{ln:prune}) and the
upper-bound scores of other candidate nodes are recalculated
(Line~\ref{ln:recalc}). The algorithm terminates, without fully evaluating
all lattice nodes, when it has obtained
at least \emph{k} answer tuples with scores higher than the highest
possible upper-bound score among all unevaluated nodes
(Line~\ref{ln:terminate}).
For an arbitrary query graph $Q$, its upper-bound score is given by the best
possible score $Q$'s answer graphs can attain.
Deriving such upper-bound score based on ${\textsf{score}_{Q}}(A)$ in Eq.~(\ref{eq:scoreFunction}) leads to loose upper-bound.
${\textsf{score}_{Q}}(A)$ sums up the structure score of $Q$ (${\textsf{s\_score}}(Q)$)
and the content score for $A$ matching $Q$ (${\textsf{c\_score}_{Q}}(A)$).
While ${\textsf{s\_score}}(Q)$ only depends on $Q$ itself, ${\textsf{c\_score}_{Q}}(A)$ captures
the matching nodes in $A$ and $Q$. Without evaluating $Q$ to get $A$,
we can only assume perfect $\textsf{match}(e, e')$ in Eq.~(\ref{eq:scoreFunction}),
which is clearly an over-optimism.
Under such a loose upper-bound, it can be difficult to achieve an early termination of
lattice evaluation.
To alleviate this problem, \system{GQBE} takes a two-stage approach.
Its query algorithm first finds the top-$k'$ answers ($k'$$>$$k$) based on the structure score
${\textsf{s\_score}}(Q)$ only, i.e., the algorithm uses a simplified answer graph
scoring function ${\textsf{score}_{Q}}(A)={\textsf{s\_score}}(Q)$.
In the second stage, \system{GQBE} re-ranks the top-$k'$ answers by the
full scoring function Eq.~(\ref{eq:scoreFunction}) and returns the top-\emph{k}\ answer tuples based
on the new scores.
Our experiments showed the best accuracy for $k$ ranging from $10$ to $25$
when $k'$ was set to around $100$. Lesser values of $k'$ lowered the accuracy and
higher values increased the running time of the algorithm. In the ensuing discussion,
we will not further distinct $k'$ and $k$.
Below we provide the algorithm details.
\subsection{Details of the Best-first Exploration Algorithm}\label{sec:alg}
\spara{\emph{(1) Selecting ${\bf {Q}_{best}}$}}
At any given moment during query lattice evaluation, the lattice nodes belong to
three mutually-exclusive sets---the evaluated, the unevaluated and the pruned.
A subset of the unevaluated nodes, denoted the \emph{lower-frontier} ($\mathcal{LF}$), are
candidates for the node to be evaluated next.
At the beginning, $\mathcal{LF}$ contains only the minimal query trees (Line~\ref{ln:init} of Alg.\ref{alg:hierarchical}). After a node is
evaluated, all its parents are added to $\mathcal{LF}$ (Line~\ref{ln:insert}). Therefore,
the nodes in $\mathcal{LF}$ either are minimal query trees or have at least one evaluated child
\begin{center}
$\mathcal{LF} = \{Q |\ Q\ \text{is not pruned}, \textsf{Children}(Q)$$=$$\emptyset \text{ or }$\\
$(\exists Q' \in \textsf{Children}(Q)\ \text{s.t.}\ Q'\ \text{is evaluated})\}$.
\end{center}
To choose $Q_{best}$ from $\mathcal{LF}$, the algorithm exploits two important properties,
dictated by the query lattice's structure
\begin{property}
\label{prop:latticeAnswerGraphs}
If $Q_1 \prec Q_2$, then $\forall A_2 \in \mathcal{A}_{Q_2}$, $\exists A_1 \in \mathcal{A}_{Q_1}$ s.t. $A_1 \prec A_2$ and $t_{A_1}$=$t_{A_2}$
\begin{proof}
If there exists an answer graph $A_2$ for a query graph $Q_2$, and there exists another query graph $Q_1$ that is a subgraph of $Q_2$, then there is a subgraph of $A_2$ that corresponds to $Q_1$. By Definition~\ref{def:ansGraph}, that corresponding subgraph of $A_2$ is an answer graph to $Q_1$. Since the two answer graphs share a subsumption relationship, the projections of the two yield the same answer tuple.
\end{proof}
\end{property}
Property~\ref{prop:latticeAnswerGraphs} says, if an answer tuple
$t_{A_2}$ is projected from answer graph $A_2$ to lattice node $Q_2$,
then every descendent of $Q_2$ must have at least one answer graph
subsumed by $A_2$ that projects to the same answer tuple.
Putting it in an informal way, an answer tuple
(graph) to a lattice node can always be ``grown'' from its descendant nodes
and thus ultimately from the minimal query trees
\begin{property}
\label{prop:scoringLatticeTuple}
If $Q_1$$\prec$$Q_2$, then $\textsf{s\_score}(Q_1)$$<$$\textsf{s\_score}(Q_2)$
\begin{proof}
If $Q_1 \prec Q_2$, then $Q_2$ contains all edges in $Q_1$ and at least one more. Thus the property holds by the definition of $\textsf{s\_score}(Q)$ in Eq.~(\ref{eq:scoreFunction}).
\end{proof}
\end{property}
Property~\ref{prop:scoringLatticeTuple} says that, if a lattice node $Q_2$ is an
ancestor of $Q_1$, $Q_2$ has a higher structure score. This can be directly
proved by referring to the definition of $\textsf{s\_score}(Q)$ in Eq.~(\ref{eq:scoreFunction}).
For each unevaluated candidate node $Q$ in $\mathcal{LF}$, we define an {\em upper-bound score},
which is the best score $Q$'s answer tuples can possibly attain.
The chosen node, $Q_{best}$, must have the highest upper-bound score among all the nodes in $\mathcal{LF}$.
By the two properties, if evaluating $Q$ returns an answer graph $A$,
$A$ has the potential to grow into an answer graph $A'$ to an ancestor node
$Q'$, i.e., $Q$$\prec$$Q'$ and $A$$\prec$$A'$. In such a case,
$A$ and $A'$ are projected to the same answer tuple $t_{A}$=$t_{A'}$.
The answer tuple always gets the better score from $A'$, under the
simplified answer scoring function
${\textsf{score}_{Q}}(A)={\textsf{s\_score}}(Q)$, which Alg.\ref{alg:hierarchical}
adopts as mentioned in Sec.~\ref{sec:bestfirst}.
Hence, $Q$'s upper-bound score depends on its \emph{upper boundary}---
$Q$'s unpruned ancestors that have no unpruned parents
\begin{definition
\label{def:ub}
The \textbf{\em upper boundary} of a node $Q$ in $\mathcal{LF}$, denoted $\mathcal{UB}(Q)$,
consists of nodes $Q'$ in the \textbf{\em upper-frontier} ($\mathcal{UF}$) that
subsume or equal to $Q$
\begin{align}
\mathcal{UB}(Q) = \{Q' |\ Q' \succeq Q, Q'\in \mathcal{UF} \}, \text{where} \nonumber
\end{align}
$\mathcal{UF}$ is the set of unpruned nodes without unpruned parents:
$\mathcal{UF}$$=$$\{Q |\ Q\ \text{is not pruned}, \nexists Q'\succ Q\ \text{s.t.}\ Q'\ \text{is not pruned}\}$.
\end{definition}
\begin{definition
\label{def:ubs}
The \textbf{\em upper-bound score} of a node $Q$ is the maximum score
of any query graph in its upper boundary
\begin{align}
\label{eq:ub}
U(Q) = \operatorname*{max}_{Q'\in \mathcal{UB}(Q)}\textsf{s\_score}(Q'
\end{align}
\end{definition}
\input{alg/alg-best-first}
\begin{figure}[tb!]
\centering
\includegraphics[width = 1.05\linewidth, keepaspectratio = true, scale=0.4]{figures/lattice-traversal.eps
\caption{Evaluating Lattice in Figure~\ref{fig:lattice}(b)}
\label{fig:lattice-traversal
\end{figure}
\begin{example}[Lattice Evaluation]
Consider the lattice in Fig.\ref{fig:lattice-traversal}(a) where the lightly shaded nodes belong to the $\mathcal{LF}$ and the darkly shaded node belongs to $\mathcal{UF}$. At the beginning, only the minimal query trees belong to the $\mathcal{LF}$ and the maximal query graph belongs to the $\mathcal{UF}$. If \emph{HL} is chosen as $Q_{best}$ and evaluating it results in matching answer graphs, all its parents (\emph{GHL}, \emph{HLP} and \emph{FHL}) are added to $\mathcal{LF}$ as shown in Fig.\ref{fig:lattice-traversal}(b). The evaluated node \emph{HL} is represented in bold dashed node.
\end{example}
\spara{\emph{(2) Pruning and Lattice Recomputation}}
A lattice node that does not have any answer graph is referred to as a {\em null node}.
If the most promising node $Q_{best}$ turns out to be a null node after evaluation, all its
ancestors are also null nodes based on Property~\ref{prop:up_close} below which follows
directly from Property~\ref{prop:latticeAnswerGraphs}.
\begin{property}[Upward Closure]
\label{prop:up_close}
If $\mathcal{A}_{Q_1}=\emptyset$, then $\forall Q_2 \succ Q_1$, $\mathcal{A}_{Q_2}=\emptyset$
\begin{proof}
Suppose there is a query node $Q_2$ such that $Q_1 \prec Q_2$ and $\mathcal{A}_{Q_1}=\emptyset$, while $\mathcal{A}_{Q_2} \neq \emptyset$. By Property~\ref{prop:latticeAnswerGraphs}, for every answer graph $A$ in $\mathcal{A}_{Q_2}$, there must exist a subgraph of $A$ that belongs to $\mathcal{A}_{Q_1}$. This is a contradiction and completes the proof.
\end{proof}
\end{property}
\input{alg/alg-recomp-bound}
\begin{figure}[tb!]
\centering
\includegraphics[width = 1.05\linewidth, keepaspectratio = true, scale=0.4]{figures/recompute-lattice_realEntity.eps
\caption{Recomputing Upper Boundary of Dirty Node $FG$}
\label{fig:recompute-lattice
\end{figure}
Based on Property~\ref{prop:up_close},
when $Q_{best}$ is evaluated to be a null node, Alg.\ref{alg:hierarchical} prunes
$Q_{best}$ and its ancestors, which changes the upper-frontier $\mathcal{UF}$.
It is worth noting that $Q_{best}$ itself may be an upper-frontier
node, in which case only $Q_{best}$ is pruned. In general, due to the
evaluation and pruning of nodes, $\mathcal{LF}$ and $\mathcal{UF}$ might overlap.
For nodes in $\mathcal{LF}$ that have at least one upper boundary
node among the pruned ones, the change of $\mathcal{UF}$ leads to changes
in their upper boundaries and, sometimes, their upper-bound scores too.
We refer to such nodes as {\em dirty nodes}. The rest of this section
presents an efficient method (Alg.~\ref{alg:recompute}) to recompute
the upper boundaries, and if changed, the upper-bound scores of the
dirty nodes.
Consider all the pairs $\langle Q, Q' \rangle$ such that
$Q$ is a dirty node in $\mathcal{LF}$, and $Q'$ is one of its pruned upper boundary nodes.
Three necessary conditions for a new candidate upper boundary node of $Q$ are that it is
(1) a supergraph of $Q$, (2) a subgraph of $Q'$ and (3) not a supergraph of $Q_{best}$.
The subsumption relationships among these graphs can be visualized in a Venn diagram, as shown in Fig.\ref{fig:venn}.
If there are $q$ edges in $Q_{best}$ but not in $Q$ (the non-intersecting region of $Q_{best}$ in Fig.\ref{fig:venn}),
we create a set of $q$ distinct graphs $Q''$. Each $Q''$ contains all edges in $Q'$ except
exactly one of the aforementioned $q$ edges (Line 8 in Alg.~\ref{alg:recompute}).
For each $Q''$, we find $Q_{sub}$ which is the weakly connected component of $Q''$ containing all the query entities (Lines 9-10).
\begin{figure}[t]
\centering
\includegraphics[width = 0.5\linewidth, keepaspectratio = true, scale=0.65]{figures/recompute-venn.eps
\caption{Venn Diagram of Edges}
\label{fig:venn}
\end{figure}
Lemma~\ref{lemma:qsub_notnull} and \ref{lemma:exist_qsub} show that $Q_{sub}$ must be one of the unevaluated nodes after pruning
the ancestor nodes of $Q_{best}$ from $\mathcal{L}$.
\begin{lemma}
\label{lemma:qsub_notnull}
$Q_{sub}$ is a {\em query graph} and it does not belong to the pruned nodes of lattice $\mathcal{L}$
\end{lemma}
\begin{proof}
$Q_{sub}$ is a query graph because it is weakly connected and it contains all the input entities.
Suppose $Q_{sub}$ is a newly generated candidate upper boundary node from pair $\langle Q,Q' \rangle$ and $Q_{sub}$ belongs to the pruned nodes of lattice $\mathcal{L}$. This can happen only due to one of
the two reasons: 1) it is a supergraph of the current null node $Q_{best}$ or 2) it is an already pruned
node. The former cannot happen since the construction mechanism of $Q_{sub}$ proposed ensures that it is not a supergraph
of $Q_{best}$. the latter implies that $Q_{sub}$ was the supergraph of an previously evaluated null node (or $Q_{sub}$ itself was a null node). In this case, since
$Q_{sub} \prec Q'$, $Q'$ would also have been pruned and thus could not have been part of the upper-boundary. Hence $\langle Q,Q' \rangle$ cannot be a valid pair for recomputing the upper boundary if $Q_{sub}$ is a pruned node. This completes the proof.
\end{proof}
\begin{lemma}
\label{lemma:exist_qsub}
$Q \preceq Q_{sub}$
\end{lemma}
\begin{proof}
Based on Alg.~\ref{alg:recompute} described above, $Q''$ is the result of deleting one edge from $Q'$ and that edge does not belong to $Q$.
Therefore, $Q$ is subsumed by $Q''$. By the same algorithm, $Q_{sub}$ is the weakly connected component of $Q''$ that contains all the query entities.
Since $Q$ already is weakly connected and contains all the query entities, $Q_{sub}$ must be a supergraph of $Q$.
\end{proof}
If $Q_{sub}$ (a candidate new upper boundary node of $Q$) is not subsumed by any node in the upper-froniter or other candidate nodes, we add $Q_{sub}$ to $\mathcal{UB}$($Q$) and $\mathcal{UF}$ (Lines 11-13).
Finally, we recompute $Q$'s upper-bound score (Line 14).
Theorem~\ref{th:correctness} justifies the correctness of the above procedure
\begin{theorem}
\label{th:correctness}
If $Q_{best}$ is evaluated to be a null node, then Alg.\ref{alg:recompute} identifies all new upper boundary nodes for every dirty node $Q$.
\end{theorem}
\begin{proof}
For any dirty node $Q$, its original upper boundary $\mathcal{UB}(Q)$ consists of two sets of nodes: (1) nodes that are not supergraphs of $Q_{best}$ and thus remain in the lattice, (2) nodes that are supergraphs of $Q_{best}$ and thus are pruned. By the definition of upper boundary node, no upper boundary node of $Q$ can be a subgraph of any node in set (1). So any new upper boundary node of $Q$ must be a subgraph of a node $Q'$ in set (2). For every pruned upper boundary node $Q'$ in set (2), the algorithm enumerates all (specifically $q$) possible children of $Q'$ that are not supergraphs of $Q_{best}$ but are supergraphs of $Q$. For each enumerated graph $Q''$, the algorithm finds $Q_{sub}$, which is the weakly connected component of $Q''$ containing all query entities. Thus all new upper boundary nodes of $Q$ are identified.
\end{proof}
\begin{example}[Recomputing Upper Boundary]
Consider the lattice in Fig.\ref{fig:recompute-lattice}(a) where nodes \emph{HL} and \emph{F} are the evaluated nodes and the lightly shaded nodes belong to the new $\mathcal{LF}$. If node \emph{GHL} is the currently evaluated null node $Q_{best}$ and \emph{FGHLP} is $Q'$, let \emph{FG} be the dirty node $Q$ whose upper boundary is to be recomputed.
The edges in $Q_{best}$ that are not present in $Q$ are \emph{H} and \emph{L}.
A new upper boundary node $Q''$ contains all edges in $Q'$ excepting exactly either \emph{H} or \emph{L}. This leads to two new upper boundary nodes, \emph{FGHP} and \emph{FGLP}, by removing \emph{L} and \emph{H} from \emph{FGHLP}, respectively.
Since \emph{FGHP} and \emph{FGLP} do not subsume each other and are not subgraphs of any other upper-frontier node, they are now part of $\mathcal{UB}$($Q$) and the new $\mathcal{UF}$.
Fig.\ref{fig:recompute-lattice}(b) shows the modified lattice where the pruned nodes are disconnected. \emph{FHLP} is another node in $\mathcal{UF}$ that is discovered
using dirty nodes such as \emph{FL} and \emph{HLP}
\end{example}
\spara{Complexity Analysis of Alg.\ref{alg:recompute}}\hspace{2mm}
The query graphs corresponding to lattice nodes are represented using bit vectors since we exactly know the edges involved in all the query graphs. The bit corresponding to an edge is set if its present in the query graph. Identifying the dirty nodes, null upper boundary nodes and building a new potential upper boundary node using a pair of nodes $\langle Q, Q' \rangle$, can be accomplished using bit operations and each step incurs $O(\lvert{E(MQG_t)}\lvert)$ time. Finding the weakly connected component of a potential upper boundary using DFS takes $O(\lvert{E(Q')}\lvert)$ time. If $\mathcal{L}_n$ is the set of all null nodes encountered in the lattice and there are $D_p$ such pairs for every null node and $q$ is the average number of potential new upper boundary nodes created per pair, the worst case time complexity of recomputing the upper-frontier is $O(\lvert \mathcal{L}_n \lvert \times D_p \times q \times \lvert{E(MQG_t)}\lvert)$. Our experimental results show low average values of $\lvert \mathcal{L}_n \lvert$, $D_p$ and $q$ with $\lvert \mathcal{L}_n \lvert$ being only 1\% of $\lvert \mathcal{L} \lvert$, $D_p$ around 8 and $q$ around 9. In practice, our upper-frontier recomputation algorithm quickly computes the dynamically changing lattice.
\spara{\emph{(3) Termination}}
After $Q_{best}$ is evaluated, its answer tuples are $\{t_A | A$$\in$$\mathcal{A}_{Q_{best}}\}$.
For a $t_A$ projected from answer graph $A$, the score assigned by
$Q_{best}$ to $A$ (and thus $t_A$) is $\textsf{s\_score}(Q_{best})$, based on
${\textsf{score}_{Q}}(A)$$=$${\textsf{s\_score}}(Q)$---the simplified scoring function adopted by Alg.\ref{alg:hierarchical}.
If $t_A$ was also projected from already evaluated nodes, it has a current score.
By Def.\ref{def:scoringAnsTuple}, the final score of $t_A$ will be from its
best answer graph. Hence, if $\textsf{s\_score}(Q_{best})$ is higher than
its current score, then its score is updated.
In this way, all found answer tuples so far are kept and their current scores
are maintained to be the highest scores they have received.
The algorithm terminates when the current score of the $k^{th}$ best answer tuple so far
is greater than the upper-bound score of the next $Q_{best}$ chosen by the algorithm,
by Theorem~\ref{th:terminate}
\begin{theorem}\label{th:terminate}
If the score of the current $k^{th}$ best answer tuple is greater than $U(Q_{best})$, then terminating the lattice evaluation guarantees that the current
top-$k$ answer tuples have scores higher than $\textsf{s\_score}(Q)$ for any unevaluated
query graph $Q$.
\end{theorem}
\begin{proof}
Suppose, upon termination, there is an unevaluated query graph $Q$ whose $\textsf{s\_score}(Q)$ is greater than the score of the $k^{th}$ answer tuple.
This implies that there exists some node in the lower-frontier whose upper-bound score is at least $\textsf{s\_score}(Q)$ and is thus greater than the score of the $k^{th}$ answer tuple. Since the termination condition precludes this, it is a contradiction to the initial assumption. We thus cannot have any unevaluated query graph whose structure score is greater than the $k^{th}$ answer tuple's score upon termination.
\end{proof}
\reminder{Arijit: I added back the earlier proof, though I think the first line of the proof is somewhat unclear.
What do i-th and k-th answer tuples mean? -- based on ranking? also the following line is unclear: ``the termination condition
which specifies that the largest \emph{upper bound} in the lattice is lesser than $t'_k$''. -- what is largest upper
bound in the lattice?}
\iffalse
\spara{(4) Complexity Analysis}\hspace{2mm}
Joins are used to evaluate the lattice nodes. Minimal query trees might require multiple joins and other lattice nodes require a single join each.
In evaluating the latter, if on average,
the number of answer graphs for a lattice node is $j$, the time to evaluate a node by joining the answers of its child node and the new edge added to form the node is $O(j)$.
If $\lvert{\mathcal{L}_{e}}\lvert$ is the actual number of lattice nodes evaluated, the worst case scenario of query processing is $O(\lvert{\mathcal{L}_{e}}\lvert \times j)$. In practice, due to the pruning power of the best-first exploration technique, $\lvert{\mathcal{L}_{e}}\lvert \ll \lvert{\mathcal{L}}\lvert$. For the queries used in our experiments on Freebase, on average only 8\% of $\lvert{\mathcal{L}}\lvert$ is evaluated. The average number of answers to a lattice node, $j$, is 6500. Thus, the time to evaluate a single lattice node has a significant role in the total query processing time. Therefore, the query processing time is not only dependent on the size of $MQG_t$, but also on the join cardinality involving the edges.
\fi
\section{Query Graph Discovery}
\label{sec:qgraph}
\subsection{Maximal Query Graph}\label{sec:mqg}
The concept of neighborhood graph $H_t$ (Def.\ref{def:ng}) was
formed to capture the features of a query tuple $t$ to be matched
by answer tuples. Given a
well-connected large data graph, $H_t$ itself can be quite large, even
under a small path length threshold $d$. For example, using Freebase
as the data graph, the query tuple \etuple{Jerry Yang}{Yahoo!} produces
a neighborhood graph with $800$K nodes and $900$K edges, for $d$=$2$.
Such a large $H_t$ makes query semantics obscure, because there might
be only few nodes and edges in it that capture important
relationships in the neighborhood of $t$.
\reminder{discuss $d$ in experiments.}
\system{GQBE}'s query graph discovery component
constructs a weighted \emph{maximal query graph} (MQG)
from the neighborhood graph $H_t$.
MQG is expected to be drastically smaller than $H_t$
and capture only important features of the query tuple.
We now define MQG and discuss its discovery algorithm.
\reminder{CL: Provide an example of maximal query graph.}
\begin{definition
\label{def:mqg}
The \textbf{\em maximal query graph} $MQG_t$, given a parameter $m$, is a weakly connected
subgraph of the neighborhood graph $H_t$ that maximizes total edge weight $\sum_{e} \textsf{w}(e)$
while satisfying (1) it contains all query entities in $t$ and
(2) it has $m$ edges. The weight of an edge $e$ in $H_t$, $\textsf{w}(e)$,
is defined in Sec.\ref{sec:edgeweight}.
\end{definition}
There are two challenges in finding $MQG_t$ by directly going after the
above definition. First, a weakly connected subgraph of $H_t$ with
exactly $m$ edges may not exist for an arbitrary $m$.
A trivial value of $m$ that
guarantees the existence of the corresponding $MQG_t$ is $|E(H_t)|$,
because $H_t$ itself is weakly connected.
This value could be too large, which is exactly why we aim to make
$MQG_t$ substantially smaller than $H_t$.
Second, even if $MQG_t$ exists for an $m$, finding it requires
maximizing the total edge weight, which is a hard problem as given in
Theorem~\ref{th:mqg}
\begin{theorem}
\label{th:mqg}
The decision version of finding the maximal query graph $MQG_t$ for
an $m$ is NP-hard
\end{theorem}
\begin{proof}
We prove the NP-hardness by reduction from the NP-hard constrained Steiner
network (CSN) problem~\cite{LZZC09}.
Given an undirected connected graph
$G_1$=$(V,E)$ with non-negative weight $w(e)$ for every $e$$\in$$E$, a
subset $V_n$$\subset$$V$, and a positive integer $m$, the CSN problem is
to find a connected subgraph $G'$=$(V',E')$ with the smallest total edge
weight, where $V_n$$\subseteq$$V'$ and $|E'|$=$m$. The polynomial-time
reduction from the CSN problem to the MQG problem is by
transforming $G_1$ to $G_2$, where each edge $e$ is given an
arbitrary direction and a new weight $w'(e)$=$W$$-$$w(e)$, where
$W$=$\sum_{e \in E} w(e)$. Let $V_n$ be the query tuple. The maximal
query graph $MQG_{V_n}$ found from $G_2$ provides a CSN in $G_1$, by
ignoring edge direction. This completes the proof.
\end{proof}
\input{alg/alg-mqg}
Based on the theoretical analysis, we present a greedy method (Alg.\ref{alg:mqg})
to find a plausible sub-optimal graph of edge cardinality \emph{close} to a given $m$.
The value of $m$ is empirically chosen to be much smaller than $|E(H_t)|$.
Consider edges of $H_t$ in descending order of weight $\textsf{w}(e)$.
We use $G_s$ to denote the graph formed by the top $s$ edges with the largest weights, which itself
may not be weakly connected. We use $M_s$ to denote the weakly connected
component of $G_s$ containing all query entities in $t$, if it exists.
Our method finds the smallest $s$ such that $|E(M_s)|$=$m$ (Line~\ref{ln:equalm}). If such an $M_s$ does not
exist, the method chooses $s_1$, the largest $s$ such that $|E(M_s)|$$<$$m$.
If that still does not exist, it chooses $s_2$, the smallest $s$ such that $|E(M_s)|$$>$$m$,
whose existence is guaranteed because $|E(H_t)|$$>$$m$.
For each $s$ value, the method employs a depth-first search (DFS) starting from
a query entity in $G_s$, if present, to check the existence of $M_s$ (Line~\ref{ln:findms}).
The $M_s$ found by this method may be unbalanced.
Query entities with more neighbors in $H_t$ likely have
more prominent representation in the resulting $M_s$.
A balanced graph should instead have a fair number of edges
associated with each query entity. Therefore, we
further propose a divide-and-conquer mechanism to construct a balanced
$MQG_t$. The idea is to break $H_t$ into $n$$+$$1$ weakly connected subgraphs.
One is the {\em core graph}, which includes all the $n$ query entities
in $t$ and all undirected paths between query entities.
Other $n$ subgraphs are for the $n$ query entities individually, where
the subgraph for entity $v_i$ includes all entities (and their incident
edges) that connect to other query entities only through $v_i$.
The subgraphs are identified by a DFS starting
from each query entity (Lines~\ref{ln:beginloop}-\ref{ln:findgraph} of Alg.\ref{alg:mqg}). During the DFS from $v_i$, all edges
on the undirected paths reaching any other query entity within distance
$d$ belong to the core graph, and other edges belong to $v_i$'s individual subgraph.
The method then applies the aforementioned greedy algorithm to find $n$$+$$1$
weakly connected components, one for each subgraph, that contain the query entities
in corresponding subgraphs. Since the core graph
connects all query entities, the $n$$+$$1$ components altogether form a weakly
connected subgraph of $H_t$, which becomes the final $MQG_t$.
For an empirically chosen small $r$ as the target size of $MQG_t$, we set
the target size for each individual component to be $\frac{r}{n+1}$, aiming at
a balanced $MQG_t$.
\spara{Complexity Analysis of Alg.\ref{alg:mqg}}\hspace{2mm}
In the aforementioned divide-and-conquer method,
if on average there are $r'$=$\frac{\lvert E(H_t) \lvert}{n+1}$ edges
in each subgraph, finding the subgraph by DFS and sorting its $r'$
edges takes $O(r'\log r')$ time.
Given the top-$s$ edges of a subgraph, checking if the weakly
connected component $M_s$ exists using DFS requires $O(s)$ time.
Suppose on average $c$ iterations are required to find the appropriate $s$.
Let $m$=$\frac{r}{n+1}$ be the average target edge cardinality of each
subgraph. Since the method initializes $s$ with $m$, the largest value
$s$ can attain is $m$$+$$c$. So the time for discovering
$M_s$ for each subgraph is $O(r' \log r'$$+$$c$$\times$$(m$$+$$c$)).
For all $n$$+$$1$ subgraphs, the total time required
to find the final $MQG_t$ is $O((n$$+$$1)\times(r' \log r'$$+$$c$$\times$$(m$$+$$c)))$.
For the queries used in our experiments on Freebase, given an empirically
chosen small $r$=$15$, $s$$\ll$$\lvert E(H_t)\lvert$ and on average $c$=$22$.
\subsection{Edge Weighting}\label{sec:edgeweight}
The definition of $MQG_t$ (Def.\ref{def:mqg}) depends on edge
weights. There can be various plausible weighting schemes.
We propose a weighting function based on several heuristic ideas.
The weight of an edge $e$ in $H_t$, $\textsf{w}(e)$,
is proportional to its inverse edge label frequency ($\textsf{ief}(e)$) and
inversely proportional to its participation degree ($\textsf{p}(e)$), given b
\begin{align}
\label{eq:edge_wt_function}
\textsf{w}(e) = \textsf{ief}(e)\;/\;\textsf{p}(e
\end{align}
\spara{Inverse Edge Label Frequency}\hspace{2mm}
Edge labels that appear frequently in the entire
data graph $G$ are often less important. For example,
edges labeled \edge{founded} (for a company's founders) can be rare and more important than edges labeled
\edge{nationality} (for a person's nationality).
We capture this by the \emph{inverse edge label frequency}
\begin{align}
\label{eq:ief}
\textsf{ief}(e) = \log\;(|E(G)|\;/\;\#label(e)
\end{align}
where $|E(G)|$ is the number of edges in $G$, and $\#label(e)$ is the
number of edges in $G$ with the same label as $e$.
\spara{Participation Degree}\hspace{2mm}
The {\em participation degree} $p(e)$ of an edge $e$=$(u,v)$ is the number of edges in $G$ that share the same label and one of $e$'s end nodes. Formally
\begin{align}
\label{eq:participation}
\hspace{-2mm} \textsf{p}(e)=\lvert\ \{e'\textrm{=}(u',v')\ |\ label(e)\textrm{=}label(e'), u'\textrm{=}u \vee v'\textrm{=}v\}\ \lver
\end{align}
While $\textsf{ief}(e)$ captures the global frequencies of edge
labels, $\textsf{p}(e)$ measures their local frequencies---an edge
is less important if there are other edges incident on the
same node with the same label. For instance, \edge{employment}
might be a relatively rare edge globally but not necessarily
locally to a company. Specifically, consider the edges representing
the \edge{employment} relationship between a company
and its \emph{many} employees and the edges for the \edge{board member}
relationship between the company and its \emph{few} board members. The latter
edges are more significant.
Note that $\textsf{ief}(e)$ and $\textsf{p}(e)$ are precomputed
offline, since they are query-independent and only rely on the
data graph $G$.
\reminder{CL: do they depend on $G$ or reduced neighborhood graph? Can they be precomputed?}
\subsection{Preprocessing: Reduced Neighborhood Graph}
The discussion so far focuses on discovering $MQG_t$ from $H_t$.
The neighborhood graph $H_t$ may have clearly unimportant edges.
As a preprocessing step, \system{GQBE} removes such edges from $H_t$
before applying Alg.\ref{alg:mqg}. The reduced size of $H_t$
not only makes the execution of Alg.\ref{alg:mqg} more efficient
but also helps prevent clearly unimportant edges from getting into $MQG_t$.
Consider the neighborhood graph $H_t$ in Fig.\ref{fig:neighbor-graph}, based
on the data graph excerpt in Fig.\ref{fig:example-graph}.
Edge $e_1$=\edgeends{Jerry Yang}{Stanford} and $label$($e_1$)=\edge{education}.
Two other edges labeled \edge{education}, $e_2$ and $e_3$, are also
incident on node \entity{Stanford}. The neighborhood
graph from a complete real-world data graph may contain many such edges
for people graduated from Stanford University.
Among these edges, $e_1$ represents an important relationship between
\entity{Stanford} and query entity \entity{Jerry Yang}, while other edges represent
relationships between \entity{Stanford} and other entities, which are
deemed unimportant with respect to the query tuple.
We formalize the definition of \emph{unimportant edges} as follows.
Given an edge $e$=$(u,v) \in E(H_t)$, $e$ is unimportant if it is
unimportant from the perspective of its either end, $u$ or $v$, i.e.,
$\text{if } e \in UE(u)\ \text{or}\ e \in UE(v)$.
Given a node $v \in V(H_t)$, $E(v)$ denotes the edges
incident on $v$ in $H_t$. $E(v)$ is partitioned into three disjoint subsets---the important edges $IE(v)$, the unimportant edges $UE(v)$ and the rest---defined as follows:
{\flushleft \vspace{-2mm}$IE(v)$=}
\vspace{-1mm}$\{e \in E(v)\ \lvert\ \exists v_i$$\in$$t, p \text{ s.t. } e$$\in$$p, ends(p)$=$\{v,v_i\},len(p)$$\leq$$d\}$;
{\flushleft \vspace{-1mm}$UE(v)$=}
\vspace{-1mm}$\{e \in E(v)\ \lvert\ e$$\notin$$IE(v), \exists e'$$\in$$IE(v) \text{ s.t. } label(e)$=$label(e')$,
$(e$=$(u,v) \wedge e'$=$(u',v)) \vee (e$=$(v,u) \wedge e'$=$(v,u'))\}$.\\
An edge $e$ incident on $v$ belongs to $IE(v)$ if
there exists a path between $v$ and any query entity in the query tuple $t$,
through $e$, with path length at most $d$.
For example, edge $e_1$ in Fig.\ref{fig:neighbor-graph} belongs to $IE(\entity{Stanford})$.
An edge $e$ belongs to $UE(v)$ if (1) it does not belong to $IE(v)$
(i.e., there exists no such aforementioned path) and (2) there exists
$e' \in IE(v)$ such that $e$ and $e'$ have the same label and they are
both either incoming into or outgoing from $v$.
By this definition, $e_2$ and $e_3$ belong to $UE(v)$ in Fig.\ref{fig:neighbor-graph},
since $e_1$ belongs to $IE(v)$. In the same
neighborhood graph, $e_4$ is in neither $IE(v)$ nor $UE(v)$.
All edges deemed unimportant by the above definition are removed from $H_t$.
The resulting graph may not be weakly connected anymore and may have
multiple weakly connected components.~\footnote{A weakly connected component of a directed graph
is a maximal subgraph where an undirected path exists for every pair of vertices.}
Theorem~\ref{th:damping-factor} states that one of the components---called
the \emph{reduced neighborhood graph}, denoted $H'_t$---contains all query
entities in $t$.
In other words, $H'_t$ is the largest weakly connected subgraph of $H_t$
containing all query entities and no unimportant edges.
Alg.\ref{alg:mqg} is applied on $H'_t$ (instead of $H_t$)
to produce $MQG_t$.
Since the techniques in the ensuing discussion only operate on $MQG_t$,
the distinction between $H_t$ and $H'_t$ will not be further noted
\begin{theorem}
\label{th:damping-factor}
Given the neighborhood graph $H_t$ for a query tuple $t$, the reduced
neighborhood graph $H'_t$ always exists
\end{theorem}
\begin{proof}
We prove by contradiction. Suppose that, after removal of all unimportant
edges, $H_t$ becomes a disconnected graph, of which none of the weakly
connected components contains all the query entities. The deletion of unimportant edges
must have disconnected at least a pair of query entities, say, $v_i$ and $v_j$.
By Def.~\ref{def:ng}, before removal of unimportant edges, $H_t$ must have at least
a path $p$ of length at most $d$ between $v_i$ and $v_j$.
By the definition of unimportant edges, every edge $e$=$(u,v)$ on $p$ belongs to
both $IE(u)$ and $IE(v)$ and thus cannot be an unimportant edge.
However, the fact that $v_i$ and $v_j$ become disconnected implies that $p$ consists of
at least one unimportant edge which is deleted.
This presents a contradiction and completes the proof.
\end{proof}
\input{sec-multituple}
\section{Related Work}\label{sec:related}
\iffalse
Our work is the first to query knowledge graphs by example entity tuples.
In the literature on graph query, the input to a query system in most cases is a
structured query, which is often graphically presented as a query graph or
pattern. Such is not what we refer to as query-by-example (QBE), because
underlyingly the query graphs and patterns are formed by using structured query
languages or other query mechanisms. For instance, \system{PathSim}~\cite{Sun+11} finds the top-\emph{k}\ similar entities that are connected to a query entity, based on a user-defined meta-path semantics in a heterogeneous network. In~\cite{YSZH12}, given a query graph as input, the system finds structurally isomorphic answer graphs with semantically similar entity nodes.
In contrast, \system{GQBE} only requires a user to provide an entity tuple, without knowing the underlying schema.
Substantial progress has been made on query mechanisms that help users construct query graphs or even do not require explicit query graphs. Such mechanisms include keyword search (e.g., \cite{KA11}), keyword-based query formulation~\cite{YCHH12}, natural language questions~\cite{Yahya+12}, interactive and form-based query formulation~\cite{Demidova+12}, and visual interface for query graph construction~\cite{GRAPHITE, GBLENDER}. Although little has been done on comparison across these mechanisms, it is an interesting future direction to conduct a usability comparison of these mechanisms and \system{GQBE}.
\fi
Lim et al.~\cite{lim_edbt13} use example tuples to find similar tuples in database tables that are coupled with ontologies.
They do not deal with graph data and the example tuples are not formed by entities.
The goal of \emph{set expansion} is to grow a set of objects starting from seed objects. Example systems include \cite{seal}, \cite{Gupta-2009}, and the now defunct \system{Google\ Sets} and \system{Squared} services
(\textsf{\scriptsize http://en.wikipedia.org/wiki/List\_of\_Google\_products}).
Chang et al.~\cite{infoNebula} identify top-\emph{k}\ correlated keyword terms from an information network given a set of terms, where each term can be an entity. These systems, except~\cite{infoNebula}, do not operate on data graphs. Instead, they find existing answers within structures in web pages such as HTML tables and lists. Furthermore, all these systems except \system{Google\ Squared} and \cite{Gupta-2009} take a set of individual entities as input. \system{GQBE} is more general in that each query tuple contains multiple entities.
Several works~\cite{TF06,Kasneci+09ming,Fang+11rex}
identify the best subgraphs/paths in a data graph to describe how several input nodes are related. The query graph discovery component of \system{GQBE} is different in important ways-- (1) The graphs in~\cite{TF06}
contain nodes of the same type and edges representing the same relationship, e.g., social networks capturing friendship between people.
The graphs in \system{GQBE} and others~\cite{Kasneci+09ming,Fang+11rex} have many different types of entities and relationships. (2) The paths discovered by their techniques only connect the input nodes. \system{REX}~\cite{Fang+11rex} has the further limitation of allowing only two input entities. Differently the maximal query graph in \system{GQBE} includes edges incident on individual query entities. (3) \system{GQBE} uses the discovered query graph to find answer graphs and answer tuples, which is not within the focus of the aforementioned works.
There are many studies on approximate/inexact subgraph matching in
large graphs, such as \system{G}-\system{Ray}~\cite{TFGE07}, \system{TALE}~\cite{tale} and
\system{NESS}~\cite{ness}.
\system{GQBE}'s query processing component is different from them on several aspects.
First, \system{GQBE} only requires to match edge labels and matching node identifiers is
not mandatory. This is equivalent to matching a query graph
with all unlabeled nodes and thereby significantly increases the problem
complexity. Only a few previous methods (e.g., \system{NESS}~\cite{ness})
allow unlabeled query nodes. Second, in \system{GQBE}, the top-\emph{k}\ query algorithm centers
around query entities.
More specifically, the weighting function gives more importance
to query entities and the minimal query trees mandate the presence of entities
corresponding to query entities.
On the contrary, previous methods give equal importance to all nodes in a query graph,
since the notion of query entity does not exist there.
Our empirical results show that this difference makes \system{NESS} produce less
accurate answers than \system{GQBE}. Finally, although the query relaxation DAG proposed
in~\cite{sihem-Yahia05} is similar to \system{GQBE}'s query lattice, the scoring mechanism
of their relaxed queries is different and depends on XML-based relaxations.
\iffalse
\textbf{keyword query}
\cite{KKD11} focuses on efficient algorithm for construction of ranked answer graphs to the keyword query on RDF graphs.
Unlike previous work such as \cite{LiOF+08}, no fix distance bound is used to define the neighborhoods of nodes. In this work, keywords are first mapped to vertex and edges in the graph. Then multiple answer graphs are constructed using a rule based component cluster creation algorithm. The answer graphs are further ranked by several relevance measures, such as compactness relevance, term relevance, node relevance, and relationship relevance.
\cite{LiOF+08} proposes an efficient and adaptive keyword search method, called EASE, for indexing and querying large collections of heterogenous data. This work models unstructured, semi-structured and structured data as graphs and proposes an extended inverted index to facilitate keyword-based search. The search is modeled as r-Radius Steiner Graph problem i.e to identify all r-radius Steiner graphs which contain all the keywords.
\textbf{with regard to (3.2): Two ways of categorizing these works. }
\textbf{(A) find similar graphs in many graphs vs. find similar subgraphs in one big graph.}
\textbf{(B) exact answers vs. approximate answers.}
\textbf{When we discuss the papers in (3.2), we should cover both RDF query processing and general graph query processing.}
The related work is in many lines of work with varying flavors. One dimension is the way data is stored in graph databases. In some work, the graph database consists of a set of graphs and the answer to a given query is a subset of these graphs. The other kind of data representation is a single large graph and the answer to a given query would be subgraphs of this large graph.
The other dimension is the similarity quotient. A lot of work has been done on retrieving (sub)graphs which are an exact match of the query graph. The other similarity metric is retrieving (sub)graphs which are approximate or inexact matches of the query graph. This section summarizes the related work based on these two dimensions.
\cite{yan04,williams1,gstring,ctree} explore exact subgraph matching in a graph database containing many target graphs. The target graphs which contain exact subgraph matches of the query graph are retrieved. The focus of \cite{yan04} was not subgraph isomorphism, but they focused on pruning the number of target graphs so as to reduce the search space for carrying out subgraph isomorphism. They use a feature called \emph{frequent substructures} to index the graphs in the database. The frequent substructures are of varying sizes and the set of all possible subgraphs of the graphs in the database is the universe of features. They use a feature size parameter to limit the size of a feature (number of edges in the feature-subgraph) as they observe that the number of feature fragments increases with the size. They prune the number of seleceted features by using a support threshold which is basically the number of graphs in the database that contain a given fragment. All feature fragments which do not have a predefined minimum support will be pruned. The high support features are further pruned using apriori-like algorithm where given two features $f_1$ and $f_2$ having the same support, if $f_1 \subset f_2$, then $f_2$ (called a redundant fragment) is pruned. They use another parameter called the discriminative fragment to select features. Finally each graph in the database is indexed based on these features. At query time, all fragments of the query graph are enumerated (here again apriori pruning techniques are applied to reduce the number of potential fragments) and only the features that are in the selected feature set is used. Intersection of the set of graphs that contain all these features is then used for subgraph isomorphism.
\cite{williams1} decomposes the graphs in the database to components and indexes each graph in terms of a component. Each component is a subgraph of the original graph $g$ and these components are nodes of a DAG. There exists a directed edge from nodes $p\rightarrow{}q$ if $p\subset{}q$ and $\nexists{}r$ s.t. $p\subset r \subset q$. Each node in the DAG is basically a subgraph of $g$ and it is represented as a in its canonical form. The DAGs of multiple graphs can also be combined to form a single DAG where some of the nodes represent the actual graphs in the database. Given a query graph $Q$, it is converted to its canonical form and is matched against the nodes in the DAG. A match indicates that $Q$ is isomorphic to a subgraph (matched node) of a graph in the database which can be reached by following the children of the matched node in the DAG. The authors also discuss about how this kind of indexing and graph representation can be used to answer similarity queries.
\cite{gstring} deals with graph specific to the domain of organic chemistry. The graphs in the database are converted and represented as strings upholding the semantics the structure of the corresponding graph has to offer. Once all graphs represented as strings, the query graph is also converted to a string (using domain specific rules) and existing string matching techniques are used to find the best match. Since this is a domain specific method,this surely cannot be used as is to solve our problem. Moreover, the rules here are specific to the structure of a graph (line, star and cycle).
\cite{ctree} discusses about representing graphs of a database in a closure-tree, a data structure developed by the authors. Each internal node in C-tree is a closure node which is obtained by merging the its children nodes (closure graphs/database graphs). The leaf nodes are the actual database graphs. The authors specify rules that can be used to merge two graphs and also define the notion of similarity and distance between graphs which can also be applied on graph closures. Given a query graph, the C-Tree is traversed to see which node in the C-tree that it is most similar to (they have a approximate definition of isomorphism). The number of nodes to be considered can be pruned as all the children nodes of a closure node can be pruned if the similarity between the query graph and the node is less than a threshold. All the graph database graphs obtained after the pruning are considered for exact/approximate subgraph isomorphism. The difference between this method and previous methods discussed is that the feature used to index the database graphs here maintain the structural information of the graph unlike in \cite{yan04} where only selected fragments of the graphs were used.
\cite{yan05} explore approximate subgraph matching in a graph database containing many target graphs. The target graphs which contain approximate subgraph matches of the query graph are retrieved. In \cite{yan05}, given a query graph all target graphs in the database which are either similar to the query graph or similar to the query graph with a permissible number of edges missing are returned as the answers. Having the provision of knocking out some edges off the original query graph brings about the approximate subgraph matching feature in this work. The graphs are modeled as a set of features and removing edges is transformed to missing features. They use a feature-graph matrix to compute the difference in features between two graphs and a edge-feature matrix to compute the permissible number of feature misses in a query graph. A feature here could be a path in a graph, discriminative fragments as in \cite{yan04} or any other feature. They prune the database graphs to obtain a smaller number of potential answer graphs using the presence of features in various graphs (common to features in the query graph). They also propose a clustering based feature selection method as they observe that neither having too many features nor having too few give a good bound on the feature misses.
Some other papers that deal with this are \cite{sapper}, \cite{saga} and \cite{tale}. SAGA \cite{saga} tries to find approximate matches for a query graph by using the measure of subgraph distance, which is a factor of structural differences (which is measured by the shortedst distance between two nodes in the graph), node mismatches and missing nodes. In order to speed up the process, the authors break down the graph into fragments and construct fragement-indexes. Given a query, that too is broken down into fragments and this index is matched against all the fragments in the target graph and the one that has the least subgraph distance is ranked higher. The same authors also proposed another system called TALE \cite{tale} which aims at doing the same for large query graphs. The indexes are constructed on every node incorporating the neigbor information in it. Given a large query graph, only the most important nodes are identified and matches for these nodes are found in the target graphs. Once corresponding node matches are found, they try to grow the complete graph following the query graph structure. Another work on similar lines is SAPPER \cite{sapper} which also used neighborhood information to index the graphs. The nodes that match vertices in the query graph are further pruned by matching random spanning trees generated from the query graph. A common feature to all these systems is that the datasource is a database of graphs and the answers is a subset of these graphs that best match a query graph.
There are fewer systems that deal with finding multiple approximate subgraphs from a single large graph. \cite{topkSubgraph} and \cite{flexibleQueryAns} deal with the problem of finding approximate subgraph matches to a given query graph from a single large data graph. \cite{topkSubgraph} aims at reducing the search space. Given a large data graph, they index the graph like a tree where all the leaf nodes are likely to have potential answer graphs to a query. Each node in the tree has a subgraph and they use the concept of node-area to denote the vertices and edges that come under a node area. They employ a feature based pruning strategy to check if the query graph is present in a node area. If not, the entire subgraph under this node in the tree is pruned out. Once they reduce the search space, subgraph isomorphism techniques are applied to find matches for a query. This work mainly concentrates on reducing the search space and not actually finding the approximate matches. \cite{flexibleQueryAns} tries to find answers to a query graph by making various approximations while conserving the semantic relatedness. Approximate subgraph matching is obtained here by carrying out label and structural approximations. The answers are ranked according to functions that evaluate approximations on data nodes, edges and query conditions. Semantic relatedness is a domain specific concept and thus the use of ontologies and taxonomies becomes necessary here. So the structural approximations is as good as the outside help that is sought after for determining the semantic relatedness. NAGA \cite{naga} is a semantic search engine. The authors propose a query language that would let users query their data source. The query model they propose is good for simple queries. But it might be a little complicated for end-users when they have to specify more complex queries. The query language has a learning curve and is not as easy as keyword search.
\cite{HuangLZ11} studies the problem of relaxing queries on RDF databases in order to acquire approximate answers. When there exist not enough answers for an exact match of RDF query, the query can be relaxed to generate more answers. For example, RDF queries such as ``x doctoralDegreeFrom Stanford'' can be relaxed to ``x degreeFrom Stanford'', or ``y typeOf proceedings'' can be relaxed to ``y typeOf books''.
This work addresses two problems: 1) ensure the quality of answers; 2) obtain top-k answers efficiently. In order to get high quality answers, similarities between relaxed queries and user queries are computed to score the answers. For the efficiency issue, two algorithms are proposed: 1) best first relaxation algorithm; 2) batch based relaxation algorithm. The first algorithm ranks all candidate relaxed queries, and execute in a best first ranking order to obtain top-k answers. This approach, however, is still not efficient enough since there are duplicate answers in the relaxed queries. In contrast, batch based algorithm would skip unnecessary queries based on the predication size of the batch by selectivity estimation.
\textbf{4. set expansion: SEISA~\cite{HeXin11}, Google Set, Google Squared, and other set expansion work}
This is similar to set expansion, but with a lot of other intricacies involved. Here we not only have to take the properties of entities like Bill\_Gates and Microsoft into account, but also the relationship between the two entities. Such a system can also be used to find answers to more complicated questions like pairs of company names and the spouse of their founders. Imagine a scenario where the spouse of only a very famous founder is known and you wish to know the spouse of many other founders and the companies they founded. Most of the work on set expansion try to expand a set of entities, but here we consider scenarios where an element in the input seed set can contain multiple entities in it. This problem of table expansion that we intend to solve here, will be able to expand a set of instances (where each instance can have multiple entities with hidden relationships between them) to a more complete set.
Recent development of Web search applications has seen large proliferation of systems that focus on querying similar entities, such as Google Sets and Google Squared\footnote{Although the demos are recently shut down by Google, both systems have been available in Google Labs Url. \url{http://en.wikipedia.org/wiki/List_of_Google_products}}. In Google Sets, a user provides several entities as a query, then the system could automatically find other entities that form a set together with the given seed entities. The new entities should either belong to the same domain or bear similar attributes with all seed entities. For example, if the user provides Honda and Toyota as seed entities, the system could find other Japanese car brand such as Mazda, Nissan, etc. If the user provides Honda Accord, Toyota Camry as seed entities, the system could find other Japanese full size sedan such as Mazda 6, Nissan Altima, etc. Another Google Labs product Google Squared has more capabilities. In Google Squared, a user could not only provide a set of entities, but also specify a set of relevant attributes of the entities. The system could then generate a table based on the given seed entities and attributes. The new entities should all be similar to the seed entities and have their attribute values showing under corresponding attributes. For example, if the user provides Honda Accord and Toyota Camry as seed entities, and specifies size, engine, and price as attributes the system could expand with other Japanese full size sedan and find the corresponding size, engine, and price for these cars. The technique used in Google Sets and Google Squared are based on information extraction. Google Sets extracts similar entities from html source files based on certain html markups such as <ul>, <ol>, and document headers. Google Squared uses extracted structured data from heterogeneous data sources available online such as Wikipedia, Freebase, and etc. for expanding the original table.
Compared with Google Sets and Google Squared, our system has following differences: first, we allow the user to use multiple entities in each seed tuple, while in both other systems there could be only one main entity in each tuple; second, we capture the relationships between the multiple entities in each seed tuple and involve the relationships into relevance ranking of the expansion, while in Google Squared such kind of relationships are not considered.
\fi
|
\section{Introduction}
In quantum gravity, classical spacetime is replaced by a quantum superposition of spacetimes, which can exhibit nontrivial properties. Here, much can be learned from the consideration of a fictitious, backreaction-free particle, probing a quantum spacetime by a random walk.
A first probe of quantum gravity effects is given by the spectral dimension, which is defined as follows: From the return probability $P_r(\sigma)$ of the random walker to his starting point after diffusion time $\sigma$, averaged over the complete spacetime, one defines a generalized spectral dimension
\begin{equation}
d_s(\sigma) =-2 \frac{\partial \ln P_r(\sigma)}{\partial \ln \sigma}.\label{d_s}
\end{equation}
This is a measure of dimensionality that agrees with the topological or Hausdorff dimension for a flat smooth spacetime for $\sigma \rightarrow 0$:
For Brownian motion on a $d$-dimensional Riemannian flat background the probability density satisfies the diffusion equation $\left(\partial_\sigma - \nabla_x^2\right)P(x,x',\sigma)=0$. The solution for the initial condition $P(x,x',0)= \delta^d(x-x')$ is then given by $P(x,x',\sigma) = (4 \pi \sigma)^{-d/2}e^{- \vert x - x' \vert^2/ (4 \sigma)}$. Accordingly $P_r(\sigma)=\frac{1}{V} \int d^d x P(x,x,\sigma) =\left( 4 \pi \sigma\right)^{-d/2} $ holds, and thus $d_s(\sigma \rightarrow 0)=d$. In contrast, the topological and the spectral dimension can differ in quantum gravity settings. The microscopic structure of spacetime can change the small-$\sigma$ behaviour of $d_s$, whereas the infrared, semiclassical limit, where $d_s = d$ must hold, is reached for large $\sigma$.
This has been observed in Causal Dynamical Triangulations (CDT) \cite{Ambjorn:2005db,Ambjorn:2005qt,Gorlich:2011ga,Benedetti:2009ge,Anderson:2011bj} and multigraph models derived from CDTs \cite{Giasemidis:2012rf,Giasemidis:2012qk}, asymptotically safe gravity \cite{Lauscher:2005qz,Lauscher:2005xz,Reuter:2011ah,Rechenberger:2012pm,Calcagni:2013vsa} and more recently in Ho\v{r}ava gravity \cite{Horava:2009if,Sotiriou:2011mu,Calcagni:2013vsa}, where a dimensional reduction from $d_s=4$ to $d_s=2$ is observed in $d=4$. Indications for the same phenomenon also exist in Loop Quantum Gravity \cite{Modesto:2008jz,Calcagni:2013vsa}, in settings with space-time noncommutativity \cite{Benedetti:2008gu} and in a model of nonlocal gravity \cite{Modesto:2011kw}. The strong-coupling limit of the Wheeler-DeWitt equation also shows a similar behaviour \cite{Carlip:2009kf}, and possible connections to Belinsky-Khalatnikov-Lifshitz-type behaviour and quantum-gravity induced focusing of light cones have been discussed in \cite{Carlip:2011tt}. An effective description in terms of fractional geometry has been attempted in \cite{Calcagni:2012rm}.
To extract the spectral dimension from a quantum gravity model, two different routes have been followed: One can either consider the flat-space diffusion equation, and modify the Laplacian accordingly to include quantum gravity effects \cite{Lauscher:2005qz}, by substituting $\nabla^2 \rightarrow \langle \nabla^2 \rangle$. A second route is open in quantum gravity models where spacetime can be represented as a graph,
consisting of vertices and edges, such as in the case of CDTs, causal sets, or spin foams. There one can study the diffusion process itself on the graph, and infer a spectral dimension from the return probability, see, e.g. \cite{Durhuus:2009zz} and \cite{Ambjorn:2005db,Ambjorn:2005qt,Gorlich:2011ga,Benedetti:2009ge,Anderson:2011bj}. This procedure can also be applied to a Lorentzian quantum gravity model, where it is nontrivial to set up a well-defined diffusion equation \cite{Debbasch}.
Here, we will present a first exploration of the spectral dimension in causal set quantum gravity. The causal set approach is based on Lorentz invariance and discreteness of quantum spacetime. A quantum spacetimes in this setting corresponds to a set of elements (spacetime points), with a relation that encodes the causal order represented by links between the elements. This partial order can be represented by a graph.
We will consider the random walk on that graph to extract the spectral dimension.
The combination of Lorentz invariance and discreteness in causal sets entails a particular form of nonlocality, which will have profound consequences for the small $\sigma$ behaviour of the spectral dimension, as we will see. In contrast to other quantum gravity approaches, we will observe an increase of the spectral dimension at small scales.
This paper is structured as follows. We introduce the basic kinematic structure of causal set quantum gravity in sec.~\ref{causet}. We then discuss analytic expressions and simulations for a random walk on different causal sets, including $d=2,3,4$ dimensional Minkowski spacetime, $1+1$ dimensional spatially compact spacetime, $2d$ de Sitter spacetime, Kleitman-Rothschild orders and transitive percolation models in sec.~\ref{ds}. We then introduce a measure of dimensionality from the meeting probability of two random walkers respecting causality in sec.~\ref{meetdim}, and conclude in sec.~\ref{conclusion}.
\section{Causal set quantum gravity}\label{causet}
Causal set quantum gravity is a fundamentally Lorentzian approach to quantum gravity, that is kinematically discrete.
The theory's basic postulate is that the fundamental nature of spacetime is a locally finite causal order \cite{Bombelli:1987aa}.
The dynamics is defined as a path-integral over causal sets, that, in common with some other approaches to quantum gravity see, e.g. \cite{Rivasseau:2012yp,Di Francesco:1993nw}, includes quantum fluctuations of topology as well as fluctuations of geometry. A proposal for the microscopic quantum dynamics in the form of a family of microscopic actions that approximate the Einstein-Hilbert action in the discrete setting has been made in \cite{Benincasa:2010ac,Benincasa:2010as, Dowker:2013vba}, allowing to evaluate the path-integral over causal sets, \cite{Surya:2011du}. For reviews on the causal set program, see, e.g. \cite{Henson:2010aq,Dowker:2005tz,Sorkin:2003bx,Dowker:aza,Surya:2011yh}.
More precisely, a causal set (causet for short) $\mathcal{C}$ is a set of elements (which are spacetime points), with a relation $\prec$, called ``precedes'', which encodes the causal structure, thus containing part of the information that is carried by the metric in the continuum. It satisfies the following axioms:
\begin{itemize}
\item transitivity: If $x \prec y$ and $y \prec z$, then $x \prec z$, $\forall x,y,z \in \mathcal{C}$.
\item non-circularity: If $x \prec y$ and $y \prec x$, then $x = y$, $\forall x,y \in \mathcal{C}$.
\item local finiteness: $\forall x, z \in \mathcal{C}$ the set $\{y \, \vert \, x \prec y \prec z, y \in \mathcal{C} \}$ is finite.
\end{itemize}
The first obviously holds for a causal relation, and the second forbids closed timelike curves.
The third axiom encodes the fundamental kinematical discreteness of causal sets:
unlike in the continuum, where an infinite number of points lie between any two given spacetime points, there is only a finite number of these in a causal set. Points at spacelike separation are unrelated in a causal set, thus the causal set corresponds to a partial order.
As the causal structure does not encode information on the conformal factor, the volume information contained in the metric is so far missing. Assigning a mean spacetime volume of $l_{\rm Pl}^d$ to each element of a causal set then recovers volume information from counting causal set elements, thus the causal set slogan: ``order+number = geometry'' \cite{RSorkin1, RSorkin2}. Here $l_{\rm Pl}=1$ is the fundamental discreteness scale, which a priori could be any scale, but most naturally lends itself to an identification with the Planck scale.
On a causal set, two distinct types of relations exist: We will call links those fundamental relations that are not implied by transitivity, i.e. those connecting ``nearest neighbors'' (where nearness refers to a notion of proper time). They are distinct from those relations implied by transitivity and constitute the transitively reduced causal set. This is represented by the Hasse diagram, which will form the central object of our investigations.
The causal structure of spacetime is nonlocal in the sense that short distances in the Lorentzian setting can translate to large distances in a Wick-rotated setting. Therefore causal sets are intrinsically nonlocal: In a causal set that is approximated by Minkowski spacetime, any point will have an infinite number of nearest neighbors. In any given frame, most of these will lie at large spatial distance \cite{Moore:1988zz,Bombelli:1988qh}. This form of nonlocality will play a crucial role in our analysis. Note that spacetime curvature can tame the causal set nonlocality: For instance in a Friedman-Robertson-Walker universe a finite number of past links exist \cite{Bombelli:1988qh,BSchmitzer}.
In causal set quantum gravity background structure such as dimensionality or topology is given up, thus in the set of all causets with $N$ elements, most causal sets do not approximate manifolds. Then the recovery of a semiclassical limit from the causal set path integral is particularly challenging and requires a nonlocal dymanics \cite{Benincasa:2010ac}. In particular, it is necessary to develop causal set estimators of continuum properties, such as the topology or the dimensionality, see, e.g. \cite{Myrheim:1978ce,Meyer,Bombelli:1987vp,Reid:2002sj, Glaser:2013pca} for estimators of the topological dimension.
\subsection{Sprinklings: Constructing causets approximated by Lorentzian manifolds}
If the fundamental structure of spacetime is a causal set, a smooth manifold emerges as a continuum approximation on semiclassical scales. As a crucial difference to, e.g. the CDT programme, causal sets are fundamentally discrete. Therefore it is not the continuum limit ($l_{\rm Pl}/l \rightarrow 0$), but the continuum approximation ($l_{\rm Pl}/ l\ll1$) that is of interest. Thus the calculation of physical quantities does not require the tuning to a second-order phase transition in the parameter space. In other words, the fundamental scale $l_{\rm Pl}$ carries a physical meaning in this case, and is not introduced as an unphysical regularisation. Observables may therefore depend on $l_{\rm Pl}$ or $N$, the total number of elements in a given causal set. We expect that if causal sets yield a realistic description of gravity, the long-distance features will approach continuum properties of (semi)classical spacetimes.
In order to find a causal set which will be approximated by a given manifold, we will employ the method of sprinklings \cite{Bombelli:1987aa}, which is based on a Poisson process: Taking a given volume $V$ in $(d+1)$ dimensional spacetime, the probability to find $n$ causal-set elements in it is given by
\begin{equation}
\mathcal{P}(n, V)= \frac{1}{n!}(\rho V)^n e^{- \rho V},\label{Poisson}
\end{equation}
where $\rho$ is the sprinkling density.
This generates a distribution that does not exhibit a preferred frame and is thus invariant under Lorentz symmetry \cite{Bombelli:2006nm}, as only a random distribution -- in contrast to a regular lattice -- can be.
A crucial property of a causal set approximating $d+1$ dimensional Minkowski spacetime is that each point has an infinite number of nearest neighbours, i.e., points it shares a link with.
This property crucially hinges on the combination of Lorentz invariance and discreteness: Discreteness implies that in any given region $R$ that is not spacelike to a given point $x$, there is a finite probability $p$ of finding $n$ elements that share a link with $x$. Lorentz invariance implies that if we boost this region $R$, we will obtain a region $R'$ that is disjoint with $R$, and must have the same probability $p$ to find $n$ elements with links to $x$. In this way be obtain an infinite number of disjoint regions and accordingly infinitely many links to $x$ \cite{Moore:1988zz,Henson:2010aq}. Of these links, many will be to points at a large spatial distance in any given frame, thus making causets non-local.
To generate a sprinkling into a region of finite volume contained in Minkowski spacetime, we first draw the number of points from the Poisson distribution \Eqref{Poisson}, and then independently choose the coordinates for each of them according to a uniform distribution. The causal relations are those induced by the underlying Minkowski spacetime, cf. fig.~\ref{sprinkle_m2}. In a first step, all causal relations are included for any time- or lightlike pair (the second has, strictly speaking, measure zero in a Poisson sprinkling), before the causet is transitively reduced.
Similarly, we can construct a sprinkling into 1+1 dimensional curved spacetime by transforming the metric to the conformally flat form $g_{\mu \nu} = \Omega^2(x) \eta_{\mu \nu}$, with $\eta_{\mu \nu}= \mathrm{diag}(-1,+1)$
and $\Omega^2(x)$ being the position-dependent conformal factor. Then, the sprinkling density needs to be rescaled by the conformal factor, which can be achieved by either constructing an appropriate mapping \cite{BSchmitzer} or rejecting excessive points \cite{Reid:2002sj}. Since the metric is conformally flat the causal links can be deduced in the same way as for Minkowski spacetime.
\section{Spectral dimension for causal sets}\label{ds}
We will now consider given classes of causal sets and study diffusion processes on these, without taking into account causal set dynamics. The spectral dimension realised in a universe described by causal set quantum gravity would be derived from the weighted return probability $\int_{\mathcal{C}} e^{i S[C]} P_r(\sigma; C)$, where $C$ denotes the causal sets that are summed over in the path-integral. Performing or approximating the path-integral is a highly challenging task \cite{Surya:2011du} which we will not attempt here.
If the dynamics implies that the expectation value over causal sets from the path integral will be a de Sitter sprinkling and the fluctuations are small, our results in sec.~\ref{deSitter} could be considered as the spectral dimension emerging from the path-integral for causal sets. Even without this dynamical information, the spectral dimension already shows very interesting behaviour.
Here, we will take causal sets as graphs, which can be probed by a random walker very much in the spirit of \cite{Durhuus:2009zz}. A random walker follows a walk, consisting of a set of causal set elements connected by links. This is in analogy to the investigations in \cite{Ambjorn:2005db,Gorlich:2011ga,Benedetti:2009ge,Anderson:2011bj}, where the spectral dimension of CDTs was determined by a random walk jumping from one simplex to the neighbouring simplices in a triangulation. A crucial difference to that setting is that steps which are purely spacelike
are not possible on a causal set, where every step is timelike (lightlike steps have measure zero in a Poisson sprinkling).
The probability $p(x,x')$ to move from a point $x$ to a point $x'$ in a causet of finite size is then given by
\begin{equation}
p(x,x') =\begin{cases}
1/n(x) &\quad \text{if } x \text{ and } x' \text{ are linked}, \\
0 &\quad \text{otherwise},
\end{cases}
\end{equation}
where $n(x)$ is the number of links starting or ending at $x$, i.e. the degree of $x$. At a given point $x$, the next point of the random walk is chosen uniformly at random from the neighbours, i.e. those points sharing a link with $x$. This random walk
does not respect the causal ordering, and diffusion can occur into the future as well as into the past. This underlines that the diffusing particle in this setting is a fictitious probe particle, and this random walk is unrelated to physical propagation of particles on a quantum spacetime\footnote{Interestingly, it has been suggested that physical particle propagation on causal sets can also be related to a particular diffusion equation \cite{Dowker:2003hb,Philpott:2008vd}.}. Note that no diffusion can occur between points at spacelike separation, as there cannot be a link between them. In this way, we get a positive definite diffusion probability even though we choose it to be proportional to the proper distance (a link on average corresponds to a proper distance of $1$). This circumvents one of the arguments why a diffusion equation on Lorentzian backgrounds is challenging to set up in the continuum, as a choice of the probability proportional to the proper time would correspond to ``negative probability'' for spacelike separations \cite{Debbasch}. Restricting the diffusion to take place within the lightcone resolves this problem.
The probability for a walk from $x$ to $x'$ in $n$ steps is then obtained as
\begin{equation}
P(x,x',n)= \begin{cases}
\sum_{p(x,x')}\Pi_{i=1}^{n}\frac{1}{n(x_i)} &\quad \mbox{ if a path $p(x,x')$ in $n$ steps exists from $x$ to $x'$, consisting of the points $x_i$},\\
0 &\quad \mbox{ otherwise}.
\end{cases}
\end{equation}
As different paths between $x$ and $x'$ constitute mutually exclusive cases, we have to sum over all possible paths. Note that for all paths from $x$ that have $n$ steps, $\sum_y\sum_{p(x,y)} \Pi_{i=1}^n \frac{1}{n(x_i)}=1$.
To evaluate the spectral dimension from $P(x,x,n)$ we must average over the starting point of the random walk. A further averaging procedure over all sprinklings into a given spacetime yields the spectral dimension associated to a class of causal sets that approximate a given spacetime.
On a discrete, lattice-like structure, a random walk will generically show the following behaviour for small diffusion times: In the first step, the return probability is zero, and becomes non-zero in the second step. Depending on the structure of the lattice, the return probability will be zero for the third step again, e.g. in the case of a two-dimensional square lattice. This oscillating behaviour of the return probability is a clear signature of a discrete setting. It implies that if the random walk is split into even and odd timesteps, and a spectral dimension is derived from the two cases, its value will differ.
For larger diffusion times, this behaviour is ``washed out'' in many settings, and the return probability becomes an approximately smooth function of the diffusion time.\footnote{This effect of the discreteness can in principle be removed by introducing a parameter that encodes the probability to remain at the same point \cite{Benedetti:2009ge}.}
To explicitly evaluate the spectral dimension, two routes are open to us, which should give the same results: Firstly, we can evaluate the spectral dimension from simulations or analytical studies of random walks on the causet, i.e., on the graph that defines it, the Hasse diagram. Secondly, we could use the embedding information in the case where the causet is approximated by a smooth Lorentzian spacetime, and derive the spectral dimension in the continuum approximation. In that case, we obtain an expression that seems rather challenging to tackle in practice \footnote{Initially the probability is peaked at the starting point and given by $P(x, x_0,0)= \delta^{d+1}(x-x_0)$.
In the next step, the random walker can jump to a given point $x_1$ if there is a link to this point. This requires, that there is a causal set element at this point, and the causal volume between the two points $x_1$ and $x_0$ is empty of causet elements (otherwise there would not be a direct link). Using the Poisson distribution, we thus obtain
\begin{equation}
P(x_1,x_0, 1) = A_0\, \rho\, e^{-\rho V(x_0,x_1)},
\end{equation}
where $V(x_0,x_1)$ denotes the volume of the causal interval between $x_0$ and $x_1$ and $P(x,x_0,1)=0$ if there is no link. $A_0$ is a normalisation constant for $x_0$ to ensure that the integrated probability is 1. In the next step, jumping to point $x_2$, the probability density must involve an integral over all possible intermediate points $x_1 \in V_1 =(J^{+}_0 \cup J^{-}_0) \cap (J_2^+ \cup J_2^-)$, where $J^{\pm}_x$ denotes the future/past of $x$, so that
\begin{equation}
P(x_2, x_0, 2)= \int_{V_1} d^{d+1} x_1 \sqrt{-g}\, A_1\, A_0\, \rho^2\, e^{- \rho V(x_0,x_1) - \rho V^{\ast}(x_1,x_2)}.
\end{equation}
This makes the construction principle clear. The only further complication arises, as the volumes other than $V(x_0,x_1)$ have to exclude the regions that have been set to be devoid of any elements in previous steps of the walk. Hence, by $V^{\ast}(x_i, x_{i+1})$ we will denote the volume $V(x_i, x_{i+1})$ minus the intersections with the $V(x_j, x_{j+1})$, $j<i$.
The probability density after $n+1$ steps will accordingly read
\begin{equation}
P(x_{n+1}, x_0, n+1)=\int_{V_n^\ast} d^{d+1} x_n \sqrt{-g} A_n \cdots \int_{V_1} d^{d+1} x_1 \sqrt{-g} A_1 A_0 \rho^{n+1} e^{-\rho V(x_0, x_1)- \ldots - \rho V^{\ast}(x_n, x_{n+1})}.
\end{equation}
}. We thus focus on simulations of random walks on causets in the following.
\subsection{Spectral dimension of sprinklings into Minkowski spacetime}
In $d+1$ dimensional Minkowski spacetime, the number of links from a given point in a causet of $N$ elements is \cite{Bombelli:1988qh}
\begin{equation}
\# {\rm links} \sim N^{(d-1)/2} \quad \mbox{($\log N$ for $d=1$)}.\label{linkno}
\end{equation}
The number of nearest neighbours thus contains global and not just local information, as $N V_{\rm Pl} = V$, where $V$ is the total volume of the spacetime approximating the causet.
This property reflects the causal-set nonlocality, and is responsible for a drastic departure of the spectral dimension from other quantum gravity settings: Due to the nonlocality a random walker will move away from the origin very quickly, cf. fig.~\ref{snap_m2}, keeping close to the lightcone.
\begin{figure}[!h]
\includegraphics[width=0.4\linewidth]{snap_m2_L100_sigma5.pdf}
\includegraphics[width=0.4\linewidth]{snap_m2_L100_sigma10.pdf}
\includegraphics[width=0.4\linewidth]{snap_m2_L100_sigma15.pdf}
\includegraphics[width=0.4\linewidth]{snap_m2_L100_sigma25.pdf}
\caption{\label{snap_m2}{Probability distribution function of a diffusion process on $1+1$ Minkowski sprinklings
with cutoff length $L=100$ at diffusion times $\sigma = 5, 10, 15, 25$. The two stages of the process are visible: low-$\sigma$ behaviour reflects the non-local nature of causal sets and yields a high value of the spectral dimension, whereas the diffusion at longer times ($\sigma \gtrsim 20$) starts to resemble the solution to the heat equation on a smooth manifold, explaining the observed asymptotic value of $d_s = 2$. Note that the colours encoding the probability correspond to different values in the different snapshots. }}
\end{figure}
More specifically, the nonlocal structure of causal sets implies that within a small number of steps
the random walker can get to a point where a typical return path consist of a large number of links, cf. fig. \ref{sprinkle_m2}. This is very different from the setting of, e.g. CDT, which is essentially local and where the number of neighbouring simplices is finite.
The consequence of the nonlocality is a strong change of the return probability in the first steps, yielding a large or even divergent spectral dimension. This follows, as the return probability is related to the inverse of the number of points at which the random walker can be at time $\sigma$. In a nonlocal setting, this number grows quickly, accordingly yielding a strong fall-off of $P_r(\sigma)$, resulting in a large $d_s$.
This behaviour is in marked contrast to other approaches to quantum gravity: There, the small-scale spectral dimension undergoes a dynamical reduction \cite{Ambjorn:2005db,Lauscher:2005qz, Modesto:2008jz, Horava:2009if, Calcagni:2013vsa}. In terms of the diffusion process this implies the property of subdiffusion, where the mean squared displacement grows slower with diffusion time than in the case of Brownian motion on a flat background. Physically, one could imagine quantum fluctuations of spacetime ``hindering'' the progress of the diffusing particle.
Here, we report the first case of superdiffusion in a quantum gravity approach. It is rooted in the fundamentally Lorentzian nature of causal sets which, combined with discreteness, yields the strong nonlocality. This allows the particle to cover larger distances (in a Euclidean sense) than in other settings, thus showing superdiffusion.
In \cite{Modesto:2009qc} it has been postulated that the presence of a smallest length scale in a Riemannian setting should lead to a decrease of the spectral dimension at small scales. Here we provide an example for the Lorentzian setting, where a fundamental length scale exists, which serves as a Lorentz-invariant cutoff, as each element of a causal set corresponds to a Planckian volume on average. The existence of this fundamental length scale together with the requirement of Lorentz invariance and discreteness actually leads to an increase of the spectral dimension on small scales.
At larger scales in our simulations, the spectral dimension drops to zero quickly. This happens on finite causets, as the random walker quickly reaches the boundary, and is ``reflected''. The probability distribution then equilibrates, yielding a constant return probability and a vanishing spectral dimension. This drop off of the spectral dimension due to the boundary is also well-known from CDT simulations \cite{Benedetti:2009ge}. In the case of causal set quantum gravity, the boundary effect is particularly severe: As explicit simulations of random walks can only be considered on finite sprinklings, there is always a finite probability to reach the boundary in a single step. Although this probability decreases with the number of points in the sprinkling,
even for large sprinklings the spectral dimension will
approach zero in a small number of diffusion steps, cf. fig.~\ref{m2_without_cutoff}.
\begin{figure}[!h]
\begin{minipage}{0.5\linewidth}
\includegraphics[width=\linewidth]{m2_without_cutoff.pdf}
\end{minipage}
\begin{minipage}{0.48\linewidth}
\caption{\label{m2_without_cutoff}Spectral dimension $d_s$ as a function of diffusion time $\sigma$ for sprinklings into $1+1$ dimensional Minkowski spacetime. Lines for three different causet sizes are shown, $N = 10^6, 10^7, 10^8$ with bands showing the $1\sigma$ statistical error. On average a random walker reaches a region close to the boundary within $\sigma \approx 30$ steps for the largest simulations. Accordingly the subsequent decrease of the spectral dimension is a boundary effect. As discussed above, the small-$\sigma$ behaviour will show a difference between $d_s$ derived from only the even or only the odd steps of a random walk, as has been observed also in CDT \cite{Benedetti:2009ge}. The two curves converge quickly on sprinklings into Minkowski spacetime. Instead of performing single random walks we compute the exact probabilities at each timestep, thus obtaining the precise values of the return probability which allows to calculate its logarithmic derivative with less uncertainty. We found it enough to average over $\sim 100$ sprinklings and starting points to achieve convergent results. Very similar volumes and averaging numbers will be used in further simulations unless stated otherwise.}
\end{minipage}
\end{figure}
The largest causets that we are able to treat due to limited computational resources have $N \approx 10^8$ elements, which corresponds to the volume of the spacetime in fundamental units. This is to be compared to a maximum volume of $2 \times 10^5$ used in the most recent CDT simulations \cite{Benedetti:2009ge}, which has proven to be large enough to avoid boundary effects. The finite, and volume independent, number of nearest neighbours in CDT in contrast to a number of nearest neighbours that grows with the volume in the case of causal sets is responsible for this major difference.
The same property implies that the small-scale value of the spectral dimension depends on $N$: Due to the nonlocality of a causet a random walker has access to the information on the total size of the causal set within the very first steps, due to \Eqref{linkno} This is crucially different from local settings, where the local movement of the random walker is independent of the total size of the quantum universe that he probes. In the causet case this dependence on $N$ is a physical one. This differs from settings where a discretisation is introduced as a regularisation, and is unphysical. Then the number of building blocks must diverge in the physical continuum limit. This limit is obtained by tuning to a second-order phase transition, where the correlation length diverges. Thus the finite size of the system will affect the results.
A finite-size scaling analysis is then in order, to obtain the continuum result from a finite simulation. In the case of causal sets, the discretisation is physical. $N$ is therefore a physical number, related to the number of spacetime points in our universe. Combined with the nonlocality of causal sets this implies that the dependence of $d_s$ on $N$ at small $\sigma$ is physical. Away from a second-order phase transition, finite-size effects in lattice-like systems do usually not play an important role, but in the causal set case, the nonlocality implies that the system size $N$ matters, and affects the small-scale value of $d_s$.
To analyze the sprinklings into Minkowski spacetime systematically, we will now introduce an infrared cutoff, see fig.~\ref{sprinkle_m2}. Its main purpose is to move the onset of the boundary effect to larger $\sigma$, thus revealing a non-trivial intermediate-$\sigma$-regime for $d_s$.
\begin{figure}[!here]
\includegraphics[width=0.35\linewidth]{sprinkle_m2.pdf}
\includegraphics[width=0.35\linewidth]{sprinkle_m2_cutoff.pdf}
\caption{\label{sprinkle_m2}{Sprinklings into $1+1$ dimensional Minkowski spacetime with unit density: without any cutoff (left), and with infrared cutoff $L=5$ (right). The causet elements and links are shown.}}
\end{figure}
This cutoff is chosen in a particular frame, where we use the Wick-rotated Euclidean metric to discard links stretching beyond the infrared cutoff distance $L$.
The effect of the cutoff is to reduce the causal set non-locality, and make it resemble a local lattice structure with a bound on the maximal number of links at a given point even for a sprinkling into Minkowski spacetime without boundaries, see fig.~\ref{sprinkle_m2}. The cutoff can thus be understood as a cutoff on the degree of the vertices, which is infinite on a Minkowski spacetime. Its additional effect is to reduce the boundary effect: Introducing a cutoff we can explicitly simulate random walks on the causal set where the random walker will not reach the boundary of the sprinkling within a given number of steps. This allows us to probe a regime of intermediate $\sigma$ without boundary effects. Without a cutoff, much larger causets would be necessary to access this regime, requiring significantly more computational resources.
Besides, using a cutoff can be understood as an approximation of the possible effect of spacetime curvature: As curvature affects the global light-cone structure,
curvature can affect the number of links \cite{Bombelli:1988qh}. Other settings with a finite number of links include finite-volume spacetimes, such as in the case of compact extra dimensions.
Taking the limit $L \rightarrow \infty$ provides a controlled way to recover the full causal set and its non-locality. In that limit, the preferred frame selected by the cutoff is removed and Lorentz invariance is restored.
\begin{figure}[!here]
\includegraphics[width=0.48\linewidth]{m2_dim.pdf} \includegraphics[width=0.48\linewidth]{m_all_dim.pdf}
\caption{\label{m2_dim}Left panel: Spectral dimension for $1+1$ dimensional Minkowski spacetime { ($N \approx 10^8$)} as a function of $\sigma$, for different values of the cutoff. After initial oscillations the curves for even and odd diffusion steps merge at $\sigma \sim 20$ with $d_s>2$ and decrease with $\sigma$.
Higher cutoff values yield larger initial values of the spectral dimension. Independently of the cutoff, all the curves approach $d_s = 2$ in the intermediate-$\sigma$ regime, before dropping to 0 at even larger values.\\
Right panel: Spectral dimension for sprinklings into Minkowski spacetimes with different dimensionality. Here we used a low cutoff value of $L=5$ in order to make the computation for higher dimensions possible. The small $\sigma$ regime exhibits the split of the spectral dimension in odd and even time steps.
The spectral dimension
approaches
the topological dimension in all the cases. { The higher dimensional results are less uncertain, as the effective volume probed becomes larger and fluctuations are less important.} Here we set $L$ constant, so $N \sim L^d$ and hence larger values of $d$ probe larger $N$ which gives a smaller uncertainty as the Poisson noise goes as $\Delta N \sim \sqrt{N}$, so the
statistical error $\Delta N/N \sim 1/\sqrt{N}$.}
\end{figure}
In our simulations with a finite cutoff,
the intermediate-$\sigma$ regime of the
spectral dimension approaches the topological dimension, cf. fig.~\ref{m2_dim}.
This is to be expected, as the spectral dimension can be used to test whether a smooth classical spacetime emerges from a quantum gravity setting at large scales, in which case $d_s \rightarrow d$.
As a major difference of the causet to a regular lattice, where the asymptotic value $d_s \rightarrow d$ is approached from below, the small-scale value for the spectral dimension on a sprinkling into Minkowski spacetime is larger than the topological dimension, due to the residual nonlocality of causal sets even in the case with a cutoff, unless we choose a cutoff $L \geq l_{\rm Pl}$, as in the right panel of fig.~\ref{m2_dim}.
In the same way, we can study higher-dimensional Minkowski spacetime, cf. fig.~\ref{m2_dim}. The higher-dimensional cases are clearly limited computationally, since the volume needed to accommodate for the diffusion without boundary effects gets much larger. Hence we are limited to simulations with small cutoff values. Nevertheless, the large-$\sigma$ asymptotic behaviour demonstrates how the spectral dimension tends to the topological one.
Let us consider the spectral dimension as a function of the cutoff and the diffusion time $\sigma$: For very low values of the cutoff, when most links are cut off, the spectral dimension is lower than $d$ for small $\sigma$.
In this case the cutoff is so severe that it effectively reduces the dimensionality of the causal set.
For larger values of the cutoff, the causal set non-locality becomes apparent. Its effect can be clearly seen by considering $d_s (\sigma_{\ast})$ as a function of $L/l_{\rm Pl}$. Here we define $\sigma_{\ast}$ as the (approximate) point where the even-$\sigma$ and odd-$\sigma$ curves are already merged. As expected, $d_s (\sigma_{\ast})$ grows for increasing $L/l_{\rm Pl}$. The rate of growth becomes slower if we choose larger values of $\sigma$ where to measure $d_s$, cf. fig.~\ref{m2_d_star}.
The limit of a Lorentz invariant sprinkling into Minkowski spacetime accordingly shows a divergent small-scale spectral dimension, as expected. For a finite causal diamond in Minkowski spacetime, $d_s>d$, with a finite value for $d_s$, holds. The large-scale behaviour indicates that the spectral dimension will indeed yield $d_s=d$ in an intermediate regime. For a finite causet, we expect the drop-off to $d_s=0$ in the limit $\sigma \rightarrow \infty$.
\begin{figure}[!h]
\includegraphics[width=0.5\linewidth]{m2_d_star.pdf}
\caption{\label{m2_d_star}{ Value of $d_s$ measured at $\sigma_\ast=20,25,30$ (right after the even and odd $\sigma$ lines merge) against the cutoff distance $L$ {{for $N \approx 10^8$}}. It is evident that the early-$\sigma$ value follows the power law $d_s(\sigma_\ast=20) \sim L^{0.24}$ and hence will be divergent for $L \rightarrow \infty$. In this limit we
expect that the large-$\sigma$ behaviour of $d_s$ will
resemble
the case with a finite cutoff.}}
\end{figure}
\FloatBarrier
\subsection{Spectral dimension of sprinklings into spacetimes with extra dimensions}
Let us consider a spacetime with a compact dimension, namely $\mathbb{R} \times S^1$. In contrast to Minkowski spacetime, its causal structure
implies that each element only has a finite number of links. This follows by considering a point $x$ and integrating the probability that it has a link to some point $x'$, $e^{- \rho V(x,x')}$ over the infinite volume of $\mathbb{R} \times S^1$, which yields a finite number. We thus expect a finite spectral dimension $d_s>d$ at small $\sigma$, even at infinite volume. For $\sigma \gg r$, where $r$ is the compactification radius, we expect an effectively one-dimensional diffusion process.
In our simulations, the spectral dimension reduces to $d_s=1$ at large $\sigma$, in accordance with our expectation, cf. fig.~\ref{cyl2d_dim}.
We expect that for large compactification radius $r \gg l_{\rm pl}$ there should be an intermediate plateau regime, where $d_s=2$, and $d_s=1$ only for larger diffusion times. In this regime, most links do not extend around the whole radius of the cylinder, but the causal set starts to resemble a sprinkling into flat Minkowski spacetime without compactification.
It turns out that for computational reasons this regime is hard to reach, as it requires very large sprinklings. We therefore use the following trick to mimic that regime: Introducing a cutoff $L$ beyond which we discard links has the effect to remove links which extend very far around the cylinder. Accordingly this approximates the causal set for a larger $r \gg l_{\rm pl}$.
In that case our simulations explicitly show the emergence of the 2-dimensional plateau before going to $d_s=1$, cf. fig.~\ref{cyl2d_dim}.
\begin{figure}[!h]
\includegraphics[width=0.48\linewidth]{cyl2d_dim.pdf} \includegraphics[width=0.48\linewidth]{cyl2d_d_star.pdf}
\caption{\label{cyl2d_dim}Left panel: Spectral dimension in the case of 2d cylinder space with $2 \pi r = 1000$ as a function of diffusion time $\sigma$. We show the spectral dimension for a cutoff $L=50$ (green dot-dashed line) and $L=100$ (red dotted line), and the case without a cutoff (blue dashed line). In all cases, the spectral dimension obeys $d_s\gg 2$ at small $\sigma$.
The introduction of the cutoff reveals a plateau that becomes more extended with lower values of $L$.\\
Right panel: Spectral dimension in 2d cylinder space evaluated at $\sigma_\ast=20,25,30$ as a function of the compactification radius. We find a fit of $d_s(\sigma_\ast=20) \sim (2\pi r)^{0.35}$ that indicates the small-scale $d_s$ would diverge as we consider larger and larger radii, thus agreeing with the spectral dimension on $1+1$ Minkowski spacetime. Note that we expect the power law fit to hold only for sufficiently large $r$, and hence we neglected $2\pi r < 200$ in the analysis.}
\end{figure}
The small-$\sigma$ value of the spectral dimension $d_s(\sigma_\ast)$ satisfies $d_s>d$, but remains finite. As the limit $r \rightarrow 0$ corresponds to the limiting case of 2$d$ Minkowski spacetime, the spectral dimension should increase as a function of compactification radius $r$. We find that a power-law with $\sigma$ dependent exponent fits the data well, see fig.~\ref{cyl2d_dim}, suggesting that the spectral dimension diverges in the limit of infinite volume and $r \rightarrow \infty$. As the small-$\sigma$ value of the spectral dimension depends on $r$, we observe that the spectral dimension is sensitive to the global topology of the spacetime. Again this is due to the nonlocality of causal sets, which implies that a random walker can access global information such as the total volume or the global topology, already in the first steps of a random walk.
\FloatBarrier
\subsection{Spectral dimension of sprinklings into de Sitter spacetime}\label{deSitter}
The spectral dimension of de Sitter spacetime is of particular interest, as its cosmological patch describes
the inflationary universe and the $\Lambda$ dominated late-time expansion of the universe.
For the cosmological or planar patch of 1+1 dimensional de Sitter, which covers half the space, namely
the region inside and including the lightcone emanating from a point on spacelike past infinity $\mathscr{I}^-$, we can use conformal coordinates $(\eta, x)$, in which the
metric is given by $g_{\mu \nu} = \left(\frac{\alpha}{t}\right)^2 \eta_{\mu\nu}$. Here $\alpha$ is the de Sitter radius and the coordinates have range $t \in [-\infty, 0]$, $x \in [-\infty,\infty]$.
Creating a sprinkling into this patch works as follows: After sprinkling into a given interval $[x_0,x_1] \times [t_0,t_1]$, the links follow from the causal structure of Minkowski spacetime. To account for the difference to de Sitter, it is only necessary to adapt the sprinkling density: As $\sqrt{-g} = (\alpha/t)^2$, the sprinkling density will not be constant, cf. fig.~\ref{sprinkle_ds2}. Following \cite{BSchmitzer, Aslanbeigi:2013fga} we achieve this by first sprinkling with constant density into a unit square, denoted by $A$ and then using a mapping $f(A) \rightarrow \Omega, f(x) \mapsto z$ for the coordinates, such that $\int_A d^2x = \frac{1}{V(\Omega)} \int_{f(A)} d^2 z \sqrt{-g}$.
To check whether global properties can affect the spectral dimension, we will consider random walks on a sprinkling into the planar patch, as well as a sprinkling into the global patch. The metric for global de Sitter spacetime can, in 2 dimensions, also be brought into a conformally flat form and reads $g_{\mu \nu}= - \frac{\alpha^2}{\cos(t)^2}\eta_{\mu \nu}$, where $t \in [- \pi/2, \pi/2]$. While a causal diamond in the cosmological patch can be mapped onto a causal diamond in the global patch due to the symmetries of de Sitter spacetime, the global properties of both patches differ \cite{Aslanbeigi:2013fga}.
Nevertheless we do not observe any qualitative differences of the spectral dimension between the two patches in our simulations, hence in the following discussion we will consider the planar patch only.
In the cosmological patch, the number of past-directed links is finite at every point in a causal interval \cite{BSchmitzer}.
As the number of future-directed links from a given point will however not be finite unless we restrict ourselves to a finite volume, we again resort to a regularisation to evaluate the spectral dimension.
Following our previous considerations, we introduce an infrared cutoff, again measuring the distance of points after a Wick-rotation. The cutoff is defined as $L(\eta) = \rho_0 \eta L^\ast$, which is position-dependent to account for the changing density by keeping the number of future-directed links from each element constant, cf. fig.~\ref{sprinkle_ds2}.
\begin{figure}[!here]
\includegraphics[width=0.33\linewidth]{sprinkle_ds2.pdf}
\includegraphics[width=0.33\linewidth]{sprinkle_ds2_cutoff.pdf}
\caption{\label{sprinkle_ds2}{Sprinklings into $(1+1)$-dimensional de Sitter spacetime (planar patch) with unit density and radius $\alpha=5$: without any cutoff (left), and with infrared cutoff $L^\ast=0.5$ (right).}}
\end{figure}
If we choose a curvature scale that is considerably larger than the fundamental length, the small $\sigma$ behaviour of the spectral dimension should approach that of Minkowski spacetime.
Accordingly, we can observe the spectral dimension to take a
value $d_s>d$ at small $\sigma$
and then approach the intermediate value $d_s=d$.
The large $\sigma$ behaviour will be affected by the presence of curvature.
We tentatively conjecture the late-$\sigma$ increase to be due to the presence of the curvature, cf. fig.~\ref{ds2_dim}. Choosing a larger value of the cutoff implies that the random walker is affected by the curvature at earlier diffusion times.
\begin{figure}[!h]
\includegraphics[width=0.48\linewidth]{ds2_dim_low.pdf}
\includegraphics[width=0.48\linewidth]{ds2_dim.pdf}
\caption{\label{ds2_dim} Spectral dimension $d_s$ as a function of the diffusion time $\sigma$ in $1+1$ de Sitter spacetime (planar patch) with radius $\alpha=10^3$, starting point $(\eta,x) = (1,0)$, and three different cutoff scales $L^\ast = 0.01, 0.03, 0.1$. Similar to the diffusion on a flat background we observe an early increase of the spectral dimension on small scales, however on large scales its value increases compared to the topological dimension of $2$ which we tentatively attribute to the presence of curvature. We observe no qualitative differences to the global patch.}
\end{figure}
\FloatBarrier
\subsection{Spectral dimension of KR orders}
Kleitman-Rothschild (KR) orders are an important subclass of all causets with $N$ elements: They consist of three layers of elements, with approximately $ N/2$ elements in the middle layer, and approximately $ N/4$ elements in the upper and lower layer. Each element in the middle layer has probability $1/2$ to share a link with any given element in the upper and lower layer. There are no direct links between the lower and the upper layer, nor within each of the layers, see fig.~\ref{KRorder}. In the limit of large $N$, KR orders are entropically favored, since their number grows as $e^{N^2/4}$, similar to the total number of causal sets of order $N$, which also grows as $e^{N^2/4}$ at leading order \cite{Kleitman}. Due to their structure, KR orders do not approximate manifolds \cite{Rideout:1999ub}, as they only contain three discrete ``moments in time".
\begin{figure}[!here]
\includegraphics[width=0.15\linewidth]{KRorder.jpg}
\caption{\label{KRorder}Example of a KR-type order for small $N$.}
\end{figure}
To test whether the spectral dimension reflects the non-manifoldlikeness, let us consider a random walk averaged over KR orders: Starting from any given point, the first step takes us into an element in the neighbouring layer. The second step can bring us back to the starting point, with a probability of $P(x,x,2)= 2/N$ ($N>4$), when averaged over KR orders. This probability is constant in the three layers, as it is the probability for the existence of a given link ($1/2$), multiplied by the average number of links, which is $N/4$ for all three layers. This pattern continues for arbitrarily large $\sigma$:
\begin{equation}
P(x,x,\sigma) = \begin{cases}
\frac{2}{N}, &\quad \sigma = 2k\\
0, &\quad \sigma =2k+1.
\end{cases}
\end{equation}
This oscillating behaviour is a manifestation of a discrete setting. In the case of KR orders, it is present for arbitrarily large diffusion times $\sigma$,
signaling the non-manifoldlikeness of this class of causets.
Let us introduce a finite probability $\chi<1$ that the diffusion particle remains at the same point. This smoothens the odd- and even-$\sigma$ curves, and changes the early-$\sigma$ behaviour of the return probability (which decays as $\chi^\sigma$), however its asymptotic value still converges to a constant, and hence the spectral dimension vanishes on large scales. By explicitly simulating KR orders we confirmed this fact, and also observed that larger orders yield a faster decay of $d_s$.
\subsection{Spectral dimension of transitive percolation models}
Much work has been done to construct dynamics for causal sets, which lead to the development of so-called Classical Sequential Growth models \cite{Rideout:1999ub, Rideout:2001kv}. Of particular interest is a specific subclass of these, namely the transitive percolation model \cite{Rideout:2000fh}.
There, a causal set is ``grown'' successively in a label-invariant way, which is the remainder of diffeomorphism invariance in the discrete setting: Given an $N$ element causal set, a new element will be included. The probability for it to have a link with any element in the existing causal set is fixed to a small, non-vanishing value $p$. These models are attractive as they define an intrinsic (albeit classical) dynamics, and can be shown not to contain KR orders, as the probability of obtaining a chain of length greater than $3$ is always non-zero \cite{Rideout:1999ub}.
Further modifications to the model aiming at obtaining a $(3+1)$-dimensional continuum limit have also been studied \cite{Rideout:2000fh}. The causal set constructed in this way contains an infinite number of so-called posts, i.e. single elements between which the universe cycles through phases of expansion, stasis and contraction \cite{Bollobas2}, and it has been suggested that the early expansion phase after the post may resemble $(3+1)$-dimensional de Sitter spacetime \cite{Ahmed:2009qm}. The resulting causet has a Myrheim-Meyer dimension of one \cite{Bollobas1}.
Let us now consider a random walk averaged over causal sets of cardinality $N$ grown in this way. Starting from any given point, the probability for a link to any other point is $p$. Thus, the probability for the random walker to return to the initial point is always $p/(Np) = N^{-1}$, independent of the diffusion time and position. Accordingly the spectral dimension will be zero for this class of models.
If we consider only a single transitive percolation, instead of averaging over the whole class, the large-$\sigma$ limit obtained from averaging over the starting point will also be $d_s=0$,
using the same argument.
We can conclude that the percolation model does not give a finite spectral dimension, which supports the point that the resulting causal sets are not manifold-like.
This has implications for the path-integral over causets: As for large $\sigma$, $d_s=0$ for transitive percolation models and KR orders, these do not yield a contribution to $\langle d_s \rangle$, irrespective of the microscopic dynamics. This suggests that the expectation value of observables such as $d_s$ could actually show a large-scale behaviour that approximates the expected semiclassical result.
Furthermore, this is an interesting case that illustrates how the spectral and the Myrheim-Meyer dimension can differ.
Our findings on KR orders and transitive percolation models suggest that the spectral dimension could be used as a measure of manifold-likeness, as it vanishes on causets which do not approximate manifolds.
\section{New probe of quantum spacetimes: Causal spectral dimension}\label{meetdim}
In the standard setup that we have considered so far, the diffusing particle explicitly violates causality, as it can propagate
forwards and backwards in time. In order to introduce a probe of quantum geometry that is closer to what a physical probe, i.e. a physical particle propagating on a quantum spacetime (neglecting backreaction) could be,
we introduce a new dimensional estimator. It is based on a random walk that respects the time-ordering of spacetime and respects causality by only propagating within the local lightcone. It is thus applicable to Lorentzian quantum gravity models and we will study it for causal sets here.
While we so far have treated the causal set as a given graph, on which a probe particle diffuses, we will focus on the interpretation of the links as causal links in this section, imposing causality on the random walks and restricting the movement of the random walker to be forward in time.
Then the return probability to the starting point is clearly zero, making the spectral dimension ill-defined.
We therefore consider two causal random walks, and extract a notion of dimensionality from the probability that the two random walks meet at a given diffusion time $\sigma$.
Let us first consider this process on a flat $d+1$ dimensional space and show how to extract the dimensionality from the meeting probability.
As we impose causality, the movement in the time-direction is restricted to be forward in time and within the local lightcone. The stochastic process underlying such a situation is well-known, and corresponds to a biased random walk, i.e. the probability to move forward and backward in time are no longer equal. {Using light-cone coordinates, we can straightforwardly set up the discrete approximation to this process on a lattice, and then take the continuum limit by taking the lattice spacing to zero.
We thereby arrive at the drift-diffusion equation in $d+1$ dimensions, where the drift term enforces causality at large diffusion times:
\begin{equation}
\partial_{\sigma} P(\vec{x},t,\sigma) = - v \partial_t P(\vec{x},t,\sigma) +D_x \nabla_{\vec{x}}^2 P(\vec{x}, t, \sigma) + D_t \partial_t^2 P(\vec{x}, t, \sigma).
\end{equation}
Herein $t$ is the Wick-rotated time coordinate, $v$ is the drift velocity, and $D_x$ and $D_t$ are diffusion constants which can be set to 1 by a choice of units. This equation is also known as a special case of the Smoluchowski equation, which describes, e.g. the diffusion process of charged particles under the influence of an electric field. The solution for $D_x=D_t=1$ and initial condition $P(\vec{x}, t ,0)= \delta^d(\vec{x})\delta(t)$ is then given by \cite{Montroll}
\begin{equation}
P(\vec{x},t,\sigma) = \frac{1}{(4 \pi \sigma)^{\frac{d+1}{2}}}e^{-\frac{\vec{x}^2}{4 \sigma}} e^{-\frac{(t-v \sigma)^2}{4 \sigma}},\label{biasedrandomwalk}
\end{equation}
which respects causality at large diffusion times (Alternatively, one can consider the Ornstein-Uhlenbeck process, which realises diffusion in the Lorentzian case by a diffusion process in momentum space, from which the spacetime-process is inferred, see, e.g. \cite{Debbasch}. The large-$\sigma$ limit then agrees with \Eqref{biasedrandomwalk}).
Then the probability for two random walkers to meet at some point after a diffusion time $\sigma$ is $P_{\rm meet}(\sigma) = \int dt \int d^d x\, P(\vec{x}, t, \sigma)^2 \sim \sigma^{\frac{-(d+1)}{2}}$. Accordingly we can define a causal spectral dimension
\begin{equation}
d_{c\, s}(\sigma) = - 2 \frac{\partial \ln P_{\rm meet}(\sigma)}{\partial \ln \sigma},
\end{equation}
for which $d_{c\, s} (\sigma) =d+1$ in the case of a flat background in analogy to the spectral dimension.
We thus propose $d_{c\, s}$ as a useful novel dimension estimator in quantum gravity models. As it has a built-in causality property, we conjecture that it can differ from the usual spectral dimension, and provides an interesting probe of quantum geometry that could also be worth studying in other approaches to quantum gravity.
In particular, the causal spectral dimension should be sensitive to nontrivial structure in the time-direction, such as a time-dependent background, or the existence of a preferred foliation.
\subsection{Causal spectral dimension on causal sets}
On the Hasse diagram representing the causet, the causal random walk is straightforward to implement: As links only exist for causal relations, a random walker following the links automatically stays within the local lightcones. The only remaining restriction is to observe the directionality of the links.
As in the case of the spectral dimension, we introduce an infrared cutoff in the case of $1+1$ dimensional Minkowski, de Sitter, and cylinder spacetimes, cf. fig.~\ref{meet_all_dim}. In all three cases the causal spectral dimension turns out to agree with the topological dimension for large $\sigma$. This provides evidence that the causal spectral dimension is a useful novel dimension estimator.
\begin{figure}[!here]
\includegraphics[width=0.49\linewidth]{meet_all_dim.pdf}
\includegraphics[width=0.49\linewidth]{meet_cyl_dim.pdf}
\caption{\label{meet_all_dim} Causal spectral dimension $d_{c\,s}$ for different causal sets: On the left, sprinklings into $1+1$ dimensional Minkowski ($L=100$, blue dashed line), de Sitter ($\alpha/l_\mathrm{Pl} = 10^3, L^\ast = 0.003$, green dot-dashed line), as well as the transitive percolation model ($N=10^6, p=0.0003$, red dotted line); on the right, cylinder spacetime ($2 \pi r = 300, L=40$). The results for the first three agree with the behaviour of spectral dimension. In contrast, the transitive percolation model yields a non-vanishing asymptotic value of $d_{c\,s} = 1$, whereas $d_s=0$ for this model. We observe that the causal spectral dimension requires up to $10^4$ averages over sprinklings and starting points, since the uncertainties in single simulations become bigger compared to the spectral dimension. Additionally, $d_{c\,s}$ at very low values of $\sigma$ carries significant errors due to calculation of the logarithmic derivative and hence is not plotted.}
\end{figure}
We next consider non-manifoldlike causets, where we find interesting differences to the spectral dimension.
The causal spectral dimension of KR orders is not well defined, as the propagation of diffusing particles is limited to just three time-steps.
We can study the behaviour of $d_{c\,s}$ on transitive percolation models by performing simulations, where we grow a causet of $N=10^6$ elements. The results indicate that after an initial increase in the causal spectral dimension it quickly converges to $d_{c\,s}=1$ on larger scales, thus agreeing with Myrheim-Meyer dimension \cite{Bollobas1}, in contrast to the spectral dimension, cf. fig.~\ref{meet_all_dim}. Due to computational limitations, we cannot study much larger causets, on which an intermediate regime with $d_{c\, s} \approx 2$ could in principle emerge.
A scale-dependent dimensionality with a large-scale dimension of one in these models has previously been discussed in \cite{Rideout:2001kv, Brightwell}.
\begin{figure}[!here]
\begin{minipage}{0.5\linewidth}
\includegraphics[width=\linewidth]{tower_dim.pdf}
\end{minipage}
\begin{minipage}{0.48\linewidth}
\caption{\label{tower_dim}Spectral dimension $d_s(\sigma)$ (blue solid line) and causal spectral dimension $d_{c\,s}(\sigma)$ (green dashed line) for a crystalline causet as functions of diffusion time $\sigma$. Due to its regular shape it was enough to perform a single simulation with $N=10^6$ causet elements. We set the size to be larger than the extent of the diffusion to avoid effects of cyclicity of $S^1$, which has been studied separately. Only the even-$\sigma$ curve of spectral dimension is plotted, since all the odd-$\sigma$ values of $P_{meet}$ vanish. After a region of reduced dimensionality for small diffusion times the dimensional estimators converge to: $d_s=2$, $d_{c\,s}=1$. This illustrates how the two can in principle give different results.}
\end{minipage}
\end{figure}
Let us now consider another interesting class of causets, namely crystalline causets, a.k.a. ``tower of crowns'', which are constructed as follows \cite{Major:2009cw}: Taking $m$ antichains (i.e. causal set elements with no links), in which the elements are labeled by $i$, then the $i$th element in the $m$th layer precedes the elements $i$ and $i+1$ mod $m$ in the $m+1$st layer. This yields a regular-lattice-like structure that embeds into $\mathbb{R} \times S^1$. This structure clearly exhibits a preferred time-foliation and is therefore not Lorentz invariant. Interestingly, in this setting the spectral and the causal spectral dimension differ. The reason is simple to understand: The allowed moves for the causal random walk are to hop from element $(m,i)$ to $(m+1,i)$ or $(m+1,i+1)$. There is a vanishing probability to remain in the same layer, or to jump over a layer. Accordingly there is no smearing of the probability density in $m$, it will simply be $\sim \delta(m-\sigma)$ in that direction. This implies, that the causal random walk is effectively only one-dimensional. Accordingly we expect to find $d_{c\, s}=1$ in contrast to $d_s=2$. We confirm this expectation by explicitly simulations of $d_s$ and $d_{c\, s}$, see fig.~\ref{tower_dim}.
This example explicitly confirms, that the causal spectral dimension and the spectral dimension can differ and are independent of each other. In particular, the value of the causal spectral dimension for the crystalline causets is clearly related to the existence of a preferred time foliation, and the corresponding restriction in the allowed steps. We speculate, that in quantum gravity settings with a preferred time foliation such a difference between the causal spectral dimension and the spectral dimension could be common.
It would be interesting to investigate the causal spectral dimension also in other settings, such as, e.g. CDTs. As the local lightcones are well-defined in every building block of a triangulation, causal diffusion can be implemented. This could provide further interesting insight into the microscopic structure of spacetime in CDTs. In particular, it would be interesting to find whether the causal spectral dimension differs when no preferred time foliation is present in that setting, as in \cite{Jordan:2013awa}.
\section{Conclusions}\label{conclusion}
We have studied the spectral dimension $d_s$ in causal set quantum gravity. This probe of quantum spacetimes has received considerable interest, as the small-scale value in several different approaches to quantum gravity has been shown to be equal, and also smaller than the topological dimension $d$. Here, we derive the spectral dimension for causal sets by simulating random walks on the Hasse diagram defining a causal set. The causal set approach to quantum gravity combines Lorentz invariance with discreteness, and posits that the fundamental quantum nature of spacetime is a causal set, i.e. a set of elements (spacetime points), with a partial order (causal relations), which is locally finite, and therefore kinematically discrete. The combination of Lorentz invariance with discreteness implies nonlocality, as elements at large spatial distance -- in a given frame -- will be connected by links.
Our study focuses on the kinematical level, where we consider causal sets that are approximated by manifolds ($\mathds{M}^2, \mathds{M}^3, \mathds{M}^4, \mathds{R} \times S^1, 1+1 \mathrm{\ de\ Sitter}$). We find that, in contrast to other approaches to quantum gravity, causal sets do not show a dynamical dimensional reduction at small scales. Instead, the spectral dimension shows an increase at small scales, $d_s>d$. We link this behaviour to the nonlocality in causal sets: A random walker has access to a large part of a causal set already within a few time-steps, in contrast to a local setting, yielding a large spectral dimension. We tentatively suggest that such an increasing spectral dimension could occur in other Lorentzian quantum gravity approaches: Whenever the proper distance is used to define a notion of nearest neighbours, an infinite number of nearest neighbours is to be expected in a spacetime of infinite volume. Then some of those nearest neighbours lie at arbitrarily large spatial distances in a given frame, and the diffusion process will exhibit superdiffusion and an increased small-scale value of the spectral dimension.
Furthermore, the nonlocality implies that global information, e.g. on the total spacetime volume, is stored in the number of nearest neighbours of an element. Accordingly the small-scale value of the spectral dimension is sensitive to global properties of the quantum spacetime, and increases with the total volume, e.g. in the case of causal sets that are approximated by Minkowski spacetime.
We observe that, using an infrared cutoff that we impose on the maximal length of a link using the Wick-rotated metric and thus selecting a preferred frame to make the evaluation computationally feasible, the large-scale spectral dimension approaches the topological dimension for manifoldlike causal sets. This is similar to the case of lattice-like discrete approximations of a continuum spacetime, where the spectral dimension equals the topological dimension for large diffusion times.
We further observe that the spectral dimension approaches zero for large diffusion times in all cases of non-manifoldlike causal sets that we have considered (KR orders, transitive percolation models). On the basis of these two observations we conjecture that the spectral dimension is a useful dimensional estimator in causal set quantum gravity, and can be used as a measure of manifoldlikeness.
We conjecture that an increase, instead of a reduction, of the spectral dimension at small scales could be common to approaches to quantum gravity which are nonlocal, although the precise form of the nonlocality is probably important \cite{Modesto:2011kw}. It would be interesting to infer properties of the spectral dimension, e.g. for the phenomenological model of relative locality \cite{AmelinoCamelia:2011bm}, which is conjectured to carry properties of the full quantum gravity regime. In this context, it would also be of interest to investigate whether additional data carried by the causal set, which can, e.g. correspond to momentum-information as in \cite{Cortes:2013pba}, can alter the spectral dimension.
\newline
Moreover, we introduce a new probe of quantum spacetimes by considering only random walks that obey causality. The allowed moves of the random walk have to be forward in time and can only lie within the local lightcone. This new probe of quantum geometry is adapted to a Lorentzian setting, and well-suited to study causal sets. A notion of dimensionality, which we call the causal spectral dimension, can then be derived from the meeting probability of two random walkers. We discuss the causal spectral dimension for continuum flat spacetime, where it equals the topological dimension. We then show by explicit simulations, that the same is true for causal sets which are approximated by manifolds, and where the large-scale causal spectral dimension equals the topological one. The small-scale value shows an increase of the causal spectral dimension, which again is connected to the causal set nonlocality. Finally, we show an explicit example where the causal spectral dimension and the spectral dimension differ, in particular $d_{c\, s} <d_s$. In this case, this difference can be connected to the existence of a preferred time foliation in these causets. The preferred foliation implies that in time direction, the diffusion process does not lead to a spread of the probability density, which instead is concentrated on one spatial hypersurface at each step. This implies a causal spectral dimension which is smaller than the spectral dimension.
We therefore propose the causal spectral dimension as a new, independent probe of quantum spacetime, that could be insightful to study in other Lorentzian approaches to quantum gravity. As its value can differ from the standard spectral dimension, a combination of both probes can yield enlightening insights into quantum gravity models.
Our investigation of the causal spectral dimension is a first implementation of diffusion processes as probes of quantum spacetimes respecting causality. As such, it can be argued to be closer to physical particle propagation on quantum spacetimes. It is thus a new interesting tool to study the fundamental structure of spacetime in different quantum gravity approaches.
{\emph{Acknowledgments:}
We thank Fay Dowker, Joe Henson, Rafael Sorkin and Sumati Surya for helpful discussions. {We would like to express our thanks to Sumati Surya for a careful reading of this manuscript and for providing helpful comments.
S.~M. thanks Perimeter Institute for hospitality during this work and Girton College, Cambridge for support.
This research was supported in part by Perimeter Institute for Theoretical Physics. Research at Perimeter Institute is supported by the Government of Canada through Industry Canada and by the Province of Ontario through the Ministry of Research and Innovation.}
|
\section{$AC^0[p]$ Frege proof systems}
\label{1}
We will work with a sequent calculus style proof system in a language
with connectives $\neg$, unbounded arity $\bigvee$ and unbounded arity
connectives $MOD_{p,i}$ for $p$ a prime and $i = 0, \dots, p-1$.
The intended meaning of the formula
$MOD_{p,i}(y_1, \dots, y_k)$ is that
$\sum_{i} y_i \equiv i\ (\mbox{mod } p)$.
The proof system has the usual structural rules (weakening, contraction
and exchange), the cut rule, the left and the right $\neg$ introduction rules
and two introduction rules for $\bigvee$
modified for the unbounded arity; the
$\bigvee : left$ rule
$$
\frac{\varphi_1, \Gamma \rightarrow \Delta\ \ \ \
\varphi_2, \Gamma \rightarrow \Delta\ \ \ \dots\ \ \
\varphi_t, \Gamma \rightarrow \Delta}{
\bigvee_{i \le t}\varphi_i, \Gamma \rightarrow \Delta}
$$
and the
$\bigvee : right$ rule
$$
\frac{\Gamma \rightarrow \Delta, \varphi_j}{\Gamma \rightarrow \Delta, \bigvee_{i \le t}\varphi_i}
$$
any $j \le t$.
There are no rules concerning the $MOD_{p,i}$ connectives but there are
new $\mbox{{\bf MOD}}_p$-{\bf axioms} (we follow \cite[Sec.12.6]{kniha}):
\begin{itemize}
\item $MOD_{p,0}(\emptyset)$
\item $
\neg MOD_{p,i}(\emptyset)\ ,\ \mbox{for}\ i=1,\dots, p-1
$
\item $
MOD_{p,i}(\Gamma,\phi) \ \equiv\ [(MOD_{p,i}(\Gamma) \wedge
\neg \phi) \vee (MOD_{p,i-1}(\Gamma) \wedge \phi)]
$
for $i = 0, \dots, p-1$, where $i-1$ means $i-1$ modulo $p$,
and where $\Gamma$ stands for a sequence (possibly empty)
of formulas.
\end{itemize}
The depth of the formula is the maximal number of alternations of the
connectives; in particular, formulas from $\fphp$ have depth
$1$ and $2$ respectively.
We have not included among the connectives the conjunction $\bigwedge$;
this is in order to decrease the number of cases
one needs to consider in the constructions later on. Note that the need
to express $\bigwedge$ using $\neg$ and $\bigvee$ may increase the depth
of $AC^0$ formulas comparing to how it is usually counted. But as
we are aiming at lower bounds for all depths this is irrelevant.
We shall denote the proof system $LK(MOD_p)$ and its depth $d$ subsystem
(operating only with formulas of depth at most $d$) $LK_d(MOD_p)$.
It is well-known that this system is polynomially equivalent to constant depth
Frege systems with $MOD_{p,i}$ connectives
(or to Tait style system as in \cite{BKZ}) and in the mutual simulation
the depth increases only by a constant as the systems have the same language
(cf.\cite{kniha}).
The size of a formula or of a proof is the total number of symbols in it.
\section{From a proof to a game with formulas}
\label{2}
In this section and in the next one we define certain games
using the specific case of the PHP as an example. This is in order
no to burden the presentation right at the beginning
with a technical discussion of the
form of formulas we allow. As it is shown in Section \ref{6}
this is without a loss of generality and, in fact, motivates
the general formulation there.
Consider the following game $G(d,n,t)$ played between two players,
Prover and Liar. At every round Prover asks a question which Liar must
answer. Allowed questions are:
\begin{enumerate}
\item[(P1)] What is the truth-value of $\varphi$?
\item[(P2)] If Liar already gave a truth-value to $\varphi = \bigvee_{i \le u} \varphi_i$,
Prover can ask as follows:
\begin{enumerate}
\item If Liar answered {\sf false} then Prover can ask an extra question
about the truth value of any one of $\varphi_j$, $j \le u$.
\item If Liar answered {\sf true} then Prover can request that Liar witnesses
his answer by giving a $j \le u$ and stating that $\varphi_j$ is
{\sf true}.
\end{enumerate}
\end{enumerate}
All formulas asked by Prover
are built from the variables of $\fphp$,
and must have the depth at most $d$
and the size at most $2^t$.
The Liar's answers must obey the following rules:
\begin{enumerate}
\item[(L0)] When asked about a formula he already gave a truth value to in an earlier
round Liar must give the same answer.
\item[(L1)] He must give $\varphi$ and $\neg \varphi$
opposite truth values.
\item[(L2)] If asked according to (P2a) about $\varphi_j$ he must give value
{\sf false}.
If asked according to (P2b) he must give value
{\sf true} also to
some $\varphi_j$ with $j \le u$.
\item[(L3)] If asked about any $MOD_p$-axiom he must say {\sf true}.
\item[(L4)] If asked about any formula from $\fphp$ he must say {\sf true}.
\end{enumerate}
The game runs for $t$ rounds of questions and Liar wins if he can always
answer while obeying the rules. Otherwise Prover wins.
\begin{lemma} \label{2.1}
For any $d \geq 2$, $n \geq 1$ and $s \geq 1$. If there is a size $s$ $LK_d(MOD_p)$
refutation of $\fphp$ then Prover has a winning strategy
for game
$$G(d + O(1), n, O(\log s))\ .$$
\end{lemma}
\prf
It is well-known that LK-proofs (or Frege proofs) can be put into a form of balanced
tree with only a polynomial increase in size and a constant increase in the depth
(cf. \cite{Kra-lower,kniha}). In particular, the hypothesis of the lemma
implies that there is a size $s^{O(1)}$ refutation $\pi$ of $\fphp$
in $LK_{d+O(1)}(MOD_p)$ that is in a form of tree whose depth is
$O(\log s)$.
The Prover will attempt - by asking Liar suitable questions -
to built a path of sequents $Z_1, Z_2, \dots$ in $\pi$ such that
\begin{itemize}
\item $Z_1$ is the end-sequent of $\pi$, i.e. the empty sequent.
\item $Z_{i+1}$ is one of the hypothesis of the inference yielding $Z_i$.
\item If $Z_i$ is $\Gamma \rightarrow \Delta$ then Prover asked all formulas in
$\Gamma, \Delta$ and Liar asserted that all formulas in $\Gamma$ are true
and all formulas in $\Delta$ are false.
\end{itemize}
Assume $Z_1, \dots, Z_i$ has been constructed. Next Prover's move
depends on the type of inference yielding $Z_i$:
\begin{itemize}
\item Structural rules: Prover asks no questions and just takes for $Z_{i+1}$
the hypothesis of the inference.
\item Cut rule: Prover asks about the truth value of the cut formula, say $\varphi$,
and if Liar asserts it to be true, Prover takes for $Z_{i+1}$ the hypothesis of
the inference having $\varphi$ in the antecedent, otherwise it takes the hypothesis
with $\varphi$ in the succedent.
\item A $\neg$ introduction rule: if $\neg \varphi$ was the formula introduced,
Prover asks $\varphi$ and takes for $Z_{i+1}$ the unique hypothesis of the inference.
\item The $\bigvee : right$ introduction rule: if the principal formula was
$\varphi = \bigvee_{i \le u}\varphi_i$ and the minor formula $\varphi_j$
Prover already asked $\varphi$ in an earlier round and got answer {\sf false}.
He now asks $\varphi_j$ and
takes for $Z_{i+1}$ the unique hypothesis of the inference.
\item The $\bigvee : left$ introduction rule:
if the principal formula was
$\varphi = \bigvee_{i \le u}\varphi_i$ Prover already asked $\varphi$ in
an earlier round and got answer {\sf true}. She now asks Liar to witness
this answer by some $\varphi_j$ and then
takes for $Z_{i+1}$ the hypothesis with the minor formula $\varphi_j$ in the antecedent.
\end{itemize}
This process either causes Liar to lose or otherwise arrives at an initial
sequent which Liar's answers claim to be false. But that contradicts
rules (L1), (L3) or (L4).
\qed
Shallow tree-like refutations of a set of axioms can be used as search trees
finding an axiom false under a given assignment: the Liar answers the truth
values determined by the assignment (see e.g. the use of such trees
in \cite{Kra-lower,kniha}).
It was an important insight
of Buss and Pudl\' ak \cite{BusPud} that when Liars are allowed not to follow an
assignment but are only required to be logically consistent then the
minimal length of Prover's winning strategy
characterizes the minimal depth of a tree-like refutation
(a form of a statement opposite to the lemma also holds as pointed out
in \cite{BusPud} in the context of unrestricted Frege systems).
\section{Algebraic formulation of $\fphp$ and a game with polynomials}
\label{3}
Let ${\bf F}_p[x_{i j}\ |\ i \in [n+1] \wedge j \in [n]]$
be the ring of polynomials over the finite field
${\bf F}_p$ with $p$ elements with the indicated variables.
Denote by
$S_{n}$ the ring
factored by the ideal generated by all polynomials $x_{i j}^2 - x_{i j}$.
Elements of $S_n$ are multi-linear polynomials.
Let $S_{n,e}$ be the ${\bf F}_p$-vector space of elements of
$S_n$ of degree at most $e$.
We shall denote monomials $x_a, \dots$ where $a, \dots$ are unordered tuples
of variable indices; the monomial is then the product of the
corresponding variables.
Beame et al.\cite{BIKPP} formulated (the negation of) PHP
as the following {\bf $\fphp$-system}
of polynomial equations in $S_n$:
\begin{itemize}
\item $x_{i_1 j} \cdot x_{i_2 j} = 0$,
for each $i_1 \neq i_2 \in [n+1]$ and $j \in [n]$.
\item $x_{i j_1} \cdot x_{i j_2} = 0$,
for each $i \in [n+1]$ and $j_1 \neq j_2 \in [n]$.
\item $1 - \sum_{j \in [n]} x_{i j} = 0$, for each $i \in [n+1]$.
\end{itemize}
The left-hand sides of these equations will be denoted $Q_{i_1, i_2; j}$,
$Q_{i; j_1, j_2}$ and $Q_i$ respectively.
The language of rings is a complete language for propositional logic
and it is easy to imagine a modification of the G-game to such a language
if the answers of Liar have to respect both the sum and the product.
The game we are going to define allows only simple questions and
requires that sums of two polynomials and products of two
monomials are respected.
We shall define the following game $H(e,n,r)$ played by two players Alice and Bob.
Alice's role will be similar to that of Prover in the G-game and Bob's to that of
Liar. In every round Alice may put to Bob a question of just one type:
\begin{enumerate}
\item[(A)] She asks Bob to give
to a polynomial $f$ from $\sne$ a value from $\fp$.
\end{enumerate}
Bob's answers must obey the following rules:
\begin{enumerate}
\item[(B0)] If asked about a polynomial whose value he gave in an earlier
round Bob must answer identically as before.
\item[(B1)] He must give to each element $c \in \fp$ the value $c$,
and to each variable either $0$ or $1$.
\item[(B2)] If he gave values to $f$, $g$ and $f+g$, the values given to
$f$ and $g$ must sum up to the value he gave to $f + g$.
\item[(B3)] If he gave values to monomials
$x_a$, $x_b$ and $x_a \cdot x_b$, the product of the
values given to $x_a$ and $x_b$ must equal to the value given to
$x_a \cdot x_b$.
\item[(B4)] He must give value $0$ to all polynomials
$Q_{i_1, i_2; j}$,
$Q_{i; j_1, j_2}$ and $Q_i$.
\end{enumerate}
The game runs for $r$ rounds and Bob wins if he can answer all questions
while obeying the rules. Otherwise Alice wins.
We consider the multiplicativity condition for
monomials rather than for polynomials
as that more clearly isolates the role of linearity.
As is shown in Section \ref{4} the two versions
of the multiplicativity condition are essentially equivalent.
In principle Bob's strategy can be adaptive (i.e. his moves depend on
the development of the game) or even may depend on Alice.
Call a strategy of Bob {\bf simple}
if it is a function $B$ assigning to elements of
$\sne$ values in $\fp$ and Bob, when asked to evaluate $f$, answers $B(f)$.
We shall abuse the language occasionally and talk about a {\bf simple Bob} rather than a
simple strategy for Bob.
\section{Five useful protocols for Alice}
\label{4}
In this section we describe five simple
protocols in which Alice can force Bob to answer various
more complicated questions,
similar to that of (P2).
\bigskip
\noindent
{\bf Protocol $M_0$:} Assume that Bob asserted that $\sum_{i\le u}f_i \neq 0$.
Alice wants to force Bob to assert that $f_j \neq 0$ for some $j \le u$
(or to lose).
\medskip
She splits the sum into halves and asks Bob to evaluate
$\sum_{i\le u/2}f_i$ and $\sum_{i > u/2}f_i$.
As he already gave to $\sum_{i\le u}f_i$ a non-zero value,
by (B0) and (B2) - unless he quits -
Bob must give to at least one of the half-sums a non-zero value.
Continuing in a binary search fashion in $\log u$
rounds she forces
Bob to assert that $f_j \neq 0$ for some $j \le u$.
\bigskip
\noindent
{\bf Protocol $M_1$:} Assume that Bob gave to some polynomials
$f$, $g$ and $f \cdot g$ values $B(f)$, $B(g)$ and
$B(f \cdot g)$ respectively, and that
$B(f) \cdot B(g) \neq B(f \cdot g)$.
Alice wants to force Bob into a contradiction with the rules.
\medskip
Alice writes polynomials $f$ and $g$ as $\fp$-linear combinations
of monomials: $f = \sum_{a \in A} c_a x_a$ and
$g = \sum_{b \in B} d_b x_b$ with $c_a, d_b \in \fp$
and $x_a, x_b$ monomials. She
splits $A$ into two halves $A = A_0 \dot\cup A_1$, and asks Bob for the values
of
$$
(\sum_{a \in A_0} c_a x_a)\ ,\
(\sum_{a \in A_0} c_a x_a)\cdot g\ ,\
(\sum_{a \in A_1} c_a x_a)\ ,\ \mbox{ and}\
(\sum_{a \in A_1} c_a x_a)\cdot g
\ .
$$
Unless Bob violates the linearity rule (B2) his answers must satisfy
$$
B(\sum_{a \in A_i} c_a x_a)\cdot B(g)\ \neq\
B((\sum_{a \in A_i} c_a x_a)\cdot g)
$$
for either $i = 0$ or $i = 1$.
Continuing in the binary search fashion Alice forces Bob to
assert
$$
B(c_a x_a)\cdot B(g)\ \neq\
B( c_a x_a\cdot g)
$$
for some monomial $x_a$. Using (B1) and (B2) she forces
$$
B(c_a x_a) = c_a B(x_a)\ \mbox{ and }\
B(c_a x_a g) = c_a B(x_a g)
$$
and hence
$$
B(x_a) \cdot B(g) \ \neq\ B(x_a \cdot g)\ .
$$
The number of variables is $n^{O(1)}$ and so the number
of monomials of degree at most $e$ is $n^{O(e)}$.
Hence all this process requires as most $O(e \log n)$
rounds of Alice's questions.
Now she analogously forces Bob to assert
$$
B(x_a) \cdot B(x_b)\ \neq \ B(x_a \cdot x_b)
$$
for some monomial $x_b$ occurring in $g$, violating thus (B3).
\bigskip
\noindent
{\bf Protocol $M_2$:} Assume that Bob asserted that $\Pi_{i\le k}f_i \neq 0$
and let $j \le k$ be arbitrary.
Alice wants to force Bob to assert that $f_j \neq 0$ (or to lose).
\medskip
She asks Bob to state the value of $f_j$ and if
Bob says $f_j \neq 0$ she stops. Otherwise
the triple $f_j, g$ and $f_j g$ for
$g := \Pi_{i \le k, i\neq j}f_i$ satisfies the hypothesis
of protocol $M_1$ and Alice can win in $O(e \log n)$ rounds.
\bigskip
\noindent
{\bf Protocol $M_3$:} Assume that Bob asserted that $\Pi_{i\le k}f_i = 0$.
Alice wants to force Bob to assert that $f_j = 0$ for some $j \le k$ (or to
lose).
\medskip
We shall describe the protocol by induction on $k$. Alice asks first
for the value of $f_k$. If Bob states that $f_k = 0$ she stops. If he states that
$f_k \neq 0$ she
asks him for the value of $\Pi_{i < k}f_i$.
If Bob says that $\Pi_{i < k}f_i = 0$, Alice
has - by the induction hypothesis - a way how to solve the task.
If he says that $\Pi_{i < k}f_i \neq 0$
Alice forces him into contradiction
using protocol $M_1$. We may assume that all polynomials $f_i$
are non-constant and thus the induction process takes at most $k \le e$
steps.
Note that again Alice needed at most
$2e + O(e \log n) = O(e t)$ rounds in total.
\bigskip
\noindent
{\bf Protocol $M_4$:} Let $g = f^{p-1}$ and assume that
Bob gave to $g$ a value different from both $0, 1$.
Alice wants to force Bob into a contradiction.
\medskip
She asks Bob for the value of $f$ and assume Bob states $f = c \in \fp$.
If $c=0$ Alice uses protocol $M_2$ to force a contradiction.
If $c \neq 0$ Alice asks Bob for values of $f^2, f^3, \dots,
f^{p-1}$ and unless Bob returns values
$c^2, c^3, \dots, c^{p-1}$ she forces him into a contradiction
by protocol $M_1$. But Bob cannot keep up these answers because
if he gave to $g$ now the value
$c^{p-1} = 1$ he would violate rule (B0).
\section{From Prover to Alice and from Bob to Liar}
\label{5}
In this section we employ the Razborov - Smolensky method
to show that the existence of many simple winning strategies
for Bob yields a winning strategy for
Liar\footnote{We could have bypassed the G-game and the explicit
use of the Razborov - Smolensky method by employing the characterization
of the size of $AC^0[p]$ Frege proofs in terms of degree of proofs in the
so called Extended Nullstellensatz of \cite{BIKPRS}. We prefer here a self-contained
presentation.}.
The reason to single out simple strategies is that we shall
apply the Razborov - Smolensky approximation method
in order to move from a G-game to an H-game,
by approximating formulas by low degree polynomials with respect to
(a set of) Bob's strategies.
The approximation process (and hence a strategy for Alice
to be constructed) depends on the set of Bob's strategies
we start with
and to avoid circularity we restrict to sets containing only (but not
necessarily all) simple strategies.
\begin{lemma} \label{5.1}
Let $d \geq 2$, $n \geq 1$ and $t \geq \log n$ be arbitrary
and take
parameters $e, r$
$$
e\ :=\ ((t^2 + 2t) p)^{d} \ \mbox{ and }\
r\ :=\ O(e t^{4})\ .
$$
Let $P$ be any strategy for Prover in game $G(d, n, t)$.
Let $\on$ be a non-empty set of simple strategies for Bob in game $H(e, n, r)$.
Then $P$ can be translated into a strategy $A$ for Alice in $H(e, n, r)$ such that
the following holds:
\begin{itemize}
\item
If
\begin{equation}\label{e}
\prob_{B \in \on}[B\ \mbox{ wins over $A$ in } H(e, n, r)]\ >\
1\ -\ 2^{-(t +1)}
\end{equation}
then there exists a Liar's strategy $L$ winning over $P$ in $G(d, n, t)$.
\end{itemize}
\end{lemma}
\prf
Let $P$ and $\on$ be given. Let $F$ be the smallest set of formulas
closed under subformulas and containing all possible $P$'s questions
according to rule (P1) in all
plays of the game $G(d, n, t)$ against all possible Liars.
The number of such (P1) questions is at most $2^{t^2}$
and each has size at most $2^t$ and so also at most $2^t$
subformulas.
Thus
the depth of all formulas in $F$ is at most $d$ and their total number is
bounded by $2^{t^2 + t}$.
We shall use the Razborov - Smolensky method to assign to all formulas
$\varphi \in F$ a polynomial $\hat \varphi \in \sne$. However, we shall
approximate with respect to Bob's
strategies from $\on$ rather than with respect to all
assignments to variables as it is usual.
Fix parameter $\ell := t^2 + 2t$. Put
$\hat x_{e} := x_{e}$, $\hat{(\neg \varphi)} := 1 - \hat \varphi$ and for
$\varphi = MOD_{p,i}(\varphi_1, \dots, \varphi_k)$ define
$$
\hat{\varphi} \ :=\ 1\ -\
((\sum_{j \le i} \hat \varphi_j) -i)^{p-1}\ .
$$
For the remaining case $\varphi = \bigvee_{i \in [u]}\varphi_i$
assume that all polynomials $\hat \varphi_i$ were already defined.
Pick $\ell$ subsets
$J_1, \dots, J_{\ell} \subseteq [u]$,
independently and uniformly at random (we shall fix them in a moment),
and define polynomial
$$
p_{\varphi}(y_1, \dots, y_u)\ :=\
1\ -\ \Pi_{j \le \ell} (1\ -\ (\sum_{i \in J_j} y_i)^{p-1})\
$$
and using $p_{\varphi}$ put
\begin{equation} \label{approx}
\hat \varphi\ :=\ p_{\varphi}(\hat \varphi_1, \dots, \hat \varphi_u)\ .
\end{equation}
The following claim is easily verified by induction on the depth of $\varphi$, using
the protocols from Section \ref{4}.
\medskip
\noindent
{\bf Claim 1:} {\em
Let $\varphi \in F$ and assume that Bob asserted that $\hat \varphi = c \in \fp$ for some
$c \neq 0, 1$. Then Alice can force Bob into a contradiction in
$O(e\log n)$ rounds.
}
\medskip
Let $b_i \in \{0,1\}$ be the truth-value of the statement $B(\hat \varphi_i)\neq 0$.
For $B \in \on$ we have that
\begin{equation}\label{rs}
\bigvee_{i \in [u]} b_i\ =\
p_{\varphi}(b_1, \dots, b_u)
\end{equation}
with the probability at least $1 - 2^{- \ell}$ (taken over the choices of sets $J$).
Hence we can select specific
sets $J_1, \dots, J_{\ell}$ such that (\ref{rs}) holds for all but
$2^{-\ell}\cdot |\on|$ simple Bobs from $\on$.
The polynomial $\hat \varphi$ in (\ref{approx}) is assumed to
have this property.
\medskip
Define in this way the polynomial $\hat \varphi$ for
all (at most $2^{t^2 + t}$)
formulas $\varphi \in F$ by induction on the depth $1, 2, \dots, d$.
Each is of degree at most $(\ell (p-1))^d \le ((t^2 + 2t)p)^d = e$ and
it holds that:
\medskip
\noindent
{\bf Claim 2:} {\em There is a subset $Err \subseteq \on$ such that
$|Err| \le 2^{- t}|\on|$ and such that (\ref{rs}) holds for all
disjunctions $\varphi \in F$ and all $B \in \on \setminus Err$.}
\medskip
Now we define, using the given
strategy $P$ for Prover, a specific strategy $A$ for
Alice in $H(e, n, r)$. We transcript $P$ into $A$ a question
by question; each question of $P$ may be replaced by a series of questions of
Alice.
If P asks according to (P1) what is the value of $\varphi$,
Alice simply asks for the value of $\hat \varphi$.
Let $\varphi = \bigvee_{i \in [u]}\varphi_i$ and assume that
P asks according to (P2); there are two cases to consider:
\begin{enumerate}
\item[(a)] $\varphi$ got value {\sf false} and P asks for the value of
one disjunct $\varphi_j$,
\item[(b)] $\varphi$ got value {\sf true} and P asks for a witness $\varphi_j$.
\end{enumerate}
Assume for the case (a) that Bob asserted in an earlier round that
$\hat\varphi = 0$. Alice asks Bob for the value of $\hat \varphi_j$. If he gives
$B(\hat \varphi_j) = 0$ the simulation of P moves to the next round. If he
replies that $B(\hat \varphi_j) = 1$,
Alice uses first protocol
$M_2$ repeatedly to force Bob to assert
$$
1\ - \ (\sum_{i \in J_v} \hat\varphi_i)^{p-1}\ \neq \ 0
$$
for all $v \le \ell$.
Then for each $v$ she uses protocol $M_4$ to force Bob to say that
$$
(\sum_{i \in J_v} \hat\varphi_i)^{p-1}\ = \ 0
$$
and further protocol $M_3$ to assert that
\begin{equation}\label{a}
\sum_{i \in J_v} \hat\varphi_i\ = \ 0\ .
\end{equation}
This needs $O(\ell e \log n) = O(t^2 e \log n) = O(e t^3)$
rounds.
As $B(\hat \varphi_j) = 1$, if
Bob uses a strategy $B \in \on \setminus Err$,
the definition of $Err$ guarantees that one of the equations
in (\ref{a}) is false when $\hat \varphi_i$'s are
evaluated by $B$:
$$
B(\sum_{i \in J_v} \hat\varphi_i)\ \neq\
\sum_{i \in J_v} B(\hat\varphi_i)\ .
$$
This itself is not a violation of rule (B2) but Alice can use this situation
and to force Bob to lose. We shall describe her strategy
as probabilistic; a deterministic one is obtained by an averaging argument.
Alice splits $J_v = K_0 \dot \cup K_1$ into halves
and asks Bob for the values of $\sum_{i \in K_0} \hat\varphi_i$
and $\sum_{i \in K_1} \hat\varphi_i$. Unless he violates (B2)
his answers must sum up to $B(\sum_{i \in J_v} \hat\varphi_i)$.
Hence for $k=0$ or $k=1$
$$
B(\sum_{i \in K_k} \hat\varphi_i)\ \neq\
\sum_{i \in K_k} B(\hat\varphi_i)\ .
$$
Alice guesses for which $k$ this happens and then proceeds analogously
with $\sum_{i \in K_k} \hat\varphi_i$, splitting it into halves,
asking Bob for the values, etc. If she always guesses right
then in $t$ steps (as the size of the sums is bounded by $2^t$)
she will reduce the sums to one term and will win.
Alice has the probability at least $2^{-t}$ to make the right
choices. She does not know a priori which of the $\ell$
sums $\sum_{i \in J_v} \hat\varphi_i$ to use so she must try all.
This takes $O(\ell t) = O(t^3)$ rounds.
There are at most $t$ simulations of a (P2a) question in the G-game
but Alice needs to employ the random strategy above only
once when the case $B(\hat \varphi_j) = 1$ occurs, and then her probability
of success is at least $2^{-t}$.
By averaging there are fixed choices that Alice can make, yielding this
success probability outside of $Err$. In particular, for a random
$B \in \on \setminus Err$, if Alice uses these choices then
either B must give to $\hat \varphi_j$ value $0$ or Alice wins
with the probability at least $2^{-t}$.
We shall describe this situation below by the phrase
that the {\em (P2a) simulation succeeded}.
\medskip
Assume for the case (b) that Bob answered earlier
that $\hat \varphi = 1$
and hence also that
$$
\Pi_{j \le \ell}(1\ -\ (\sum_{i \in J_j}\hat\varphi_i)^{p-1})\ =\ 0.
$$
Alice uses protocols $M_3$ and $M_4$
to force Bob to state that
$\sum_{i \in J_v}\hat \varphi_i = 1$ for some $v \le \ell$.
This uses $O(e \log n) = O(e t)$ rounds.
Then she uses protocol $M_0$
to force Bob to say that
$\hat \varphi_j \neq 0$ for some $j \in J_v$
and by Claim 1 the value has to be $1$ ($O(e \log n) =
O(e t)$ rounds are used in Claim 1).
The number of formulas $\varphi_i$ is bounded by the size of
$\varphi$, i.e. by $2^t$, and so this uses at most $t$ rounds in protocol $M_0$,
i.e. still $O(e t)$ in total.
This describes the strategy $A$.
\bigskip
By Claim 2 with the probability at least $1 - 2^{-t}$
a random $B \in \on$ is outside $Err$, and for these Alice's
simulations of (P2a) questions succeed with the probability at least
$2^{-t}$. Thus the
inequality (\ref{e}) from the
hypothesis of the lemma implies that there is at least one
$B \in \on \setminus Err$ winning over the particular Alice's strategy
A and for which A's simulations of (P2a) questions succeed.
Use B to define a strategy L for Liar in the original game $G(d, n, t)$
simply by giving to $\varphi$ the truth value
$B(\hat\varphi)$ when asked a (P1) type question, and
giving a witness $\varphi_j$ constructed in the case (b) above
when asked a (P2b) type question.
From the construction of A (and rules for Bob) it follows that L
satisfies the rules for Liar. In particular, by (B4) all polynomials
from the $\fphp$-system
get $0$ by B and so all axioms of $\fphp$
get by L value
{\sf true}.
Note that one question of P is transcribed into at most
$O(e t^3)$ Alice's questions.
Hence in every play of the H-game transcribing a play of the $G$-game
there are in total at most
$r = O(e t^4)$ rounds.
\qed
\section{A general reduction to a search problem}
\label{6}
The reduction of the lengths-of-proofs problem to a question
about the $H$-games in Sections \ref{2} - \ref{5} is not specific to
$\fphp$ and works in a fairly general situation that we shall describe now.
Then we reduce the proof complexity problem further to a question
about the computational complexity of a certain task involving computations
with search trees.
The only specific thing in the $\fphp$ case is
the transcription of the axioms of $\fphp$ into the $\fphp$
polynomial system in Section \ref{3}. This is not a mere mechanical translation
from DeMorgan language into the language of rings (in that the
axioms $\bigvee_{j \in [n]} p_{i j}$
would translate into polynomials of degree about $n$
and not into degree $1$ polynomials $Q_i$).
In order to avoid inevitable technicalities when trying to define
suitable translations from a general set of axioms to
a polynomial system we simply take
as our starting point an unsolvable system of polynomial equations
of a constant degree. The
truth value of an equation $f(x_1, \dots, x_m)=0$ for Boolean variables $x_i$,
$f$ a degree $O(1)$ polynomial over $\fp$, can be defined by a
depth $2$, size $m^{O(1)}$
$AC^0[p]$ formula. Namely, writing $f$ as an $\fp$-linear combination
$\sum_{a \in A} c_a x_a$ of monomials $x_a$, with $c_a \in \{1, \dots, p-1\}$,
consider the formula
\begin{equation}
\varphi \ :=\
MOD_{p,0}(\psi_1, \dots, \psi_k)
\end{equation}
where $k = \sum_{a \in A} c_a$ and $\psi_i$'s are
conjunctions of variables corresponding
to monomials from $f$, each monomial $x_a$ being represented
$c_a$-times. Clearly\footnote{Instead of assuming degree $O(1)$ it would suffice
to assume that $f$ is an $\fp$-linear combination of polynomially
many monomials.} $\varphi$ represents the truth value of $f=0$ on
Boolean variables.
The polynomial system can thus be also thought of
as an unsatisfiable set of $AC^0[p]$ formulas and we can speak about
its $LK_d(MOD_p)$ refutations.
\bigskip
We shall now consider the following general set-up. For $n = 1, 2, \dots$
let ${\cal F}_n$ be a sequence of sets of polynomials over $\fp$
in variables $Var({\cal F}_n)$. We shall
assume that:
\begin{enumerate}
\item
polynomials in sets ${\cal F}_n$ have $O(1)$ degree,
\item the size of both ${\cal F}_n$ and
$Var({\cal F}_n)$ is $n^{O(1)}$,
\item the polynomial system
$$
f\ =\ 0\ ,\ \mbox{ for } \ f \in {\cal F}_n
$$
contains equations $x^2 - x = 0$ for all
$x \in Var({\cal F}_n)$
and is unsolvable in $\fp$.
\end{enumerate}
Let $\sfn$ be the $\fp$ - vector space of multi-linear polynomials
in variables of ${\cal F}_n$ and of degree at most $e$.
We want to replace games and strategies considered in previous sections by
a more direct computational model, namely that of search trees.
Define an {\bf $\sfn$ - search tree} $T$ to be a $p$-ary tree whose
inner nodes (non-leaves) are labelled by polynomials from
$\sfn$, the $p$ edges leaving a node labelled by $g$
are labelled by $g=0, g=1, \dots, g = p-1$, and leaves
are labelled by elements of a set $X$.
Any function $B : \sfn \rightarrow \fp$ determines a path
$P_T(B)$ in $T$
consisting of edges labelled by $g = B(g)$
and thus it also determines an element of $X$: the label of the
unique leaf on $P_T(B)$. Hence $T$ defines a function assigning
to any map $B : \sfn \rightarrow \fp$ an element of $X$ to be denoted $T(B)$.
Let $\eer$ be the set of pairs and triples of the form
$(B1, c)$ for $c \in \fp$ or $(B1, x)$ for $x \in Var({\cal F}_n)$,
$(B2, f, g)$, $(B3, x_a, x_b)$ or $(B4, f)$ for $f \in {\cal F}_n$,
with $f,g, x_a, x_b$ of degree at most $e$. These are intended to
indicate what instance of which
rule did Bob violate. We say that
$(B1, c)$ is an {\bf error for} $B$ iff $B(c) \neq c$, $(B1, x)$
is an error for $B$ iff $B(x) \neq 0, 1$, and similarly for
the other pairs and triples\footnote{We ignore errors for (B0) as that rule cannot
be violated by a simple Bob
and hence search trees do not need to ask anything twice on any path.}.
In the following statement we talk about refutations of equations $f=0$,
$f \in {\cal F}_n$. As pointed out earlier, we can view them also as depth $2$,
polynomial size formulas with $MOD_{p,0}$ connectives and hence it makes
a prefect sense to talk about their $LK_d(MOD_p)$-refutations.
The reductions of Sections \ref{2} - \ref{5}
used the example of $\fphp$ (see the beginning of Section \ref{2}) but nothing
specific to it was used. Hence we can employ
the reductions to derive the following general statement.
In it we replace degree $e$ by (bigger) $r$
in order to avoid the need to define here
the relation between them implicit
in Lemma \ref{5.1}.
\begin{theorem} \label{6.2}
Let $r = r(n) \geq (\log n)^{\omega(1)}$ be a function and
let ${\cal F}_n$ be sets of polynomials obeying the restrictions
1., 2. and 3. listed above.
Then for every $d \geq 2$ there are $\epsilon_d > 0$
and $n_d \geq 1$ such that for an arbitrary non-empty
set $\om$ of maps from $\sfnr$ to $\fp$
the following
implication (I) holds for all $n \geq n_d$ and all
$0 < \epsilon \le \epsilon_d$:
\begin{itemize}
\item[(I)]
If for every $\sfnr$ - search tree $T$
of depth $r$ and with
leaves labelled by elements of $\eerr$
it holds that
\begin{equation} \label{key}
\prob_{B \in \om}[\mbox{ $T(B)$ is not an error for $B$ }]
\ > \ 1 -
2^{- r^{\epsilon}}
\end{equation}
then
$LK_d(MOD_p)$ does not refute the set of
formulas $f=0$, $f \in {\cal F}_n$, by a proof of size
less than $2^{\Omega(r^{\epsilon})}$.
\end{itemize}
\end{theorem}
\prf
Assume that $LK_d(MOD_p)$ does refute the set of
formulas $f=0$, $f \in {\cal F}_n$, by a proof of size
$s = s(n)$. By Lemma \ref{2.1} Prover has a winning strategy P
for game $G(d+c, n, t)$, where $t = t(n) = O(\log s)$ and $c$ is
an absolute constant.
Put $\epsilon_d := \frac{1}{2(d+c)+5}$ and let $0 < \epsilon \le
\epsilon_d$. If it were that $t+1 \le r^\epsilon$ then
the parameters $e', r'$ of the game $H(e',n,r')$ constructed
in Lemma \ref{5.1}
satisfy
$e' \le r' < r$ and, in particular, the game is an $H(r,n,r)$ game.
The strategy A defined in Lemma \ref{5.1} for the game
defines an $\sfnr$ - search tree $T$
of depth $r$ and with
leaves labelled by elements of $\eerr$ in a natural way:
a path in $T$ corresponds to possible answers of a simple Bob
and the path stops as soon as a violation of one of the rules (B1)-(B4)
occurs (rule (B0) cannot be broken by a simple Bob).
The label of the resulting leaf is the instance of the rule
that was broken
(if a violation did not occur we use any element of $\eerr$).
Assume that $\om$ is a set of simple Bobs
for which the inequality (\ref{key}) holds. Then also
the inequality (\ref{e}) from Lemma \ref{5.1} holds and thus
by that lemma there is a strategy L for Liar that wins over P in
the original $G$-game. That is a contradiction and thus
$s \geq 2^{\Omega(r^{\epsilon})}$.
\qed
\bigskip
To conclude the paper let us discuss informally the construction
underlying Lemma \ref{5.1} and Theorem \ref{6.2}. In particular, we
see these formal statements as templates for a possible variety
of analogous reductions, and it is not clear which one - if any - will
be eventually useful.
The strategy $A$ is constructed in Lemma \ref{5.1} by a randomized
process from strategy $P$ and from set $\on$. Let us call the class
of all strategies
$A$ that can occur in this way the class of {\em $(P, \on)$-generated
strategies}. One such class contains only a few of
all possible Alice's strategies.
Moreover, we can pick $\on$ depending on $P$. Hence one can weaken the
hypothesis in these statements and, for example, Theorem \ref{6.2}
could be reformulated as follows:
\begin{itemize}
\item
Let $r = r(n) \geq (\log n)^{\omega(1)}$ be a function and
let ${\cal F}_n$ be sets of polynomials obeying the restrictions
1., 2. and 3. listed above.
Then for every $d \geq 2$ there are $\epsilon_d > 0$
and $n_d \geq 1$ such that the following holds:
If for every Prover's strategy $P$ for game $G(d,n,r^{\Omega(1)})$
there exists a non-empty
set $\om(P)$ of maps from $\sfnr$ to $\fp$
then the following
implication (I') holds for all $n \geq n_d$ and all
$0 < \epsilon \le \epsilon_d$:
\begin{itemize}
\item[(I')]
If for every $\sfnr$ - search tree $T$
of depth $r$ and with
leaves labelled by elements of $\eerr$
originating from a $(P, \om(P))$-generated
$A$
it holds that
\begin{equation}
\prob_{B \in \om(P)}[\mbox{ $T(B)$ is not an error for $B$ }]
\ > \ 1 -
2^{- r^{\epsilon}}
\end{equation}
then
$LK_d(MOD_p)$ does not refute the set of
formulas $f=0$, $f \in {\cal F}_n$, by a proof of size
less than $2^{\Omega(r^{\epsilon})}$.
\end{itemize}
\end{itemize}
\medskip
\noindent
This formulation stains the combinatorially clean original
formulation by a reference to $P$ but (I') may be a weaker hypothesis
to arrange.
Another issue is the discouragingly high probability required in
(\ref{e}) and (\ref{key}). This is due solely by Alice's simulation
of the (P2a) move of $P$. At that point she found $\ell \le
O(t^2) \le e \le r$ sets $K$, $|K| \le 2^t$, of degree $e$ polynomials
such that for one of them $B$ fails linearity:
\begin{equation}\label{concl}
B(\sum_{i \in K} g_i)\ \neq\
\sum_{i \in K} B(g_i)\
\end{equation}
and her strategy worked up to this point
for all $B \notin Err$ (as long as $P$
was a winning strategy for the Prover). Getting from this situation to
a violation of rule (B2) costs her the drop of the success probability by
the multiplicative factor $2^{-t}$. Hence we could redefine the rules
for the H-game and, in particular,
the error sets $\eer$ for the search problems
to be solved by the trees, and include that situation
(i.e. $A$ producing $\ell$ sets $K$ such that one of them satisfies
(\ref{concl})) among the stopping Bob's errors. Let us call
$*\eer$ the set of errors with this new type of an error added.
Then we could reformulate Theorem \ref{6.2}
differently as follows:
\begin{itemize}
\item
Let $r = r(n) \geq (\log n)^{\omega(1)}$ be a function and
let ${\cal F}_n$ be sets of polynomials obeying the restrictions
1., 2. and 3. listed above.
Then for every $d \geq 2$ there are $\epsilon_d > 0$
and $n_d \geq 1$ such that for an arbitrary non-empty
set $\om$ of maps from $\sfnr$ to $\fp$
the following
implication (I'') holds for all $n \geq n_d$ and all
$0 < \epsilon \le \epsilon_d$:
\begin{itemize}
\item[(I'')]
If for every $\sfnr$ - search tree $T$
of depth $r$ and with
leaves labelled by elements of $*\eerr$
it holds that
\begin{equation}
\prob_{B \in \om}[\mbox{ $T(B)$ is not an error for $B$ }]
\ > \
2^{- r^{\epsilon}}
\end{equation}
then
$LK_d(MOD_p)$ does not refute the set of
formulas $f=0$, $f \in {\cal F}_n$, by a proof of size
less than $2^{\Omega(r^{\epsilon})}$.
\end{itemize}
\end{itemize}
\medskip
\noindent
Let us stress that the culprit property is the linearity
by observing that simple Bobs can be without a loss of generality assumed
to satisfy all rules except possibly (B2). First, having $B$ we can
define $B'$ by correcting all values of $B$ that violate rules (B1) or
(B4). If $B'$ is asked by Alice for one of these new values, the original $B$
would lose. Hence $B'$ is as good as $B$ against any $A$.
Then define $B''$ by giving to every monomial $x_a = \Pi_i x_i$
the value $\Pi_i B'(x_i)$. Enhance any $A$ to vigilant $A^*$
that whenever she asks for the value of a monomial, she asks also for the
values of all its variables (this enlarges the number of round $e$-times at
most). Clearly, $B''$ fares as well as $B'$ against a vigilant $A^*$.
Finally, let us remark that
it would be interesting and possibly quite useful to
modify the construction so that
adaptive Bobs are allowed.
\bigskip
\noindent
{\bf Acknowledgements.}
I thank L.~Kolodziejczyk (Warsaw),
S.~M\"{u}ller (Tokyo), P.~Pudl\' ak (Prague) and N.~Thapen (Prague)
for critical comments and discussions.
I am also indebted to the anonymous referee for helpful suggestions.
|
\chapter{Chaotic Interference Versus Decoherence: External Noise, State Mixing and Quantum-Classical Correspondence}
\pacs{05.45.Mt, 03.65.Nk, 73.23.-b, 03.65.Nk, 24.30.-v, 03.65.Nk}
\maketitle
\section{Outline}
\label{sec:outline}
\noindent
\looseness-1
The famous Nils Bohr's quantum-classical correspondence
principle states that the classical mechanics is a limiting case of the
more general quantum mechanics. This implies that ``under certain
conditions" quantum laws of motion become equivalent to classical
ones. One of the conditions is fairly obvious: the corresponding
classical action should be very large as compared with the Planck's
constant $\hbar$. But is this the \textit{sufficient} condition? In fact,
\textit{it is not!}
The quantum laws show up in two different, although not entirely independent ways:
\begin{enumerate}
\item discrete spectrum of finite motion,
\item interference phenomena.
\end{enumerate}
Even if the energy spectrum of a finite closed quantum system
becomes continuous in the formal limit of vanishing $\hbar$, the
interference effects cannot disappear in similar manner. Indeed, a
quantum wave functions has no definite classical counterpart.
Meanwhile, suppression of effects of quantum interference
("decoherence") is a key requirement for the classical laws to appear.
Being, in essence, of quite general nature, this problem takes on
special significance in the non-trivial case of non-linear classically
chaotic quantum systems.
A number of typical manifestations of the quantum coherence in the
time evolution as well as eigenstates' properties are widely discussed
in the scientific literature:
\begin{itemize}
\item Wave packet dynamics and decay of quantum fidelity.
\item Universal local spectral fluctuations.
\item Scars in the stationary eigenfunctions.
\item Elastic enhancement in chaotic resonance scattering.
\item Weak localization in transport phenomena.
\end{itemize}
The specific features of quantum dynamics of classically chaotic
systems seem to be in striking contrast with those of
genuine classical chaos. Since these features can even question the
validity of the quantum-classical correspondence principle by itself,
a more profound analysis is needed for understanding
the bridge between the classical and quantum chaotic worlds.
\subsection{Chaotic time evolution}
\label{sec:ChaoTimeEvolution}
In the case of regular classical dynamics, the system's response to a
weak external perturbation is proportional to its strength and the
system may be still treated as a \textit{closed} one during
sufficiently long time. In contrast, chaotic classical dynamics is
exponentially unstable and therefore it is extremely sensitive to any
uncontrollable external influence. We can never neglect
the influence of the environment. This therefore stipulates the
\textit{self-mixing} property of classical dynamical chaos and, as a
consequence, a very fast decay of the phase correlations (here and
throughout the paper we use the language of action-angle variables).
Whereas the exponential decay of the phase correlations is an
underlying feature of the classical dynamical chaos \cite{Chiri79},
the so called "quantum chaos", i.e. quantum dynamics
of classically chaotic systems is not by itself capable of destroying
the \textit{quantum phase} coherence. Strictly speaking, any
initially pure quantum state remains pure during arbitrary long
evolution. Quantization of the phase space removes exponential
instability and makes the quantum dynamics to be substantially more
stable than the classical one. There exists a threshold $\sigma_c(t)$
of the external noise strength $\sigma$ below which the coherence
survives up to the time $t$ \cite{Sokol08,Sokol08a,Sokol09}
Only appreciably strong noise or finite measurement accuracy can
produce a mixture of quantum states sufficient for noticeable
suppression of quantum coherence. A number of appropriate
characteristics: Peres fidelity $F(t)$, purity $\mathbb{P}(t)$,
Shannon (information) ${\cal I}(t)$ and von Neumann (correlational)
${\cal S}(t)$ entropies are used to demonstrate the gradual loss of
quantum coherence during system's evolution in the presence of
noisy background. Being sensitive to the quantum coherence, the
von Neumann entropy remains smaller than the Shannon entropy but
runs up monotonically to the latter when the evolution time
approaches some moment $t_{(dec)}$ when decoherence becomes
complete. After this time, the system occupies the whole phase space
volume accessible at the running degree of excitation energy thus
reaching a sort of equilibrium. Henceforth the phase volume expands
``adiabatically'': the both entropies remain almost constant. The
evolution after the time
$t_{(dec)}$ is Markovian \cite{Sokol09}.
\subsection{2D billiards}
\label{sec:2Dbilliards}
The plain ($2D$) areas with closed irregular borders called billiards
are pet systems often used to illustrate characteristic features of the
classical dynamical chaos. With the advent of the ability to fabricate
mesoscopic analogs of the classical billiards the opportunity appeared
to observe experimentally the signatures of chaos in classically
chaotic quantum systems. An excellent possibility thus has arisen to
verify the theoretical concepts developed in numerous theoretical
investigations of "quantum chaos" phenomena. Extensive study of the
electron transport through ballistic $2D$ meso-structures
\cite{Jalab90,Marcu92} (see also \cite{Beena97,Alhas00} and
references therein) have fully confirmed correctness of the basic
ideas \cite{Izrai90,Gutzw90,Haake91} of the theory of the chaotic
quantum interference and relevance \cite{Baran94,Beena97,Alhas00}
of the so called random-matrix approach
\cite{Verba85,Mello85,Mehta91} to the problem of the universal
spectral fluctuations as well as conductance fluctuations in open
mesoscopic set-ups (see sec.\ref{sec:Billiards&Transport}).
The energy $E$ is the only integral of motion in a closed billiard.
Repeating reflections from the border produces a countless manifold
of unclosed exponentially unstable trajectories and, as a consequence,
chaotic dynamics. At the same time, there exists a countable set of
specific, but also exponentially unstable, periodic orbits. The number
$N(T)$ of the latter trajectories grows exponentially with the period
$T$ as $N(T)\sim e^{\frac{T}{\tau_c}}$ with the characteristic time
$\tau_c$ directly connected, $\tau_c\thicksim 1/\Lambda$, to the
Lyapunov exponent $\Lambda$ that describes the exponential
instability \cite{Zasla85}. Meanwhile, in contravention with the
classical exponential instability, stable interference maxima (scars)
along the periodic classical orbits (alongside with irregularly
scattered point-like interference maxima (specks)) are discovered
numerically \cite{Helle84} in the stationary billiard's eigenstates
even in the very deep semi-classical region . Does this fact
compromises the quantum-classical correspondence principle?\\
To answer afore-posed question let us consider first an elementary
example of $1D$ finite motion. A particle with a mass $m$ and
energy $E$ moves in a potential well. The classical position
probability density is proportional to the fraction of the period of
oscillations the particle spends near a point $x$ and is easily found to
be:
\begin{equation}
w_c(x)=
\left\{
\begin{array}{ll}
0 &x\notin [a, b],\\
\frac{m\,\omega_c(E)}{\pi\, p_c(x)},\,\,\, &x\in[a, b]\,.\\
\end{array}
\right.
\label{eq:classW}
\end{equation}
where $p_c(x)=\sqrt{2m\left[E-U(x)\right]}$ is the particle's classical
momentum at the point $x$ and $\omega_c(E)$ is the frequency of
classical oscillations between the turning points $a$ and $b$.
On the other hand the semi-classical solution of the corresponding
quantum problem yields
\begin{eqnarray}
w_n(x)&=&|\psi_n(x)|^2 \nonumber \\
&=&\left\{
\begin{array}{ll}
e^{\{-\frac{2}{\hbar}\int_x^a dx' |p_c(x')|\}},
e^{\{-\frac{2}{\hbar}\int_b^x dx' |p_c(x')|\}} &\rightarrow 0,\\
\frac{m\,\omega_c(E_n)}{\pi\, p_c(x)}\,2\cos^2\{\frac{1}{\hbar}
\int_a^x dx'p_c(x')-\frac{\pi}{4}\} &\rightarrow ?
\end{array}
\right.
\label{eq:quantW}
\end{eqnarray}
\begin{figure}[h]
\begin{center}
\includegraphics[width=70mm,angle=0,
keepaspectratio=true]{1DWell.eps}
\caption{\textit{Top}: wave function of harmonic oscillator for the
energy level $n=25$; \textit{bottom}: corresponding probability
distributions are shown by thick red and thin black lines for classical and quantum cases respectively.}
\label{fig:1DWell}
\end{center}
\end{figure}
As expected (see the simplest of linear oscillator,
Fig.\ref{fig:1DWell}, bottom frame),
the probability density (\ref{eq:quantW}) vanishes in
the classically forbidden region when the semi-classical parameter
goes to infinity but there exists no reasonable result in the
\textit{classically allowed} interval. The wildly fluctuating without
approaching a certain limit cosine-square factor appears from the
interference of two waves running in the opposite directions. To
attain the classical result (\ref{eq:classW}) an
additional \textit{averaging} over some finite either
position $\Delta x$ or energy $\Delta n$ intervals around fixed
$|x\rangle$ or $|n\rangle$ states is necessary.
\begin{eqnarray}
\label{eq:avWn_x}
\overline{w_n(x)}^{(x)}&=&\langle n|\left(\int_{x'\in\Delta x} d x'\,
\mathbf{p}_x(x')|x'\rangle\langle x'|\right)|n\rangle=
\langle n|{\hat\rho}^{(x)}|n\rangle\\
\label{eq:avWn_n}
\overline{w_n(x)}^{(n)}&=&
\langle x|\left(\sum_{n'\in \Delta n}\mathbf{p}_{n}(n')
|n'\rangle\langle n'|\right)|x\rangle=
\langle x|{\hat\rho}^{(n)}|x\rangle\,.
\end{eqnarray}
The density matrices ${\hat\rho}^{(q)}$ $(q=n,x)$ describe
{\it incoherent mixtures} of the quantum states within the indicated
intervals. The real and normalized to unity quantities $\mathbf{p}$
characterize the weights of these states. Obviously, the range of
averaging $\Delta x$ should satisfy the condition
$\Delta x\geqslant \pi\hbar/p_c(x)$ to meet the classical behavior.
Similar reasoning leads to the condition
$\Delta E\approx \Delta n\,\hbar \omega_c\geqslant 2\pi\hbar/T$
where $T=2\int_a^b d x/v_c(x)$ is the period of classical
oscillations. Similarly, averaging over the energy interval $\Delta E$
wipes off all scars of eigenstates of a $2D$ billiard with the periods
$T>\hbar/\Delta E$, whereas those with smaller periods still survive
\cite{Bogom88}.
Strictly speaking, decoherence is not perfect as long as off-diagonal
matrix elements of the density matrix still exist. They are complex
and their phases carry some information on the more subtle
interference effects. Complete decoherence is achieved only when the
number of mixing states is so large that the density matrix becomes
proportional to the unit matrix and ceases to depend on the basis in
the Hilbert space of states. As a matter of fact, decoherence can
originate only from: (i) the process of preparation of initially mixed
state, and (ii) mixing induced during the time evolution by a
persisting external noise.
\subsection{The basics of quantum mixed states.}
\label{sec:basicsQuMixStates}
The concept of mixed states plays a paramount role in the problem of
decoherence. A pure quantum state is specified at some moment of
time $t$ by its wave vector $|\psi(t)\rangle$ in the Hilbert space. This
allows, in particular, calculation of the mean value
$O(t)=\langle\psi(t)|{\hat O}|\psi(t)\rangle$
(in what follows we suggest this vector to be normalized to unity,
$\langle\psi(t)|\psi(t)\rangle$=1) of any dynamical quantity
represented by a Hermitian operator ${\hat O}$. Equivalently, the
same state can be described by the density matrix
${\hat\rho (t)=|\psi(t)\rangle}\langle\psi(t)|$
that satisfies two obvious conditions:
\begin{equation}\label{P_rho}
Tr{\hat\rho }(t)=1,\quad {\hat\rho}^2(t)={\hat\rho}(t).
\end{equation}
The second relation is the necessary and sufficient condition of purity
of a quantum state. With these definitions at hands, the mean values
can be presented in one of the following forms
\begin{equation}\label{mean_O}
O(t)=\langle\psi(t)|{\hat O}|\psi(t)\rangle=
Tr\left[{\hat O}\,{\hat\rho} (t)\right]=
\sum_{m,n}O_{n m}(t)\rho_{m n}(t).
\end{equation}
The expression in the right hand side represents the same mean value
in the basis of a complete set of motionless vectors $|n\rangle$ in the
Hilbert space. Since the density matrix is Hermitian, all its diagonal
elements $\rho_{n n}(t)$ are real.
Like any Hermitian matrix, the density matrix $\rho_{m n}(t)$ can be
diagonalized with the help of some unitary transformation. On
account of the conditions (\ref{P_rho}) the density matrix of a pure
quantum state has only one nonzero eigenvalue that equals one.
Obviously, the diagonalization returns us to the original form
${\hat\rho(t)= |\psi(t)\rangle}\langle\psi(t)|$. The transformation
matrix depends on time if the system undergo time evolution. In
practice, a fixed basis of states $|n\rangle$ is, as a rule, a more
relevant choice from physical point of view.
A mixed quantum state is described at any moment of time $t$ by an
incoherent sum
\begin{equation}\label{mixed_rho}
{\hat\rho}(t)=\sum_k \mathbf{p}_k(t)|v_k(t)\rangle\langle v_k(t)|
\end{equation}
of binary contributions where the states $|v_k(t)\rangle$ are the
eigenvectors of the density matrix when $0\leq\mathbf{p}_k(t)\leq 1$
stand for the weights of the corresponding pure fragments. These
weights do not actually depend on time during dynamical evolution
described by some unitary evolution matrix ${\hat U}(t)$. It is not the
case, however, if the system interacts with a noisy background.
In the fixed basis $|n\rangle$, an initially
diagonal incoherent mixture develops off-diagonal elements even
during unitary dynamical evolution. There exists, however, a
convenient invariant measure of state purity called \textit{Purity}
\begin{equation}\label{Purity}
\mathbb{P}(t)=\mathrm{Tr}\left[{\hat\rho}^2(t)\right]=
\sum_k\mathbf{p}^2_k= \mathbb{P}(0)
\end{equation}
that remains constant as long as the noise is absent. The Purity is
restricted to the interval $0\leq\mathbb{P}\leq 1$ and is mounting to
one in the limit of a perfectly pure state. In many respects similar
properties are inherent in the invariant von Neumann entropy
\begin{equation}\label{v_N_ent}
{\cal S}(t)=-\mathrm{Tr}\left[{\hat\rho}(t)\ln{\hat\rho}(t)\right]=
{\cal S}(0)
\end{equation}
which vanishes only in the case of a perfectly pure state.
It is very useful to transfer the story on the, generally mixed,
quantum states evolution to the language of the phase space
\cite{Agarw70,Agarw70a,Sokol08a}. This elucidates analogy and
distinctions between classical and quantum dynamics. A double
Fourier transformation of the density matrix
\begin{equation}\label{Wfunc}
\begin{array}{c}
W(\alpha^*,\alpha;t)=
\frac{1}{\pi^2\hbar}\int d^2\eta\,
\exp\left(\eta\frac{\alpha^\star}{\sqrt\hbar}-
\eta^\star\frac{\alpha}{\sqrt\hbar}\right)
\tilde{\rho}(\eta^\star,\eta;t)=\\
\frac{1}{\pi^2\hbar}\int d^2\eta\,
\exp\left(\eta^*\frac{\alpha}{\sqrt\hbar}-
\eta\frac{\alpha^*}{\sqrt\hbar}\right) \mathrm{Tr}\left[{\hat\rho(t)}\,
{\hat D(\eta)}\right]\,,
\end{array}
\end{equation}
with the operator
${\hat D(\eta)}=\exp(\eta\,{\hat a}^{\dag}-\eta^*{\hat a})$
being the coherent states displacement
defines the {\it Wigner function} $W(\alpha^*,\alpha;t)$
that is a direct quantum counterpart of the classical phase space
distribution function $W^{(c)}(\alpha^*,\alpha;t)$. The complex
variables $\alpha^*,\alpha$ are connected with the standard
action-angle variables $I, \theta$ by the canonical transformations
$\alpha=\sqrt{I}e^{-i\theta},\,\alpha^*=\sqrt{I}e^{i\theta}$.
Deep and detailed analysis of properties of the density matrix and its
significant role for the problem of decoherence has been presented in
the review \cite{Zurek03}.
\subsection{Peres fidelity}
\label{sec:PeresFidelity}
Response of an evolving in time quantum system to a weak external
perturbation is of prime interest in the context of the problem of
stability and reversibility of quantum motion. The customary
quantitative characteristic of sensitivity of classically chaotic
quantum dynamics to such perturbations is the Peres fidelity
\cite{Peres84}:
\begin{equation}\label{Fidelity}
F(t)=\frac{\mathrm{Tr}\left[{\hat\rho}_H(t)\,
{\hat\rho}_{H'}(t)\right]}{\mathbb{P}(t)}=
\frac{\sum_{k l}{\bf p}_k{\bf p}'_l\big|\langle
v_k(t)|v'_l(t)\rangle\big|^2}{\mathbb{P}(t)}\,.
\end{equation}
Fidelity measures the weighted mean distance between two,
generally mixed, quantum states evolving according to slightly
different Hamiltonians ${\hat H}$ and ${\hat H}'$, is bounded in the
interval $[0, 1]$ and equals one when ${\hat H}'={\hat H}$. Its time
decay due to diminishing of the overlap of $v$-eigenvectors
elucidates the sensitivity of the motion to an external influence. In
particular, this quantity enable one to directly connect stability and
reversibility of quantum dynamics with complexity of the quantum
Wigner function. More than that, the notion of the Peres fidelity
directly extends to the classical mechanics
(see sec.II in \cite{Benen03} and \cite{Benen03a}).
The number of $\theta$-harmonics can serve (see below) as a natural
quantitative measure of complexity of the both, classical as well as
quantum phase space (quasi-)distributions
\cite{Sokol08a} (see, however, \cite{Benen12} where another measure of complexity has been proposed that is relevant in the case of systems with more than one degrees of freedom).
\subsection{Open mesoscopic billiards and electron quantum transport}
\label{sec:Billiards&Transport}
Scattering of quantum particles by 2D billiard-like mesoscopic
structures connected to the continuum by two opening leads each supporting
$M$ channels is an intriguing issue that attracted a lot of attention of
theorists as well as experimentalists for at least the last two decades.
If the mean free path of such a particle exceeds the typical size of the
structure the particle's dynamics inside it strongly depends on the shape
of its border. The classical motion in this case becomes stochastic and
quantum scattering is expected to be described well within the
framework of the random matrix approach to the resonance
scattering theory. Intensive experimental studies confirmed in many
respects these expectations. Nevertheless, electron transport
experiments with ballistic quantum dots reveal noticeable and
persisting up to zero temperature loss of the quantum-mechanical
coherence in contravention of predictions of the random-matrix as
well as semiclassical scattering theories.
There exists a number of different methods of accounting for the
decoherence effect in the ballistic quantum transport processes.As
a matter of fact, all of them originate from the pioneering
B\"uttiker's ideas \cite{Buetti86}. The dot is supposed to be
connected someway with a bath of electrons so that the
dot and the bath can exchange electrons in such a manner that the
mean exchange electric current vanishes. Since the incoming electron
carries no information of the phase of the wave function in the dot,
the coherence turns our to be suppressed. The cost payed is that the
number of particle is conserved only in average during the scattering
so that the scattering matrix is not, strictly speaking, unitary.
An alternative approach has been therefore proposed in
\cite{Beena05} with a closed long stub instead of an opening lead.
Thereby unitarity of the $S$-matrix is guaranteed and none of the
electrons is lost at any individual act of scattering.
Decoherence
takes place in this case because of a spatially random time-dependent
external electric field that acts in the stub. As a result, an electron
once penetrated the stub return back in the dot without whatever phase
memory.
In spite of the advantages of the stub model the necessity of
introducing {\it ad hoc} an external time dependent potential seems
to be somewhat artificial. Still another possibility arises if a
time-independent weak interaction is taken into account with a
relatively rare irregular impurities in the semiconductor
heterostructure to whose interface region the electrons are confined.
At that, unavoidable experimental averaging over the scattering energy
on account of finite experimental accuracy in measuring the cross sections
is a point of primary importance.
Due to the interaction with environment, each ``doorway'' resonance
state excited in the structure via external channels gets fragmented
onto a large number $\sim\Gamma_s/d$ (the
spreading width $\Gamma_s$ characterizes the strength of the
coupling to the environment when $d$ is the single-quasi-particle
mean level spacing) of very narrow resonances \cite{Sokol10}.
Only the cross sections averaged over the fine $(\thicksim d)$ structure scale are
observable. Due to such an averaging the doorway resonance states
are damped not only because of escape through such channels but
also due to the ulterior population of the relatively long-lived environmental
states. As a result, transmission of an electron with a given incoming
energy $E_{in}$ through the structure turns out to be an incoherent
sum of the flow formed by the interfering damped doorway
resonances and the retarded flow of the particles re-emitted into the
structure by the environment. Being delayed, the returning electrons
do not interfere with those that escape directly through the external
leads.
The temporal characteristics of the transport are described in detail by the $M\times M$ Smith time delay matrix $Q=-iS^{\dag}dS/dE=Q^{(s)}+Q^{(e)}$
that consists of two contributions $Q^{(s,e)}$. They correspond respectively to the
(modified due to the interaction with the background) time delay within the dot and
additional delay because of the temporary transitions into the background.
We suppose the temperature of the environment to be zero whereas
the energy of the incoming particle $E_{in}$ can be close to or
somewhat above the Fermi surface of electrons in the environment.
Therefore, though the number of the particles is definitely conserved
in each individual event of transmission, there exists a probability
that some part of the electron's energy can be absorbed due to
environmental many-body effects. The both decoherence and
absorption phenomena can be naturally treated within the framework of a
unit microscopic model based on the general theory of the resonance
scattering. Both these effects are controlled by the only parameter:
the spreading width of the doorway resonances.
If the energy $E_{in}$ noticeably exceeds the environment's Fermi
surface and the doorway resonances overlap, the random matrix
approach becomes relevant and ensemble averaging in the doorway
sector is appropriate. Such an averaging, being equivalent to the
energy averaging over the doorway scale $D$, suppresses all
interference effects save the elastic enhancement phenomenon. The latter
is a direct consequence of the time reversal symmetry and manifests itself
in the so called {\it weak localization effect}. However the energy
absorption in the environment violates this symmetry and suppresses the weak
localization.
\section{Particulars}
\label{sec:Particulars}
\subsection{An example of chaotic classical dynamics}
\label{sec:exmplClassChaoDyn}
As an instructive and typical example of chaotic classical system
we consider below a periodically kicked quartic nonlinear oscillator
\begin{equation}\label{H_c}
H(\alpha^{*},\alpha;t)=H^{(0)}+H^{(k)},
\end{equation}
where the unperturbed Hamiltonian function reads
\begin{equation}\label{H^0_c}
H^{(0)}(\alpha^{*},\alpha)=
\frac{p^2}{2m}+\frac{\omega^2_0}{4}\tan^2(\sqrt{2m}\, x)=
\omega_0\,|\alpha|^2+|\alpha|^4 ,
\end{equation}
and the time-dependent perturbation
\begin{equation}
H^{(k)}(\alpha^{*},\alpha;t)=g(t)=
g_0\sum_{\tau}\delta(t-\tau)\,(\alpha^*+\alpha) .
\end{equation}
constitutes a sequence of periodic kicks with strength $g_0$ that
are acting at the instances $\tau=\pm 1, \pm 2, ...\,$.
There are two convenient choices of the canonical variables in this
case: $I, \theta$ or $\alpha, i\alpha^*$ related by the canonical
transformation
$\alpha=\sqrt{I}\,e^{-i\theta}$, $\alpha^*=\sqrt{I}\,e^{i\theta}$.
The action-angle variables $I, \theta$ are ordinary used in the
classical considerations. On the other hand, the advantage of
$\alpha$-variables is that they are directly related to the quantum
creation-annihilation operators ${\hat a}^{\dagger}, {\hat a}$.
The action-angle variables satisfy along a given phase trajectory two
coupled nonlinear equations \cite{Sokol84}:
\begin{eqnarray}
I(t)&=&\left|{\sqrt{{\overset{\circ}I}}
+i\int_0^t d\tau
g(\tau)e^{i[\theta(\tau)-\overset{\circ}\theta]}}\right|^2\\
\theta(t)&=&\int_0^t d\tau[\omega_0+2I(\tau)]
\end{eqnarray}
(here and in what follow we mark the initial values by the overset
circle). A great majority of trajectories becomes exponentially
unstable and, the corresponding motion is globally chaotic when the
strength of the kicks exceeds 1: $|g_0|>1$. Under this condition the
phase correlations decay with time exponentially fast \cite{Sokol07},
\begin{equation}\label{ClasCorrDecay}
\Big|\int d \overset{\circ}I\,d\overset{\circ}\theta\,
W^{(c)}(\overset{\circ}I\,,\overset{\circ}\theta;t=0)\,
e^{i\left(\theta(t)-\overset{\circ}\theta\right)}\Big|^2=
\exp(-t/\tau_c),
\end{equation}
where the characteristic time $\tau_c\sim 1/\Lambda$ is directly
connected to the Lyapunov exponent $\Lambda$. Here
$W^{(c)}(I\,,\theta;t=0)$ is the initial probability distribution in the
phase space. This exponential decay of phase correlations is an
universal fingerprint of the classical dynamical chaos.
In fact, almost all trajectories are alike when the motion is globally
chaotic and, actually, only behavior of manifolds of them is of the real
interest. Therefore the phase space methods appear to be most
relevant in the chaotic regime. The classical phase-space distribution
function $W^{(c)}(\alpha^{*},\alpha;t)$ satisfies the linear Liouville
equation
\begin{equation}\label{L_c_eqn}
i\frac{\partial}{\partial t}\,W^{(c)}(\alpha^*,\alpha;t)=
{\cal\hat L}_c(t)\,W^{(c)}(\alpha^*,\alpha;t)
\end{equation}
with the unitary Liouville operator ${\cal\hat L}_c(t)$ that is a sum
of two, stationary and time-dependent, parts:
${\cal\hat L}_c(t)={\cal\hat L}_c^{(0)}+{\cal\hat L}^{(k)}(t)$.
The first, unperturbed part reads \cite{Sokol08a}
\begin{equation}\label{L_0_c}
{\cal\hat L}_c^{(0)}=
\left(\omega_0+2|\alpha|^2\right)
\left(\alpha^*\frac{\partial}{\partial\alpha^*}-
\alpha\frac{\partial}{\partial\alpha}\right),
\end{equation}
where the operator
$ \left(\alpha^*\frac{\partial}{\partial\alpha^*}-
\alpha\frac{\partial}{\partial\alpha}\right)
=-i\frac{\partial}{\partial\theta}$
formally coincides with the quantum-mechanical angular momentum
operator $\frac{1}{\hbar}{\hat L}_z$. In fact, this operator describes
rotation in the phase-space around the origin with a local angular
velocity $\left(\omega_0+2|\alpha|^2\right)$.
The time-dependent perturbation (kick) operator
\begin{equation}\label{L_k}
{\cal\hat L}^{(k)}(t)=
g_0\sum_{\tau}\delta(t-\tau)
\left(\frac{\partial}{\partial\alpha^*}-
\frac{\partial}{\partial\alpha}\right)
\end{equation}
describes sequence of instant shifts by the distance $g_0$ in the
$\alpha$-plane. The alternating twists and shifts develop
unpredictably complicated pattern of the density distribution
$W^{(c)}(\alpha^*,\alpha;t)$ when the perturbation strength
constant $|g_0|>1$ \cite{Chiri79}. It is of primary importance here
that the unperturbed part ${\cal\hat L}_c^{(0)}$ of the Liouville
operator has \textit{continuous} spectrum of eigenvalues. As a
consequence, the classical phase distribution is structuring
exponentially fast on finer and finer scale during chaotic evolution.
Such a behavior is the paramount property of the classical dynamical
chaos.
Fourier analysis in the phase plane provides a natural tool for
elucidating the process of this structuring. Taking into account
periodicity of the distribution function $W^{(c)}(\alpha^*,\alpha;t)$
with respect to the angle $\theta$,
\begin{equation}\label{Four_repr}
W^{(c)}(I,\theta;t)=
\frac{1}{2\pi}\sum_{m=-\infty}^{\infty}
W_m^{(c)}(I;t)\,e^{i m\theta}\,,
\end{equation}
the most simple and efficient idea is just to follow the upgrowth of the
number of its $\theta$-harmonics. It can be easily quantified with the
help of notion of the Peres fidelity. We suppose for simplicity that the
initial distribution has been isotropic,
$W_m^{(c)}(I;0)=0$ if $m\neq 0$
and let the system to evolve autonomously during some time $t_r$.
At this moment, we probe the system with the help of an instant weak
perturbation $\xi I\delta(t-t_r)$ that produces rotation in the phase
plane by some angle $\xi$. To simplify slightly the subsequent
formulae we suggest the parameter $\xi$ to be a Gaussian random
variable. The probe results then in the instant change of the
distribution
$W^{(c)}(I,\theta;t_r-0)\rightarrow W^{(c)}(I,\theta+\sigma;t_r+0)$
where the parameter $\sigma$ sets the mean squared strength
of the perturbation. The Peres fidelity defined in this case as
\begin{equation}\label{F_c_sen}
F_{(sen)}(\sigma;t_r)=\frac{\int
d^2\alpha\,W^{(c)}\left(\alpha^*,\alpha;t_r\right)
W^{(c)}\left(\sigma\big|\alpha^*,\alpha;t_r+0\right)}
{\int d^2\alpha\,[W^{(c)}\left(\alpha^*,\alpha;t_r\right)]^2}
=\sum_{m=0}^{\infty} e^{-\frac{1}{2}\sigma^2 m^2}
\mathcal{W}^{(c)}_m(t_r)
\end{equation}
characterizes sensitivity of the motion to an external influence.
The quantities $\mathcal{W}^{(c)}_m(t)$
\begin{equation}\label{m_weights_c}
\mathcal{W}^{(c)}_m(t)=(2-\delta_{m0})\,
\frac{\int_0^{\infty}d I\,\Big|W^{(c)}_m(I;t)\Big|^2}
{\sum_{m=-\infty}^{+\infty}\,\int_0^\infty dI |W^{(c)}_m(I;t)|^2}
\end{equation}
that are expressed in the terms of the Fourier harmonics and satisfy
the normalization condition
\begin{equation}\label{W_norm_c}
\sum_{m=0}^{\infty}\mathcal{W}^{(c)}_m(t)=1,
\end{equation}
can naturally be interpreted in the probabilistic manner as the
weights of the corresponding $\theta$-harmonics. Therefore we can
define in the spirit of the linear response approach the mean number
of them $\langle |m|\rangle_t=\sqrt{\langle m^2\rangle_t}$ at a
moment of time $t$ with the aid of the relation
\begin{equation}\label{Mean_m_c}
\langle m^2\rangle_t=
\sum_{m=0}^{\infty} m^2\, \mathcal{W}^{(c)}_m(t)=
-\frac{d^2 F_{(sen)}(\sigma;t)}{d\sigma^2}\big|_{\sigma=0}\,.
\end{equation}
The number $\langle |m|\rangle_t$ can serve as a convenient
quantitative measure of complexity of a phase space distribution at a
given moment of time $t$. Due to exponential instability of classical
dynamics the number of harmonics proliferates exponentially
$\langle |m|\rangle_t\propto e^{t/\tau_c}$ \cite{Sokol08a}.
Let us suppose now that the motion has been reversed at the
moment $t_r$, immediately following after the
time of probing. In view of the fact that the
Liouville evolution operator is unitary one we can transform the
expression ({\ref{F_c_sen}}) into the form
\begin{equation}\label{F_c_rev}
F_{(sen)}(\sigma;t_r)=
\frac{\int d^2\alpha\,W^{(c)}\left(\alpha^*,\alpha;0\right)
W^{(c)}\left(\sigma\big|\alpha^*,\alpha;t_r+0,0\right)}
{\int d^2\alpha\,[W^{(c)}\left(\alpha^*,\alpha;0\right)]^2}
=F_{(rev)}(\sigma;t_r)\equiv F(\sigma;t_r)\,.
\end{equation}
Transformed in this form, fidelity measures the overlap of the
isotropic initial distribution function and the distribution
$W^{(c)}\left(\sigma\big|\alpha^*,\alpha;t_r+0,0\right)$
that, after being changed by the probing perturbation at the moment
$t_r$, has reverted to the initial moment $t=0$. This formula
connects the reversibility of the motion with sensitivity to external
perturbations or, by other words, with complexity of the distribution
function at the reversal moment $t_r$.
\subsubsection{Deterministic diffusion and onset of irreversibility}
\label{sec:Diffusion&classIrrevers}
The phenomenon of the so called deterministic diffusion is one of the
simplest manifestation of the classical dynamical chaos. The mean
value of the action $I(t)$ at any given time $t$ is calculated as follows:
\begin{equation}\label{I_c_mean}
\langle I\rangle_t=\int_0^{\infty}dI\,I\,W^{(c)}_0(I;t)=
\int_0^{\infty}dI\,I\,\overline{W^{(c)}_0(I;t)}=
\langle I\rangle_0+g_0^2\,t
\end{equation}
see, e.g. \cite{Chiri79,Sokol84a,Abarb82}. The averaging is performed in two steps
here. We have smoothed first
the very irregular function $W^{(c)}_0(I;t)$ over a small interval
surrounding a fixed value $I$ within which the factor $I$ does not
appreciably change. Successive $I$-integration with the ``coarse
grained'' distribution $\overline{W^{(c)}_0(I;t)}$ results in the
diffusive increase of the mean action. Similarly, we can also calculate
the moments $\langle I^k\rangle_t,\, k=2, 3,...,k_t$. The longer is the
duration $t$ of the evolution the larger the power $k_t$ at which the
two-step integration procedure is still valid.
It is obvious that, formally, the deterministic diffusion
(\ref{I_c_mean}) is perfectly time-reversible.
However the exponential instability makes this statement impractical.
Even a very weak external noise destroys reversibility and turns the
motion into irreversible process.
To illustrate this statement, we add to our Hamiltonian function a new
term
\begin{equation}\label{noise_c}
H^{(noise)}(\alpha^*,\alpha;t)=
|\alpha|^2\sum_{\tau}\xi_{\tau}\,\delta(t-\tau),
\quad \langle\xi_{\tau}\rangle=0,\,
\langle\xi_{\tau}\xi_{\tau'}\rangle=\sigma^2\delta_{\tau\tau'}\,.
\end{equation}
that describes a stationary Gaussian noise with the strength
parameter $\sigma$. Each period of unperturbed evolution is
immediately followed by a phase plane rotation by a random angle
$\xi$. Averaging over the noise realizations sets up the coarse grained
distribution function
\begin{equation}\label{c_g_W}
W^{(c)}(\sigma|I,\theta;t)=
\overline{W^{(c)}(\{\xi\}|I,\theta;t)}^{(noise)}.
\end{equation}
The averaging suppresses Fourier harmonics with respect to the both
canonical variables thus extenuating irregular oscillations of the
distribution. They are perfectly smoothed away in the limit of very
strong noise
$\sigma\rightarrow\infty$ :
$W^{(c)}_{|m|\geqslant 1}(\infty|I,;t)=0$
and
\begin{equation}\label{Inf_noise}
W^{(c)}_{0}(\infty|I;t)=\frac{1}{\langle I\rangle_0+g_0^2\,t}
\exp\left(\frac{I}{\langle I\rangle_0+g_0^2\,t}\right).
\end{equation}
In this case the fidelity (\ref{F_c_rev}) equals
$F_{(rev)}(\infty|t_r)=e^{-\frac{t_r}{\tau_c}}\,.$
Indeed, the backward evolution starts with the isotropic distribution
(\ref{Inf_noise}) at the moment $t_r$ and develops
$m(t_r)\propto e^{t_r/\tau_c}$
$\theta$-harmonics by the time $t=0$.
Thus the overlap with the initial isotropic distribution
$W^{(c)}\left(\alpha^*,\alpha;0\right)$ is exponentially small. The
standard diffusion takes place in the backward evolution as well.
Under influence of the noise of a moderate level $\sigma$ the
backward diffusion is delayed for some time $\tau_d\propto\ln\sigma$
\cite{Iked95}
during which the system partly recovers its preceding states. But
after that the diffusion recommences.
\subsection{Quantum evolutions of classically chaotic system}
\label{sec:QuantumEvolutionOfChaos}
The Hamiltonian of the quantum counterpart of the classical oscillator
considered above is immediately obtained by the substitutions:
$\alpha\Rightarrow\sqrt{\hbar}\,{\hat a}$,
$\alpha^*\Rightarrow\sqrt{\hbar}\,{\hat a}^{\dag}.$
\begin{equation}
{\hat H}({\hat a}^{\dag},{\hat a};t)=\hbar\omega_0 {\hat n}+
\hbar^2{\hat n}^2+g_0\sum_{\tau}\delta(t-\tau)\,
\sqrt{\hbar}\,({\hat a}^{\dag}+{\hat a}).
\end{equation}
(In the chosen units all parameters: $\hbar$, $\omega_0$, $g_0$
are dimensionless.)
The quantum evolution
$\hat\rho(t)=\hat{\cal U}(t)\hat\rho(0)\hat{\cal U}^{\dag}(t)$
of the density matrix is described by the unitary operator
$\hat{\cal U}(t)$ that is $t$ successive repetitions of the one period
Floquet operator $\hat U$: $\hat{\cal U}(t)={\hat U}^t$,
the latter being a sequence of instant change of the excitation
number (coherent state displacement operator) and subsequent free
rotation in the phase plane induced by the time-independent
Hamiltonian
${\hat H}^{(0)} ({\hat a}^{\dag},{\hat a})=
\hbar\omega_0 {\hat n}+\hbar^2{\hat n}^2$:
\begin{equation}
\hat U=e^{-\frac{i}{\hbar}
\hat{H}^{(0)}}\hat D\left(i\frac{g_0}{\sqrt\hbar}\right)
\label{eq:U}
\end{equation}
The Wigner function (\ref{Wfunc}) satisfies the quantum Liouville
equation:
\begin{equation}\label{L_q_eqn}
i\frac{\partial}{\partial t}\,W(\alpha^*,\alpha;t)=
{\cal\hat L}_q(t)\,W(\alpha^*,\alpha;t)\\
\end{equation}
where the quantum Liouville operator is, similar to the classical one,
the sum
${\cal\hat L}_q(t)={\cal\hat L}_q^{(0)}+{\cal\hat L}^{(k)}(t)$.
The only, but of primary importance, distinction from the classical
case is appearance of a new second-derivative term in the rotation
operator
\begin{equation}\label{L_0_q}
{\cal\hat L}_q^{(0)}=\left(\omega_0-\hbar-\frac{1}{2}\hbar^2
\frac{\partial^2}{\partial\alpha^*\partial\alpha}+2|\alpha|^2\right)
\left(\alpha^*\frac{\partial}{\partial\alpha^*}-
\alpha\frac{\partial}{\partial\alpha}\right).
\end{equation}
The combination
$-\frac{1}{2}\hbar^2\frac{\partial^2}{\partial\alpha^*\partial\alpha}
+2|\alpha|^2$
formally coincides with the quantum Hamiltonian of a 2D isotropic
linear oscillator. As a result, the spectrum of the operator
(\ref{L_0_q}) becomes discrete, in contrast to his classical
counterpart. This important fact is a manifestation of {\it the
quantization of the phase space}.
Below we will consider evolution of an initially isotropic and,
generally, mixed state. If we choose the density matrix in the form
\begin{equation}\label{In_rho}
{\hat\rho}(0)=
\frac{\hbar}{\Delta+\hbar}
\sum_{n=0}^{\infty}\left(\frac{\Delta}{\Delta+\hbar}\right)^n
|n\rangle\langle n|
\end{equation}
the corresponding Wigner function
\begin{equation}\label{W_q_in_Delta}
W(\alpha^*,\alpha;0)=
\frac{1}{\Delta+\hbar/2}\,e^{-\frac{|\alpha|^2}{\Delta+\hbar/2}}.
\end{equation}
is a Poissonian distribution with respect to the action variable
$I=|\alpha|^2$.
In particular, in the case $\Delta=0$ this state turns into the pure
ground state ${\hat\rho}(0)=|0\rangle\langle 0|$ whose Wigner
function
$W(\alpha^*,\alpha;0)=
\frac{2}{\hbar}\, e^{-\frac{2|\alpha|^2}{\hbar}}$
occupies the minimal quantum cell $\hbar/2$.
As compared with the classical Liouville equation the quantum one
\begin{equation}
i\frac{\partial}{\partial t}\,W(\alpha^*,\alpha;t)=
{\cal\hat L}_c\,W(\alpha^*,\alpha;t)
-\frac{1}{2}\hbar^2\frac{\partial^2}{\partial\alpha^*\partial\alpha}
W(\alpha^*,\alpha;t);
\end{equation}
contains an additional proportional to $\hbar^2$ term with higher
(second) order derivative. Even being initially very small, this term
can be neglected only, because of the classical exponential instability,
during a short time interval which is called the Ehrenfest time
$t_E=\tau_c\,\ln\frac{2\langle I\rangle_{t_E}}{\hbar}$. After this
interval the quantum effects dominate and difference between the
two dynamics becomes crucial. Indeed, meshing of the Wigner
function stops quite soon on account of the quantization of the phase
space.
\subsection{Stability and reversibility vs complexity of quantum states.}
\label{sec:quStabRev&Complexity}
Relying upon the analogy mentioned above between the classical
phase space distribution function on the one hand and the quantum
Wigner function on the another we will confront below the typical
features of the two corresponding "chaotic" dynamics. Quantization of
the phase space implies a much simpler structure of the Wigner
function than that of the corresponding classical distribution. The
Liouvillian phase space approach that we use allows us to measure
complexity of this function using the same tools that have been
utilized in the classical case considered in subsection \ref{sec:exmplClassChaoDyn}.
\subsubsection{Response to a single probe}
\label{sec:SingleProbe}
In much the same fashion as in eqs. (\ref{F_c_sen}-\ref{F_c_rev}),
the response of a quantum system to an instant probe at a moment of
time $t$ is measured with the help of the quantum Peres fidelity
(\ref{Fidelity}) defined by the relation
\begin{equation}\label{F_q}
F(\sigma|t)=\frac{\mathrm{Tr}\left[{\hat\rho(t)}\,
{\hat\rho(\sigma|t)}\right]}{\mathbb{P}(t)}=
\sum_{m=0}^{\infty}
e^{-\frac{1}{2}\sigma^2 m^2}\mathcal{W}_m(t)
\end{equation}
where the quantities
\begin{equation}\label{calW_m_q}
\mathcal{W}_m(t)=
\frac{(2-\delta_{m0})}{\mathbb{P}(t)}\sum_{n=0}^{\infty}
\Big|\langle n+m\big|{\hat\rho}(t)\big|n\rangle\Big|^2,\quad\,
\sum_{m=0}^{\infty}\mathcal{W}_m(t)=1
\end{equation}
are direct quantum analogs of the classical weights
(\ref{m_weights_c}) of the $\theta$-harmonics. Indeed, they can be
expressed \cite{Sokol08a,Sokol08} in the terms of Fourier amplitudes
of the Wigner function by the formulae:
\begin{equation}\label{m_weights_q}
\mathcal{W}_m(t)=
\frac{(2-\delta_{m0})}{\mathbb{P}(t)}\,
\hbar\int_0^{\infty}d I\,\Big|W_m(I;t)\Big|^2
\end{equation}
(see Fig.\ref{fig:quWm}).
\begin{equation}\label{Purity_q}
\mathbb{P}(t)=
\mathrm{Tr}\left[{\hat\rho}^2(t)\right]=
\sum_{m=-\infty}^{+\infty}\,
\int_0^\infty dI |W_m(I;t)|^2=\mathbb{P}(0).
\end{equation}
These relations differ from eqs. (\ref{m_weights_c}, \ref{W_norm_c})
only by replacement $W_m^{(c)}(I;t)\Rightarrow W_m(I;t)$ of the
Fourier amplitudes of the classical distribution function by those of
the quantum Wigner function. It is worth noting that the phases of
the off-diagonal matrix elements of the density matrix at the moment
$t$ fall out of the weights (\ref{calW_m_q}). Nevertheless such
phases do play a certain, though not significant (see below), role
during the preceding evolution.
\begin{figure}[!h]
\centering
\includegraphics[width=80mm,
keepaspectratio=true]{quWm.eps}
\caption{Distribution of harmonics ${\cal W}_m(t)$ as a function
of $m$, at different times $t=10, 30$ and $50$ from bottom to top
(these curves are scaled by a factor 0.01, 0.1 and 1, respectively).
Left and right panels correspond to pure ($\Delta=0$) and mixed
($\Delta=25$) initial states, see eq.(\ref{In_rho}). Other parameters
of simulations are $g_0=2$, $\hbar=1$.}
\label{fig:quWm}
\end{figure}
\subsubsection{Complexity of a quantum state.}
\label{sec:ComplexityQuState}
Just as it has been in the case of a classical phase space distribution,
complexity of a quantum state can be characterized by the number
$\langle |m|\rangle_t$ of $\theta$-harmonics of the Wigner function.
The corresponding weights are given now by eqs. (\ref{m_weights_q},
\ref{Purity_q}) and again
\begin{equation}\label{mean_m_q}
\langle m^2\rangle_t=
\sum_{m=0}^{\infty} m^2\,\mathcal{W}_m(t)=
-\frac{\rm d^2 F(\sigma;t)}{\rm d\sigma^2}\Big|_{\sigma=0}.
\end{equation}
It should be emphasized that the chosen measure is {\it equally valid}
in the quantum as well as classical cases. This allows direct
comparison of the main features of the both dynamics. In particular,
whereas the number of harmonics of the classical distribution
function increases, due to the exponential instability, exponentially
during the {\it whole} time of evolution, in the quantum case the
exponential regime is, generally speaking, restricted to the Ehrenfest
time interval. These statements are illustrated in the
Fig.\ref{fig:Wquantclass} where the dependence of the quantity
$\langle m^2\rangle_t$ on the time is shown for a set of different
values of the effective Plank's constant.
\begin{figure}
\centering
\includegraphics[width=8.cm,angle=0
]{avMs.eps}
\caption{$\langle m^2\rangle_t$ before and after the Ehrenfest time.
Triangles, diamonds and squares: $\hbar = 1,\, 0.1$ and $0.01$
respectively. Classical dynamics is shown by empty red circles,
the gray line presents an exponential fit.
The initial phase area {\it holds constant} $1/2$ in all 3 cases.
The kick strength used here is $g_0=1.5$.}
\label{fig:Wquantclass}
\end{figure}
The Plank's constant plays here a twofold role: on the one part
it fixes the phase volume of the elementary quantum cell that is
occupied in the phase space by the pure ground state
${\hat\rho}(0)=|0\rangle\langle 0|$, and, on the other part, it governs
the dynamics via the evolution equation. In the
Fig.\ref{fig:Wquantclass} we keep constant the size 1/2 of the initial
Wigner distribution (\ref{W_q_in_Delta}) by choosing
$\Delta=1/2-\hbar/2$. The initial state is pure ($\Delta=0$) when
$\hbar=1$ (triangles) but becomes more and more mixed
($\Delta\approx 1/2\gg\hbar$) in the cases $\hbar=0.1$ (diamonds)
and $\hbar=0.01$ (squares) correspondingly. It is clearly seen that
the smaller is the value of {\it dynamical} Plank's constant the longer
the classical exponential regime lasts. By other words, the mixing of
the initial state suppresses the quantum interference effect and
restores the classical behavior.
\subsubsection{Information entropy vs von Neumann entropy}
\label{sec:Entropies}
The probabilistic meaning of the quantities $\mathcal{W}_m(t)$
allows us to introduce the Shannon or {\it information entropy}:
\begin{eqnarray}\label{Inf_entropy}
{\cal I}(t)&=&
-\sum_{m=0}^{\infty}\mathcal{W}_m(t)\,\ln\mathcal{W}_m(t)\\
&=&
\left\{
\begin{array}{ll}
0, & t=0,\\
\ln\langle |m|\rangle_t+1-\frac{\ln 2}{2}+O(1/\langle |m|\rangle_t), \qquad &\langle |m|\rangle_t\gg 1.\\
\end{array}
\right.
\end{eqnarray}
This is another, though equivalent, possible way to characterize the
complexity of a quantum state. Such a choice turns out to be even
more convenient for our further purposes. This entropy starts from
zero (the initial state has no harmonics but zero), increases linearly
(with a slope defined by the classical characteristic time $\tau_c$)
during the Ehrefest time and then slows down to the quantum
logarithmic regime. Since the number of harmonics at sufficiently
large time weakly depends on the peculiar properties of the initial
state, being practically the same for pure and mixed ones, this
entropy is practically insensitive to quantum correlations.
On the contrary, the invariant von Neumann entropy
\begin{equation}
{\cal S}(t)=
-\mathrm{Tr} \left[{\hat\rho}(t)\,\ln{\hat\rho}(t)\right]={\cal S}(0)
\end{equation}
is {\it perfectly} sensitive to quantum correlations (hence the name
``correlational'' \cite{Sokol98}). This entropy does not depend on
time as long as the evolution remains unitary and equals zero
${\cal S}(t)=0$ when the state is pure. The coherence and quantum
correlations can be destroyed only in the presence of an persistent
external noise or during the process of preparation of the initial state
(see below).
\subsubsection{Persistent noise.}
\label{sec:Noise}
The stationary noise is described in our model by the Hamiltonian
operator
\begin{equation}\label{noise_q}
{\hat H}^{(noise)}({\hat a}^{\dag},{\hat a};t)=
\hbar\,{\hat n}\sum_{\tau}\xi_{\tau}\,\delta(t-\tau),
\quad \langle\xi_{\tau}\rangle=0,\,
\langle\xi_{\tau}\xi_{\tau'}\rangle=
\sigma^2\delta_{\tau\tau'}\,.
\end{equation}
As a result, the evolution operator takes the form
\begin{equation}\label{U(ksi,t)}
\hat{\cal U}(\{\xi\};t)=
\prod_{\tau=1}^{\tau=t}\left[e^{-i\xi_{\tau}{\hat n}}\,\hat U\right].
\end{equation}
This operator remains unitary for any fixed noise realization $\{\xi\}$
(history). Accordingly, a pure initial state remains pure during
the whole time of evolution. At a running moment $t$, the
excitation of the oscillator and the degree of
anisotropy of the Wigner function are characterized by the
probability distributions \cite{Sokol09}
\begin{equation}
w_n({\xi}; t)=\langle n | \hat{\rho}({\xi}; t) |n\rangle,
\label{eq:rho_xi}
\end{equation}
and
\begin{equation}\label{calW_m_q_xi}
\mathcal{W}_m(t)=
\frac{(2-\delta_{m0})}{\mathbb{P}(t)}\sum_{n=0}^{\infty}
\Big|\langle n+m\big|{\hat\rho}({\xi}t)\big|n\rangle\Big|^2,
\quad\,
\sum_{m=0}^{\infty}\mathcal{W}_m(t)=1
\end{equation}
where the density matrix
\begin{equation}
\hat{\rho}({\xi};t)=
\hat{\cal U}(\{\xi\};t)\hat\rho(t=0)\hat{\cal U}^{\dagger}(\{\xi\};t)
\end{equation}
is defined for some fixed noise realization $\{\xi\}$. The initial
state can be chosen to be the ground one
$ \hat\rho(t=0)=\big|0\rangle\langle 0\big|$, because after a few first
kicks the distributions (\ref{eq:rho_xi},\ref{calW_m_q_xi}) aquire
a practically general exponential form (see Fig.\ref{fig:ns_wn_Wm}).
Averaging over noise realizations keeps the
slopes of these distributions
unchanged but kills the wild fluctuations around their regular
exponential decay. In the limit of strong noise $\sigma\gg1 $
(see thick dashed lines in Fig.\ref{fig:ns_wn_Wm}) the
noise acts as a ``coarse graining'' that reproduces the behavior of
the corresponding classical distributions \cite{Sokol09}.
There exist "self-averaging" quantities
like mean excitation number $\langle n\rangle_t$ or the mean
number $\langle |m|\rangle_t$ of $\theta$-harmonics that do not
depend, in fact, on the noise realization \cite{Sokol09}
\begin{equation}
\langle n(\{\xi\};t)\rangle=n(\sigma;t),\,\,\,
\langle m^2(\{\xi\};t)\rangle=
m^2(\sigma;t)=n(\sigma;t)\left(n(\sigma;t)+1\right).
\end{equation}
\begin{figure}[!h]
\centering
\includegraphics[width=80mm, keepaspectratio=true]
{ns_wn_Wm.eps}
\caption{Probability distributions
(\ref{eq:rho_xi}), (\ref{calW_m_q_xi}) at the moment $t=80$
with no noise ($\sigma=0$, thin black lines), weak noise
($\sigma=0.001$, thin green lines) and strong noise ($\sigma =1$,
thick dashed red line). In the two latter cases the distributions are
averaged over $10^3$ realizations. The initial state is pure ($\Delta=0$, see eq.(\ref{In_rho})). Other parameters
of simulations are $g_0=2$, $\hbar=1$.}
\label{fig:ns_wn_Wm}
\end{figure}
On the contrary, the quantum Peres fidelity defined as
\begin{equation}\label{Fid_q}
F(\{\xi\}; t)=\frac{\mathrm{Tr}
\left[{\hat\rho(t)}\, {\hat\rho(\{\xi\};t)}\right]}{\mathbb{P}(t)}.
\end{equation}
is not a self-averaging quantity and wildly fluctuates from one noise
history to
another. Therefore averaging over all possible noise realizations is
necessary to obtain a reasonably simple and adequate measure:
\begin{equation}\label{Fid_q_av}
F(\sigma;t)=\overline{F(\{\xi\};t)}^{\{\xi\}}
=\mathrm{Tr}
\left[{\hat\rho(t)}\,{\hat\rho^{(av)}(\sigma;t)}\right] .
\end{equation}
This procedure brings into consideration the average density matrix
$\rho^{(av)}(t)$ whose one step evolution is described by the
transformation
\begin{equation}\label{1step}
\langle n'|{\hat\rho^{(av)}(\sigma;\tau)}|n\rangle=
e^{-\frac{1}{2}\sigma^2(n'-n)^2}\,
\langle n'|{\hat U}{\hat\rho^{(av)}(\sigma;\tau-1)
{\hat U}^{\dag}}|n\rangle.
\end{equation}
The noise suppresses off-diagonal matrix elements of the density
matrix thus gradually cutting down the number of harmonics of the
corresponding Wigner function. The evolution is not unitary
anymore. The latter entails state mixing, loss of memory on the
initial state and suppression of the quantum interference. These
effects show up in the behavior of the von Neumann entropy defined
in the terms of the averaged density matrix as
\begin{equation}\label{N_entropy_ev}
{\cal S}(\sigma;t)= -\mathrm{Tr}
\left[{\hat\rho}^{(av)}(\sigma;t)\,\ln{\hat\rho}^{(av)}(\sigma;t)
\right].
\end{equation}
The Fig.\ref{fig:two_entropies} illustrates this behavior in
comparison with that of the information Shannon entropy
${\cal I}(t)$ (see (\ref{Inf_entropy})). The entropy
${\cal S}(\sigma;t)$ increases the faster the larger is the level of
noise (the full lines from bottom to top) and approaches the
information Shannon ${\cal I}(\sigma;t)$ entropy (circles) from
below. At some time $t_{(dec)}$, both the entropies coincide and go
after that together. All coherent effects are washed away by this
time. Therefore the entropy (\ref{N_entropy_ev}) is suitable
{\it for tracing the gradual loss of the quantum coherence.}
\begin{figure}[!h]
\centering
\includegraphics[width=80mm,
keepaspectratio=true]{entropy.eps}
\caption{Von Neumann entropy:
$\sigma= (0.125,1, 8, 64, 512)\cdot 10^{-3}$,
solid (black, blue, green, magenta, red) lines from bottom to top.
Circles: information entropy
${\cal I}(t);\,\,\,{\cal S}(\sigma;t>t_{(dec)}) \to {\cal I}(t)\,$.
In these simulations we use $g_0=2$, $\hbar=1$.}
\label{fig:two_entropies}
\end{figure}
The decoherence time $t_{(dec)}$ can be estimated as
\cite{Sokol09}
\begin{equation}
t_{(dec)}(\sigma)\thicksim\sqrt{\frac{\hbar}{\sigma^2 D}}
\end{equation}
where $D$ is the classical diffusion coefficient. Henceforth the
system occupies the maximal phase area accessible at the running
value of excitation thus reaching a sort of equilibrium. Finally, the
system expands "adiabatically", the both entropies being almost
constant. The further evolution turns out to be \textit{Markovian}
\cite{Sokol09}.
\subsubsection{Time independent perturbation. Mixed initial state.}
\label{sec:TimeIndependPertirb}
As it has already been noted above, the quantum interference turns
out to be somewhat suppressed if the evolution started from a mixed
initial state. We consider below another kind of decoherence that
takes place in the case of a time-independent perturbation.
\begin{equation}
\hat{H}_V=\hat{H}^{(0)}+\varepsilon {\hat V},
\end{equation}
where the unperturbed Hamiltonian describes, as before, dynamics
of the classically chaotic nonlinear oscillator. The perturbed and
unperturbed motions are juxtaposed by means of the unitary
Loschmidt echo operator ${\hat f}(t)$ \cite{Sokol07}
\begin{eqnarray}
{\hat f}(t)&=&{\hat U}^{\dag}(t){\hat U}_V(t)\\
{\hat f}(t)&=&T\exp\left[-i\frac{\varepsilon}{\hbar}\int_0^t d\tau
{\hat{\cal V}}(\tau)\right]\\
{\hat{\cal V}}(\tau)&=&{\hat U}^{\dag}(t)\,{\hat V}{\hat U}(t).
\end{eqnarray}
In spite of the fact that the perturbation ${\hat{\cal V}}(\tau)$
evolves chaotically, the quantum coherence {\it is in no way
spoiled} as long as the initial state is pure. More than that, one
might think that even if the initial state is an incoherent mixture
quantum coherence can be rapidly generated by producing complex
off diagonal matrix elements during dynamical evolution.
Nevertheless, we will show below that if the system is classically
chaotic and the evolution \emph{starts from a wide incoherent
mixed state}, then the initial incoherence persists due to the
intrinsic classical chaos so that the quantum phases remain
irrelevant \cite{Sokol07}.
In the case of a pure initial state
${\hat\rho}(0)=
|\overset{\circ}\psi\rangle\langle\overset{\circ}\psi|$
the Peres fidelity (\ref{Fidelity}) is simply the probability
\begin{equation}\label{Pure_in}
F_{\overset{\circ}\psi}(t)=
|\langle \overset{\circ}\psi|{\hat f(t)}|\overset{\circ}\psi\rangle|^2=
\left|\mathrm{Tr}\left[{\hat f(t)}\overset{\circ}\rho\right]\right|^2
\end{equation}
to survive in this state under influence of a chaotically evolving
perturbation ${\hat{\cal V}}(\tau)$ till the time $t$. When the
evolution starts from a mixed state
${\hat\rho}(0)=
\sum_k p_k
|\overset{\circ}\psi_k\rangle \langle\overset{\circ}\psi_k|,\,\,
\sum_k p_k=1$,
the expression (\ref{Pure_in}) can be generalized in two different
ways. The first of them leads to the standard definition
(\ref{Fidelity}) that can be rewritten identically in the form
\begin{equation}\label{Fidelity+}
F(t)=\frac{1}{\mathrm{Tr}{\hat\rho^2(0)}}
\sum_{k,k^\prime} p_{k} p_{k^\prime}W_{k k^\prime}(t)
\end{equation}
where the quantities
$W_{k k^\prime}(t)=
|\langle \overset{\circ}\psi_k|{\hat f}(t)|
\overset{\circ}\psi_{k^\prime}\rangle |^2$
are probabilities of transitions induced by the unitary
transformation ${\hat f}(t)$.
The influence of coherent effects is hidden in the dynamics of the
complex matrix elements
$f_{k k'}(t)=\langle \overset{\circ}\psi_k|{\hat f}(t)|
\overset{\circ}\psi_{k^\prime}\rangle$.
Another way of generalization is suggested by the experimental
configuration with periodically kicked ion traps proposed in
\cite{Gardi97}. In such Ramsey type interferometry experiments
one directly accesses the fidelity amplitudes (see \cite{Gardi97}
rather than their square moduli. Motivated by this consideration,
we will consider further the quantity
\begin{equation}\label{Aledg}
{\cal F}(t)=
\left|\mathrm{Tr}\left[{\hat f(t)}{\hat\rho}(0)\right]\right|^2=
\Big|\sum_k p_k f_k(t)\Big|^2=
\sum_k p^2_k F_k(t)+\sum_{k,k'}(1-\delta_{kk'}) p_k
p_{k'}f_k(t)f_{k'}^*(t)\,,
\end{equation}
that is obtained by directly extending the formula (\ref{Pure_in}).
Below we refer to this new quantity as {\it allegiance}. The first
term in the r.h.s. is the sum of fidelities
$F_k=|f_k|^2=|\langle \overset{\circ}\psi_k|{\hat f}|
\overset{\circ}\psi_k\rangle|^2$
of the individual pure initial states with weights $p_k^2$, while the
second, interference term, depends on the relative phases of fidelity
amplitudes. If the number $K$ of pure states
$|\overset{\circ}\psi_{k}\rangle$ that form the initial mixed state
is large, $K\gg 1$, so that $p_k\backsimeq 1/K$ for $k\leqslant K$
and zero otherwise, the first term is $\backsim 1/K$ at the initial
moment $t=0$ while the second term $\backsim 1$. Therefore, in
the case of a wide mixture, the decay of the function ${\cal F}(t)$ is
determined by the second sum of interfering contributions.
Therefore, in contrast to the standard Peres fidelity (\ref{Fidelity},
\ref{Fidelity+}), the allegiance ${\cal F}$ directly {\it accounts for
the quantum interference} and can be expected to retain quantal
features even in the deep semiclassical region. This is not, however,
the case as will be shown below.
Analytical calculation of the pure-state fidelity
$F_{\overset{\circ}\alpha}(t)$ for a pure coherent quantum state
$|\overset{\circ}\alpha\rangle$ as well as the allegiance
${\cal F}(t)$ for an incoherent mixed state can be performed with
the help of expressing the both quantities in terms of the Feynman's
path integral in the oscillator's phase plane. A method of
semiclassical evaluation of this integral has been worked out in
\cite{Sokol84}. Referring the reader to this paper for all technical
details we present below the main results of the calculations
\cite{Sokol07}.
These results are quite different it the two cases of our interest. If
the initial state is a pure coherent one
$|\overset{\circ}\alpha\rangle$ the fidelity amplitude is found to be
\begin{equation}\label{fidocoherent}
f_{\overset{\circ}\alpha}(t)=\frac{2}{\pi\hbar}\int d^2\delta\,
e^{-\frac{2}{\hbar}|\delta|^2}\exp\left\{i\frac{\sigma}{2}
\left[{\tilde\theta_c(t)}-
{\overset{\circ}{\tilde\theta_c}}\right]\right\},
\end{equation}
where the phase
${\tilde\theta_c(t)}=\theta_c(\omega_0-2|\delta|^2;
\overset{\circ}\alpha^*+\delta^*,
\overset{\circ}\alpha+\delta;t)=\int_0^t
d\tau[\omega_0-2|\delta|^2+2{\tilde I}_c(\tau)$.
It should be stressed that the fidelity
$F_{\overset{\circ}\alpha}(t)=|f_{\overset{\circ}\alpha}(t)|^2$
does not decay in time at all if the quantum fluctuations described
by the integral over $\delta$ in (\ref{fidocoherent}) are neglected.
On the initial stage of the evolution
$t\ll\frac{1}{\Lambda}\ln\frac{2}{\varepsilon}$, while the phases
${\tilde\theta_c(t)}$ are not yet perfectly randomized the fidelity
$F_{\overset{\circ}\alpha}$
decays, because of classical exponential instability,
super-exponentially:
$$F_{\overset{\circ}\alpha}(t)\approx
\exp\left(-\frac{\varepsilon^2}{4\hbar}e^{\Lambda t}\right),$$
\cite{Silve02, Iomin04}. During this time the contribution of
the averaging over the initial Gaussian distribution in the
classical $\overset{\circ}\alpha$ phase plane dominates while the
influence of the quantum fluctuations of the linear frequency
remains negligible. Such a decay has, basically, a classical nature
\cite{Eckha03} and the Planck's constant appears only as the size of
the initial distribution. For larger times the quantum fluctuations
reduce the fidelity decay to exponential law
$F_{\overset{\circ}\alpha}(t)=\exp(-2\Lambda t).$
The situation changes dramatically if the initial state is an
incoherent mixture. More precisely, we consider a mixed initial state
represented by a Glauber's diagonal expansion
\cite{Glaub63,Glaub63a}
${\hat\rho}(0)=~\int d^2\overset{\circ}\alpha {\cal
P}(|\overset{\circ}\alpha-
\overset{\circ}\alpha_c|^2)|\overset{\circ}\alpha\rangle
\langle\overset{\circ}\alpha|$ with a wide positive definite weight
function ${\cal P}$ which covers a large number of quantum cells.
Note that here and in the following we assume that the initial
mixture is isotropically distributed in the phase plane around a fixed
point $\overset{\circ}\alpha_c\,$,
with the density
${\cal P}_{\overset{\circ}\alpha_c}(\overset{\circ}\alpha^*,
\overset{\circ}\alpha)={\cal P}(|\overset{\circ}\alpha-
\overset{\circ}\alpha_c|^2)$. Then allegiance equals \cite{Sokol07}
${\cal F}(t;{\overset{\circ}\alpha_c})=
|f(t;{\overset{\circ}\alpha_c})|^2$,
where
\begin{equation}\label{ampmx}
\begin{array}{c}
f(t; {\overset{\circ}\alpha_c})=\int d^2\overset{\circ}\alpha\,
{\cal P}(|\overset{\circ}\alpha-\overset{\circ}\alpha_c|^2)
f_{\overset{\circ}\alpha}(t)\\
\approx\int d^2\overset{\circ}\alpha\,
{\cal P}(|\overset{\circ}\alpha-
\overset{\circ}\alpha_c|^2)\exp\left\{i\frac{\varepsilon}{2\hbar}
\left[\theta_c(t)-\overset{\circ}\theta_c(0)\right]\right\}.
\end{array}
\end{equation}
This formula directly relates the decay of a \textit{quantum}
quantity, the allegiance, to that of correlation function of the
\textit{classical} phases. No quantum features are present in the
r.h.s. of (\ref{ampmx}).
Summarizing,
the decay pattern of the allegiance ${\cal F}(t)$ depends on the
value of the parameter $\sigma=\varepsilon/\hbar$. In particular, for
$\sigma\ll 1$, we recover the well known Fermi Golden Rule (FGR)
regime. Indeed, in this case the cumulant expansion can be used,
$\ln f(t; {\overset{\circ}\alpha_c})=
\sum_{\kappa=1}^{\infty} \frac{(i\sigma)^{\kappa}}{\kappa!}
\chi_{\kappa}(t)\,.$
All the cumulants are real, hence, only the even ones are significant.
The lowest of them,
\begin{equation}\label{qum}
\begin{array}{c}
\chi_2(t)=\int_0^t d\tau_1\int_0^t d\tau_2\langle \left[
I_c(\tau_1)- \langle I_c(\tau_1)\rangle\right]
\left[ I_c(\tau_2)-\langle I_c(\tau_2)
\rangle\right]\rangle\equiv\int_0^t d\tau_1 \int_0^t
d\tau_2 K_I(\tau_1,\tau_2)\;,
\end{array}
\end{equation}
is positive. Assuming that the classical autocorrelation function
decays exponentially,
$K_I(\tau_1,\tau_2)=
\langle\left(\Delta I_c\right)^2\rangle
\exp\left(-|\tau_1-\tau_2|/\tau_I\right)$
with some characteristic time $\tau_I$, we obtain
$\chi_2(t)=2\langle\left(\Delta I_c\right)^2\rangle\tau_It=2Kt$
for the times $t>\tau_I$ and arrive, finally, at the FGR decay law
${\cal F}(t; {\overset{\circ}\alpha}_c) =\exp(-2\sigma^2Kt)$
\cite{Cerru02,Jacqu01,Prose02}. Here
$K=\int_0^{\infty}d\tau K_I(\tau,0)=
\langle\left(\Delta I_c\right)^2\rangle\tau_I\,.$
The significance of the higher connected correlators
$\chi_{\kappa\geq 4}(t)$ grows with the increase of the parameter
$\sigma$. When this parameter roughly exceeds one, the cumulant
expansion fails and the FGR approximation is no longer valid. In the
regime $\sigma\gtrsim 1$, the decay rate of the function ${\cal
F}(t; {\overset{\circ}\alpha}_c)=\big|f(t;
{\overset{\circ}\alpha_c})\big|^2$ ceases to depend on $\sigma$
\cite{Sagde88} and coincides with the decay rate $1/\tau_c$ of
the classical correlation function (\ref{ClasCorrDecay}),
\begin{equation}\label{Ldecay}
{\cal F}(t; {\overset{\circ}\alpha}_c)=\exp(-t/\tau_c)\,.
\end{equation}
This rate is intimately related to the local instability of the chaotic
classical motion though it is not necessarily given by the Lyapunov
exponent $\Lambda$ itself. Quantum interference {\it does not show
up} at all.
\subsection{Ballistic electron quantum transport in the presence of weakly disordered background.}
\label{sec:BallisticElTransport}
Finally, we will discuss the decoherence phenomenon in the
electron transport through an open ballistic quantum dot as this
problem is seen from the point of view of the general resonance
scattering theory \cite{Sokol10}. Peculiarities of this transport
reflect the properties of eigenstates of the quantum billiards, whose
spectra are highly nontrivial in the classicaly chaotic regimes.
The loss of coherence that is the main problem of our concern is
attributed below to interaction with a weakly disordered many-body
environment (``walls''). So, the whole our system consists of an open
cavity and walls and is described by the following non-Hermitian
effective Hamiltonian
\begin{equation}
{\cal \hat H}=
\left(
\begin{array}{cc}
{\cal H}^{(s)} & V^{\dag}\\
V & H^{(e)}\\
\end{array}%
\right).
\end{equation}
The upper left block stands for the non-Hermitian effective
Hamiltonian of the irregularly shaped cavity (dot) with two similar
leads supporting each $M/2$ equivalent channels.
\begin{equation}
{\cal H}^{(s)} = H^{(s)}-\frac{i}{2}A A^{\dag}
\end{equation}
This non-Hermitian effective Hamiltonian describes a set of
$N^{(s)}$ electron doorway resonance states with complex
eigenenergies ${\cal E}_n=E_n-\frac{i}{2}\Gamma_n$ separated
by mean level spacing $D$.
The Hermitian matrix $H^{(e)}$ represents the environment with a
very dense discrete spectrum of $N^{(e)}$ ($>>> N^{(s)}$) real
energy levels $\epsilon_{e}$ (mean level spacing
$\delta<<<D$). These states get excess to the continuum
only due to the coupling $V$ to the doorway states in the cavity.
We exploit further the single particle approximation in the
environmental sector: $H^{(e)}\Rightarrow H^{(e)}_{sp}$. The
mean level spacing of a quasi-electron $d\propto 1/N^{(e)}_{sp}$
is much greater that the many-body spacing $\delta$ but still much
smaller than the doorway spacing, $ \delta\ll d\ll D$. The
interaction $V$ with to irregular impurities is
described by a rectangular $N^{(e)}_{sp}\times N^{(s)}$
matrix with random matrix elements
\begin{equation}\label{av_V}
\langle V_{\nu n}\rangle=0, \quad
\langle V_{\mu m}^* V_{\nu n}\rangle =
\frac{1}{2} \Gamma_s \frac{d}{\pi}\delta_{\mu\nu}\delta_{m n}
\end{equation}
The second relation defines the spreading width
\begin{equation}\label{Gamma_s}
\Gamma_s=2\pi\frac{\langle | V|^2\rangle}{d}
\end{equation}
that satisfies the condition ${\Gamma_s}\gg d$, so that the influence
of the disorder lies beyond validity of the standard perturbation
theory.
The unitary $M\times M$ scattering matrix has the form
\begin{equation}
S(E)=I -i {\cal T}(E)=I-i {A^{\dag}}{\cal G}_D(E)A.
\end{equation}
Therefore the evolution of a scattered electron inside the cavity is
described by the $(N^{(s)}\times N^{(s)})$ doorway resolvent
(doorway propagator)
\begin{equation}\label{D_way_prop}
{\cal G}_D(E)=\frac{ I}{E-{\cal H}^{(s)}-\Sigma(E)}.
\end{equation}
Here the Hermitian $N^{(s)}\times N^{(s)}$-matrix
\begin{equation}\label{Sigma}
\Sigma(E)= V^{\dag}\frac{I}{E - H_{sp}^{(e)}}V=V^{\dag}G_{sp}^{(e)}(E)V
\end{equation}
accounts for transitions cavity $\leftrightarrow$ environment. Being
averaged over the random coupling amplitudes $V$, this matrix is,
with accuracy $1/N^{(e)}_{sp}$, diagonal and is proportional to
the trace in the single-quasi-particle space.
\begin{equation}\label{Sigma_av}
\Sigma(E)\Rightarrow\frac{1}{2}{\Gamma_s}\,g(E);\quad g(E)=\frac{d}{\pi}Tr G_{sp}^{(e)}(E).
\end{equation}
As a result a given doorway resonance is fragmented in a large
number of narrow resonances whose complex energies are found by
solving the equation
\begin{equation}\label{Fine_str}
{\cal E}_{\nu}^n -{\cal E}_n-\frac{1}{2}\Gamma_s
g({\cal E}_{\nu}^n)=0,
\end{equation}
so that $\Gamma_s/d$ fine structure resonances originate from any
given doorway doorway state.
The transition matrix transforms now to
\begin{equation}\label{T_exact}
{\cal T}^{ab}(E)=\sum_n
\frac{{\cal A}_n^a {\cal A}_n^b}{E-{\cal E}_n
-\frac{1}{2}{\Gamma_s} g(E)}
=\sum_{\nu}
\frac{\tilde{\cal A}_{\nu}^a \tilde{\cal A}_{\nu}^b}{E-
{\cal E}_{\nu}}.
\end{equation}
The resulting transition amplitudes are now sums of interfering
contributions of all narrow fine-structure resonances. The new pole
residues are complex and, therefore, interfere! \textit{No loss of
coherence on this stage!}\\
\subsubsection{Time delay.}
\label{sec:Timedelay}
The resonant Smith time delay matrix $Q=-iS^{\dag}dS/dE$ can be expressed
\cite{Sokol97} in terms of the vectors $b^{(s,e)}$ of the intrinsic part of
the total scattering wave function. A straightforward calculation gives $Q\!=\!{b^{(s)}}^{\dag}b^{(s)}+{b^{(e)}}^{\dag}b^{(e)} \!=\!Q^{(s)} + Q^{(e)}$ where the vectors
\begin{equation}\label{bb}
\quad b^{(s)}(E)={\cal G}^{(s)}_{\textsc{d}}(E)A^{(s)}\,;
\quad b^{(e)}(E)=G_{sp}^{(e)}(E)Vb^{(s)}(E)
\end{equation}
have dimensions $N^{(s)}\times M$ and $N_{sp}^{(e)}\times M$
respectively. The two contributions $Q^{(s,e)}$ correspond to the modified due to the
interaction with the background time delay within the dot and delay because of the
virtual transitions into the background. In particular, the diagonal elements of the
resonant Smith matrix give the norms of the internal parts of the scattering wave
function initiated in specific channels. Finally, the typical scattering duration called
the Wigner delay time is represented by the following weighted-mean quantity
\begin{equation}\label{Wdt}
\tau_W(E)=\frac{1}{M}{\rm Tr}\, Q(E)\,.
\end{equation}
After averaging over the interaction $V$ we arrive in the main approximation to
\begin{equation}\label{Lam}
Q(E)=\Lambda(E)\,{b^{(s)}}^{\dag}b^{(s)}=\Lambda(E)
Q^{(s)}(E)\,;\quad \Lambda(E)=1+\frac{1}{2}\Gamma_s
l^{(e)}\;(E)
\end{equation}
where $l(E)$=$\frac{d}{\pi}$$\rm Tr$ $\left[{G_{sp}^{(e)}}^{\dag}(E) G_{sp}^{(e)}(E)\right]$.
In fact, however, the spectrum of the fine-structure resonances is
extremely dense so that this structure cannot be resolved experimentally. Only
quantities averaged over some energy interval $d\ll\Delta E\ll D$ are observed.
\begin{equation}\label{Csecn, Tdelay_av}
\begin{array}{c}
\overline{\sigma^{ab}(E)}=
\frac{1}{\Delta E}\int_{E-\frac{1}{2}\Delta E}^{E+\frac{1}{2}\Delta E}
dE'\,\big|{\cal T}^{ab}(E')\big|^2\,;\\
\overline{\tau_W(E)}=
\frac{1}{\Delta E}\int_{E-\frac{1}{2}\Delta E}^{E+\frac{1}{2}\Delta E}
dE'\,\tau_W(E')\,.
\end{array}
\end{equation}
Neglecting unobservable spectral fluctuations of the background we assume a rigid
spectrum with equidistant levels $\epsilon_{\mu}\!=\!\varepsilon_0\!+\!\mu d\!$
(the {\it picket fence} approximation). One of the advantages of this uniform model
is that the loop functions $g(E), l(E)$ can be calculated explicitly \cite{Sokol97}
\begin{equation}\label{g,l}
g(E)=\cot\left(\frac{\pi E}{d}\right)\,,
\quad l(E)=\frac{\pi}{d}\sin^{-2}\left(\frac{\pi E}{d}\right)\,.
\end{equation}
\subsubsection{Isolated doorway resonance near the Fermi energy.}
\label{sec:IsolatedDoorwayRes}
If the incoming electron with the energy $E\approx E_{res}$ excites
an isolated ($\Gamma=\sum_c\Gamma^c\ll D$) resonance state with
the energy $E_{res}\approx 0$ very close to the Fermi surface in
the environment, the transition cross section turns out to be equal to
\begin{equation}\label{TrCrossSec}
\sigma^{ab}(E)=\Big|{\cal T}^{ab}(E)\Big|^2=
\frac{\Gamma^a\Gamma^b}{\left[E-
\frac{1}{2}\Gamma_s\cot\left(\frac{\pi E}{d}\right)\right]^2+
\frac{1}{4}\Gamma^2}\,,
\end{equation}
when the Wigner time delay looks in this case as
\begin{equation}\label{Q1}
\tau_W(E)=\Gamma\,\frac{1+\frac{\pi\Gamma_s}{2d}\sin^{-2}\left(\frac{\pi
E}{d}\right)} {\left[E- \frac{1}{2}\Gamma_s\cot\left(\frac{\pi E}{d}\right)\right]^2+ \frac{1}{4}\Gamma^2}\,.
\end{equation}
The both quantities reveal strong fine-structure fluctuations on the typical scale
of the background level spacing $d$.
The fine-scale energy averaging yields then
\begin{equation}\label{AvTrCrossSec}
\overline{\sigma^{ab}(E)}=\frac{\Gamma^a\Gamma^b}{E^2+
\frac{1}{4}\left(\Gamma+\Gamma_s\right)^2}+
\frac{\Gamma^a\Gamma^b}{\Gamma}\frac{\Gamma_s}{E^2+
\frac{1}{4}\left(\Gamma+\Gamma_s\right)^2}=\left(1+\frac{\Gamma_s}{\Gamma}\right)
\frac{\Gamma^a\Gamma^b}{E^2+\frac{1}{4}\left(\Gamma+\Gamma_s\right)^2}\,
\,.
\end{equation}
The averaging destroyed the coherence and decomposed the cross section into a sum
of two incoherent contributions. The first of them corresponds to excitation and
subsequent decay of the doorway resonance (widened because of leaking into the
environment) through one of the $M$ outer channels. The leakage effect is described by
additional shift in the upper part of the complex energy plane by the
distance $\frac{1}{2}\Gamma_s$. The second term accounts for the
particles re-injected back in the cavity from the background. There is no net
loss of the electrons. All of them escape finally via outer channels.
The environment looks from outside as a black box which swallows a particle and
spits it back in the cavity after some time. This time is characterized by the mean
Wigner time delay that also consists of two contributions,
\begin{equation}\label{av_t_delay}
\overline{\tau_W}(E)=\frac{\Gamma+\Gamma_s}{E^2
+\frac{1}{4}(\Gamma+\Gamma_s)^2}+\frac{2\pi}{d}\;.
\end{equation}
The first term describes the delay on the damped by the internal ``friction''
resonance level inside the dot when the second one accounts for the electrons
delayed in the environment by the time $\tau_d=\frac{2\pi}{d}$ proportional to
the quasi-particle's level density.
The conductivity of the device is proportional to the transport cross section
\begin{eqnarray}\label{Transp1}
G(E)&=&\sum_{a\in 1,b\in 2}\overline{\sigma^{a b}(E)}=
\frac{\Gamma_1\,\Gamma_2}{\Lambda(E)}+
\frac{\Gamma_ 1\,\Gamma_2}{\Gamma_1+\Gamma_2}\,
\frac{\Gamma_s}{\Lambda(E)}\\
&=&T_{1 2}+\frac{T_{1 s}\,T_{s 2}}{T_{1 s}+T_{s 2}}
\end{eqnarray}
where $\Gamma_k=\sum_{c\in k}\Gamma^c$, $k=1,2$;
$\Gamma_1+\Gamma_2=\Gamma$ and
\begin{equation}\label{SubTrProb1}
T_{s k}(E)=\frac{\Gamma_s\,\Gamma_k}{\Lambda(E)}\,,\quad\,\,\,\,
\Lambda(E)=E^2+\frac{1}{4}\left(\Gamma+\Gamma_s\right)^2.
\end{equation}
The term $T_{1 2}$ describes transition from the first to the second
lead via the broadened intermediate doorway resonance when the
additional contribution incorporates the interchanges with the
environment. The latter can be naturally interpreted by introducing
an additional fictitious $(M+1)th$ channel that connects the
resonance state with the environment. The corresponding extended
scattering matrix remains unitary.
The found expression is formally identical to that obtained within the
framework of the B\"{u}ttiker's voltage-probe model \cite{Buetti86,Buetti88}
of the decoherence phenomenon. The corresponding dimensionless decoherence
rate equals in our case to $\gamma_s=\frac{2\pi}{D}\Gamma_s=\Gamma_s\tau_D$.
\subsubsection{Many-body effects.}
\label{sec:ManyBodyeff}
The single-particle approximation used up to now is well justified
only when the scattering energy $E$ is very close to the Fermi
surface in the environment. For higher scattering energies,
many-body effects should be taken into account. They show up, in
particular, in a finite lifetime of the quasi-electron with the energy
$E>E_F=0$. The simplest way to account for this effect is to
attribute some imaginary part to the quasi-particle's energy,
$\varepsilon_{\mu}=\mu d - \frac{i}{2}\Gamma_e$. The resonant
denominator looks then as \cite{Sokol97}
\begin{equation}\label{AbsProp}
{\cal D}_{res}(E)=E-E_{res}
-\frac{1}{2}\Gamma_s(1-\xi^2)\frac{\eta}{1+\xi^2\eta^2}+
\frac{i}{2}\left(\Gamma+\Gamma_s\xi\frac{1+\eta^2}{1+
\xi^2\eta^2}\right)
\end{equation}
where $E_{res}$ is the position of the doorway resonance and
the following notations
$$\xi=\tanh\left(\frac{\pi\Gamma_e}{2 d}\right),
\quad \eta=\cot\left(\frac{\pi E}{d}\right)$$
has been used.
The transport cross section $G(E)$ retains still its
form (\ref{Transp1}) but the subsidiary transition
probabilities looks now as
\begin{equation}\label{SubTrProb1Abs}
T_{sk}(E)\Rightarrow T_{sk}(E;\kappa)=
\frac{\Gamma_s\,\Gamma_k}{\Lambda(E;\kappa)}
\end{equation}
instead of (\ref{SubTrProb1}). The factor
\begin{equation}\label{eq4}
\frac{1}{\Lambda(E;\kappa)}=\frac{1}{\Lambda(E)}\,\frac{1}{1+
\kappa\frac{\Lambda(E)}{\Gamma\Gamma_s}},\quad \Lambda(E)=(E-E_{res})^2+\frac{1}{4}(\Gamma+\Gamma_s)^2
\end{equation}
depends on the new parameter $\kappa$ that accounts for
inelastic effects in the background,
\begin{equation}\label{kappa}
\begin{array}{c}
\kappa=\frac{4\xi}{(1-\xi)^2}=e^{\gamma_e}-1\approx\\
\left\{\begin{array}{l}
\gamma_e\ll 1,\,\,\,\,\,\,\,\textrm{if}\,\,\,\,\,\,\,\,\,\,{\tau}_e\gg\tau_d\,,\\
e^{\gamma_e}\gg 1,\,\,\,\,\, \textrm{if}\,\,\,\,\,\,\,\,\,\,\tau_e<\tau_d\,,\\
\end{array}\right.
\quad(\gamma_e=\tau_d\Gamma_e)\,
\end{array}
\end{equation}
where ${\tau}_e=1/\Gamma_e$ is the lifetime of the quasi-electron
in the environment.
Strictly speaking, the assumed quasi-electron decay, that implies
infinite density of the final states in the background, seems to
destroys the unitarity of the scattering matrix in contradiction with
what has been stated before. In fact, a single-particle state once
excited in the environment with a very dense but, nevertheless,
discrete spectrum evolves after that quite similar to a quasi-stationary
state only till the time $2\pi/\delta\gg \tau_e=1/\Gamma_e$.
After this time, recovery of the initial non-stationary state begins.
An electron that carry in particular electric charge preserves to a certain
extent its individuality in the environment. It can lose, because of the
many-body effects, only a part of its energy but not the charge and inevitably
returns sooner or later in the cavity and escapes finally via one of the
outer channels. There exists a good probability for the electron to be re-emitted
in the cavity with some intermediate energy $E_{out}<E_{in}\approx E_{res}$ within
the much shorter time interval $\tau_d=\frac{2\pi}{d}$. The portion of energy lost
by such a retarded electron dissipates inside the environment. As a result, the
background temperature jumps slightly up during each act of the scattering. However,
supposing that the environment system is bulky enough, we can disregard this
very slow increase of the environment temperature. Alternatively, we can suppose
that a special cooling technique is in use.
Near the doorway resonance energy $E_{res}$ the influence of the
finite lifetime effects is negligible within the range
$0\leqslant\kappa\lesssim\kappa_c=
\frac{4\Gamma\Gamma_s}{(\Gamma+\Gamma_s)^2}$. The critical
value $\kappa_c$ reaches it's maximum possible, $\kappa_c=1$,
when $\Gamma=\Gamma_s$ and becomes small if one out of the two
widths noticeably exceeds another. In these cases the interval of
weak absorption is very restricted and the absorption begins to play
an important role. If the resonance is so narrow that
$\Gamma\ll\Gamma_s$, then $\kappa_c\approx
4\frac{\Gamma}{\Gamma_s}\ll 1$ and the subsidiary probabilities
(\ref{SubTrProb1Abs}) at the resonance energy $E=E_{res}$
and $\kappa\gtrsim\kappa_c$ are small,
$T_{sk}(E=E_{res};\kappa)\approx\frac{16}{\kappa}
\frac{\Gamma\Gamma_k}{\Gamma_s^2}\lesssim
\frac{16}{\kappa_c}
\frac{\Gamma\Gamma_k}{\Gamma_s^2}\approx
4\frac{\Gamma_k}{\Gamma_s}\ll 1$.
On the other hand, the very quasi-particle concept is self-consistent
only if $\gamma_e=\tau_d\Gamma_e\lesssim 1$ so that the
physically feasible interval of the strong absorption regime is
$\kappa_c\approx
4\frac{\Gamma}{\Gamma_s}\lesssim\kappa\lesssim 1$.
In this interval, the subsidiary transition probabilities (\ref{SubTrProb1Abs})
are much smaller than the direct transport probability $T_{1 2}$ thus signifying
suppression of quantum coherence, the degree of suppression being quantitatively
characterized by the decoherence rate $\gamma_s\equiv\frac{2\pi}{D}\Gamma_s$.
Another possible method of accounting for the absorbtion has been proposed
in \cite{Efeto95}. The absorption is modelled in this case by including in
the Hamiltonian a spatially uniform imaginary potential with the strength $-\frac{i}{2}\gamma_{\phi}$. Results of this two approaches become equivalent
with the obvious identification $\gamma_{\phi}=\gamma_s$.
\subsubsection{Overlapping doorway resonances.}
\label{sec:OverlapDoorwayRes}
A number of overlapping doorway states can be excited if the
incoming electron energy $E_{in}$ appreciably exceeds the Fermi
energy. As before, the cross sections averaged over the fine
structure scale consist of incoherent contributions of directly
scattered and penetrated into the environment and then re-emitted
particles.
\begin{equation}
\overline{\sigma^{{ab}}(E)}=
\sigma_d^{ab}(E)+\sigma_r^{ab}(E),
\end{equation}
where the direct and re-emitted contributions are
\begin{eqnarray}
\sigma_d^{ab}(E)&=&
\Big|\sum_n \frac{{\cal A}_n^a {\cal A}_n^b}{{\cal D}_n(E)}\Big|^2,\\
\sigma_r^{ab}(E)&=&\Gamma_s \int_0^\infty dt_r\,\sigma_r^{ab}(E;t_r),\\
&&\sigma_r^{ab}(E;t_r)=
\Big|\sum_n\frac{{\cal A}_n^a {\cal A}_n^b}{{\cal D}_n(E)}\,e^{-i{\cal E}_n t_r}\Big|^2,\\
&&{{\cal D}_n(E)}=E-E_n+\frac{i}{2}\left(\Gamma_n+\Gamma_s\right).
\end{eqnarray}
The particles delayed within environment for different times
contribute incoherently.
Since the electron motion in the cavity is supposed to be
classically chaotic the ensemble averaging $\langle ...\rangle$
in the doorway sector is appropriate. It is easy to see that, as
long as the inelastic effects in the background are fully neglected,
such an averaging perfectly eliminates dependence of all mean cross
sections on the spreading width. Indeed, the ensemble averaged
cross section is expressed in terms of the S-matrix two-point
correlation function
$C_V^{a b}(\varepsilon)=C_0^{a b}(\varepsilon-i\Gamma_s)$ as
\begin{equation}\label{AnsEvCrossSec_d}
\langle\sigma_d^{ab}(E)\rangle=
C_V^{a b}(0)=C_0^{a b}(-i\Gamma_s)=
\int_0^{\infty} dt\,e^{-\Gamma_s t}\,K_0^{a b}(t)\,.
\end{equation}
The subscript $V$ indicates the coupling to the background and
the function $K_0^{a b}(t)$ is the Fourier transform of the
correlation function $C_0^{a b}(\varepsilon)$. On the other
hand, it is easy to show that
\begin{equation}\label{AnsEvCrossSec_r}
\langle\sigma_r^{ab}(E;t_r)\rangle=\int_0^{\infty} dt\,
e^{-\Gamma_s t}\,K_0^{a b}(t+t_r)\,.
\end{equation}
Therefore, finally,
\begin{equation}\label{AnsEvCrossSec}
\begin{array}{c}
\langle\overline{\sigma^{ab}(E)}\rangle=
\int_0^{\infty} dt\,e^{-\Gamma_s t}\,K_0^{a b}(t)\\+
\Gamma_s\int_0^{\infty} dt_r\int_0^{\infty} dt\,e^{-\Gamma_s t}\,
K_0^{a b}(t+t_r)=
\int_0^{\infty} dt\,\,K_0^{a b}(t)=\langle\sigma_0^{ab}(E)\rangle \,.
\end{array}
\end{equation}
The ensemble averaging, being in fact equivalent to the energy
averaging over the doorway scale $D$, suppresses all interference
effects save the elastic enhancement because of the time reversal
symmetry. The latter effect manifests itself in the weak localization phenomenon. The time reversal symmetry is violated owing to the energy absorption in the environment.
\subsubsection{Energy absorption and suppression of the weak localization.}
\label{sec:EnergyAbsorp&WeakLoc}
Taking into account eq. (\ref{AnsEvCrossSec}) we rewrite the
ensemble-averaged cross sections as
$\langle\overline{\sigma^{ab}(E)}\rangle=
\langle\sigma_0^{ab}(E)\rangle+\Delta\sigma^{ab}(E;\kappa)$
where the second contribution can be reduced \cite{Sokol10}
to the following compact expression
\begin{equation}\label{Delta_sigma}
\Delta\sigma^{ab}(E;\kappa)=
-\sqrt{\frac{\kappa \Gamma_s}{4}}\int_0^{\infty}
\frac{dt}{\sqrt{-\frac{d}{dt}+\frac{\kappa}{4\Gamma_s}
\left(\frac{d}{dt}-\Gamma_s\right)^2}}\,K_0^{a b}(t)\,.
\end{equation}
Being presented in such a form, this result is equally valid for both
the orthogonal (GOE, time reversal symmetry) as well as the unitary
(GUE, no time reversal symmetry) cases.
To simplify further calculation we will consider the case of an appreciably
large number $M\gg 1$ of statistically equivalent scattering channels, all
of them with the maximal transmission coefficient $T=1$. Then, first, the
channel indices $a, b$ can be dropped. And, second, the characteristic decay
time $t_W=1/\Gamma_W=\tau_D/M$ (this time is called the {\it dwell time} when
the inverse quantity is known as the {\it Weisskopf width}) of the function
$K_0(t)$ is much shorter than the mean delay time $\tau_D=2\pi/D$.
Independently of time reversal symmetry, the function $K_0(t)$ is
real, positive definite, monotonously decreases with time $t$ and
satisfies the conditions $K_0(t<0)=0,\,\,K_0(0)=1$. This allows us to
represent this function in the form of the mean-weighted decay
exponent \cite{Sokol08b}:
\begin{equation}\label{LaplRepr}
K_0(t)=\int_0^{\infty}d\Gamma\,
e^{-\Gamma t}\,w(\Gamma),\quad
\int_0^{\infty}d\Gamma\,w(\Gamma)=
K_0(0)=1\,.
\end{equation}
Rigorously speaking, the weight functions $w(\Gamma)$ have
different forms before ($t<\tau_D$) and after ($t>\tau_D$)
the Heisenberg time $\tau_D$. However contribution of the latter
interval is as small as $e^{-M}$ \cite{Sokol08b}.
Neglecting such a contribution we obtain in any inelastic channel
\begin{equation}\label{Final_2}
\Delta\sigma(E;\kappa)=
-\sqrt{\frac{\kappa \Gamma_s}{4}}
\int_0^{\infty}
d\Gamma\,\frac{w(\Gamma)}{\Gamma}\frac{1}{\sqrt{\Gamma+
\frac{\kappa}{4\Gamma_s}(\Gamma+\Gamma_s)^2}}\,.
\end{equation}
In the strong absorption limit
$\kappa\gg
\frac{4\Gamma_s\Gamma_W}{(\Gamma_s+\Gamma_W)^2}$
the parameter $\kappa$ disappears from the found expression and
the latter reduces to
\begin{equation}\label{InfAbsCor}
\Delta\sigma(E;\kappa)\Rightarrow -\Gamma_s\int_0^{\infty}
d\Gamma\,\frac{w(\Gamma)}{\Gamma\,(\Gamma+\Gamma_s)}
=-\Gamma_s\int_0^{\infty} dt_r\int_0^{\infty} dt\,
e^{-\Gamma_s t}\,
K_0^{a b}(t+t_r)\,.
\end{equation}
According to Eqs. (\ref{AnsEvCrossSec_d}, \ref{AnsEvCrossSec})
this brings us to the result
\begin{equation}\label{Efet's_Limit}
\langle\overline{\sigma(E)}\rangle=\int_0^{\infty} dt\,
e^{-\Gamma_s t}\,K_0(t)=
\int_0^{\infty}d\Gamma\frac{w(\Gamma)}{\Gamma+\Gamma_s}=
\langle\sigma_d(E)\rangle .
\end{equation}
The averaged cross section approaches in this limit the value
$1/\Gamma_s$ {\it independently of the symmetry class} when the
spreading width $\Gamma_s$ noticeably exceeds the typical widths
contributing to the integral over $\Gamma$.
For the case of time reversal symmetry (GOE) the asymptotic
expansion \cite{Verba85} of the two-point correlation function gives
\cite{Sokol07}
\begin{equation}\label{W_TRS}
w^{(GOE)}(\Gamma)=\delta(\Gamma-\Gamma_W)-
\frac{2}{t_H}\delta'(\Gamma-\Gamma_W)
+\frac{M}{2t_H^2}\delta''(\Gamma-\Gamma_W)+...
\end{equation}
whereas for the case of absence of such a symmetry (GUE) similar
expansion result in
\begin{equation}\label{W_no_TRS}
w^{(GUE)}(\Gamma)=\delta(\Gamma-\Gamma_W)+...\,.
\end{equation}
In both these cases contributions of the omitted terms are estimated
as $O(1/(M^{-7/2}))$. With such an accuracy, the formula
(\ref{Final_2}) yields for the weak localization the expression
\begin{equation}\label{weak_loc_gen}
\begin{array}{c}
\Delta G\equiv G^{(GUE)}-G^{(GOE)}\\
= M_1M_2\left(2\frac{d}{d\mu}+
\frac{\mu}{2}\frac{d^2}{d\mu^2}\right)
\left\{\frac{1}{\mu}\left[1-\frac{\sqrt{\frac{\kappa\gamma_s}{4}}}
{\sqrt{\mu+\frac{\kappa}{4\gamma_s}(\mu+\gamma_s)^2}}\right]\right\}\Big|_{\mu=M}\\
\end{array}
\end{equation}
which is valid for arbitrary values of the parameters $\kappa$,
$\gamma_s$ and $M$. The unfolded explicit expression is a bit too
lengthy. Visualization of this result is presented in
Fig.~\ref{fig:WeakLoc} for two different values of the
(dimensionless) spreading width $\gamma_s$ and $M_1=M_2=2;
M=4$. The reduction of the difference $\Delta G$ displays
suppression of the quantum coherence. Note that the effect
becomes more pronounced as the number of channels decreases.
\begin{figure}
\begin{center}
\includegraphics[width=80mm,
keepaspectratio=true]{weakLoc.eps}
\caption{Weak localization versus absorption parameter $\kappa$.
Lines correspond to $\gamma_s=25$ (blue) and
$\gamma_s=64$ (red); $M=4$.}
\label{fig:WeakLoc}
\end{center}
\end{figure}
\section{Acknowledgments}
We are very much obliged to Giuliano Benenti, Giulio Casati, and Yaroslav Kharkov
with whom we had the advantage of cooperation over a period of years. V.V.S. is
especially grateful to Vladimir Zelevinsky for long lasting friendship and collaboration.
This work is supported by the Ministry of Education and Science of the
Russian Federation (contract 14.B37.21.8408). Also we greatly appreciate
countenance by the RAS Joint scientific program "Nonlinear dynamics
and Solitons".
\bibliographystyle{abbrv}
|
\section{Introduction}
Mapping a problem involving discrete variables into continuous variables often results in a more tractable formulation.
For the case of probabilistic inference, in this paper we present a new approach to sample from distributions over binary variables, based on mapping
the original discrete distribution into a continuous one with a piecewise quadratic log-likelihood, from which we can sample efficiently using
exact Hamiltonian Monte Carlo (HMC).
The HMC method is a Markov Chain Monte Carlo algorithm that usually has better performance over Metropolis or Gibbs samplers,
because it manages to propose transitions in the Markov chain which lie far apart in the sampling space, while maintaining a
reasonable acceptance rate for these proposals. But the implementations of HMC algorithms generally involve the non-trivial tuning
of numerical integration parameters to obtain such a reasonable acceptance rate (see~\cite{neal2011mcmc} for a review).
The algorithms we present in this work are special because the Hamiltonian equations of motion can be integrated exactly, so there is no
need for tuning a step-size parameter and the Markov chain always accepts the proposed moves. Similar ideas have been used recently to sample
from truncated Gaussian multivariate distributions~\cite{pakman2013exact}, allowing much faster sampling than other methods.
It should be emphasized that despite the apparent complexity of deriving the new algorithms,
their implementation is very simple.
Since the method we present transforms a binary sampling problem into a continuous one, it is natural to extend it to distributions
defined over mixtures of binary and Gaussian or exponential variables, transforming them into purely continuous distributions.
Such a mixed binary-continuous problem arises in Bayesian model selection with a spike-and-slab prior
and we illustrate our technique by focusing on this case. In particular, we show how to sample from the posterior of linear and probit regression models
with spike-and-slab priors, while also imposing truncations in the parameter space (e.g., positivity).
The method we use to map binary to continuous variables consists in simply identifying a binary variable with the sign of a continuous one.
An alternative relaxation of binary to continuous variables, known in statistical physics as the ``Gaussian integral trick''~\cite{hertz1991introduction},
has been used recently to apply HMC methods to binary distributions~\cite{zhang2012continuous}, but the details of
that method are different than ours. In particular, the HMC in that work is not `exact' in the sense used above and the algorithm only works for
Markov random fields with Gaussian potentials.
\section{Binary distributions}
We are interested in sampling from a probability distribution $p( {\bf s} )$ defined over $d$-dimensional binary vectors $ {\bf s} \in \{-1,+1\}^{d}$,
and given in terms of a function $f( {\bf s} )$ as
\begin{eqnarray}
p( {\bf s} ) = \frac{1}{Z}f( {\bf s} ) \,.
\end{eqnarray}
Here $Z$ is a normalization factor, whose value will not be needed.
Let us augment the distribution~$p( {\bf s} )$ with continuous variables $ {\bf y} \in \mathbb{R}^{d}$ as
\begin{eqnarray}
p( {\bf s} , {\bf y} ) = p( {\bf s} )p( {\bf y} | {\bf s} )
\label{psy}
\end{eqnarray}
where $p( {\bf y} | {\bf s} )$ is non-zero only in the orthant defined by
\begin{eqnarray}
s_i = sign(y_i) \qquad \qquad i =1, \ldots, d.
\label{sy}
\end{eqnarray}
The essence of the proposed method is that we can sample from $p( {\bf s} )$ by sampling $ {\bf y} $ from
\begin{eqnarray}
p( {\bf y} ) &=& \sum_{ {\bf s} '} p( {\bf s} ') p( {\bf y} | {\bf s} ') \,,
\label{ysum}
\\
&=& p( {\bf s} ) p( {\bf y} | {\bf s} ) \,,
\label{ynosum}
\end{eqnarray}
and reading out the values of $ {\bf s} $ from (\ref{sy}).
In the second line we have made explicit that for each $ {\bf y} ,$ only one term in the sum in~(\ref{ysum}) is non-zero,
so that $p( {\bf y} )$ is piecewise defined in each orthant.
In order to sample from $p( {\bf y} )$ using the exact HMC method of~\cite{pakman2013exact},
we require $\log p( {\bf y} | {\bf s} )$ to be a quadratic function of $ {\bf y} $ on its support.
The idea is to define a potential energy function
\begin{eqnarray}
U( {\bf y} ) = - \log p( {\bf y} | {\bf s} ) - \log f( {\bf s} ) \,,
\end{eqnarray}
introduce momentum variables $q_i$, and consider the piecewise continuous Hamiltonian
\begin{eqnarray}
H( {\bf y} , {\bf q} ) &= U( {\bf y} ) + \frac{ {\bf q} \cdot {\bf q} }{2} \,,
\label{Hh}
\end{eqnarray}
whose value is identified with the energy of a particle moving in a $d$-dimensional space.
Suppose the particle has initial coordinates $ {\bf y} (0)$. In each iteration of the sampling algorithm, we sample initial values $ {\bf q} (0)$ for the momenta from a standard Gaussian distribution
and let the particle move during a time $T$ according to the equations of motion
\begin{eqnarray}
\dot{ {\bf y} }(t) = \frac{\partial H}{\partial {\bf q} (t)} \,, \qquad \quad \dot{ {\bf q} }(t) = - \frac{\partial H}{\partial {\bf y} (t)} \,.
\label{eom}
\end{eqnarray}
The final coordinates, $ {\bf y} (T)$, belong to a Markov chain with invariant distribution $p( {\bf y} )$, and are used as the initial coordinates of the next iteration.
The detailed balance condition follows directly from the conservation of energy and $( {\bf y} , {\bf q} )$-volume along the trajectory dictated by (\ref{eom}),
see~\cite{neal2011mcmc,pakman2013exact} for details.
The restriction to quadratic functions of $ {\bf y} $ in $\log p( {\bf y} | {\bf s} )$ allows us to solve the differential equations~(\ref{eom}) exactly in each orthant.
As the particle moves, the potential energy $U( {\bf y} )$ and the kinetic energy $\frac{ {\bf q} \cdot {\bf q} }{2}$ change in tandem, keeping
the value of the Hamiltonian (\ref{Hh}) constant. But this smooth interchange gets interrupted when any coordinate reaches zero.
Suppose this first happens at time~$t_j$ for coordinate $y_j$, and assume that
$y_j < 0$ for $t < t_j$.
Conservation of energy imposes now a jump on the momentum $q_j$ as a result of the discontinuity in $U( {\bf y} )$.
Let us call $q_j(t_j^-)$ and~$q_j(t_j^+)$ the value of the momentum $q_j$ just before and after the coordinate hits $y_j=0$.
In order to enforce conservation of energy, we equate the Hamiltonian at both sides of the $y_j=0$ wall, giving
\begin{eqnarray}
\frac{q_j^2(t_j^+)}{2} = \Delta_j + \frac{q_j^2(t_j^-)}{2}
\label{kin}
\end{eqnarray}
with
\begin{eqnarray}
\Delta_j = U( y_j=0, s_{j}=-1) - U( y_j=0, s_{j}=+1)
\label{deltaj}
\end{eqnarray}
If eq. (\ref{kin}) gives a positive value for $q_j^2(t_j^+)$, the coordinate $y_j$ crosses the boundary and continues its trajectory in the new orthant.
On the other hand, if eq.(\ref{kin}) gives a negative value for~$q_j^2(t_j^+)$, the particle is reflected from the $y_j=0$ wall
and continues its trajectory with $q_j(t_j^+) = - q_j(t_j^-)$.
When $\Delta_j < 0$, the situation can be understood as the limit of a scenario in which the particle faces an upward hill in the potential energy, causing it to diminish its velocity
until it either reaches the top of the hill with a lower velocity or stops and then reverses. In the limit in which the hill has finite height but infinite slope,
the velocity change occurs discontinuously at one instant. Note that we used in (\ref{kin}) that the momenta $q_{i\neq j}$ are continuous, since this sudden
infinite slope hill is only seen by the $y_j$ coordinate.
Regardless of whether the particle bounces or crosses the $y_j=0$ wall, the other coordinates move unperturbed until the next boundary hit,
where a similar crossing or reflection occurs, and so on, until the final position $ {\bf y} (T)$.
The framework we presented above is very general and in order to implement a particular sampler we need to
select the distributions~$p( {\bf y} | {\bf s} )$. Below we consider in some detail two particularly simple choices that illustrate the diversity of options here.
\subsection{Gaussian augmentation}
Let us consider first for $p( {\bf y} | {\bf s} )$ the truncated Gaussians
\begin{eqnarray}
p( {\bf y} | {\bf s} ) =
\left\{
\begin{array}{ll}
(2/\pi)^{d/2}\, e^{- \frac{ {\bf y} \cdot {\bf y} }{2}} & \textrm{for} \,\, sign(y_i)=s_i, \qquad i =1, \ldots, d
\\
0 & \textrm{otherwise} \,,
\end{array}
\right.
\label{pys}
\end{eqnarray}
The equations of motion (\ref{eom}) lead to $\ddot{ {\bf y} }(t) = - {\bf y} (t), \ddot{ {\bf q} }(t) = - {\bf q} (t)$, and have a solution
\begin{eqnarray}
y_i(t) &=& y_i(0)\cos(t) + q_i(0) \sin(t) \,,
\label{yt}
\\
&=& u_i \sin(\phi_i + t) \,,
\label{yt2}
\\
q_i(t) &=& -y_i(0)\sin(t) + q_i(0) \cos(t) \,,
\\
&=& u_i \cos(\phi_i + t) \,.
\end{eqnarray}
This setting is similar to the case studied in~\cite{pakman2013exact} and from $\phi_i = \tan^{-1}( y_i(0)/q_i(0) )$ the boundary hit times $t_i$ are easily obtained.
When a boundary is reached, say $y_j=0$, the coordinate $y_j$ changes its trajectory for $t > t_j$ as
\begin{eqnarray}
y_j(t) &=& q_j(t_j^+) \sin(t-t_j) \,,
\label{yj}
\end{eqnarray}
with the value of $q_j(t_j^+)$ obtained as described above.
Choosing an appropriate value for the travel time $T$ is crucial when using HMC algorithms~\cite{hoffman2011no}.
As is clear from (\ref{yt2}), if we let the particle travel during a time $T>\pi$, each coordinate reaches zero at least once, and the hitting times can be ordered as
\begin{eqnarray}
0 < t_{j_1} \leq t_{j_2} \leq \cdots \leq t_{j_d} < \pi \,.
\label{cyc}
\end{eqnarray}
Moreover, regardless of whether a coordinate crosses zero or gets reflected, it follows from (\ref{yj}) that the successive hits occur at
\begin{eqnarray}
t_{i} + n\pi, \quad n=1,2,\ldots
\end{eqnarray}
and therefore the hitting times only need to be computed once for each coordinate in every iteration.
If we let the particle move during a time $T=n\pi$, each coordinate reaches zero $n$ times, in the cyclical order (\ref{cyc}),
with a computational cost of $O(nd)$ from wall hits. But choosing \emph{precisely} $T=n\pi$ is not recommended
for the following reason. As we just showed, between $y_j(0)$ and $y_j(\pi)$ the coordinate touches the boundary $y_j=0$ once,
and if $y_j$ gets reflected off the boundary, it is easy to check that we have $y_j(\pi) = y_j(0)$.
If we take $T=n\pi$ and the particle gets reflected all the $n$ times it hits the boundary,
we get $y_j(T) = y_j(0)$ and the coordinate $y_j$ does not move at all. To avoid these singular situations, a good choice is~$T=(n + \frac12) \pi$,
which generalizes the recommended travel time $T=\pi/2$ for truncated Gaussians in~\cite{pakman2013exact}.
The value of $n$ should be chosen for each distribution, but we expect optimal values for $n$ to grow with $d$.
With $T=(n + \frac12) \pi$, the total cost of each sample is $O((n+1/2)d)$ on average from wall hits,
plus $O(d)$ from the sampling of $ {\bf q} (0)$ and from the $d$ inverse trigonometric functions to obtain the hit times $t_i$.
But in complex distributions, the computational cost is dominated by the the evaluation of~$\Delta_i$ in (\ref{deltaj}) at each wall hit.
Interestingly, the rate at which wall $y_i=0$ is crossed coincides with the acceptance rate in a Metropolis algorithm that
samples uniformly a value for $i$ and makes a proposal of flipping the binary variable~$s_i$.
See the Appendix for details.
Of course, this does not mean that the HMC algorithm is the same as Metropolis, because in HMC the order in which the walls are hit is fixed given the initial velocity,
and the values of $q^2_i$ at successive hits of $y_i=0$ within the same iteration are not independent.
\subsection{Exponential and other augmentations}
Another distribution that allows one an exact solution of the equations of motion (\ref{eom}) is
\begin{eqnarray}
p( {\bf y} | {\bf s} ) = \left\{
\begin{array}{ll}
e^{- \sum_{i=1}^d |y_i|} & \textrm{for} \,\, sign(y_i)=s_i, \qquad i =1, \ldots, d
\\
0 & \textrm{otherwise} \,,
\end{array}
\right.
\label{pyse}
\end{eqnarray}
which leads to the equations of motion $\ddot{y_i}(t) = -s_i$, with solutions of the form
\begin{eqnarray}
y_i(t) &=& y_i(0) + q_i(0)t -\frac{s_it^2}{2} \,.
\end{eqnarray}
In this case, the initial hit time for every coordinate is the solution of the quadratic equation~$y_i(t)=0$.
But, unlike the case of the Gaussian augmentation, the order of successive hits is not fixed. Indeed, if coordinate $y_j$ hits zero at time $t_j$,
it continues its trajectory as
\begin{eqnarray}
y_j(t> t_j) = q(t_j^+)(t-t_j) - \frac{s_j}{2}(t-t_j)^2 \,,
\end{eqnarray}
so the next wall hit $y_j=0$ will occur at a time $t'_j$ given by
\begin{eqnarray}
(t'_j-t_j) = 2|q_j(t_j^+)| \,,
\label{nextt}
\end{eqnarray}
where we used $s_j = sign(q_j(t_j^+))$. So we see that the time between successive hits of the same coordinate depends only on its momentum
after the last hit. Moreover, since the value of $|q_j(t^+)|$ is smaller than $|q_j(t^-)|$ if the coordinate crosses to an orthant of lower probability,
equation (\ref{nextt}) implies that the particle moves away faster from areas of lower probability.
This is unlike the Gaussian augmentation, where a coordinate `waits in line' until all the other coordinates
touch their wall before hitting its wall again.
The two augmentations we considered above have only scratched the surface of interesting possibilities.
One could also define $f( {\bf y} | {\bf s} )$ as a uniform distribution in a box such that the computation of the times for wall hits
would becomes purely linear and we get a classical `billiards' dynamics.
More generally, one could consider different augmentations in different orthants
and potentially tailor the algorithm to mix faster in complex and multimodal distributions.
\section{Spike-and-slab regression with truncated parameters}
The subject of Bayesian sparse regression has seen a lot of work during the last decade.
Along with priors such as the Bayesian Lasso~\cite{park2008bayesian} and the Horsehoe~\cite{carvalho2010horseshoe},
the classic spike-and-slab prior~\cite{mitchell1988bayesian, george1993variable}
still remains very competitive, as shown by its superior performance in the recent works~\cite{mohamed2011bayesian,goodfellow2012spike, chen2012bayesian}.
But despite its successes, posterior inference remains a computational challenge for the spike-and-slab prior.
In this section we will show how the HMC binary sampler can be extended to sample from the posterior of these models.
The latter is a distribution over a set of binary and continuous variables, with the
binary variables determining whether each coefficient should be included in the model or not.
The idea is to map these indicator binary variables into continuous variables as we did above, obtaining a
distribution from which we can sample again using exact HMC methods.
Below we consider a regression model with Gaussian noise but the extension to exponential noise (or other scale-mixtures of Gaussians) is immediate.
\subsection{Linear regression}
Consider a regression problem with a log-likelihood that depends quadratically on its coefficients, such as
\begin{eqnarray}
\log p(D| {\bf w} ) = -\frac12 {\bf w} '{\bf M} {\bf w} + {\bf r}\cdot {\bf w} + const.
\end{eqnarray}
where $D$ represents the observed data. In a linear regression model ${\bf z} = X {\bf w} + {\bf \varepsilon}$, with ${\bf \varepsilon} \sim {\cal N}(0,\sigma^2)$,
we have ${\bf M} = X'X/\sigma^2$ and ${\bf r}={\bf z}'X/\sigma^2$.
We are interested in a spike-and-slab prior for the coefficients $ {\bf w} $ of the form
\begin{eqnarray}
p( {\bf w} , {\bf s} |a,\tau^2) = \prod_{i=1}^{d} p(w_i|s_i,\tau^2) p(s_i|a) \,.
\label{sns}
\end{eqnarray}
Each binary variable $s_i = \pm 1$ has a Bernoulli prior $p(s_i|a) = a^{\frac{(1+s_i)}{2}} (1-a)^{\frac{(1-s_i)}{2}} $
and determines whether the coefficient $w_i$ is included in the model. The prior for $w_i$, conditioned on $s_i$, is
\begin{eqnarray}
p(w_i|s_i,\tau^2) = \left\{
\begin{array}{ll}
\frac{1}{\sqrt{2 \pi \tau^2}} e^{- \frac{w_i^2}{2 \tau^2}} & \textrm{for} \,\, s_i=+1,
\label{slab}
\\
\\
\delta(w_i) & \textrm{for} \,\, s_i=-1
\end{array}
\right.
\end{eqnarray}
We are interested in sampling from the posterior, given by
\begin{eqnarray}
p( {\bf w} , {\bf s} |D,a,\tau^2) &\varpropto & p(D| {\bf w} ) p( {\bf w} , {\bf s} |a,\tau^2)
\\
&\varpropto &
\frac{
e^{-\frac12 {\bf w} ' {\bf M} {\bf w} + {\bf r} \cdot {\bf w} }
e^{-\frac12 {\bf w} '_{+} {\bf w} _{+} \tau^{-2} }}
{(2 \pi \tau^2)^{|\st^{{}_{+}}|/2}}
\delta( {\bf w} _{-}) a^{|\st^{{}_{+}}|} (1-a)^{|\st^{{}_{-}}|}
\label{ssz}
\\
&\varpropto&
\frac{
e^{-\frac12 {\bf w} '_{+} \left( {\bf M}_{+} + \tau^{-2} \right) {\bf w} _{+} + {\bf r}_{+}\cdot {\bf w} _{+} }}
{(2 \pi \tau^2)^{|\st^{{}_{+}}|/2}}
\delta( {\bf w} _{-}) a^{|\st^{{}_{+}}|} (1-a)^{|\st^{{}_{-}}|}
\label{sso}
\end{eqnarray}
where $\st^{{}_{+}}$ are the variables with $s_i=+1$ and $\st^{{}_{-}}$ those with $s_i=-1$. The notation~${\bf r}_{+}$,~${\bf M}_{+}$ and $ {\bf w} _{+}$ indicates a restriction
to the $\st^{{}_{+}}$ subspace and $ {\bf w} _{-}$ indicates a restriction to the $\st^{{}_{-}}$ space.
In the passage from (\ref{ssz}) to (\ref{sso}) we see that the spike-and-slab prior shrinks the dimension of the Gaussian
likelihood from $d$ to $|\st^{{}_{+}}|$.
In principle we could integrate out the weights $ {\bf w} $ and
obtain a collapsed distribution for $ {\bf s} $, but we are interested in cases in which the space of $ {\bf w} $ is truncated and therefore the integration is not feasible.
An example would be when a non-negativity constraint $w_i \geq 0$ is imposed.
In these cases, one possible approach is to sample from~(\ref{sso}) with a block Gibbs sampler over the pairs~$\{w_i,s_i\}$, as proposed in~\cite{mohamed2011bayesian}.
Here we will present an alternative method, extending the ideas of the previous section. For this, we consider a new distribution, obtained in two steps.
Firstly, we replace the delta functions in (\ref{sso}) by a factor similar to the slab (\ref{slab})
\begin{eqnarray}
\delta(w_i) \rightarrow \frac{1}{\sqrt{2 \pi \tau^2}}e^{- \frac{w_i^2}{2 \tau^2}} \qquad \qquad s_i = -1
\label{rjs}
\end{eqnarray}
The introduction of a non-singular distribution for those $w_i$'s that are excluded from the model in~(\ref{sso})
creates a Reversible Jump sampler~\cite{green1995reversible}:
the Markov chain can now keep track of all the coefficients, whether they belong or not to the model in a given state of the chain,
thus allowing them to join or leave the model along the chain in a \emph{reversible} way.
Secondly, we augment the distribution with $ {\bf y} $ variables as in (\ref{psy})-(\ref{ynosum}) and sum over $ {\bf s} $.
Using the Gaussian augmentation (\ref{pys}), this gives a distribution
\begin{eqnarray}
p( {\bf w} , {\bf y} |D,a,\tau^2) \varpropto e^{-\frac12 {\bf w} '_{+} \left( {\bf M}_{+} + \tau^{-2} \right) {\bf w} _{+} + {\bf r}_{+}\cdot {\bf w} _{+} }
e^{-\frac{ {\bf w} _{-}\cdot {\bf w} _{-} }{2 \tau^2}} e^{-\frac{ {\bf y} \cdot {\bf y} }{2}} a^{|\st^{{}_{+}}|} (1-a)^{|\st^{{}_{-}}|}
\label{ssn}
\end{eqnarray}
where the values of $ {\bf s} $ in the rhs are obtained from the signs of $ {\bf y} $.
This is a piecewise Gaussian, different in each orthant of $ {\bf y} $, and possibly truncated in the $ {\bf w} $ space.
Note that the changes in $p( {\bf w} , {\bf y} |D,a,\tau^2)$ across orthants of $ {\bf y} $ come both from the factors $a^{|\st^{{}_{+}}|} (1-a)^{|\st^{{}_{-}}|}$ and from the functional dependence on the $ {\bf w} $ variables.
Sampling from~(\ref{ssn}) gives us samples from the original distribution (\ref{sso}) using a simple rule: each pair $(w_i,y_i)$ becomes~$(w_i,s_i =+1)$ if $y_i \geq 0$ and
$(w_i=0,s_i =-1)$ if $y_i < 0$.
This undoes the steps we took to transform (\ref{sso}) into (\ref{ssn}): the identification $s_i = sign(y_i)$
takes us from $p( {\bf w} , {\bf y} |D,a,\tau^2)$ to $p( {\bf w} , {\bf s} |D,a,\tau^2)$, and setting $w_i=0$ when $s_i =-1$ undoes the replacement in (\ref{rjs}).
Since (\ref{ssn}) is a piecewise Gaussian distribution, we can sample from it again using the methods of~\cite{pakman2013exact}.
As in that work, the possible truncations for $ {\bf w} $ are given as $g_n( {\bf w} ) \geq 0$ for $n=1, \ldots, N$, with $g_n( {\bf w} )$ any product of linear and quadratic functions of $ {\bf w} $.
The details are a simple extension of the purely binary case and are not very illuminating, so we leave them for the Appendix.
\subsection{Probit regression}
Consider a probit regression model in which binary variables $b_i=\pm 1$ are observed with probability
\begin{eqnarray}
p(b_i| {\bf w} , {\bf x} _i) = \frac{1}{\sqrt{2\pi}}\int_{z_ib_i \geq 0} dz_i e^{-\frac12(z_i + {\bf x} _i {\bf w} )^2}
\end{eqnarray}
Given a set of $N$ pairs $(b_i, {\bf x} _i)$, we are interested in the posterior distribution of the weights $ {\bf w} $ using the spike-and-slab
prior~(\ref{sns}).
This posterior is the marginal over the $z_i$'s of the distribution
\begin{eqnarray}
p( {\bf z} , {\bf w} , {\bf s} | {\bf x} ,a,\tau^2) \varpropto \prod_{i=1}^N e^{-\frac12(z_i + {\bf x} _i {\bf w} )^2} p( {\bf w} , {\bf s} |a,\tau^2) \quad \quad z_ib_i \geq 0 \,,
\end{eqnarray}
and we can use the same approach as above to transform this distribution into a truncated piecewise Gaussian, defined over
the $(N+2d)$-dimensional space of the vector $( {\bf z} , {\bf w} , {\bf y} )$. Each $z_i$ is truncated according to the sign of $b_i$ and we can
also truncate the $ {\bf w} $ space if we so desire. We omit the details of the HMC algorithm, since it is very similar to the linear regression
case.
\section{Examples}
We present here three examples that illustrate the advantages of the proposed HMC algorithms over Metropolis or Gibbs samplers.
\subsection{1D Ising model}
We consider a 1D periodic Ising model, with $p( {\bf s} ) \varpropto e^{-\beta E[ {\bf s} ]}$,
where the energy is $E[ {\bf s} ] = - \sum_{i=1}^d s_{i} s_{i+1}$, with $s_{d+1}=s_{1}$ and $\beta$ is the inverse temperature. Figure~\ref{ising1d} shows the first $1000$ iterations of
both the Gaussian HMC and the Metropolis\footnote{As is well known (see e.g.\cite{newman1999monte}), for binary distributions, the Metropolis sampler that
chooses a random spin and makes a proposal of flipping its value, is more efficient than the Gibbs sampler. }
sampler on a model with $d=400$ and $\beta=0.42$, initialized with all spins $s_i=1$.
In HMC we took a travel time $T=12.5\pi$ and, for the sake of comparable
computational costs, for the Metropolis sampler we recorded the value of $ {\bf s} $ every $d \times 12.5$ flip proposals.
The plot shows clearly that the Markov chain mixes much faster with HMC than with Metropolis.
A useful variable that summarizes the behavior of the Markov chain is
the magnetization
$ m = \frac{1}{d}\sum_{i=1}^d s_i $\,,
whose expected value is $\langle m \rangle=0$. The oscillations of both samplers around this value illustrate the superiority of the HMC sampler.
In the Appendix we present a more detailed comparison of the HMC Gaussian and exponential and the Metropolis samplers, showing that the Gaussian
HMC sampler is the most efficient among the three.
\begin{figure}[t]
\begin{center}
\includegraphics[scale=0.53]{1DIsing.pdf}
\end{center}
\caption{ {\bf 1D Ising model.}
First $1000$ iterations of Gaussian HMC and Metropolis samplers on a model with $d=400$ and $\beta=0.42$, initialized with all spins $s_i=1$ (black dots).
For HMC the travel time was $T=12.5\pi$ and in the Metropolis sampler we recorded the state of the Markov chain once every $d \times 12.5$ flip proposals.
The lower two panels show the state of $ {\bf s} $ at every iteration for each sampler. The plots show clearly that the HMC model mixes faster than Metropolis in this model.}
\label{ising1d}
\end{figure}
\subsection{2D Ising model }
We consider now a 2D Ising model on a square lattice of size $L\times L$ with periodic boundary conditions below the critical temperature. Starting from a completely disordered state,
we compare the time it takes for the sampler to reach one of the two low energy states with magnetization $m \simeq \pm 1$.
Figure~\ref{ising2d} show the results of 20 simulations of such a model with $L=100$ and inverse temperature~$\beta=0.5$.
We used a Gaussian HMC with $T=2.5 \pi$ and a Metropolis sampler recording values of $ {\bf s} $ every~$ 2.5 L^2$ flip proposals.
In general we see that the HMC sampler reaches higher likelihood regions faster.
Note that these results of the 1D and 2D Ising models illustrate the advantage of the HMC
method in relation to two different time constants relevant for Markov chains~\cite{sokal1989monte}. Figure~\ref{ising1d} shows
that the HMC sampler explores faster the sampled space once the chain has reached its equilibrium distribution, while Figure~\ref{ising2d} shows that the HMC sampler
is faster in reaching the equilibrium distribution.
\begin{figure}[t]
\begin{center}
\includegraphics[scale=0.50]{2DIsing.pdf}
\end{center}
\caption{ {\bf 2D Ising model.} First samples from $20$ simulations in a 2D Ising model in a square lattice of side length $L=100$ with periodic boundary conditions
and inverse temperature~$\beta = 0.5$. The initial state is totally disordered. We do not show the first 4 samples in order to ease the visualization.
For the Gaussian HMC we used $T=2.5 \pi$ and for Metropolis we recorded
the state of the chain every $2.5 L^2 $ flip proposals. The plots illustrate that in general HMC reaches equilibrium faster than Metropolis in this model. }
\label{ising2d}
\end{figure}
\subsection{Spike-and-slab linear regression with positive coefficients}
We consider a linear regression model ${\bf z} = X {\bf w} + {\bf \varepsilon}$ with the following synthetic data. $X$ has $N=700$ rows,
each sampled from a $d=150$-dimensional Gaussian whose covariance matrix has $3$ in the diagonal and $0.3$ in the nondiagonal entries. The noise is
${\bf \varepsilon} \sim {\cal N}(0,\sigma^2=100)$. The data ${\bf z}$ is generated with a coefficients vector $ {\bf w} $, with 10 non-zero entries with values between $1$ and $10$.
The spike-and-slab hyperparameters are set to $a=0.1$ and $\tau=10$. Figure~\ref{hmc_spike} compares the results of the proposed HMC method versus the Gibbs
sampler used in~\cite{mohamed2011bayesian}. In both cases we impose a positivity constraint on the coefficients.
For the HMC sampler we use a travel time $T=\pi/2$. This results in a number of wall hits (both for $ {\bf w} $ and $ {\bf y} $ variables) of $\simeq 150$,
which makes the computational cost of every HMC and Gibbs sample similar.
The advantage of the HMC method is clear, both in exploring regions of higher probability and in the mixing speed of the sampled coefficients.
This impressive difference in the efficiency of HMC versus Gibbs is similar to the case of truncated multivariate Gaussians studied
in~\cite{pakman2013exact}.
\begin{figure}[t]
\begin{center}
\includegraphics[scale=0.56]{hmc_spike_and_slab.pdf}
\end{center}
\caption{ {\bf Spike-and-slab linear regression with constraints.}
Comparison of the proposed HMC method with the Gibbs sampler of~\cite{mohamed2011bayesian} for the posterior of a linear regression model with spike-and-slab
prior, with a positivity constraint on the coefficients. See the text for details of the synthetic data used.
Above: log-likelihood as a function of the iteration. Middle: samples of the first coefficient. Below: ACF of the first coefficient.
The plots shows clearly that HMC mixes much faster than Gibbs and is more consistent in exploring areas of high probability.
}
\label{hmc_spike}
\end{figure}
\section{Conclusions and outlook}
We have presented a novel approach to use exact HMC methods to sample from generic binary distributions and certain distributions over mixed binary and continuous variables,
Even though with the HMC algorithm is better than Metropolis or Gibbs in the examples we presented,
this will clearly not be the case in many complex binary distributions for which specialized sampling algorithms have been developed,
such as the Wolff or Swendsen-Wang algorithms for 2D Ising models near the critical temperature~\cite{newman1999monte}.
But in particularly difficult distributions, these HMC algorithms could be embedded as inner loops inside more powerful algorithms of Wang-Landau type~\cite{wang2001efficient}.
We leave the exploration of these newly-opened realms for future projects.
\subsubsection*{Acknowledgments}
This work was supported by an NSF CAREER award and by the US Army Research Laboratory and the US Army Research Office under contract number W911NF-12-1-0594.
{\bf \large Appendix}
|
\section{Introduction}
\label{S_intro_Sud}
Let $\mu,\nu \in \mathcal P(\mathbb{R}^d)$ with $\mu \ll \mathcal L^d$ and consider the Monge optimal transportation problem
\begin{equation}
\label{E_transpo_Sud}
\min \bigg\{ \int \d{\mathtt T(x) - x} d\mu(x), \ \mathtt T : \mathbb{R}^d \rightarrow \mathbb{R}^d, \ \nu = \mathtt T_\# \mu \bigg\},
\end{equation}
where $\d{\cdot}$ is a \emph{convex norm} in $\mathbb{R}^d$, namely a positively $1$-homogeneous function whose unit ball $\{x\in\mathbb{R}^d:\,\d{x}\leq1\}$ is a closed $d$-dimensional convex set $D$ with $0\in\mathrm{int} \,D$. The $\mu$-measurable maps $\mathtt T:\mathbb{R}^d\rightarrow\mathbb{R}^d$ satisfying $\mathtt T_\#\mu=\nu$ are called \emph{transport maps}.
Well known examples show that if $\mu$ is not absolutely continuous w.r.t. $\mathcal L^d$, there may be no optimal transport maps (see Theorem 8.3 of \cite{conf:optcime}).
Due to the nonlinearity of the constraint $\mathtt T_\#\mu$, the classical approach to solve \eqref{E_transpo_Sud} is first to consider the relaxed problem
of finding optimal \emph{transference plans} $\pi \in \Pi^{\mathrm{opt}}_{\d{\cdot}}(\mu,\nu)$ defined by
\begin{equation}
\label{E_ptimal_intro}
\int \d{y-x} d\bar\pi(x,y) = \min \bigg\{ \int \d{y-x} d\pi(y,x), \pi \in \Pi(\mu,\nu) \bigg\},
\end{equation}
where
\begin{equation}
\Pi(\mu,\nu) := \Big\{ \pi \in \mathcal P(\mathbb{R}^d \times \mathbb{R}^d) : (\mathtt p_1)_\# \pi = \mu, (\mathtt p_2)_\# \pi = \nu \Big\}
\end{equation}
and $\mathtt p_i : \underset{j}{\prod} X_j \rightarrow X_i$ is the projection on the $i$-coordinate in the product space $\underset{j}{\prod} X_j$.
Assuming that
\begin{equation}
\label{E_transport_prob_Sud}
\inf_{\pi \in \Pi(\mu,\nu)} \int_{\mathbb{R}^d \times \mathbb{R}^d} \d{y-x} \,d\pi(x,y) < +\infty,
\end{equation}
by standard theorems in optimal transportation there always exists an optimal transference plan, without being in the degenerate situation where every plan $\pi \in \Pi(\mu,\nu)$ is optimal.
Then, if one can show that there exists at least an optimal transport plan $\pi$ which is concentrated on a graph of a $\mu$-measurable map $\mathtt T$, i.e. $\pi:=(\mathbb{I}\times\mathtt T)_\#\mu$, then $\mathtt T$ is an optimal transport map solving \eqref{E_transpo_Sud}.
The first strategy to show the existence of such a transference plan was proposed by Sudakov in \cite{sudak} and consists in decomposing via disintegration of measures the optimal transportation problem \eqref{E_ptimal_intro} into a family of transportation problems on $Z_\mathfrak a\times\mathbb{R}^d$, where $\{Z_\mathfrak a\}_{\mathfrak a\in\mathfrak{A}} \subset \mathbb{R}^d$ are disjoint regions where the construction of an optimal map $\mathtt T_\mathfrak a : Z_\mathfrak a \rightarrow \mathbb{R}^d$ is simpler than in the original problem, and then to obtain $\mathtt T$ by piecing together the maps $\mathtt T_\mathfrak a$.\\
With additional regularity properties on the densities of $\mu$, $\nu$ or on the norm, such as uniform convexity, an approach partially equal to the one proposed by Sudakov was successfully followed in \cite{ambr:lecttrans}, \cite{conf:optcime}, \cite{caffafeldmc} and \cite{trudiwang}. The most general case where up to now this approach has been successfully implemented (see \cite{Car:strictly}) is the case in which $\d{\cdot}$ is strictly convex, namely when the set $D$ is strictly convex.
Other approaches have also been used. In \cite{evagangbo}, the problem \eqref{E_transpo_Sud} for strictly convex norms has been solved using PDE methods under the assumption that the marginals $\mu$, $\nu$ have Lipschitz continuous densities w.r.t. $\mathcal L^d$. The problem \eqref{E_transpo_Sud} was solved for crystalline norms in \cite{ambprat:crist}.
In \cite{champdepasc:MongeS, champdepasc:Monge}, the authors solved the Monge problem first with strictly convex and then with general convex norms using a different method, which does not pass through a geometric/measure theoretic decomposition of the optimal transportation problem \eqref{E_ptimal_intro} into simpler ones, but is based on the selection among the optimal transference plans $\pi \in \Pi^{\mathrm{opt}}_{\d{\cdot}}(\mu,\nu)$ of a transference plan $\check \pi$ which is also minimizing a secondary cost: more precisely, one selects the (unique) transference plan $\check \pi$ such that
\[
\check \pi \ \text{is a minimizer of} \ \inf \bigg\{ \int |x - y|^2 d\pi(x,y):\, \pi \in \Pi^\mathrm{opt}_{\d{\cdot}}(\mu,\nu) \bigg\}.
\]
and the main issue consists in proving that $\check \pi$ is actually induced by a transport map $\mathtt T$, which clearly satisfies \eqref{E_transpo_Sud}.
However, the problem of whether Sudakov's strategy could be successfully implemented also in the case of general convex norms has remained open for a long time. The aim of this paper is to show how this problem can be solved. In order to introduce the notation that we need to state our main results and explain the new ideas and concepts in the case of general convex norms, we first resume briefly how Sudakov's strategy works for strictly convex norms.
The first step of Sudakov's approach consists in finding a suitable partition in $\mathbb{R}^d$ on which the transport occurs, namely s.t. the optimal plans move the initial mass inside the elements of the partition.
By duality (see e.g. \cite{villa:Oldnew}), there exists a function $\psi:\mathbb{R}^d\rightarrow\mathbb{R}$, called \emph{Kantorovich potential}, which satisfies
\begin{align}
\psi(y)-\psi(x)&\leq\d{y-x},\quad\forall\,x,y\in\mathbb{R}^d,\label{0_psi1}\\
\psi(y)-\psi(x)&=\d{y-x},\quad\text{for $\pi$-a.e. $(x,y)$, $\forall\,\pi\in\Pi^{\mathrm{opt}}_{\d{\cdot}}(\mu,\nu)$}.\label{0_psi2}
\end{align}
Observe that, by \eqref{0_psi1}, for all $(x,y)$ as in \eqref{0_psi2} and $\forall\,0\leq s\leq t\leq1$
\begin{equation}
\label{0_ztzs}
\psi(z_t)-\psi(z_s)=\d{z_t-z_s},\quad z_t:=(1-t)x+ty.
\end{equation}
The open oriented segments $Z^1_\mathfrak a:=]x,y[\,\subset\mathbb{R}^d$ (where $\mathfrak a\in\mathfrak{A}^1$ is a continuous parameter, $1$ referring to the dimension of the elements) whose extreme points satisfy \eqref{0_psi2} and which are maximal w.r.t. set inclusion are called \emph{optimal rays}.
By strict convexity, if $(x,y)$ and $(y,z)$, with $x,z \not= y$, satisfy \eqref{0_psi2}, then
\begin{equation}
\label{0_non_branch}
y \in ]x,z[.
\end{equation}
In particular, if $(x,y)$ and $(x',y')$ satisfy \eqref{0_psi2} but $\mathbb{R}^+(y-x)\neq\mathbb{R}^+(y'-x')$, then
\begin{equation}
\label{0_nointers}
]x,y[\,\,\cap\,\,]x',y'[=\emptyset.
\end{equation}
Hence, the optimal rays $\{Z^1_\mathfrak a\}_{\mathfrak a\in\mathfrak{A}^1}$ form a Borel partition of $\mathbb{R}^d$ into $1$-dimensional open segments, up to the set of their \emph{initial points} $\underset{\mathfrak a\in\mathfrak{A}^1}{\cup}\mathcal I(Z^1_\mathfrak a)\subset\mathbb{R}^d$ and of their \emph{final points} $\underset{\mathfrak a\in\mathfrak{A}^1}{\cup}\mathcal E(Z^1_\mathfrak a)\subset\mathbb{R}^d$, defined for every $\mathfrak a\in\mathfrak{A}^1$ by
\begin{equation}
Z^1_\mathfrak a=\bigl\{(1-t)\mathcal I(Z^1_\mathfrak a)+t\mathcal E(Z^1_\mathfrak a)\bigr\},\qquad \mathcal E(Z^1_\mathfrak a)-\mathcal I(Z^1_\mathfrak a)\in C^1_\mathfrak a,
\end{equation}
being $C^1_\mathfrak a$ the half-line in $\mathbb{R}^d$ giving the direction on $Z^1_\mathfrak a$ along which the transport occurs, i.e.
\[
\psi(y)-\psi(x)=\d{y-x}, \quad x,y\in Z^1_\mathfrak a\quad\Rightarrow\quad y-x\in C^1_\mathfrak a.
\]
The partition into optimal rays with directions of transport $\{Z^1_\mathfrak a, C^1_\mathfrak a\}_{\mathfrak a\in\mathfrak{A}^1}$ is the simplest example of what we will call \emph{directed locally affine partition}.
Moreover, the set of initial/final points of the optimal rays is $\mathcal L^d$-negligible (and then also $\mu$-negligible). Indeed, if $(x,y)$ satisfies \eqref{0_psi2} and $\psi$ is differentiable at $x$ --notice that this happens $\mathcal L^d$-a.e. (and then $\mu$-a.e.) since $\psi$ is Lipschitz--, then
\begin{equation}
\label{0_yxpartial}
y\in x+ \big( \partial\d{\cdot} \big)^{-1}(\nabla\psi(x)),
\end{equation}
where $\partial\d{\cdot}$ is the subdifferential of the convex norm, and, by strict convexity of $\d{\cdot}$, $(\partial\d{\cdot})^{-1}(\nabla\psi(x))$ is an half-line corresponding to a unique $C^1_\mathfrak a$.
We recall that the convex cones of the form $(\partial\d{\cdot})^{-1}(\ell)$ for some $\ell\in D^*$ are called \emph{exposed faces} of $\d{\cdot}$, while more generally the \emph{extremal faces} of $\d{\cdot}$ are by definition the projections on $\mathbb{R}^d$ of the extremal faces of the convex cone $\mathrm{epi}\d{\cdot}\subset\mathbb{R}^{d+1}$. In the strictly convex case, both concepts coincide and are given by half-lines.
\begin{figure
\centering
\subfloat[The construction of optimal rays through the potential $\psi$ and the epigraph of $\d{\cdot}$.]
{
\label{F:sf:alwaysbad}
\begin{minipage}[c]{8cm}
\centering\resizebox{8cm}{5cm}{\input{casostrettintrobis.pdf_t}}
\end{minipage}
}
\hskip 1cm
\subfloat[A set of optimal rays and a simple cone vector field.]
{
\label{F:sf:intermdep}
\begin{minipage}[c]{6cm}
\centering\resizebox{6cm}{5cm}{\input{sheafconeintrobis.pdf_t}}
\end{minipage}
}
\label{Fi_casostrett_intro}
\end{figure}
Assume w.l.o.g. that $\mu\perp\nu$ --hence $\pi\{(x,y):\,y\neq x\}=1$-- and that, for the moment, also $\nu\ll\mathcal L^d$.
Then, this first step yields that the optimal rays $\{Z^1_\mathfrak a\}_{\mathfrak a\in\mathfrak{A}^1}$ on which $\psi$ is differentiable form a partition in $\mathbb{R}^d$ --up to the $\mathcal L^d$-negligible set (thus also ($\mu+\nu$)-negligible) where $\psi$ is not differentiable-- s.t.
\[
\pi\Big(\underset{\mathfrak a\in\mathfrak{A}^1}{\bigcup}Z^1_\mathfrak a\times Z^1_\mathfrak a\Big)=1,\quad\forall\,\pi\in\Pi^{\mathrm{opt}}_{\d{\cdot}}(\mu,\nu).
\]
The second step of the strategy consists in decomposing the transport problem in the sets $\{Z^1_\mathfrak a \times \mathbb{R}^d\}_{\mathfrak a \in \mathfrak{A}^1}$.
More precisely, for any given cone $C\subset\mathbb{R}^d$, let us denote by $\mathtt c_{C}$ the cost function
\[
\mathtt c_{C}(x,y):=\mathbbm 1_{C}(y-x),
\]
where $\mathbbm 1_A$ is the indicator function of the set $A$ (see \eqref{E_indicator_function_A}).
For notational convenience, if $\mathtt c : \mathbb{R}^d \times \mathbb{R}^d \rightarrow [0,\infty]$ is a Borel cost function, we use the notation
\[
\Pi^f_\mathtt c(\mu,\nu) := \bigg\{ \pi \in \Pi(\mu,\nu) : \int \mathtt c(x,y) \,d\pi(x,y) < \infty \bigg\}.
\]
Then, by \eqref{0_yxpartial} and \eqref{0_nointers} it follows that if $\pi=\int\pi^1_\mathfrak a\,dm(\mathfrak a)$, $\mu=\int\mu^1_\mathfrak a\,dm(\mathfrak a)$ and $\nu=\int\nu^1_\mathfrak a\,dm(\mathfrak a)$ denote the strongly consistent disintegrations (see Definition \ref{D_dis}) of $\pi$ w.r.t. $\{Z^1_\mathfrak a\times\mathbb{R}^d\}_{\mathfrak a \in \mathfrak{A}^1}$ and of $\mu$ and $\nu$ w.r.t. $\{Z^1_\mathfrak a\}_{\mathfrak a \in \mathfrak{A}^1}$, one has
\begin{equation}
\label{0_opt_char}
\pi\in\Pi^{\mathrm{opt}}_{\d{\cdot}}(\mu,\nu)\quad\Leftrightarrow\quad \pi=\int\pi^1_\mathfrak a\,dm(\mathfrak a),\quad \text{$\pi^1_\mathfrak a\in\Pi^{f}_{\mathtt c_{C^1_\mathfrak a}}(\mu^1_\mathfrak a,\nu^1_\mathfrak a)$},
\end{equation}
being $\Pi^{f}_{\mathtt c_{C^1_\mathfrak a}}(\mu_\mathfrak a,\nu_\mathfrak a)$ the plans of finite $\mathtt c_{C^1_\mathfrak a}$-cost between $\mu_\mathfrak a$ and $\nu_\mathfrak a$.
In other words, the transport problem on $\mathbb{R}^d$ reduces to a family of independent $1$-dimensional transport problems with linear cost and prescribed direction. If $\mu^1_\mathfrak a$ has no atoms, then the unique transference plan concentrated on a monotone graph in $Z^1_\mathfrak a \times Z^1_\mathfrak a$ is actually concentrated on a map $\mathtt T^1_\mathfrak a$. In this setting, monotone means monotone w.r.t. the order induced by $C^1_\mathfrak a$ on $Z^1_\mathfrak a$, and the statement is a well known and simple result for 1-dimensional problems, which can be seen as a particular case of a more general structure result for optimal transportation problems with quadratic cost (see for example \cite{Bre:polarrear}).
Then, the main problem in \cite{Car:strictly} was to prove that the disintegration of $\mathcal L^d$ (and thus of $\mu$) on the optimal rays has non-atomic conditional measures. Indeed, for a general Borel partition into segments this might not be true, as discovered in a counterexample to the original Sudakov's proof by Alberti, Kirchheim and Preiss (see personal communication in \cite{ambprat:crist}). The main issue was then to prove that the optimal rays satisfy an additional regularity property which guarantees that the conditional measures of $\mathcal L^d$ are not atomic. In \cite{ambr:lecttrans}, \cite{conf:optcime}, \cite{caffafeldmc} and \cite{trudiwang}, due to the additional regularity assumptions either on the measures $\mu$, $\nu$ or on the norm, the unit vector field giving at each point of an optimal ray the direction of transport is locally Lipschitz. Then, via changes of variables using the classical Coarea Formula one can reduce to study the disintegration of the Lebesgue measure on families of parallel segments, namely Fubini theorem, which gives the absolute continuity of the conditional measures w.r.t. the $1$-dimensional Hausdorff measure $\mathcal H^1$ on the segments on which they are concentrated. The absolute continuity of the conditional measures w.r.t. the $1$-dimensional Hausdorff measure on the optimal rays --thus implying the solvability of the Monge problem-- for general strictly convex norms was proved in \cite{Car:strictly}. Since in the general case no Lipschitz regularity is available, the author used a technique first introduced for a partition into segments arising from a different variational problem in \cite{BiaGlo}. Such a technique is based on the validity for the family of segments (in this case, the optimal rays) of an approximation property via sequences of cone vector fields, that we call \emph{cone approximation property} (with the same terminology used in the first part of \cite{Dan:Phd}).
We point out that, compared to the approach followed in \cite{champdepasc:MongeS, champdepasc:Monge}, Sudakov's approach for the Monge problem gives and relies upon a deeper geometric characterization of the transport via optimal plans, namely the existence of a family of lower dimensional regions (in the strictly convex case, $1$-dimensional) on which the transport occurs and on which the existence of optimal maps becomes easier to prove.
It remained unclear if the original strategy of Sudakov can be successful not only in the case of strictly convex norms, thus giving a complete geometric characterization of the optimal transport plans via decomposition into lower dimensional transportation problems.
The aim of this paper is to show how Sudakov's approach can be carried on also in the general convex case. In the next section we define new concepts, which in the strictly convex case (i.e. when the extremal faces of $\d{\cdot}$ are 1-dimensional) are trivially satisfied by the decomposition in optimal rays $Z^1_\mathfrak a$, and state our main results, giving an overall idea of the whole construction.
\subsection{Sudakov's strategy in the general convex case}
\label{Ss_main_Sud}
Recall that, for all $(x,y)$ as in \eqref{0_psi2}, \eqref{0_ztzs} holds. In the strictly convex case, we have seen that \eqref{0_ztzs} and \eqref{0_non_branch} imply that whenever
\begin{equation}
\label{0_xy'y''}
\exists\,y',y''\neq x \quad \text{s.t.} \qquad \psi(y')-\psi(x)=\d{y'-x} \ \ \text{and} \ \ \psi(x)-\psi(y'')=\d{x-y''},
\end{equation}
then $x$ belongs to a segment $Z^1_\mathfrak a$ called optimal ray, which belongs to a partition on $\mathbb{R}^d$ on which the transport occurs along the direction $C^1_\mathfrak a=\mathbb{R}^+(y'-x)=\mathbb{R}^+(x-y'')$.
However, for general convex norms, the optimal rays do not satisfy \eqref{0_nointers} and then do not form a partition in $\mathbb{R}^d$. Actually, $\forall\,x\in\mathbb{R}^d$, the sets
\begin{equation}
\label{0_subsuper}
\partial^+\psi(x):= \big\{y':\,\psi(y')-\psi(x)=\d{y'-x} \big\},\quad\partial^-\psi(x):= \big\{y'':\,\psi(x)-\psi(y'')=\d{x-y''} \big\},
\end{equation}
called respectively \emph{superdifferential and subdifferential of $\psi$ at $x$}, may be contained in one or even more higher dimensional cones corresponding to extremal faces of $\d{\cdot}$. Now, unlike in the strictly convex case, an extremal face is not in general a $1$-dimensional half line but a $k$-dimensional cone, with $k=1,\dots,d$. Hence Sudakov claimed that the regions on which the transport occurs are relatively open subsets of affine planes whose dimension is equal to $k$. However, even when considering the set of points in the super/subdifferential of $\psi$ at a certain point $x$ which are contained in a single $k$-extremal cone of $\mathrm{epi}\d{\cdot}$, it may not be a convex $k$-dimensional set or more generally a set with a well defined affine dimension (see Figure \ref{Fi_intro_supersub}).
\begin{figure}
\centering{\resizebox{16cm}{6cm}{\input{sudakov_complete_intro.pdf_t}}}
\caption{In the left picture a possible superdifferential $\partial^+\mathrm{graph}\,\psi$ at a point $z$ of $\mathrm{graph}\,\psi$ is depicted in black and different blue colors. Notice that it is not convex, not even inside extremal faces of the norm. We also underline in dark blue a set $\mathtt{O}(z,w)$ for some $w\in\partial^+\mathrm{graph}\,\psi(z)$, in order to show the completeness property.
In the right picture we depict in red the super/subdifferential at a point $z$ of the regular set (yellow region) in $\mathrm{graph}\,\psi$. The cone $z + \mathrm{epi} \d{\cdot}$ is also represented.
}
\label{Fi_intro_supersub}
\end{figure}
When we faced this problem for general convex norms, the first main issue was to find other conditions which determine that a point $x$ belongs to one of the desired $k$-dimensional regions, thus generalizing the property that whenever $y'\in\partial^+\psi(x)\setminus \{x\}$, $y''\in\partial^-\psi(x)\setminus \{x\}$ then $\mathbb{R}^+(y'-x)=\mathbb{R}^+(x-y'')$ and $x$ belongs to the optimal ray containing the segment $]y'',y'[$.
The natural generalization of the partition into optimal rays for strictly convex norms is to look for a \emph{directed locally affine partition} $\{Z^k_\a,C^k_\a\}_{\nfrac{k=1,\dots,d}{\mathfrak a \in \mathfrak{A}^k}}$ of $\mathbb{R}^d$ (see Definition \ref{D_locaffpart}), namely a Borel partition of $\mathbb{R}^d$ into sets $Z^k_\a$ which are locally affine and $k$-dimensional, i.e. relatively open in their affine hull whose linear dimension is $k$, together with an extremal cone $C^k_\mathfrak a$ of $\d{\cdot}$ that will correspond to the union of directions of the optimal rays starting from $x \in Z^k_\a$.
The first key idea is to observe that Kantorovich duality \eqref{0_psi1}-\eqref{0_psi2} can be rewritten as follows (see Section \ref{Ss_convex_norm_cone}). Let $\hat\mu=(\mathbb{I}\times\psi)_\#\mu$, $\hat\nu=(\mathbb{I}\times\psi)_\#\nu$ and $\hat\pi=((\mathbb{I}\times\psi)\times(\mathbb{I}\times\psi))_\#\pi$. One has
\begin{align}
\pi\in\Pi^{\mathrm{opt}}_{\d{\cdot}}(\mu,\nu)\quad&\Leftrightarrow\quad\hat\pi\in\Pi^f_{\mathtt c_{\mathrm{epi}\d{\cdot}}}(\hat\mu,\hat\nu)\label{0_lifting1}\\
&\Leftrightarrow\quad\hat\pi\bigl(\partial^+\mathrm{graph}\,\psi\bigr)=1,\label{0_lifting2}
\end{align}
where
\begin{align}
\partial^+ \mathrm{graph}\,\psi:=&~\bigl(\mathrm{graph}\,\psi\times\mathrm{graph}\,\psi\bigr)\cap\bigl\{\mathtt c_{\mathrm{epi}\d{\cdot}}<+\infty\bigr\}\notag\\
=&~\mathrm{graph}\,\psi\times\mathrm{graph}\,\psi\cap\mathtt p_{\mathbb{R}^d}^{-1}(\partial^+\psi). \label{0_superdiff_graph}
\end{align}
is the \emph{superdifferential of the set $\mathrm{graph}\,\psi\subset\mathbb{R}^{d+1}$}.
In other words, \eqref{0_lifting1} tells us that studying the optimal transportation problem between $\mu$ and $\nu$ in $\mathbb{R}^d$ is equivalent to study the finite cost transportation problem in $\mathbb{R}^{d+1}$ for a convex cone cost (precisely $\mathtt c_{\mathrm{epi}\d{\cdot}}$) between measures ($\hat\mu$, $\hat\nu$) concentrated on a $\d{\cdot}$-Lipschitz graph (namely $\mathrm{graph}\,\psi$)
or, by \eqref{0_lifting2}, to study transport plans which are concentrated on the superdifferential of the graph of the $\d{\cdot}$-Lipschitz function $\psi$.
The advantage of this point of view is that the properties of the super/subdifferential of $\psi$ which permit to generalize \eqref{0_xy'y''}, and then to find a locally affine directed partition, can be more naturally expressed in terms of geometric properties of the super/subdifferential of $\mathrm{graph}\,\psi$ --where the subdifferential of $\mathrm{graph}\,\psi$ is the set $\partial^-\mathrm{graph}\,\psi:=\big(\partial^+\mathrm{graph}\,\psi\big)^{-1}$.
First we will find a directed locally affine partition $\{\tilde Z^k_\mathfrak a,\tilde C^k_\mathfrak a\}_{\nfrac{k=1,\dots,d}{\mathfrak a \in \mathfrak{A}^k}}$ in $\mathbb{R}^{d+1}$ for this transportation problem, whose direction cones $\tilde C^k_\mathfrak a$ are extremal faces of $\mathrm{epi}\d{\cdot}$ and on which the disintegration of the $d$-dimensional Hausdorff measure $\mathcal H^d$ on $\mathrm{graph}\,\psi$ has conditional measures which are absolutely continuous w.r.t. $\mathcal H^k \llcorner \tilde Z^k_\mathfrak a$, and then we will find the desired locally affine partition $\{Z^k_\mathfrak a,C^k_\mathfrak a\}_{\nfrac{k=1,\dots,d}{\mathfrak a \in \mathfrak{A}^k}}$ simply projecting it on $\mathbb{R}^d$. Indeed, the extremal faces of $\d{\cdot}$ are by definition the projections on $\mathbb{R}^d$ of the extremal faces of $\mathrm{epi}\d{\cdot}$ and the ``lifting map'' $\mathbb{I}\times\psi$ is bi-Lipschitz, thus mapping negligible sets into negligible sets.
The crucial properties of the super/subdifferential $\partial^\pm\mathrm{graph}\,\psi$ that we will use to find the partition are the so-called \emph{transitivity property}
\begin{equation}
\label{0_trans}
w'\in\partial^\pm\mathrm{graph}\,\psi(w)\quad\Rightarrow\quad\partial^\pm\mathrm{graph}\,\psi(w')\subset\partial^\pm\mathrm{graph}\,\psi(w)
\end{equation}
and the \emph{completeness property} of the $\d{\cdot}$-Lipschitz graph $\mathrm{graph}\,\psi$, that we define below.
Let $F$ be an extremal face of the convex cone $\mathrm{epi}\d{\cdot}$ and denote by $\inter_{\mathrm{rel}} F$ its relative interior, namely its interior w.r.t. its affine hull. Moreover, for any $z,w\in\mathbb{R}^{d+1}$ let
\begin{equation}
\mathtt{O}(z,w):= z+\mathrm{epi}\d{\cdot} \cap w-\mathrm{epi}\d{\cdot}.
\end{equation}
The completeness property of $\mathrm{graph}\,\psi$ is the following:
\begin{align}
&w\in\partial^+\mathrm{graph}\,\psi(z),\quad w-z\in\inter_{\mathrm{rel}} F\quad\Rightarrow\quad\mathtt{O}(z,w)=z+F\cap w-F\subset\partial^+\mathrm{graph}\,\psi(z),\label{0_compl1}\\\
&w\in\partial^-\mathrm{graph}\,\psi(z),\quad z-w\in\inter_{\mathrm{rel}} F\quad\Rightarrow\quad\mathtt{O}(w,z)=w+F\cap z-F\subset\partial^-\mathrm{graph}\,\psi(z),\label{0_compl2}
\end{align}
where $z+ F\cap w-F$ is convex and satisfies $\mathbb{R}^+\bigl((z+F\cap w-F)-z\bigr)=\mathbb{R}^+\bigl(w-(z+F\cap w-F)\bigr)=F$ (see Proposition \ref{P_parall}).
In the strictly convex case, the extremal faces $F$ of $\mathrm{epi}\d{\cdot}$ are half-lines. Moreover, whenever \eqref{0_compl1} (resp. \eqref{0_compl2}) holds $\mathtt p_{\mathbb{R}^d} F$ is the extremal face of $\d{\cdot}$ giving the direction of an optimal ray starting (resp. arriving) at $x=\mathtt p_{\mathbb{R}^d} z$, and $\mathtt p_{\mathbb{R}^d}\mathtt{O}(z,w)=\bigl[\mathtt p_{\mathbb{R}^d} z,\,\mathtt p_{\mathbb{R}^d} w\bigr]$.
In the general convex case, the completeness property \eqref{0_compl1}-\eqref{0_compl2} then implies that whenever the directions of the optimal rays starting/arriving at a point $z$ are contained in a certain face $F$ and there exists a direction in $\inter_{\mathrm{rel}} F$, then they are a cone of directions coinciding with $F$. Moreover, by the transitivity property \eqref{0_trans}, whenever the same thing happens also for two points each belonging to one of the sets $B(x,\delta)\cap (z\pm \inter_{\mathrm{rel}} F)$, then $z$ has a locally affine neighborhood, of the same dimension as $F$ and contained in $z + \mathrm{aff} F$, made of points for which the admissible directions of transport coincide with the directions of $F$.
Roughly speaking, the relative interior of the extremal face $F$ plays the role of a direction of an optimal ray and the set $\mathtt p_{\mathbb{R}^d}\mathtt{O}(z,w)$ the role of the segment $[x,y]$ inside such optimal ray in the strictly convex case (see Figure \ref{Fi_intro_supersub}).
The suitable generalization of \eqref{0_xy'y''} and its implications can then be found in the concept of what we call \emph{regular transport set} $\mathcal R \theta_\psi$. The notation will be clear in Section \ref{S_foliations}
when we study the more general transport problem for $\mathtt c_{\tilde{\mathbf C}}$-Lipschitz foliations, namely a family of graphs of $|\cdot|_{D(\mathfrak a)^*}$-Lipschitz functions depending on a continuous parameter $\mathfrak a$ (Section \ref{S_foliations} and Proposition \ref{P_ex_fol}). The study of $\mathtt c_{\tilde{\mathbf C}}$-Lipschitz foliations will be one of the main issues to complete the construction of a suitable directed locally affine partition (Theorem \ref{T_final}) on which to solve \eqref{E_transpo_Sud}. The points in $\mathcal R\theta_\psi$ are the points $z$ such that
\begin{enumerate}
\item \label{Point_intro_loafpr1} the set of directions
\[
\mathcal D^+ \theta_\psi(z) = \bigg\{ \frac{w-z}{|w-z|}:\, w \in \partial^+\mathrm{graph}\, \psi(z) \setminus \{z\} \bigg\},
\]
of the optimal rays starting in $z$ is convex in $\mathbb S^{d-1}$, and the same for the set of directions
\[
\mathcal D^- \theta_\psi(z) = \bigg\{ \frac{z-w}{|z-w|}:\, w\in \partial^- \mathrm{graph}\,\psi(z) \setminus \{z\} \bigg\},
\]
of the optimal rays arriving in $z$,
\item \label{Point_intro_loafpr2} the two sets $\mathcal D^+ \theta_\psi(z)$, $\mathcal D^- \theta_\psi(z)$ coincide,
\item \label{Point_intro_loafpr3} there are points $w'$, $w''$ such that
\[
\frac{z-w'}{|z-w'|} \in \inter_{\mathrm{rel}} \mathcal D^- \theta_\psi(z), \qquad \frac{w''-z}{|w''-z|} \in \inter_{\mathrm{rel}} \mathcal D^+ \theta_\psi(z)
\]
and Points (\ref{Point_intro_loafpr1}-\ref{Point_intro_loafpr2}) hold for $w'$, $w''$ too.
\end{enumerate}
Then the sets $\tilde Z^k_\a$, $\tilde C^k_\a$ are now determined by
\[
z \in \tilde Z^k_\a \quad \Longrightarrow \quad
\left\{
\begin{aligned}
\tilde Z^k_\a =~& \mathcal R \theta_\psi \cap \mathrm{aff}\partial^+\mathrm{graph}\, \psi(z), \crcr
\tilde C^k_\a =&~\mathrm{epi}\, \d{\cdot} \cap \Big( \mathrm{aff} \partial^+ \mathrm{graph}\,\psi(z)-z\Big).
\end{aligned}
\right.
\]
Such a directed locally affine partition will be called \emph{differential partition}.
One can see that the sets $\tilde Z^k_\a$ are relatively open in their affine hull, and that $\tilde C^k_\a$ are extremal faces of $\mathrm{epi}\d{\cdot}$.
Recall that the index $k$ denotes the affine dimension of $\tilde Z^k_\a$, which coincides with the linear dimension of $\tilde C^k_\a$, while $\mathfrak a \in \mathfrak{A}^k$ is an index of continuum cardinality.
The second step in the strategy is then to show that the transport problem $\Pi^f_{\mathtt c_{\mathrm{epi}\d{\cdot}}}(\hat\mu,\hat\nu)$ can be decomposed, via disintegration of measures, into a family of finite cost transport problems on $\{\tilde Z^k_\a\times\mathbb{R}^{d+1}\}_{\nfrac{k=1,\dots,d}{\mathfrak a \in \mathfrak{A}^k}}$ with first marginals which are absolutely continuous w.r.t. the Hausdorff measure $\mathcal H^k$ on the $k$-dimensional set $\tilde Z^k_\a$ on which they are concentrated.
Since the definition of ``good points'', i.e. of the regular transport set $\mathcal R \theta_\psi$, is definitely more complicated than in the strictly convex case, it is perfectly understandable that the proof of the absolute continuity w.r.t. to the Hausdorff measure on the $k$-dimensional sets $\{\tilde Z^k_\a\}_{k,\mathfrak a}$ of the conditional probabilities of the disintegration of $\mathcal H^d$ (and then of $\hat\mu$) are considerably more intricate. The main reference for the approach used in this part is \cite{CarDan}, where the so-called \emph{cone approximation property} introduced in \cite{BiaGlo} (with the terminology used in \cite{Dan:Phd}) was first generalized to partitions into higher dimensional sets, showing the absolute continuity property for the conditional probabilities of the disintegration of the surface measure on the graph of a convex function w.r.t. the partition induced by the relative interior of the extremal faces. In particular, in Section \ref{S_disintechnique} it is shown that the differential partition satisfies both the \emph{forward} and the \emph{backward cone approximation property}, namely the cone approximation property holds both for the optimal rays starting at a point $z$ and for the points arriving at $z$, thus giving that the conditional measures of $\mathcal H^d$ are indeed equivalent to the $k$-dimensional Hausdorff measure on the set on which they are concentrated.
As for the proof of the $\mathcal H^d$-negligibility of the set $\mathbb{R}^{d+1} \setminus \mathcal R\theta_\psi$, since for general convex norms the extremal faces may be more than the exposed ones, it is not possible to use the same reasoning as in the strictly convex case.
However, we will show that the set $\mathbb{R}^{d+1} \setminus \mathcal R\theta_\psi$ is made of \emph{initial/final points} for two other partitions (the \emph{super/subdifferential partitions} introduced in \ref{Ss_regu_resi_set}), which satisfy the (initial/final) \emph{forward/backward cone approximation property}. Hence, the same disintegration technique used in \cite{CarDan} permits to show that they are $\mathcal H^d\llcorner\mathrm{graph}\,\psi$-negligible (see Theorem \ref{T_FC_no_initial}).
Denoting with $\{Z^k_\a,C^k_\a\}_{\nfrac{k=1,\dots,d}{\mathfrak a \in \mathfrak{A}^k}}$ the projection of the differential partition $\{\tilde Z^k_\a,\tilde C^k_\a\}_{\nfrac{k=1,\dots,d}{\mathfrak a \in \mathfrak{A}^k}}$ on $\mathbb{R}^d$, in Section \ref{S_theorem_1_proof} we deduce the following theorem. The statement includes also the points which do not belong to any optimal ray, and in that case the dimension $k$ of the elements of the directed locally affine partition they belong to, is $k=0$, as well as $C^0_\mathfrak a = \{0\}$.
Since we will often write the graph of a directed locally affine partition $\{Z^k_\a,C^k_\a\}_{k,\mathfrak a}$ as
\[
\mathbf D:=\Big\{(k,\mathfrak a,z,C^k_\a):\,k\in\{0,\dots,d\},\,\mathfrak a\in\mathfrak{A}^k,\,z\in Z^k_\a\Big\},
\]
we will use also the notation
\begin{equation}
\label{E_cost_all_cone_intro}
\mathtt c_\mathbf D(x,y) :=
\begin{cases}
\mathbbm 1_{C^k_\a}(y-x) & \text{if} \ \exists\, k,\mathfrak a \ \text{s.t.} \ x \in Z^k_\a, \crcr
+\infty & \text{otherwise}.
\end{cases}
\end{equation}
Notice that for costs $\mathtt c$ of the form \eqref{E_cost_all_cone_intro}, one has clearly $\Pi^\mathrm{opt}_\mathtt c(\mu,\nu) = \Pi^f_\mathtt c(\mu,\nu)$, since the only values of $\mathtt c$ are $0$, $\infty$.
\begin{theorem}
\label{T_1}
Let $\mu,\nu \in \mathcal P(\mathbb{R}^d)$ with $\mu \ll \mathcal L^d$ and let $\d{\cdot}$ be a convex norm in $\mathbb{R}^d$. Then there exists a locally affine directed partition $\{Z^k_\mathfrak a,C^k_\mathfrak a\}_{\overset{k=0,\dots,d}{\mathfrak a\in\mathfrak{A}^k}}$ in $\mathbb{R}^d$ with the following properties:
\begin{enumerate}
\item \label{Point_1_T_1} for all $\mathfrak a \in \mathfrak{A}^k$ the cone $C^k_\mathfrak a$ is a $k$-dimensional extremal face of $\d{\cdot}$;
\item \label{Point_2_T_1} $\displaystyle{\mathcal L^d \biggl( \mathbb{R}^d \setminus \underset{k,\mathfrak a}{\bigcup}\; Z^k_\mathfrak a \biggr)=0}$;
\item \label{Point_3_T_1} the disintegration of $\mathcal L^d$ w.r.t. the partition $\{Z^k_\a\}_{k,\mathfrak a}$, $\displaystyle{\mathcal L^d \llcorner_{\underset{k,\mathfrak a}{\cup} Z^k_\a} = \int v^k_\mathfrak a\,d\eta(k,\mathfrak a)}$, satisfies
\[
v^k_a \simeq \mathcal H^k \llcorner_{Z^k_\mathfrak a};
\]
\item \label{Point_4_T_1} for all $\pi \in \Pi^{\mathrm{opt}}_{\d{\cdot}}(\mu,\nu)$, the disintegration $\displaystyle{\pi = \int \pi^k_\a\, dm(k,\mathfrak a)}$ w.r.t. the partition $\{Z^k_\a \times \mathbb{R}^d\}_{k,\mathfrak a}$ satisfies
\[
\pi^k_\mathfrak a \in \Pi^ f_{\mathtt c_{C^k_\mathfrak a}} \big( \mu^k_\a, (\mathtt p_2)_\# \pi^k_\a \big),
\]
where $\displaystyle{\mu = \int \mu^k_\a\,dm(k,\mathfrak a)}$ is the disintegration w.r.t. the partition $\{Z^k_\a\}_{k,\mathfrak a}$, and moreover
\[
(\mathtt p_2)_\#\pi^k_\a \biggl( Z^k_\a \cup \biggl( \mathbb{R}^d \setminus \underset{(k',\mathfrak a') \not= (k,\mathfrak a)}{\bigcup} Z^{k'}_{\mathfrak a'} \biggr) \biggr) = 1.
\]
\end{enumerate}
If also $\nu \ll \mathcal L^d$, then for all $\pi\in\Pi^{\mathrm{opt}}_{\d{\cdot}}(\mu,\nu)$
\[
(\mathtt p_2)_\#\pi^k_\a = \nu^k_\a
\]
where $\displaystyle{\nu = \int \nu^k_\a \,dm(k,\mathfrak a)}$ is the disintegration w.r.t. the partition $\{Z^k_\a\}_{k,\mathfrak a}$, and the converse of Point \eqref{Point_4_T_1} holds:
\begin{equation*}
\pi^k_\a \in \Pi^ f_{\mathtt c_{C^k_\a}}(\mu^k_\a,\nu^k_\a) \quad \Longrightarrow \quad \pi \in \Pi^{\mathrm{opt}}_{\d{\cdot}}(\mu,\nu).
\end{equation*}
\end{theorem}
A locally affine directed partition satisfying Point \eqref{Point_3_T_1} is called \emph{Lebesgue-regular} (see Definition \ref{D_disint_regular}). This concludes the first part of the paper.
A remark is in order here: in Point \eqref{Point_4_T_1}, the conditional second marginals $(\mathtt p_2)_\# \pi^k_\a$ are independent on the potential $\psi$ but \emph{depend} on the particular transference plan $\pi$ which we are decomposing. This can be seen with elementary examples (see Example \ref{Ex_2ndmarg} in Section \ref{Ss_partitions_intro}). From now on the analysis will be done in a class of transference plans which have the same conditional second marginals: in fact, we will see in a moment that the partition of Theorem \ref{T_1} needs to be refined and by inspection one sees that such refinement changes when changing the conditional marginals.
We will consider then nonempty subsets of the optimal plans of the form
\[
\Pi^f_{\mathtt c_{\mathbf D}}(\mu,\{\bar\nu_\mathfrak a\}):=\Big\{\pi\in\Pi^f_{\mathtt c_{\mathbf D}}(\mu,\nu):\, (\mathtt p_2)_\#\pi_\mathfrak a=\bar\nu_\mathfrak a\Big\},
\]
that is equivalent to fix a transport plan of finite $\mathtt c_{\mathbf D}$-cost $\check\pi$ and consider all transport plans $\pi\in\Pi^f_{\mathtt c_{\mathbf D}}(\mu,\{(\mathtt p_2)_\#\check \pi_\mathfrak a\})$.
In the strictly convex case, Theorem \ref{T_1} has been proven in \cite{Car:strictly}. There the dimensions of the sets of the locally affine partition is equal to one, and it is classical and fairly easy to see that the optimal transportation problems
\[
\Pi^{\mathrm{opt}}_{\mathtt c^1_{\mathfrak a,2}}(\mu^1_\mathfrak a,\nu^1_\mathfrak a),\quad\mu^1_\mathfrak a(Z^1_\mathfrak a)=1,
\]
where
\[
\mathtt c^1_{\mathfrak a,2}(x,y)= \begin{cases}
|y-x|^2 & \text{if $x\in Z^1_\mathfrak a$, $y-x\in C^1_\mathfrak a$}, \\
+\infty & \text{otherwise},
\end{cases}
\]
have a solution induced by a map $\mathtt T^1_\mathfrak a:Z^1_\mathfrak a\rightarrow\mathbb{R}^d$. More precisely, one shows that any $\mathtt c^1_{\mathfrak a,2}$-cyclically monotone transference plan is induced by a unique transport map $\mathtt T^1_\mathfrak a$. Since the dependence of the maps $\mathtt T^1_\mathfrak a$ on $\mathfrak a$ is $m$-measurable, the map $\mathtt T(x):=\sum_{\mathfrak a\in\mathfrak{A}}\mathtt T^1_\mathfrak a(x)\chi_{Z_\mathfrak a^1}(x)$ is an optimal map for \eqref{E_transpo_Sud}.
Actually, $\mathtt T$ is the unique optimal transport map relative to the cost
\begin{equation}
\label{0_c2}
\mathtt c_{2}(x,y):= \begin{cases}
|y-x|^2 & \text{if $\mathtt c_{\mathbf D}(x,y)<+\infty$}, \\
+\infty & \text{otherwise}.
\end{cases}
\end{equation}
In the general convex case, the analogous way to solve \eqref{E_transpo_Sud} would be to prove that the optimal transportation problems on the sets of the partition of Theorem \ref{T_1}
\[
\Pi^{\mathrm{opt}}_{\mathtt c^k_{\mathfrak a,2}}(\mu^k_\mathfrak a,\nu^k_\mathfrak a),\quad\mu^k_\mathfrak a(Z^k_\mathfrak a)=1,
\]
where
\[
\mathtt c^k_{\mathfrak a,2}(x,y)=\begin{cases}
|y-x|^2 & \text{if $x\in Z^k_\mathfrak a$, $y-x\in C^k_\mathfrak a$},\\
+\infty & \text{otherwise},
\end{cases}
\]
have a solution induced by a map $\mathtt T^k_\mathfrak a:Z^k_\mathfrak a\rightarrow\mathbb{R}^d$ whose graph is the support of any $\mathtt c^k_{\mathfrak a,2}$-cyclically monotone transference plan, and then to glue together the maps $\mathtt T^k_\a$.
This fact would be true, by classical results in optimal transportation, if there existed a pair of optimal potentials $\phi^k_\a$, $\psi^k_\a$ for the cost $\mathtt c_{C^k_\a}$. Recall that, for a cost $\mathtt c : \mathbb{R}^d \times \mathbb{R}^d \rightarrow [0,\infty]$, one calls optimal potentials a pair of functions $\phi$, $\psi$ s.t.
\begin{align*}
&\phi,\,\psi:\mathbb{R}^d\rightarrow[-\infty,+\infty), \quad \text{ $\phi$ $\mu$-measurable and $\psi$ $\nu$-measurable}, \\
&\phi(x)+\psi(y)\leq \mathtt c(x,y), \quad \forall\, x,y \in \mathbb{R}^d, \\
&\phi(x)+\psi(y)= \mathtt c(x,y), \quad \text{ $\pi$-a.e. for some $\pi\in \Pi(\mu,\nu)$}.
\end{align*}
Recall also that, if $\Gamma\subset \mathbb{R}^d\times \mathbb{R}^d$ is a carriage for $\pi$ and $(x_0,y_0) \in \Gamma$, then
\begin{align}
\label{E_general_phi_intro}
\phi(x) &:= \inf \bigg\{ \sum_{i=0}^I \mathtt c(x_{i+1},y_i) - \mathtt c(x_i,y_i):\, I \in \mathbb{N},\, (x_i,y_i) \in \Gamma,\, x_{I+1} = x \bigg\},\\
\psi(x)&:=\mathtt c(x,y)-\phi(x)\
\end{align}
yield a pair $\phi$, $\psi$ of optimal potentials provided $\phi$ is $\mu$-a.e. finite. When $\mathtt c$ is a convex norm, then $\psi=-\phi$ is a Kantorovich potential.
Indeed, by formula \eqref{E_general_phi_intro}, if $\exists\,\phi^k_\a,\, \psi^k_\a$ optimal potentials w.r.t. $\mathtt c_{C^k_\a}$ then there exist also $\phi^k_{\mathfrak a,2}$, $\psi^k_{\mathfrak a,2}$ optimal potentials for $\mathtt c^k_{\mathfrak a,2}$ and it is then classical to show that any $\mathtt c^k_{\mathfrak a,2}$-cyclically monotone transference plan is unique and induced by an optimal map $T^k_\a$.
However, as shown in \cite{Car1}, in general the transport problem in $\Pi^f_{\mathtt c_{C^k_\a}}(\mu^k_\a,\{(\mathtt p_2)_\# \pi^k_\a\})$ on $Z^k_\a$ with cost $\mathtt c_{C^k_\a}$ does not have a potential $\phi^k_\a$ (see the final example of \cite{Car1}), thus the directed locally affine partition of Theorem \ref{T_1} is not refined enough to give immediately the existence of transport maps in each of the sets $Z^k_\a \times \mathbb{R}^d$. Another approach that has been used at this point to show the existence of an optimal map assuming the existence of a directed locally affine partition is the one adopted in \cite{JimSan:quadratic}, which though uses techniques similar to \cite{champdepasc:Monge}, and then is not really simplifying the problem in the spirit of Sudakov's strategy.
What we show in the second part of the paper, more precisely in Section \ref{S_cfibr_cfol}, is that the directed locally affine partition of Theorem \ref{T_1} can be refined into another directed locally affine partition $\{\check Z^{',\ell}_\b, \check C^{',\ell}_\b\}_{\nfrac{\ell = 0,\dots,d}{\mathfrak b \in \mathfrak B^\ell}}$ such that, given a carriage of any $\mathtt c^k_{\mathfrak a,2}$-cyclically monotone transference plan, a pair of optimal potentials $\check \phi^{',\ell}_\b$, $\check \psi^{',\ell}_\b$ can be constructed on each of its elements $\check Z^{',\ell}_\b$.
In order to explain what we mean by a ``refinement'' of the partition of Theorem \ref{T_1}, referring to Section \ref{Ss_linear_pre_unique_transf} for wider motivations and more precise statements, let us consider formula \eqref{E_general_phi_intro}. The sequence of points
\[
(x_0,y_0), (x_1,y_0), (x_1,y_1), (x_2,y_1), \dots, (x_i,y_i), ( x_{i+1},y_i), (x_{i+1},y_{i+1}), \dots, (x,y_I), \qquad (x_i,y_i) \in \Gamma,
\]
is an \emph{axial path}, and we say that the axial path is a \emph{$(\Gamma,\mathtt c)$-axial path} if $\mathtt c(x,y) < \infty$ for all couples $(x,y)$ in the axial path: since we can assume that $\Gamma \subset \{\mathtt c < \infty\}$, this condition is equivalent to $\mathtt c(x_{i+1},y_i), \mathtt c(x,y_I) < \infty$. It is a well know fact that if $\mu$-a.a. points belong to an axial path starting from and ending in $(x_0,y_0)$ (which will be called a \emph{$(\Gamma, \mathtt c)$-cycle}), then formula \eqref{E_general_phi_intro} yields a $\mu$-a.e. finite potential $\phi$. Its dual $\psi$ turns out then to be finite and independent on $x$ for $\nu$-a.e. $y \in \mathbb{R}^d$.
It becomes then natural to ask for a directed locally affine partition $\{Z^k_\a,C^k_\a\}_{k,\mathfrak a}$ that, in addition to \eqref{Point_1_T_1}, \eqref{Point_2_T_1}, \eqref{Point_3_T_1} and \eqref{Point_4_T_1} of Theorem \ref{T_1} for all $\pi\in\Pi^f_{\mathtt c_{\mathbf D}}(\mu,\{\bar\nu_\mathfrak a\})$, it satisfies the following property. For all carriages $\Gamma\subset\{\mathtt c_{\mathbf D}<+\infty\}$ s.t. $\pi(\Gamma)=1$ for some $\pi\in\Pi^f_{\mathtt c_{\mathbf D}}(\mu,\{\bar\nu_\mathfrak a\})$, the sets $Z^k_\a$ are contained in a $(\Gamma,\mathtt c_{C^k_\a})$-cycle up to a $\mu^k_\a$-negligible set (eventually depending on $\Gamma$): the cost in each $Z^k_\a$ is the \emph{cone cost} given by $\mathtt c_{C^k_\a}(x,y) = \mathbbm 1_{C^k_\a}(y-x)$.
This cyclical connectedness condition is called in this paper \emph{$\Pi^f_{\mathtt c_{\mathbf D}}(\mu,\{\bar\nu_\mathfrak a\})$-cyclical connectedness} (see Definition \ref{D_pimunuconn}) and, as discussed above, when verified it guarantees the existence of optimal potentials.
The second main result of this paper claims the existence of such a partition. The fact that it is a refinement of an already existing locally affine partition, such as the one of Theorem \ref{T_1}, namely that each of its sets is contained in some $Z^k_\a$
and the corresponding cone of directions is an extremal face of the cone $C^k_\a$, is expressed by saying that it is a \emph{subpartition} of $\{Z^k_\a, \,C^k_\a\}$ (see Definition \ref{D_dir_subpart}). Recall Definition \ref{D_disint_regular} of Lebesgue-regular partition.
\begin{theorem}
\label{T_subpart_final}
Let $\{Z^k_\a, C^k_\a\}_{\nfrac{k = 0,\dots,d}{\mathfrak a \in \mathfrak A^k}}$ be a \emph{Lebesgue-regular} directed locally affine partition in $\mathbb{R}^d$ and let $\mu\ll\mathcal L^d$, $\nu\in\mathcal P(\mathbb{R}^d)$ such that $\Pi^f_{\mathtt c_{\mathbf D}}(\mu,\nu)\neq\emptyset$.
Then, for all $\check \pi\in\Pi^f_{\mathtt c_{\mathbf D}}(\mu,\nu)$ there exists a directed locally affine subpartition $\{\check Z^{',\ell}_\b, \check C^{',\ell}_\b\}_{\nfrac{\ell = 0,\dots,d}{\mathfrak b \in \mathfrak B^\ell}}$ of $\{Z^k_\a, C^k_\a\}_{\nfrac{k = 0,\dots,d}{\mathfrak a \in \mathfrak A^k}}$, up to a $\mu$-negligible set $N'_{\check \pi}$, such that
\begin{equation*}
\{\check Z^{',\ell}_\b, \check C^{',\ell}_\b\}_{\ell,\mathfrak b}\quad\text{is Lebesgue-regular},
\end{equation*}
and if $\bar\nu^{',\ell}_\b:=(\mathtt p_2)_\#\check \pi^{',\ell}_\b$, where $\check \pi^{',\ell}_\b$ is the conditional probability on the partition $\{\check Z^{',\ell}_\mathfrak b \times \mathbb{R}^d\}_{\ell,\mathfrak b}$, then each set $\check Z^{',\ell}_\b$ is $\Pi^f_{\mathtt c_{\mathbf D}}(\mu,\{\bar\nu^{',\ell}_\b\})$-cyclically connected, for all $\ell$, $\mathfrak b$.
\end{theorem}
Applying Theorem \ref{T_subpart_final} to the directed locally affine partition given by Theorem \ref{T_1}, one obtains immediately the following result. As in the case of Theorem \ref{T_1}, the second part of Point \eqref{Point_4_T_final} of the next theorem is a consequence of the precise analysis of the regions where the mass transport occurs.
\begin{theorem}
\label{T_final}
Let $\mu, \nu\in\mathcal P(\mathbb{R}^d)$ with $\mu\ll\mathcal L^d$ and let $\d{\cdot}$ be a convex norm in $\mathbb{R}^d$. Then, for all $\check \pi\in\Pi^{\mathrm{opt}}_{\d{\cdot}}(\mu,\nu)$ there exists a locally affine directed partition $\{\check Z^k_\mathfrak a,\check C^k_\mathfrak a\}_{\overset{k=0,\dots,d}{\mathfrak a\in\mathfrak{A}^k}}$ in $\mathbb{R}^d$ with the following properties:
\begin{enumerate}
\item for all $\mathfrak a \in \mathfrak{A}^k$ the cone $\check C^k_\mathfrak a$ is a $k$-dimensional extremal face of $\d{\cdot}$;
\item $\displaystyle{\mu \biggl(\mathbb{R}^d\setminus\underset{k,\mathfrak a}{\bigcup}\; \check Z^k_\mathfrak a \biggr) = 0}$;
\item the partition is \emph{Lebesgue-regular};
\item \label{Point_4_T_final} the disintegration $\displaystyle{\check \pi = \int \check \pi^k_\a\, dm(k,\mathfrak a)}$ w.r.t. the partition $\{\check Z^k_\a \times \mathbb{R}^d\}_{k,\mathfrak a}$ satisfies
\[
\check \pi^k_\mathfrak a \in \Pi^f_{\mathtt c_{C^k_\mathfrak a}} \big( \check \mu^k_\a, (\mathtt p_2)_\# \check \pi^k_\a \big),
\]
where $\displaystyle{\mu = \int \check \mu^k_\a\,dm(k,\mathfrak a)}$ is the disintegration w.r.t. the partition $\{\check Z^k_\a\}_{k,\mathfrak a}$, and moreover
\[
(\mathtt p_2)_\#\check \pi^k_\a \biggl( \check Z^k_\a \cup \biggl( \mathbb{R}^d \setminus \underset{(k',\mathfrak a') \not= (k,\mathfrak a)}{\bigcup} \check Z^{k'}_{\mathfrak a'} \biggr) \biggr) = 1;
\]
\item the partition
$\{\check Z^k_\a\}_{k,\mathfrak a}$ is $\Pi^f_{\mathtt c_{\bar{\mathbf D}}}(\mu,\{(\mathtt p_2)_\#\check \pi^k_\a\})$-cyclically connected.
\end{enumerate}
\end{theorem}
\begin{remark}
\label{R_remark_str_conv}
We note that the elements of the locally affine partition $\{\check Z^k_\mathfrak a,\check C^k_\mathfrak a\}_{\nfrac{k=1,\dots,d}{\mathfrak a \in \mathfrak{A}^k}}$ given by the above theorem have maximal linear dimension
\[
\max \big\{ k : \check Z^k_\mathfrak a \not= \emptyset \big\} \leq \max \big\{ \dim C : C \text{ extremal face of } \mathrm{epi} \d{\cdot} \big\}.
\]
In particular, if $D$ is strictly convex, the locally affine decomposition is made only of directed rays, and one recovers the results of \cite{Car:strictly} for strictly convex norms.
\end{remark}
In the case $\nu \ll \mathcal L^d$, the decomposition does not depends on the transference plan, as in the strictly convex case. In particular, we can say that it is universal, i.e. it is independent on the particular transference plan $\pi \in \Pi^\mathrm{opt}_{\d{\cdot}}(\mu,\nu)$ used.
\begin{theorem}
\label{T_final_nu}
Assume that $\nu\ll\mathcal L^d$. Then the directed locally affine partition of Theorem \ref{T_final} satisfies the following properties:
\begin{enumerate}
\item for all $\mathfrak a \in \mathfrak{A}^k$ the cone $\check C^k_\mathfrak a$ is a $k$-dimensional extremal face of $\d{\cdot}$;
\item[(2')] $\displaystyle{\mu \biggl(\mathbb{R}^d\setminus\underset{k,\mathfrak a}{\bigcup}\; \check Z^k_\mathfrak a \biggr) = \nu \biggl(\mathbb{R}^d\setminus\underset{k,\mathfrak a}{\bigcup}\; \check Z^k_\mathfrak a \biggr) = 0}$;
\item the partition is \emph{Lebesgue-regular};
\item[(4')] for all $\check \pi \in \Pi^{\mathrm{opt}}_{\d{\cdot}}(\mu,\nu)$, the disintegration $\displaystyle{\check \pi = \int \check \pi^k_\a\, dm(k,\mathfrak a)}$ w.r.t. the partition $\{\check Z^k_\a \times \mathbb{R}^d\}_{k,\mathfrak a}$ satisfies
\[
\check \pi^k_\mathfrak a \in \Pi^ f_{\mathtt c_{C^k_\mathfrak a}}(\check \mu^k_\a, \check \nu^k_\a),
\]
where $\displaystyle{\mu = \int \check \mu^k_\a\,dm(k,\mathfrak a)}$, $\displaystyle{\nu = \int \check \nu^k_\a\,dm(k,\mathfrak a)}$ are the disintegration of $\mu$, $\nu$ w.r.t. the partition $\{\check Z^k_\a\}_{k,\mathfrak a}$;
\item[(5')] $\{\check Z^k_\a\}_{k,\mathfrak a}$ is $\Pi^ f_{\mathtt c_{\bar{\mathbf D}}}(\mu,\nu)$-cyclically connected.
\end{enumerate}
\end{theorem}
In particular $\displaystyle{\check \pi \biggl( \underset{k,\mathfrak a}{\bigcup} \check Z^k_\a \times \check Z^k_\a \bigg)=1}$.
The main step in the proof of Theorem \ref{T_subpart_final} is the following
\begin{theorem}
\label{T_subpart_step}
Let $\{Z^k_\a, C^k_\a\}_{\nfrac{k = 0,\dots,d}{\mathfrak a \in \mathfrak A^k}}$ be a Lebesgue-regular directed locally affine partition in $\mathbb{R}^d$ and let $\mu$, $\nu$ be probability measures in $\mathcal P(\mathbb{R}^d)$ such that $\mu\ll\mathcal L^d$ and $\Pi^f_{\mathtt c_{\mathbf D}}(\mu,\nu)\neq\emptyset$.
Then, for all fixed $\check \pi \in \Pi^ f_{\mathtt c_{\mathbf D}}(\mu,\nu)$, there exists a directed locally affine subpartition $\{\check Z^\ell_{\mathfrak b}, \check C^\ell_{\mathfrak b}\}_{\nfrac{\ell = 0,\dots,d}{\mathfrak b \in \mathfrak B^\ell}}$ of $\{Z^k_\a, C^k_\a\}_{\nfrac{k = 0,\dots,d}{\mathfrak a \in \mathfrak A^k}}$, up to a $\mu$-negligible set $N_{\check \pi}$, such that
\begin{equation*}
\big\{ \check Z^\ell_\mathfrak b,\check C^\ell_\mathfrak b \big\}_{\ell,\mathfrak b} \quad \text{is Lebesgue-regular},
\end{equation*}
and setting $\check \nu^\ell_\mathfrak b := (\mathtt p_2)_\#\check \pi^\ell_\mathfrak b$, where $\check \pi^\ell_\mathfrak b$ is the conditional probability on the partition $\{\check Z^{\ell}_\mathfrak b \times \mathbb{R}^d\}_{\ell,\mathfrak b}$, then the sets
\begin{equation}
\label{E_subp_ell_cycl_intro}
\Bigl\{ \check Z^\ell_\mathfrak b:\,\check Z^\ell_\mathfrak b\subset Z^\ell_\mathfrak a\text{ for some $\mathfrak a\in\mathfrak{A}^\ell$}, \ell=1,\dots,d \Bigr\}
\end{equation}
form a $\Pi^ f_{\mathtt c_{\mathbf D}}(\mu,\{\check \nu^\ell_\mathfrak b\})$-cyclically connected partition.
\end{theorem}
Theorem \ref{T_subpart_step} allows to construct a locally directed affine subpartition $\{\check Z^\ell_{\mathfrak b}, \check C^\ell_{\mathfrak b}\}_{\nfrac{\ell = 0,\dots,d}{\mathfrak b \in \mathfrak B^\ell}}$ to a directed locally affine partition $\{Z^k_\a, C^k_\a\}_{\nfrac{k = 0,\dots,d}{\mathfrak a \in \mathfrak A^k}}$ such that the sets which do not lower their affine dimensions (i.e. for which $\check Z^\ell_\mathfrak b \subset Z^k_\a$ and $\ell = k$) are $\Pi^f_{\mathtt c_{\mathbf D}}(\mu,\{\check \nu^\ell_\mathfrak b\})$-cyclically connected.
Since the subpartition
\[
\big\{\check Z^\ell_{\mathfrak b}, \check C^\ell_{\mathfrak b} \big\}_{\nfrac{\ell = 0,\dots,d-1}{\mathfrak b \in \bar{\mathfrak B}^\ell}} \quad \text{such that if} \ \check Z^\ell_\mathfrak b \subset Z^k_\mathfrak a \ \text{then} \ \ell < k \ \text{(equivalently neglecting the sets of \eqref{E_subp_ell_cycl_intro})}
\]
is a Lebesgue-regular directed locally affine partition, and as a subpartition of $\{Z^k_\a,C^k_\a\}_{k,\mathfrak a}$ the index $\ell$ is decreasing of at least $1$ in each $Z^k_\a$, by a finite iterative argument one immediately obtains Theorem \ref{T_subpart_final}.
The proof of Theorem \ref{T_subpart_step} relies on nonstandard tools in measure theory --namely, the sufficient condition for uniqueness/optimality of transference plans based on the existence of suitable Borel linear preorders given in \cite{BiaCar}-- and on the existence of Lebesgue-regular directed locally affine partitions for one parameter families of graphs of Lipschitz functions w.r.t. convex norms (called \emph{$\mathtt c_{\tilde{\mathbf C}}$-Lipschitz foliations}), whose construction generalizes the one of the differential partition of Theorem \ref{T_1}.
We give now a brief scheme of the main steps of the proof of Theorem \ref{T_subpart_step}. Then, we will go on stating its consequences, ending with the solution of Monge's problem in Theorem \ref{T_Monge_final}.
First of all, one can reduce to study the finite cost transportation problem on directed locally affine partitions with fixed dimension $k$ and whose cones of directions are close to a given one, called \emph{$k$-directed sheaf sets} (see Section \ref{Ss_mapping_sheaf_to_fibration}, Definition \ref{D_sheaf_set}). Moreover, by a change of variables which preserves the characteristics of the optimal transportation problem $\Pi^f_{\mathtt c}(\mu,\{\bar\nu^k_\a\})$, one can assume that the sets $Z^k_\a$ of the sheaf set are contained in distinct parallel planes, thus studying the so called \emph{$k$-directed fibrations} with cones of directions $\tilde{\mathbf C}\subset\mathfrak{A}\times\mathcal C(k,\mathbb{R}^k)$ (see Definition \ref{D_fibration}). $\tilde{\mathbf C}(\mathfrak a)$ is the cone of directions of the region $Z^k_\a$.
To give an idea of how the subpartition is constructed, in this introduction we assume that $\mathfrak{A}^k=\{\mathfrak a\}$, so that the finite cost transportation problem on such $\tilde{\mathbf C}$-directed fibration is a finite transportation problem for a single $k$-dimensional cone cost in $\mathbb{R}^k$. In the paper, the variable $\mathfrak a\in\mathfrak{A}$ plays the role of a parameter and is kept in all the constructions and definitions in order to show that the sufficient measurability conditions w.r.t. $\mathfrak a$, which are needed in order to define global objects, are satisfied.
By the discussion made before Theorem \ref{T_subpart_final}, it is natural to fix a carriage $\Gamma\subset\bigl\{ \mathtt c_{\mathbf C^k_\a}<+\infty\bigr\}$ of some $\pi^k_\a\in\Pi^f_{\mathtt c_{C^k_\a}}(\mu^k_\a,\nu^k_\a)$ and to see whether the partition of $\mathbb{R}^k$ into $(\Gamma,\mathtt c_{C^k_\a})$-cycles satisfies our requirements. It turns out that in general this is not true, the first main reason being that not all the other transport plans are necessarily concentrated on its sets.
However --as proven for general cost functions $\mathtt c$ in \cite{BiaCar} in order to give very general sufficient conditions for uniqueness/optimality of transport plans-- a partition on which all the transport plans $\pi \in \Pi^\mathrm{opt}_{\mathtt c_{C^k_\a}}(\mu^k_\a,\nu^k_\a)$ are concentrated exists provided one can find a Borel linear preorder (i.e., a transitive relation such that every two points can be compared) which contains the set $\bigl\{\mathtt c_{C^k_\a}<+\infty\bigr\}$ (i.e, it is \emph{$\mathtt c_{C^k_\a}$-compatible} according to Definition \ref{D_compatible}) and extends the linear preorder $\preccurlyeq_{(\Gamma, \mathtt c_{C^k_\a})}$ whose equivalence classes are the $(\Gamma, \mathtt c_{C^k_\a})$-cycles (i.e., $x \preccurlyeq_{(\Gamma, \mathtt c_{C^k_\a})} y$ if there is a $(\Gamma,\mathtt c_{C^k_\a})$-axial path of finite cost connecting $y$ to $x$).
In Section \ref{Ss_gamma_order}, Theorem \ref{T_order_gamma}, we show how to construct such a preorder. The preorder will be denoted in the following by $\preccurlyeq_{\Gamma,\mathtt W^\Gamma}$, where $\mathtt W^\Gamma$ is related to the countable procedure to construct the Borel preorder (see Section \ref{S_cfibr_cfol}). Its equivalence classes turn out to be either families of graphs of $|\cdot|_{\mathbf D(\mathfrak a)^*}$-Lipschitz functions $\{\mathtt h_\mathfrak t^\pm(\mathfrak a)\}_{\mathfrak t\in\mathfrak T}$ --being $\mathrm{epi}\,|\cdot|_{\mathbf D(\mathfrak a)^*}=C^k_\a$-- or $k$-dimensional equivalence classes. Such families of sets are called $\mathtt c_{\tilde{\mathbf C}}$-Lipschitz foliations and are studied in Section \ref{S_foliations}. In particular, the finite cost transportation problem w.r.t. $\mathtt c_{C^k_\a}$ in $\mathbb{R}^k$ reduces to a family of finite cost transportation problems w.r.t. $\mathtt c_{C^k_\a}$ on the sets of this $\mathtt c_{\tilde{\mathbf C}(\mathfrak a)}$-Lipschitz foliation.
The equivalence classes $\{\check Z^k_\mathfrak b\}_{\mathfrak b}$ which do not lower the affine dimension of the $Z^k_\a$ are connected by $(\Gamma, \mathtt c_{C^k_\a})$-cycles, up to a $\mu$-negligible set, and then in principle they are candidate to be the $k$-dimensional sets of Theorem \ref{T_subpart_step}. However, they are not necessarily connected by $(\Gamma', \mathtt c_{C^k_\a})$-cycles for the other carriages $\Gamma'$ of plans in $\Pi^f_{\mathtt c_{C^k_\a}}(\mu^k_\a,\nu^k_\a)$. In fact, changing $\Gamma$, the Borel preorder $\preccurlyeq_{\Gamma,\mathtt W^\Gamma}$ also varies. Hence we need to use an abstract result on measure theory \cite{BiaCar} (recalled here in Appendix \ref{A_minimal_equivalence}), assuring that there is a minimal Borel linear preorder among the ones of the type $\preccurlyeq_{\Gamma,\mathtt W^\Gamma}$: for this one, the sets which do not lower the dimension of the $Z^k_\a$ and are of positive $\mu^k_\a$-measure are $\Pi^f_{\mathtt c_{C^k_\a}}(\mu^k_\a,\nu^k_\a)$-cyclically connected (see Theorem \ref{T_cfibrcfol}). Notice that this $\Pi^f_{\mathtt c_{C^k_\a}}(\mu^k_\a,\nu^k_\a)$-cyclically connectedness property can be indeed interpreted as a minimality or ``indecomposability'' property of the new $k$-dimensional sets.
As for the finite cost transportation problem for the cost $\mathtt c_{C^k_\a}$ on the classes of the minimal equivalence relation which are graphs of $|\cdot|_{\mathbf D(\mathfrak a)^*}$-Lipschitz functions $\{\mathtt h_\mathfrak t(\mathfrak a)\}_{\mathfrak t\in\mathfrak T}$, one uses the same tools as in the proof of Theorem \ref{T_1} to show the existence of a \emph{differential} locally affine partition $\{\check Z^\ell_\mathfrak b,\check C^\ell_\mathfrak b\}_{\ell<k,\mathfrak b}$ on which the transportation problem decomposes (see Section \ref{S_foliations}). Indeed, the sets of such partition which are contained in $\mathrm{graph}\,\mathtt h_\mathfrak t(\mathfrak a)$ are constructed as the sets of the differential partition for the Kantorovich potential $\psi$: the only difference now is that one has to take care of the measurability of these sets w.r.t. the parameters $\mathfrak a,\mathfrak t$.
The only missing point in the proof that the union of the differential partition and of the $k$-dimensional equivalence classes $\{\check Z^\ell_\mathfrak b,\check C^\ell_\mathfrak b\}_{\ell,\mathfrak b}$ satisfies the conclusions of Theorem \ref{T_subpart_step} is then the Lebesgue-regularity.
Notice that now the directed locally affine partition is obtained applying the same reasoning as in Theorem \ref{T_1} but for a family of norm-Lipschitz graphs depending on a continuous parameter. Hence one would be tempted to deduce the Lebesgue-regularity property first disintegrating the Lebesgue measure $\mathcal L^k$ on such graphs and then using the cone approximation property (as for the Kantorovich potential) for each conditional measure of $\mathcal L^k$ on a single graph. However, as we will show in a counterexample (see Remark \ref{R_not_gener_potential}), this is not possible because in general the conditional measures of $\mathcal L^k$ on a family of Lipschitz graphs might to be absolutely continuous w.r.t. the $(k-1)$-dimensional Hausdorff measure on the graphs on which they are concentrated.
In fact, the Lebesgue-regularity property for the sets $\{\check Z^\ell_\mathfrak b,\check C^\ell_\mathfrak b\}_{\ell,\mathfrak b}$ follows by the fact that the Lipschitz graphs of $\{\mathtt h_\mathfrak t(\mathfrak a)\}$ are the equivalence classes of a $\mathtt c_{C^k_\a}$-compatible Borel linear preorder on which all the transport plans in $\Pi^f_{\mathtt c_{C^k_\a}}(\mu^k_\a,\nu^k_\a)$ are concentrated. Indeed, by the uniqueness theorem stated in \cite{BiaCar} one can prove the cone approximation property for the subpartition. The procedure is similar to the procedure followed in the case a single potential $\psi$ is present: however, the convergence of the cone approximating vector fields is now due to the uniqueness of a suitable transference plan (see Section \ref{S_cone_approx_folia}).
As discussed before, Theorem \ref{T_final} gives as an application the possibility to construct optimal potentials w.r.t. secondary cost functions such as $\mathtt c_2$ \eqref{0_c2} on each set of the partition $\check Z^k_\a$. In the case in which the secondary cost function is obtained by minimizing the original transport problem w.r.t. another convex norm $|\cdot|_{(D')^*}$, one obtains a refinement of the directed locally affine partition of Theorem \ref{T_final} with cones of directions given by intersections of extremal faces of $|\cdot|_{(D')^*}$ and $\d{\cdot}$.
More precisely, let $|\cdot|_{(D')^*}$ be a convex norm with unit ball $D'$, and consider the secondary minimization problem
\begin{equation}
\label{E_seconry_min_prob}
\min \bigg\{ \int |y - x|_{(D')^*}\, d\pi(x,y):\, \pi \in \Pi^{\mathrm{opt}}_{|\cdot|_{D^*}}(\mu,\nu) \bigg\}.
\end{equation}
If $\check \pi$ is a minimizer of the above problem, by the fact that $\check \pi$ is also a minimizer of
\[
\int \mathtt c_{D,D'}(x,y)\,d \pi(x,y), \qquad \mathtt c_{D,D'}(x,y) :=
\begin{cases}
|y-x|_{(D')^*} & x \in \check Z^k_\mathfrak a, y-x \in \check C^k_\mathfrak a, \crcr
+\infty & \text{otherwise},
\end{cases}
\]
and that each $\check Z^k_\a$ is $\Pi^f_{\check C^k_\a}(\mu^k_\a,\nu^k_\a)$-cyclically connected, Proposition \ref{P_second_cost} yields that in each $\check Z^k_\mathfrak a$ there exists a potential pair $\phi^k_\a$, $\psi^k_\a$, and since $\mathtt c_{D,D'}$ satisfies the triangle inequality we can take $\phi^k_\a = - \psi^k_\a$. By the existence of such a potential, restricting then to a single set $\check Z^k_\mathfrak a$ one can prove as in the proof of Theorem \ref{T_1} the existence of a directed locally affine partition $\{\check Z^{k,\ell}_{\mathfrak a,\mathfrak b},\check C^{k,\ell}_{\mathfrak a,\mathfrak b}\}_{\nfrac{\ell = 0,\dots,k}{\mathfrak b \in \mathfrak B^{k,\ell}_\mathfrak a}}$. In the resulting statement one has to replace the ambient space $\mathbb{R}^d$ with $\check Z^k_\a$, the measure $\mathcal L^d$ with $\mathcal H^k\llcorner_{\check Z^k_\a}$, the marginals $\mu$, $\nu$ with $\mu^k_\a$, $\nu^k_\a$ and the cost $\d{\cdot}$ with
\begin{equation}
\label{E_cka_str_intro}
\mathtt c^k_\mathfrak a := \mathtt c_{D,D'} \llcorner_{\check Z^k_\mathfrak a \times \mathbb{R}^d} =
\begin{cases}
|y-x|_{(D')^*} & y-x \in \check C^k_\mathfrak a, \crcr
+\infty & \text{otherwise}.
\end{cases}
\end{equation}
More precisely, we obtain the following theorem.
\begin{theorem}
\label{T_final_seconry}
Let $\mu, \nu\in\mathcal P(\mathbb{R}^d)$ with $\mu\ll\mathcal L^d$ and let $\check \pi$ be an optimal transport plan for the problem \eqref{E_seconry_min_prob}. Then there exists a locally affine directed partition $\{\check Z^{k,\ell}_{\mathfrak a,\mathfrak b},\check C^{k,\ell}_{\mathfrak a,\mathfrak b}\}_{\nfrac{k=0,\dots,d,\mathfrak a\in\mathfrak{A}^k}{\ell = 0,\dots,k,\mathfrak b \in \mathfrak B^{k,\ell}_\mathfrak a}}$ in $\mathbb{R}^d$ with the following properties:
\begin{enumerate}
\item for all $k$, $\mathfrak a \in \mathfrak{A}^k$, the cone $\check C^{k,\ell}_{\mathfrak a,\mathfrak b}$ is an $\ell$-dimensional extremal face of the cost $\mathtt c^k_\a$ given by \eqref{E_cka_str_intro}, i.e. the intersection of a $k$-dimensional face of $\d{\cdot}$ with an extremal face of $|\cdot|_{(D')^*}$;
\item $\displaystyle{\mu \biggl( \mathbb{R}^d \setminus \underset{k,\mathfrak a,\ell,\mathfrak b}{\bigcup} \check Z^{k,\ell}_{\mathfrak a,\mathfrak b} \biggr)=0}$;
\item the partition is \emph{Lebesgue-regular};
\item the disintegration $\displaystyle{\check \pi = \int \check \pi^{k,\ell}_{\mathfrak a,\mathfrak b}\, dm(k,\mathfrak a,\ell,\mathfrak b)}$ w.r.t. the partition $\{\check Z^{k,\ell}_{\mathfrak a,\mathfrak b} \times \mathbb{R}^d\}_{k,\mathfrak a,\ell,\mathfrak b}$ satisfies
\[
\check \pi^{k,\ell}_{\mathfrak a,\mathfrak b} \in \Pi^f_{\mathtt c_{C^{k,\ell}_{\mathfrak a,\mathfrak b}}} \big( \check \mu^{k,\ell}_{\mathfrak a,\mathfrak b}, \big\{ (\mathtt p_2)_\# \check \pi^{k,\ell}_{\mathfrak a,\mathfrak b} \big\} \big),
\]
where $\displaystyle{\mu = \int \check \mu^{k,\ell}_{\mathfrak a,\mathfrak b}\,dm(k,\mathfrak a,\ell,\mathfrak b)}$ is the disintegration w.r.t. the partition $\{\check Z^{k,\ell}_{\mathfrak a,\mathfrak b}\}_{k,\mathfrak a,\ell,\mathfrak b}$, and moreover
\[
(\mathtt p_2)_\#\check \pi^{k,\ell}_{\mathfrak a,\mathfrak b} \biggl( \check Z^{k,\ell}_{\mathfrak a,\mathfrak b} \cup \biggl( \mathbb{R}^d \setminus \underset{(k',\mathfrak a',\ell',\mathfrak b') \not=(k,\mathfrak a,\ell,\mathfrak b)}{\bigcup} \check Z^{k',\ell'}_{\mathfrak a',\mathfrak b'} \biggr) \biggr) = 1;
\]
\item the partition
$\{\check Z^{k,\ell}_{\mathfrak a,\mathfrak b}\}_{k,\mathfrak a,\ell,\mathfrak b}$ is $\Pi^f_{\mathtt c_{\bar{\mathbf D}}}(\mu,\{(\mathtt p_2)_\#\check \pi^{k,\ell}_{\mathfrak a,\mathfrak b}\})$-cyclically connected.
\end{enumerate}
\end{theorem}
A completely similar extension can be given to Theorem \ref{T_final_nu}.
A particular case is when each extremal face of $|\cdot|_{(D')^*}$ is contained in an extremal face of $|\cdot|_{D^*}$: in this case condition (1) becomes
\begin{enumerate}[{\it (1')}]
\itemindent .55cm
\item \label{Point_1_prime} {\it for all $k$, $\mathfrak a \in \mathfrak{A}^k$, the cone $\check C^{k,\ell}_{\mathfrak a,\mathfrak b}$ is an $\ell$-dimensional extremal face of $|\cdot|_{(D')^*}$.}
\end{enumerate}
The only difference w.r.t. Theorem \ref{T_final} is that now $\check \pi$ is a minimum for the secondary minimization problem \eqref{E_seconry_min_prob}, not a transference plan in $\Pi^{\mathrm{opt}}_{|\cdot|_{(D')^*}}(\mu,\nu)$.
The case (1') above happens if for example $|\cdot|_{(D')^*}$ is strictly convex, so that the $\check Z^{k,\ell}_{\mathfrak a,\mathfrak b}$ are now directed segments, i.e. $\ell = 1$. By the standard analysis on transportation problems in 1-d, and the measurable dependence on $k$, $\mathfrak a$, $\mathfrak b$, the existence of an optimal transport map $\mathtt T$ for the Monge problem \eqref{E_transpo_Sud} follows as a simple corollary. In particular, the restriction $\mathtt T \llcorner_{\check Z^{k,1}_{\mathfrak a,\mathfrak b}}$ is a monotone increasing map in the direction of $\check C^{k,1}_{\mathfrak a,\mathfrak b}$ on $\mathrm{aff}\, \check Z^{k,1}_{\mathfrak a,\mathfrak b}$, for all $k$, $\mathfrak a$, $\mathfrak b$.
\begin{theorem}
\label{T_Monge_final}
Let $\mu, \nu\in\mathcal P(\mathbb{R}^d)$, $\mu\ll\mathcal L^d$. Then, there exists an optimal transport map $\mathtt T$ for the Monge problem \eqref{E_transpo_Sud}.
\end{theorem}
\subsection{Structure of the paper}
\label{Ss_struct_Sud}
The paper is organized as follows.
In Section \ref{S_setting} we collect the main notations, definitions and the basic tools we will need in the paper. After recalling some standard definitions of commonly used sets and $\sigma$-algebras, we introduce some notations for functions and multifunctions in Section \ref{Ss_souslin_multifunction}. The basic tools in convex analysis as well as the definitions of the Polish spaces $\mathcal A(k,V)$ made of $k$-dimensional affine subspaces of the affine space $V \subset \mathbb{R}^d$ and the Polish space $\mathcal C(k,V)$ made of non degenerate $k$-dimensional cones of a vector space $V \subset \mathbb{R}^d$ are listed in Section \ref{Ss_intro_affine_subspaces_cones}. \\
The fundamental tools on measure theory and the disintegration theorem are recalled in Section \ref{Ss_measure_disintegration}, while the definition of the optimal transportation problem with the classical sufficient conditions for optimality of transference plans are listed in Section \ref{Ss_transference_plans}. The key analysis on transportation problems for which no potential is available and the role of Borel linear preorders is presented in Section \ref{Ss_linear_pre_unique_transf}. In particular, we define $\Pi^f_{\mathtt c}(\mu,\nu)$-cyclically connected partitions (see Definition \ref{D_pimunuconn}), we state Proposition \ref{P_second_cost} yielding the existence of a family of potentials on the elements of a partition for secondary costs under the assumption of $\Pi^f_{\mathtt c}(\mu,\nu)$-cyclical connectedness, and we show in Theorem \ref{T_A2} \cite{BiaCar} that whenever a Borel linear preorder is $\mathtt c$-compatible and extends a $\preccurlyeq_{(\Gamma,\mathtt c)}$-preorder, then all the transport plans are concentrated on its equivalence classes.
In Section \ref{S_conetransport} we analyze the optimal transportation for cone costs of the form \eqref{E_cost_all_cone_intro}. In Section \ref{Ss_convex_norm_cone} we show the equivalence between optimal transportation problems in $\mathbb{R}^d$ with cost $\d{\cdot}$ and marginals $\mu,\nu \in \mathcal P(\mathbb{R}^d)$ and optimal transportation problems in $\mathbb{R}^{d+1}$ with cost $\mathtt c_{\mathrm{epi}\,\d{\cdot}}$ and marginals $\hat \mu, \hat \nu \in \mathcal P(\mathbb{R}^{d+1})$, where the measures $\hat \mu$, $\hat \nu$ are supported on a $\d{\cdot}$-Lipschitz graph $\mathrm{graph}\,\psi$ and $\mu = (\mathtt p_{\mathbb{R}^d})_\# \hat \mu$, $\nu = (\mathtt p_{\mathbb{R}^d})_\# \hat \mu$. \\
In Section \ref{Ss_partitions_intro} we generalize the single cone cost transportation problem to the transportation problem on a directed locally affine partition. Here we introduce also the notion of initial and final points of a directed locally affine partition and the notion of conditional second marginals, as well as an example of their dependence w.r.t. the transference plan (Example \ref{Ex_2ndmarg}) and a special case where the conditional second marginals correspond to the disintegration of $\nu$ (Proposition \ref{P_dispiani_2}). \\
A standard covering argument allows to decompose a directed locally affine partition into $k$-directed \emph{sheaf sets}, i.e. directed locally affine partitions whose components $Z^k_\a$ and cones $C^k_\a$ are close to a given reference plane $V^k$ and cone $C^k$, and their projection on $V^k$ contains a given open $k$-dimensional cube, Proposition \ref{P_countable_partition_in_reference_directed_planes} and Definition \ref{D_sheaf_set}. This allows to map these sets into \emph{directed fibrations}, where the $Z^k_\a$ are contained in the planes $\{\mathfrak a\} \times \mathbb{R}^k$, $\mathfrak a \in \mathbb{R}^{d-k}$ (Proposition \ref{P_map_sheaf_set_into_fibration}). In this case the transport problem splits into a family of transport problems, each one moving mass on a plane of the form $\{\mathfrak a\} \times \mathbb{R}^k$ and with cost
\begin{equation}
\label{E_cone_C_a_intro}
\mathtt c_{\tilde{\mathbf C}(\mathfrak a)}(w,w') := \mathbbm 1_{\tilde{\mathbf C}(\mathfrak a)}(w'-w),
\end{equation}
where $\mathfrak a \mapsto \tilde{\mathbf C}(\mathfrak a)$ a $\sigma$-compact map with values in $\mathcal C(k,\mathbb{R}^k)$: since $k$ is fixed on a fibration, we can skip it in order to simplify the notation. \\
The final part of the section shows that the mapping of a sheaf set into a fibration preserves the key structures of the optimal transportation problem needed in the proofs of Theorems \ref{T_1} and \ref{T_subpart_final}, and thus allows us to work from now onwards on fibrations.
In Section \ref{S_foliations} we present a technique in order to find the so called \emph{differential} directed locally affine subpartition of a given \emph{$\mathtt c_{\tilde{\mathbf C}}$-Lipschitz foliation} of a $\tilde{\mathbf C}$-directed fibration. The reason why we introduce and study {$\mathtt c_{\tilde{\mathbf C}}$-Lipschitz foliations} is that they are the natural generalization of the notions of graphs of cone-Lipschitz functions --as the Kantorovich potential $\psi$-- and equivalence classes of a $\mathtt c_{C^k_\a}$-compatible Borel linear preorder (see Proposition \ref{P_ex_fol}). For the terminology used to briefly list the content of this section we refer also to the discussion made in this introduction at the beginning of Section \ref{Ss_main_Sud}.\\
Due to the results of the previous section, when the differential subpartition is mapped back from the $\mathtt c_{\tilde{\mathbf C}}$-Lipschitz foliation to the $k$-directed sheaf set, one obtains subpartitions of the sheaf sets covering a given directed locally affine partition, and thus we have a method which yields a subpartition of a given directed locally affine partition. \\
In Section \ref{Ss_cone_lipschitz_graph} we first analyze the simplest example of $\mathtt c_{\tilde{\mathbf C}}$-Lipschitz foliation: a \emph{$\mathtt c_{\tilde{\mathbf C}(\mathfrak a)}$-Lipschitz graph}, namely a graph of a $|\cdot|_{\mathbf D(\mathfrak a)^*}$-Lipschitz function (with $\tilde{\mathbf C}(\mathfrak a)=\mathrm{epi}\,|\cdot|_{\mathbf D(\mathfrak a)^*}$) in a fibration consisting of a single fiber $\{\mathfrak a\} \times \mathbb{R}^k$, whose \emph{super/subdifferential} satisfy the \emph{completeness property} \eqref{0_compl1}-\eqref{0_compl2} (as the graph of the Kantorovich potential for the cost $\mathtt c_{\mathrm{epi}\,\d{\cdot}}$).\\
In Section \ref{Ss_cone_lipschitz} we consider general $\mathtt c_{\tilde{\mathbf C}}$-Lipschitz foliations, namely partitions of $\mathfrak{A}\times\mathbb{R}^k$ whose sets are contained in $\{\mathfrak a\}\times\mathbb{R}^k$ as $\mathfrak a$ varies in $\mathfrak{A}$ and are given by collections of complete $\mathtt c_{\tilde{\mathbf C}(\mathfrak a)}$-Lipschitz graphs (see Proposition \ref{P_fol_char}). We extend to these families of sets the notion of super/subdifferential (see Definition \ref{D_partial_theta}) and in Section \ref{Ss_regu_resi_set} we show that its completeness and transitivity properties permit to select regions called \emph{forward/backward regular set} and \emph{regular set}.\\
These regions are respectively partitioned in the so called \emph{super/subdifferential partition} and \emph{differential partition} in Section \ref{Ss_partition_transport_set} (see Theorem \ref{T_partition_E+-} and Corollary \ref{C_v}).\\
In Section \ref{Ss_analysis_residual_set} we analyze the \emph{residual set}, namely the complementary of the regular set, and characterize it as the union of the initial/final points of the {super/subdifferential partition} (Theorem \ref{T_partition_E}).\\
In Section \ref{Ss_optim_folia} we give a descriptive characterization of the support of the optimal transportation problem on a $\mathtt c_{\tilde{\mathbf C}}$-Lipschitz foliation in terms of the forward/backward regular set and initial/final points.
Section \ref{S_disintechnique} concerns the problem of Lebesgue-regularity of the disintegration on directed locally affine partitions. The main property which allows to deduce the regularity is the \emph{cone approximation property}.\\
First we consider $1$-directed sheaf sets made of segments whose projection on the line generated by the reference cone is a given interval. These particular sheaf sets are called \emph{model sets of directed segments} (see Section \ref{Ss_model_dir_segm}). In the case of strictly convex norms it is sufficient to analyze this special case.\\
The analysis is then extended to \emph{$k$-dimensional model sets}, namely $k$-directed sheaf sets whose projection on the $k$-dimensional plane generated by the reference cone is a given rhomboid (Section \ref{Ss_k_dim_model_set}). In this case, the \emph{cone approximation property} refers to the cone approximation property of any of its $1$-dimensional slices, the latter being sections of a $k$-dimensional model set with $(d-k+1)$-dimensional planes transversal to the reference plane (see Definition \ref{D_1_dim_slice_sheaf}): by transversality, on each of these planes the $k$-dimensional model set becomes a model set of directed segments. \\
Next the analysis is extended to $k$-dimensional sheaf sets (Section \ref{Ss_k_dim_sheaf}). The main observation is that one can partition the sheaf set into countably many $k$-dimensional model sets (Theorem \ref{T_one_d_slicing_FC}). \\
Finally, the property of approximation by cone vector fields also for initial/final points yields the Lebesgue-negligibility of the initial/final points by means of a technique developed first in \cite{celper:euler}, and then extended in \cite{BiaGlo,Car:strictly,CarDan} (see Section \ref{Ss_neglig_init_fin}).
At this point all the techniques needed in order to prove Theorem \ref{T_1} are presented, and its proof is done in Section \ref{S_theorem_1_proof}. Indeed, in Section \ref{S_foliations} we develop a technique to find directed locally affine subpartitions by means of $\mathtt c_{\tilde{\mathbf C}}$-Lipschitz foliations, and the graph of the potential $\psi$ is in particular a $\mathtt c_{\tilde{\mathbf C}(\mathfrak a)}$-Lipschitz graph. The only point which remains to be proved is that the disintegration of the Lebesgue measure is regular, which is a consequence of the cone approximation property. The section is thus devoted to the proof of the cone approximation property for cone-Lipschitz graphs (Theorem \ref{T_cone_graph}).
Let $\tilde {\mathbf D}=\{Z^k_\a,C^k_\a\}$ be a directed fibration with the associated transportation problem; as said before, we assume that $\Pi^f_{\mathtt c_{\tilde{\mathbf{D}}}}(\mu,\nu) \not= \emptyset$. In Section \ref{S_cfibr_cfol} we show how to
further partition $\{Z^k_\a,C^k_\a\}$ into a ${\mathtt c_{\tilde{\mathbf{C}}}}$-Lipschitz foliation, whose $k$-dimensional sets satisfy the assumptions of Theorem \ref{T_subpart_step} (see Theorem \ref{T_cfibrcfol}). The key results are stated in Theorem \ref{T_order_gamma} and Proposition \ref{P_equiv_coun}. The sets of this ${\mathtt c_{\tilde{\mathbf{C}}}}$-Lipschitz foliation are given by the equivalence classes of a ${\mathtt c_{\tilde{\mathbf{C}}}}$-compatible linear preorder on which all the transport plans in $\Pi^f_{\mathtt c_{\tilde{\mathbf{D}}}}(\mu,\nu)$ are concentrated (called \emph{$(\mathtt c_{\tilde{\mathbf{C}}},\mu,\nu)$-compatible} linear preorder in Definition \ref{D_cmunucomp}), as anticipated after the statement of Theorem \ref{T_subpart_step}.
In Section \ref{S_cone_approx_folia} we prove the cone approximation property for the differential partition of a $\mathtt c_{\tilde{\mathbf C}}$-Lipschitz foliation whose sets are given by equivalence classes of a Borel $({\mathtt c_{\tilde{\mathbf{C}}}}, \mu,\nu)$-compatible linear preorder. Since we do not have a potential, we need to use the uniqueness theorem of the linear preorder (Theorem \ref{T_coneappr_fol}): as a corollary, one immediately obtains the Lebesgue-regularity of the disintegration (Corollary \ref{C_infinnegl}).\\
The section is concluded which an example (Remark \ref{R_not_gener_potential}) which shows that this result cannot be deduced as a consequence of the analysis of Section \ref{S_theorem_1_proof}, even if the level sets of $\theta$ are Lipschitz graphs. In fact, the disintegration of the Lebesgue measure on a Lipschitz foliation is in general not absolutely continuous w.r.t. the natural Hausdorff measures on the level sets: we show an example where the level sets of the function $\theta : [0,1]^2 \rightarrow [0,1]$ generating the foliation are $C^\infty$, nevertheless the disintegration of $\mathcal L^2 \llcorner_{[0,1]^2}$ on the level sets of $\theta$ have Dirac masses.
The proof of Theorems \ref{T_subpart_step} and \ref{T_final_nu} are done in Section \ref{S_proff_main_Th}, and they are obtained as direct consequences of the results proved so far.
Finally, in Appendix \ref{A_minimal_equivalence} we present one of the main result of \cite{BiaCar} about the minimality of equivalence relations and prove a key consequence used in our proof, Corollary \ref{C_constant_for_minimal_equivalence}.
In Appendix \ref{A_appendix_nota} we collect the notations used in this paper.
\section{General notations and definitions}
\label{S_setting}
As standard notation, we will write $\mathbb{N}$ for the natural numbers, $\mathbb{N}_0 = \mathbb{N} \cup \{0\}$, $\mathbb{Q}$ for the rational numbers and $\mathbb{R}$ for the real numbers. The sets of positive rational and real numbers will be denoted by $\mathbb{Q}^+$ and $\mathbb{R}^+$ respectively.
The $d$-dimensional real vector space will be denoted by $\mathbb{R}^d$. $B^d(x,r)$ is the open unit ball in $\mathbb{R}^d$ with center $x$ and radius $r>0$ and $\mathbb S^{d-1}$ is the $(d-1)$-dimensional unit sphere. The scalar product of two vectors $x,y \in \mathbb{R}^d$ will be denoted by $x \cdot y$, and the Euclidean norm of $x \in \mathbb{R}^d$ by $|x| = \sqrt{x \cdot x}$. To avoid the analysis of different cases when parameters are in $\mathbb{R}^k$ for $k=1,\dots,d$ or in $\mathbb{N}$, we set $\mathbb{R}^0 := \mathbb{N}$.
We denote the first infinite ordinal number by $\omega$.
Given a set $X$, $\mathtt{P}(X)$ is the power set of $X$. The closure of a set $A$ in a topological space $X$ will be denoted by $\mathrm{clos}\,A$, and its interior by $\mathrm{int}\,A$. If $A \subset Y \subset X$, then the relative interior of $A$ in $Y$ is $\inter_{\mathrm{rel}} A$: in general the space $Y$ will be clear from the context. The topological boundary of a set $A$ will be denoted by $\partial A$, and the relative boundary is $\partial_\mathrm{rel} A$.
If $A$, $A'$ are subsets of a real vector space, we will write
\begin{equation}
\label{E_vector_sum}
A + A' := \big\{ x + x':\, x \in A, x' \in A' \big\}.
\end{equation}
If $T \subset \mathbb{R}$, then we will write
\begin{equation}
\label{E_product_scalar_vector_def}
T A := \big\{ t x:\, t \in T, x \in A \big\}.
\end{equation}
In particular $A - A' = A + (-A')$.
If $\prod_i X_i$ is the product space of the spaces $X_i$, we will denote the projection on the $\bar i$-component either as $\mathtt p_{\bar i}$ or $\mathtt p_{x_{\bar i}}$ or $\mathtt p_{X_{\bar i}}$: in general no ambiguity will occur.
\subsection{Functions and multifunctions}
\label{Ss_souslin_multifunction}
A multifunction $\mathbf f$ will be considered either as a map $\mathbf f:\,X \subset \mathrm{dom}\,\mathbf f \rightarrow \mathtt{P}(Y)$ or as a set $\mathbf f \subset X \times Y$. The set $\mathrm{dom}\, \mathbf f$ is called the \emph{domain} of $\mathbf f$. For every $x\in\mathrm{dom}\,\mathbf f$ we will write
\[
\mathbf f(x) = \big\{ y \in Y: (x,y) \in \mathbf f \big\}.
\]
The inverse of $\mathbf f$ will be denoted by
\begin{equation}
\label{E_inverse_multi_function}
\mathbf f^{-1} = \big\{ (y,x) \in Y \times X : (x,y) \in \mathbf f \big\}.
\end{equation}
Similarly, if $A\subset X\times Y$, then $A^{-1}:=\{(y,x):\,(x,y)\in A\}$.
In the same spirit, we will often not distinguish between a single valued function $\mathtt f$ and its graph, denoted by $\mathrm{graph}\, \mathtt f$. We say that the function $\mathtt f$ (or the multifunction $\mathbf f$) is \emph{$\sigma$-continuous} if the set $\mathrm{graph}\,\mathtt f$ (or $\mathbf f\subset X\times Y$) is $\sigma$-compact. Note that we do not require its domain to be the entire space.
If $\mathtt f$, $\mathtt g$ are two functions, their composition will be denoted by $\mathtt g \circ \mathtt f$, with domain $\mathtt f^{-1}(\mathrm{dom}\,\mathtt g)$.
If $\mathtt f:X\rightarrow Y$, $\mathtt g:X\rightarrow Z$, then the product map is denoted by $\mathtt f\times \mathtt g:X\rightarrow Y\times Z$.
The epigraph of a function $\mathtt f : X \rightarrow \mathbb{R}$ is the set
\begin{equation}
\label{E_epigraph_function}
\mathrm{epi}\,\mathtt f := \big\{ (x,t)\in X\times\mathbb{R} : \mathtt f(x) \leq t \big\}.
\end{equation}
The identity map will be written as $\mathbb I$, the characteristic function of a set $A$ will be denoted by
\begin{equation}
\label{E_char_funct_set_A}
\chi_A(x) :=
\begin{cases}
1 & x \in A, \crcr
0 & x \notin A,
\end{cases}
\end{equation}
and the indicator function of a set $A$ is defined by
\begin{equation}
\label{E_indicator_function_A}
\mathbbm 1_A(x) :=
\begin{cases}
0 & x \in A, \crcr
+\infty & x \notin A.
\end{cases}
\end{equation}
\subsection{Affine subspaces, convex sets and norms}
\label{Ss_intro_affine_subspaces_cones}
For $k,k',d \in \mathbb{N}$, $k' \leq k \leq d$, define $\mathcal G(k,\mathbb{R}^d)$ to be the set of $k$-dimensional subspaces of $\mathbb{R}^d$ and let $\mathcal A(k,\mathbb{R}^d)$ be the set of $k$-dimensional affine subspaces of $\mathbb{R}^d$. If $V \in \mathcal A(k,\mathbb{R}^d)$, we define $\mathcal A(k',V) \subset \mathcal A(k',\mathbb{R}^d)$ to be the set of $k'$-dimensional affine subspaces of $V$. We also denote by $\mathtt p_V:\mathbb{R}^d\rightarrow V$ the projection map.
If $A \subset \mathbb{R}^d$, then define its \emph{affine span} as
\begin{equation}
\label{E_affine_span_set}
\mathrm{aff}\,A := \bigcap_{\nfrac{V\in\mathcal A(k,\mathbb{R}^d)}{\,k\in\mathbb{N},\,A\subset V}} V,
\end{equation}
and its convex hull $\mathrm{conv}\,A$ as the set given by the intersection of all convex sets of $\mathbb{R}^d$ containing $A$. Given $k$ vectors $\{\mathtt e_1,\dots,\mathtt e_k\}\subset\mathbb{R}^d$, their linear span $\mathrm{aff}\,\{\mathtt e_1,\dots,\mathtt e_k,0\}$ is denoted by $\langle\mathtt e_1,\dots,\mathtt e_k\rangle$, and the orthogonal space to $V\in\mathcal G(k,\mathbb{R}^d)$ is denoted by $V^\perp$. Notice that
\[
\mathrm{aff}\,A=\mathrm{conv}\,A+\mathbb{R}(\mathrm{conv}\,A-\mathrm{conv}\,A) \qquad \text{and} \qquad \mathrm{aff}\,A\in\underset{k\in\mathbb{N}}{\bigcup}\mathcal A(k,\mathbb{R}^d),
\]
unless $A$ consists of a single point. If we set by convention $\mathcal A(0,\mathbb{R}^d)=\mathbb{R}^d$, then the above formula holds also when $A$ is a singleton.
The \emph{linear dimension} of an affine subspace $V$ is denoted by $\dim\,V$, and we set accordingly $\dim\{x\}=0$ for all $x\in\mathcal A(0,\mathbb{R}^d)$.
If $A$ is convex, then its relative interior in $\mathrm{aff}\, A$, denoted by $\inter_{\mathrm{rel}} A$, is nonempty and $A\subset\mathrm{clos}\,\inter_{\mathrm{rel}} A$. Hence we define $\dim\,A:=\dim \mathrm{aff}\,A$.
An \emph{extremal face} of a convex set $A\subset\mathbb{R}^d$ is by definition any convex set $A'\subset A$ such that
\[
\inter_{\mathrm{rel}}\,[x,y]\cap A'\neq\emptyset\quad\Rightarrow\quad[x,y]\in A',\quad \forall\,x,y\in A.
\]
A convex set $C\subset\mathbb{R}^d$ is a \emph{convex cone} if
\[
C = \{0\} \cup \mathbb{R}^+C.
\]
In particular, $0$ is called the \emph{vertex} of the convex cone. For all $k\in\{1,\dots,d\}$, we let $\mathcal C(k,\mathbb{R}^d)$ to be the set of closed $k$-dimensional convex cones in $\mathbb{R}^d$ which are \emph{nondegenerate}, meaning that
\[
C\in\mathcal C(k,\mathbb{R}^d)\quad\Rightarrow\quad C\setminus\{0\}\subset \inter_{\mathrm{rel}}\, H_C
\]
for some $k$-dimensional half-plane $H_C \subset \mathbb{R}^d$.
The extremal faces of a convex cone are still convex cones called \emph{extremal cones}.
If $D$ is a $d$-dimensional compact convex set in $\mathbb{R}^d$ and $0\in\mathrm{int}\,D$, then one can define the \emph{(convex) norm} (or \emph{Minkowski functional}) $\d{\cdot}$ generated by $D$ as
\begin{equation}
\label{E_norm}
\d{x} = \min \big\{ t \in \mathbb{R}^+:\,x \in t D \big\} = \max \big\{ x' \cdot x:\,x' \in D^* \big\},
\end{equation}
where
\[
D^* := \Big\{ x' \in \mathbb{R}^d:\,x' \cdot x \leq 1 \text{ for all $x\in D$} \Big\}
\]
is the \emph{convex dual} to $D$. Equivalently, $\d{\cdot}:\mathbb{R}^d\rightarrow\mathbb{R}$ is identified by the following properties:
\begin{align*}
& \{x\in\mathbb{R}^d:\,\d{x}\leq1\}=D;\\
& \d{t x} = t \d{x},\quad\forall\,t \geq 0 \qquad \text{positively $1$-homogeneous};\\
& \d{x+y}\leq\d{x}+\d{y}\qquad\text{subadditive}.
\end{align*}
In particular, since $\d{\cdot}$ is positively $1$-homogeneous, subadditivity can be equivalently replaced by convexity.
\begin{remark}
\label{R_cone_epi}
Notice that the set $\mathrm{epi}\,\d{\cdot}$ belongs to $\mathcal C(d+1;\mathbb{R}^{d+1})$, with the identification $\mathbb{R}^{d}\times\mathbb{R}\overset{\sim}{=}\mathbb{R}^{d+1}$. Viceversa, given a convex cone $C\in\mathcal C(k,\mathbb{R}^d)$, if we fix a system of coordinates $(x_1,\dots,x_k,x_{k+1},\dots,x_d)\in\mathbb{R}^d$ such that $H_C=\{x_k\geq0,\,x_{k+1}=\dots=x_d=0\}$, then $C$ is the epigraph of a convex norm on $\mathbb{R}^{k-1}\overset{\sim}{=}\{x_{k}=\dots=x_d=0\}$.
\end{remark}
We will call \emph{extremal cones} of a convex norm $\d{\cdot}:\mathbb{R}^d\rightarrow\mathbb{R}$ either the extremal cones of $\mathrm{epi}\,\d{\cdot}$ in $\mathbb{R}^{d+1}$ or their projections on $\mathbb{R}^d$, being the distinctions between the two cases clear from the context.
For $C \in \mathcal C(k,\mathbb{R}^d)$ we call $C\cap\mathbb S^{d-1}$ the \emph{set of directions} of $C$. For $V \in \mathcal G(k,\mathbb{R}^d)$ we will also write
\begin{equation}
\label{E_cones_and_directions_in_subspace}
\mathcal C(k',V) := \big\{ C \in \mathcal C(k',\mathbb{R}^d): C \subset V \big\}.
\end{equation}
If $C \in \mathcal C(k,\mathbb{R}^d)$, $r>0$, we also define the cone
\begin{equation}
\label{E_epsilon_neigh_of_Cka}
\mathring C(r) := \{0\} \cup \mathbb{R}^+ \Big( \big( C\cap\mathbb S^{d-1} + B^d(0,r) \big) \cap \mathrm{aff}\,C \Big).
\end{equation}
Clearly $C(r) := \mathrm{clos}\,\mathring C(r) \in \mathcal C(k,\mathbb{R}^d)$ for $0 < r \ll 1$ and $C = \underset{n}{\cap}\, \mathring C(2^{-n})$. For $r > 0$ we also define the cone
\begin{equation}
\label{E_inverse_neigh_Cone}
\mathring C(-r) := \{0\} \cup \mathbb{R}^+ \Big\{ x \in \mathbb S^{d-1} \cap \mathrm{aff}\,C :\, B^d(x,r) \cap \mathrm{aff}\,C \subset C \Big\},
\end{equation}
so that $\{0\} \cup \inter_{\mathrm{rel}} C = \underset{n}{\cup}\, \mathring C(-2^{-n})$: as before $C(-r) := \mathrm{clos}\,\mathring C(-r) \in \mathcal C(k,R^d)$ for $0 < r \ll 1$. More generally, for $C \in \mathcal C(k,\mathbb{R}^d)$ we will use the notation
\begin{equation}
\label{E_open_cone_iterior_def}
\mathring C := \{0\} \cup \inter_{\mathrm{rel}} C = \bigcup_{n \in \mathbb{N}} \mathring C(-2^{-n}).
\end{equation}
By convention we set $\mathcal C(0,\mathbb{R}^d)=\mathbb{R}^d$ and we will often denote a convex cone $C$ as $C^{\dim\,C}$ to emphasize its dimension.
On $\mathcal G(k,\mathbb{R}^d)$, $\mathcal A(k,\mathbb{R}^d)$ and $\mathcal C(k,\mathbb{R}^d)$ we consider the topology given by the Hausdorff distance $\mathtt d_{\mathrm H}$ in every closed ball $\mathrm{clos}\, B^d(0,r)$ of $\mathbb{R}^d$, i.e.
\begin{equation}
\label{E_distance_cones}
\mathtt d(A,A') := \sum_n 2^{-n} \mathtt d_{\mathrm H} \big( A \cap B^d(0,n), A' \cap B^d(0,n) \big).
\end{equation}
for two generic elements $A$, $A'$.
If $A \subset \mathbb S^{d-1}$, its spherical convex envelope is defined as
\begin{equation}
\label{E_convex_envelope_on_sphere}
\mathrm{conv}_{\mathbb S^{d-1}} A := \mathbb S^{d-1} \cap \big( \mathbb{R}^+ \mathrm{conv}\,A \big).
\end{equation}
\subsection{Measures and disintegration}
\label{Ss_measure_disintegration}
We will denote the Lebesgue measure on $\mathbb{R}^d$ by $\mathcal L^d$, and the $k$-dimensional Hausdorff measure on $V \in \mathcal A(k,\mathbb{R}^d)$ as $\mathcal H^k \llcorner_V$. In general, the restriction of a function/measure to a set $A \in \mathbb{R}^d$ will be denoted by the symbol $\llcorner_A$ (or sometimes $\llcorner A$) following the function/measure. The product of two locally finite Borel measures $\varpi_0$, $\varpi_1$ will be denoted by $\varpi_0 \otimes \varpi_1$.
The Lebesgue points $\mathrm{Leb}(A)$ of a set $A \subset \mathbb{R}^d$ are the points $z \in A$ such that
\begin{equation}
\label{E_lebesgue_point_definition}
\lim_{r \rightarrow 0} \frac{\mathcal L^d(A \cap B^d(z,r))}{\mathcal L^d(B^d(z,r))} = 1.
\end{equation}
If $\varpi$ is a locally bounded Borel measure on $\mathbb{R}^d$, we will write $\varpi \ll \mathcal L^d$ if $\varpi$ is \emph{absolutely continuous} (a.c. for brevity) w.r.t. $\mathcal L^d$.
For a generic \emph{Polish space} $X$ (i.e., a separable and complete metric space), the Borel sets and the set of Borel probability measures will be respectively denoted by $\mathcal B(X)$ and $\mathcal P(X)$.
The \emph{Souslin sets $\Sigma^1_1$} of a Polish space $X$ are the projections on $X$ of the Borel sets of $X \times X$. The $\sigma$-algebra generated by the Souslin sets will be denoted by $\varTheta$.
Two Radon measures $\varpi_0$, $\varpi_1$ on $X$ are \emph{equivalent} if for all $B \in \mathcal B(X)$
\begin{equation}
\label{E_equiv_varpi}
\varpi_0(B) = 0 \quad \Longleftrightarrow \quad \varpi_1(B)=0,
\end{equation}
and we denote this property by $\varpi_0 \simeq \varpi_1$.
If $\varpi$ is a measure on a measurable space $X$ and $\mathtt f : X \rightarrow Y$ is an $\varpi$-measurable map, then the \emph{push-forward of $\varpi$} by $\mathtt f$ is the measure $\mathtt f_\# \varpi$ on $Y$ defined by
\begin{equation}
\label{E_push_forward}
\mathtt f_\#\varpi(B)=\varpi(\mathtt f^{-1}(B)),\quad\text{for all $B$ in the $\sigma$-algebra of $Y$.}
\end{equation}
Finally we briefly recall the concept of disintegration of a measure over a partition.
\begin{definition}[Partitions]
\label{D_part}
A \emph{partition} in $\mathbb{R}^d$ is a family $\{Z_{\mathfrak a}\}_{\mathfrak a\in\mathfrak{A}}$ of disjoint subsets of $\mathbb{R}^d$. We say that $\{Z_{\mathfrak a}\}_{\mathfrak a\in\mathfrak{A}}$ is a \emph{Borel partition} if $\mathfrak{A}$ is a Polish space, $\underset{\mathfrak a \in \mathfrak A}{\cup} Z_\mathfrak a$ is Borel and the \emph{quotient map} $\mathtt h : \underset{\mathfrak a\in\mathfrak{A}} {\cup} Z_\mathfrak a \rightarrow \mathfrak{A}$, $\mathtt h : z \mapsto \mathtt h(z) = \mathfrak a$ such that $z \in Z_{\mathfrak a}$, is Borel-measurable. We say that $\{Z_{\mathfrak a}\}_{\mathfrak a\in\mathfrak{A}}$ is \emph{$\sigma$-compact} if $\mathfrak{A}\subset\mathbb{R}^k$ for some $k\in\mathbb{N}$, $\underset{\mathfrak a \in \mathfrak{A}}{\cup}\, Z_\mathfrak a$ is $\sigma$-compact and $\mathtt h$ is $\sigma$-continuous.
\end{definition}
The sets in the $\sigma$-algebra $\{\mathtt h^{-1}(F):\, F \in \mathcal B(\mathfrak{A})\}$ are also called in the literature \emph{saturated sets}. Notice that we do not require $\{Z_{\mathfrak a}\}_{\mathfrak a\in\mathfrak{A}}$ to be a covering of $\mathbb{R}^d$.
\begin{definition}[Disintegration]
\label{D_dis}
Given a Borel partition in $\mathbb{R}^d$ into sets $\{Z_\mathfrak a\}_{\mathfrak a\in\mathfrak{A}}$ with quotient map $\mathtt h : \underset{\mathfrak a\in\mathfrak{A}}{\cup} Z_\mathfrak a \rightarrow \mathfrak{A}$ and a probability measure $\varpi \in \mathcal P(\mathbb{R}^d)$ s.t. $\varpi\bigl(\underset{\mathfrak a\in\mathfrak{A}}{\cup}Z_\mathfrak a\bigr)=1$, a \emph{disintegration} of $\varpi$ w.r.t. $\{Z_\mathfrak a\}_{\mathfrak a\in\mathfrak{A}}$ is a family of probability measures $\{\varpi_\mathfrak a\}_{\mathfrak a\in\mathfrak{A}}\subset\mathcal P(\mathbb{R}^d)$ such that
\begin{align}
& \mathfrak{A} \ni \mathfrak a \mapsto \varpi_\mathfrak a(B) \quad \text{is an $\mathtt h_{\#} \varpi$-measurable map $\forall\,B\in\mathcal B(\mathbb{R}^d)$}, \label{E_disint1}\\
& \varpi \bigl( B \cap \mathtt h^{-1}(F) \bigr) = \int_F \varpi_{\mathfrak a}(B)\,d\mathtt h_{\#}\varpi(\mathfrak a), \quad \forall\,B\in\mathcal B(\mathbb{R}^d),\,F\in\mathcal B(\mathfrak{A}). \label{E_disint2}
\end{align}
\end{definition}
As proven in Appendix A of \cite{BiaCar} (for a more comprehensive analysis see \cite{Fre}), we have the following theorem.
\begin{theorem}
\label{T_disint}
Under the assumptions of Definition \ref{D_dis}, the disintegration $\{\varpi_\mathfrak a\}_{\mathfrak a\in\mathfrak{A}}$ is \emph{unique} and \emph{strongly consistent}, namely
\begin{align}
&\text{if }\mathfrak a\mapsto\varpi^1_{\mathfrak a},\:\mathfrak a\mapsto\varpi^2_{\mathfrak a} \text{ satisfy \eqref{E_disint1}-\eqref{E_disint2}} \quad \Longrightarrow \quad \varpi^1_\mathfrak a=\varpi^2_\mathfrak a \ \text{for $\mathtt h_{\#}\varpi$-a.e. $\mathfrak a\in\mathfrak{A}$}; \notag \\
&\varpi_{\mathfrak a}(Z_\mathfrak a)=1 \quad \text{ for $\mathtt h_{\#}\varpi$-a.e. $\mathfrak a\in\mathfrak{A}$}.\label{E_disintsc}
\end{align}
\end{theorem}
The measures $\{\varpi_\mathfrak a\}_{\mathfrak a\in\mathfrak{A}}$ are also called \emph{conditional probabilities}.
To denote the (strongly consistent) disintegration $\{\varpi_\mathfrak a\}_{\mathfrak a\in\mathfrak{A}}$ of a probability measure $\varpi\in\mathcal P(\mathbb{R}^d)$ on a Borel partition $\{Z_\mathfrak a\}_{\mathfrak a\in\mathfrak{A}}$ we will often use the formal notation
\begin{equation}
\label{E_dis_not}
\varpi=\int_\mathfrak A \varpi_\mathfrak a \,dm(\mathfrak a),\quad \varpi_\mathfrak a(Z_\mathfrak a)=1,
\end{equation}
with $m = \mathtt h_\# \varpi$, $\mathtt h$ being the quotient map.
Since the conditional probabilities $\varpi_\mathfrak a$ are defined $m$-a.e., many properties (such as $\varpi_\mathfrak a(Z_\mathfrak a)=1$) should be considered as valid only for $m$-a.e. $\mathfrak a \in \mathfrak{A}$: for shortness, we will often consider the $\varpi_{\mathfrak a}$ redefined on $m$-negligible sets in order to have statements valid $\forall \mathfrak a \in \mathfrak{A}$.
We also point out the fact that, according to Definition \ref{D_dis}, in order that a disintegration of $\varpi$ over a partition can be defined, $\varpi$ has to be concentrated on the union of the sets of the partition (which do not necessarily cover the whole $\mathbb{R}^d$). In general, if we remove this assumption, since the formulas \eqref{E_disint1}-\eqref{E_disint2} make sense nonetheless for $B \subset \underset{\mathfrak a\in\mathfrak{A}}{\cup}Z_\mathfrak a$, by means of formula \eqref{E_dis_not} we ``reconstruct'' only $\varpi\llcorner_{\underset{\mathfrak a\in\mathfrak{A}}{\cup}Z_\mathfrak a}$ .
Let $m'\in\mathcal P(\mathfrak{A})$, $\{\varpi'_\mathfrak a\}_{\mathfrak a\in\mathfrak{A}}\subset\mathcal P(\mathbb{R}^d)$ such that
\begin{equation*}
\mathfrak{A} \ni \mathfrak a \mapsto \varpi'_\mathfrak a(B) \quad \text{ is $m'$-measurable, $\forall\,B\in\mathcal B(\mathbb{R}^d)$}.
\end{equation*}
Then, one can define the probability measure $\varpi'$ on $\mathbb{R}^d$ by
\begin{equation}
\label{E_int_measure}
\varpi'(B)=\int_\mathfrak A \varpi'_\mathfrak a(B)\,dm'(\mathfrak a),\quad\text{ $\forall\,B\in\mathcal B(\mathbb{R}^d)$}.
\end{equation}
The measure defined in \eqref{E_int_measure} will be denoted as
\[
\varpi' = \int_\mathfrak A \varpi'_\mathfrak a\,dm'.
\]
Notice that, despite the notation is the same as in \eqref{E_dis_not}, the family $\{\varpi'_\mathfrak a\}_{\mathfrak a \in \mathfrak A}$ in the above definition is not necessarily a disintegration of $\varpi'$, both because the measure $m'$ is not necessarily a quotient measure of a Borel partition and because the measures $\varpi'_\mathfrak a$ are not necessarily concentrated on the sets of a partition. In the rest of the paper, such an ambiguity will not occur, since we will always point out whether a measurable family of probability measures is generated by a disintegration or not.
\begin{remark}
\label{R_disint_lebesgue}
If instead of $\varpi \in \mathcal P(\mathbb{R}^d)$ we consider the Lebesgue measure $\mathcal L^d$ (more generally, a Radon measure) a disintegration $\{\upsilon_\mathfrak a\}_{\mathfrak a\in\mathfrak{A}}$ is to be considered in the following sense. First choose a partition $\{A_i\}_{i\in\mathbb{N}}$ of $\mathbb{R}^d$ into sets with unit Lebesgue measure, then let
\[
\mathcal L^d \llcorner_{A_i} = \int \upsilon_{\mathfrak a,i} d\eta_i(\mathfrak a), \quad \eta_i := \mathtt h_\# \mathcal L^d \llcorner_{A_i},
\]
be the standard disintegration of the probability measure $\mathcal L^d \llcorner_{A_i}$, and finally
\[
\upsilon_\mathfrak a := \sum_i 2^i \upsilon_{\mathfrak a,i}, \quad \eta := \sum_i 2^{-i} \eta_i.
\]
Clearly, in this definition the ``conditional probabilities'' $\upsilon_\mathfrak a$ and the ``image measure'' $\eta$ depend on the choice of the sets $\{A_i\}_{i\in\mathbb{N}}$.
\end{remark}
\subsection{Optimal transportation problems}
\label{Ss_transference_plans}
For a generic Polish space $X$, measures $\mu,\nu \in \mathcal P(X)$ and Borel \emph{cost function} $\mathtt c : X \times X \rightarrow [0,\infty]$, we will consider the sets of probability measures
\begin{equation}
\label{E_Pi_mu_nu}
\Pi(\mu,\nu) := \Big\{ \pi \in \mathcal P(X \times X) : (\mathtt p_1)_\# \pi = \mu, (\mathtt p_2)_\# \pi = \nu \Big\},
\end{equation}
\begin{equation}
\label{E_Pi_mu_nu_fin}
\Pi^f_{\mathtt c}(\mu,\nu) := \bigg\{ \pi \in \Pi(\mu,\nu) : \int_{X \times X} \mathtt c \,d\pi < +\infty \bigg\},
\end{equation}
\begin{equation}
\label{E_Pi_mu_nu_optimal}
\Pi^{\mathrm{opt}}_{\mathtt c}(\mu,\nu) := \bigg\{ \pi \in \Pi(\mu,\nu) : \int_{X \times X} \mathtt c \,d\pi = \inf_{\pi' \in \Pi(\mu,\nu)} \int_{X \times X} \mathtt c \,d\pi' \bigg\}.
\end{equation}
The elements of the set defined in \eqref{E_Pi_mu_nu} are called \emph{transference} or \emph{transport plans} between $\mu$ and $\nu$, those in \eqref{E_Pi_mu_nu_fin} \emph{transference} or \emph{transport plans with finite cost} and the set defined in \eqref{E_Pi_mu_nu_optimal} is the set of \emph{optimal plans}. The quantity
\begin{equation}
\label{E_transport_prob}
\mathtt C(\mu,\nu) := \inf_{\pi \in \Pi(\mu,\nu)} \int_{X \times X} \mathtt c \,d\pi
\end{equation}
is the \emph{transportation cost}.
In the following we will always consider costs and measures s.t. $\mathtt C(\mu,\nu)<+\infty$, thus $\Pi^f_{\mathtt c}(\mu, \nu)\neq\emptyset$.
The problem of showing that $\Pi^{\mathrm{opt}}_{\mathtt c}(\mu,\nu)\neq\emptyset$ is called Monge-Kantorovich problem.
We recall (see e.g. \cite{BiaCar,kel:duality}) that any optimal plan $\pi\in\Pi^{\mathrm{opt}}_{\mathtt c}(\mu,\nu)$ is \emph{$\mathtt c$-cyclically monotone}, i.e. there exists a $\sigma$-compact \emph{carriage} $\Gamma\subset X\times X$ such that $\pi(\Gamma)=1$ and for all $I\in \mathbb{N}$, $\{(x_i,y_i)\}_{i=1}^I\subset\Gamma$,
\[
\sum_{i=1}^I\mathtt c(x_i,y_i)\leq\sum_{i=1}^I\mathtt c(x_{i+1},y_i),
\]
where we set $x_{I+1}:=x_1$. Any such $\Gamma$ is called \emph{$\mathtt c$-cyclically monotone carriage}. However, in order to deduce the optimality of a transference plan the $\mathtt c$-cyclical monotonicity condition itself is not sufficient and one has to impose additional conditions. Most of the conditions in the literature exploit the dual formulation of Monge-Kantorovich problem (see \cite{villa:Oldnew}), namely
\[
\mathtt C(\mu,\nu)=\sup_{\underset{\phi\:\mu\text{-meas. and }\psi\:\nu\text{-meas.}}{\phi,\,\psi:X\rightarrow[-\infty,+\infty)}}\Big\{\int \phi(x)\,d\mu(x)+\int\psi(y)\,d\nu(y):\,{\phi(x)+\psi(y)\leq \mathtt c(x,y)}\Big\}.
\]
For example (see Lemma 5.3 of \cite{BiaCar}) if there exists a pair of functions
\begin{align}
&\phi,\,\psi:X\rightarrow[-\infty,+\infty), \quad \text{ $\phi$ $\mu$-measurable and $\psi$ $\nu$-measurable}, \label{E_phipsi1} \\
&\phi(x)+\psi(y)\leq \mathtt c(x,y), \quad \forall\, x,y \in X, \label{E_phipsi2} \\
&\phi(x)+\psi(y)= \mathtt c(x,y), \quad \text{ $\pi$-a.e. for some $\pi\in \Pi(\mu,\nu)$}, \label{E_phipsi3}
\end{align}
then $\phi,\,\psi$ are optimizers for the dual problem and $\pi\in\Pi^{\mathrm{opt}}_{\mathtt c}(\mu,\nu)$. Conditions on the cost guaranteeing the existence of such potentials (and indeed of more regular ones) are e.g. the following ones:
\begin{enumerate}
\item \label{Point_boun_fg} $\mathtt c$ is l.s.c. and satisfies $\mathtt c(x,y) \leq \mathtt f(x) + \mathtt g(y)$ for some $\mathtt f \in L^1(\mu)$, $\mathtt g \in L^1(\nu)$ (\cite{rachrusch});
\item $\mathtt c$ is real-valued and satisfies the following assumption (\cite{conf:optcime})
\[
\nu\Big(\Big\{y:\,\int\mathtt c(x,y)\,d\mu(x)<+\infty\Big\}\Big)>0,\quad \mu\Big(\Big\{x:\,\int\mathtt c(x,y)\,d\nu(y)<+\infty\Big\}\Big)>0;
\]
\item $\{\mathtt c<+\infty\}$ is an open set $O$ minus a $\mu\otimes\nu$-negligible set $N$ (\cite{beigolmarsch}).
\end{enumerate}
The weakest sufficient condition for optimality, which does not rely on the existence of global potentials and implies the results recalled above, has been given in \cite{BiaCar}. Since such condition will be needed and of fundamental importance for the proofs of our main results (in particular, Theorem \ref{T_subpart_step}), in the next section we give a brief explanation of the approach followed in \cite{BiaCar} and we state it in a form which will be more convenient for our purposes.
\subsection{Linear preorders, uniqueness and optimality}
\label{Ss_linear_pre_unique_transf}
Let $\mathtt c : X \times X \rightarrow [0,+\infty]$ be a Borel cost function on a Polish space $X$ such that $\mathtt c(x,x) = 0$ for all $x \in X$, let $\mu$, $\nu \in \mathcal P(X)$ be such that $\Pi^f_{\mathtt c}(\mu,\nu) \neq \emptyset$ and let $\Gamma \subset X \times X$ be a $\mathtt c$-cyclically monotone carriage of some $\pi \in \Pi^f_{\mathtt c}(\mu,\nu)$ satisfying w.l.o.g. $\{(x,x) : x \in X\} \subset \Gamma$.
A standard formula for constructing a pair of optimal potentials is the following: for fixed $(x_0,y_0) \in \Gamma$ and $(x,y)\in\Gamma$, define
\begin{align}
\label{E_pot_form}
\phi(x) &:= \inf \bigg\{ \sum_{i=0}^I \mathtt c(x_{i+1},y_i) - \mathtt c(x_i,y_i):\, (x_i,y_i) \in \Gamma, I \in \mathbb{N}, x_{I+1} = x \bigg\},\\
\psi(y) &:=\mathtt c(x,y)-\phi(x). \notag
\end{align}
If one of the assumptions $(1)$-$(3)$ holds, then this $\phi$, $\psi$ satisfy \eqref{E_phipsi1}-\eqref{E_phipsi3}. However, for general Borel costs $\mathtt c$, the assumptions $(1)$-$(3)$ are not satisfied. In particular, for any choice of $(x_0,y_0)$, there may be a set of positive $\mu$-measure on which $\phi$ is not well defined (namely, the infimum in \eqref{E_pot_form} is taken over an empty set) or takes the value $-\infty$ (see the examples in \cite{BiaCar}).
To explain why this can happen and briefly recall the strategy adopted in \cite{BiaCar} to overcome this problem in a more general setting, we need the following definition.
\begin{definition}[Axial paths and cycles]
\label{D_axpath}
An \emph{axial path with base points $\{(x_i,y_i)\}_{i=1}^I \subset \Gamma$}, $I\in\mathbb{N}$, starting at $x = x_1$ and ending at $x'$ is the sequence of points
\[
(x,y_1) = (x_1,y_1),(x_2,y_1),\dots,(x_i,y_{i-1}),(x_i,y_i),(x_{i+1},y_i),\dots,(x_I,y_I),(x',y_I).
\]
We will say that the axial path \emph{goes from} $x$ to $x'$: note that $x \in \mathtt p_1 \Gamma$. A \emph{closed axial path} or \emph{cycle} is an axial path with base points in $\Gamma$ such that $x = x'$. A \emph{$(\Gamma, \mathtt c)$-axial path} is an axial path with base points in $\Gamma$ whose points are contained in $\{\mathtt c < \infty\}$ and a \emph{$(\Gamma,\mathtt c)$-cycle} is a closed $(\Gamma, \mathtt c)$-axial path.
\end{definition}
Notice that, in order that \eqref{E_pot_form} is well defined, for $\mu$-a.e. point $x\in\mathtt p_1\Gamma$ there must be a $(\Gamma,\mathtt c)$-axial path going from $x_0$ to $x$. Moreover, being $\Gamma$ $\mathtt c$-cyclically monotone, $\phi$ is surely finite valued in the case in which for $\mu$-a.e. point $x\in\mathtt p_1 \Gamma$ there exists also a $(\Gamma,\mathtt c)$-axial path going from $x$ to $x_0$ (and thus to a.a. any other point in $\Gamma$). In particular, $x$ and $x_0$ are connected by a $(\Gamma,\mathtt c)$-cycle.
The first idea in \cite{BiaCar} is then to partition $X$ into the equivalence classes $\{Z_\mathfrak a\}_{\mathfrak a\in \mathfrak{A}}$ induced by the $(\Gamma,\mathtt c)$-cycle equivalence relation and disintegrate $\mu$, $\nu$ over $\{Z_\mathfrak a\}_{\mathfrak a\in \mathfrak{A}}$ and $\pi$ over $\{Z_\mathfrak a\times Z_\mathfrak b\}_{\mathfrak a,\mathfrak b\in\mathfrak{A}}$.
Since $\mathtt c(x,x)=0$ $\forall\,x\in X$ and $\Gamma\supset \mathrm{graph}\,\mathbb{I}$, then $(x,y)\in\Gamma$ implies that $x$ and $y$ belong to the same $(\Gamma,\mathtt c)$-cycle (consider the path $(x,y)$, $(y,y)$, $(y,y)$, $(x,y)$) and in particular that
\begin{equation}
\label{E_piconc}
\pi\bigg(\underset{\mathfrak a\in\mathfrak{A}}{\bigcup}Z_\mathfrak a\times Z_\mathfrak a\bigg)=1.
\end{equation}
If the disintegration is strongly consistent (see Theorem \ref{T_disint}), we get
\begin{align}
\mu&=\int\mu_\mathfrak a\,dm(\mathfrak a),\quad \mu_\mathfrak a(Z_\mathfrak a)=1, \label{E_mudisi}\\
\nu&=\int\nu_\mathfrak a\,dm(\mathfrak a),\quad\nu_\mathfrak a(Z_\mathfrak a)=1, \label{E_nudisi}\\
\pi&=\int\pi_{\mathfrak a\a}\,d(\mathbb{I}\times\mathbb{I})_\#m(\mathfrak a),\quad\pi_\mathfrak a(Z_\mathfrak a\times Z_\mathfrak a)=1, \label{E_pidisint}
\end{align}
where $m=\mathtt h_\#\mu=\mathtt h_\#\nu$ because there exists at least a plan in $\Pi^f_\mathtt c(\mu,\nu)$ --in this case $\pi$-- such that \eqref{E_piconc} is satisfied.
Notice that the fact that $\pi$ is concentrated on the diagonal equivalence classes $\{Z_\mathfrak a \times Z_\mathfrak a\}_{\mathfrak a \in \mathfrak{A}}$, i.e. formula \eqref{E_piconc}, is equivalent to say that the quotient measure $(\mathtt h \times \mathtt h)_\#\pi$
satisfies
\[
(\mathtt h \times \mathtt h)_\#\pi=(\mathbb{I}\times\mathbb{I})_\#m,
\]
i.e. it is concentrated on the diagonal of $\mathfrak{A}\times \mathfrak{A}$ (see \eqref{E_pidisint}).
Now, as a consequence of the fact that $\mu_\mathfrak a$-a.a. points in $Z_\mathfrak a$ can be connected to $\mu_\mathfrak a$-a.a. other points in $Z_\mathfrak a$ by a $(\Gamma\cap Z_\mathfrak a\times Z_\mathfrak a, \mathtt c)$-cycle and $\exists\,\pi_{\mathfrak a\mathfrak a}\in \Pi^f_\mathtt c(\mu_\mathfrak a,\nu_\mathfrak a)$ $\mathtt c$-cyclically monotone which is concentrated on $\Gamma\cap Z_\mathfrak a\times Z_\mathfrak a$, using \eqref{E_pot_form} we are able to construct optimal potentials $\phi_\mathfrak a$, $\psi_\mathfrak a:Z_\mathfrak a\rightarrow[-\infty,+\infty)$ for the transportation problem in $\Pi(\mu_\mathfrak a,\nu_a)$ and conclude that
\[
\pi_{\mathfrak a\a} \in \Pi^{\mathrm{opt}}_\mathtt c(\mu_\mathfrak a,\nu_\mathfrak a), \quad\text{ for $m$-a.e. $\mathfrak a$.}
\]
Let us then consider another $\pi'\in \Pi^f_\mathtt c(\mu,\nu)$. After the disintegration w.r.t. $\{Z_\mathfrak a\times Z_\mathfrak b\}_{\mathfrak a,\mathfrak b \in \mathfrak{A}}$ we get
\[
\pi' = \int\pi'_{\mathfrak a\mathfrak b}dm'(\mathfrak a,\mathfrak b), \quad \pi'_{\mathfrak a\mathfrak b}(Z_\mathfrak a\times Z_\mathfrak b)=1,
\]
with
\begin{equation}
\label{E_quotient_cost}
m'\in\Pi^f_{(\mathtt h \times \mathtt h)_\# \mathtt c}(m,m), \quad \text{where} \quad (\mathtt h \times \mathtt h)_\# \mathtt c(\mathfrak a,\mathfrak b) = \inf_{Z_\mathfrak a \times Z_\mathfrak b} \mathtt c(x,y).
\end{equation}
Hence one has the following theorem, which gives a sufficient condition for optimality based on behavior of optimal transport plans w.r.t. disintegration on $(\Gamma,\mathtt c)$-cycle equivalence relations.
\begin{theorem}
\label{T_A1}
Let $\Gamma$ be a $\mathtt c$-cyclically monotone carriage of a transference plan $\pi \in \Pi^f_\mathtt c(\mu,\nu)$. If the partition $\{Z_\mathfrak a\}_{\mathfrak a\in\mathfrak{A}}$ w.r.t. the $(\Gamma,\mathtt c)$-cycle equivalence relation satisfies
\begin{align}
&\text{the disintegration on $\{Z_\mathfrak a\}_{\mathfrak a\in\mathfrak{A}}$ is strongly consistent,} \label{E_teoA1} \\
&\pi'\bigg(\underset{\mathfrak a}{\bigcup}\; Z_\mathfrak a\times Z_\mathfrak a\bigg)=1,\quad\forall\,\pi'\in\Pi^f_\mathtt c(\mu,\nu), \label{E_teoA2}
\end{align}
then $\pi$ is an optimal transference plan.
\end{theorem}
Indeed, if \eqref{E_teoA1} and \eqref{E_teoA2} are satisfied, then $\pi'=\int\pi'_{\mathfrak a\a}\,d(\mathbb{I}\times\mathbb{I})_\#m(\mathfrak a)$ with $\pi'_{\mathfrak a\a} \in \Pi^f_\mathtt c(\mu_\mathfrak a,\nu_\mathfrak a)$ and one obtains the conclusion by integrating w.r.t. $m$ the optimality of the conditional plans $\pi_{\mathfrak a\a}$, namely
\[
\int \mathtt c(x,y)\,d\pi_{\mathfrak a\a}(x,y)\leq\int\mathtt c(x,y)\,d\pi'_{\mathfrak a\a}(x,y).
\]
The second crucial point in \cite{BiaCar} is then to find weak sufficient conditions such that the assumptions of Theorem \ref{T_A1} are satisfied.
Before introducing them, we show how the request that the sets of a Borel partition satisfying \eqref{E_teoA2} coincide with the equivalence classes of the $(\Gamma,\mathtt c)$-cycle relation can be weakened, yet yielding the possibility of constructing optimal potentials on each class --and then, as a corollary, to prove the optimality of a $\mathtt c$-cyclically monotone plan $\pi$.
First, we need the following
\begin{definition}
\label{def:gccyclconn}
A set $E\subset \mathtt p_1\Gamma$ is \emph{$(\Gamma,\mathtt c)$-cyclically connected} if $\forall\, x,y\in E$ there exists a $(\Gamma,\mathtt c)$-cycle connecting $x$ to $y$.
\end{definition}
According to the above definition, the equivalence classes of $\preccurlyeq_{(\Gamma,\mathtt c)}$ are maximal $(\Gamma,\mathtt c)$-cyclically connected sets, namely $(\Gamma,\mathtt c)$-cyclically connected sets which are maximal w.r.t. set inclusion.
Then notice that, given a Borel partition $\{Z_{\mathfrak b}'\}_{\mathfrak b\in\mathfrak B}\subset\mathbb{R}^d$ such that
\[
\pi\bigg(\underset{\mathfrak b}{\bigcup}Z_\mathfrak b'\times Z_\mathfrak b'\bigg)=1,\quad\forall\,\pi\in\Pi^f_{\mathtt c}(\mu,\nu)
\]
and whose sets are $(\Gamma,\mathtt c)$-cyclically connected but not necessarily maximal, then it is still possible to define on each of them a pair of optimal potentials and prove the optimality of $\pi$ such that $\pi(\Gamma)=1$.
Moreover, one can weaken this condition by removing a $\mu$-negligible set in the following way.
Let $\mu=\int\mu'_{\mathfrak b}\,dm'(\mathfrak b)$, $\mu_{\mathfrak b}'(Z_\mathfrak b)=1$.
\begin{definition}
\label{D_gammaconn}
The partition $\{Z_\mathfrak b'\}_{\mathfrak b\in\mathfrak B}$ is \emph{$(\mu,\Gamma, \mathtt c)$-cyclically connected} if $\exists\,F\subset X$ $\mu$-conegligible s.t. $Z_\mathfrak b'\cap F$ is $(\Gamma, \mathtt c)$-cyclically connected $\forall\, \mathfrak b\in\mathfrak B$. Equivalently, $\exists$ an $m'$-conegligible set $\mathfrak B'\subset \mathfrak B$ s.t. $\forall\,\mathfrak b'\in\mathfrak B'$ $\exists\,N_\mathfrak b'\subset Z_\mathfrak b'$, with` $\mu_{\mathfrak b}'(N_\mathfrak b')=0$, s.t. $Z_\mathfrak b'\setminus N_\mathfrak b'$ is $(\Gamma, \mathtt c)$-cyclically connected.
\end{definition}
When the $(\mu,\Gamma, \mathtt c)$-cyclically connectedness property holds for all $\mathtt c$-cyclically monotone carriages of all transport plans of finite cost --hence it is possible to construct optimal potentials starting from any $\mathtt c$-cyclically monotone $\Gamma$-- we have the following
\begin{definition}
\label{D_pimunuconn}
We say that $\{Z_\mathfrak b'\}$ is \emph{$\Pi^f_\mathtt c(\mu,\nu)$-cyclically connected} if it is $(\mu,\Gamma, \mathtt c)$-cyclically connected $\forall\,\Gamma$ $\mathtt c$-cyclically monotone s.t. $\pi(\Gamma)=1$ for some $\pi\in\Pi^f_\mathtt c(\mu,\nu)$.
\end{definition}
Notice that the $\mu$-conegligible set $F$ in the definition of $(\mu,\Gamma,\mathtt c)$-cyclically connected partition depends on the set $\Gamma$.
In this paper, in particular for the proof of Theorems \ref{T_final_seconry} and \ref{T_Monge_final}, the importance of $\Pi^f_\mathtt c(\mu,\nu)$-cyclically connected partitions is given by the following proposition.
\begin{proposition}
\label{P_second_cost}
Let $\{Z'_\mathfrak b\}_{\mathfrak b\in \mathfrak B}$ be a $\Pi^f_\mathtt c(\mu,\nu)$-cyclically connected Borel partition satisfying
\begin{equation}
\label{epseccost}
\pi\bigg(\underset{\mathfrak b}{\bigcup}\; Z_\mathfrak b'\times Z_\mathfrak b'\bigg)=1,\quad\forall\,\pi\in\Pi^f_\mathtt c(\mu,\nu)
\end{equation}
for a cost function of the form
\begin{equation}
\label{E_cost_ind}
\mathtt c(x,y)=\mathbbm 1_M(x,y), \qquad M \supset \big\{ (x,x) : x \in X \big\}.
\end{equation}
Let $\mathtt c_\mathtt m : X \times X \rightarrow [0,+\infty]$ be any \emph{secondary cost} of the form
\begin{equation}
\label{E_cs}
\mathtt c_\mathtt m(x,y) =
\begin{cases}
\mathtt m(x,y) & \mathtt c(x,y)<+\infty, \\
+\infty & \text{otherwise,}
\end{cases}
\end{equation}
where $\mathtt m$ is l.s.c. and there exist $\mathtt f \in L^1(\mu),\, \mathtt g \in L^1(\nu)$ s.t. $\mathtt m(x,y) \leq \mathtt f(x) + \mathtt g(y)$. Then, any $\mathtt c_\mathtt m$-cyclically monotone plan $\pi_\mathtt m\in\Pi^f_{\mathtt c_\mathtt m}(\mu,\nu)$ is optimal for $\mathtt c_\mathtt m$. More precisely, for any $\mathtt c_\mathtt m$-cyclically monotone set $\Gamma_\mathtt m$ with $\pi_\mathtt m(\Gamma_\mathtt m)=1$, there exist Borel functions $\phi^{\mathtt m}$, $\psi^\mathtt m$ such that the restrictions
\begin{equation}
\label{E_phipsi_m_restr}
\phi^\mathtt m_\mathfrak b := \phi^\mathtt m \llcorner_{Z_\mathfrak b'}, \qquad \psi^\mathtt m_\mathfrak b := \psi^\mathtt m \llcorner_{Z_\mathfrak b'}
\end{equation}
are Borel optimal potentials for $\Pi^{\mathrm{opt}}_{\mathtt c_\mathtt m}(\mu_\mathfrak b',\nu_\mathfrak b')$, for all $\mathfrak b$ in an $m'$-conegligible set $\mathfrak B'\subset \mathfrak B$.
\end{proposition}
\begin{proof}
Notice that $\Pi^f_{\mathtt c_{\mathtt m}}(\mu,\nu)\subset\Pi^f_\mathtt c(\mu,\nu)$.
Let $\Gamma_\mathtt m\subset \underset{\mathfrak b}{\cup}\; Z_\mathfrak b'\times Z_\mathfrak b'$ be a $\mathtt c_\mathtt m$-cyclically monotone carriage for $\pi_\mathtt m\in \Pi^f_{\mathtt c_{\mathtt m}}(\mu,\nu)$. Then, there exists a conegligible set $F \subset X$ such that $Z_\mathfrak b' \cap F$ is $(\Gamma_\mathtt m,\mathtt c)$-cyclically connected for all $\mathfrak b\in \mathfrak B$. Hence, formula \eqref{E_pot_form}, together with the validity of the Point \eqref{Point_boun_fg} at page \pageref{Point_boun_fg}, yields potentials $\phi^\mathtt m_\mathfrak b$, $\psi^\mathtt m_\mathfrak b$ for the transport problem in $\Pi^f_{\mathtt c_{\mathtt m}}(\mu_\mathfrak b,\nu_\mathfrak b)$ with cost $\mathtt c_\mathtt m$. In particular, the conditional probability $\pi_{\mathtt m,\mathfrak b\b}$ is optimal in $\Pi^f_{\mathtt c_{\mathtt m}}(\mu_\mathfrak b,\nu_\mathfrak b)$, and thus by \eqref{epseccost} it follows as in Theorem \ref{T_A1} that $\pi_\mathtt m$ is optimal in $\Pi^f_{\mathtt c_{\mathtt m}}(\mu,\nu)$.
The fact that one can find Borel functions $\phi^\mathtt m$, $\psi^\mathtt m$ such that \eqref{E_phipsi_m_restr} holds is an application of standard selection principles, and it can be found in \cite{BiaCar}.
\end{proof}
In order to state the main result of \cite{BiaCar} which is at the core of their sufficient condition concerning optimality, we need the concept of (linear) preorder.
\begin{definition}[(Linear) Preorder]
\label{D_preorder}
A \emph{preorder} on $X$ is a set $A\subset X\times X$ s.t.
\begin{align*}
&(x,x) \in A, \quad \forall\, x \in X\\
&(x,y) \in A \quad \wedge \quad (y,z) \in A \quad \Longrightarrow \quad (x,z) \in A.
\end{align*}
A preorder $A\subset X\times X$ is \emph{linear} if
\[
X\times X = A \cup A^{-1}.
\]
\end{definition}
The statement $(x,y)\in A$ will also be denoted by $x\preccurlyeq_A y$ and $A$ is also called the \emph{graph of the (linear) preorder $\preccurlyeq_A$}. Any preorder $\preccurlyeq_A$ induces the equivalence relation $\simeq_A$ on $X$
\[
x \simeq_A y \qquad \Longleftrightarrow \qquad x\preccurlyeq_Ay\quad\text{and}\quad y\preccurlyeq_A x.
\]
We also denote the graph of the equivalence relation $\simeq_A$ by
\[
A \cap A^{-1} \quad \text{ or } \quad \preccurlyeq_A \cap \, (\preccurlyeq_A)^{-1}.
\]
Going back to our problem, one can see that the $(\Gamma,\mathtt c)$-axial relation gives a Borel preorder on $X$, namely
\begin{equation}
\label{E_axpreorder}
x \preccurlyeq_{(\Gamma,\mathtt c)} y \quad\text{ if there exists a $(\Gamma,\mathtt c)$-axial path going from $y$ to $x$.}
\end{equation}
\begin{figure}
\input{orderaxialnotatbis.pdf_t}
\caption{The graph of the cost $\mathtt c$ is given by the indicator function of the region inside the blue curve. The graph of a $\mathtt c$-compatible linear preorder $\preccurlyeq_A$ is given by the union of the pink and of the red region. The red region corresponds to the graph of the induced equivalence relation $\simeq_A$. We draw also an axial path connecting $x_5$ to $x_3$ with base points $(x_5,y_5)$, $(x_4,y_4)$, and a $(\Gamma, \mathtt c)$-cycle connecting $(x_1,y_1)$ to $(x_2,y_2)$.}
\label{Fi_orderaxial.notat}
\end{figure}
The reason for introducing (linear) preorders in this context is given by the following theorem \cite{BiaCar}.
\begin{theorem}
\label{T_A2}
Let $A \subset X \times X$ be a Borel graph of a linear preorder on $X$ with equivalence classes $\{Z^A_\mathfrak c\}_{\mathfrak c \in \mathfrak C}$
satisfying
\begin{align}
& \{\mathtt c < +\infty\} \subset A, \label{E_tA22} \\
& \preccurlyeq_{(\Gamma,\mathtt c)} \subset A,\text{ for some $\mathtt c$-cyclically monotone set $\Gamma$ s.t. $\pi(\Gamma)=1$, $\pi\in\Pi^f_{\mathtt c}(\mu,\nu)$.} \label{E_tA21}
\end{align}
Then, the disintegration w.r.t. the partition $\{Z^A_\mathfrak c\}_{\mathfrak c \in \mathfrak C}$ is strongly consistent and
\begin{equation}
\label{E_eqca}
\pi'\biggl( \bigcup_\mathfrak c Z^A_\mathfrak c \times Z^A_\mathfrak c \biggr) = 1, \qquad \forall\,\pi'\in\Pi^f_\mathtt c(\mu,\nu).
\end{equation}
\end{theorem}
For future convenience we give the following definition.
\begin{definition}
\label{D_compatible}
A preorder $\preccurlyeq_A$ on $X$ is \emph{$\mathtt c$-compatible} if \eqref{E_tA22} holds.
\end{definition}
\begin{remark}
\label{rem_pre}
Let $A$ be a $\mathtt c$-compatible linear preorder. Whenever a carriage $\Gamma$ satisfies \eqref{E_tA21} the $\preccurlyeq_{(\Gamma,\mathtt c)}$-equivalence classes are contained in the equivalence classes of $\simeq_A$ and then, as noticed before, since $\Gamma \supset \mathrm{graph}\,\mathbb{I}$ and $\mathtt c(x,x)=0$ for all $x$,
\[
\Gamma \subset \underset{\mathfrak c}{\bigcup} \,Z^A_\mathfrak c \times Z^A_\mathfrak c, \quad\pi\bigg(\underset{\mathfrak c}{\bigcup}Z^A_\mathfrak c\times Z^A_\mathfrak c\bigg)=1.
\]
Viceversa, if $\pi'\bigl(\underset{\mathfrak c}{\cup}Z^A_\mathfrak c\times Z^A_\mathfrak c\bigl)=1$ for some $\pi'\in\Pi^f_{\mathtt c}(\mu,\nu)$ and $\pi'(\Gamma')=1$, then by the $\mathtt c$-compatibility of $A$
\[
\preccurlyeq_{(\Gamma'\cap\underset{\mathfrak c}{\cup}Z^A_\mathfrak c\times Z^A_\mathfrak c,\mathtt c)}\subset A
\]
and then also its equivalence classes are contained in the equivalence classes of $\simeq_A$.
In particular, \eqref{E_tA21} could also be rewritten as $\pi\bigl(\underset{\mathfrak c}{\cup}Z^A_\mathfrak c\times Z^A_\mathfrak c\bigr)=1$.
We point out that, while a $\mathtt c$-compatible linear preorder satisfying \eqref{E_tA21} for some $\Gamma$ can always be constructed using the axiom of choice, \eqref{E_eqca} may not hold if the linear preorder is not Borel (see \cite{BiaCar}): hence, the main assumption of the theorem is the Borel regularity.
Finally, notice that the partition into equivalence classes of $\preccurlyeq_{(\Gamma'\cap\underset{\mathfrak c}{\cup}Z^A_\mathfrak c\times Z^A_\mathfrak c,\mathtt c)}$ with $\Gamma'$ as above is $(\mu,\Gamma',\mathtt c)$-cyclically connected in the sense of Definition \ref{D_gammaconn}.
\end{remark}
In order to prove Theorem \ref{T_subpart_step}, in Section \ref{S_cfibr_cfol} we will look --for a particular class of cost functions of the form \eqref{E_cost_ind} called \emph{cone-Lipschitz costs associated to a directed fibration}-- for $\Pi^f_{\mathtt c}(\mu,\nu)$-cyclically connected partitions satisfying \eqref{epseccost}. Therefore, by Theorem \ref{T_A2} and Remark \ref{rem_pre}, we will construct a Borel $\mathtt c$-compatible linear preorder $A$ such that, for any carriage of finite cost $\Gamma'$, the equivalence classes of $\preccurlyeq_{(\Gamma'\cap \underset{\mathfrak c}{\cup}Z^A_\mathfrak c\times Z^A_\mathfrak c,\mathtt c)}$ coincide up to a $\mu$-negligible set with those of $\simeq_A$.
For convenience we give also the following
\begin{definition}
\label{D_cmunucomp}
If $\preccurlyeq_A$ is $\mathtt c$-compatible and \eqref{E_tA21} holds for every $\pi\in\Pi^f_\mathtt c(\mu,\nu)$, then $A$ is called \emph{$(\mathtt c,\mu,\nu)$-compatible}.
\end{definition}
Hence, Theorem \ref{T_A2} can also be restated saying that whenever $A$ is a Borel $\mathtt c$-compatible linear preorder satisfying \eqref{E_tA21} for some $\Gamma$ of finite cost, then it is $(\mathtt c,\mu,\nu)$-compatible.
According to the terminology used in \cite{BiaCar}, $(\mathtt c,\mu,\nu)$-compatibility can also be restated saying that the diagonal in the quotient space
\begin{equation}
\label{E_push_f_A}
(\mathbb{I}\times\mathbb{I})\circ \mathtt h\circ \mathtt p_1 (A)
\end{equation}
is a \emph{set of uniqueness} for $\Pi^f_{(\mathtt h \times \mathtt h)_\# \mathtt c}(m,m)$, where $\mathtt h$ is the quotient map associated to the partition $\simeq_A$: this means that there exists a unique transference plan in $\Pi^f_{(\mathtt h \times \mathtt h)_\# \mathtt c}(m,m)$, namely $(\mathbb{I} \times \mathbb{I})_\# m$.
\section{Optimal transportation problems with convex norm and cone costs}
\label{S_conetransport}
Let $\d{\cdot}:\mathbb{R}^d\rightarrow\mathbb{R}$ be a convex norm as defined in \eqref{E_norm} and $\mu,\nu\in\mathcal P(\mathbb{R}^d)$. The transport plans with finite $\d{\cdot}$-cost $\Pi^f_{\d{\cdot}}(\mu,\nu)$ and the optimal plans w.r.t. $\d{\cdot}$ $\Pi^{\mathrm{opt}}_{\d{\cdot}}(\mu,\nu)$ are respectively given by the transference plans with finite cost and the optimal plans w.r.t. the cost function
\begin{equation}
\label{E_norm_cost_1}
\mathtt c(x,y) = \d{y - x}.
\end{equation}
Since the cost is a norm, we have the following well known results \cite{ambrgigli:userguide}: if $\Pi^f_{\d{\cdot}}(\mu,\nu) \not= \emptyset$, then
\begin{enumerate}
\item there exists at least one optimal transference plan $\bar \pi$;
\item if $\Gamma$ is a $|\cdot|_{D^*}$-cyclically monotone carriage of $\bar \pi$, then for $(x_0,y_0) \in \Gamma$ the function given by \eqref{E_pot_form},
\[
\phi(x) := \inf \bigg\{ \sum_{i=0}^I \big| y_i - x_{i+1} \big|_{D^*} - \big| y_i - x_i \big|_{D^*}: I \in \mathbb{N}, (x_i,y_i) \in \Gamma, x_{I+1} = x \bigg\},
\]
is Lipschitz continuous on $\mathbb{R}^d$ and
\begin{equation}
\label{E_potential_norm_leq}
\phi(x) - \phi(y) \leq |y - x|_{D^*},\quad\forall x,y \in \mathbb{R}^d,
\end{equation}
\begin{equation}
\label{E_potential_norm_equal}
\int_{\mathbb{R}^d \times \mathbb{R}^d} |y-x|_{D^*}\, d\bar \pi(x,y) = \int_{\mathbb{R}^d} \phi(x)\, d\mu(x) - \int_{\mathbb{R}^d} \phi(y)\, d\nu(x).
\end{equation}
\end{enumerate}
In particular, $\pi$ is an optimal plan if and only if
\[
\pi \Big( \Big\{ (x,y) : \phi(x) - \phi(y) = |y - x|_{D^*} \Big\} \Big) = 1.
\]
In the following we will denote by $\psi$ the dual potential
\begin{equation}
\label{E_dual_pot_psi}
\psi(x) := - \phi(x),
\end{equation}
which will be called \emph{Kantorovich potential}.
Clearly
\begin{equation}
\label{E_psi_forward_norm}
\psi(y) - \psi(x) \leq |y-x|_{D^*}, \qquad \pi \in \Pi^{\mathrm{opt}}_{\d{\cdot}}(\mu,\nu) \ \Leftrightarrow \ \pi \Big( \Big\{ (x,y) : \psi(y) - \psi(x) = |y - x|_{D^*} \Big\} \Big) = 1.
\end{equation}
\begin{definition}
\label{D_super_sub_diff}
A function $\varphi : \mathrm{dom}\, \varphi\subset\mathbb{R}^d\rightarrow\mathbb{R}$ is \emph{$\d{\cdot}$-Lipschitz} if it satisfies
\[
\varphi(y) - \varphi(x) \leq \d{y - x}, \qquad \forall\,x,y\in\mathrm{dom}\,\varphi.
\]
The \emph{superdifferential} of $\varphi$ is the set
\[
\partial^+\varphi := \Big\{ (x,y) : \varphi(y) - \varphi(x) = |y - x|_{D^*} \Big\},
\]
while its \emph{subdifferential} is the set
\[
\partial^-\varphi := \big(\partial^+\varphi\big)^{-1}.
\]
\end{definition}
Hence, \eqref{E_psi_forward_norm} can be rephrased as
\begin{equation}
\label{E_psi_opt}
\exists\,\psi : \mathbb{R}^d \rightarrow \mathbb{R} \text{ $\d{\cdot}$-Lipschitz s.t. }\pi \in \Pi^{\mathrm{\mathrm{opt}}}_{\d{\cdot}}(\mu,\nu) \ \Leftrightarrow \ \pi \big( \partial^+\psi \big) = 1.
\end{equation}
Let now $C^k\in\mathcal C(k;\mathbb{R}^k)$.
\begin{definition}
\label{D_cone_cost}
We define the \emph{convex cone cost associated to $C^k$} as the function $\mathtt c_{C^k}:\mathbb{R}^k\times\mathbb{R}^k\rightarrow[0,+\infty]$ given by
\begin{equation}
\label{E_cone_cost}
\mathtt c_{C^k}(x,y) =
\begin{cases}
0 & y-x \in C^k, \\
+\infty & \text{otherwise.}
\end{cases}
\end{equation}
\end{definition}
Given $\mu,\nu \in \mathcal P(\mathbb{R}^k)$, let $\Pi^f_{\mathtt c_{C^k}}(\mu,\nu)$ be the set of transport plans of finite cone cost. Notice that
\[
\Pi^f_{\mathtt c_{C^k}}(\mu,\nu)=\Pi^{\mathrm{opt}}_{\mathtt c_{C^k}}(\mu,\nu)= \Big\{ \pi\in\Pi^f_{\mathtt c_{C^k}}(\mu,\nu):\,\pi\text{ is $\mathtt c_{C^k}$-cyclically monotone} \Big\}.
\]
\subsection{Transportation problems with convex norms and cone costs on Lipschitz graphs}
\label{Ss_convex_norm_cone}
The optimal transport problem w.r.t. $\d{\cdot}$ can be casted as a convex cone optimal transportation problem on $\mathbb{R}^{d+1}\simeq\mathbb{R}^d\times\mathbb{R}$ w.r.t. the convex cone cost $\mathtt c_{\mathrm{epi}\,{\d{\cdot}}}$ associated to
\[
\mathrm{epi}\,\d{\cdot}\in\mathcal C(d+1;\mathbb{R}^{d+1})
\]
(see Definition \ref{D_cone_cost} and Remark \ref{R_cone_epi}).
Define in fact the measures in $\mathcal P(\mathrm{graph}\,\psi)$
\begin{equation}
\label{E_measure_on_graph}
\hat \mu := (\mathbb{I}\times\psi)_\# \mu, \qquad \hat \nu := (\mathbb{I}\times\psi)_\# \nu,
\end{equation}
where $\psi$ is the Kantorovich potential of $\Pi^{\mathrm{opt}}_{\d{\cdot}}(\mu,\nu)$, formula \eqref{E_dual_pot_psi}, and for $\pi \in \Pi(\mu,\nu)$ consider the plan in $\mathcal P(\mathrm{graph}\,\psi \times \mathrm{graph}\,\psi)$
\begin{equation}
\label{E_graph_plan}
\hat \pi := \big( (\mathbb{I}\times\psi) \times (\mathbb{I}\times\psi) \big)_\# \pi.
\end{equation}
The fundamental observations are \eqref{E_psi_opt} and the following: if $\varphi$ is $\d{\cdot}$-Lipschitz, then
\begin{equation}
\label{E_partvarphi_form}
\partial^+\varphi = \mathtt p_{\mathbb{R}^d\times\mathbb{R}^d} \Big( \mathrm{graph}\,\varphi \times \mathrm{graph}\,\varphi \cap \bigl\{ \mathtt c_{\mathrm{epi}\,\d{\cdot}} < +\infty \bigr\} \Big).
\end{equation}
\begin{definition}
\label{D_subsuperdiff}
If $\mathrm{graph}\,\varphi\subset\mathbb{R}^{d+1}$ is the graph of a $\d{\cdot}$-Lipschitz function, define its \emph{superdifferential} and \emph{subdifferential} respectively as
\begin{equation}
\label{E_sdiff_varphi}
\partial^+\mathrm{graph}\,\varphi=\mathrm{graph}\,\varphi\times\mathrm{graph}\,\varphi\cap\big\{\mathtt c_{\mathrm{epi}\,\d{\cdot}}<+\infty\big\},\qquad\partial^-\mathrm{graph}\,\varphi=\big(\partial^+ \mathrm{graph}\,\varphi\big)^{-1}.
\end{equation}
\end{definition}
Then \eqref{E_partvarphi_form} can be rewritten as
\[
\partial^\pm\varphi=\mathtt p_{\mathbb{R}^d\times\mathbb{R}^d} \big( \partial^\pm\mathrm{graph}\,\varphi \big).
\]
Hence the following proposition holds true.
\begin{figure}
\input{sudakovliftingbis.pdf_t}
\caption{The equivalence of Proposition \ref{P_equivalence_lifting}.}
\label{Fi_sudakov.lifting}
\end{figure}
\begin{proposition}
\label{P_equivalence_lifting}
The following statements are equiveridical:
\begin{enumerate}
\item \label{PCond_1_setting} \hskip 1cm $\pi \in \Pi^{\mathrm{opt}}_{\d{\cdot}}(\mu,\nu)$;
\item \label{PCond_2_setting} \hskip 1cm $\hat \pi := \big( (\mathbb{I}\times\psi) \times (\mathbb{I}\times\psi) \big)_\# \pi \in \Pi^f_{\mathtt c_{\mathrm{epi}\,\d{\cdot}}}(\hat \mu,\hat \nu)$, with $\hat \mu$, $\hat \nu$ given by \eqref{E_measure_on_graph};
\item \label{PCond_3_setting} \hskip 1cm $\pi = \big( \mathtt p_{\mathbb{R}^d \times \mathbb{R}^d} \big)_\# \hat \pi$ for some $\hat \pi \in \Pi^f_{\mathtt c_{\mathrm{epi}\,\d{\cdot}}}(\hat \mu,\hat \nu)$, with $\hat \mu$, $\hat \nu$ given by \eqref{E_measure_on_graph};
\item \label{E_hat_pi_graph} \hskip 1cm $\hat \pi := \big( (\mathbb{I}\times\psi) \times (\mathbb{I}\times\psi) \big)_\# \pi$ satisfies $\hat{\pi}(\partial^+\mathrm{graph}\,\psi) = 1$.
\end{enumerate}
\end{proposition}
Observe that, since $(\mathbb{I}\times\psi):\mathbb{R}^d\rightarrow\mathrm{graph}\,\psi$ is bi-Lipschitz, then if $\varpi \in \mathcal P(X)$ and $\hat \varpi := (\mathbb{I} \times \psi)_\# \varpi$,
\begin{equation}
\label{E_graphpsimu}
\varpi(B) = 0 \quad \Longleftrightarrow \quad \hat \varpi \bigl( (\mathbb{I}\times\psi)(B) \bigr) = 0, \quad \forall\, B \in \mathcal B(\mathbb{R}^d).
\end{equation}
\subsection{Optimal transportation problems on directed locally affine partitions}
\label{Ss_partitions_intro}
We first give the definition of directed locally affine partition.
\begin{definition}
\label{D_locaff}
We say that a nonempty subset $Z \subset \mathbb{R}^d$ is \emph{locally affine} if there exist $k\in\{0,\dots,d\}$ and $V \in \mathcal A(k,\mathbb{R}^d)$ such that $Z \subset V$ and $Z$ is relatively open in $V$, i.e. $Z=\inter_{\mathrm{rel}} Z \not= \emptyset$.
\end{definition}
Notice that, in the above definition, $V=\mathrm{aff}\,Z$. Whenever $Z$ is a locally affine set of dimension $k$ we will often denote it as $Z^{k}$ to emphasize its dimension.
\begin{definition}
\label{D_locaffpart}
A \emph{directed locally affine partition} in $\mathbb{R}^d$ is a partition into locally affine sets $\{Z^k_\mathfrak a\}_{\nfrac{k = 0,\dots,d}{\mathfrak a \in \mathfrak A^k}}$, endowed with a family of closed nondegenerate convex cones $\{C^k_\mathfrak a\}_{\nfrac{k = 0,\dots,d}{\mathfrak a \in \mathfrak A^k}}$ such that
\begin{enumerate}
\item the set
\begin{equation}
\label{E_bD}
\mathbf D:=\Big\{ (k,\mathfrak a,z,C^k_\a) : k\in\{0,\dots,d\},\, \mathfrak a \in \mathfrak{A}^k,\, z\in Z^k_\a \Big\}\subset \bigcup_{k=0}^{d} \{k\} \times \mathfrak{A}^k\times\mathbb{R}^d\times\mathcal C(k,\mathbb{R}^d)
\end{equation}
is $\sigma$-compact;
\item $\mathrm{aff}\,(z+C^k_\a)=\mathrm{aff}(Z^k_\a)$ for all $z\in Z^k_\a$.
\end{enumerate}
\end{definition}
For shortness we will use the notation
\begin{align}
\label{E_mathbf_Z_base_partition}
\mathbf Z^k &:= \mathtt p_z \mathbf D(k) = \bigcup_{\mathfrak a \in \mathfrak A^k} Z^k_\mathfrak a, \qquad \mathbf Z := \mathtt p_z \mathbf D = \bigcup_k \mathbf Z^k = \bigcup_{k=0}^d \bigcup_{\mathfrak a \in \mathfrak A^k} Z^k_\mathfrak a,\qquad
\bar{\mathbf Z}^k:=\underset{\mathfrak a\in\mathfrak{A}^k}{\bigcup}\mathrm{clos}\,Z^k_\a.
\end{align}
For the conditional probabilities of a measure $\mu$ over a locally affine partition we will use the notation $\{\mu^k_\mathfrak a\}_{\nfrac{k=0,\dots,d}{\mathfrak a\in\mathfrak{A}^k}}$, with $\mu^k_\mathfrak a(Z^k_\mathfrak a)=1$: the fact that the disintegration is strongly consistent is a consequence of the fact that the function $\mathbf Z\ni z \mapsto (k,\mathfrak a)$ has $\sigma$-compact graph $\mathtt p_{(z,k,\mathfrak a)} \mathbf D$.
Notice that the quotient space of the partition is given by
\begin{equation}
\label{E_disj_unio_A}
\mathfrak{A} := \bigsqcup_{k} \mathfrak{A}^k,
\end{equation}
where $\sqcup$ denotes the disjoint union of sets.
Given a locally affine directed partition $\{Z^k_\a, C^k_\a\}_{k,\mathfrak a}$ one can define the sets of \emph{initial} and \emph{final points} as follows.
\begin{definition}
\label{D_initial_final}
Define for $k = 1,\dots,d$, $\mathfrak a \in \mathfrak{A}^k$ the \emph{initial points of $Z^k_\a$} as
\begin{equation*}
\mathcal I(Z^k_\a) := \Bigl\{ z \in \mathbb{R}^d \setminus \mathbf Z : \exists\,r>0 \text{ s.t. } z + \inter_{\mathrm{rel}} C^k_\a \cap B^d(z,r) \subset Z^k_\a \Bigr\},
\end{equation*}
and the \emph{final points of $Z^k_\a$} as
\begin{equation*}
\mathcal E(Z^k_\a) := \Bigl\{ z \in \mathbb{R}^d \setminus \mathbf Z : \exists\,r>0 \text{ s.t. } z - \inter_{\mathrm{rel}} C^k_\a \cap B^d(z,r) \subset Z^k_\a \Bigr\}.
\end{equation*}
Finally, we call \emph{sets of initial points} and \emph{sets of final points} of the locally affine directed partition $\{Z^k_\a, C^k_\a\}_{k,\mathfrak a}$ the sets given respectively by
\begin{equation}
\label{E_initial_final}
\mathcal I := \bigcup_{k,\mathfrak a} \mathcal I(Z^k_\a), \qquad \mathcal E := \bigcup_{k,\mathfrak a} \mathcal E(Z^k_\a).
\end{equation}
\end{definition}
\begin{figure}
\input{sudakovdirefibrbis28.pdf_t}
\caption{A directed locally affine partition in $\mathbb{R}^3$ into three $2$-dimensional sets with quotient space $\mathfrak{A}^2$, five $1$-dimensional sets with quotient space $\mathfrak{A}^1$ and a $3$-dimensional set with quotient space $\mathfrak{A}^3$. In each $k$-dimensional subpartition, for $k=1,\dots,3$, we denote a locally affine set as $Z^k_\mathfrak a$ with cone of directions $C^k_\mathfrak a$ (colored in blue), initial points $\mathcal I(Z^k_\a)$ (colored in green), final points $\mathcal E(Z^k_\a)$ (colored in red) and quotient point $\mathfrak a$ (colored in purple), chosen as in \eqref{E_mathfrak_A_k_def}.}
\label{Fi_direct_fibra}
\end{figure}
Notice that the sets $\mathcal I(Z^k_\a)$, $\mathcal I(Z^{k'}_{\mathfrak a'})$ do not need to be disjoint even if $k\neq k'$ and $\mathfrak a\neq \mathfrak a'$, and the same for $\mathcal E(Z^k_\a)$, $\mathcal E(Z^{k'}_{\mathfrak a'})$.
Moreover,
\begin{equation}
\label{E_infin_incl}
\mathcal I(Z^k_\a) \cup \mathcal E(Z^k_\a)\subset \partial_{\mathrm{rel}}Z^k_\a,
\end{equation}
but the inclusion \eqref{E_infin_incl} may be strict (see Figure \ref{Fi_direct_fibra} and Figure \ref{Fi_infin}).
The measurability of the sets of initial/final points is proven in the Lemma \ref{L_regularity_initial_final_points}. In the proof we use the concept of \emph{completeness} of a directed locally affine partition, whose meaning will be clear in Section \ref{S_foliations} and whose definition is given below. Since up to that section, when it will become crucial for our analysis of the super/subdifferential partitions, such a property will be used only in order to prove measurability issues, more precisely in the proofs of Lemma \ref{L_regularity_initial_final_points} and of Proposition \ref{P_countable_partition_in_reference_directed_planes}, a deeper understanding of its meaning is up to then not necessary and can be for the moment neglected.
\begin{definition}
\label{D_part_compl}
A directed locally affine partition $\{Z^k_\a,C^k_\a\}_{k,\mathfrak a}$ is \emph{complete} if
\begin{equation}
\label{E_part_compl}
x+C^k_\a\,\cap\, y-C^k_\a\subset Z^k_\a,\qquad\forall\,x,y\in Z^k_\a.
\end{equation}
\end{definition}
In the proof of Proposition \ref{P_parall}, we will see that the set defined in \eqref{E_part_compl} is a convex set satisfying
\[
\mathbb{R}^+\bigl((x+C^k_\a\,\cap\, y-C^k_\a)-x\bigr)=\mathbb{R}^+\bigl(y-(x+C^k_\a\,\cap\, y-C^k_\a)\bigr)=C^k_\a.
\]
\begin{definition}
\label{D_dir_subpart}
We will say that a directed locally affine partition $\{Z^{',\ell}_\mathfrak b,\,C^{',\ell}_\mathfrak b\}_{\nfrac{\ell = 0,\dots,d}{\mathfrak b \in \mathfrak B^\ell}}$ in $\mathbb{R}^d$ is a \emph{directed locally affine subpartition} of $\{Z^k_\mathfrak a,\,C^k_\mathfrak a\}_{\nfrac{k = 0,\dots,d}{\mathfrak a \in \mathfrak A^k}}$ if the following holds:
\begin{enumerate}
\item \label{E_same_base} $\mathbf Z=\mathbf Z'$, where $\mathbf Z'$ is the set given by \eqref{E_mathbf_Z_base_partition} for $\{Z^{',\ell}_\mathfrak b,\,C^{',\ell}_\mathfrak b\}_{\ell,\mathfrak b}$;
\item $\forall\,\ell,\mathfrak b$ there exists $k,\mathfrak a$ s.t. $Z^{',\ell}_\mathfrak b\subset Z^k_\mathfrak a$ and $C^{',\ell}_\mathfrak b$ is an extremal face of $C^k_\mathfrak a$.
\end{enumerate}
\end{definition}
\begin{definition}
\label{D_disint_regular}
We say that a locally affine ($\sigma$-compact) partition $\{Z^k_{\mathfrak a}\}_{\nfrac{k = 0,\dots,d}{\mathfrak a \in \mathfrak A^k}}$ is \emph{Lebesgue-regular} if the conditional probabilities $\{\upsilon^k_\mathfrak a\}_{k,\mathfrak a}$ of the disintegration of $\mathcal L^d$ on the partition $\{Z^k_{\mathfrak a}\}_{k,\mathfrak a}$ (see Remark \ref{R_disint_lebesgue}) satisfy
\begin{equation}
\label{E_disint_regular}
\upsilon^k_\mathfrak a \simeq \mathcal H^k \llcorner_{Z^k_\mathfrak a}, \quad \text{for $\eta$-a.e. $(k,\mathfrak a) \in \mathfrak{A}$}.
\end{equation}
\end{definition}
From the definition of disintegration of a Radon measure given in Remark \ref{R_disint_lebesgue}, it is not difficult to check that the validity of \eqref{E_disint_regular} is independent on the partition into unit measure sets $\{A_i\}$, hence Definition \ref{D_disint_regular} is consistent.
To a directed locally affine partition $\{Z^k_\mathfrak a,\,C^k_\mathfrak a\}_{\nfrac{k = 0,\dots,d}{\mathfrak a \in \mathfrak A^k}}$ in $\mathbb{R}^d$, we associate the cost function
\begin{equation}
\label{E_c_bD}
\mathtt c_{\mathbf D}(x,y) :=
\begin{cases}
0 & x\in Z^k_\a, \mathtt c_{C^k_\a}(x,y)<+\infty \ \text{for some} \ (k,\mathfrak a)\in\mathfrak{A}, \\
+\infty & \text{otherwise.}
\end{cases}
\end{equation}
Notice that, since $\mathbf D$ is $\sigma$-compact, $\mathtt c_{\mathbf D}$ is $\sigma$-continuous. Indeed,
\[
\{\mathtt c_{\mathbf D}<+\infty\} = \mathtt p_{x,y} \bigl( \big\{ (k,\mathfrak a,x,y), (k,\mathfrak a,x,y-x)\in\mathbf D \big\} \bigr).
\]
Let us consider $\mu,\nu\in\mathcal P(\mathbb{R}^d)$ satisfying
\[
\quad\Pi^f_{\mathbf c_{\mathbf D}}(\mu,\nu)\neq\emptyset.
\]
By definition of $\mathtt c_{\mathbf D}$, one can easily see that $\mu(\mathbf Z)=1$ and
\[
\Pi^{\mathrm{opt}}_{\mathtt c_{\mathbf D}}(\mu,\nu) = \Pi^f_{\mathtt c_{\mathbf D}}(\mu,\nu)= \Big\{ \pi\in\Pi^f_{\mathtt c_{\mathbf D}}(\mu,\nu):\,\pi\text{ is $\mathtt c_{\mathbf D}$-cyclically monotone} \Big\}.
\] Let
\[
\mu=\int\mu^k_\a\,dm(k,\mathfrak a), \qquad \mu^k_\a(Z^k_\a)=1,
\]
be the disintegration of $\mu$ w.r.t. the partition $\{Z^k_\a\}_{k,\mathfrak a}$.
We have the following characterization.
\begin{proposition}
\label{P_dispiani}
$\pi\in\Pi^f_{\mathtt c_{\mathbf{D}}}(\mu,\nu)$ if and only if the strongly consistent disintegration $\{\pi^k_\mathfrak a\}_{k,\mathfrak a}\subset\mathcal P(\mathbb{R}^d\times\mathbb{R}^d)$ of $\pi$ w.r.t. the partition $\{Z^k_\mathfrak a\times\mathbb{R}^d\}_{k,\mathfrak a}$ satisfies the following properties:
\begin{subequations}
\label{E_L_dispiani}
\begin{equation}
\label{E_L_dispiani2}
\pi^k_\mathfrak a \in \Pi^f_{\mathtt c_{C^k_\mathfrak a}}(\mu^k_\mathfrak a,(\mathtt p_2)_\#\pi^k_\mathfrak a) \quad \text{for $m$-a.e. $(k,\mathfrak a)$},
\end{equation}
\begin{equation}
\label{E_L_dispiani3}
\int (\mathtt p_2)_\# \pi^k_\a\,dm(k,\mathfrak a) = \nu,
\end{equation}
where the measure on the l.h.s. of \eqref{E_L_dispiani3} is defined as in \eqref{E_int_measure}.
\end{subequations}
\end{proposition}
\begin{proof}
If $\pi \in \Pi^f_{\mathtt c_{\mathbf D}}(\mu,\nu)$, then up to an $m$-negligible set one has $\pi^k_\a \in \Pi^f_{\mathtt c_{\mathbf D}}(\mu^k_\a,(\mathtt p_2)_\# \pi^k_\a)$, and since $\mathtt c_{\mathbf D} \llcorner_{Z^k_\a\times\mathbb{R}^d} = \mathtt c_{C^k_\a}$ one deduces \eqref{E_L_dispiani2}. The equality \eqref{E_L_dispiani3} is a fairly easy consequence of $(\mathtt p_2)_\# \pi = \nu$.
Conversely, if $\pi$ satisfies \eqref{E_L_dispiani}, then the two formulas
\[
(\mathtt p_2)_\# \pi = \int (\mathtt p_2)_\# \pi^k_\a dm(k,\mathfrak a), \quad \int \mathtt c_{\mathbf D} d\pi = \int \bigg( \int \mathtt c_{\mathbf D} d\pi^k_\a \bigg) dm(k,\mathfrak a) = \int \bigg( \int \mathtt c_{C^k_\a} d\pi^k_\a \bigg) dm(k,\mathfrak a)
\]
yield $\pi \in \Pi^f_{\mathtt c_{\mathbf D}}(\mu,\nu)$.
\end{proof}
In other words any optimal transference plan w.r.t. the cost associated to a directed locally affine partition can be decomposed as a family of transference plans on the $k$-dimensional affine hulls of the $k$-dimensional sets of the partition, moving the mass along the cones of directions, and viceversa it can be reconstructed given a family $\{\pi^k_\mathfrak a\}_{k,\mathfrak a}$ satisfying \eqref{E_L_dispiani2}-\eqref{E_L_dispiani3}.
In general \eqref{E_L_dispiani3} is not a disintegration (see Section \ref{Ss_measure_disintegration} before Remark \ref{R_disint_lebesgue}), as the following example shows.
\begin{example}
\label{Ex_2ndmarg}
\begin{figure}
\input{sudakovnotuniquebis.pdf_t}
\caption{The transport problem on the directed locally affine partition described in Example \ref{Ex_2ndmarg}.}
\label{Fi_notuniqe}
\end{figure}
For $d=3$ let
\[
\mu := \mathcal H^1 \llcorner_{\{-1\} \times [0,1/2] \times \{0\} \cup \{1\} \times [0,1/2] \times \{0\}}, \quad \nu := 2 \mathcal H^1 \llcorner_{\{0\} \times [0,1/2] \times \{1\}},
\]
and consider the directed locally affine partition
\[
Z^2_1 := \big\{ (z_1,z_2,z_3), z_1 < 0, z_1 = z_3 - 1 \big\}, \quad C^2_1 := \big\{ (z_1,z_2,z_3) : |z_2| \leq z_1, z_1 = z_3 \big\},
\]
\[
Z^2_2 := \big\{ (z_1,z_2,z_3), z_1 > 0, z_1 = - z_3 + 1 \big\}, \quad C^2_2 := \big\{ (z_1,z_2,z_3) : |z_2| \leq - z_1, z_1 = - z_3 \big\}.
\]
Then, for every decomposition $2\nu = \nu_1 + \nu_2$ with $\nu_1, \nu_2 \in \mathcal P(\mathbb{R}^2)$,
\[
\Pi(\mu,\{\nu_1,\nu_2\}) := \Bigl\{ \pi \in \Pi^f_{\mathtt c_{\mathbf D}}(\mu,\nu) : (\mathtt p_2)_{\#}\pi^2_1 = \nu_1,\,(\mathtt p_2)_{\#}\pi^2_2 = \nu_2 \Bigr\} \neq \emptyset,
\]
and clearly $\Pi(\mu,\{\nu_1,\nu_2\}) \subset \Pi^f_{\mathtt c_\mathbf D}(\mu,\nu)$.
\end{example}
Example \ref{Ex_2ndmarg} motivates the following definition.
\begin{definition}
\label{D_pinu}
Given a transference plan $\bar\pi\in\Pi^f_{\mathtt c_{\mathbf D}}(\mu,\nu)$, we define the \emph{conditional second marginals} of $\bar\pi$ w.r.t. $\{Z^k_\a,C^k_\a\}_{k,\mathfrak a}$ as
\[
\bar\nu^k_\mathfrak a := (\mathtt p_2)_\#\bar\pi^k_\a, \quad \text{ for } (k,\mathfrak a) \in \mathfrak{A}.
\]
We also set
\begin{equation}
\label{E_optplan_dirpar}
\Pi^f_{\mathtt c_\mathbf D}(\mu,\{\bar\nu^k_\a\}) = \Bigl\{ \pi\in\Pi^f_{\mathtt c_{\mathbf D}}(\mu,\nu):\, (\mathtt p_2)_\#\pi^k_\a=\bar\nu^k_\mathfrak a \text{ for $m$-a.e. }(k,\mathfrak a) \Bigr\},
\end{equation}
and we call \eqref{E_optplan_dirpar} the set of \emph{optimal transport plans on the directed locally affine partition $\mathbf D$} w.r.t. $\mu$ and $\{\bar\nu^k_\a\}$.
\end{definition}
Therefore, in the following by \emph{optimal transportation problem on a directed locally affine partition $\mathbf D$} we mean an optimal transportation problem w.r.t. the cost $\mathtt c_{\mathbf D}$ between measures $\mu$ and $\{\bar\nu^k_\a\}_{k,\mathfrak a}$, being the latter \emph{admissible second marginals}, namely conditional second marginals of at least one transference plan $\pi\in \Pi^f_{\mathtt c_\mathbf D}(\mu,\nu)$.
Notice that in Example \ref{Ex_2ndmarg} the existence of more than one family of admissible second marginals for the given optimal transportation problem would be avoided provided
\begin{equation}
\label{E_nuz1}
\nu(\mathbf Z)=1.
\end{equation}
What \eqref{E_nuz1} implies in general is that any family of admissible second marginals $\{\bar\nu^k_\a\}$ is given by a relabeling of the disintegration of $\nu$ on $Z^k_\a$, but it may not necessarily occur that
\begin{equation}
\label{E_barnuza1}
\bar\nu^k_\a(Z^k_\a)=1
\end{equation}
(see Figure \ref{Fi_notuniqe}). In the next proposition we give a criterion --namely, condition \eqref{E_more_than_complet}-- in order that condition \eqref{E_barnuza1} is satisfied and then there exists just one family of admissible second marginals. Condition \eqref{E_more_than_complet} will be indeed satisfied by directed locally affine partitions called fibrations and $\mathtt c_{\tilde{\mathbf C}}$-foliations given by single Lipschitz graphs (see Corollary \ref{C_plan_fibr} and Proposition \ref{P_hat_bf_D_graph}).
\begin{proposition}
\label{P_dispiani_2}
Assume that
\begin{equation}
\label{E_more_than_complet}
z \in Z^k_\a, z' \in Z^{k'}_{\mathfrak a'} \ \text{for} \ (k,\mathfrak a) \not= (k',\mathfrak a') \quad \Longrightarrow \quad \mathbf Z \cap ( z + C^k_\a ) \cap ( z' + C^{k'}_{\mathfrak a'} ) = \emptyset.
\end{equation}
Hence,
\begin{equation}
(\mathtt p_2)_\#\pi^k_\a\Big(Z^k_\a\cup\mathbb{R}^d\setminus\underset{(\mathfrak a',k')\neq(\mathfrak a,k)}{\bigcup}Z^{k'}_{\mathfrak a'}\Big)=1, \quad\text{for all $\pi\in\Pi^f_{\mathtt c_{\mathbf{D}}}(\mu,\nu)$, $m$-a.e. $(k,\mathfrak a)\in\mathfrak{A}$.}
\end{equation}
Moreover, if $\nu(\mathbf Z)=1$, one has that
\begin{equation*}
(\mathtt p_2)_\#\pi^k_\a = \nu^k_\a,\quad \text{being $\nu = \int \nu^k_\a\,dm(k,\mathfrak a)$, $\nu^k_\a(Z^k_\a)=1$ the disintegration of $\nu$ w.r.t. $\mathbf D$.}
\end{equation*}
\end{proposition}
Hence the conditional second marginals of $\pi \in \Pi^f_{\mathtt c_{\mathbf{D}}}(\mu,\nu)$ are equal to the conditional probabilities of $\nu$, computed via disintegration on $Z^k_\a$. Notice that part of the statement is that the quotient measure $m$ of $\nu$ is the same as for $\mu$.
\begin{proof}
It is fairly easy to see that \eqref{E_more_than_complet} implies that
\[
\mathbf Z \times \mathbb{R}^d\cap \big\{ \mathtt c_{\mathbf D} < +\infty \big\} \subset \bigcup_{\mathfrak a,k} Z^k_\a \times \Big(Z^k_\a\cup\mathbb{R}^d\setminus \mathbf Z\Big),
\]
so that each $\pi \in \Pi^f_{\mathtt c_{\mathbf D}}(\mu,\nu)$ is concentrated on
\[
\bigcup_{\mathfrak a,k} Z^k_\a \times \Big(Z^k_\a\cup\mathbb{R}^d\setminus \mathbf Z\Big),
\]
and this concludes the proof.
\end{proof}
\subsection{From directed partitions to directed fibrations}
\label{Ss_mapping_sheaf_to_fibration}
In the first part of this section we show that a directed locally affine partition is a countable union of directed locally affine partitions whose elements are locally affine sets having the same dimension and whose direction cones are ``close'' to a fixed reference cone. This kind of partitions will be called \emph{sheaf sets}. Then, we will see that the optimal transportation problem on a \emph{$k$-directed sheaf set} --with $k$ denoting the dimension of its locally affine sets-- can be equivalently reformulated as an optimal transportation problem on a $k$-directed sheaf set whose sets are contained in distinct parallel $k$-dimensional planes, called \emph{$k$-directed fibration}. The advantage of this reformulation is that on a $k$-directed fibration all the supports of the second marginals are disjoint, condition \eqref{E_more_than_complet} holds and then (by Corollary \ref{C_plan_fibr}) one can consider the quotient variables of the partition as parameters for a family of independent convex-cone optimal transportation problems in $\mathbb{R}^k$.
Since $0$-dimensional sets, i.e. single points in $\mathbb{R}^d$, are obviously not further partitionable, from now on we will consider partitions into sets of dimension $k\geq1$.
Let $\{Z^k_\a, C^k_\a\}_{\nfrac{k = 1,\dots,d}{\mathfrak a \in \mathfrak A^k}}$ be a locally affine directed partition in $\mathbb{R}^d$. If $\{\mathtt e^k_i\}_{i=1}^k$ are vectors in $\mathbb{R}^d$, define the sets
\begin{equation}
\label{E_cone_e_k_i}
C(\{\mathtt e^k_i\}) := \bigg\{ \sum_{i=1}^k t_i \mathtt e^k_i:\, t_i \in \mathbb{R}^+ \cup \{0\} \bigg\}
\end{equation}
and
\begin{equation}
\label{E_U_square_mathrm_e}
U(\{\mathtt e^k_i\}) := \bigg\{ \sum_{i=1}^k t_i \mathtt e^k_i:\, t_i \in [0,1]\bigg\}.
\end{equation}
Recalling the definitions given in Sections \ref{Ss_intro_affine_subspaces_cones} and \ref{Ss_partitions_intro}, and the completeness property of Definition \ref{D_part_compl}, we have the following result.
\begin{proposition}
\label{P_countable_partition_in_reference_directed_planes}
There exists a countable covering of $\mathbf D$ into disjoint directed locally affine partitions $\{\mathbf D^k_n\}_{\nfrac{k = 1,\dots,d}{n \in \mathbb{N}}}$, with the following properties: $\forall\,n\in\mathbb{N}$., set
\[
\mathfrak A^k_n := \mathtt p_\mathfrak a \mathbf D_n^k, \quad C^k_\a = \mathtt p_{\mathcal C(k,\mathbb{R}^d)} \mathbf D^k_n(\mathfrak a),\quad Z^k_\a=\mathtt p_{\mathbb{R}^d}\mathbf D_n^k(\mathfrak a).
\]
Then, $\mathtt p_{\{1,\dots,k\}}\mathbf D^k_n=\{k\}$ for all $n \in \mathbb{N}$, i.e. the elements $Z^k_\a$, $C^k_\a$ have linear dimension $k$, for $\mathfrak a \in \mathfrak{A}^k_n$, and there exist
\begin{itemize}
\item linearly independent vectors $\{\mathtt e^k_i(n)\}_{i=1}^k\subset \mathbb S^{d-1}$, with linear span
\[
V^k_n=\langle \mathtt e_1^k(n),\dots,\mathtt e_k^k(n)\rangle,
\]
\item a given point $z^k_n \in V^k_n$,
\item constants $r^k_n,\lambda^k_n \in \mathbb{R}^+$,
\item a non degenerate cone $C^k \in \mathcal C(k,\mathbb{R}^d)$, with $C(\{\mathtt e_i^k\}) \subset \mathring C^k$,
\end{itemize}
such that it holds:
\begin{enumerate}
\item \label{Point_1_Proposition_P_countable_partition} the enlargement of the cone $C^k$ by a factor $2r^k_n$ is non-degenerate
\[
C^k(2r^k_n) \in \mathcal C(k,V^k_n);
\]
\item \label{Point_2_Proposition_P_countable_partition} the projections on $V^k_n$ of the cones $C^k_\a$, $\mathfrak a \in \mathfrak{A}^k_n$, have a uniform opening containing $C^k$
\[
C^k(r^k_n)\subset \mathtt p_{V^k_n} \mathring C^k_\mathfrak a;
\]
\item \label{Point_3_Proposition_P_countable_partition} the projections on $V^k_n$ of the cones $C^k_\a$, $\mathfrak a \in \mathfrak{A}^k_n$, are strictly contained in the given cone $C^k(2r^k_n)$
\[
\mathtt p_{V^k_n} C^k_\mathfrak a \subset C^k(2r^k_n);
\]
\item \label{Point_4_Proposition_P_countable_partition} the projection map on $V^k_n$ is nondegenerate
\[
| \mathtt p_{V^k_n} z | \geq 1/\sqrt{2} \quad \text{ for all $z \in C^k_\mathfrak a\cap \mathbb S^{d-1}$, $\mathfrak a \in \mathfrak A^k_n$};
\]
\item the projection of $Z^k_\a$ on $V^k_n$ contains a given cube
\[
z^k_n + \lambda^k_n \, U(\{\mathtt e^k_i(n)\}) \subset \mathtt p_{V^k_n} Z^k_\mathfrak a.
\]
\end{enumerate}
Moreover, if $\mathbf D$ is complete, then the sets $\{\mathbf D^k_n\}$ are Borel.
\end{proposition}
Observe that from Point \eqref{Point_2_Proposition_P_countable_partition} and Point \eqref{Point_4_Proposition_P_countable_partition} above it follows that there exists $\rho > 0$ such that
\begin{equation*}
|\mathtt p_{V^k_n} (z-z')| \geq \rho |z-z'|, \qquad \forall\, z ,z'\in \mathrm{aff}\, Z^k_\mathfrak a,\,\forall \mathfrak a \in \mathfrak A^k_n.
\end{equation*}
\begin{proof
If $V \in \mathcal G(k,\mathbb{R}^d)$, $C \in \mathcal C(k,V)$ and given two real numbers $0<\delta, r<1$ such that $C(2r)\in\mathcal C(k,\mathbb{R}^d)$, consider the subset $L(k,C,r,\delta)$ of $\mathcal C(k,\mathbb{R}^d)$ defined by
\begin{align}
\label{E_L_V_r_def}
L(k,C,r,\delta) := \Big\{ C' \in \mathcal C(k,\mathbb{R}^d) : \ (i)&~ C(r) \subset \mathtt p_V \mathring C', \crcr
(ii)&~ \mathtt p_V C' \subset C(2r), \crcr
(iii)&~ \inf \big\{ |\mathtt p_V z|: z \in C'\cap \mathbb S^{d-1} \big\} > 1 - \delta \Big\}.
\end{align}
If is fairly easy to see that for all $0<\delta<1$ as above the family
\begin{equation}
\label{E_base_of_C_k_R_ell}
\mathfrak L(k,\delta) := \Big\{ L(k,C,r',\delta): C \in \mathcal C(k,\mathbb{R}^d), 0 < r' < 1 \text{ s.t. }C(2r')\in\mathcal C(k,\mathbb{R}^d) \Big\}
\end{equation}
generates a base of neighborhoods of $\mathcal C(k,\mathbb{R}^d)$. In particular, we can find a countable family of sets $\{L(k,C^k_n,r^k_n,1\slash\sqrt 2)\}_{n \in \mathbb{N}}$, covering $\mathcal C(k,\mathbb{R}^d)$ --being the latter separable.
Notice that
\[
\mathrm{clos}\,L(k,C,r,\delta) = \bigg\{ C' \in \mathcal C(k,\mathbb{R}^d) :\,C(r) \subset \mathtt p_V C',\,\mathtt p_V C' \subset C(2r),\,\inf \big\{ |\mathtt p_V z|: z \in C'\cap \mathbb S^{d-1} \big\} \geq 1 - \delta \bigg\}
\]
is compact.
Then, define
\begin{equation*}
\mathbf D^k_n := \bigg\{ \big( \mathfrak a,z,C^k_\mathfrak a \big) \in \mathbf D(k): C^k_\mathfrak a \in \mathrm{clos}\,L \big( k,C^k_n,r^k_n,1/\sqrt{2} \big) \setminus \bigcup_{n' < n} \mathrm{clos}\,L \big( k,C^k_{n'},r^k_{n'},1/\sqrt{2} \big) \bigg\}.
\end{equation*}
Clearly $\{\mathbf D^k_n\}_{n \in \mathbb{N}}$ is a covering of $\mathbf D(k)$ into disjoint sets, and it is fairly straightforward to prove that these sets are $\sigma$-compact, because the sets $\mathrm{clos}\,L(k,C,r,\delta)$ are compact.
For each $k$, $C^k_n$, $r^k_n$, consider a family of $k$ linearly independent unit vectors $\{\mathtt e^k_i(n)\}_{i =1}^k$ in $\mathbb{R}^d$ such that
\begin{equation*}
C(\{\mathtt e^k_i(n)\}) \subset \mathring C^k_n.
\end{equation*}
Being the family of sets $\bigl\{\{z + \lambda \, \inter_{\mathrm{rel}} U(\{\mathtt e^k_i(n)\})\}\bigr\}_{z \in V^k_n,\,\lambda \in \mathbb{R}^+}$ a base of the topology of $V^k_n$, let $\bigl\{\{z_m + \lambda_m \, \inter_{\mathrm{rel}} U(\{\mathtt e^k_i(n)\})\}\bigr\}_{m \in \mathbb{N}}$ be a countable base. Define thus
\begin{equation*}
\mathbf D^k_{n,m} := \Big\{ (\mathfrak a,z,C^k_\mathfrak a) \in \mathbf D^k_n : z_m + \lambda_m U(\{\mathtt e^k_i(n)\}) \subset \mathtt p_{V^k_n} Z^k_\mathfrak a \Big\} \setminus \bigcup_{m' < m} \mathbf D_{n,m'}^k.
\end{equation*}
Since the directed partition is complete (see Definition \ref{D_part_compl}) and Point \eqref{Point_2_Proposition_P_countable_partition} holds then
\[
z_m + \lambda_m U(\{\mathtt e^k_i(n)\}) \subset \mathtt p_{V^k_n} Z^k_\mathfrak a\quad\Leftrightarrow\quad \Big\{z_m,\,z_m+\lambda_m\sum_{i=1}^k\mathtt e_i^k(n)\Big\}\subset \mathtt p_{V^k_n} Z^k_\mathfrak a.
\]
Let $\mathbf f^k_n:\,\mathtt p_{\mathfrak{A}\times\mathbb{R}^d}\mathbf D^k_n \rightarrow\mathtt p_{V^k_n}\bigl(\mathtt p_{\mathfrak{A}\times\mathbb{R}^d}\mathbf D^k_n\bigr)$ be the $\sigma$-continuous map
\[
\mathbf f^k_n(\mathfrak a,z)=\mathtt p_{V^k_n}(Z^k_\a).
\]
One has that
\begin{equation}
\label{E_sheafcomp}
\begin{split}
\bigg\{ (\mathfrak a,z) \in\mathtt p_{\mathfrak{A}\times\mathbb{R}^d}\mathbf D^k_n :\, \bigg\{z_m,\,z_m+\lambda_m& \sum_{i=1}^k\mathtt e_i^k(n) \bigg\}\subset \mathtt p_{V^k_n} Z^k_\mathfrak a \bigg\} \\
=&~ \mathtt p_{12} \Big(\mathrm{graph}\, \mathbf f^k_n\cap \big\{(\mathfrak a,z,z_m):\,\mathfrak a\in\mathfrak{A},z\in\mathbb{R}^d \big\} \Big) \\
&\cap\mathtt p_{12} \bigg( \mathrm{graph}\, \mathbf f^k_n \cap \bigg\{ \bigg( \mathfrak a,z, z_m+\lambda_m \sum_{i=1}^k\mathtt e_i^k(n) \bigg) :\, \mathfrak a\in\mathfrak{A}, z\in\mathbb{R}^d \bigg\} \bigg),
\end{split}
\end{equation}
is a $\sigma$-compact set, thus $\mathbf D^k_{n,m}$ is Borel. Relabeling the sets $\mathbf D^k_{n,m}$ as $\mathbf D^k_n$, the proof is completed.
\end{proof}
\begin{remark}
\label{R_sheaf_negl}
In the rest of this section, without further comments, we will assume that the directed locally affine partitions are complete, according to Definition \ref{D_part_compl}. Indeed, this will be always the case for the partitions we analyze in the paper. Since we will be interested into directed locally affine partitions up to sets which are $\varpi$-negligible w.r.t. some fixed measure $\varpi$, we will also consider the sets of the countable partition $\{\mathbf D^k_n\}$ as $\sigma$-compact, which is always the case provided we remove an $\varpi$-negligible set.
\end{remark}
\begin{figure}
\input{sudakovsheafdefbis.pdf_t}
\caption{The locally affine sets $Z^2_1$, $Z^2_2$, $Z^2_3$ with cones of directions $C^2_1$, $C^2_2$, $C^2_3$ (given by the union of the cyan and the yellow triangles) form a $2$-dimensional directed sheaf set with reference plane $V^2$, base vectors $\mathtt e^2_1,\,\mathtt e^2_2$, reference rectangle $z^2+\lambda^2 U(\{\mathtt e^2_1,\,\mathtt e^2_2\})$ and base cones of directions $C(\{\mathtt e^2_1,\mathtt e^2_2\})(r^2)\subset V^2$ (colored in cyan) and $C(\{\mathtt e^2_1,\mathtt e^2_2\})(2r^2)\subset V^2$ (colored in blue). In the picture, we underline with the corresponding color the counterimages of the reference cones on the affine spans of the locally affine sets. The remaining locally affine set of the partition does not belong to the sheaf set since the projection of its cone of directions (colored in yellow) on the reference plane does not contain/is contained in the reference cone.}
\label{Fi_sheafdef}
\end{figure}
\begin{definition}
\label{D_sheaf_set}
For $k=1,\dots,d$, we call \emph{$k$-(dimensional) directed sheaf set} a $\sigma$-compact directed locally affine partition into $k$-dimensional sets $\mathbf D^k$ which satisfy the same properties of the sets $\mathbf D^k_n$ in Proposition \ref{P_countable_partition_in_reference_directed_planes}: there exist
\begin{itemize}
\item linearly independent vectors $\{\mathtt e^k_i\}_{i=1}^k\subset \mathbb S^{d-1}$, with linear span
\[
V^k = \langle \mathtt e_1^k,\dots,\mathtt e_k^k \rangle,
\]
\item a given point $z^k \in V^k$,
\item constants $r^k,\lambda^k \in \mathbb{R}^+$,
\item a non degenerate cone $C^k \in \mathcal C(k,\mathbb{R}^d)$, with $C(\{\mathtt e_i^k\})(r_k) \subset \mathring C^k$,
\end{itemize}
such that, denoting
$\mathfrak A^k := \mathtt p_\mathfrak a \mathbf D^k$, $C^k_\a = \mathtt p_{\mathcal C(k,\mathbb{R}^d)} \mathbf D^k(\mathfrak a)$, $Z^k_\a=\mathtt p_{\mathbb{R}^d}\mathbf D^k(\mathfrak a)$,
it holds:
\begin{enumerate}
\item
$C^k(r^k) \in \mathcal C(k,V^k)$;
\item
$C^k \subset \mathtt p_{V^k} \mathring C^k_\mathfrak a$;
\item
$\mathtt p_{V^k} C^k_\mathfrak a \subset C^k(r^k)$ for all $\mathfrak a \in \mathfrak A^k$
\item
$|\mathtt p_{V^k} z| \geq 1/\sqrt{2}$ for all $z \in C^k_\mathfrak a\cap \mathbb S^{d-1}$, $\mathfrak a \in \mathfrak A^k$;
\item
$z^k + \lambda^k \, U(\{\mathtt e^k_i\}) \subset \mathtt p_{V^k} Z^k_\mathfrak a$.
\end{enumerate}
The $k$-dimensional plane $V^k=\langle \mathtt e_1^k,\dots \mathtt e_k^k\rangle$ will be called \emph{reference plane}, the cones $C^k \subset C^k(r^k) =:C^{',k}$ \emph{base cones of directions}, $z^k$ \emph{base point} and $z^k + \lambda^k \, U(\{\mathtt e^k_i\})$ \emph{base rectangle} of the sheaf set.
\end{definition}
Moreover, we can choose
\begin{equation}
\label{E_mathfrak_A_k_def}
\mathfrak A^k := \mathbf Z^k \cap \big( z^k + (V^k)^\perp \big).
\end{equation}
In this way the quotient space $\mathfrak A^k$ is a subset of a $(d-k)$-dimensional affine space. This will be our default choice for the quotient space $\mathfrak{A}^k$ of $\mathbf D^k$.
Before going on, we prove the following lemma, announced in Section \ref{Ss_partitions_intro}.
\begin{lemma}
\label{L_regularity_initial_final_points}
If the directed locally affine partition is $\sigma$-compact and complete, then the sets $\mathcal I$, $\mathcal E$ are Souslin.
\end{lemma}
\begin{proof}
We prove the statement only for $\mathcal I$, since the proof for $\mathcal E$ is analogous. Moreover, we can consider w.l.o.g. a directed locally affine partition given by a $\sigma$-compact directed sheaf set as in Definition \ref{D_sheaf_set}.
Let $\mathfrak{A} = \underset{l\in\mathbb{N}}{\cup}\, \mathfrak{A}_l$ such that the sets $\mathtt p_{\mathfrak{A}_l\times\mathcal C(k,\mathbb{R}^d)}\mathbf D(k)$ are compact,
and for $n\in\mathbb{N}$ define
\[
\mathcal I^k_{n,l} := \Bigl\{(\mathfrak a,z)\in\mathfrak{A}_l\times\mathbb{R}^d :\, z\in\mathrm{clos}\,Z^k_\mathfrak a,\quad z + \inter_{\mathrm{rel}} C^k_{\mathfrak a,l} \cap B^d(z,2^{-n}) \subset Z^k_{\mathfrak a} \Bigr\}.
\]
By the completeness property of the sheaf set, there exists $n'$ such that
\begin{equation}
\label{E_iknl}
\mathcal I^k_{n,l}:= \Bigl\{(\mathfrak a,z)\in\mathfrak{A}_l\times\mathbb{R}^d :\, z\in\mathrm{clos}\,Z^k_\mathfrak a,\quad z + \inter_{\mathrm{rel}} C^k_{\mathfrak a,l} \cap \partial B^d(z,2^{-n'})\cap Z^k_\a\neq\emptyset \Bigr\}.
\end{equation}
Since the set $\{(\mathfrak a,z):\,z\in\mathrm{clos}\,Z^k_\mathfrak a\}$ is Borel, then by \eqref{E_iknl}, reasoning as in \eqref{E_sheafcomp} the sets $\mathcal I^k_{n,l}$ are Borel too and finally
\[
\mathcal I=\mathtt p_{\mathbb{R}^d}\Big(\underset{k,n,l}{\bigcup}\,\mathcal I^k_{n,l}\Big)\setminus \mathbf Z
\]
is Souslin.
\end{proof}
Now we show that the graph $\mathbf D^k$ of a $k$-directed sheaf set in $\mathbb{R}^d$ can be mapped injectively into a subset of $\mathbb{R}^{d-k}\times\mathbb{R}^k$ called \emph{fibration}, consisting of a family of parallel $k$-dimensional locally affine sets. In this section, points in $\mathbb{R}^{d-k}\times\mathbb{R}^k$ will be denoted as $(\mathfrak q,w)$.
\begin{figure}
\centerline{\resizebox{14cm}{12cm}{\input{sudakovmappingfibrabis.pdf_t}}}
\caption{The mapping of a $2$-dimensional sheaf set into a fibration, Proposition \ref{P_map_sheaf_set_into_fibration}. The pink region denotes the support of the conditional measures $\mu^2_\mathfrak a$ (resp. $\tilde\mu^2_\mathfrak q$), the yellow one the support of the conditional measures $\nu^2_\mathfrak a$ (resp. $\tilde\nu^2_\mathfrak q$), and the blue cones $C^2_\mathfrak a$ (resp. $\tilde {\mathbf C}^2(\mathfrak q)$) are the cones of directions of the locally affine sets $Z^2_\mathfrak a$ (in gray color). The reference cones $C^2\subset C^{',2}$ and $\bar C^2\subset\bar C^{',2}$ are also depicted.}
\label{Fi_mappingfibra}
\end{figure}
\begin{definition}
\label{D_fibration}
A \emph{$k$-(dimensional) directed fibration} is a $\sigma$-compact set $\tilde{\mathbf D}^k\subset\mathbb{R}^{d-k}\times\mathbb{R}^k$ endowed with a $\sigma$-continuous map $\tilde{\mathbf C}^k:\mathtt p_{\mathbb{R}^{d-k}}(\tilde{\mathbf D}^k)\rightarrow\mathcal C(k,\mathbb{R}^k)$, $\mathfrak q \mapsto \tilde{\mathbf C}^k(\mathfrak q)$, such that
\begin{align}
&\tilde{\mathbf D}^k(\mathfrak q)\text{ is open in $\mathbb{R}^k$,} \notag \\
&\exists\, \bar C^k \subset \bar C^{',k}, \ \text{with} \ \bar C^k, \bar C^{',k} \in \mathcal C(k,\mathbb{R}^k), \quad \text{ s.t. } \quad \bar C^k \subset \mathring{\tilde{\mathbf C}}^k(\mathfrak q) \subset \tilde{\mathbf C}^k(\mathfrak q) \subset \mathring{\bar{C}}^{',k}, \label{E_cones_fibration}
\end{align}
for all $\mathfrak q \in \mathtt p_{\mathbb{R}^{d-k}}(\tilde{\mathbf D}^k)$.
\end{definition}
To a directed fibration $\tilde\mathbf D^k$ we associate the cost
\begin{equation}
\label{E_cost_fibr}
\mathtt c_{\tilde \mathbf D^k}(\mathfrak q,w,\mathfrak q',w') =
\begin{cases}
0 & \mathfrak q = \mathfrak q', w'-w \in \tilde {\mathbf C}^k(\mathfrak q), \\
+\infty & \text{otherwise.}
\end{cases}
\end{equation}
Recalling the definition given in \eqref{E_cone_cost}, notice that
\[
\mathtt c_{\tilde{\mathbf D}^k}(\mathfrak q,w,\mathfrak q',w')=
\begin{cases}
\mathtt c_{\tilde{\mathbf C}^k(\mathfrak q)}(w,w') & \text{ if $\mathfrak q = \mathfrak q'$}, \\
+\infty & \text{otherwise.}
\end{cases}
\]
\begin{proposition}
\label{P_map_sheaf_set_into_fibration}
Let $\mathbf D^k$ be a $k$-dimensional directed sheaf set with base cones $C^k\subset C^{',k}\in\mathcal C(k,\mathbb{R}^d)$, reference plane $V^k$, base point $z^k$ and quotient space $\mathfrak{A}^k=Z^k\cap(z^k+(V^k)^{\perp})$ and define the map
\begin{equation}
\label{E_r_map}
\mathbf r:\mathfrak{A}^k \times \mathbb{R}^d \ni (\mathfrak a,z) \mapsto \Big( \mathtt i_{d-k}(\mathfrak a), \mathtt i_k \circ \mathtt p_{V^k}(z) \Big) \in \mathbb{R}^{d-k} \times \mathbb{R}^k,
\end{equation}
where
\begin{equation}
\label{E_identif_map}
\mathtt i_{d-k}:z^k+(V^k)^\perp\rightarrow\mathbb{R}^{d-k},\quad\mathtt i_k:V^k\rightarrow\mathbb{R}^k
\end{equation}
are the identification maps.
Then, $\mathbf r_{|_{\underset{\mathfrak a\in\mathfrak{A}^k}{\cup}\{\mathfrak a\}\times \mathrm{aff} Z^k_\a}}$ is a bijection onto $\underset{\mathfrak q\in\mathtt i_{d-k}(\mathfrak{A}^k)}{\cup}\{\mathfrak q\}\times\mathbb{R}^k$ and $\mathbf r_{|_{\underset{\mathfrak a\in\mathfrak{A}^k}{\cup}\{\mathfrak a\}\times Z^k_\a}}=\mathbf r_{|_{\mathtt p_{\mathfrak{A}^k\times\mathbb{R}^d}(\mathbf D^k)}}$ maps the sheaf set $\mathtt p_{\mathfrak{A}^k\times\mathbb{R}^d}(\mathbf D^k)$ into the elements of the fibration $\tilde{\mathbf D}^k:=\mathbf r(\mathtt p_{\mathfrak{A}^k\times\mathbb{R}^d}(\mathbf D^k))$ endowed with the direction map
\begin{equation*}
\tilde{\mathbf C}^k:\mathtt i_{d-k}(\mathfrak{A}^k)\ni\mathfrak q\mapsto \mathtt i_k\circ\mathtt p_{V^k}\circ\mathtt p_{\mathcal C(k,\mathbb{R}^d)}\bigl(\mathbf D^k(\mathtt i_{d-k}^{-1}(\mathfrak q))\bigr).
\end{equation*}
Moreover,
\begin{equation}
\label{E_rxr}
(\mathbf r\times\mathbf r)_\#\Pi^f_{\mathtt c_{\mathbf D^k}}(\mu,\{\nu^k_\mathfrak a\}) = \Pi^f_{\mathtt c_{\tilde\mathbf D^k}}(\hat \mu,\hat \nu),
\end{equation}
where
\begin{align}
\label{E_hat_mu_nu}
\tilde \mu = \int (\mathtt i_{k} \circ \mathtt p_{V^k})_\#\mu^k_{\mathtt i_{d-k}^{-1}(\mathfrak q)}\,d\mathtt i_{d-k}\#m(\mathfrak q), \quad \tilde \nu = \int (\mathtt i_{k} \circ \mathtt p_{V^k})_\#\nu^k_{\mathtt i_{d-k}^{-1}(\mathfrak q)}\,d\mathtt i_{d-k}\#m(\mathfrak q).
\end{align}
Finally
\begin{align*}
&(\mathtt i_{k} \circ \mathtt p_{V^k})_\# \mu^k_{\mathtt i_{d-k}^{-1}(\mathfrak q)} \big( \mathtt i_{k} \circ ( \mathbf r(B) \cap \mathbb{R}^{d-k} \times \{\mathfrak q\} ) \big) = 0 \quad \Longleftrightarrow \quad \mu^k_{\mathtt i_{d-k}^{-1}(\mathfrak q)}(B) = 0, \quad \forall\, B \in \mathcal B(\mathbb{R}^d) \cap \mathrm{aff} Z^k_\a, \\
&(\mathtt i_{k} \circ \mathtt p_{V^k})_\# \nu^k_{\mathtt i_{d-k}^{-1}(\mathfrak q)} \big( \mathtt i_{k} \circ ( )\mathbf r(B) \cap \mathbb{R}^{d-k} \times \{\mathfrak q\} ) \big) = 0 \quad \Longleftrightarrow \quad \nu^k_{\mathtt i_{d-k}^{-1}(\mathfrak q)}(B) = 0, \quad \forall\, B \in \mathcal B(\mathbb{R}^d) \cap \mathrm{aff} Z^k_\a.
\end{align*}
\end{proposition}
\begin{proof}
The functions $\mathbf r$ and $\tilde{\mathbf C}^k$ are $\sigma$-continuous, because their graphs are projections of $\sigma$-compact sets.
The facts that $\mathbf r$ is a bijection and that $\mathbf r(\mathtt p_{\mathfrak{A}^k\times\mathbb{R}^d}\mathbf D^k)$ is a fibration are straightforward, observing that \eqref{E_cones_fibration} is satisfied by the cones $\bar{C}^k=\mathtt i_k(C^k)$, $\bar{C}^{',k}=\mathtt i_k(C^{',k})$, thanks to $(2)$, $(3)$ and $(4)$ of Definition \ref{D_sheaf_set}.
As for the last statements, it is sufficient to observe that
\[
\mathtt c_{\mathbf D^k}((\mathfrak a,z),(\mathfrak a',z'))=\mathtt c_{\tilde\mathbf D^k}(\mathbf r(\mathfrak a,z),\mathbf r(\mathfrak a',z')) \cdot \mathbbm 1_{\tilde \mathbf D^k \times \mathbb{R}^{d-k} \times \mathbb{R}^k}(\mathbf r(\mathfrak a,z),\mathbf r(\mathfrak a',z')),
\]
and that $\mathbf r_{|_{\{\mathfrak a\}\times \mathrm{aff} Z^k_\a}}$ is bi-Lipschitz, $\forall\,\mathfrak a\in\mathfrak{A}^k$.
\end{proof}
In the following, we set
\begin{equation}
\label{E_tilde_mu_nu_k_a}
\tilde\mu^k_\mathfrak q = (\mathtt i_{k} \circ \mathtt p_{V^k})_\# \mu^k_{\mathtt i_{d-k}^{-1}(\mathfrak q)}, \quad \tilde\nu^k_\mathfrak q = (\mathtt i_{k} \circ \mathtt p_{V^k})_\# \nu^k_{\mathtt i_{d-k}^{-1}(\mathfrak q)}, \quad \tilde\pi^k_\mathfrak q = (\mathtt i_{k} \circ \mathtt p_{V^k} \times \mathtt i_{k} \circ \mathtt p_{V^k})_\# \pi^k_{\mathtt i_{d-k}^{-1}(\mathfrak q)}.
\end{equation}
\begin{remark}
\label{rem_sheaffibr}
Notice that, once we fix an orthogonal basis $\{\mathtt e_i\}_{i=1}^d\subset\mathbb{R}^d$ and identify $\mathbb{R}^{d-k}\times\mathbb{R}^k\simeq \langle \mathtt e_1,\dots,\mathtt e_{d-k}\rangle\times\langle \mathtt e_{d-k+1},\dots,\mathtt e_{d}\rangle=\mathbb{R}^d$, a $k$-directed fibration is the image through the map $\mathtt i_{d-k}\times\mathtt i_k$ of a $k$-directed sheaf set whose reference $k$-plane is $\{0\}\times\langle \mathtt e_{d-k+1},\dots,\mathtt e_{d}\rangle$. Therefore, we can think of a $k$-directed fibration as a $k$-directed sheaf set whose sets are contained in disjoint parallel $k$-planes. In particular, when we speak about directed locally affine subpartitions of a fibration (as e.g. in Proposition \ref{P_sub_sheaf_fol}) we mean the image through the map $\mathtt i_{d-k}\times\mathtt i_k$ of directed locally affine subpartitions of the corresponding $k$-directed sheaf set.
\end{remark}
\begin{proposition}
\label{P_sub_sheaf_fol}
Let $\{Z^\ell_{\mathfrak a,\mathfrak b},C^\ell_{\mathfrak a,\mathfrak b}\}_{\nfrac{\ell=0,\dots,k}{\mathfrak a \in \mathfrak{A}^k,\mathfrak b \in \mathfrak B^\ell}}$ be a locally affine directed subpartition of the sheaf set $\{Z^k_\a,C^k_\a\}_{k,\mathfrak a}$. Then the sets
\[
\tilde Z^\ell_{\mathfrak q,\mathfrak b} = \mathbf r(\mathfrak a,Z^\ell_{\mathfrak a,\mathfrak b}),\quad \tilde C^\ell_{\mathfrak q,\mathfrak b}=\mathtt i_{k}\circ\mathtt p_{V^k}(C^\ell_{\mathfrak a,\mathfrak b})
\]
form a directed locally affine subpartition of the fibration $\tilde{\mathbf D}^k=\mathbf r(\mathtt p_{\mathfrak{A}^k\times\mathbb{R}^d}(\mathbf D^k))$, and viceversa.
If moreover the subpartition $\{Z^\ell_{\mathfrak a,\mathfrak b},C^\ell_{\mathfrak a,\mathfrak b}\}_{\ell,\mathfrak a,\mathfrak b}$ is regular, then also the subpartition $\{\tilde Z^\ell_{\mathfrak q,\mathfrak b},\tilde C^\ell_{\mathfrak q,\mathfrak b}\}_{\ell,\mathfrak q,\mathfrak b}$ is regular, and viceversa.
\end{proposition}
\begin{proof}
Since $\mathtt r_{|_{\{\mathfrak a\} \times \mathrm{aff} \, Z^k_\a}}$ is an invertible projection, the first part of the statement is obvious.
The same reasoning holds for the regularity of the measures.
\end{proof}
Recalling Propositions \ref{P_dispiani} and \ref{P_dispiani_2}, it is fairly easy to prove the following
\begin{corollary}
\label{C_plan_fibr}
Let $\tilde\mathbf D^k$ be a $k$-directed fibration and $\tilde\mu$, $\tilde\nu\in\mathcal P(\mathbb{R}^{d-k}\times\mathbb{R}^k)$ s.t. $\Pi^f_{\mathtt c_{\tilde\mathbf D^k}}(\tilde\mu,\tilde\nu)\neq\emptyset$. Then,
\begin{equation}
\tilde\pi\in\Pi^f_{\mathtt c_{\tilde\mathbf D^k}}(\tilde\mu,\tilde\nu)\quad\Leftrightarrow\quad\tilde\pi^k_\mathfrak q\in\Pi^f_{\mathtt c_{\tilde{\mathbf C}^k(\mathfrak q)}}(\tilde\mu^k_\mathfrak q,\tilde\nu^k_\mathfrak q),
\end{equation}
being $\tilde\mu=\int\tilde\mu^k_\mathfrak q \, d\tilde m(\mathfrak q)$, $\tilde\nu=\int\tilde\nu^k_\mathfrak q \, d\tilde m(\mathfrak q)$ and $\tilde\pi=\int\tilde\pi^k_\mathfrak q \, d\tilde m(\mathfrak q)$ be respectively the disintegrations of $\tilde\mu$, $\tilde\nu$ w.r.t. the partition $\{\{\mathfrak q\}\times \mathbb{R}^k\}_{\mathfrak q\in\mathtt i_{d-k}(\mathfrak{A}^k)}$ and the disintegration of $\tilde\pi$ w.r.t. $\{\{\mathfrak q\}\times \mathbb{R}^k\times \{\mathfrak q\}\times \mathbb{R}^k\}_{\mathfrak q\in\mathtt i_{d-k}(\mathfrak{A}^k)}$.
\end{corollary}
\section{Directed locally affine partitions on cone-Lipschitz foliations}
\label{S_foliations}
In the first part of this section we generalize the notion of graph of a $\d{\cdot}$-Lipschitz function up to the definition of a $\mathtt c_{\tilde{\mathbf C}}$-Lipschitz foliation, where $\tilde{\mathbf C}$ is the family of cones of directions associated to a given $k$-dimensional fibration $\tilde\mathbf D\subset\mathbb{R}^{d-k}\times\mathbb{R}^k$. From now on we fix $k$ and we drop the superscript $k$ in the notation for a $k$-directed fibration or $k$-dimensional cone. Moreover we will replace the variable $\mathfrak q$ with $\mathfrak a$, since it is clear from Proposition \ref{P_map_sheaf_set_into_fibration} and Remark \ref{rem_sheaffibr} that the quotient spaces of a sheaf set and of the corresponding fibration can be identified.
\noindent In particular, for any fixed $\mathfrak a\in\mathfrak{A}=\mathtt p_{\mathbb{R}^{d-k}}(\tilde\mathbf D)$, the intersection of a $\mathtt c_{\tilde{\mathbf C}}$-Lipschitz foliation with $\{\mathfrak a\}\times\mathbb{R}^k$ will be a suitable collection of disjoint \emph{(complete) $\mathtt c_{\tilde{\mathbf C}(\mathfrak a)}$-Lipschitz graphs} --namely, graphs of $|\cdot|_{\tilde D(\mathfrak a)^*}$-Lipschitz functions where $\tilde D(\mathfrak a)$ is convex set s.t. $\tilde{ \mathbf C}(\mathfrak a)=\mathrm{epi}\,|\cdot|_{\tilde D(\mathfrak a)^*}$ (see Definition \ref{D_complete_G})-- and at most countably many sets with nonempty interior.
Next, we generalize the notion of super/subdifferentials given in Definition \ref{D_subsuperdiff} for single graphs of $\d{\cdot}$-Lipschitz functions to this new class of objects: at each point $w\in\tilde\mathbf D(\mathfrak a)$, the superdifferential will be the intersection of the cone $w+\tilde{\mathbf C}(\mathfrak a)$ with the $\mathtt c_{\tilde{\mathbf C}(\mathfrak a)}$-Lipschitz graph to which $w$ belongs.
Our main result is Theorem \ref{T_partition_E+-}, in which we prove that, up to a residual set, a $\mathtt c_{\tilde{\mathbf C}}$-Lipschitz foliation can be decomposed into a directed locally affine partition whose cone of directions at each point is given by the super/subdifferential.
Moreover, in Theorem \ref{T_partition_E} we characterize the residual set as the set of \emph{initial}/\emph{final points} of the super/subdifferential partitions (see Definition \ref{D_initial_final}).
\subsection{Convex cone-Lipschitz graphs}
\label{Ss_cone_lipschitz_graph}
Let $\tilde{C}$ be the epigraph of a convex norm $|\cdot|_{\tilde{D}^*}:\mathbb{R}^{k-1}\rightarrow\mathbb{R}$: by Remark \ref{R_cone_epi}, as a subset of $\mathbb{R}^k=\mathbb{R}^{k-1}\times\mathbb{R}$, $\tilde{C}\in\mathcal C(k,\mathbb{R}^k)$. We
denote variables in $\mathbb{R}^k$ as $w=(x,y)\in\mathbb{R}^{k-1}\times\mathbb{R}$ and we let $\mathtt c_{\tilde{C}} : \mathbb{R}^k \times \mathbb{R}^k \rightarrow [0,+\infty]$ be the related \emph{convex cone cost} (see Definition \eqref{D_cone_cost}).
Now we introduce a class of subsets of $\mathbb{R}^k$ which includes the graphs of $|\cdot|_{\tilde{D}^*}$-Lipschitz functions $\varphi:\mathbb{R}^{k-1}\rightarrow\mathbb{R}$. In particular, when $k=d+1$ and $\tilde{C}=\mathrm{epi}\,\d{\cdot}$, this class contains the graphs of the Kantorovich potentials $\psi$ for the transport problem with cost \eqref{E_norm_cost_1} (see Section \ref{Ss_convex_norm_cone}).
\begin{definition}
\label{D_complete_G}
A set $G\subset \mathbb{R}^k$ is a \emph{$\mathtt c_{\tilde{C}}$-Lipschitz graph} if
\begin{subequations}
\label{E_G}
\begin{equation}
\label{E_G1}
G \times G \cap \big\{ \mathtt c_{\tilde{C}} < +\infty \big\} \subset \big\{ (w,w') : w'-w \in \partial \tilde{C} \big\}.
\end{equation}
Moreover, a $\mathtt c_{\tilde{C}}$-Lipschitz graph $G\subset\mathbb{R}^k$ is \emph{complete} if
\begin{equation}
\label{E_G2}
\mathtt{O}(w,w') : =\big\{ \mathtt c_{\tilde{C}}(w,\cdot) < +\infty \big\} \cap \big\{ \mathtt c_{\tilde{C}}(\cdot,w') < +\infty \big\} \subset G, \quad \forall\, w,w' \in G.
\end{equation}
\end{subequations}
\end{definition}
Notice that \eqref{E_G1} is equivalent to
\[
w'\notin w\,\pm\,\mathrm{int}\, \tilde{C}, \qquad \forall\,w,w'\in G,
\]
which can be rephrased as
\begin{equation}
\label{E_phi_graph_G}
G = \mathrm{graph}\,\varphi_G \quad \text{ for some $\varphi_G:\mathrm{dom}\,\varphi_G\subset\mathbb{R}^{k-1}\rightarrow\mathbb{R}$, $\varphi_G$ $\d{\cdot}$-Lipschitz}.
\end{equation}
The second condition \eqref{E_G2} yields
\[
G \supset \mathtt{O}(G\times G) := \bigcup_{w,w'\in G} \mathtt{O}(w,w').
\]
\begin{remark}
\label{R_lipgraph}
If $G=\mathrm{graph}\,\varphi_G$ with $\varphi_G$ $|\cdot|_{\tilde{D}^*}$-Lipschitz and $\mathrm{dom}\,\varphi_G=\mathbb{R}^{k-1}$, then $G$ is a complete $\mathtt c_{\tilde{C}}$-Lipschitz graph. The analysis of this particular case will be sufficient for the proof of Theorem \ref{T_1}. If $\mathrm{dom}\,\varphi_G\neq\mathbb{R}^{k-1}$, we anticipate that the ``completeness'' property \eqref{E_G2} is what we need to construct sets which preserve the properties of the $|\cdot|_{\tilde{D}^*}$-super/subdifferentials of Lipschitz functions on $\mathbb{R}^{k-1}$: these properties are fundamental for our later purposes, culminating with the proof of Theorem \ref{T_partition_E+-}.
\end{remark}
Recalling Definition \ref{D_subsuperdiff} of super/subdifferential of the graph of a $\d{\cdot}$-Lipschitz function $\varphi$ we give the following definition.
\begin{definition}
\label{D_tilde_C_diff}
Given a set $G\subset\mathbb{R}^k$, define the \emph{$\mathtt c_{\tilde{C}}$-superdifferential} and the \emph{$\mathtt c_{\tilde{C}}$-subdifferential} of $G$ respectively as
\begin{equation*}
\partial^+G := G \times G \cap \bigl\{ \mathtt c_{\tilde{C}} < +\infty \bigr\}, \qquad \partial^-G: = G \times G \cap \bigl( \bigl\{ \mathtt c_{\tilde{C}} < +\infty \bigr\} \bigr)^{-1}.
\end{equation*}
\end{definition}
Notice that $\partial^-G=(\partial^+G)^{-1}$ and since
\begin{equation}
\label{E_sum_of_cones_inside_cone}
(w+C)+C\subset w+C, \qquad \forall\,w\in\mathbb{R}^k,\ C \in \bigcup_{\ell=1}^k \mathcal C(\ell,\mathbb{R}^k),
\end{equation}
one deduces the \emph{transitivity property}
\begin{equation*}
w' \in \partial^\pm G(w) \quad \Longrightarrow \quad \partial^\pm G(w') \subset \partial^\pm G(w).
\end{equation*}
The property of a set of being a complete $\mathtt c_{\tilde{C}}$-Lipschitz graph can be equivalently restated in terms of its $\mathtt c_{\tilde{C}}$-super/subdifferentials as follows.
\begin{proposition}
\label{P_compl_para_sd}
$G\subset\mathbb{R}^k$ is a complete $\mathtt c_{\tilde{C}}$-Lipschitz graph if and only if
\begin{subequations}
\label{E_gprop_para_sd}
\begin{equation}
\label{E_Gproper}
\partial^\pm G \subset \mathrm{graph}\,\bigr(\mathbb{I}\pm \partial \tilde{C}\bigl),
\end{equation}
\begin{equation}
\label{E_parallelogram_subdiff}
\mathtt{O}(w,w') \subset \partial^+G(w) \cap \partial^-G(w'), \quad \forall(w,w') \in G\times G.
\end{equation}
\end{subequations}
\end{proposition}
\begin{proof}
By Definition \ref{D_tilde_C_diff}, property \eqref{E_Gproper} is a rephrasing of \eqref{E_G1}, and \eqref{E_parallelogram_subdiff} is a rephrasing of \eqref{E_G2}.
\end{proof}
Recalling \eqref{E_partvarphi_form} and \eqref{E_sdiff_varphi}, we notice that if $G$ is a $\mathtt c_{\tilde{C}}$-Lipschitz graph and $\varphi_G$ is the function satisfying \eqref{E_phi_graph_G}, then
\[
\partial^\pm G = \partial^\pm\mathrm{graph}\,\varphi_G=\mathbb{I}\times\varphi_G \bigl( \partial^\pm\varphi_G \bigr).
\]
Now we state a simple geometric characterization of the set $\mathtt{O}(w,w')$ which will be fundamental in our study of the $\mathtt c_{\tilde{C}}$-super/subdifferentials of complete $\mathtt c_{\tilde{C}}$-Lipschitz graphs. First we give the following definition.
\begin{definition}
\label{D_cww'}
For $w,\,w'\in\mathbb{R}^k$, we let $C(w,w')$ be the extremal cone of $\tilde{C}$ satisfying
\begin{equation}
\label{E_cww'}
w' - w \in \inter_{\mathrm{rel}} C(w,w').
\end{equation}
Equivalently, $C(w,w')$ is the minimal cone w.r.t. set inclusion among the extremal cones of $\tilde{C}$ containing $w'-w$.
\end{definition}
Notice that, by \eqref{E_G1}, if $w,w'\in G$ and $\mathtt c_{\tilde C}(w,w') < \infty$ then $C(w,w')\subset\partial\tilde{C}$, i.e. $C(w,w')$ is a proper extremal cone of $\tilde{C}$.
\begin{proposition}\label{P_parall}
$\mathtt{O}(w,w')$ is the convex set given by
\[
\mathtt{O}(w,w')=w+C(w,w')\cap w'-C(w,w')
\]
and there exists $\delta>0$ such that
\begin{equation}
\label{E_O_ball}
B^k(w,\delta) \cap \big( w + C(w,w') \big) \subset \mathtt{O}(w,w') \quad \text{and} \quad B^k(w',\delta) \cap \big( w'-C(w,w') \big) \subset \mathtt{O}(w,w').
\end{equation}
In particular,
\begin{equation*}
\mathbb{R}^+ (\mathtt{O}(w,w') - w) = \mathbb{R}^+ (w' - \mathtt{O}(w,w')) = C(w,w').
\end{equation*}
\end{proposition}
\begin{proof}
By Definition, $\mathtt{O}(w,w')\supset w+C(w,w')\cap w'-C(w,w')$ and since $w'-w\in\inter_{\mathrm{rel}} C(w,w')$, \eqref{E_O_ball} follows.
Let us assume that $\exists\,z\in \mathtt{O}(w,w')\setminus w+ C(w,w')$ and let $F$ be the smallest face of $\tilde C$ such that $z-w$, $w'-w\in F$. Notice that $w'-w\in\partial F$. W.l.o.g. we set $w=0$ and $z=(z_1,z_2)\in \mathbb{R}^{\ell-1}\times\mathbb{R}$, $w'=(w'_1,w'_2)\in \mathbb{R}^{\ell-1}\times\mathbb{R}$
w.r.t. coordinates s.t. $F$ is the epigraph of a convex norm $|\cdot|_{E^*}:\mathbb{R}^{\ell-1}\rightarrow\mathbb{R}$. Hence, $w'_2=|w_1'|_{E^*}$ and either $z_2>|z_1|_{E^*}$, or $z_2=|z_1|_{E^*}$ and $w'_2-z_2>|w_1'-z_1|_{E^*}$.
In the first case, $w'_2-z_2<|w'_1|_{E^*}-|z_1|_{E^*}\leq|w'_1-z_1|_{E^*}$, which implies $w'-z\in\mathrm{aff} F\setminus F\subset\mathbb{R}^k\setminus \tilde C$, contradicting the fact that $z\in\mathtt{O}(w,w')$.
In the second case, we get $w'_2-z_2>|w'_1-z_1|_{E^*}\geq|w'_1|_{E^*}-|z_1|_{E^*}=w'_2-z_2$, thus leading again to a contradiction.
\end{proof}
Observe that, with the characterization of Proposition \ref{P_parall} and by projecting on $\mathbb{R}^{k-1}$, \eqref{E_parallelogram_subdiff} can be equivalently restated in terms of the super/subdifferentials of $\varphi_G$ saying that, for all $x,x'\in\mathrm{dom}\,\varphi_G$, the set $\partial^+\varphi_G(x) \cap \partial^-\varphi_G(x')$ contains the convex set $\mathtt{O}_{\mathtt p_{\mathbb{R}^{k-1}}}(x,x')$ such that
\[
\mathtt{O}_{\mathtt p_{\mathbb{R}^{k-1}}}(x,x')=x+C_{\mathtt p_{\mathbb{R}^{k-1}}}(x,x')\cap x'-C_{\mathtt p_{\mathbb{R}^{k-1}}}(x,x')
\]
and
\begin{equation}
\label{E_proj_Oww'}
\mathbb{R}^+ \bigl( \mathtt{O}_{\mathtt p_{\mathbb{R}^{k-1}}}(x,x')- x \bigr) = \mathbb{R}^+ \bigl( x' - \mathtt{O}_{\mathtt p_{\mathbb{R}^{k-1}}}(x,x') \bigr) = C_{\mathtt p_{\mathbb{R}^{k-1}}}(x,x'),
\end{equation}
being $C_{\mathtt p_{\mathbb{R}^{k-1}}}(x,x')$ the minimal extremal cone of $|\cdot|_{\tilde D^*}$ containing $x'-x$ (see Section \ref{Ss_intro_affine_subspaces_cones} for the definition of extremal cone).
\subsection{Convex cone-Lipschitz foliations}
\label{Ss_cone_lipschitz}
Let $\tilde{\mathbf C}:\mathfrak{A}\times\mathbb{R}^k\rightarrow\mathcal C(k;\mathbb{R}^k)$, $\mathfrak{A}\subset\mathbb{R}^{d-k}$, be the convex cone direction map of a $k$-directed fibration satisfying \eqref{E_cones_fibration} and let $\mathtt c_{\tilde{\mathbf C}}$ be the cost function defined in \eqref{E_cost_fibr}. Recall that, by \eqref{E_cost_fibr}
\begin{equation}
\label{E_caa'}
\mathtt c_{\tilde{\mathbf C}}(\mathfrak a,w,\mathfrak a',w') < +\infty \quad \Longrightarrow \quad \mathfrak a=\mathfrak a',
\end{equation}
and
\begin{equation}
\mathtt c_{\tilde{\mathbf C}}(\mathfrak a,w,\mathfrak a,w)=\mathtt c_{\tilde{\mathbf C}(\mathfrak a)}(w,w').
\end{equation}
Moreover, set
\begin{equation}
\label{E_D_star_q}
\tilde{\mathbf C}(\mathfrak a)=\mathrm{epi}\,|\cdot|_{D(\mathfrak a)^*},
\end{equation}
where $D(\mathfrak a) \subset \mathbb{R}^{k-1}$ for some suitable orthonormal coordinates independent of $\mathfrak a$.
\begin{definition}
\label{D_Lip_fol}
A $\mathtt c_{\tilde{\mathbf C}}$-Lipschitz foliation is a $\sigma$-compact partition in $\mathfrak{A} \times \mathbb{R}^k$ with quotient map $\theta : \mathrm{dom}\,\theta \subset \mathfrak{A} \times \mathbb{R}^k \rightarrow \mathfrak T$ such that
\begin{subequations}
\label{E_compl_fol0}
\begin{equation}
\label{E_compl_fol}
(\mathfrak a,w),(\mathfrak a',w') \in \{\theta=\mathfrak t\} \quad \Longrightarrow \quad \big\{ \mathtt c_{\tilde{\mathbf C}} (\mathfrak a,w,\cdot,\cdot) < +\infty \big\} \cap \big\{ \mathtt c_{\tilde{\mathbf C}}(\cdot,\cdot, \mathfrak a',w') < +\infty \big\} \subset \{ \theta = \mathfrak t \},
\end{equation}
\begin{equation}
\label{E_compl_fol1}
\theta(\mathfrak a,w) = \theta(\mathfrak a',w') \quad \Longrightarrow \quad \mathfrak a=\mathfrak a'.
\end{equation}
\end{subequations}
\end{definition}
By \eqref{E_caa'}, \eqref{E_compl_fol} and recalling \eqref{E_G2}, we set
\begin{equation}
\label{E_O_q_ww'}
\mathtt{O}(\mathfrak a)(w,w') := \bigl\{ \mathtt c_{\tilde{\mathbf C}(\mathfrak a)}(w,\cdot) < +\infty \bigr\} \cap \bigl\{ \mathtt c_{\tilde{\mathbf C}(\mathfrak a)}(\cdot,w') < +\infty \bigr\} = (w+\tilde{\mathbf C}(\mathfrak a)) \cap (w'-\tilde{\mathbf C}(\mathfrak a)).
\end{equation}
We note moreover that from \eqref{E_compl_fol1} one has $\mathtt p_\mathfrak{A} \{\theta=\mathfrak t\}= \{\mathfrak a\}$ for some $\mathfrak a\in\mathfrak{A}$, thus in general $\mathfrak T=\mathfrak{A}\times\mathfrak S$ for some Polish space $\mathfrak S$ and $\mathtt p_\mathfrak{A}\circ\theta\circ\mathtt p_\mathfrak{A}^{-1}=\mathbb{I}_\mathfrak{A}$.
Hence, for simplicity of notation, from now onwards we will write --when not leading to confusion-- $\mathfrak a=\mathtt p_\mathfrak{A}\{\theta=\mathfrak t\}$.
The following definition is given to simplify the notation in Proposition \ref{P_fol_char} (see Remark \ref{R_nondeg}).
\begin{definition}
\label{D_non_dege}
We call \emph{non-degeneracy set} of a $\mathtt c_{\tilde{\mathbf C}}$-Lipschitz foliation the set
\begin{equation}
\label{E_nondeg}
\begin{split}
\biggl\{ \mathfrak t \in \mathfrak T :&~ \exists \, w, w' \in \mathtt p_{\mathbb{R}^k} \{\theta=\mathfrak t\} \text{ s.t. } \crcr
&~ \mathrm{int} \big\{ \mathtt c_{\tilde{\mathbf C}(\mathfrak a)}(\cdot,w') < +\infty \big\} \cup \mathrm{int} \big\{\mathtt c_{\tilde{\mathbf C}(\mathfrak a)}(w,\cdot) < +\infty \big\} \subset \bigl( \mathbb{R}^k \setminus \mathtt p_{\mathbb{R}^k}\{\theta=\mathfrak t\} \bigr) \biggr\}.
\end{split}
\end{equation}
We say that the partition $\{\theta^{-1}(\mathfrak t)\}_{\mathfrak t\in\mathfrak T}$ is \emph{non-degenerate} if the set $\{\mathfrak t:\,\{\theta=\mathfrak t\}\neq\emptyset\}$ coincides with its non-degeneracy set.
\end{definition}
In other words, the set $\{\theta = \mathfrak t\}$ is non-degenerate if there are two points $w$, $w'$ in $\mathtt p_{\mathbb{R}^k} \{\theta = \mathfrak t\}$ such that
\[
( w' - \tilde{\mathbf C}(\mathfrak a) ) \cap \{\theta = \mathfrak t\} \subset w' - \partial \tilde{\mathbf C}(\mathfrak a), \qquad ( w + \tilde{\mathbf C}(\mathfrak a) ) \cap \{\theta = \mathfrak t\} \subset w + \partial \tilde{\mathbf C}(\mathfrak a).
\]
\begin{proposition}
\label{P_fol_char}
Let $\{\theta^{-1}(\mathfrak t)\}_{\mathfrak t\in\mathfrak T}$ be a non-degenerate $\mathtt c_{\tilde{\mathbf C}}$-Lipschitz foliation. Then, there exist two Borel functions
\[
\mathtt h^-,\mathtt h^+ : \Bigl\{ (\mathfrak t,x)\in \mathfrak T \times \mathbb{R}^{k-1} : x \in \mathtt p_{\mathbb{R}^{k-1}}(\{\theta=\mathfrak t\}) \Bigr\} \rightarrow \mathbb{R}
\]
such that
\begin{enumerate}
\item \label{Prop_1_fol_char} $x \mapsto \mathtt h^+(\mathfrak t,x),\mathtt h^-(\mathfrak t,x)$ are $|\cdot|_{D(\mathfrak a)^*}$-Lipschitz functions for all $\mathfrak t$, where $\{\mathfrak a\} = \mathtt p_{\mathfrak{A}} \{\theta=\mathfrak t\}$ and $D(\mathfrak a)^*$ is given by \eqref{E_D_star_q};
\item \label{Prop_2_fol_char} $\mathrm{clos}\, \{\theta=\mathfrak t\}(\mathfrak a) \subset \Big\{ (x,y) \in \mathbb{R}^{k-1}\times \mathbb{R} : \mathtt h^-(\mathfrak t,x) \leq y \leq \mathtt h^+(\mathfrak t,x) \Big\}$;
\item \label{Prop_3_fol_char} $\inter_{\mathrm{rel}} \{\theta=\mathfrak t\}(\mathfrak a) = \Big\{ (x,y)\in \times \mathbb{R}^{k-1} \times \mathbb{R} : \mathtt h^-(\mathfrak t,x) < y < \mathtt h^+(\mathfrak t,x) \Big\}$.
\end{enumerate}
In particular,
\begin{equation}
\label{E_Lipschitzgraph}
\begin{split}
\inter_{\mathrm{rel}} \{\theta=\mathfrak t\}(\mathfrak a) = \emptyset \quad & \Longleftrightarrow \quad \mathtt h^-(\mathfrak t,x) = \mathtt h^+(\mathfrak t,x) \ \text{for} \ x \in \mathtt p_{\mathbb{R}^{k-1}} \{\theta=\mathfrak t\}(\mathfrak a) \\
& \Longleftrightarrow \quad \{\theta=\mathfrak t\}(\mathfrak a) \text{ is a complete $\mathtt c_{\tilde{\mathbf C}(\mathfrak a)}$-Lipschitz graph} \\
& \Longleftrightarrow \quad \{\theta=\mathfrak t\}(\mathfrak a) = \mathrm{graph}\,\mathtt h^\pm(\mathfrak t) \llcorner_{\mathtt p_{\mathbb{R}^{k-1}} \{\theta=\mathfrak t\}(\mathfrak a) \times \mathbb{R}}.
\end{split}
\end{equation}
\end{proposition}
\begin{proof}
By \eqref{E_compl_fol}, $\{\theta=\mathfrak t\}(\mathfrak a)$ satisfies \eqref{E_G2}. In particular, for all $x \in \mathtt p_{\mathbb{R}^{k-1}} (\{\theta=\mathfrak t\}(\mathfrak a))$ the set
\[
\big\{ y \in \mathbb{R}: (x,y) \in \{\theta=\mathfrak t\}(\mathfrak a) \big\}
\]
is a segment. Thus define for $x \in \mathtt p_{\mathbb{R}^{k-1}} (\{\theta=\mathfrak t\}(\mathfrak a))$
\begin{subequations}
\label{E_h_def}
\begin{equation}
\label{E_h-def}
\mathtt h^-(\mathfrak t,x) := \inf \Big\{ y : \mathtt c_{\tilde{\mathbf C}(\mathfrak a)}(x',y',x,y) < +\infty \text{ for some } (x',y') \in \{\theta=t\}(\mathfrak a)\Big\},
\end{equation}
\begin{equation}
\label{E_h+def}
\mathtt h^+(\mathfrak t,x) := \sup \Big\{ y : \mathtt c_{\tilde{\mathbf C}(\mathfrak a)}(x,y,x',y') < +\infty \text{ for some } (x',y') \in \{\theta=t\}(\mathfrak a) \Big\}.
\end{equation}
\end{subequations}
Since $\{\theta^{-1}(\mathfrak t)\}_{\mathfrak t \in \mathfrak T}$ is non degenerate, then for all $x \in \mathtt p_{\mathbb{R}^{k-1}} (\{\theta=\mathfrak t\}(\mathfrak a))$ it follows that
\[
(x,\mathtt h^-(\mathfrak t,x)) \cap \mathrm{int} \big\{ \mathtt c_{\tilde{\mathbf C}(\mathfrak a)}(\cdot,w') < +\infty \big\} =(x,\mathtt h^+(\mathfrak t,x)) \cap \mathrm{int} \big\{ \mathtt c_{\tilde{\mathbf C}(\mathfrak a)}(w,\cdot) < +\infty \big\} = \emptyset,
\]
where $w$, $w'$ are the points of non-degeneracy \eqref{E_nondeg}, so that $\mathtt h^+$, $\mathtt h^-$ are real valued functions.
Using again property \eqref{E_G2}, one has also
\[
\mathtt h^-(\mathfrak t,x) = \inf \Big\{ y' + |x - x'|_{D(\mathfrak a)^*}: (x',y') \in \{\theta=\mathfrak t\}(\mathfrak a) \} \Big\},
\]
\[
\mathtt h^+(\mathfrak t,x) = \sup \Big\{ y' - |x' - x|_{D(\mathfrak a)^*}: (x',y') \in \{\theta=\mathfrak t\}(\mathfrak a)\} \Big\},
\]
which show that $\mathtt h^+$, $\mathtt h^-$ are $|\cdot|_{D(\mathfrak a)^*}$-Lipschitz, proving Point \eqref{Prop_1_fol_char}. Points \eqref{Prop_2_fol_char} and \eqref{Prop_3_fol_char} of the statement are an immediate corollary of the definitions \eqref{E_h_def} and property \eqref{E_G2}.
Finally \eqref{E_Lipschitzgraph} is a straightforward consequence of the first part of the statement.
\end{proof}
\begin{remark}
\label{R_nondeg}
From the proof of Proposition \ref{P_fol_char} it is clear that out of the non-degeneracy set there are three possibilities: either the function defined in \eqref{E_h-def} is identically $-\infty$ or the function in \eqref{E_h+def} is identically $+\infty$ or both things happen. Hence, for all $\mathfrak a\in\mathfrak{A}$ there exist at most countably many $\{\mathfrak t_{\mathfrak a_n}\}_{n\in\mathbb{N}}\subset\mathfrak T$ s.t.
\[
\inter_{\mathrm{rel}}\{\theta=\mathfrak t_{\mathfrak a_n}\}(\mathfrak a)\neq\emptyset.
\]
\end{remark}
In the following proposition we give the two examples of $\mathtt c_{\tilde{\mathbf C}}$-Lipschitz foliations we will deal with in the rest of the paper.
\begin{proposition}
\label{P_ex_fol}
(1) Let $G=\mathrm{graph}\,\varphi\subset\mathbb{R}^k$ be a complete $\mathtt c_{\tilde{C}}$-Lipschitz graph. Then,
the trivial equivalence relation on $\{\mathfrak a_0\}\times G$ given by the constant quotient map $\theta_\varphi(\{\mathfrak a_0\}\times G)=\mathfrak a_0$
determines a $\mathtt c_{\tilde{\mathbf C}}$-Lipschitz foliation in $\{\mathfrak a_0\}\times G\subset\{\mathfrak a_0\}\times\mathbb{R}^k$, being $\tilde{\mathbf C}$ the constant cone direction map $\tilde{\mathbf C}:\{\mathfrak a_0\}\times G\ni (\mathfrak a_0,w)\mapsto\tilde C\in\mathcal C(k, \mathbb{R}^k)$, which satisfies \eqref{E_Lipschitzgraph}.
(2) Let $\{\theta^{-1}(\mathfrak t)\}_{\mathfrak t\in\mathfrak T}$ be the equivalence classes of a $\mathtt c_{\tilde{\mathbf C}}$-compatible linear preorder $\preccurlyeq$ on $\mathfrak{A}\times\mathbb{R}^k$ with $\sigma$-compact graph (see Definition \ref{D_compatible}). Then they form a $\mathtt c_{\tilde{\mathbf C}}$-Lipschitz foliation.
\end{proposition}
\begin{proof}
To prove the first part of the proposition, it is sufficient to notice that property \eqref{E_compl_fol} corresponds to \eqref{E_G2} and \eqref{E_Lipschitzgraph} is a consequence of \eqref{E_G1}.
As for the second example, by \eqref{E_tA22} observe that if $w,w'\in\{\theta=\mathfrak t\}(\mathfrak a)$ then
\[
\big\{ \mathtt c_{\tilde{\mathbf C}(\mathfrak a)}(w,\cdot) < +\infty \big\} \cap \big\{ \mathtt c_{\tilde{\mathbf C}(\mathfrak a)}(\cdot,w') < +\infty \big\} \subset \big\{ w'' : w \preccurlyeq w'' \text{ and } w'' \preccurlyeq w' \big\} \subset \{\theta=\mathfrak t\}(\mathfrak a).
\]
\end{proof}
In view of \eqref{E_Lipschitzgraph}, we extend Definition \ref{D_tilde_C_diff} to $\mathtt c_{\tilde{\mathbf C}}$-Lipschitz foliations.
\begin{definition}
\label{D_partial_theta}
We define the \emph{superdifferential} of a $\mathtt c_{\tilde{\mathbf C}}$-Lipschitz foliation $\{\theta^{-1}(\mathfrak t)\}_{\mathfrak t \in \mathfrak T}$ as the set $\partial^+ \theta \subset \mathfrak{A} \times \mathbb{R}^k \times \mathfrak{A} \times \mathbb{R}^k$ defined by
\begin{equation*}
\partial^+\theta(\mathfrak a,\mathfrak a') = \Big\{ (w,w') : \theta(\mathfrak a,w) = \theta(\mathfrak a',w') \quad \text{and} \quad \mathtt c_{\tilde{\mathbf C}}(\mathfrak a,w,\mathfrak a',w') < +\infty \Big\}.
\end{equation*}
Analogously, we define its \emph{subdifferential} as the set $\partial^-\theta \subset \mathfrak{A} \times \mathbb{R}^k \times \mathfrak{A} \times \mathbb{R}^k$ given by
\begin{equation*}
\partial^-\theta(\mathfrak a,\mathfrak a') = \Big\{ (w,w') : \theta(\mathfrak a,w) = \theta(\mathfrak a',w') \quad \text{and} \quad \mathtt c_{\tilde{\mathbf C}}(\mathfrak a',w',\mathfrak a,w) < +\infty \Big\}.
\end{equation*}
\end{definition}
Clearly
\[
\partial^-\theta=(\partial^+\theta)^{-1}
\]
and since by \eqref{E_caa'}
\begin{equation}
\label{E_subaa'}
\partial^\pm \theta(\mathfrak a,\mathfrak a') \not= \emptyset \quad \Longrightarrow \quad \mathfrak a = \mathfrak a',
\end{equation}
then for simplicity we will use the notation
\[
\partial^\pm\theta(\mathfrak a) = \partial^\pm\theta(\mathfrak a,\mathfrak a).
\]
\begin{remark}
Recalling Definition \ref{D_tilde_C_diff}, notice that
\[
\partial^{\pm}\theta(\mathfrak a) = \bigcup_{\nfrac{\mathfrak t \in \mathfrak T}{\mathtt p_{\mathfrak{A}} \{\theta=\mathfrak t\} = \mathfrak a}} \mathtt p_{\mathbb{R}^k \times \mathbb{R}^k} \partial^\pm (\{\theta=\mathfrak t\}(\mathfrak a)),
\]
and, in particular, as for $\mathtt c_{\tilde C}$-Lipschitz graphs, we have the \emph{transitivity property}
\begin{equation}
\label{E_transitivity_subdifferential_a}
w' \in \partial^\pm \theta(\mathfrak a,w) \quad \Rightarrow \quad \partial^\pm \theta(\mathfrak a,w') \subset \partial^\pm \theta(\mathfrak a,w).
\end{equation}
\end{remark}
\subsection{Regular transport sets and residual set}
\label{Ss_regu_resi_set}
In this section we consider only the elements of a $\mathtt c_{\tilde{\mathbf C}}$-foliation whose $\{\mathfrak a\}$-sections are partitions into complete $\mathtt c_{\tilde{\mathbf C}(\mathfrak a)}$-Lipschitz graphs, namely level sets of $\theta$ which satisfy \eqref{E_Lipschitzgraph}.
We go through a careful analysis of the geometric properties of the super/subdifferentials of $\mathtt c_{\tilde{\mathbf C}}$-Lipschitz foliations, which finally leads to partition them into \emph{forward/backward regular sets} and a \emph{residual set}.
By \eqref{E_subaa'}, in the following definitions the variable $\{\mathfrak a\}$ simply plays the role of a parameter. Then, to understand the geometric structure of the problem one can also think from now onwards that $\mathfrak{A}=\{\mathfrak a_0\}$. The dependence on $\mathfrak a$ is kept in order to show that all the sets and functions constructed below depend measurably (resp. Borel or $\sigma$-continuously) on the parameter $\mathfrak a$.
We will define the following sets through their $\{\mathfrak a\}$-sections.
\begin{description}
\item[Forward/backward transport set] the \emph{forward/backward transport sets} are respectively defined by
\begin{subequations}
\label{E_forward_backward_set}
\begin{equation}
\label{E_forward_transport_set}
\mathcal T^+\theta(\mathfrak a) := \big\{ w : \partial^+ \theta(\mathfrak a,w)\not= \{w\} \big\} = \mathtt p_1 \big( \partial^+ \theta(\mathfrak a)\setminus \mathbb{I} \big),
\end{equation}
\begin{equation}
\label{E_backward_transport_set}
\mathcal T^-\theta(\mathfrak a) := \big\{ w : \partial^-\theta(\mathfrak a,w)\not= \{w\} \big\} = \mathtt p_1 \big( \partial^-\theta(\mathfrak a) \setminus \mathbb{I} \big).
\end{equation}
\end{subequations}
\item[Set of fixed points] the \emph{set of fixed points} is given by
\begin{equation}
\label{E_set_of_fixed_points}
\mathcal F\theta(\mathfrak a):= \mathbb{R}^k \setminus \big( \mathcal T^-\theta(\mathfrak a) \cup \mathcal T^+\theta(\mathfrak a)
\big).
\end{equation}
\item[Forward/backward direction multifunction] The \emph{forward/backward direction multifunction} are respectively given by
\begin{subequations}
\label{E_forw_back_dir_theta}
\begin{equation}
\label{E_forward_directions_bar_theta}
\mathcal D^+ \theta(\mathfrak a):= \bigg\{ \bigg( w, \frac{w'-w}{|w'-w|} \bigg) : w \in \mathcal T^+\theta(\mathfrak a), w' \in \partial^+\theta(\mathfrak a,w) \setminus \{w\} \bigg\},
\end{equation}
\begin{equation}
\label{E_backward_directions_bar_theta}
\mathcal D^- \theta(\mathfrak a):= \bigg\{ \bigg( w, \frac{w-w'}{|w-w'|} \bigg) : w \in \mathcal T^-\theta(\mathfrak a), w' \in \partial^- \theta(\mathfrak a,w) \setminus \{w\} \bigg\}.
\end{equation}
\end{subequations}
\end{description}
The following proposition collects the fundamental properties of the super/subdifferentials of $\mathtt c_{\tilde{\mathbf C}}$-Lipschitz foliations. The most striking feature is that, due to the completeness property \eqref{E_G2} of its $\mathtt c_{\tilde{\mathbf C}(\mathfrak a)}$-Lipschitz graphs, the forward/backward direction multifunctions at a point of $\{\theta=\mathfrak t\}(\mathfrak a)$ contain all the information about the super/subdifferential at that point and also in a ``neighborhood'' of it (see Remark \ref{R_directionfaces}). Whenever $w,w'\in\{\theta=\mathfrak t\}(\mathfrak a)$, we will define $C(\mathfrak a)(w,w')$ and $\mathtt{O}(\mathfrak a)(w,w')$ as in Definitions \ref{D_cww'} and \eqref{E_G2} for the convex cone $\tilde C=\tilde{\mathbf C}(\mathfrak a)$.
\begin{proposition}
\label{Pdirectionfaces1}
Let $F \subset \tilde{\mathbf C}(\mathfrak a)$ be an extremal cone, $w\in\mathcal T^+\theta(\mathfrak a)$. The following conditions are equivalent:
\begin{enumerate}
\item \label{item1Pfaces} $F\cap \mathbb S^{k-1}\subset \mathcal D^+\theta(\mathfrak a,w)$;
\item \label{item2Pfaces} there exists $w' \in \partial^+ \theta(\mathfrak a,w)$ such that $F=C(\mathfrak a)(w,w')$;
\item \label{item4Pfaces} $\mathtt{O}(\mathfrak a)(w,w') \subset \partial^+ \theta(\mathfrak a,w)$ for some $w'$ such that $F=C(\mathfrak a)(w,w')$;
\item \label{item3Pfaces} there exists $\delta = \delta(w,F)>0$ such that
\begin{equation*}
B^k(w,2\delta)\cap w+F\subset\partial^+\theta(\mathfrak a,w).
\end{equation*}
\end{enumerate}
In particular, if $F$ satisfies one of the above conditions, then for all $w'$ as in (\ref{item2Pfaces}-\ref{item4Pfaces})
\begin{equation}
\label{E_ballF}
F \cap \mathbb S^{k-1} \subset \mathcal D^+\theta(\mathfrak a,\bar w) \qquad \forall\, \bar w \in \inter_{\mathrm{rel}} \mathtt{O}(\mathfrak a)(w,w').
\end{equation}
Finally, if $F$ is maximal w.r.t. set inclusion among the extremal cones of $\tilde{\mathbf C}(\mathfrak a)$ satisfying (\ref{item1Pfaces}-\ref{item3Pfaces}), then
\begin{equation}
\label{E_ballFmax}
F \cap \mathbb S^{k-1} = \mathcal D^+\theta(\mathfrak a,\bar w) = \mathrm{conv}_{\mathbb S^{k-1}} \mathcal D^+\theta(\mathfrak a,\bar w) \qquad \forall\, \bar w \in \inter_{\mathrm{rel}} \mathtt{O}(\mathfrak a)(w,w'),
\end{equation}
where $w'$ is chosen as in (\ref{item2Pfaces}-\ref{item4Pfaces}).
\end{proposition}
\begin{proof}
{\it\eqref{item1Pfaces}$\,\Rightarrow\,$\eqref{item2Pfaces}.} It is sufficient to take $w'\in\partial^+\theta(\mathfrak a,w)\setminus\{w\}$ s.t.
\[
\frac{w'-w}{|w'-w|}\in\inter_{\mathrm{rel}} F\cap \mathbb S^{k-1}
\]
and recall Definition \ref{D_cww'}.
{\it\eqref{item2Pfaces}$\,\Rightarrow\,$\eqref{item4Pfaces}.} If $w'$ satisfies \eqref{item2Pfaces}, then \eqref{item4Pfaces} follows from the completeness assumption \eqref{E_compl_fol} in the definition of $\mathtt c_{\tilde{\mathbf C}}$-Lipschitz foliation.
{\it\eqref{item4Pfaces}$\,\Rightarrow\,$\eqref{item3Pfaces}.} It follows immediately from \eqref{E_O_ball}.
{\it\eqref{item3Pfaces}$\,\Rightarrow\,$\eqref{item1Pfaces}.} It is a direct consequence of the definition of $\mathcal D^+\theta(\mathfrak a)$.
{\it Proof of \eqref{E_ballF}.}
Let $\bar w\in\inter_{\mathrm{rel}}\mathtt{O}(\mathfrak a)(w,w')$. Then, by the geometric properties of $\mathtt{O}(\mathfrak a)(w,w')$ given in Proposition \ref{P_parall}, it is fairly easy to see that $w'\in\partial^+\theta(\mathfrak a,\bar w)$, and $C(\mathfrak a)(\bar w,w')\cap \mathbb{S}^{k-1}\subset\mathcal D^+\theta(\mathfrak a,\bar w)$. Since $\bar w\in\inter_{\mathrm{rel}}\mathtt{O}(\mathfrak a)(w,w')$, $C(\mathfrak a)(\bar w,w')=F$.
{\it Proof of \eqref{E_ballFmax}.}
Let now $F$ be maximal and $\bar w\in \inter_{\mathrm{rel}}\mathtt{O}(\mathfrak a)(w,w')$. By \eqref{E_ballF} we already know that $F\cap \mathbb{S}^{k-1}\subset\mathcal D^+\theta(\mathfrak a,\bar w)$. Let us assume that there exists $\hat w\in\partial^+\theta(\mathfrak a,\bar w)$ such that $\hat w - \bar w \in \mathbb{R}^k \setminus F$: being $F$ an extremal cone, then one has also $\hat w - \bar w \in \mathbb{R}^k \setminus \mathrm{aff}\,F$. By the transitivity property \eqref{E_transitivity_subdifferential_a}, $\hat w \in \partial^+\theta(\mathfrak a,w) \setminus (w+F)$ and, by simple geometrical considerations similar to those made in the proof of Proposition \ref{P_parall},
\[
F \subsetneq \mathbb{R}^+(\mathtt{O}(\mathfrak a)(w,\hat w)-w)
\]
with strict inclusion. Hence, by the completeness assumption \eqref{E_compl_fol}, this contradicts the maximality of $F$.
\end{proof}
A completely similar proposition can be proved for $\partial^- \theta$: we state it without proof.
\begin{proposition}
\label{Pdirectionfaces1-}
Let $F \subset \tilde{\mathbf C}(\mathfrak a)$ be an extremal cone, $w\in\mathcal T^-\theta(\mathfrak a)$. The following conditions are equivalent:
\begin{enumerate}
\item \label{item1Pfaces-} $F \cap \mathbb S^{k-1} \subset \mathcal D^-\theta(\mathfrak a,w)$;
\item \label{item2Pfaces-} there exists $w'' \in \partial^- \theta(\mathfrak a,w)$ such that $F=C(\mathfrak a)(w'',w)$;
\item \label{item4Pfaces-} $\mathtt{O}(\mathfrak a)(w'',w) \subset \partial^- \theta(\mathfrak a,w)$ for some $w''$ such that $F=C(\mathfrak a)(w'',w)$;
\item \label{item3Pfaces-} there exists $\delta = \delta(w,F) > 0$ such that
\begin{equation*}
B^k(w,2\delta) \cap w - F \subset \partial^- \theta(\mathfrak a,w).
\end{equation*}
\end{enumerate}
In particular, if $F$ satisfies one of the above conditions, then for all $w''$ as in (\ref{item2Pfaces-}-\ref{item4Pfaces-})
\begin{equation*}
F \cap \mathbb S^{k-1} \subset \mathcal D^- \theta(\mathfrak a,\bar w) \qquad \forall\,\bar w\in \inter_{\mathrm{rel}}\mathtt{O}(\mathfrak a)(w'',w).
\end{equation*}
Finally, if $F$ be maximal w.r.t. set inclusion among the extremal cones of $\tilde{\mathbf C}(\mathfrak a)$ satisfying (\ref{item1Pfaces-}-\ref{item3Pfaces-}), then
\begin{equation}
\label{E_ballFmax-}
F \cap \mathbb S^{k-1} = \mathcal D^- \theta(\mathfrak a,\bar w) = \mathrm{conv}_{\mathbb S^{k-1}} \mathcal D^- \theta(\mathfrak a,\bar w) \qquad \forall\, \bar w\in\inter_{\mathrm{rel}}\mathtt{O}(\mathfrak a)(w'',w),\end{equation}
where $w''$ is chosen as in (\ref{item2Pfaces-}-\ref{item4Pfaces-}).
\end{proposition}
\begin{remark}
\label{R_directionfaces}
The radii of the balls $\delta=\delta(F,w)$ satisfying Point \eqref{item3Pfaces} in Propositions \ref{Pdirectionfaces1}-\ref{Pdirectionfaces1-} for a fixed $w \in \mathcal T^\pm \theta(\mathfrak a)$ might actually change as $F$ varies in the set of extremal cones satisfying Point \eqref{item1Pfaces} and even tend to zero for some sequence of distinct $\{F_n\}_{n\in\mathbb{N}}$.
\end{remark}
Finally, we define the \emph{($\ell$-dimensional) forward/backward transport sets} and the \emph{residual sets}: they are defined in terms
of properties of the forward/backward direction multifunctions, i.e. of ``local'' (see Remark \ref{R_directionfaces}) properties of their super/subdifferentials.
\begin{description}
\item[$\ell$-dimensional forward/backward regular transport set] for $\ell = 1,\dots,k-1$, the \emph{$\ell$-dimensional forward/backward regular transport sets} are defined respectively as
\begin{subequations}
\label{E_ell_prime_forward_backward_regular_transport_set}
\begin{equation}
\label{E_forward_regular_transport_set}
\begin{split}
\mathcal R^{+,\ell} \theta(\mathfrak a) := \bigg\{ w \in \mathcal T^+ \theta(\mathfrak a) : (i)&~ \mathcal D^+ \theta(\mathfrak a,w) = \mathrm{conv}_{\mathbb S^{k-1}} \mathcal D^+ \theta(\mathfrak a,w) \crcr
(ii)&~ \dim \big( \mathrm{conv}_{\mathbb S^{k-1}} \mathcal D^+ \theta(\mathfrak a,w) \big) = \ell-1 \crcr
(iii)&~ \exists\, w'' \in \mathcal T^+ \theta(\mathfrak a) \cap \Big( w - \mathrm{int}_{\mathrm{rel}} \big( \mathbb{R}^+\mathrm{conv}_{\mathbb S^{k-1}} \mathcal D^+\theta(\mathfrak a,w) \big) \Big) \crcr
&~ \text{such that } \theta(\mathfrak a,w'') = \theta(\mathfrak a,w) \text{ and } (i),(ii) \ \text{hold for} \ w'' \bigg\}.
\end{split}
\end{equation}
\begin{equation}
\label{E_backward_regular_transport_set}
\begin{split}
\mathcal R^{-,\ell} \theta(\mathfrak a) := \bigg\{ w \in \mathcal T^- \theta(\mathfrak a) : (i)&~ \mathcal D^- \theta(\mathfrak a,w) = \mathrm{conv}_{\mathbb S^{k-1}} \mathcal D^- \theta(\mathfrak a,w) \crcr
(ii)&~ \dim \big( \mathrm{conv}_{\mathbb S^{k-1}} \mathcal D^- \theta(\mathfrak a,w) \big) = \ell-1 \crcr
(iii)&~ \exists\, w'' \in \mathcal T^- \theta(\mathfrak a) \cap \Big( w + \mathrm{int}_{\mathrm{rel}} \big( \mathbb{R}^+\mathrm{conv}_{\mathbb S^{k-1}} \mathcal D^-\theta(\mathfrak a,w) \big) \Big) \crcr
&~ \text{such that } \theta(\mathfrak a,w'') = \theta(\mathfrak a,w) \text{ and } (i),(ii) \ \text{hold for} \ w'' \bigg\}.
\end{split}
\end{equation}
\end{subequations}
\item[Forward/backward regular transport sets] the \emph{forward/backward regular transport sets} are defined respectively by
\begin{equation}
\label{E_forward_backward_regular_transport_set}
\mathcal R^+ \theta(\mathfrak a):= \bigcup_{\ell = 1}^{k-1} \mathcal R^{+,\ell} \theta(\mathfrak a), \qquad \mathcal R^- \theta(\mathfrak a):= \bigcup_{\ell = 1}^{k-1} \mathcal R^{-,\ell} \theta(\mathfrak a).
\end{equation}
\item[Regular transport set] the \emph{regular transport set} is defined by
\begin{equation}
\label{E_regular_transport_set}
\mathcal R \theta(\mathfrak a) := \mathcal R^- \theta(\mathfrak a) \cap \mathcal R^+\theta(\mathfrak a).
\end{equation}
\item[Residual set] the \emph{residual set} is defined by
\begin{equation}
\label{E_residual_set_N}
\mathcal N \theta(\mathfrak a):= \bigl(\mathcal T^+\theta(\mathfrak a)\cup\mathcal T^-\theta(\mathfrak a)\bigr) \setminus \mathcal R \theta(\mathfrak a).
\end{equation}
\end{description}
Property \eqref{E_forward_regular_transport_set} $(iii)$ (\eqref{E_backward_regular_transport_set} $(iii)$) will be crucial in Theorem \ref{T_partition_E+-} in order to prove that the sets of points in $\mathcal R^{+,\ell} \theta(\mathfrak a)$ ($\mathcal R^{-,\ell} \theta(\mathfrak a)$) which belong to the same level set of $\theta$ and whose superdifferentials (subdifferentials) have the same affine span are $\ell$-dimensional locally affine sets (see Definition \ref{D_locaff}).
\begin{figure}
\centering{\resizebox{14cm}{6cm}{\input{sudakovfoliationbis.pdf_t}}}
\caption{A possible decomposition of a level set of $\theta$ (or equivalently of a complete $\mathtt c_{\tilde{\mathbf{C}}(\mathfrak a)}$-Lipschitz graph), with the various sets introduced in Section \ref{Ss_regu_resi_set}. More precisely, given the $3$-dimensional cone of directions $\tilde{\mathbf C}(\mathfrak a)$, the yellow region represents the set of regular points $\mathcal R^{+,2}(\mathfrak a)\cap\mathcal R^{-,2}(\mathfrak a)$ with $2$-dimensional forward/backward cones of directions $\mathcal D^{+}\theta(\mathfrak a)=\mathcal D^-\theta(\mathfrak a)$, the black line the set of regular points $\mathcal R^{+,1}(\mathfrak a)\cap\mathcal R^{-,1}(\mathfrak a)$ with $1$-dimensional forward/backward cones of directions, the blue line the set $\mathcal R^{-,1}\theta(\mathfrak a)$ of points with $1$-dimensional backward cone $\mathcal D^-\theta(\mathfrak a)$ and $2$-dimensional forward cone $\mathcal D^{+}\theta(\mathfrak a)$ and the purple line the set $\mathcal R^{+,1}\theta(\mathfrak a)$ of points with $1$-dimensional forward cone $\mathcal D^+\theta(\mathfrak a)$ and $2$-dimensional backward cone $\mathcal D^{-}\theta(\mathfrak a)$. The brown region represents a set $\mathtt O(\mathfrak a)(w',w'')$ as in \eqref{E_O_q_ww'}. As we will see in Section \ref{Ss_analysis_residual_set}, the red curve represents the set of final points $\mathcal E\theta(\mathfrak a)\setminus \mathcal T^+\theta(\mathfrak a)$, the green curve the set of initial points $\mathcal I\theta(\mathfrak a)\setminus\mathcal T^-(\mathfrak a)$, and also the blue line is contained in $\mathcal I\theta(\mathfrak a)$ and the purple line in $\mathcal E\theta (\mathfrak a)$.}
\label{Fi_decotheta}
\end{figure}
We now prove that all the above sets, except $\mathcal F \theta$ and $\mathcal N \theta$, are $\sigma$-compact.
\begin{proposition}
\label{P_borel_regu_transport_sets}
The sets $\partial^\pm \theta$, $\mathcal T^\pm\theta$, $\mathcal D^\pm \theta$, $\mathcal R^{\pm,\ell}\theta$, $\mathcal R^\pm\theta$, $\mathcal R\theta$ are $\sigma$-compact.
The sets $\mathcal F \theta$, $\mathcal N \theta$ are Borel.
\end{proposition}
\begin{proof}
For every set in the first statement of the proposition, we will construct $\sigma$-compact subsets of a Polish space whose projection corresponds to that particular set.
{\it Subdifferential:} consider the following sets:
\begin{enumerate}
\item $\{(\mathfrak a,w,\mathfrak a',w') : \mathtt c_{\tilde{\mathbf C}}(\mathfrak a,w,\mathfrak a',w') < \infty\}$: since the function $\mathtt c_{\tilde{\mathbf C}}$ is $\sigma$-continuous, it follows that this set is $\sigma$-compact;
\item $\{(\mathfrak a,w,\mathfrak a',w') : \theta(\mathfrak a,w) = \theta(\mathfrak a',w')\}$: the same reasoning of the previous point applies here, being $\theta$ a $\sigma$-continuous function in Definition \ref{D_Lip_fol} of $\mathtt c_{\tilde{\mathbf C}}$-Lipschitz foliation.
\end{enumerate}
It follows that the set
\begin{align*}
\partial^- \theta =&~ \Big\{ (\mathfrak a,w,\mathfrak a',w') \in \mathfrak{A} \times \mathbb{R}^k \times\mathfrak{A}\times \mathbb{R}^k : {\mathtt c_{\tilde{\mathbf C}}}(\mathfrak a',w',\mathfrak a,w) < +\infty, \theta(\mathfrak a,w) = \theta(\mathfrak a,w') \Big\} \crcr
=&~ \big\{ (\mathfrak a,w,\mathfrak a',w') : \mathtt c_{\tilde{\mathbf C}}(\mathfrak a',w',\mathfrak a,w) < \infty \big\} \cap \big\{ (\mathfrak a,w,\mathfrak a',w') : \theta(\mathfrak a,w) = \theta(\mathfrak a',w') \big\}
\end{align*}
is $\sigma$-compact.
{\it Backward transport set:} the set $\mathcal T^-\theta$ is the projection of the $\sigma$-compact set
\[
\bigcup_{n \in \mathbb{N}} \partial^- \theta \cap \big\{ (\mathfrak a,w,\mathfrak a',w') : |w-w'| \geq 2^{-n} \big\},
\]
and thus it is $\sigma$-compact.
{\it Backward directions:} since
\[
\{w \not= w'\} \ni (w,w') \mapsto \frac{w-w'}{|w-w'|} \in \mathbb S^{k-1}
\]
is continuous, it follows that $\mathcal D^- \theta$ is $\sigma$-compact, being the image of a $\sigma$-compact set by a continuous function.
{\it Backward regular transport sets:} first notice that the map $A \mapsto \mathrm{conv} A$ is continuous w.r.t. the Hausdorff topology and that the sets
\begin{equation}
\label{E_C_m_def}
\mathcal C_m(\ell,\mathbb{R}^k) := \bigg\{ C\in\mathcal C(\ell;\mathbb{R}^k): \mathring C(-1\slash m) \supset \big( B^k(w_m,1\slash(2m)) \cap\, \mathrm{aff}\, C \big), \ \mathrm{dist} \big( w_m, \mathrm{aff}\, C) \leq 1/(4m) \bigg\}
\end{equation}
are closed w.r.t. the Hausdorff topology, for all $\ell=1,\dots,k-1$, $w_m \in \mathbb{Q}^k$, $m\in\mathbb{N}$. Since the function $\mathcal C(\ell,\mathbb{R}^k)\ni C \mapsto \mathrm{dim}\,C$ is constant on these sets, then it is $\sigma$-continuous.
Let us now prove that the set
\[
\Big\{ (w,w',C) \in \mathbb{R}^k \times \mathbb{R}^k \times \mathcal C(\ell,\mathbb{R}^k) : w' \in w - \inter_{\mathrm{rel}} C \Big\}
\]
is $\sigma$-compact. This follows by considering the closed sets $C(-r) \setminus B^k(0,r)$, observing that
\[
C_n \rightarrow C \quad \Longrightarrow \quad C_n(-r) \setminus B^k(0,r) \rightarrow C(-r) \setminus B^k(0,r),
\]
and writing the previous set as the union of countably many $\sigma$-compact sets in the following way:
\[
\bigcup_{n \in \mathbb{N}} \bigg\{ (w,w',C) \in \mathbb{R}^k \times \mathbb{R}^k \times \mathcal C(\ell,\mathbb{R}^k) : w' \in w - \big( C(-2^{-n}) \setminus B^k(0,2^{-n}) \bigr) \bigg\}.
\]
From Proposition \ref{Pdirectionfaces1-}, we have moreover that
\[
\mathcal D^- \theta(\mathfrak a,w) = \mathrm{conv}_{\mathbb S^{k-1}}\, \mathcal D^- \theta(\mathfrak a,w)\quad\Leftrightarrow\quad \mathcal D^- \theta(\mathfrak a,w) \cap\inter_{\mathrm{rel}} \mathrm{conv}_{\mathbb S^{k-1}}\, \mathcal D^- \theta(\mathfrak a,w)\neq\emptyset.
\]
Hence the set
\begin{align*}
\bigg\{ (\mathfrak a,w,\mathfrak a', w',C):\ (i)&~ (\mathfrak a,w),(\mathfrak a',w') \in \mathcal T^-\theta \crcr
(ii)&~ \theta(\mathfrak a,w) = \theta(\mathfrak a',w')\crcr
(iii)&~ C = \mathbb{R}^+\mathrm{conv}_{\mathbb S^{k-1}}\mathcal D^\pm \theta(\mathfrak a,w)\crcr
(iv)&~ w' \in w - \mathrm{int}_{\mathrm{rel}}\, C \crcr
(v)&~ \dim \big( \mathrm{conv}_{\mathbb S^{k-1}}\, \mathcal D^- \theta(\mathfrak a,w) \big) = \dim \big( \mathrm{conv}_{\mathbb S^{k-1}}\, \mathcal D^- \theta(\mathfrak a',w') \big) = \ell-1 \crcr
(vi)&~ \mathcal D^- \theta(\mathfrak a,w) = \mathrm{conv}_{\mathbb S^{k-1}}\, \mathcal D^- \theta(\mathfrak a,w), \mathcal D^- \theta(\mathfrak a,w') = \mathrm{conv}_{\mathbb S^{k-1}}\, \mathcal D^- \theta(\mathfrak a,w') \bigg\}
\end{align*}
is $\sigma$-compact, being the finite intersection of $\sigma$-compact sets, and thus $\mathcal R^{-,\ell}\theta$ is $\sigma$-compact too.
The proof for $\partial^+ \theta$, $\mathcal T^+\theta$, $\mathcal D^+ \theta$ and $\mathcal R^{+,\ell}\theta$ is analogous, and hence the $\sigma$-compactness of $\mathcal R^+\theta$, $\mathcal R^-\theta$ and $\mathcal R\theta$ follows.
Being the difference of two $\sigma$-compact sets a Borel set, the Borel measurability of $\mathcal F \theta$ and $\mathcal N \theta$ is proved.
\end{proof}
\subsection{Super/subdifferential directed partitions of regular sets}
\label{Ss_partition_transport_set}
In this section we construct directed locally affine partitions of the forward regular sets and backward regular sets of a $\mathtt c_{\tilde{\mathbf C}}$-Lipschitz foliation, which will be respectively called \emph{superdifferential directed partitions} and \emph{subdifferential directed partitions}. As we will see, these partitions coincide on the regular set, thus giving a directed locally affine partition which will be called \emph{$\mathtt c_{\tilde{\mathbf C}}$-differential directed partition}.
Define the maps
\begin{subequations}
\label{E_partition_on_T}
\begin{equation}
\label{E_partition_on_T+}
\begin{array}{ccccc}
\mathtt v^+ &:& \mathcal R^+ \theta&\rightarrow& \mathfrak T \times \overset{k-1}{\underset{\ell=1}{\cup}} \mathcal A(\ell,\mathbb{R}^k) \crcr
&& (\mathfrak a,w) &\mapsto& \mathtt v^+(\mathfrak a,w) := \big( \theta(\mathfrak a,w), \mathrm{aff}\, \partial^+ \theta(\mathfrak a,w) \big)
\end{array}
\end{equation}
\begin{equation}
\label{E_partition_on_T-}
\begin{array}{ccccc}
\mathtt v^- &:& \mathcal R^- \theta&\rightarrow& \mathfrak T \times \overset{k-1}{\underset{\ell=1}{\cup}} \mathcal A(\ell,\mathbb{R}^k) \crcr
&& (\mathfrak a,w) &\mapsto& \mathtt v^-(\mathfrak a,w) := \big( \theta(\mathfrak a,w), \mathrm{aff}\, \partial^- \theta(\mathfrak a,w) \big)
\end{array}
\end{equation}
\end{subequations}
In the following, when clear from the context, we identify sets $E_\mathfrak a\subset\{\mathfrak a\}\times\mathbb{R}^k$ with $\mathtt p_{\mathbb{R}^k}E_\mathfrak a$.
\begin{theorem}
\label{T_partition_E+-}
The map $\mathtt v^+$ induces a (complete) directed locally affine partition on $\mathcal R^+\theta$ into sets
\begin{equation*}
\Big\{ Z^{\ell,+}_{\mathfrak a,\mathfrak b}\Big\}_{\overset{\ell=1,\dots,k-1}{\mathfrak a\in\mathfrak{A},\mathfrak b\in\mathfrak B_{\ell,+}(\mathfrak a)}},\quad Z^{\ell,+}_{\mathfrak a,\mathfrak b}\subset \{\mathfrak a\}\times \mathbb{R}^k
\end{equation*}
with direction cones
\[
\Big\{ C^{\ell,+}_{\mathfrak a,\mathfrak b} \Big\}_{\overset{\ell=1,\dots,k-1}{\mathfrak a\in\mathfrak{A},\mathfrak b\in\mathfrak B_{\ell,+}(\mathfrak a)}},\quad C^{\ell,+}_{\mathfrak a,\mathfrak b} \subset \{\mathfrak a\}\times \mathcal C(\ell,\mathbb{R}^k)
\]
such that the following holds:
\begin{align}
\label{item2TpartE+}
&C^{\ell,+}_{\mathfrak a,\mathfrak b}\text{ is the extremal face of $\tilde{\mathbf C}(\mathfrak a)$ s.t. $\mathcal D^+\theta(\mathfrak a)(w)= C^{\ell,+}_{\mathfrak a,\mathfrak b}\cap \mathbb S^{k-1}$, $\forall\,w\in Z^{\ell,+}_{\mathfrak a,\mathfrak b}$. }
\end{align}
Analogously, the map $\mathtt v^-$ induces a (complete) directed locally affine partition of $\mathcal R^-\theta$ into sets
\begin{equation*}
\Big\{ Z^{\ell,-}_{\mathfrak a,\mathfrak b}\Big\}_{\overset{\ell=1,\dots,k-1}{\mathfrak a\in\mathfrak{A},\mathfrak b\in\mathfrak B_{\ell,-}(\mathfrak a)}},\quad Z^{\ell,-}_{\mathfrak a,\mathfrak b}\subset \{\mathfrak a\}\times \mathbb{R}^k
\end{equation*}
with direction cones
\[
\Big\{C^{\ell,-}_{\mathfrak a,\mathfrak b}\Big\}_{\overset{\ell=1,\dots,k-1}{\mathfrak a\in\mathfrak{A},\mathfrak b\in\mathfrak B_{\ell,-}(\mathfrak a)}},\quad C^{\ell,-}_{\mathfrak a,\mathfrak b} \subset \{\mathfrak a\}\times \mathcal C(\ell,\mathbb{R}^k)
\]
such that the following holds:
\begin{align}
\label{item2TpartE-}
&C^{\ell,-}_{\mathfrak a,\mathfrak b}\text{ is the extremal face of $\tilde{\mathbf C}(\mathfrak a)$ s.t. $\mathcal D^-\theta(\mathfrak a)(w)= C^{\ell,-}_{\mathfrak a,\mathfrak b}\cap \mathbb S^{k-1}$, $\forall\,w\in Z^{\ell,-}_{\mathfrak a,\mathfrak b}$.}
\end{align}
\end{theorem}
As a corollary one obtains the following decomposition of $\mathcal R \theta$.
\begin{corollary}
\label{C_v}
The two maps $\mathtt v^+$, $\mathtt v^-$ coincide on $\mathcal R \theta$, i.e.
\begin{equation}
\label{E_v}
\mathtt v:=\mathtt v^+_{|_{\mathcal R\theta}} = \mathtt v^-_{|_{\mathcal R\theta}},
\end{equation}
and the map $\mathtt v : \mathcal R \theta \rightarrow \mathfrak T \times \overset{k-1}{\underset{\ell=1}{\cup}} \mathcal A(\ell,\mathbb{R}^k)$ defined above induces on $\mathcal R \theta$ a (complete) directed locally affine partition $\{Z^\ell_{\mathfrak a,\mathfrak b},\,C^\ell_{\mathfrak a,\mathfrak b}\}_{\nfrac{\ell=1,\dots,k-1}{\mathfrak a \in \mathfrak{A}, \mathfrak b \in \mathfrak B_\ell(\mathfrak a)}}$, where, for all $\mathfrak b \in \mathfrak B_\ell(\mathfrak a)$ and $\ell=1,\dots,k-1$,
\[
Z^\ell_{\mathfrak a,\mathfrak b} = Z^{\ell,+}_{\mathfrak a,\mathfrak b} \cap Z^{\ell,-}_{\mathfrak a,\mathfrak b}, \qquad C^\ell_{\mathfrak a,\mathfrak b} = C^{\ell,+}_{\mathfrak a,\mathfrak b} = C^{\ell,-}_{\mathfrak a,\mathfrak b}.
\]
In particular, both \eqref{item2TpartE+} and \eqref{item2TpartE-} are satisfied.
\end{corollary}
In the following we will use the following definitions.
\begin{definition}
\label{D_vpm}
Given a $\mathtt c_{\tilde{\mathbf C}}$-Lipschitz foliation $\theta$, the directed locally affine partition induced on $\mathcal R^+\theta$ by $\mathtt v^+$ is called the \emph{superdifferential directed partition}, while the directed locally affine partition induced on $\mathcal R^-$ by $\mathtt v^-$ is called the \emph{subdifferential directed partition}.
The partition induced by $\mathtt v$ on the regular points $\mathcal R\theta$ is called \emph{$\mathtt c_{\tilde{\mathbf C}}$-differential directed partition}.
\end{definition}
We prove the part of Theorem \ref{T_partition_E+-} which regards $\mathcal R^+\theta$ and $\mathtt v^+$, being the one about $\mathcal R^-\theta$ and $\mathtt v^-$ completely symmetric. Notice that in the proof is also shown that the map $\mathfrak a \mapsto \mathfrak B_{\ell,\pm}(\mathfrak a)$ is $\sigma$-continuous.
\begin{proof}[Proof of Theorem \ref{T_partition_E+-}]
Being a single-valued map, $\mathtt v^+$ clearly induces a partition of $\mathcal R^+\theta$. Moreover, $\mathtt v^+$ is $\sigma$-continuous by Proposition \ref{P_borel_regu_transport_sets} and the fact that the affine envelope of compact sets is $\sigma$-continuous w.r.t. Hausdorff topology. Since, by \eqref{E_compl_fol1}
\[
\mathtt v^+(\mathfrak a,w)=\mathtt v^+(\mathfrak a',w')\quad\Leftrightarrow\quad\mathfrak a=\mathfrak a',
\]
let
\[
\mathfrak B_{\ell,+}(\mathfrak a) := \mathtt v^+( \{\mathfrak a\} \times \mathcal R^+\theta(\mathfrak a))\cap \mathfrak T\times \mathcal A(\ell,\mathbb{R}^k)
\]
and $\forall\,\mathfrak b\in \mathfrak B_{\ell,+}(\mathfrak a)$ let
\[
Z^{\ell,+}_{\mathfrak a,\mathfrak b} := \Bigl\{ w \in \mathcal R^+\theta(\mathfrak a) : \mathtt v^+(\mathfrak a,w) = \mathfrak b \Bigr\}.
\]
By \eqref{E_forward_regular_transport_set} (i), for all $w \in Z^{\ell,+}_{\mathfrak a,\mathfrak b}$
\[
\mathbb{R}^+\mathcal D^+\theta(\mathfrak a,w) = \mathrm{aff}\,\partial^+\theta(\mathfrak a,w) -w\cap \tilde{\mathbf C}(\mathfrak a) = \mathtt p_2 \mathfrak b-w \cap \tilde{\mathbf C}(\mathfrak a)\in\mathcal C(\ell,\mathbb{R}^k).
\]
Thus, by \eqref{E_forward_regular_transport_set} (iii), $Z^{\ell,+}_{\mathfrak a,\mathfrak b}\subset\mathcal R^{+,\ell}\theta$ and
\[
C^{\ell,+}_{\mathfrak a,\mathfrak b} := \mathtt p_2 \mathfrak b-w \cap \tilde{\mathbf C}(\mathfrak a)
\]
satisfies \eqref{item2TpartE+}.
Let us now show that $Z^{\ell,+}_{\mathfrak a,\mathfrak b}$ is relatively open in $\mathcal R^{+,\ell}\theta \cap \mathtt p_2 \mathfrak b$. More precisely, we prove the following
\begin{claim}
\label{Cl_oww'neigh}
For all $w \in Z^{\ell,+}_{\mathfrak a,\mathfrak b}$, there exist
\[
w' \in \partial^+ \theta(\mathfrak a,w) \cap w + \inter_{\mathrm{rel}} C^{\ell,+}_{\mathfrak a,\mathfrak b} \quad \text{and} \quad w'' \in \partial^- \theta(\mathfrak a,w) \cap w - \inter_{\mathrm{rel}} C^{\ell,+}_{\mathfrak a,\mathfrak b}
\]
such that
\begin{equation*}
\inter_{\mathrm{rel}} \mathtt{O}(\mathfrak a)(w'',w') \subset Z^{\ell,+}_{\mathfrak a,\mathfrak b}.
\end{equation*}
\end{claim}
By \eqref{E_forward_regular_transport_set} $(i)$ and \eqref{E_ballFmax} there exists $w'\in \partial^+ \theta(\mathfrak a,w) \cap w + \inter_{\mathrm{rel}} C^{\ell,+}_{\mathfrak a,\mathfrak b}$ s.t. $w'\in Z^{\ell,+}_{\mathfrak a,\mathfrak b}$ and by \eqref{E_forward_regular_transport_set} $(iii)$ there exists $w'' \in \partial^- \theta(\mathfrak a,w) \cap w - \inter_{\mathrm{rel}} C^{\ell,+}_{\mathfrak a,\mathfrak b}$ s.t. $w''\in Z^{\ell,+}_{\mathfrak a,\mathfrak b}$. By Proposition \ref{P_parall}, $\inter_{\mathrm{rel}}\mathtt{O}(\mathfrak a)(w'',w')$ is a relatively open neighborhood of $w$ in $\mathtt p_2 \mathfrak b$.
Let now $\bar w \in \inter_{\mathrm{rel}}\mathtt{O}(\mathfrak a)(w'',w')$. By the completeness of the superdifferential \eqref{E_compl_fol},
\[
\bar w \in \partial^+ \theta(\mathfrak a,w'') \quad \text{and} \quad w' \in \partial^+\theta(\mathfrak a,\bar w).
\]
Hence,
\[
C^{\ell,+}_{\mathfrak a,\mathfrak b} = \mathbb{R}^+\mathcal D^+\theta(\mathfrak a,w')\subset\mathbb{R}^+\mathcal D^+\theta(\mathfrak a,\bar w)\subset\mathbb{R}^+\mathcal D^+\theta(\mathfrak a,w'')=C^{\ell,+}_{\mathfrak a,\mathfrak b},
\]
thus implying that $\bar w\in Z^{\ell,+}_{\mathfrak a,\mathfrak b}$.
\end{proof}
The proof of Corollary \ref{C_v} follows easily from the proof of Theorem \ref{T_partition_E+-}.
\begin{proof}[Proof of Corollary \ref{C_v}]
If $w \in \mathcal R\theta(\mathfrak a)$, then
\[
w \in Z^{\ell,+}_{\mathfrak a,\mathfrak b}\cap Z^{\ell',-}_{\mathfrak a,\mathfrak b'}\subset \mathcal R^{+,\ell} \theta(\mathfrak a) \cap \mathcal R^{-,\ell'} \theta(\mathfrak a).
\]
By the transitivity property \eqref{E_transitivity_subdifferential_a} one has that
\[
\ell = \ell', \qquad \mathcal D^+ \theta(\mathfrak a,w) = \mathcal D^- \theta(\mathfrak a,w).
\]
Hence, as in the proof of Claim \ref{Cl_oww'neigh},
\[
w \in \inter_{\mathrm{rel}} \mathtt{O}(\mathfrak a)(w'',w') \subset Z^{\ell,+}_{\mathfrak a,\mathfrak b} \cap Z^{\ell,-}_{\mathfrak a,\mathfrak b} , \quad C^{\ell,+}_{\mathfrak a,\mathfrak b} = C^{\ell,-}_{\mathfrak a,\mathfrak b},
\]
yielding the conclusion.
\end{proof}
In the following we will use the notation $\hat\mathbf D^+$, $\hat\mathbf D^-$ and $\hat\mathbf D$ to denote the $\sigma$-compact graphs of the directed locally affine partitions induced respectively by $\mathtt v^+$, $\mathtt v^-$ and $\mathtt v$, namely
\begin{subequations}
\label{E_sub_directed_partition_all}
\begin{align}
\label{E_sub_directed_partition_minus}
\hat{\mathbf D}^+ := \Big\{ \big( \ell,\mathfrak a,\mathfrak b,w,C \big) : \mathtt v^+(\mathfrak a,w) = \mathfrak b, C = \mathtt p_2 \mathfrak b-w \cap \tilde{\mathbf C}(\mathfrak a), w \in Z^{\ell,+}_{\mathfrak a,\mathfrak b} \Big\},
\end{align}
\begin{align}
\label{E_sub_directed_partition_plus}
\hat{\mathbf D}^- := \Big\{ \big( \ell,\mathfrak a,\mathfrak b,w,C \big) : \mathtt v^-(\mathfrak a,w) = \mathfrak b, C = \mathtt p_2 \mathfrak b-w \cap \tilde{\mathbf C}(\mathfrak a), w \in Z^{\ell,-}_{\mathfrak a,\mathfrak b} \Big\},
\end{align}
\begin{align}
\label{E_sub_directed_partition}
\hat{\mathbf D} := \Big\{ \big( \ell,\mathfrak a,\mathfrak b,w,C \big) : \mathtt v(\mathfrak a,w) = \mathfrak b, C = \mathtt p_2 \mathfrak b-w\cap \tilde{\mathbf C}(\mathfrak a), w \in Z^\ell_{\mathfrak a,\mathfrak b} \Big\}.
\end{align}
\end{subequations}
We will also use the notations $\mathfrak c = (\mathfrak a,\mathfrak b)$, $Z^{\ell,+}_{\mathfrak c}$, $C^{\ell,+}_{\mathfrak c}$. Recalling that $\mathfrak{A}\subset\mathbb{R}^{d-k}$ and observing that, after the partition into sheaf sets as in Proposition \ref{P_countable_partition_in_reference_directed_planes} and the injection into a fibration, we can take as in \eqref{E_mathfrak_A_k_def}
\[
\mathfrak B_{\ell,+} = \bigcup _{\mathfrak a \in \mathfrak{A}} \mathfrak B_{\ell,+}(\mathfrak a) \subset \mathbb{R}^{k-\ell},
\]
we let $\mathfrak c \in \mathfrak C_{\ell,+} \subset \mathbb{R}^{d-\ell}$, $\ell=1,\dots,k-1$.
\subsection{\texorpdfstring{Analysis of the residual set}{Analysis of the residual set}}
\label{Ss_analysis_residual_set}
Now we give a characterization of the residual set as the union of \emph{initial} and \emph{final points} respectively for the superdifferential partition and the subdifferential partition. Moreover, we fully characterize the super/subdifferentials at each point of the super/subdifferential partitions in terms of the regular and initial/final points.
Recalling Definition \ref{D_initial_final} of initial and final points of a directed locally affine partition, let
\begin{equation}
\label{E_init_Z+}
\mathcal I^+\theta := \bigcup_{\ell,\mathfrak a,\mathfrak b} \mathcal I \bigl( Z^{\ell,+}_{\mathfrak a,\mathfrak b} \bigr)
\end{equation}
be the \emph{sets of initial points of the superdifferential partition $\{Z^{\ell,+}_{\mathfrak a,\mathfrak b}, C^{\ell,+}_{\mathfrak a,\mathfrak b}\}_{\ell,\mathfrak a,\mathfrak b}$} and let
\begin{equation}
\label{E_final_Z-}
\mathcal E^-\theta := \bigcup_{\ell,\mathfrak a,\mathfrak b} \mathcal E \bigl( Z^{\ell,-}_{\mathfrak a,\mathfrak b} \bigr)
\end{equation}
be the \emph{set of final points of the subdifferential partition $\{Z^{\ell,-}_{\mathfrak a,\mathfrak b}, C^{\ell,-}_{\mathfrak a,\mathfrak b}\}_{\ell,\mathfrak a,\mathfrak b}$}.
\begin{theorem}
\label{T_partition_E}
The following holds:
\begin{equation}
\label{Point_residual_N_is_initial_final}
\mathcal N \theta = \mathcal I^+\theta \cup \mathcal E^-\theta,
\end{equation}
and moreover
\begin{subequations}
\label{Point_sudiff_is_partition_in_fin}
\begin{equation}
\label{Point_sudiff_is_partition_in}
\partial^+ \theta(\mathfrak a,w) = \big( w + C^{\ell,+}_{\mathfrak a,\mathfrak b} \big) \cap \big( Z^{\ell,+}_{\mathfrak a,\mathfrak b} \cup \mathcal E^- \theta(\mathfrak a) \big), \qquad \forall\, w \in Z^{\ell,+}_{\mathfrak a,\mathfrak b},
\end{equation}
\begin{equation}
\label{Point_sudiff_is_partition_fin}
\partial^- \theta(\mathfrak a,w) = \big( w - C^{\ell,-}_{\mathfrak a,\mathfrak b} \big) \cap \big( Z^{\ell,-}_{\mathfrak a,\mathfrak b} \cup \mathcal I^+ \theta(\mathfrak a) \big), \qquad \forall\, w \in Z^{\ell,-}_{\mathfrak a,\mathfrak b}.
\end{equation}
\end{subequations}
\end{theorem}
\begin{remark}
\label{R_not_only_R}
Notice that in general the points of the set $\mathcal N \theta$, i.e. the complement of the set of regular points $\mathcal R\theta$, may not belong to the set $\mathcal I \theta \cup \mathcal E \theta$, i.e. the set of initial and final points for the directed locally affine partition $\{Z^\ell_{\mathfrak a,\mathfrak b},C^\ell_{\mathfrak a,\mathfrak b}\}_{\ell,\mathfrak a,\mathfrak b}$ induced by $\mathtt v$ on $\mathcal R\theta$ (see Figure \ref{Fi_infin}).
\end{remark}
\begin{figure}
\centering{\resizebox{9cm}{7cm}{\input{sudakovinitialfinalbis.pdf_t}}}
\caption{The yellow region is made of points in the regular set $\mathcal R\theta$. The points $z$ and $z'$ in the figure belong to $\mathcal I^+\theta\setminus (\mathcal I\theta\cup\mathcal E\theta)$. The points on the red segments belong instead to $\mathcal E\theta$ and the points on the green segments to $\mathcal I\theta$.}
\label{Fi_infin}
\end{figure}
The main observation in the proof of Theorem \ref{T_partition_E} is contained in the following
\begin{remark}
\label{R_t+t-i+e-}
We observe the following properties of $\mathcal T^+ \theta$, $\mathcal T^- \theta$:
\begin{align*}
& w \in \mathcal T^+\theta(\mathfrak a) \quad \Longrightarrow \quad \exists\,r>0, \, Z^{\ell,+}_{\mathfrak a,\mathfrak b} \text{ such that } B^k(w,r) \cap \big( w + \inter_{\mathrm{rel}} C^{\ell,+}_{\mathfrak a,\mathfrak b} \big) \subset Z^{\ell,+}_{\mathfrak a,\mathfrak b}; \\
& w \in \mathcal T^-\theta(\mathfrak a) \quad \Longrightarrow \quad \exists\,r>0, \, Z^{\ell,-}_{\mathfrak a,\mathfrak b} \text{ such that } B^k(w,r) \cap \big( w - \inter_{\mathrm{rel}} C^{\ell,-}_{\mathfrak a,\mathfrak b} \big) \subset Z^{\ell,-}_{\mathfrak a,\mathfrak b}.
\end{align*}
The statements follow respectively from \eqref{E_ballFmax} and \eqref{E_ballFmax-}, taking $C^{\ell,\pm}_{\mathfrak a,\mathfrak b}$ equal to any maximal extremal cone of $\mathbb{R}^+\mathcal D^\pm\theta(\mathfrak a,w)$.
\end{remark}
\begin{proof}[Proof of Theorem \ref{T_partition_E}]
Let $w \in \mathcal N\theta(\mathfrak a)$: then from definition \eqref{E_residual_set_N},
\[
w \in \mathcal T^+\theta(\mathfrak a) \setminus \mathcal R\theta (\mathfrak a) \quad \text{or} \quad w \in \mathcal T^-\theta(\mathfrak a) \setminus \mathcal R\theta(\mathfrak a).
\]
If $w \in \mathcal T^+\theta(\mathfrak a) \setminus \mathcal R^+\theta (\mathfrak a)$, then by Remark \ref{R_t+t-i+e-} and Definition \ref{D_initial_final} it follows that $w\in\mathcal I^+\theta(\mathfrak a)$. If $w \in \bigl(\mathcal T^+\theta(\mathfrak a)\cap\mathcal R^+\theta (\mathfrak a)\bigr)\setminus \mathcal R^-\theta (\mathfrak a)$, $w\in\mathcal T^-\theta(\mathfrak a)$ by \eqref{E_forward_regular_transport_set} (iii) and then again by Remark \ref{R_t+t-i+e-} we conclude that $w\in\mathcal E^-\theta(\mathfrak a)$. \\
Analogously, $\mathcal T^-\theta(\mathfrak a) \setminus \mathcal R^-\theta (\mathfrak a)\subset\mathcal E^-\theta(\mathfrak a)$ and $\bigl(\mathcal T^-\theta(\mathfrak a)\cap\mathcal R^-\theta (\mathfrak a)\bigr)\setminus \mathcal R^+\theta (\mathfrak a)\subset\mathcal I^+\theta(\mathfrak a)$. This concludes the proof of \eqref{Point_residual_N_is_initial_final}.
We prove \eqref{Point_sudiff_is_partition_in}, being the proof of \eqref{Point_sudiff_is_partition_fin} analogous. We already know by \eqref{item2TpartE+} that
\[
\partial^+\theta(\mathfrak a,w)\subset w + C^{\ell,+}_{\mathfrak a,\mathfrak b} \quad \text{for all } w \in Z^{\ell,+}_{\mathfrak a,\mathfrak b}.
\]
Let now $w\in Z^{\ell,+}_{\mathfrak a,\mathfrak b}$ and $w''\in\partial^-\theta(\mathfrak a,w)\cap w-\inter_{\mathrm{rel}} C^{\ell,+}_{\mathfrak a,\mathfrak b}$ as in \eqref{E_forward_backward_regular_transport_set} (iii). Then, as in Claim \ref{Cl_oww'neigh}, for all $w'\in\partial^+\theta(\mathfrak a,w)$ one has
\[
\inter_{\mathrm{rel}}\mathtt{O}(\mathfrak a)(w'',w')\subset Z^{\ell,+}_{\mathfrak a,\mathfrak b}.
\]
Notice that now we do not need to specify that $w'-w\in\inter_{\mathrm{rel}} C^{\ell,+}_{\mathfrak a,\mathfrak b}$ because we already know $Z^{\ell,+}_{\mathfrak a,\mathfrak b}$ to be open in $\mathrm{aff}(w+C^{\ell,+}_{\mathfrak a,\mathfrak b})$. Then there are two possibilities: either
\begin{equation}
\label{E_firt_T_ele1}
\mathcal D^+\theta(\mathfrak a,w') = C^{\ell,+}_{\mathfrak a,\mathfrak b} \cap \mathbb S^{k-1},
\end{equation}
or by transitivity
\begin{equation}
\label{E_firt_T_ele2}
\mathcal D^+\theta(\mathfrak a,w') \subsetneq C^{\ell,+}_{\mathfrak a,\mathfrak b} \cap \mathbb S^{k-1}.
\end{equation}
In case \eqref{E_firt_T_ele1} holds, it follows that $w'\in Z^{\ell,+}_{\mathfrak a,\mathfrak b}$. Otherwise, one has
\begin{equation}
\label{E_firt_T_ele3}
w' \in \mathcal T^-\theta(\mathfrak a) \setminus \mathcal R^- \theta(\mathfrak a),
\end{equation}
and then as seen before $w'\in\mathcal E^-\theta(\mathfrak a)$. \\
To prove \eqref{E_firt_T_ele3}, it is sufficient to take a maximal backward extremal cone $F$ for $w'$ as in \eqref{E_ballFmax-} containing $C^{\ell,+}_{\mathfrak a,\mathfrak b}$: by \eqref{E_firt_T_ele2},
\[
\partial^+\theta(\mathfrak a,w') \cap \big( w'+\inter_{\mathrm{rel}} F \big) = \emptyset,
\]
and then $w'$ cannot be a backward regular point.
\end{proof}
\subsection{\texorpdfstring{Optimal transportation on $\mathtt c_{\tilde{\mathbf C}}$-Lipschitz foliations}{Optimal transportation on c-Lipschitz foliations}}
\label{Ss_optim_folia}
Let $\{ \theta^{-1}(\mathfrak t) \}_{\mathfrak t\in\mathfrak T}$ be a $\mathtt c_{\tilde{\mathbf C}}$-Lipschitz foliation on $\mathfrak{A}\times\mathbb{R}^k\subset\mathbb{R}^{d-k}\times\mathbb{R}^k$ and consider two probability measures in $\mathcal P(\mathfrak{A}\times\mathbb{R}^k)$ which can be disintegrated as
\[
\tilde \mu = \int \tilde\mu_\mathfrak a d\tilde m(\mathfrak a), \quad \tilde\nu = \int \tilde \nu_\mathfrak a d\tilde m(\mathfrak a), \qquad \tilde m = (\mathtt p_{\mathbb{R}^{d-k}})_\# \tilde \mu = (\mathtt p_{\mathbb{R}^{d-k}})_\# \tilde \nu,
\]
and satisfying
\begin{subequations}
\label{E_541-2}
\begin{equation}
\label{E_541}
\tilde \mu \big( \theta^{-1}(\mathfrak T) \big) = \tilde \nu \big( \theta^{-1}(\mathfrak T) \big) = 1,
\end{equation}
\begin{equation}
\label{E_542}
\emptyset \neq \Pi^f_{\mathtt c_{\tilde{\mathbf C}},\theta}(\tilde \mu,\tilde \nu) := \bigg\{ \pi \in \Pi^f_{\mathtt c_{\tilde{\mathbf C}}}(\tilde \mu,\tilde \nu) : \pi \bigg( \bigcup_{\mathfrak t \in \mathfrak T} \{\theta=\mathfrak t\} \times \{\theta=\mathfrak t\} \bigg) = 1 \bigg\}.
\end{equation}
\end{subequations}
Notice that, by Definition \ref{D_partial_theta} of superdifferential of $\theta$,
\begin{equation}
\label{E_pifct}
\Pi^f_{\mathtt c_{\tilde{\mathbf C}},\theta}(\tilde \mu,\tilde \nu) = \Pi(\tilde \mu,\tilde \nu) \cap \bigl\{ \pi : \pi(\partial^+ \theta) = 1 \bigr\}.
\end{equation}
By Theorem \ref{T_partition_E},
\[
\bigcup_{\mathfrak t \in \mathfrak T} \{\theta=\mathfrak t\} = \mathcal I^+ \theta \cup \mathcal E^- \theta \cup \mathcal R \theta \cup \mathcal F \theta,
\]
hence $\pi\in\Pi^f_{\mathtt c_{\tilde{\mathbf C}},\theta}(\tilde\mu,\tilde\nu)$ if and only if it has $\tilde\mu$ and $\tilde\nu$ as first and second marginal respectively and it is concentrated on the set
\begin{equation}
\label{E_description}
\partial^+ \theta \cap \Bigl[ \big( \mathcal I^+ \theta \cup \mathcal E^- \theta \cup \mathcal R \theta \cup \mathcal F \theta \big) \times \big( \mathcal I^+ \theta \cup \mathcal E^- \theta \cup \mathcal R \theta \cup \mathcal F \theta \big) \Bigr].
\end{equation}
First notice that
\begin{subequations}
\label{E_descr}
\begin{equation}
\label{E_descr1}
\partial^+ \theta \cap \big( \mathcal F \theta \times (\mathfrak{A} \times \mathbb{R}^k) \big) = \partial^+ \theta \cap \big( (\mathfrak{A} \times \mathbb{R}^k) \times \mathcal F \theta \big) = \mathrm{graph}\,\mathbb{I}\, \llcorner_{\mathcal F \theta},
\end{equation}
\begin{equation}
\label{E_descr2}
\partial^+ \theta \cap \big( (\mathfrak{A} \times \mathbb{R}^k) \times (\mathcal T^+ \theta \setminus \mathcal T^- \theta) \big) = \mathrm{graph}\,\mathbb{I}\, \llcorner_{(\mathcal T^+ \theta \setminus \mathcal T^-\theta)},
\end{equation}
\begin{equation}
\label{E_descr3}
\partial^+ \theta \cap \big( (\mathcal T^- \theta \setminus \mathcal T^+ \theta) \times (\mathfrak{A} \times \mathbb{R}^k) \big) = \mathrm{graph}\,\mathbb{I}\, \llcorner_{(\mathcal T^- \theta \setminus \mathcal T^+ \theta)},
\end{equation}
\end{subequations}
and thus the restrictions of all the transport plans in \eqref{E_pifct} to the sets \eqref{E_descr} are concentrated on the diagonal $\{(\mathfrak a,w,\mathfrak a,w):\,(\mathfrak a,w)\in\mathfrak{A}\times\mathbb{R}^k\}$. Hence, we will assume w.l.o.g. that
\begin{equation}
\label{E_543}
\tilde\mu(\mathcal F\theta)=\tilde\nu(\mathcal F\theta)=0, \quad\tilde\mu(\mathcal T^-\theta \setminus \mathcal T^+\theta)=0,\quad\tilde\nu(\mathcal T^+\theta \setminus \mathcal T^-\theta)=0.
\end{equation}
This is true e.g. if $\tilde\mu\perp\tilde\nu$.
Now we give, in the same spirit of Theorem \ref{T_partition_E}, a more accurate description of the set \eqref{E_description} (we will neglect the subsets \eqref{E_descr} by the above observation), independently of the measures $\tilde\mu$, $\tilde\nu$. As a consequence (see Proposition \ref{P_disint_fol} below) we will get that if $\tilde\mu(\mathcal I^+\theta)=0$, then the class of transport plans \eqref{E_pifct} coincides with the class of transport plans of finite cost on the directed locally affine partition of $\mathcal R^{+}\theta$ with quotient map $\mathtt v^+$ defined in Theorem \ref{T_partition_E+-} and, if $\tilde\nu(\mathcal E^-\theta)=0$, they are moreover of finite cost on the directed partition on $\mathcal R\theta$ induced by the map $\mathtt v$.
In Table \ref{Tab_poss_inter} we put on the horizontal and vertical line the sets of a partition of $\mathcal I^+\theta\cup\mathcal R\theta\cup\mathcal E^-\theta$. If $A$ is a set belonging to the horizontal line and $B$ belongs to the vertical, in the square $(A,B)$ we write $Y$ in case possibly $\partial^+\theta \cap (A\times B)\neq\emptyset$, and $N$ in case always $\partial^+\theta \cap (A\times B)=\emptyset$. Recall that $\mathcal R\theta=\mathcal R^+\theta\cap\mathcal R^-\theta$, $\mathcal R^+\theta\cap\mathcal E^-\theta=\mathcal R^+\theta\setminus \mathcal R^-\theta$, $\mathcal R^-\theta\cap\mathcal I^+\theta=\mathcal R^-\theta\setminus \mathcal R^+\theta$, $\mathcal I^-\theta\setminus\mathcal T^+\theta=\mathcal T^-\theta\setminus\mathcal T^+\theta$,
\[
\begin{split}
\mathcal I^+ \theta =&~ (\mathcal I^+ \theta \setminus \mathcal T^- \theta) \cup (\mathcal I^+ \theta \cap \mathcal E^- \theta) \cup (\mathcal R^- \theta\cap \mathcal I^+ \theta ), \crcr
\mathcal E^- \theta =&~ (\mathcal E^- \theta \setminus \mathcal T^+ \theta) \cup (\mathcal I^+ \theta \cap \mathcal E^- \theta) \cup (\mathcal R^+ \theta\cap \mathcal E^- \theta ),
\end{split}
\]
are disjoint unions.
\begin{table}[h]
\caption{The possible intersection of $\partial^+ \theta$ with a partition of $\mathcal T \theta$}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
& $\mathcal R\theta$ & $\mathcal R^+ \theta\cap \mathcal E^-\theta$ & $\mathcal R^- \theta\cap \mathcal I^+\theta$ & $\mathcal I^+ \theta\setminus\mathcal T^-\theta$ & $\mathcal E^- \theta\setminus \mathcal T^+\theta$ & $\mathcal I^+\theta \cap \mathcal E^-\theta$ \\
\hline
$\mathcal R\theta$ & Y & Y & N & N & Y & Y \\
\hline
$\mathcal R^+\theta \cap \mathcal E^-\theta$ & N & Y & N & N & Y & Y\\
\hline
$\mathcal R^-\theta \cap \mathcal I^+\theta$ & Y & Y & Y & N & Y & Y \\
\hline
$\mathcal I^+ \theta\setminus \mathcal T^-\theta$ & Y & Y & Y & Y ($\mathrm{graph}\, \mathbb{I}$) & Y & Y \\
\hline
$\mathcal E^-\theta \setminus \mathcal T^+\theta$ & N & N & N & N & Y ($\mathrm{graph} \,\mathbb{I})$ & N \\
\hline
$\mathcal I^+ \theta\cap \mathcal E^-\theta$ & Y & Y & Y & N & Y & Y \\
\hline
\end{tabular}
\label{Tab_poss_inter}
\end{table}
\begin{proof}[Proof of Table \ref{Tab_poss_inter}]
First of all, \eqref{E_descr2} and \eqref{E_descr3} yield immediately the row of $\mathcal E^- \theta \setminus \mathcal T^+ \theta$ and the column of $\mathcal I^+ \theta \setminus \mathcal T^- \theta$, that we write just for symmetry. Moreover, by Remark \ref{R_t+t-i+e-}, also the row of $\mathcal I^+ \theta \setminus \mathcal T^- \theta$ and the column of $\mathcal E^- \theta \setminus \mathcal T^+ \theta$ easily follow.
In order to prove the other squares we will use the following facts, which have already been proven and used in Sections \ref{Ss_partition_transport_set} and \ref{Ss_analysis_residual_set} and are a consequence of \eqref{E_transitivity_subdifferential_a} and definitions \eqref{E_forward_regular_transport_set} and \eqref{E_backward_regular_transport_set}:
\begin{align}
&w\in\mathcal R\theta(\mathfrak a)\quad\Rightarrow \quad\mathcal D^+\theta(\mathfrak a,w)=\mathcal D^-\theta(\mathfrak a,w);\label{E_table1}\\
&w\in\mathcal R^-\theta(\mathfrak a)\cap \mathcal I^+\theta(\mathfrak a)\quad\Rightarrow \quad\mathcal D^+\theta(\mathfrak a,w)\supsetneq\mathcal D^-\theta(\mathfrak a,w);\label{E_table2}\\
&w\in\mathcal R^+\theta(\mathfrak a)\cap \mathcal E^-\theta(\mathfrak a)\quad\Rightarrow \quad\mathcal D^-\theta(\mathfrak a,w)\supsetneq\mathcal D^+\theta(\mathfrak a,w).\label{E_table3}
\end{align}
Let us first prove the relations given by the squares containing the letter $N$.
Let $w\in\mathcal R^+\theta(\mathfrak a)$ and $w'\in\partial^+\theta(\mathfrak a,w)$. Then,
\[
\mathcal D^+\theta(\mathfrak a,w')\overset{\eqref{E_transitivity_subdifferential_a}}{\subset}\mathcal D^+\theta(\mathfrak a,w)\overset{\eqref{E_table1},\eqref{E_table3}}{\subset}\mathcal D^-\theta(\mathfrak a,w)\overset{\eqref{E_transitivity_subdifferential_a}}{\subset}\mathcal D^-\theta(\mathfrak a,w').
\]
By \eqref{E_table1} and \eqref{E_table2} we then conclude that $\partial^+\theta\cap (\mathcal R\theta\times \mathcal R^-\theta \cap \mathcal I^+\theta)=\emptyset$ and by \eqref{E_table1}-\eqref{E_table3} that $\partial^+\theta\cap(\mathcal R^+\theta \cap \mathcal E^-\theta\times \mathcal R\theta)=\emptyset$, $\partial^+\theta\cap(\mathcal R^+\theta \cap \mathcal E^-\theta\times\mathcal R^-\theta \cap \mathcal I^+\theta)=\emptyset$.
Let us now deal with the relations given by the squares containing the letter $Y$. We refer to Figure \ref{Fi_structsubdiff}.
\begin{figure}
\centering{\resizebox{12cm}{10cm}{\input{sudakovstructsubdiffbis.pdf_t}}}
\caption{The relations given by the letter $Y$.}
\label{Fi_structsubdiff}
\end{figure}
First we claim that
\begin{equation}
\label{E_Y1}
\partial^+\theta\cap (\mathcal R^+\theta\cap \mathcal E^-\theta\times \mathcal I^+\theta\cap\mathcal E^-\theta)\neq\emptyset.
\end{equation}
Indeed, let e.g. $w\in \mathcal R^+\theta(\mathfrak a)\cap \mathcal E^-\theta(\mathfrak a)$ such that
\[
\begin{split}
& \dim(\mathbb{R}^+\mathcal D^+\theta(\mathfrak a,w))=2, \\
& \mathcal D^+\theta(\mathfrak a,w)\subsetneq \mathcal D^-\theta(\mathfrak a,w)\subsetneq \mathrm{conv}_{\mathbb{S}^{k-1}} \mathcal D^-\theta(\mathfrak a,w), \\
& \dim (\mathbb{R}^+\mathrm{conv}_{\mathbb{S}^{k-1}} \mathcal D^-\theta(\mathfrak a,w))=3,
\end{split}
\]
and $w'\in\partial^+\theta(\mathfrak a,w)$ s.t.
\[
\mathcal D^-\theta(\mathfrak a,w')= \mathrm{conv}_{\mathbb{S}^{k-1}} \mathcal D^-\theta(\mathfrak a,w) \quad \text{but} \quad \mathcal D^+\theta(\mathfrak a,w')\subsetneq \mathrm{conv}_{\mathbb{S}^{k-1}}\mathcal D^+\theta(\mathfrak a,w')=\mathcal D^+\theta(\mathfrak a,w).
\]
Next we claim that
\begin{equation}
\label{E_Y2}
\partial^+\theta\cap (\mathcal R\theta\times \mathcal R^+\theta\cap \mathcal E^-\theta)\neq\emptyset.
\end{equation}
Indeed, it is sufficient to take e.g. $w\in\mathcal R\theta(\mathfrak a)$ with $\dim (\mathbb{R}^+\mathcal D^+\theta(\mathfrak a,w))=2$ and $w'\in \partial^+\theta(\mathfrak a,w)$ s.t.
\[
\mathcal D^-\theta(\mathfrak a,w')=\mathcal D^+\theta(\mathfrak a,w) \quad \textrm{but} \quad \mathcal D^+\theta(\mathfrak a,w')=\mathrm{conv}_{\mathbb{S}^{k-1}}\mathcal D^+\theta(\mathfrak a,w')\subsetneq \mathcal D^-\theta(\mathfrak a,w').
\]
Now let us prove that
\begin{equation}
\label{E_Y3}
\partial^+\theta\cap (\mathcal R^-\theta\cap \mathcal I^+\theta\times \mathcal R\theta)\neq\emptyset.
\end{equation}
Take for example $w\in\mathcal R^-\theta(\mathfrak a)\cap \mathcal I^+\theta(\mathfrak a)$ and $\bar w\in\partial^+\theta(\mathfrak a,w)\cap w+\inter_{\mathrm{rel}} F$ as in \eqref{E_ballFmax} with $F$ maximal face in $\mathbb{R}^+\mathcal D^+\theta(\mathfrak a,w)$ containing $\mathbb{R}^+\mathcal D^-\theta(\mathfrak a,w)$.
Finally we claim that
\begin{equation}
\label{E_Y4}
\partial^+\theta\cap (\mathcal I^+\theta\cap\mathcal E^-\theta\times\mathcal R^-\theta\cap \mathcal I^+\theta)\neq\emptyset.
\end{equation}
Hence, by the transitivity property \eqref{E_transitivity_subdifferential_a}, \eqref{E_Y1}-\eqref{E_Y4} immediately give all the remaining $Y$ squares in the table.
In order to show \eqref{E_Y4}, let $w\in\mathcal I^+\theta(\mathfrak a)\cap\mathcal E^-\theta(\mathfrak a)$ with
\[
\begin{split}
& \mathcal D^+\theta(\mathfrak a,w)=\mathrm{conv}_{\mathbb{S}^{k-1}}\mathcal D^+\theta(\mathfrak a,w), \\
& \dim\mathbb{R}^+\mathcal D^+\theta(\mathfrak a,w)=3 \quad \text{and} \\
& \mathcal D^-\theta(\mathfrak a,w)\subsetneq\mathrm{conv}_{\mathbb{S}^{k-1}}\mathcal D^-\theta(\mathfrak a,w)\subsetneq\mathcal D^+\theta(\mathfrak a,w).
\end{split}
\]
Then, let $w'\in \partial^+\theta(\mathfrak a,w)\,\cap\, w+\inter_{\mathrm{rel}} (\mathbb{R}^+\mathrm{conv}_{\mathbb{S}^{k-1}}\mathcal D^-\theta(\mathfrak a,w))$ s.t.
\[
\mathcal D^+\theta(\mathfrak a,w')=\mathcal D^+\theta(\mathfrak a,w) \quad \text{and} \quad \mathcal D^-\theta(\mathfrak a,w')=\mathrm{conv}_{\mathbb{S}^{k-1}}\mathcal D^-\theta(\mathfrak a,w).
\]
\end{proof}
Neglecting \eqref{E_descr}, we conclude that
\begin{align}
\partial^+ \theta \cap \Big[ \big( \mathcal I^+ \theta \cup \mathcal R \theta \cup \mathcal E^- \theta \big) \times \big( \mathcal I^+ \theta \cup \mathcal R \theta \cup \mathcal E^- \theta \big) \Big] =&~ \Bigl[ \partial^+ \theta \cap \big( \mathcal R \theta \times (\mathcal R \theta \cup \mathcal E^- \theta) \big) \Bigr] \label{E_subdiff_dec}\\
&~ \cup \Bigl[ \partial^+ \theta \cap \big( (\mathcal R^+ \theta \cap \mathcal E^- \theta) \times \mathcal E^- \theta \big) \Bigr] \notag\\
&~ \cup \Bigl[ \partial^+ \theta \cap (\mathcal I^+ \theta \times \mathcal T^- \theta) \Bigr].\notag
\end{align}
Using \eqref{E_subdiff_dec}, Theorem \ref{T_partition_E+-} and Corollary \ref{C_v} we have the following.
\begin{proposition}
\label{P_disint_fol}
Let $\{\theta^{-1}(\mathfrak t)\}_{\mathfrak t\in\mathfrak T} \subset \mathtt P(\mathfrak{A} \times\mathbb{R}^k)$ be a $\mathtt c_{\tilde{\mathbf C}}$-Lipschitz foliation and $\tilde \mu,\tilde \nu \in \mathcal P(\mathfrak{A} \times \mathbb{R}^k)$ such that \eqref{E_541-2} and \eqref{E_543} hold.
If $\tilde\mu(\mathcal I^+\theta)=0$, then
\[
\tilde \pi \in \Pi^f_{\mathtt c_{\tilde{\mathbf C}},\theta}(\tilde \mu,\tilde \nu) \quad \Longleftrightarrow \quad \tilde \pi \in \Pi^f_{\mathtt c_{\hat{\mathbf D}^+}}(\tilde \mu,\tilde \nu),
\]
where $\hat{\mathbf D}^+$ is the locally affine partition induced by $\mathtt v^+$ on $\mathcal R^+\theta$ and $\mathtt c_{\hat{\mathbf D}^+}$ the related cost, as defined in \eqref{E_c_bD}.
Moreover, if $\tilde\mu(\mathcal I^+\theta)=0$ and $\tilde\mu(\mathcal R^+\theta\cap\mathcal E^-\theta)=0$ then
\[
\tilde \pi \in \Pi^f_{\mathtt c_{\tilde{\mathbf C}},\theta}(\tilde \mu,\tilde \nu) \quad \Longleftrightarrow \quad \tilde \pi \in \Pi^f_{\mathtt c_{\hat{\mathbf D}}}(\tilde\mu,\tilde\nu),
\]
where $\hat{\mathbf D}$ is the locally affine partition induced by $\mathtt v=\mathtt v^+_{|_{\mathcal R\theta}}$ on $\mathcal R\theta$ and $\mathtt c_{\hat{\mathbf D}}$ the related cost.
\end{proposition}
We end this section with the following special case.
\begin{proposition}
\label{P_hat_bf_D_graph}
Let $\mathfrak t\in\mathfrak T$, $\mathfrak a=\mathtt p_\mathfrak{A}(\{\theta=\mathfrak t\})$. Then \eqref{E_more_than_complet} holds for the differential directed locally affine partition of the regular set of $\{\theta=\mathfrak t\}$
\[
\{Z^\ell_{\mathfrak a,\mathfrak b}, C^\ell_{\mathfrak a,\mathfrak b}\}_{\overset{\ell=1,\dots,k-1}{\mathfrak b\in\mathfrak B_{\ell}(\mathfrak a)}}.
\]
\end{proposition}
\begin{proof}
For $(\ell,\mathfrak b) \not= (\ell',\mathfrak b')$, let
\[
\bar z \in (Z^\ell_{\mathfrak a,\mathfrak b} + C^\ell_{\mathfrak a,\mathfrak b}) \cap (Z^{\ell'}_{\mathfrak a,\mathfrak b'} + C^{\ell'}_{\mathfrak a,\mathfrak b'})\cap \mathtt p_{\mathbb{R}^k}\hat{\mathbf D}(\mathfrak t),
\]
where $\{Z^\ell_{\mathfrak a,\mathfrak b},C^\ell_{\mathfrak a,\mathfrak b}\}_{\ell,\mathfrak b}$ is the directed locally affine partition of $\hat{\mathbf D}(\mathfrak t)$. Hence, $\bar z\in w+C^\ell_{\mathfrak a,\mathfrak b}\cap w'+C^{\ell'}_{\mathfrak a,\mathfrak b'}$ for some $w\in Z^\ell_{\mathfrak a,\mathfrak b}$, $w'\in Z^{\ell'}_{\mathfrak a,\mathfrak b'}$ and, since by assumption $\theta$ is constant (namely, equal to $\mathfrak t$) on $\mathtt p_{\mathbb{R}^k}\hat{\mathbf D}(\mathfrak t)$, then by definition of superdifferential
\[
\bar z\in \partial^+\theta(\mathfrak a,w)\cap\partial^+\theta(\mathfrak a,w').
\]
By the transitivity property and the fact that $\bar z$ is regular
\[
C^\ell_{\mathfrak a,\mathfrak b}\cup C^{\ell'}_{\mathfrak a,\mathfrak b'}\subset\mathbb{R}^+\mathcal D^-\theta(\mathfrak a,\bar z)=\mathbb{R}^+\mathcal D^+\theta(\mathfrak a,\bar z)\subset C^\ell_{\mathfrak a,\mathfrak b}\cap C^{\ell'}_{\mathfrak a,\mathfrak b'},
\]
which implies that $C^\ell_{\mathfrak a,\mathfrak b}=C^{\ell'}_{\mathfrak a,\mathfrak b'}$, thus $\ell=\ell'$ and $\mathrm{aff} Z^\ell_{\mathfrak a,\mathfrak b}=\mathrm{aff} Z^{\ell'}_{\mathfrak a,\mathfrak b'}$.
Hence, by definition of $\mathtt v$, $Z^\ell_{\mathfrak a,\mathfrak b}=Z^{\ell'}_{\mathfrak a,\mathfrak b'}$ contradicting our initial assumption.
\end{proof}
\begin{remark}
\label{R_4.27}
By Proposition \ref{P_hat_bf_D_graph}, the differential partition of a single complete $\mathtt c_{\tilde C}$-Lipschitz graph satisfies \eqref{E_more_than_complet}.
\end{remark}
From Proposition \ref{P_hat_bf_D_graph} and Proposition \ref{P_dispiani_2}, one obtains immediately the following
\begin{corollary}
\label{C_transp_graph}
If $\tilde\mu$, $\tilde\nu$ are as in Proposition \ref{P_disint_fol} and $\tilde\mu(\mathcal I^+\theta)=0$, $\tilde\nu(\mathcal E^-\theta)=0$, then
\[
\tilde\pi\in\Pi^f_{\mathtt c_{\tilde{\mathbf C}}, \theta}(\tilde\mu,\tilde\nu)\quad\Leftrightarrow\quad\tilde\pi=\int\tilde\pi^\ell_{\mathfrak c} \, d\tilde m(\mathfrak c),\quad\tilde\pi^\ell_\mathfrak c\in\Pi^f_{\mathtt c_{C^\ell_\mathfrak c}}(\tilde\mu^\ell_\mathfrak c,\tilde\nu^\ell_\mathfrak c),
\]
where $\{\tilde\pi^\ell_\mathfrak c\}$, $\{\tilde\mu^\ell_\mathfrak c\}$ and $\{\tilde\nu^\ell_\mathfrak c\}$ are respectively the disintegrations of $\tilde\pi$, $\tilde\mu$, $\tilde\nu$ w.r.t. the partition induced by $\mathtt v$.
\end{corollary}
\section{Dimensional reduction on directed partitions via cone approximation property}
\label{S_disintechnique}
In this section we recall, in an abstract and more general setting, the main steps of the disintegration technique first introduced in \cite{BianchGlo} for partitions into segments and then extended to locally affine partitions of any dimension in \cite{CarDan}. This technique allows to prove the absolute continuity of the conditional probabilities of the Lebesgue measure and to deduce that the \emph{initial} and \emph{final points} of a directed locally affine partition are Lebesgue negligible, provided the direction map satisfies a suitable regularity assumption that we call (\emph{initial/final}) \emph{forward/backward cone approximation property}. For more details on the proofs of the results contained in this section, we refer to \cite{CarDan}, Section 4.
\subsection{Model sets of directed segments}
\label{Ss_model_dir_segm}
We first deal with \emph{model sets of directed segments}, namely $1$-dimensional sheaf sets whose projection on their reference line is a given segment. At the end of the paragraph, the forward/backward cone approximation property for these model sets will be introduced as a sufficient condition in order to have absolutely continuous disintegrations.
\begin{definition}
\label{D_modeldire_segm}
A \emph{model set of directed segments} or \emph{1-dimensional model set} is a 1-dimensional directed sheaf set $\{Z^1_\mathfrak a,C^1_\mathfrak a\}_{\mathfrak a\in\mathfrak{A}^1}$ with $\sigma$-continuous direction vector field
\begin{equation}
\label{E_mathit_v_def}
\mathtt d : \bigcup_{\mathfrak a \in \mathfrak{A}^1} Z^1_\mathfrak a \rightarrow \mathbb \mathbb{S}^{d-1}, \qquad \mathtt d : Z^1_\mathfrak a \ni z \mapsto C^1_\mathfrak a \cap \mathbb{S}^{d-1},
\end{equation}
and reference line $\langle\mathtt e\rangle$ for which there exist $h^-, h^+ \in \mathbb{R}$, $h^- < h^+$, such that
\[
\mathtt p_{\langle \mathtt e \rangle} (Z^1_\mathfrak a) = (h^-,h^+) \mathtt e \qquad \forall\, \mathfrak a \in \mathfrak{A}^1.
\]
\end{definition}
We will also call \emph{model set} the set $\mathbf Z^1= \underset{\mathfrak a \in \mathfrak{A}^1}{\cup}\, Z^1_\mathfrak a$, and we say that the triple $(\mathtt e, h^-,h^+)$ is a \emph{reference configuration}. Moreover we assume that
\[
\mathcal L^{d}(\mathbf Z^1) < +\infty.
\]
For shortness we will sometimes use the notation $\mathbf Z^1(\mathtt d,\mathfrak{A}^1,\mathtt e,h^-,h^+)$.
We also set $\overline{\mathbf Z}^1=\overline{\mathbf Z^1}$ as in \eqref{E_mathbf_Z_base_partition}
\begin{equation}
\label{E_clos_Z_1a}
\overline{\mathbf Z}^1 := \bigcup_{\mathfrak a\in\mathfrak{A}^1} \mathrm{clos}\, Z^1_\mathfrak a.
\end{equation}
Notice that
\[
\overline{\mathbf Z}^1 \cap \mathtt p^{-1}_{\langle\mathtt e\rangle} \big( (h^-,h^+) \mathtt e \big) = \mathbf Z^1.
\]
Given a $1$-dimensional model set $\{Z^1_\mathfrak a,C^1_\mathfrak a\}_{\mathfrak a\in\mathfrak{A}^1}$ with reference configuration $(\mathtt e,h^-,h^+)$, define the perpendicular \emph{sections}
\begin{equation}
\label{E_P_t_section}
P_t := \overline{\mathbf Z}^1 \cap \mathtt p^{-1}_{\langle \mathtt e \rangle}(t \mathtt e), \qquad t \in [h^-,h^+].
\end{equation}
Clearly from Definition \ref{D_initial_final} one has
\begin{equation*}
P_{h^-} = \mathcal I(\mathbf Z^1),\qquad P_{h^+} = \mathcal E(\mathbf Z^1),
\end{equation*}
where $\mathcal I(\mathbf Z^1)$, $\mathcal E(\mathbf Z^1)$ are the initial/final points of $\{Z^1_\mathfrak a,C^1_\mathfrak a\}_{\mathfrak a\in\mathfrak{A}^1}$.
For all $t\in[h^-,h^+]$, denote also
\begin{equation}
\label{E_d_t_def}
\mathtt d^t := \mathtt d \llcorner_{P_t},
\end{equation}
where $\mathtt d^{h^-} : P_{h^-} \rightarrow \mathbb{S}^{d-1}$, $\mathtt d^{h^+} : P_{h^+} \rightarrow \mathbb{S}^{d-1}$ are the multivalued extensions of $\mathtt d$ defined by
\begin{equation}
\label{E_tt_d_ext}
\mathtt d^{h^-}(z) = \bigl\{ C^1_\mathfrak a \cap \mathbb{S}^{d-1} : z \in \mathcal I(Z^1_\mathfrak a) \bigr\}, \qquad \mathtt d^{h^+}(z) = \bigl\{ C^1_\mathfrak a \cap \mathbb{S}^{d-1} : z \in \mathcal E(Z^1_\mathfrak a) \bigr\}.
\end{equation}
\begin{lemma}
\label{L_reg_tt_d_h_pm}
The sets $\mathrm{graph}\,\mathtt d^{h^-}$, $\mathrm{graph}\,\mathtt d^{h^+}$ are $\sigma$-compact, and hence there exist Borel sections
\[
\mathcal I(\mathbf Z^1) \ni z \mapsto \tilde{\mathtt d}^{h^-}_+(z) \in \mathtt d^{h^-}(z), \quad \mathcal E(\mathbf Z^1) \ni z \mapsto \tilde{\mathtt d}^{h^+}_-(z) \in \mathtt d^{h^+}(z).
\]
\end{lemma}
\begin{proof}
W.l.o.g., we prove the result only for $\mathrm{graph}\,\mathtt d^{h^-}$.
Since $\mathtt d$ is $\sigma$-compact, let $\mathbf Z^1 = \underset{l}{\cup}\, \mathbf Z^1_l$ such that $\mathbf Z^1_l$ are compact and $\mathtt d \llcorner_{\mathbf Z^1_l}$ is continuous. Then it is fairly easy to see that the multivalued maps
\[
\mathtt p^{-1}_{\langle \mathtt e \rangle} (h^- \mathtt e) \cap \big( z + \mathbb{R} \mathtt d(z) \big) \mapsto \mathtt d(z), \quad z \in \mathbf Z^1_l,
\]
are compact, and thus the regularity of $\mathrm{graph}\,\mathtt d^{h^-}$ follows.
The existence of sections $\tilde{\mathtt d}^{h^-}_+$, $\tilde{\mathtt d}^{h^+}_-$ with Borel regularity is standard for compact multifunctions (see for example Theorem 5.2.1, page 189 of \cite{Sri:courseborel}) and an easy argument yields the conclusion.
\end{proof}
We will define the vector fields
\begin{subequations}
\label{E_tilde_tt_d_pm}
\begin{equation}
\label{E_tilde_tt_d_+}
\tilde{\mathtt d}_+ (z) := \tilde{\mathtt d}^{h^-}_+(z') \qquad \text{if} \quad z \in (z' + \mathbb{R} \tilde{\mathtt d}^{h^-}_+(z')) \cap \mathtt p^{-1}_{\langle \mathtt e \rangle}([h^-,h^+]), \ z' \in \mathrm{dom}\,\tilde{\mathtt d}^{h^-}_+,
\end{equation}
\begin{equation}
\label{E_tilde_tt_d_-}
\tilde{\mathtt d}_- (z) := \tilde{\mathtt d}^{h^+}_-(z') \qquad \text{if} \quad z \in (z' + \mathbb{R} \tilde{\mathtt d}^{h^+}_-(z')) \cap \mathtt p^{-1}_{\langle \mathtt e \rangle}([h^-,h^+]), \ z' \in \mathrm{dom}\,\tilde{\mathtt d}^{h^+}_-.
\end{equation}
\end{subequations}
In particular,
\[
\tilde{\mathtt d}_+ \llcorner_{P_{h^-}} = \tilde{\mathtt d}_+^{h^-}, \qquad \tilde{\mathtt d}_- \llcorner_{P_{h^+}}=\tilde{\mathtt d}_-^{h^+},
\]
and
\[
\tilde{\mathtt d}_\pm \llcorner_{\mathtt p^{-1}_{\langle\mathtt e\rangle} ((h^-,h^+) \mathtt e)} = \mathtt d \llcorner_{\mathrm{dom}\,\tilde{\mathtt d}_\pm \cap \mathtt p^{-1}_{\langle\mathtt e\rangle} ((h^-,h^+) \mathtt e) }.
\]
Define for $s \in (h^-,h^+)$, $t \in [h^-,h^+]$ the map
\begin{equation}
\label{E_sigma_s_t}
\sigma^{s,t} : P_s \rightarrow P_t,
\qquad
z \mapsto \displaystyle{z + (t-s) \frac{\mathtt d(z)}{\mathtt d(z) \cdot \mathrm e}},
\end{equation}
which sends each point of the section $P_s$ in the unique point of $P_t$ which belongs to the same segment of the model set. It is a bijection for $t \in (h^-,h^+)$, with inverse $\sigma^{t,s}$.
Given Borel measurable selections $\tilde{\mathtt d}_\pm$, as in \eqref{E_tilde_tt_d_+} and \eqref{E_tilde_tt_d_-}, one defines for $s \in [h^-,h^+)$ ($s \in (h^-,h^+]$), $t \in [h^-,h^+]$ the map
\begin{equation}
\label{E_tilde_sigma_s_t}
\tilde \sigma^{s,t}_\pm : P_s\cap \mathrm{dom}\, \tilde{\mathtt d}_\pm \rightarrow P_t,
\qquad
z \mapsto \displaystyle{z + (t-s) \frac{\tilde{\mathtt d}_\pm(z)}{\tilde{\mathtt d}_\pm(z) \cdot \mathrm e}}.
\end{equation}
Notice that $\tilde \sigma^{s,t}_\pm $ coincides with $\sigma^{s,t}_\pm\llcorner_{\mathrm{dom}\,\tilde{\mathtt d}_\pm \cap \mathtt p^{-1}_{\langle\mathtt e\rangle} ((h^-,h^+) \mathtt e) }$ for $s\neq h^\mp$.
\subsubsection{Cone approximation property and absolute continuity}
\label{Sss_cone_approx}
We recall that our problem is the following: if
\[
\mathcal L^{d} \llcorner_{\mathbf Z^1} = \int \upsilon_\mathfrak a \,d\eta(\mathfrak a)
\]
is the disintegration of $\mathcal L^{d} \llcorner_{\mathbf Z^1}$ w.r.t. the partition $\{Z^1_\mathfrak a\}_{\mathfrak a\in\mathfrak{A}^1}$, then we ask whether
\[
\upsilon_\mathfrak a \ll \mathcal H^1 \llcorner_{Z^1_\mathfrak a} \quad \text{ and/or } \quad \upsilon_\mathfrak a \simeq \mathcal H^1 \llcorner_{Z^1_\mathfrak a}.
\]
By the next lemma, the absolute continuity problem along the segments $\{Z^1_\mathfrak a\}_\mathfrak a$ can be reduced to an absolute continuity problem for the push-forward of $\mathcal H^{d-1}$ on the sections through the maps $\sigma^{s,t}$.
\begin{lemma}
\label{L_equivarea}
Let us fix a section $P_t$ of $\mathbf Z^1$, $t \in (h^-,h^+)$. Then, the following two statements are equivalent:
\begin{equation}
\label{E_equivarea1}
\sigma^{s,t}_\# \big( \mathcal H^{d-1} \llcorner_{P_s} \big) \ll \mathcal H^{d-1} \llcorner_{P_t},\quad\text{for $\mathcal L^1$-a.e. $s \in(h^-,h^+)$;}
\end{equation}
\begin{equation}
\label{E_equivarea4}
\eta \ll \mathcal H^{d-1} \llcorner_{P_t} \quad \text{and} \quad \upsilon_\mathfrak a \ll \mathcal H^1 \llcorner_{Z^1_\mathfrak a}, \quad\text{for $\eta$-a.e. $\mathfrak a$.}
\end{equation}
Moreover, also the following two statements are equivalent:
\begin{equation}
\label{E_equivarea2}
\sigma^{t,s}_\# \big( \mathcal H^{d-1} \llcorner_{P_t} \big) \ll \mathcal H^{d-1} \llcorner_{P_s} \quad\text{and formula \eqref{E_equivarea1} hold for $\mathcal L^1$-a.e. $s\in(h^-,h^+)$;}
\end{equation}
\begin{equation}
\label{E_equivarea3}
\eta \simeq \mathcal H^{d-1} \llcorner_{P_t} \quad \text{and} \quad \upsilon_\mathfrak a \simeq \mathcal H^1 \llcorner_{Z^1_\mathfrak a}, \quad\text{for $\eta$-a.e. $\mathfrak a$.}
\end{equation}
\end{lemma}
In particular, whenever \eqref{E_equivarea3} holds, $\mathbf Z^1$ is a regular partition according to Definition \ref{D_disint_regular}.
The proof of Lemma \ref{L_equivarea} is an application of Fubini-Tonelli theorem w.r.t. the projection on $\langle\mathtt e\rangle$ and the change of variables formula for the maps $\sigma^{s,t}$. We give a short proof for completeness.
\begin{proof}
By \eqref{E_equivarea1}
\[
\sigma^{s,t}_\# \big( \mathcal H^{d-1} \llcorner_{P_s} \big) = \mathtt f(s,t,\cdot) \mathcal H^{d-1} \llcorner_{P_t}
\]
for some Borel non-negative function $\mathtt f$, for a.e. $s\in(h^-,h^+)$ and thus by Fubini-Tonelli theorem we can write for a compactly supported function $\phi : \mathbb{R}^d \rightarrow \mathbb{R}$
\[
\begin{split}
\int \phi \mathcal L^d \llcorner_{\mathbf Z^1} =&~ \int_{h^-}^{h^+} \bigg[ \int_{P_s} \phi(z) d\mathcal H^{d-1}(z) \bigg] ds = \int_{h^-}^{h^+} \bigg[ \int_{P_t} \phi(\sigma^{t,s}(z)) d\big( \sigma^{s,t}_\# \mathcal H^{d-1} \llcorner_{P_s} \big)(z) \bigg] ds \crcr
=&~ \int_{h^-}^{h^+} \bigg[ \int_{P_t} \phi(\sigma^{t,s}(z)) \mathtt f(s,t,z) d\mathcal H^{d-1}(z) \bigg] ds = \int_{P_t} \bigg[ \int_{h^-}^{h^+} \phi(\sigma^{t,s}(z)) \mathtt f(s,t,z) ds \bigg] d\mathcal H^{d-1}(z).
\end{split}
\]
By the uniqueness of the disintegration (see Theorem \ref{T_disint}), this shows that \eqref{E_equivarea4} is true. Repeating the argument starting from the end, one can prove the equivalence of \eqref{E_equivarea1} and \eqref{E_equivarea4}.
By using the additional assumption \eqref{E_equivarea2}, $\mathtt f(s,t,\cdot)$ is $\mathcal H^{d-1}$-a.e. strictly positive on $P_t$, for a.e. $s\in (h^-,h^+)$ and the equivalence of \eqref{E_equivarea2} with \eqref{E_equivarea3} follows immediately.
\end{proof}
Now we are ready to introduce the forward/backward cone approximation properties, which will imply the assumptions of Lemma \ref{L_equivarea}. For notational convenience, we will state the definition of cone vector field.
\begin{figure}
\centering{\resizebox{12cm}{10cm}{\input{sudakovmodelsetbis.pdf_t}}}
\caption{A model set of directed segments and a union of two cone vector fields.}
\label{Fi_modelset}
\end{figure}
\begin{definition}
\label{D_cone_vector}
The \emph{cone vector field} with base in $E_1 \subset \mathbb{R}^{d}$, and vertex $\bar z \in E_2 \subset \mathbb{R}^{d} \setminus E_1$ is defined as
\begin{equation*}
\begin{array}{ccccc}
\mathtt d &:& E_1 \supset \mathrm{dom}\,\mathtt d &\rightarrow& \mathbb S^{d-1} \crcr
&& z &\mapsto& \mathtt d(z) := \frac{\bar z- z}{\|\bar z- z\|}
\end{array}
\end{equation*}
We say that $\mathtt d$ is a \emph{finite union of cone vector fields} with base in $E_1$ and vertices in $E_2$ if there exist finitely many cone vector fields $\{\mathtt d_i\}_{i=1}^I$ with bases in $E_1$ and vertices $\{\bar z_i\}_{i=1}^I$ in $E_2 \subset \mathbb{R}^{d} \setminus E_1$ such that the sets
\begin{equation*}
E_{\mathtt d_i} := \big\{ (1-t) \bar z_i + t z, t \in [0,1], z \in \mathrm{dom}\, \mathtt d_i \big\}, \quad i = 1,\dots,I,
\end{equation*}
satisfy $E_{\mathtt d_i} \cap E_{\mathtt d_j} = \emptyset$, for all $i \neq j$.
\end{definition}
\begin{definition}
\label{D_forw_back}
We say that the model set of directed segments $\mathbf Z^1(\mathtt d,\mathfrak{A}^1,\mathtt e,h^-,h^+)$ has the \emph{forward cone approximation property} if there exists $\epsilon > 0$ such that for all $t \in (h^-,h^+)$ there exists $\{\mathtt d^t_j\}_{j \in \mathbb{N}}$ finite union of cone vector fields with base in $\mathtt p^{-1}_{\langle\mathtt e\rangle}(t \mathtt e)$ and vertices in $\mathtt p^{-1}_{\langle\mathtt e\rangle}((h^+ + \epsilon)\mathtt e)$ such that
\[
\mathcal H^{d-1}(P_t\setminus\mathrm{dom}\, \mathtt d^t_j) = 0
\]
and $\mathtt d^t_j \rightarrow \mathtt d^t$ $\mathcal H^{d-1} \llcorner_{P_t}$-a.e..
Analogously, we say that the model set of directed segments $\mathbf Z^1(\mathtt d,\mathfrak{A}^1,\mathrm e,h^-,h^+)$ has the \emph{backward cone approximation property} if there exists $\epsilon > 0$ such that for all $t \in (h^-,h^+)$ there exists $\{\mathtt d^t_j\}_{j \in \mathbb{N}}$ finite union of cone vector fields with base in $\mathtt p^{-1}_{\langle\mathtt e\rangle}(t \mathtt e)$ and vertices in $\mathtt p^{-1}_{\langle\mathtt e\rangle}((h^- - \epsilon) \mathtt e)$ such that
\[
\mathcal H^{d-1}(P_t\setminus\mathrm{dom}\, \mathtt d^t_j) = 0
\]
and $\mathtt d^t_j \rightarrow -\mathtt d^t$ $\mathcal H^{d-1} \llcorner_{P_t}$-a.e..
\end{definition}
\begin{lemma}
\label{L_forward_esti_area}
If $\mathbf Z^1(\mathtt d,\mathfrak{A}^1,\mathtt e,h^-,h^+)$ has the forward cone approximation property, then for $h^- < s \leq t \leq h^+$
\begin{equation}
\label{E_forward_esti_area1}
\sigma^{s,t}_\# \mathcal H^{d-1} \llcorner_{P_s} \leq \bigg( \frac{h^+ + \epsilon - s}{h^+ + \epsilon - t} \bigg)^{d-1} \mathcal H^{d-1} \llcorner_{P_t}.
\end{equation}
Analogously, if $\mathbf Z^1(\mathtt d,\mathfrak{A}^1,\mathtt e,h^-,h^+)$ has the backward cone approximation property, then for $h^- \leq t < s <h^+$
\begin{equation}
\label{E_forward_esti_area2}
\sigma^{s,t}_\# \mathcal H^{d-1} \llcorner_{P_s} \leq \bigg( \frac{s-h^-+ \epsilon}{t-h^-+ \epsilon} \bigg)^{d-1} \mathcal H^{d-1} \llcorner_{P_t}.
\end{equation}
\end{lemma}
\begin{proof}
We prove only the first estimate, since the proof of the second is completely similar.
It is fairly easy to see that the estimates hold for the map $\sigma^{s,t}$ associated to a cone vector field with bases in $P_s$ and vertices in $P_{h^++\epsilon}$ as in Definition \ref{D_cone_vector} by similitude criteria for triangles or equivalently by the polar change of coordinates in $\mathbb{R}^d$. Hence, the same estimate holds also for the finite unions of cone vector fields approximating $\mathtt d^s$ as in Definition \ref{D_forw_back}.
Restricting by Egorov's Theorem to continuous and uniformly convergent sequences $\{\mathtt d^s_j\}_{j\in\mathbb{N}}$ on compact subsets of $P_s$, by the u.s.c. of $\mathcal H^{d-1}$ on the hyperplanes perpendicular to $(h^-,h^+)\mathtt e$ w.r.t. the Hausdorff convergence of compact sets, the inequality immediately passes to the limit.
\end{proof}
It is straightforward to observe that \eqref{E_forward_esti_area1} implies \eqref{E_equivarea1} and \eqref{E_forward_esti_area2} implies the first part of \eqref{E_equivarea2}.
Hence we have the following
\begin{corollary}
\label{C_disinte_ac}
If $\mathbf Z^1(\mathtt d,\mathfrak{A}^1,\mathtt e,h^-,h^+)$ has either the forward cone approximation property or the backward cone approximation property, then
\[
\eta \ll \mathcal H^{d-1} \llcorner_{\mathfrak{A}^1} \quad \text{and} \quad \upsilon_\mathfrak a \ll \mathcal H^1 \llcorner_{Z^1_\mathfrak a} \ \text{for $\eta$-a.e. $\mathfrak a \in \mathfrak{A}^1$.}
\]
If both the forward cone approximation and the backward cone approximation properties hold, then $\mathbf Z^1$ is a regular partition, i.e.
\[
\eta \simeq \mathcal H^{d-1} \llcorner_{\mathfrak{A}^1} \quad \text{and} \quad \upsilon_\mathfrak a \simeq \mathcal H^1 \llcorner_{Z^1_\mathfrak a} \ \text{for $\eta$-a.e. $\mathfrak a \in \mathfrak{A}^1$.}
\]
\end{corollary}
We can extend the forward/backward cone approximation properties to Borel sections of initial/final points. This will be useful later (see Theorem \ref{T_FC_no_initial}), when we will give conditions ensuring that $\mathcal L^d(\mathcal I(\mathbf Z))=0$/$\mathcal L^d(\mathcal E(\mathbf Z))=0$ for a directed locally affine partition $\mathbf Z$.
\begin{definition}
\label{D_initial_forward}
We say that $\mathbf Z^1(\mathtt d,\mathfrak{A}^1,\mathtt e,h^-,h^+)$ satisfies the \emph{initial forward cone approximation property} if there exists a Borel section $\tilde{\mathtt d}_+$ which satisfies the assumptions of the forward cone approximation property of Definition \ref{D_forw_back} for all $t \in [h^-,h^+)$.
Similarly, $\mathbf Z^1(\mathtt d,\mathfrak{A}^1,\mathtt e,h^-,h^+)$ satisfies the \emph{final backward cone approximation property} if there exists a Borel section $\tilde{\mathtt d}_-$ which satisfies the assumptions of the backward cone approximation property of Definition \ref{D_forw_back} for all $t\in(h^-,h^+]$.
\end{definition}
The next lemma is the analogue of Lemma \ref{L_forward_esti_area} for the Borel sections $\tilde{\mathtt d}_\pm$.
\begin{lemma}
\label{Cinfin}
If $\mathbf Z^1(\mathtt d,\mathfrak{A}^1,\mathtt e,h^-,h^+)$ satisfies the initial (final) forward (backward) cone approximation property, then \eqref{E_forward_esti_area1} (resp. \eqref{E_forward_esti_area2}) holds for $\tilde\sigma^{s,t}_+$ ($\tilde\sigma^{s,t}_-$), for all $h^- \leq s \leq t \leq h^+$ (resp. $h^- \leq t \leq s\leq h^+$).
\end{lemma}
\subsection{\texorpdfstring{$k$-dimensional model sets}{k-dimensional model sets}}
\label{Ss_k_dim_model_set}
In this section we extend the results for $1$-dimensional model sets to the $k$-dimensional model sets defined below.
\begin{definition}
\label{D_k_dim_model}
A \emph{$k$-dimensional model set} is a $k$-dimensional directed sheaf set $\{Z^k_\a,C^k_\a\}_{\mathfrak a\in\mathfrak{A}^k}$ with reference plane $V^k = \langle \mathtt e^k_1,\dots,\mathtt e^k_k \rangle$, for which there exist $\mathrm h^- = (h^-_1,\dots,h^-_k)$, $\mathrm h^+ = (h^+_1,\dots,h^+_k) \in \mathbb R^k$, with $h^-_j < h^+_j$ for all $j=1,\dots,k$, such that
\begin{equation}
\label{E_pkdimen}
\mathtt p_{V^k} Z^k_\a = \inter_{\mathrm{rel}} U \big( \{\mathtt e^k_i\},\mathrm h^-,\mathrm h^+ \big) := \bigg\{ \sum_{j=1}^k t_j \mathtt e^k_j : t_j \in (h^-_j,h^+_j) \bigg\}.
\end{equation}
\end{definition}
We will also call \emph{$k$-dimensional model set} the set $\mathbf Z^k=\underset{\mathfrak a\in\mathfrak{A}^k}{\cup}Z^k_\a$. Setting
\begin{equation*}
\mathcal D : \mathbf Z^k \rightarrow \mathcal C(k,\mathbb{R}^d) \cap \mathbb{S}^{d-1}, \quad \mathcal D : Z^k_\a \ni z \mapsto \mathcal D(z) := C^k_\a \cap \mathbb{S}^{d-1}
\end{equation*}
for the direction map, sometimes we will use the more precise notation $\mathbf Z^k(\mathcal D,\mathfrak{A}^k,\mathtt e^k_1,\dots,\mathtt e^k_k,\mathrm h^-,\mathrm h^+)$ and we say that the $3k$-tuple $(\mathtt e^k_1,\dots,\mathtt e^k_k, \mathrm h^-,\mathrm h^+)$ is a \emph{reference configuration} for the model set.
The absolute continuity problem for the disintegration of the Lebesgue measure on a $k$-dimensional model set $\mathbf Z^k$ can be reduced to the absolute continuity problem for $1$-dimensional model sets obtained cutting $\mathbf Z^k$ with suitable $(d-k+1)$-dimensional planes, called \emph{slices}.
Indeed, for all $\mathtt e \in C(\{\mathtt e^k_i\})\cap\mathbb{S}^{d-1}$ and $w\in \inter_{\mathrm{rel}} U(\{\mathtt e^k_i\},\mathrm h^-,\mathrm h^+)$, set
\begin{subequations}
\begin{equation*}
h^-(w,\mathtt e) := \inf \Big\{ t \in \mathbb{R} : w + t \mathtt e \in U \big( \{\mathtt e^k_i\},\mathrm h^-,\mathrm h^+ \big) \Big\},
\end{equation*}
\begin{equation*}
h^+(w,\mathtt e) := \sup \Big\{ t \in \mathbb{R} : w + t \mathtt e \in U \big( \{\mathtt e^k_i\},\mathrm h^-,\mathrm h^+ \big) \Big\}.
\end{equation*}
\end{subequations}
\begin{definition}
\label{D_1_sim_slice}
Given a $k$-dimensional model set $\mathbf Z^k(\mathcal D,\mathfrak{A}^k,\mathrm e^k_1,\dots,\mathrm e^k_k,\mathrm h^-,\mathrm h^+)$ we define \emph{$1$-dimensional slice} of $\mathbf Z^k$ in the direction $\mathtt e\in C(\{\mathtt e^k_i\})\cap\mathbb{S}^{d-1}$ any set of the form
\begin{equation}
\label{E_slicekdim}
\mathbf Z^k \cap \mathtt p^{-1}_{V^k} \Big( w + \big( h^-(w,\mathtt e),h^+(w,\mathtt e) \big) \mathtt e \Big), \quad w \in U \big( \{\mathtt e^k_i\},\mathrm h^-,\mathrm h^+ \big).
\end{equation}
\end{definition}
The important observation is the following:
\begin{remark}
\label{Prop_1dim_slice}
By \eqref{E_pkdimen} and Point $(3)$ of Definition \ref{D_sheaf_set}, the $1$-dimensional slice \eqref{E_slicekdim} is a model set of directed segments in the $(d+1-k)$-dimensional space $\mathtt p^{-1}_{V^k}(w + \langle\mathtt e\rangle)$ with direction vector field
\begin{equation}
\label{E_mathtt_d_e}
\mathtt d_\mathtt e:=\mathcal D\cap\mathtt p^{-1}_{V^k}(\langle \mathtt e\rangle),
\end{equation}
quotient space $\mathfrak{A}^k$ and reference configuration $(\mathtt e, h^-(w,\mathtt e),h^+(w,\mathtt e))$.
\end{remark}
As a consequence we obtain the following $k$-dimensional version of Lemma \ref{L_forward_esti_area}.
\begin{lemma}
\label{L_one_d_slicing_FC}
If the forward (or backward) cone approximation property holds for all the $1$-dimensional slices of a $k$-dimensional model set $\mathbf Z^k$ in the directions $\mathtt e^k_1,\dots\mathtt e^k_k$, then
\begin{equation*}
\mathcal L^d \llcorner_{\mathbf Z^k} = \int \upsilon^k_\a\, d\eta(k,a), \quad \text{with} \ \upsilon^k_\a \ll \mathcal H^{k}\llcorner_{Z^k_\a} \ \text{and} \ \eta(k) \ll \mathcal H^{d-k} \llcorner_{\mathfrak{A}^k}.
\end{equation*}
If both the forward and the backward properties hold for all the above slices, then $\mathbf Z^k$ is regular, namely
\begin{equation}
\label{E_equiv_Zk}
\eta(k) \simeq \mathcal H^{d-k} \llcorner_{\mathfrak{A}^k} \quad \text{and} \quad \upsilon^k_\a \simeq \mathcal H^{k} \llcorner_{Z^k_\a} \ \text{for $\eta$-a.e. $(k,\mathfrak a)$}.
\end{equation}
\end{lemma}
\begin{proof}
W.l.o.g. we assume $\mathtt e^k_i$ to be the first $k$ unit vectors of a standard orthonormal base in $\mathbb{R}^d$, $h_j^-<0<h_j^+$ $\forall\,j=1,\dots,k$, and $\mathfrak{A}^k = \mathbf Z \cap \mathtt p_{V^k}^{-1}(0)$.
For $x \in U(\{\mathtt e^k_i\},\mathtt h^-,\mathtt h^+) \cap \mathtt p^{-1}_{\langle \mathtt e^k_1 \rangle}(0)$, consider the $1$-dimensional slice
\[
\mathbf Z^1_x := \mathbf Z^k \cap \mathtt p^{-1}_{V^k} \big( x + (h^-_1,h^+_1) \mathtt e^k_1 \big) = \bigcup_{\mathfrak a \in \mathfrak{A}^k} Z^k_\mathfrak a \cap \mathtt p^{-1}_{V^k} \big( x + (h^-_1,h^+_1) \mathtt e^k_1 \big).
\]
By Fubini-Tonelli theorem
\[
\mathcal L^d \llcorner_{\mathbf Z^k} = \int _{\mathbf Z^k\cap\mathtt p^{-1}_{\langle \mathtt e_1^k\rangle}(0)}\mathcal L^{d-k+1} \llcorner_{\mathbf Z^1_{x}} \mathcal L^{k-1}(dx)
\]
and applying Lemma \ref{L_forward_esti_area} in the case of forward (or backward) approximation property, we obtain the disintegration
\[
\mathcal L^d \llcorner_{\mathbf Z^k} = \int \upsilon^1_{\mathfrak a,x} d\eta_1(x,\mathfrak a), \quad \text{with} \ \upsilon^1_{\mathfrak a,x} \ll \mathcal H^1 \llcorner_{Z^k_\mathfrak a \cap \mathtt p^{-1}_{V^k} ( x + (h^-_1,h^+_1) \mathtt e^k_1 )}, \ \eta_1 \ll \mathcal H^{d-1} \llcorner_{\mathbf Z^k \cap\, \mathtt p_{\langle \mathtt e^k_1 \rangle}^{-1}(0)}.
\]
Now one repeats the procedure starting with the measure $\mathcal L^{d-1}$ restricted to the $(k-1)$-dimensional model set in $\mathbb{R}^{d-1}$ given by
\[
\mathbf Z^{k-1} = \mathbf Z^k \cap \mathtt p_{\langle \mathtt e^k_1 \rangle}^{-1}(0)
\]
and considering directions along $\mathtt e^k_2$, which allows to write
\[
\eta_1 = \int_{\mathbf Z^k\cap\mathtt p_{\langle \mathtt e^k_1,\mathtt e^k_2 \rangle}^{-1} (0)} \eta_{1,\mathfrak a,y} d\eta_2(y,\mathfrak a), \quad \text{with} \ \eta_{1,\mathfrak a,y} \ll \mathcal H^1 \llcorner_{Z^k_\mathfrak a \cap \mathtt p^{-1}_{V^k}(y+(h_2^-,h_2^+)\mathtt e^k_2)}, \ \eta_2 \ll \mathcal H^{d-2} \llcorner_{\mathbf Z^k \cap \mathtt p^{-1}_{\langle\mathtt e^k_1,\mathtt e^k_2 \rangle}(0)}
\]
Hence by composing the two disintegrations one obtains
\[
\mathcal L^d \llcorner_{\mathbf Z^k} = \int \upsilon^2_{\mathfrak a,y} d\eta_2(y,\mathfrak a), \quad \text{with} \ \upsilon^2_{\mathfrak a,y} \ll \mathcal H^2 \llcorner_{Z^k_{\mathfrak a} \cap \mathtt p^{-1}_{V^k}(y+(h^-_1,h^+_1) \mathtt e^k_1+(h_2^-,h_2^+)\mathtt e^k_2)}.
\]
Iterating the process $k$-times, one obtains the result.
In case both approximation properties holds, the same analysis shows \eqref{E_equiv_Zk}.
\end{proof}
\subsection{\texorpdfstring{$k$-dimensional sheaf sets and $\mathcal D$-cylinders}{k-dimensional sheaf sets and D-cylinders}}
\label{Ss_k_dim_sheaf}
Now we apply the cone approximation property technique also to general $k$-directed sheaf sets.
Let $\{Z^k_\a,C^k_\a\}_{\mathfrak a\in\mathfrak{A}^k}$ be a $k$-directed sheaf set with reference plane $V^k = \langle \mathtt e^k_1,\cdots,\mathtt e^k_k \rangle$, base rectangle $U(\{\mathtt e_i^k\},\mathtt h^-,\mathtt h^+)$ and direction map
\begin{equation}
\label{E_k_dim_dire_map}
\mathcal D^k(\mathfrak a) := C^k_\a \cap \mathbb S^{d-1}.
\end{equation}
For $\mathfrak{A}^{k,'} \subset \mathfrak{A}^k$ $\sigma$-compact, set
\begin{equation}
\label{E_bf_Zk'_def}
\mathbf Z^{k,'} := \bigcup_{\mathfrak a \in \mathfrak{A}^{k,'}} Z^k_\a.
\end{equation}
\begin{definition}
\label{D_mathcal_D_cyl}
Any $k$-dimensional model set $\mathbf Z^k(\mathcal D, \mathfrak{A}^{k,'},\mathtt e^k_1,\dots,\mathtt e^k_k,\mathrm h^-, \mathrm h^+)$ of the form
\begin{equation*}
\mathbf Z^k \big( \mathcal D, \mathfrak{A}^{k,'},\mathtt e^k_1,\dots,\mathtt e^k_k,\mathrm h^-, \mathrm h^+ \big) = \mathbf Z^{k,'} \cap \mathtt p_{V^k}^{-1} \big(\inter_{\mathrm{rel}} U(\{\mathtt e^k_i\},\mathrm h^-,\mathrm h^+) \big)
\end{equation*}
for which there exists $\epsilon>0$ such that
\[
\begin{split}
&\mathbf Z^k \Big( \mathcal D,\mathfrak{A}^{k,'},\mathtt e^k_1,\dots,\mathtt e^k_k, \mathrm h^- -(\epsilon,\dots,\epsilon), \mathrm h^+ + (\epsilon,\dots,\epsilon) \Big) \\
& \qquad \qquad \qquad = \mathbf Z^{k,'} \cap \mathtt p_{V^k}^{-1} \Big( \inter_{\mathrm{rel}} U \big( \{\mathtt e^k_i\},\mathrm h^--(\epsilon,\dots,\epsilon),\mathrm h^++(\epsilon,\dots,\epsilon) \big) \Big)
\end{split}
\]
is also a $k$-dimensional model set, will be called \emph{$k$-dimensional $\mathcal D$-cylinder}.
\end{definition}
In particular, by the above definition and Definition \ref{D_k_dim_model}
\[
\mathtt p_{V^k} Z ^k_\a \supset \inter_{\mathrm{rel}} U \Big( \{\mathtt e^k_i\},\mathrm h^--(\epsilon,\dots\,\epsilon), \mathrm h^++(\epsilon,\dots\,\epsilon) \Big) \qquad \forall\, \mathfrak a \in \mathfrak{A}^{k,'}.
\]
\begin{remark}
\label{R_property}
Notice that, since any $Z^k_\a$ is a relatively open set, then the sheaf set can be covered by a countable disjoint collection of $k$-dimensional $\mathcal D$-cylinders
\begin{equation}
\label{E_bf_Z_k'_n}
\mathbf Z^{k,'}_n = \mathbf Z^k \big( \mathcal D, \mathfrak{A}^{k,'}_n,\mathtt e^k_1,\dots,\mathtt e^k_k,\mathrm h^-_n, \mathrm h^+_n \big), \qquad n \in \mathbb{N},
\end{equation}
up to the points which belong to the perpendicular sections
\[
\mathbf Z^{k,'}_n \cap \mathtt p^{-1}_{V^k} \big( \partial U(\{\mathtt e^k_i\},\mathrm h_n^-, \mathrm h_n^+) \big).
\]
In particular, the $k$-dimensional $\mathcal D$-cylinders as in \eqref{E_bf_Z_k'_n} define a partition of the sheaf set $\mathbf Z^{k,'}$ up to an $\mathcal L^d$-negligible set.
\end{remark}
\begin{definition}
\label{D_1_dim_slice_sheaf}
Define \emph{$1$-dimensional slices} of a directed locally affine partition $\{Z^k_\a, C^k_\a\}_{\underset{\mathfrak a\in\mathfrak{A}^k}{k=1,\dots,d}}$ the $1$-dimensional slices of any of the $k$-dimensional $\mathcal D$-cylinders given by \eqref{E_bf_Z_k'_n}, for any of the countably many $k$-directed sheaf sets $\mathbf Z^k$ given by Proposition \eqref{P_countable_partition_in_reference_directed_planes}.
\end{definition}
\begin{remark}
\label{R_slice0}
Notice that, for any $1$-dimensional slice of a $\mathcal D$-cylinder
\[
\mathbf Z^{k,'}_n \cap \mathtt p^{-1}_{V^k} \Big( w_n + \big( h^-(w_n,\mathtt e_n),h^+(w_n,\mathtt e_n) \big) \mathtt e_n \Big), \quad w_n \in \inter_{\mathrm{rel}} U \big( \{\mathtt e^k_i\},\mathrm h^-_n,\mathrm h^+_n \big), \ \mathtt e_n \in C(\{\mathtt e^k_i\}),
\]
as in \eqref{E_slicekdim}, there exists $\epsilon = \epsilon_n >0$ such that the set
\[
\mathbf Z^{k,'}_n \cap \mathtt p^{-1}_{V^k} \Big( w_n + \big( h^-(w_n,\mathtt e_n) - \epsilon_n,h^+(w_n,\mathtt e_n) + \epsilon_n \big) \mathtt e_n \Big)
\]
is still a $1$-dimensional model set. This assures that the extreme points of the segments of the slice do not contain relative boundary points of the sets of the partition, and in particular the vector field $\mathtt d_{\mathtt e_n}$ is single-valued up to the boundary of $Z^{k,'}_{n,\mathfrak a}$.
\end{remark}
The main result of this section is then the following theorem.
\begin{theorem}
\label{T_one_d_slicing_FC}
If either the forward cone approximation property or the backward cone approximation property holds for the $1$-dimensional slices of a directed locally affine partition $\{Z^k_\a, C^k_\a\}_{k,\mathfrak a\in\mathfrak{A}^k}$, then
\[
\mathcal L^d \llcorner_{\mathbf Z^k} = \int \upsilon^k_a d\eta(k,\mathfrak a) \quad \text{with} \ \eta(k) \ll \mathcal H^{d-k} \llcorner_{\mathfrak{A}^k} \ \text{and} \ \upsilon^k_\a \ll \mathcal H^{k} \llcorner_{Z^k_\a} \ \text{for $\eta(k)$-a.e.} \ \mathfrak a \in \mathfrak{A}^k.
\]
If both properties hold, then $\mathbf Z^k$ is regular, i.e.
\[
\eta(k) \simeq \mathcal H^{d-k} \llcorner_{\mathfrak{A}^k} \ \text{and} \ \upsilon^k_\a \simeq \mathcal H^{k} \llcorner_{Z^k_\a} \ \text{for $\eta(k)$-a.e.} \ \mathfrak a \in \mathfrak{A}^k.
\]
\end{theorem}
\begin{proof}
By Proposition \ref{P_countable_partition_in_reference_directed_planes} and Remark \ref{R_property}, using a simple covering argument and $\sigma$-additivity of measures one can reduce to study the absolute continuity of the disintegrations on $\mathcal D$-cylinders. Then, by Definition \ref{D_1_dim_slice_sheaf}, one concludes using Lemma \ref{L_one_d_slicing_FC} .
\end{proof}
\subsection{Negligibility of initial/final points}
\label{Ss_neglig_init_fin}
Now we deal with the other measure-theoretic problem connected to directed locally affine partitions, namely to establish whether
\[
\mathcal L^d(\mathcal I) = 0 \quad \text{ and/or } \quad \mathcal L^d(\mathcal E) = 0.
\]
It turns out that these properties are implied by the validity of the initial/final cone approximation properties for \emph{initial/final $1$-dimensional slices}.
\begin{definition}
\label{L_init_fin_1_dim_slice}
An \emph{initial $1$-dimensional slice} of a directed locally affine partition $\{Z^k_\a, C^k_\a\}_{\underset{\mathfrak a\in\mathfrak{A}^k}{k=1,\dots,d}}$ is a $1$-dimensional model set of the form
\[
\mathbf Z^k \cap \mathtt p_{V^k}^{-1} \big(w+(h^-,h^+)\mathtt e \big)
\]
for a $k$-directed sheaf set $\mathbf Z^k $ with reference plane $V^k=\langle\mathtt e_1,\dots,\mathtt e_k\rangle$, $\mathtt e\in C(\{\mathtt e_i^k\})\cap\mathbb S^{d-1}$, for which there exists $\epsilon > 0$ such that the set
\[
\mathbf Z^k \cap \mathtt p_{V^k}^{-1} \big( w+(h^-,h^++\epsilon) \mathtt e \big)
\]
is still a $1$-dimensional model set.
Similarly, a \emph{final $1$-dimensional slice} of a directed locally affine partition $\{Z^k_\a, C^k_\a\}_{\underset{\mathfrak a\in\mathfrak{A}^k}{k=1,\dots,d}}$ is a $1$-dimensional model set of the form
\[
\mathbf Z^k \cap \mathtt p_{V^k}^{-1} \big( w+(h^-,h^+)\mathtt e \big)
\]
for which there exists $\epsilon>0$ such that the set
\[
\mathbf Z^k \cap \mathtt p_{V^k}^{-1} \big( w+(h^--\epsilon,h^+) \mathtt e \big)
\]
is still a $1$-dimensional model set.
\end{definition}
By Remark \ref{R_slice0}, the vector field $\mathtt d_\mathtt e$ of an initial slice can be multivalued only at the points of the section $P_{h^-}$, while for a final slice it can be multivalued only on $P_{h^+}$.
\begin{remark}
\label{R_infinslices}
Notice that a $1$-dimensional slice of a directed locally affine partition according to Definition \ref{D_1_dim_slice_sheaf} is both an initial/final $1$-dimensional slice. In particular, since the direction vector field of a $1$-dimensional slice is single-valued, its Borel-measurable sections coincide trivially with itself and then the initial/final forward/backward cone approximation properties (see Definition \ref{D_initial_forward}) are simply an extension of the forward/backward cone approximation property (see Definition \ref{D_forw_back}) to the initial/final points of the slice. Hence, saying that the initial/final $1$-dimensional slices of a directed locally affine partition satisfy the initial forward/final backward cone approximation property implies that the $1$-dimensional slices of that directed locally affine partition satisfy the forward/backward cone approximation property.
\end{remark}
The following theorem follows from Corollary \ref{Cinfin}, as in the density Lemma 4.19 proved in \cite{CarDan} for the relative boundary points of the locally affine partition into the faces of a convex function.
\begin{theorem}
\label{T_FC_no_initial}
If the initial $1$-dimensional slices of a directed locally affine partition satisfy the initial forward cone approximation property, then
\[
\mathcal L^d(\mathcal I) = 0.
\]
Similarly, if the final $1$-dimensional slices of a directed locally affine partition satisfy the final backward cone approximation property, then
\[
\mathcal L^d(\mathcal E) = 0.
\]
\end{theorem}
We give only a sketch of the proof, since the details have already been given in Lemma 4.19 of \cite{CarDan}.
\begin{proof}
W.l.o.g. we can restrict to a $k$-directed sheaf set $Z^k$ with reference $k$-plane $V^k=\langle\mathtt e^k_1,\dots,\mathtt e^k_k\rangle$. Let us then consider the map
\begin{equation}
\label{E_r_init_map}
\mathcal I \ni z \mapsto \mathtt l(z) := \sup \Big\{ r : z + \big( \inter_{\mathrm{rel}}\, C^k_\a \cap B^d(0,r) \big) \subset Z^k_\a, \ \text{for some} \ \mathfrak a \in \mathfrak{A}^k \Big\}.
\end{equation}
By Definition \ref{D_initial_final}, $\mathtt l(z) > 0$ for all $z \in \mathcal I$, and then by a countable covering argument we need only to prove the negligibility of the set
\begin{equation}
\label{E_cal_I_r}
\mathcal I^{\bar r} := \mathcal I \cap \mathtt l^{-1}(\bar r),
\end{equation}
with $\bar r > 0$ fixed.
Assume that $\mathcal L^d(\mathcal I^{\bar r})>0$ and that $\bar z \in \bar{\mathbf Z}^k$ is a Lebesgue point of $\mathcal I^{\bar r}$. Then if $\mathtt e\in C(\{\mathtt e_i^k\})\cap\mathbb S^{d-1}$, at least one of the sets
\[
P_{w + t \mathtt e} = \mathcal I^{\bar r} \cap \mathtt p^{-1}_{V^k} ( w + t \mathtt e), \quad w \in \langle\mathtt e\rangle^\perp \cap V^k,
\]
has $\mathcal H^{d-k}$-positive measure, and we can assume that $\bar z$ is also a Lebesgue point for $\mathcal H^{d-k} \llcorner_{P_{w+t\mathtt e}}$. For definiteness, we will assume that $P_{w + t \mathtt e}=P_{\bar w + h^-(\bar w, \mathtt e) \mathtt e}=\overline{\mathbf Z}^{k,'}\cap\mathtt p^{-1}_{V^k}(\bar w+h^-(\bar w,\mathtt e)\mathtt e)$ for some $\bar w$ in the relative interior of the base rectangle of $Z^{k,'}\subset Z^k$, and let
\[
P_{\bar w + h^-(\bar w, \mathtt e) \mathtt e}\ni z\mapsto\mathtt d^{h^-}(z) := \Big\{ C^k_\a \cap \mathtt p^{-1}_{V^k}\langle\mathtt e\rangle: z \in \mathcal I(Z^k_\a) \Big\}
\]
be the multivalued maps defined in \eqref{E_tt_d_ext} for the $1$-dimensional slice $\mathbf Z^{k,'}_{\bar w, \mathtt e}$ defined by
\[
\mathbf Z^{k,'}_{\bar w,\mathtt e}:=\Big\{ Z^k_\a \cap \mathtt p^{-1}_{V^k} \big( \bar w + (h^-,h^+) \mathtt e \big), C^k_\a \cap \mathtt p^{-1}_{V^k}\langle\mathtt e\rangle\Big\}_{\mathfrak a \in \mathfrak{A}^{k,'}}, \quad h^+ := h^- + \frac{\bar r}{2},
\]
where $\mathfrak{A}^{k,'}$ is the sets of $\mathfrak a$ such that $Z^k_\a$ has an initial point $z$ on $P_{\bar w + h^-(\bar w, \mathtt e) \mathtt e}$ and
\[
z + \inter_{\mathrm{rel}}\, C^k_\a \cap B^d(0,\bar r/2) \subset Z^k_\a.
\]
If $\tilde{\mathtt d}^{h^-}_+$ is a Borel section of $\mathtt d^{h^-}$ chosen accordingly to Definition \ref{D_initial_forward}, then consider the $1$-dimensional slice $\mathbf Z^{k,''}_{\bar w, \mathtt e} \subset \mathbf Z^{k,'}_{\bar w, \mathtt e}$ defined by
\[
\mathbf Z^{k,''}_{\bar w,\mathtt e}:=\Big\{ Z^k_\a \cap \mathtt p^{-1}_{V^k} \big( \bar w + (h^-,h^+) \mathtt e \big), C^k_\a \cap \mathtt p^{-1}_{V^k}(\langle\mathtt e\rangle) \Big\}_{\mathfrak a \in \mathfrak{A}^{k,''}}, \quad h^+ := h^- + \frac{\bar r}{4},
\]
where $\mathfrak{A}^{k,''} \subset \mathfrak{A}^{k,'}$ is the set of $\mathfrak a$ satisfying
\[
z \in P_{\bar w + h^- (\bar w,\mathtt e)\mathtt e} \quad \Longrightarrow \quad \tilde{\mathtt d}^{h^-}_+(z) = C^k_\a \cap \mathtt p^{-1}_{V^k}\langle\mathtt e\rangle \cap \mathbb S^{d-1}.
\]
In other words, $\mathbf Z^{k,''}_{\bar w,\mathtt e}$ is the $1$-dimensional slice whose initial points belong to $P_{\bar w + h^-(\bar w,\mathtt e) \mathtt e}$ and whose segments are given by $\mathrm{dom}\,\tilde{\mathtt d}_+\cap\mathtt p_{V^k}^{-1}\big(\bar w+\big(h^-(\bar w,\mathtt e),h^-(\bar w,\mathtt e)+\frac{\bar r}{4}\big)\big)\mathtt e$, where $\tilde{\mathtt d}_+$ was defined in \eqref{E_tilde_tt_d_+}. Clearly, by restricting to a $\sigma$-compact $\mathcal H^{d-k}$-conegligible subset of $P_{\bar w + h^-(\bar w, \mathtt e) \mathtt e}$ so that $\tilde{\mathtt d}^{h-}_+$ is $\sigma$-continuous, this procedure defines an initial $1$-dimensional slice.
By the initial forward cone approximation property, Lemma \ref{Cinfin} implies that if $\bar z \in \mathcal I(Z^{k,''}_{\bar w,\mathtt e})$ is a Lebesgue point of $\mathcal H^{d-k} \llcorner_{P_{\bar w + h^-(\bar w,\mathtt e) \mathtt e}}$, then
\[
\lim_{r \rightarrow 0} \bigg[ \lim_{t \searrow h^-} \mathcal H^{d-k} \big( \mathbf Z^{k,''}_{\bar w,\mathtt e} \cap P_{\bar w + t \mathtt e} \cap B^d(\bar z,r) \big) \bigg] = 1
\]
Since $\mathbf Z^{k,''}\cap\mathcal I^{\bar r}=\emptyset$, this clearly contradicts the fact that $\bar z$ is a Lebesgue point of $\mathcal I^{\bar r}$.
\end{proof}
In view of Theorems \ref{T_one_d_slicing_FC} and \ref{T_FC_no_initial} and recalling Remark \ref{R_infinslices}, for future convenience we give the following
\begin{definition}
\label{D_coneapprpart}
A directed locally affine partition satisfies the (initial/final) forward/backward cone approximation property if its (initial/final) $1$-dimensional slices satisfy the
(initial/final) forward/backward cone approximation property.
\end{definition}
\section{Proof of Theorem \ref{T_1}}
\label{S_theorem_1_proof}
This section is devoted to the proof of Theorem \ref{T_1}, which we recall below.
\begin{theorem1}
Let $\mu,\nu \in \mathcal P(\mathbb{R}^d)$ with $\mu \ll \mathcal L^d$ and let $\d{\cdot}$ be a convex norm in $\mathbb{R}^d$. Then there exists a locally affine directed partition $\{Z^k_\mathfrak a,C^k_\mathfrak a\}_{\overset{k=0,\dots,d}{\mathfrak a\in\mathfrak{A}^k}}$ in $\mathbb{R}^d$ with the following properties:
\begin{enumerate}
\item for all $\mathfrak a \in \mathfrak{A}^k$ the cone $C^k_\mathfrak a$ is a $k$-dimensional extremal face of $\d{\cdot}$;
\item $\mathcal L^d \Bigl( \mathbb{R}^d \setminus \underset{k,\mathfrak a}{\bigcup}\, Z^k_\mathfrak a \Bigr)=0$;
\item $\{Z^k_\a\}_{k,\mathfrak a}$ is regular, namely the disintegration of the measure $\mathcal L^d$ w.r.t. the partition $\{Z^k_\a\}_{k,\mathfrak a}$, $\displaystyle{\mathcal L^d \llcorner_{\underset{k,\mathfrak a}{\cup} Z^k_\a} = \int v^k_\mathfrak a\,d\eta(k,\mathfrak a)}$, satisfies
\[
v^k_a \simeq \mathcal H^k \llcorner_{Z^k_\mathfrak a}\quad\text{for $\eta(k)$-a.e. $\mathfrak a\in\mathfrak{A}^k$.}
\]
\item \label{Point_4_T_1_h} for all $\pi \in \Pi^{\mathrm{opt}}_{\d{\cdot}}(\mu,\nu)$, the disintegration $\displaystyle{\pi = \int \pi^k_\a\, dm(k,\mathfrak a)}$ w.r.t. the partition $\{Z^k_\a \times \mathbb{R}^d\}_{k,\mathfrak a}$ satisfies
\[
\pi^k_\mathfrak a \in \Pi^f_{\mathtt c_{C^k_\mathfrak a}}(\mu^k_\a, (\mathtt p_2)_\# \pi^k_\a),
\]
where $\displaystyle{\mu = \int \mu^k_\a\,dm(k,\mathfrak a)}$ is the disintegration w.r.t. the partition $\{Z^k_\a\}_{k,\mathfrak a}$, and moreover
\[
(\mathtt p_2)_\#\pi^k_\a \biggl( Z^k_\a \cup \biggl( \mathbb{R}^d \setminus \underset{(k',\mathfrak a') \not= (k,\mathfrak a)}{\bigcup} Z^{k'}_{\mathfrak a'} \biggr) \biggr) = 1.
\]
\end{enumerate}
If also $\nu \ll \mathcal L^d$, then for all $\pi\in\Pi^{\mathrm{opt}}_{\d{\cdot}}(\mu,\nu)$
\[
(\mathtt p_2)_\#\pi^k_\a = \nu^k_\a
\]
where $\displaystyle{\nu = \int \nu^k_\a \,dm(k,\mathfrak a)}$ is the disintegration w.r.t. the partition $\{Z^k_\a\}_{k,\mathfrak a}$, and the converse of Point \eqref{Point_4_T_1_h} holds:
\begin{equation*}
\pi^k_\a \in \Pi^f_{\mathtt c_{C^k_\a}}(\mu^k_\a,\nu^k_\a) \quad \Longrightarrow \quad \pi \in \Pi^{\mathrm{opt}}_{\d{\cdot}}(\mu,\nu).
\end{equation*}
\end{theorem1}
We start the proof by recalling that, by Proposition \ref{P_equivalence_lifting},
\begin{equation}
\label{E_pioptgraph}
\pi \in \Pi^{\mathrm{opt}}_{\d{\cdot}}(\mu,\nu) \quad \Longleftrightarrow \quad \hat \pi \in \Pi(\hat\mu,\hat\nu),\ \hat{\pi}(\partial^+\mathrm{graph}\,\psi)=1,
\end{equation}
where $\psi:\mathbb{R}^d\rightarrow\mathbb{R}$ is the $\d{\cdot}$-Lipschitz function given by a Kantorovich potential and $\hat\mu$, $\hat\nu$, $\hat\pi$ are the push-forwards of $\mu$, $\nu$, $\pi$ on $\mathbb{R}^{d+1}$ through the map $(\mathbb{I}\times\psi)$.
By Remark \ref{R_lipgraph}, $\mathrm{graph}\,\psi \subset \mathbb{R}^{d+1}$ is a \emph{complete $\mathtt c_{\mathrm{epi}\d{\cdot}}$-Lipschitz graph}, according to Definition \ref{D_complete_G}. Then, by Proposition \ref{P_ex_fol}, call $\theta_\psi$ the trivial $\mathtt c_{\tilde{\mathbf C}}$-Lipschitz foliation on $\mathbb{R}^{d+1}$ associated to $\mathrm{graph}\,\psi$.
We now show that Theorem \ref{T_1} follows from \eqref{E_pioptgraph} , thanks to the results of Sections \ref{S_foliations} and \ref{S_disintechnique} and the following theorem.
\begin{theorem}
\label{T_cone_graph}
Let $\mathrm{graph}\,\varphi\subset\mathbb{R}^{d+1}$ be a complete $\mathtt c_{\mathrm{epi}\d{\cdot}}$-Lipschitz graph. Then, the superdifferential partition satisfies the initial forward cone approximation property and the subdifferential partition satisfies the final backward approximation property.
\end{theorem}
Indeed, recalling Definition \ref{D_coneapprpart}, first notice that if Theorem \ref{T_cone_graph} holds, then by Theorem \ref{T_FC_no_initial}
\begin{equation}
\label{E_hd0}
\mathcal H^d(\mathcal I^+\theta_\varphi)= \mathcal H^d(\mathcal E^-\theta_\varphi) = 0.
\end{equation}
Moreover, by Remark \ref{R_infinslices} and the fact that $\hat{\mathbf D}=\hat{\mathbf D}^+\cap\hat{\mathbf D}^-$ (see \eqref{E_sub_directed_partition_all}), Theorem \ref{T_one_d_slicing_FC} implies that the disintegration of the $d$-dimensional Hausdorff measure on the differential partition $\hat{\mathbf D}$ of $\theta_\varphi$ is regular, namely
has conditional probabilities equivalent to the Hausdorff measures on the locally affine sets on which they are concentrated.
Therefore, denoting the locally affine partition $\hat{\mathbf D}$ as
\[
\big\{ \hat Z^{k}_{\mathfrak \mathfrak a},\hat C^k_\a \big\}_{\overset{k=1,\dots\,d}{\mathfrak a\in \mathfrak A^k}} \subset \mathtt P \bigg( \mathbb{R}^{d+1} \times \bigcup_{k=1}^d \mathcal C(k,\mathbb{R}^{d+1}) \bigg),
\]
and setting $\{\hat Z^0_\mathfrak a\}_{\mathfrak a\in\mathfrak A^0}$ for the $0$-dimensional partition of the fixed points $\mathfrak A^0=\mathcal F\theta_\varphi$,
by \eqref{E_graphpsimu} the sets
\[
\Big\{ Z^k_\mathfrak a = \mathtt p_{\mathbb{R}^d} \hat Z^k_\mathfrak a, C^k_\mathfrak a = \mathtt p_{\mathbb{R}^d} \hat C^k_\mathfrak a \Big\}_{\overset{k=0,\dots\,d}{\mathfrak a\in \mathfrak A^k}} \subset \mathtt P \bigg( \mathbb{R}^{d} \times \bigcup_{k=0}^d \mathcal C(k,\mathbb{R}^{d}) \bigg),
\]
define a locally affine directed partition of $\mathbb{R}^{d}$ satisfying $(1)$, $(2)$ and $(3)$.
Let us now use the fact that $\varphi=\psi$ is a Kantorovich potential for $\Pi^{\mathrm{opt}}_{\d{\cdot}}(\mu,\nu)$ and that $\mu\ll\mathcal L^d$. By \eqref{E_graphpsimu} and \eqref{E_hd0},
\begin{align}
& \mu \ll \mathcal L^d \quad \Longrightarrow \quad \hat \mu \ll \mathcal H^d \llcorner_{\mathrm{graph}\,\psi} \quad \Longrightarrow \quad \hat \mu(\mathcal I^+\theta_\psi) = \hat \mu(\mathcal E^-\theta_\psi) = 0, \label{E_mu0}
\end{align}
Then, by Remark \ref{R_4.27}, Proposition \ref{P_disint_fol} applies to the locally affine partition $\big\{ \hat Z^{k}_{\mathfrak \mathfrak a},\hat C^k_\a \big\}_{\overset{k=1,\dots\,d}{\mathfrak a\in \mathfrak A^k}}$, giving that any transport plan $\hat\pi$ as in \eqref{E_pioptgraph} satisfies $\hat\pi\in\Pi^f_{\mathtt c_{\hat{\mathbf D}}}(\hat\mu,\hat\nu)$. In particular, by Proposition \ref{P_dispiani}, the disintegration $\displaystyle{\hat\pi = \int \hat\pi^k_\a\, dm(k,\mathfrak a)}$ w.r.t. the partition $\{\hat Z^k_\a \times \mathbb{R}^{d+1}\}_{k,\mathfrak a}$ satisfies
\[
\hat\pi^k_\mathfrak a \in \Pi^f_{\mathtt c_{\hat C^k_\mathfrak a}}(\hat\mu^k_\a, (\mathtt p_2)_\# \hat\pi^k_\a),
\]
where $\displaystyle{\hat\mu = \int \hat\mu^k_\a\,dm(k,\mathfrak a)}$ is the disintegration w.r.t. the partition $\{\hat Z^k_\a\}_{k,\mathfrak a}$. Moreover, by Proposition \ref{P_hat_bf_D_graph}, the partition $\{\hat Z^k_\a\}_{k,\mathfrak a}$ satisfies condition \eqref{E_more_than_complet}, which gives
\[
(\mathtt p_2)_\#\hat\pi^k_\a \biggl( \hat Z^k_\a \cup \biggl( \mathrm{graph}\,\psi \setminus \underset{(k',\mathfrak a') \not= (k,\mathfrak a)}{\bigcup} \hat Z^{k'}_{\mathfrak a'} \biggr) \biggr) = 1.
\]
Then it is not difficult to see that the directed locally affine partition of $\mathbb{R}^d$ given by $\big\{ Z^{k}_{\mathfrak \mathfrak a},C^k_\a \big\}_{\overset{k=0,\dots\,d}{\mathfrak a\in \mathfrak A^k}}$ satisfies also Point $(4)$ of Theorem \ref{T_1}. Finally, if $\nu\ll\mathcal L^d$, by \eqref{E_hd0} and \eqref{E_graphpsimu} we have also
\begin{equation}
\nu \ll \mathcal L^d \quad \Longrightarrow \quad \hat \nu \ll \mathcal H^d \llcorner_{\mathrm{graph}\,\psi} \quad \Longrightarrow \quad \hat\nu(\mathcal E^-\theta_\psi) = 0.\label{E_nu0}
\end{equation}
Then $\hat\nu(\mathtt p_{\mathbb{R}^{d+1}}(\hat{\mathbf D}))=1$ and Corollary \ref{C_transp_graph} gives, when projected on $\mathbb{R}^d$, the last part of Theorem \ref{T_1}.
\begin{remark}
\label{E_more_nat_form}
Observe that the characterization given by Proposition \ref{P_disint_fol} of the optimal transport plans for the $\mathtt c_{\mathrm{epi}\d{\cdot}}$-Lipschitz set $\mathrm{graph}\,\psi$ seems more natural than the one given by Theorem \ref{T_1} for their projections on $\mathbb{R}^d$, namely the optimal transport plans for the original convex norm problem. Indeed, in the first case we have a complete (namely, if and only if) geometric characterization of the transport plans by disintegrations into transport plans of finite cone cost w.r.t. their conditional marginals, even in the case in which $\nu$ is not absolutely continuous. This is due to the geometric condition \eqref{E_more_than_complet}, which is satisfied by the partition $\{\hat Z^k_\a, \hat C^k_\a\}_{k,\mathfrak a}$ and not by its projection on $\mathbb{R}^d$.
In particular, there might be decompositions $\{\nu^k_\a\}$ of $\nu$ which are not obtained by projections of second marginals of disintegrations of $\hat\pi\in \Pi(\hat\mu,\hat\nu),\ \hat{\pi}(\partial^+\mathrm{graph}\,\psi)=1$ and such that $\Pi^f_{\mathtt c_{\mathtt p_{\mathbb{R}^d}\hat{\mathbf D}}}(\mu,\{\nu^k_\a\}) \not= \emptyset$.
\end{remark}
\begin{proof}[Proof of Theorem \ref{T_cone_graph}]
The proof will be given in two steps.
We prove the initial forward cone approximation property for the superdifferential partition of the forward regular set, being the proof of the final backward cone approximation property for the subdifferential partition analogous.
By Definition \ref{D_coneapprpart}, let us consider an initial $1$-dimensional slice
\[
\mathbf Z^{k,+} \cap \mathtt p^{-1}_{V^k} \bigl( (h^-, h^+) \mathtt e \bigr)=\underset{k,\mathfrak a}{\bigcup}\,Z^{k,+}_\mathfrak a\cap \mathtt p^{-1}_{V^k} \bigl( (h^-, h^+) \mathtt e \bigr)
\]
where
\begin{enumerate}
\item $V^k = \langle \mathtt e^k_1,\dots,\mathtt e^k_k\rangle\in \mathcal G(k,\mathbb{R}^{d+1})$ reference plane of the sheaf set $\mathbf Z^{k,+}$,
\item $\mathtt e \in \mathbb{S}^d \cap C(\{\mathtt e^k_i\})$,
\item there exists $\varepsilon > 0$ for which
\[
\mathbf Z^{k,+} \cap \mathtt p^{-1}_{V^k} \bigl( ( h^-, h^+ + \varepsilon)\mathtt e \bigr)
\]
is still a $1$-dimensional model set. Let $\mathtt d_\mathtt e$ be the direction vector field.
\end{enumerate}
{\it Step 1.} Assume $\mathtt d_\mathtt e^{h^-}$ is injective.
Then, it is sufficient to prove the forward cone approximation property for the vector field $\mathtt d_\mathtt e$ on a fixed perpendicular section, say e.g. $P_{h^-}$.
First recall that, by the general properties of sheaf sets, i.e. Point $(3)$ of Definition \ref{D_sheaf_set},
\[
\{0\} \cup \mathbb{R}^+ \mathtt e \subset C(\{\mathtt e^k_i\}) \subset \mathtt p_{V^k}(C^{k,+}_\mathfrak a), \quad \forall\,\mathfrak a\in\mathfrak{A}^k.
\]
By definition of $1$-dimensional slice, for all $w_{\mathfrak a,h^-} = (x_{\mathfrak a,h^-},\varphi(x_{\mathfrak a,h^-})) \in P_{h^-}$ one has
\begin{equation}
\label{E_tt_d_e_sigma}
\mathtt d_\mathtt e(w_{\mathfrak a,h^-}) = \frac{\sigma^{h^-,h^++\varepsilon}(w_{\mathfrak a,h^-})-w_{\mathfrak a,h^-}}{|\sigma^{h^-,h^++\varepsilon}(w_{\mathfrak a,h^-})-w_{\mathfrak a,h^-}|},
\end{equation}
being
\[
\sigma^{h^-,h^++\varepsilon}(w_{\mathfrak a,h^-}) = w_{\mathfrak a,h^{+}+\varepsilon} = \big( y_{\mathfrak a,h^{+}+\varepsilon},\varphi(y_{\mathfrak a,h^{+}+\varepsilon}) \big)
\]
the unique point s.t.
\[
P_{h^++\varepsilon} \cap \big( w_{\mathfrak a,h^{-}} + C^{k,+}_\mathfrak a \big) = \{ w_{\mathfrak a,h^{+}+\varepsilon} \}.
\]
Since now we are dealing with the superdifferential partition, (see Theorems \ref{T_partition_E+-} and \ref{T_partition_E}) for all $w_\mathfrak a \in Z^{k,+}_\mathfrak a$
\begin{equation*}
C^{k,+}_{\mathfrak a} = \mathbb{R}^+ \mathcal D^+{\theta}_{\varphi}(\mathfrak a,w_{\mathfrak a}) \quad \text{and} \quad \partial^+\theta_{\varphi}(\mathfrak a,w_{\mathfrak a}) \cap \mathcal R^{+,k} \theta_{\varphi} = \big( w_{\mathfrak a} + C^{k,+}_{\mathfrak a} \big) \cap Z^{k,+}_\mathfrak a.
\end{equation*}
Then we conclude that
\[
y_{\mathfrak a,h^{+}+\varepsilon}=\mathtt p_{\mathbb{R}^d}(w_{\mathfrak a,h^{+}+\varepsilon})
\]
is the unique point of $\mathtt p_{\mathbb{R}^d}(P_{h^++\varepsilon})$ s.t.
\[
\varphi(y_{\mathfrak a,h^{+}+\varepsilon}) - \varphi(x_{\mathfrak a,h^{-}}) = \big| y_{\mathfrak a,h^{+}+\varepsilon} - x_{\mathfrak a,h^{-}} \big|_{D^*},
\]
namely $y_{\mathfrak a,h^{+}+\varepsilon}$ is the unique maximizer of
\begin{equation}
\label{E_ya_unique}
\varphi(x_{\mathfrak a,h^-}) = \underset{y \in \mathtt p_{\mathbb{R}^d}(P_{h^++\varepsilon})} \max \Big\{ \varphi(y) - \d{y-x_{\mathfrak a,h^-}} \Big\}.
\end{equation}
Hence one can construct the finite cone approximations of $\mathtt p_{\mathbb{R}^d} \mathtt d_e$ as in \cite{Car1}, namely discretizing the set $\mathtt p_{\mathbb{R}^d}(P_{h^++\varepsilon})$ and taking the cones given by the differential partition of an optimal potential w.r.t. a strictly convex cost obtained by perturbating the norm cost $\d{\cdot}$ and whose second marginals are Dirac deltas centered at the points of the discretization (see \cite{Car1}). The convergence of the approximations to $\mathtt p_{\mathbb{R}^d}\mathtt d_e$ at a.e. point $x_{\mathfrak a,h^-}$ as the cost perturbation goes to $0$ and the points of the discretization become dense is given by the uniqueness of the $y_{\mathfrak a,h^{+}+\varepsilon}\in\mathtt p_{\mathbb{R}^d}(P_{h^++\varepsilon})$ satisfying \eqref{E_ya_unique}.
Lifting the approximating cones with the map $\mathbb{I}\times\varphi$, one gets finite cones approximations of $\mathtt d_e$ as required.
{\it Step 2.} Let now $\mathtt d_\mathtt e^{h^-}$ be possibly multivalued. In order to prove the initial forward cone approximation property, we build as in Step 1 finite cone approximations given by the differential partition of optimal potentials w.r.t. strictly convex approximating costs and second marginals given by Dirac deltas in $\mathtt p_{\mathbb{R}^d}(P_{h^++\varepsilon})$. These will converge to a Borel section $\tilde{\mathtt d}^{h^-}_{+,\mathtt e}$ of the direction vector field $\mathtt d_\mathtt e^{h^-}$ which by construction satisfies the cone approximation property.
\end{proof}
\begin{remark}
\label{R_part_case_Th7}
The above theorem can also be proved as a particular case of the analysis done in Section \ref{S_cone_approx_folia}: in this case we have a single cone-Lipschitz graph, and the uniqueness role of the linear order is trivial.
\end{remark}
\section{\texorpdfstring{From $\tilde{\mathbf C}^k$-fibrations to linearly ordered $\tilde{\mathbf C}^k$-Lipschitz foliations}{From C-fibrations to linearly ordered C-Lipschitz foliations}}
\label{S_cfibr_cfol}
This section is devoted to the proof of Theorem \ref{T_cfibrcfol} stated below, that will be the building block for proving of Theorem \ref{T_subpart_step}.
Let $\tilde{\mathbf C}^k : \mathbb{R}^{d-k} \times \mathbb{R}^k \supset \mathfrak{A} \times \mathbb{R}^k \rightarrow \mathcal C(k,\mathbb{R}^k)$ be the $\sigma$-compact direction map of a $k$-dimensional fibration $\tilde{\mathbf D}^k$ and $\mathtt c_{\tilde{\mathbf C}^k}$ be the associated cost function \eqref{E_cost_fibr}. Let
\[
\tilde \mu = \int \tilde \mu_\mathfrak a d\tilde m(\mathfrak a), \qquad \tilde \nu = \int \tilde \nu_\mathfrak a d\tilde m(\mathfrak a)
\]
be probability measures on $\mathbb{R}^d$
such that
\begin{equation}
\label{cfibrcfol1}
\Pi^f_{\mathtt c_{\tilde{\mathbf C}^k}}(\tilde\mu,\tilde\nu)\neq\emptyset
\end{equation}
and
\begin{equation}
\label{cfibrcfol2}
\tilde \mu_\mathfrak a \ll \mathcal H^k \llcorner_{\tilde Z^k_\a} \quad \text{ for $\tilde m$-a.e. $\mathfrak a \in \mathfrak{A}$.}
\end{equation}
Recall Definition \ref{D_compatible} of $(\mathtt c,\mu,\nu)$-compatible preorder, Definition \ref{D_pimunuconn} of $\Pi^f_{\mathtt c}(\mu,\nu)$-cyclically connected partition and let $\{0,1\}^\mathbb{N}$ be the Polish space of sequences in $\{0,1\}$ endowed with the product topology.
Our main result is the following theorem. Recall that $\omega$ is the first countable ordinal.
\begin{theorem}
\label{T_cfibrcfol}
If \eqref{cfibrcfol1} and \eqref{cfibrcfol2} hold, then there exists a $(\mathtt c_{\tilde{\mathbf C}^k},\tilde \mu,\tilde \nu)$-compatible linear preorder $\bar \preccurlyeq$ with Borel graph on $\mathfrak{A} \times \mathbb{R}^k$
\begin{equation}
\label{E_bar_theta_th}
\bar \preccurlyeq = (\bar \theta \times \bar \theta)^{-1}(\trianglelefteq_\omega), \qquad \bar \theta : \mathfrak{A} \times \mathbb{R}^{k} \rightarrow \mathfrak{A} \times \{0,1\}^\mathbb{N}, \ \trianglelefteq_\omega \ \text{linear order,}
\end{equation}
and equivalence classes $\{\bar\theta^{-1}(\mathfrak a,\mathfrak t)\}_{\nfrac{\mathfrak a \in \mathfrak{A}}{\mathfrak t \in \{0,1\}^\mathbb{N}}}$ such that the subcollection of sets $\{\bar Z^k_{\mathfrak a,\mathfrak t}\}_{\nfrac{\mathfrak a \in \mathfrak{A}}{\mathfrak t\in\mathfrak T^k(\mathfrak a)}}$ defined by
\begin{equation}
\label{E_subcollection}
\bar Z^k_{\mathfrak a,\mathfrak t} = \inter_{\mathrm{rel}}\,\bar\theta^{-1}(\mathfrak a,\mathfrak t) \quad \text{and} \quad \tilde\mu_\mathfrak a(\bar\theta^{-1}(\mathfrak a,\mathfrak t))>0
\end{equation}
is $\Pi^f_{\mathtt c_{\tilde{\mathbf C}^k}}(\tilde \mu,\tilde \nu)$-cyclically connected.
\end{theorem}
As noticed in Proposition \ref{P_ex_fol}, the equivalence classes of a $\mathtt c_{\tilde{\mathbf C}^k}$-compatible linear preorder on $\mathfrak{A}\times\mathbb{R}^k$ with $\sigma$-compact graph form a $\mathtt c_{\tilde{\mathbf C}^k}$-Lipschitz foliation.
Then, by definition of $(\mathtt c_{\tilde{\mathbf C}^k},\tilde \mu,\tilde \nu)$-compatible linear preorder and by disintegration of measures, Theorem \ref{T_cfibrcfol} claims that we can reduce the optimal transportation problem on a $\tilde{\mathbf C}^k$-fibration to a family of optimal transportation problems on the level sets of a $\mathtt c_{\tilde{\mathbf C}^k}$-Lipschitz foliation, whose $k$-dimensional classes of positive $\tilde\mu_\mathfrak a$ measure (see the characterization of $\mathtt c_{\tilde{\mathbf C}^k}$-Lipschitz foliations given in Proposition \ref{P_fol_char}) satisfy the cyclically connectedness property w.r.t. $\Pi^f_{\mathtt c_{\tilde{\mathbf C}^k}}(\tilde\mu,\tilde\nu)$.
As noticed in the Proposition \ref{P_ex_fol}, since
\[
\mathtt c_{\tilde{\mathbf C}^k}(\mathfrak a,w,\mathfrak a',w') < +\infty \quad \Rightarrow \quad \mathfrak a = \mathfrak a',
\]
the equivalence classes of the preorder $\bar \preccurlyeq$ constructed in Theorem \ref{T_cfibrcfol} will be contained in sections $\{\mathfrak a\}\times\mathbb{R}^k$. At a first reading, the geometry which lies behind the construction of $\bar\preccurlyeq$ will be clear to the reader even assuming that $\mathfrak{A}=\{\mathfrak a_0\}$ for some point $\mathfrak a_0$, and thus
$\mathtt c_{\tilde{\mathbf C}^k}$ is equal to a single convex cone cost. The variable $\mathfrak a\in\mathfrak{A}$ plays in fact the role of a parameter on which the maps used to define the preorder have to depend in a suitably measurable way.
As a preliminary, let us define the sets of $\sigma$-compact \emph{carriages} as follows: for $\tilde \pi \in \Pi^f_{\mathtt c_{\tilde{\mathbf C}^k}}(\tilde \mu,\tilde \nu)$ set
\begin{equation}
\label{E_varGamma_def_pi}
\varGamma(\tilde \pi) := \Big\{ \tilde \Gamma \subset (\mathfrak{A}\times\mathbb{R}^k) \times (\mathfrak{A}\times\mathbb{R}^k) : \tilde \Gamma \subset \{{\mathtt c_{\tilde{\mathbf C}^k}} < \infty\},\text{ $\tilde\Gamma$ $\sigma$-compact, }\tilde \pi(\tilde \Gamma) = 1 \Big\},
\end{equation}
and define
\begin{equation}
\label{E_varGamma_def}
\varGamma := \bigcup_{\tilde \pi \in \Pi^f_{\mathtt c_{\tilde{\mathbf C}^k}}(\tilde \mu,\tilde \nu)} \varGamma(\tilde \pi).
\end{equation}
The section of a carriage $\tilde \Gamma(\mathfrak a,\mathfrak a)$ will be also denoted as $\tilde\Gamma(\mathfrak a)\subset\mathbb{R}^k\times\mathbb{R}^k$.
\subsection{\texorpdfstring{Construction of a $(\mathtt c_{\tilde{\mathbf C}^k},\tilde\mu,\tilde\nu)$-compatible linear preorder}{Construction of a (c,m,n)-compatible linear preorder}}
\label{Ss_gamma_order}
The main result of this section, which is the first step of the proof of Theorem \ref{T_cfibrcfol}, is the following theorem.
\begin{theorem}
\label{T_order_gamma}
For any $\tilde \Gamma \in \varGamma$ there exists a $(\mathtt c_{\tilde{\mathbf C}^k},\tilde \mu,\tilde \nu)$-compatible linear preorder $\preccurlyeq_{\tilde\Gamma,\mathtt W^{\tilde{\Gamma}}}$ with Borel graph
\begin{equation}
\label{E_bar_theta_th_2}
\preccurlyeq_{\tilde\Gamma,\mathtt W^{\tilde{\Gamma}}} = \big( \theta_{\tilde\Gamma,\mathtt W^{\tilde{\Gamma}}} \times \theta_{\tilde\Gamma,\mathtt W^{\tilde{\Gamma}}} \big)^{-1}(\trianglelefteq_\omega), \qquad \theta_{\tilde\Gamma,\mathtt W^{\tilde{\Gamma}}} : \mathfrak{A} \times \mathbb{R}^{d-k} \rightarrow \mathfrak{A} \times \{0,1\}^\mathbb{N}, \ \trianglelefteq_\omega \ \text{linear order,}
\end{equation}
whose equivalence classes $\{\theta^{-1}_{\tilde{\Gamma},\mathtt W^{\tilde{\Gamma}}}(\mathfrak a,\mathfrak t)\}_{\nfrac{\mathfrak a \in \mathfrak{A}}{\mathfrak t\in\{0,1\}^\mathbb{N}}}$ satisfy
\begin{equation}
\label{E_leb_open}
\mathrm{Leb} \Big( \mathtt p_1(\tilde{\Gamma}(\mathfrak a)) \cap \{\theta_{\tilde{\Gamma},\mathtt W^{\tilde{\Gamma}}}(\mathfrak a,\cdot)=(\mathfrak a,\mathfrak t)\} \Big) \quad \text{ is $(\tilde{\Gamma}(\mathfrak a),\mathtt c_{\tilde{\mathbf C}^k(\mathfrak a)})$-cyclically connected}.
\end{equation}
\end{theorem}
By the characterization of $\mathtt c_{\tilde{\mathbf C}^k}$-Lipschitz foliations given in Proposition \ref{P_fol_char}, \eqref{E_leb_open} must refer to $k$-dimensional equivalence classes. Moreover, by Remark \ref{rem_pre}, $\preccurlyeq _{\tilde\Gamma\cap\cup\theta^{-1}(\mathfrak a,\mathfrak t)\times\theta^{-1}(\mathfrak a,\mathfrak t)}\subset\preccurlyeq_{\tilde{\Gamma},\mathtt W^{\tilde{\Gamma}}}$ and then $\forall\,x,y\in \mathrm{Leb} \big( \mathtt p_1(\tilde{\Gamma}(\mathfrak a)) \cap \{\theta_{\tilde{\Gamma},\mathtt W^{\tilde{\Gamma}}}(\mathfrak a,\cdot)=(\mathfrak a,\mathfrak t)\} \big)$ the $(\tilde\Gamma(\mathfrak a),\mathtt c_{\tilde{\mathbf C}^k(\mathfrak a)})$-cycle connecting $x$ to $y$ must be contained in $\theta^{-1}(\mathfrak a,\mathfrak t)\times\theta^{-1}(\mathfrak a,\mathfrak t)$.
The first step to prove Theorem \ref{T_order_gamma} is to select an $\tilde m$-conegligible set $\tilde{\mathfrak{A}}' \subset \mathfrak{A}$ and a $\sigma$-compact subset of
\[
\mathtt p_1 \tilde{\Gamma} \cap \tilde{\mathfrak{A}}' \times \mathbb{R}^k
\]
with $\mathfrak a$-sections countable and dense in $\mathtt p_1\tilde{\Gamma}(\mathfrak a)$.
\begin{lemma}
\label{L_z_n_dense_selections}
There exist an $\tilde m$-conegligible $\sigma$-compact set $\tilde{\mathfrak A}' \subset \mathfrak{A}\subset\mathbb{R}^{d-k}$ and a countable family $\mathtt W^{\tilde{\Gamma}}$ of $\sigma$-continuous functions $\mathtt w_n^{\tilde{\Gamma}}: \tilde{\mathfrak A}' \rightarrow \mathbb{R}^k$, $n \in \mathbb{N}$, such that for all $\mathfrak a \in \tilde{\mathfrak{A}}'$
\begin{equation}
\label{E_z_n_dense_sel}
\big\{ \mathtt w_n^{\tilde{\Gamma}}(\mathfrak a) \big\}_{n \in \mathbb{N}} \subset \mathtt p_1 \tilde \Gamma(\mathfrak a) \subset
\mathrm{clos}\,\{\mathtt w_n^{\tilde{\Gamma}}(\mathfrak a)\}_{n \in \mathbb{N}}.
\end{equation}
\end{lemma}
\begin{proof
For shortness we use the notation
\begin{equation*}
\Lambda := \mathtt p_1 \tilde \Gamma = \Big\{ (\mathfrak a,w) : \exists\, \mathfrak a', w' \text{ s.t. } (\mathfrak a,w,\mathfrak a',w') \in \tilde \Gamma \Big\} \subset \mathbb{R}^{d-k} \times \mathbb{R}^k.
\end{equation*}
{\it Step 1.} Let $\mathfrak Q := \mathtt p_\mathfrak a(\Lambda) \subset \mathbb{R}^{d-k}$ and fix $\varepsilon > 0$. By standard selection theorems (for example, Theorem 5.2.1 of \cite{Sri:courseborel} is sufficient in this setting), there exists $\mathtt w^\varepsilon_0 : \mathfrak Q \mapsto \mathbb{R}^k$ Borel such that $\mathrm{graph}\,\mathtt w^\varepsilon_0 \subset \Lambda$. By Lusin Theorem (134Yd of \cite{MR2462519}) we obtain an $\tilde m$-conegligible set $\mathfrak Q^\varepsilon_0$ such that $\mathtt w^\varepsilon_0 \llcorner_{\mathfrak Q^\varepsilon_0}$ is $\sigma$-continuous.
Define
\[
\Lambda^\varepsilon_0 := \Lambda \cap (\mathtt p_\mathfrak a)^{-1}(\mathfrak Q^\varepsilon_0), \quad (\Lambda^\varepsilon_1)' := \Lambda^\varepsilon_0 \setminus \big\{ (\mathfrak a,w) : \big| w - \mathtt w^\varepsilon_0(\mathfrak a) \big| < \varepsilon \big\}.
\]
These are clearly Borel sets.
Let $(\mathfrak Q^\varepsilon_1)' := \mathtt p_\mathfrak a((\Lambda^\varepsilon_1)')$, and define $(\mathfrak Q^\varepsilon_0)'' \subset \mathfrak Q^\varepsilon_0 \setminus (\mathfrak Q^\varepsilon_1)'$ as a $\sigma$-compact set with the same $\tilde m$-measure of $\mathfrak Q^\varepsilon_0 \setminus (\mathfrak Q^\varepsilon_1)'$.
{\it Step 2.} If the Borel set $(\Lambda^\varepsilon_n)' \subset \mathbb{R}^{d-k} \times \mathbb{R}^k$ and Souslin set $(\mathfrak Q^\varepsilon_n)' := \mathtt p_\mathfrak a((\Lambda^\varepsilon_n)') \subset \tilde{\mathfrak{A}}'$ are given, let $\mathtt w^\varepsilon_n : (\mathfrak Q^\varepsilon_n)' \rightarrow \mathbb{R}^{k}$ be a $\varTheta$-measurable selection s.t. $\mathrm{graph}\,\mathtt w^\varepsilon_n \subset (\Lambda^\varepsilon_n)' $, where $\varTheta$ is the $\sigma$-algebra generated by Souslin sets: its existence is guaranteed by Theorem 5.5.2 of \cite{Sri:courseborel}. As in Step 1, find an $\tilde m$-conegligible set $\mathfrak Q^\varepsilon_n \subset (\mathfrak Q^\varepsilon_n)'$ such that $\mathtt w^\varepsilon_n \llcorner_{\mathfrak Q^\varepsilon_n}$ is $\sigma$-continuous.
Define the Borel sets
\[
\Lambda^\varepsilon_{n+1} := (\Lambda^\varepsilon_n)' \cap (\mathtt p_\mathfrak a)^{-1}(\mathfrak Q^\varepsilon_n), \quad (\Lambda^\varepsilon_{n+1})' := \Lambda^\varepsilon_{n+1} \setminus \big\{ (\mathfrak a,w) : \big| w - \mathtt w^\varepsilon_n(\mathfrak a) \big| < \varepsilon \big\}.
\]
If $(\mathfrak Q^\varepsilon_{n+1})' := \mathtt p_\mathfrak a((\Lambda^\varepsilon_{n+1})')$, let $(\mathfrak Q^\varepsilon_n)'' \subset \mathfrak Q^\varepsilon_n \setminus (\mathfrak Q^\varepsilon_{n+1})'$ be a $\sigma$-compact set with the same measure of $\mathfrak Q^\varepsilon_n \setminus (\mathfrak Q^\varepsilon_{n+1})'$.
Extend also the $\sigma$-compact function $\mathtt w^\varepsilon_n$ to an $\tilde m$-conegligible set by
\[
\mathtt w^\varepsilon_n(\mathfrak a) :=
\begin{cases}
\mathtt w^\varepsilon_n(\mathfrak a) & \mathfrak a \in \mathfrak Q^\varepsilon_n, \crcr
\mathtt w^\varepsilon_{m}(\mathfrak a) & \mathfrak a \in (\mathfrak Q^\varepsilon_m)'',\quad m=0,\dots, n-1.
\end{cases}
\]
{\it Step 3.} By repeating the above procedure countably many times, we obtain a countable family of $\sigma$-continuous functions $\mathtt w^\varepsilon_n : \overset{\infty}{\underset{m=0}{\cup}} (\mathfrak Q^\varepsilon_m)'' \rightarrow \mathbb{R}^k$, $n \in \mathbb{N}_0$, such that
\[
\Lambda \cap (\mathtt p_\mathfrak a)^{-1} \bigg( \bigcup_{m=0}^\infty (\mathfrak Q^\varepsilon_m)'' \bigg) \subset \Big\{ (\mathfrak a,w) : \mathrm{dist} \big( w, \{ \mathtt w^\varepsilon_n(\mathfrak a)\}_{n \in \mathbb{N}_0} \big\} \big) < \varepsilon \Big\}.
\]
Taking a countable sequence $\varepsilon_i \searrow 0$ as $i \rightarrow \infty$, the functions $\{\mathtt w^{\varepsilon_i}_n\}_{i,n \in \mathbb{N}_0}$ satisfy clearly the statement when restricted to an $\tilde m$-conegligible $\sigma$-compact subset $\tilde{\mathfrak A}'$ of $\underset{i \in \mathbb{N}}{\cap} \underset{n \in \mathbb{N}_0}{\cup} (\mathfrak Q^{\varepsilon_i}_n)''$.
\end{proof}
Now we associate to each $(\mathfrak a,\mathtt w_n^{\tilde{\Gamma}}(\mathfrak a))$ the subset of $\{\mathfrak a\}\times\mathbb{R}^k$ of all the points $(\mathfrak a,w)$ s.t. $\exists$ an axial path of finite $\mathtt c_{\tilde{\mathbf C}^k}$-cost in $\tilde\Gamma$ going from $(\mathfrak a,\mathtt w_n^{\tilde{\Gamma}}(\mathfrak a))$ to $(\mathfrak a,w)$ (see Definition \ref{D_axpath}).
\noindent Define
\begin{equation}
\label{E_h_n}
H_{\tilde{\Gamma},n} := \bigg\{ (\mathfrak a,w) : \exists\, (\bar w,\bar w') \in \tilde{\Gamma} \ \text{s.t.} \ \mathtt c_{\tilde{\mathbf C}^k(\mathfrak a)}(w,\bar w') < \infty \text{ and } (\mathfrak a,\bar w) \preccurlyeq_{(\tilde{\Gamma},\mathtt c_{\tilde{\mathbf C}^k})} (\mathfrak a,\mathtt w^{\tilde{\Gamma}}_n(\mathfrak a)) \bigg\},
\end{equation}
where $\preccurlyeq_{(\tilde{\Gamma},\mathtt c_{\tilde{\mathbf C}^k})}$ is the $(\tilde{\Gamma}, \mathtt c_{\tilde{\mathbf C}^k})$-axial preorder relation defined in \eqref{E_axpreorder}. Notice that
\begin{equation}
\label{E_ordergamma}
H_{\tilde{\Gamma},n} \cap \mathtt p_1\tilde\Gamma=\bigg\{(\mathfrak a,w):\,(\mathfrak a,w)\preccurlyeq_{(\tilde{\Gamma},\mathtt c_{\tilde{\mathbf C}^k})}(\mathfrak a,\mathtt w^{\tilde{\Gamma}}_n(\mathfrak a))\bigg\}.
\end{equation}
Observe that, despite the notation, $H_{\tilde{\Gamma},n}$ does not depend only on $\tilde\Gamma$ but also on the sections $\{\mathtt w_n^{\tilde{\Gamma}}\}_{n\in\mathbb{N}}$ selected in Lemma \ref{L_z_n_dense_selections}.
\begin{figure}
\centering{\resizebox{12cm}{8cm}{\input{sudakovfibrationbis28.pdf_t}}}
\caption{The construction of the set $H_{\tilde{\Gamma},n}$.}
\label{Fi_fibration}
\end{figure}
\begin{proposition}
\label{P_Hn_sigma_cpt_compatibility}
The set $H_{\tilde{\Gamma},n}$ is $\sigma$-compact in $\tilde{\mathfrak{A}}' \times \mathbb{R}^k$ and the set $\tilde{\mathfrak A}:=\bigl\{\mathfrak a\in\tilde{\mathfrak{A}}':\,H_{\tilde{\Gamma},n}(\mathfrak a)\neq\emptyset\bigr\}$ is Borel. Moreover
\begin{equation}
\label{E_conecomp_hn}
w'\in H_{\tilde{\Gamma},n}(\mathfrak a) \quad \Longrightarrow \quad \bigl\{\mathtt c_{\tilde{\mathbf C}^k(\mathfrak a)}(\cdot,w')<+\infty\bigr\}\subset H_{\tilde{\Gamma},n}(\mathfrak a).
\end{equation}
\end{proposition}
\begin{proof
We prove the $\sigma$-compactness of $H_{\tilde{\Gamma},n}$, since \eqref{E_conecomp_hn} is clear from the definition \eqref{E_h_n}. Observing that
\[
\begin{split}
H_{\tilde{\Gamma},n} = \Big\{ (\mathfrak a,w) :&~ \exists\, I \in \mathbb{N}, \big\{ (w_i,w'_i) \big\}_{i = 1}^I \subset \tilde \Gamma(\mathfrak a) \text{ s.t. } \crcr
&~ w_1 = \mathtt w_n^{\tilde{\Gamma}}(\mathfrak a) \ \text{and} \ {\mathtt c}_{\tilde{\mathbf C}^k(\mathfrak a)}(w_{i+1},w'_{i}),\,{\mathtt c}_{\tilde{\mathbf C}^k(\mathfrak a)}(w,w'_I) < \infty \Big\}, \crcr
\end{split}
\]
write
\[
H_{\tilde{\Gamma},n} = \bigcup_{I \in \mathbb{N}} H_{\tilde{\Gamma},n}^I \qquad \text{where} \qquad H_{\tilde{\Gamma},n}^I = \mathtt p_{(\mathfrak a_{I+1},w_{I+1})}(\tilde H^I_{\tilde{\Gamma},n})
\]
and $\tilde H^I_{\tilde{\Gamma},n} \subset (\mathbb{R}^{d-k} \times \mathbb{R}^k)^{2I+1}$ is given by
\begin{align*}
\tilde H^I_{\tilde{\Gamma},n} :
=&~ \big\{ w_1 = \mathtt w_n^{\tilde{\Gamma}}(\mathfrak a_1) \big\} \cap \bigg[ \bigcap_{i=1}^I \big\{ (\mathfrak a_i,w_i,\mathfrak a'_i,w'_i) \in \tilde \Gamma \big\} \bigg] \cap \bigg[ \bigcap_{i=1}^{I} \big\{ {\mathtt c}_{\tilde{\mathbf C}^k}(\mathfrak a_{i+1},w_{i+1},\mathfrak a'_{i},w'_{i}) < \infty \big\} \bigg].
\end{align*}
Since $\mathfrak a \mapsto \tilde{\mathbf C}^k({\mathfrak a})$ is $\sigma$-continuous, it follows that the set $\{{\mathtt c}_{\tilde{\mathbf C}^k} < \infty\}$ is $\sigma$-compact in $(\mathbb{R}^{d-k} \times \mathbb{R}^k) \times (\mathbb{R}^{d-k} \times \mathbb{R}^k)$. Hence, being $\mathtt w_n^{\tilde{\Gamma}}$ $\sigma$-continuous and $\tilde \Gamma$ $\sigma$-compact, the set $\tilde H^I_{\tilde{\Gamma},n}$ is $\sigma$-compact, thus also $H^I_{\tilde{\Gamma},n}$, and finally $H_{\tilde{\Gamma},n}$ too.
\end{proof}
We are now ready to define the Borel linear preorder of Theorem \ref{T_order_gamma}. If $\mathtt W^{\tilde{\Gamma}} = \{\mathtt w_n^{\tilde{\Gamma}}\}_n\in\mathbb{N}$ is the countable family of sections constructed in Lemma \ref{L_z_n_dense_selections}, define the function
\begin{equation*}
\begin{array}{ccccc}
\theta_{\tilde{\Gamma},\mathtt W^{\tilde{\Gamma}}} &:& \tilde{\mathfrak A}' \times \mathbb{R}^k &\rightarrow& \tilde{\mathfrak A}' \times \{0,1\}^\mathbb{N} \crcr
&& (\mathfrak a,w) &\mapsto& \theta_{\tilde{\Gamma},\mathtt W^{\tilde{\Gamma}}}(\mathfrak a,w) := \big( \mathfrak a,\{\chi_{\mathbb{R}^k \setminus H_{\tilde{\Gamma},n}(\mathfrak a)}(w)\}_{n \in \mathbb{N}} \big)
\end{array}
\end{equation*}
Since each component $\mathtt p_i\circ\theta_{\tilde{\Gamma},\mathtt W^{\tilde{\Gamma}}}$ of $\theta_{\tilde{\Gamma},\mathtt W^{\tilde{\Gamma}}}$ is Borel, also $\theta_{\tilde{\Gamma},\mathtt W^{\tilde{\Gamma}}}$ is Borel in the product topology.
On the space $\mathbb{R}^{d-k} \times \{0,1\}^\alpha$, $\alpha$ ordinal number, let us consider the natural linear order given by the lexicographic order. Namely, for $\mathfrak a = (\mathfrak a_1,\dots,\mathfrak a_{d-k}) \in \mathbb{R}^{d-k}$ set
\begin{equation}
\label{E_lexico_on_R_d_k}
\mathfrak a <_{\mathrm{lexi}_{\mathbb{R}^{d-k}}} \mathfrak a' \quad \Longleftrightarrow \quad \exists \,i \in \{1,\dots,d-k\} \text{ s.t. } \forall j < i, \, \mathfrak a_j = \mathfrak a'_j \ \text{and} \ \mathfrak a_i < \mathfrak a'_i,
\end{equation}
and define
\begin{equation}
\label{E_lexico_on_Q_01}
\begin{split}
\big( \mathfrak a,\{s_\beta\}_{\beta < \alpha} \big) \vartriangleleft_\alpha \big( \mathfrak a',\{s'_\beta\}_{\beta < \alpha} \big) \ \Longleftrightarrow &\text{ either } \mathfrak a <_{\mathrm{lexi}_{\mathbb{R}^{d-k}}} \mathfrak a' \text{ or } \crcr
& \ \mathfrak a = \mathfrak a' \ \text{and} \ \exists \bar \beta < \alpha:\, s_\beta = s_\beta' \ \forall\, \beta < \bar \beta,\quad s_{\bar \beta} = 0 \text{ and } s'_{\bar \beta} = 1.
\end{split}
\end{equation}
Let then $\trianglelefteq_\omega$ be the lexicographic linear order on $\mathbb{R}^{d-k}\times\{0,1\}^\mathbb{N}$ and define the linear preorder on $\tilde{\mathfrak{A}}'\times\mathbb{R}^k$ as
\begin{equation}
\label{E_order_on_fibration_by_Z}
\preccurlyeq_{\tilde{\Gamma},\mathtt W^{\tilde{\Gamma}}} := \big( \theta_{\tilde{\Gamma},\mathtt W^{\tilde{\Gamma}}} \otimes \theta_{\tilde{\Gamma},\mathtt W^{\tilde{\Gamma}}} \big)^{-1}(\trianglelefteq_\omega).
\end{equation}
The induced equivalence relation on $\tilde{\mathfrak A}' \times \mathbb{R}^k$ is given by
\begin{equation*}
\preccurlyeq_{\tilde{\Gamma},\mathtt W^{\tilde{\Gamma}}} \cap \preccurlyeq^{-1}_{\tilde{\Gamma},\mathtt W^{\tilde{\Gamma}}} = \bigl\{ \theta^{-1}_{\tilde{\Gamma},\mathtt W^{\tilde{\Gamma}}}(\mathfrak a,\mathfrak t) \bigr\}_{(\mathfrak a,\mathfrak t) \in \tilde{\mathfrak{A}}' \times \{0,1\}^\mathbb{N}}.
\end{equation*}
\begin{figure}
\centering{\resizebox{14cm}{11cm}{\input{sudakovthetafunctbis.pdf_t}}}
\caption{The function $\theta_{\tilde{\Gamma},\mathtt W^{\tilde{\Gamma}}}$ and the cyclical connectedness of the Lebesgue points of $\mathtt p_1(\tilde \Gamma(\mathfrak a))$.}
\label{Fi_thetafunct}
\end{figure}
\begin{proof}[Proof of Theorem \ref{T_order_gamma}]
The proof is given in two steps.
{\it Step 1.} In this step we prove that the relation $\preccurlyeq_{\tilde{\Gamma},\mathtt W^{\tilde{\Gamma}}}$ defined in \eqref{E_order_on_fibration_by_Z} is a $(\mathtt c_{\tilde{\mathbf C}^k},\tilde \mu,\tilde\nu)$-compatible linear preorder with Borel graph. By Remark \ref{rem_pre}, this amounts to prove that $\preccurlyeq_{\tilde{\Gamma},\mathtt W^{\tilde{\Gamma}}}$ is Borel, $\mathtt c_{\tilde{\mathbf C}^k}$-compatible and \eqref{E_tA21} holds for the carriage $\tilde\Gamma$.
First of all, $\preccurlyeq_{\tilde{\Gamma},\mathtt W^{\tilde{\Gamma}}}$ is Borel because it is the preimage under $\theta_{\tilde{\Gamma},\mathtt W^{\tilde{\Gamma}}}$, which is a Borel map, of the lexicographic order $\trianglelefteq_\omega$.
Moreover
\begin{equation*}
\begin{split}
\mathtt c_{\tilde{\mathbf C}^k(\mathfrak a)}(w,w') < +\infty \quad \overset{\eqref{E_conecomp_hn}}{\Longrightarrow} \quad &\text{if $w\in \mathbb{R}^k \setminus H_{\tilde{\Gamma},n}(\mathfrak a)$, then $w' \in \mathbb{R}^k \setminus H_{\tilde{\Gamma},n}(\mathfrak a)$} \crcr
\quad \Longrightarrow \, \quad &\theta_{\tilde{\Gamma},\mathtt W^{\tilde{\Gamma}}}(\mathfrak a,w) \trianglelefteq_\omega \theta_{\tilde{\Gamma},\mathtt W^{\tilde{\Gamma}}}(\mathfrak a,w'),
\end{split}
\end{equation*}
i.e. $\preccurlyeq_{\tilde{\Gamma},\mathtt W^{\tilde{\Gamma}}}$ is $\mathtt c_{\tilde{\mathbf C}^k}$-compatible. Formula \eqref{E_tA21} follows directly from \eqref{E_ordergamma}.
{\it Step 2.} Now we prove \eqref{E_leb_open}. Let
\[
w,w' \in \mathrm{Leb} \Big( \mathtt p_1(\tilde \Gamma(\mathfrak a)) \cap \big\{ \theta_{\tilde{\Gamma},\mathtt W^{\tilde{\Gamma}}}(\mathfrak a,\cdot) = (\mathfrak a, \mathfrak t) \big\} \Big).
\]
Since $\preccurlyeq_{\tilde{\Gamma},\mathtt W^{\tilde{\Gamma}}}$ is $\mathtt c_{\tilde{\mathbf C}^k}$-compatible, by Proposition \ref{P_ex_fol} its equivalence classes form a $\mathtt c_{\tilde{\mathbf C}^k}$-Lipschitz foliation and then, from Point \eqref{Prop_3_fol_char} of Proposition \ref{P_fol_char}, there exists $r>0$ such that
\[
B^k(w,r), B^k(w',r) \subset \big\{ \theta_{\tilde{\Gamma},\mathtt W^{\tilde{\Gamma}}}(\mathfrak a,\cdot) = (\mathfrak a,\mathfrak t) \big\}.
\]
Hence, by the density of $\{\mathtt w_n^{\tilde{\Gamma}}(\mathfrak a)\}_{n\in\mathbb{N}}$ stated in Lemma \ref{L_z_n_dense_selections}, there exist $\mathtt w_{\bar n}^{\tilde{\Gamma}}$, $\mathtt w_{\bar n'}^{\tilde{\Gamma}}$ such that
\begin{equation}
\label{E_cycle_contrct}
\theta_{\tilde{\Gamma},\mathtt W^{\tilde{\Gamma}}}(\mathfrak a,\mathtt w_{\bar n}^{\tilde{\Gamma}}(\mathfrak a)) = \theta_{\tilde{\Gamma},\mathtt W^{\tilde{\Gamma}}}(\mathfrak a,\mathtt w_{\bar n'}^{\tilde{\Gamma}}(\mathfrak a)) = (\mathfrak a, \mathfrak t) \qquad \text{and} \qquad {\mathtt c}_{\tilde{\mathbf C}^k(\mathfrak a)}(w,\mathtt w_{\bar n}^{\tilde{\Gamma}}(\mathfrak a)), {\mathtt c}_{\tilde{\mathbf C}^k(\mathfrak a)}(\mathtt w^{\tilde{\Gamma}}_{\bar n'}(\mathfrak a),w') < \infty.
\end{equation}
The first condition in \eqref{E_cycle_contrct} implies that
\[
\mathtt w_{\bar n}^{\tilde{\Gamma}}(\mathfrak a) \in H_{\tilde{\Gamma},n'}(\mathfrak a), \qquad \text{i.e.} \qquad \mathtt w_{\bar n}^{\tilde{\Gamma}}(\mathfrak a) \preccurlyeq_{(\tilde{\Gamma},\mathtt c_{\tilde{\mathbf C}^k(\mathfrak a)})} \mathtt w_{\bar n'}^{\tilde{\Gamma}}(\mathfrak a).
\]
The second condition implies that
\[
w \preccurlyeq_{(\tilde{\Gamma},\mathtt c_{\tilde{\mathbf C}^k})} \mathtt w_{\bar n}^{\tilde{\Gamma}}(\mathfrak a) \qquad \text{and} \qquad \mathtt w_{\bar n'}^{\tilde{\Gamma}}(\mathfrak a)\preccurlyeq_{(\tilde{\Gamma},\mathtt c_{\tilde{\mathbf C}^k(\mathfrak a)})}w'.
\]
Hence, composing the three axial paths, $w\preccurlyeq w$, and exchanging their roles we obtain a $(\tilde{\Gamma},\mathtt c_{\tilde{\mathbf C}^k(\mathfrak a)})$-cycle, thus concluding the proof.
\end{proof}
Notice that actually the subset of a $k$-dimensional class $\{\theta_{\tilde{\Gamma},\mathtt W^{\tilde{\Gamma}}}(\mathfrak a,\cdot)=(\mathfrak a,\mathfrak t)\}$ which is contained in the $(\Gamma,\mathtt c_{\tilde{\mathbf C}^k(\mathfrak a)})$-cycle above is
\[
\begin{split}
\Bigl\{ w \in \mathtt p_1\tilde{\Gamma}(\mathfrak a) \cap \{\theta_{\tilde{\Gamma},\mathtt W^{\tilde{\Gamma}}}(\mathfrak a,\cdot)=(\mathfrak a,\mathfrak t)\} :&~ \big( w - \inter_{\mathrm{rel}} \tilde{\mathbf C}^k(\mathfrak a) \big) \cap \mathtt p_1\tilde{\Gamma}(\mathfrak a) \cap \{\theta_{\tilde{\Gamma},\mathtt W^{\tilde{\Gamma}}}(\mathfrak a,\cdot)=(\mathfrak a,\mathfrak t)\}\neq \emptyset \crcr
&~ \text{and} \ \big( w + \inter_{\mathrm{rel}} \tilde{\mathbf C}^k(\mathfrak a) \big) \cap \mathtt p_1\tilde{\Gamma}(\mathfrak a) \cap \{\theta_{\tilde{\Gamma},\mathtt W^{\tilde{\Gamma}}}(\mathfrak a,\cdot)=(\mathfrak a,\mathfrak t)\}\neq \emptyset \Bigr\}.
\end{split}
\]
In fact, by \eqref{E_z_n_dense_sel} every couple of points $w$, $w'$ in the above set satisfy \eqref{E_cycle_contrct} for some $\mathtt w_{\bar n}^{\tilde{\Gamma}}$, $\mathtt w_{\bar n'}^{\tilde{\Gamma}}$.
\subsection{\texorpdfstring{Minimal $(\mathtt c_{\tilde{\mathbf C}^k},\tilde\mu,\tilde\nu)$-compatible linear preorder}{Minimal (c,m,n)-compatible linear preorder}}
\label{Ss_minial_order}
Now we apply Theorem \ref{T_minimal_equival} to the class of linear preorders $\{\theta_{\tilde{\Gamma},\mathtt W^{\tilde{\Gamma}}}\}_{\tilde{\Gamma},\mathtt W^{\tilde{\Gamma}}}$ constructed in Theorem \ref{T_order_gamma}, in order to find a Borel $(\mathtt c_{\tilde{\mathbf C}^k},\tilde\mu,\tilde\nu)$-compatible linear preorder as in Theorem \ref{T_cfibrcfol}.
Recall the definition of lexicographic order $\trianglelefteq_\alpha$ on $\mathbb{R}^{d-k}\times\{0,1\}^\alpha$, $\alpha$ ordinal number, given in \eqref{E_lexico_on_Q_01}, and recall also the definition of closure under countable intersection of a family of equivalence relations, Definition \ref{D_clos_sigma_E} and Remark \ref{R_min_contain_E}.
\begin{proposition}
\label{P_equiv_coun}
The class of equivalence relations
\begin{align}
\label{E_equivcount}
\Big\{ \preccurlyeq_{\tilde{\Gamma},\mathtt W^{\tilde{\Gamma}}}\cap\preccurlyeq_{\tilde{\Gamma},\mathtt W^{\tilde{\Gamma}}}^{-1}:\, \preccurlyeq_{\tilde{\Gamma},\mathtt W^{\tilde{\Gamma}}} \text{ as in Theorem \ref{T_order_gamma} for some $\tilde{\Gamma}\in\varGamma$} \Big\}
\end{align}
is closed under countable intersections.
\end{proposition}
\begin{proof}
Let $\{\tilde{\Gamma}_n\}_{n\in\mathbb{N}}\subset\varGamma$ and $\big\{\preccurlyeq_{\tilde{\Gamma}_n,\mathtt W^{\tilde{\Gamma}_n}}=(\theta_{\tilde{\Gamma}_n,\mathtt W^{\tilde{\Gamma}_n}}\times \theta_{\tilde{\Gamma}_n,\mathtt W^{\tilde{\Gamma}_n}})^{-1}(\trianglelefteq_\omega)\big\}_{n\in\mathbb{N}}$ be a countable family of $(\mathtt c_{\tilde{\mathbf C}^k}, \tilde\mu,\tilde\nu)$-compatible Borel linear preorders as in Theorem \ref{T_order_gamma}. Fix $\tilde{\Gamma}_i\subset \varGamma(\tilde\pi_i)$ for some $i\in\mathbb{N}$. By Definition \ref{D_cmunucomp} of $(\mathtt c_{\tilde{\mathbf C}^k}, \tilde\mu,\tilde\nu)$-compatibility,
\[
\tilde\pi_i\bigg(\tilde{\Gamma}_i\underset{n\in\mathbb{N}}{\bigcap}\preccurlyeq_{\tilde{\Gamma}_n,\mathtt W^{\tilde{\Gamma}_n}}\cap(\preccurlyeq_{\tilde{\Gamma}_n,\mathtt W^{\tilde{\Gamma}_n}})^{-1}\bigg)=1
\]
and by Remark \ref{rem_pre}
\begin{equation}
\label{E_gammaicomp}
\preccurlyeq_{\tilde{\Gamma}_i\underset{n\in\mathbb{N}}{\bigcap}\preccurlyeq_{\tilde{\Gamma}_n,\mathtt W^{\tilde{\Gamma}_n}}\cap(\preccurlyeq_{\tilde{\Gamma}_n,\mathtt W^{\tilde{\Gamma}_n}})^{-1},\mathtt c_{\tilde{\mathbf C}^k}}\subset \underset{n\in\mathbb{N}}{\bigcap}\preccurlyeq_{\tilde{\Gamma}_n,\mathtt W^{\tilde{\Gamma}_n}}.
\end{equation}
Let then $\bar{\mathtt W}^{\tilde{\Gamma}_i}$ be a countable family of $\sigma$-compact sections of $\mathtt p_1\big(\tilde{\Gamma}_i\underset{n\in\mathbb{N}}{\bigcap}\preccurlyeq_{\tilde{\Gamma}_n,\mathtt W^{\tilde{\Gamma}_n}}\cap(\preccurlyeq_{\tilde{\Gamma}_n,\mathtt W^{\tilde{\Gamma}_n}})^{-1}\big)$ as in Lemma \ref{L_z_n_dense_selections}.
Then, by \eqref{E_gammaicomp}, it follows immediately that the $(\mathtt c_{\tilde{\mathbf C}^k}, \tilde\mu,\tilde\nu)$-compatible Borel linear preorder $\preccurlyeq_{\tilde{\Gamma}_i\underset{n\in\mathbb{N}}{\cap}\preccurlyeq_{\tilde{\Gamma}_n,\mathtt W^{\tilde{\Gamma}_n}}\cap(\preccurlyeq_{\tilde{\Gamma}_n,\mathtt W^{\tilde{\Gamma}_n}})^{-1},\bar{\mathtt W}^{\tilde{\Gamma}_i}}$ constructed as in the proof of Theorem \ref{T_order_gamma} satisfies
\[
\preccurlyeq_{\tilde{\Gamma}_i\underset{n\in\mathbb{N}}{\cap}\preccurlyeq_{\tilde{\Gamma}_n,\mathtt W^{\tilde{\Gamma}_n}}\cap(\preccurlyeq_{\tilde{\Gamma}_n,\mathtt W^{\tilde{\Gamma}_n}})^{-1},\bar{\mathtt W}^{\tilde{\Gamma}_i}}\subset\underset{n\in\mathbb{N}}{\bigcap}\preccurlyeq_{\tilde{\Gamma}_n,\mathtt W^{\tilde{\Gamma}_n}}.
\]
\end{proof}
We now have all the tools to prove Theorem \ref{T_cfibrcfol}.
\begin{proof}[Proof of Theorem \ref{T_cfibrcfol}]
Let $\bar \preccurlyeq\cap\bar \preccurlyeq^{-1}=\cup_{\mathfrak a,\mathfrak t'} \big\{ \bar\theta^{-1}(\mathfrak a,\mathfrak t') \big\} \times \big\{ \bar\theta^{-1}(\mathfrak a,\mathfrak t') \big\}$ be the minimal equivalence relation in the class \eqref{E_equivcount} w.r.t. the measure $\tilde \mu$, whose existence is guaranteed by Theorem \ref{T_minimal_equival}.
We claim that it satisfies the conclusions of Theorem \ref{T_cfibrcfol}. Thanks to Theorem \ref{T_order_gamma}, we only have to prove \eqref{E_subcollection}.
Recalling Definitions \ref{D_pimunuconn} and \ref{D_gammaconn}, let $\tilde \pi \in \Pi^f_{\mathtt c_{\tilde{\mathbf C}^k}}(\tilde\mu,\tilde\nu)$, $\tilde{\Gamma} \in \varGamma(\tilde \pi)$. By Remark \ref{rem_pre}, we can consider the carriage
\begin{equation}
\label{E_check_Gamma'}
\check\Gamma := \tilde\Gamma\cap \bigcup_{\mathfrak a,\mathfrak t'} \{ \bar\theta^{-1}(\mathfrak a,\mathfrak t') \} \times \{ \bar\theta^{-1}(\mathfrak a,\mathfrak t') \}\subset\tilde\Gamma
\end{equation}
and prove that the subcollection $\{\bar Z^k_{\mathfrak a,\mathfrak t}\}_{\nfrac{\mathfrak a \in \mathfrak{A}}{\mathfrak t\in\mathfrak T^k(\mathfrak a)}}$ of the equivalence classes of $\bar\preccurlyeq$ defined by
\begin{equation}
\label{E_subcollection2}
\bar Z^k_{\mathfrak a,t} = \inter_{\mathrm{rel}}\,\bar\theta^{-1}(\mathfrak a,\mathfrak t) \quad \text{and} \quad \tilde\mu_\mathfrak a(\bar\theta^{-1}(\mathfrak a,\mathfrak t))>0
\end{equation}
is $(\tilde\mu,\check\Gamma,\mathtt c_{\tilde{\mathbf C}^k})$-cyclically connected.
Let $\mathtt W^{\check\Gamma}$ be a countable family of $\sigma$-compact sections of $\mathtt p_1\check\Gamma$ as in Lemma \ref{L_z_n_dense_selections}. Hence, by \eqref{E_check_Gamma'}, reasoning as in the proof of Proposition \ref{P_equiv_coun}, the equivalence classes of $\preccurlyeq_{\check\Gamma,{\mathtt W}^{\check\Gamma}}$ are contained in those of $\bar \preccurlyeq$. By minimality of $\{\bar \theta^{-1}(\mathfrak a,\mathfrak t)\}$ and Corollary \ref{C_constant_for_minimal_equivalence}, there exists a $\tilde \mu$-conegligible $\sigma$-compact set $\check B \subset \mathbb{R}^{d-k} \times \mathbb{R}^k$ and a Borel function
\[
\mathtt s : \mathbb{R}^{d-k} \times \{0,1\}^{\mathbb{N}} \rightarrow \mathbb{R}^{d-k} \times \{0,1\}^\mathbb{N}
\]
such that
\[
\theta_{\check \Gamma,{\mathtt W^{\check \Gamma}}} = \mathtt s \circ \bar \theta \qquad \text{ on } \check B.
\]
The set $\check B$ depends on $\theta_{\check \Gamma,{\mathtt W^{\check\Gamma}}}$.
In particular, using this result for the equivalence classes of positive $\tilde \mu_\mathfrak a$-measure of $\bar\theta$, we deduce that there exists a set
\[
\mathfrak{A}'' \subset \Big\{ \mathfrak a : \exists\, \mathfrak t \ \text{s.t.}\ \tilde \mu_\mathfrak a \big(\bar \theta^{-1}(\mathfrak a,\mathfrak t) \big) > 0 \Big\}
\]
such that
\[
m \Big( \Big\{ \mathfrak a : \exists\, \mathfrak t \ \text{s.t.}\ \tilde \mu_\mathfrak a \big(\bar \theta^{-1}(\mathfrak a,\mathfrak t) \big) > 0 \Big\} \setminus \mathfrak{A}'' \Big) = 0
\]
and for all $\mathfrak a \in \mathfrak{A}''$, for all $\mathfrak t$ such that $\tilde \mu_\mathfrak a(\bar \theta^{-1}(\mathfrak a,\mathfrak t)) > 0$, the function $\theta_{\check\Gamma,\mathtt W^{\check\Gamma}}$ is $\tilde\mu_\mathfrak a$-a.e. constant on $\{\bar \theta^{-1}(\mathfrak a,\mathfrak t)\}$.
Hence, using the assumption \eqref{cfibrcfol2} and condition \eqref{E_leb_open} for $\preccurlyeq_{\check\Gamma,{\mathtt W}^{\check\Gamma}}$, the sets $\{\bar \theta^{-1}(\mathfrak a,\mathfrak t)\}$ with $\mathfrak a \in \mathfrak{A}''$ and of positive $\tilde \mu_\mathfrak a$ measure are open (see Proposition \ref{P_fol_char}) and their set of Lebesgue points is of full $\tilde \mu_\mathfrak a$-measure and $(\check\Gamma,\mathtt c_{\tilde{\mathbf C}(\mathfrak a)})$-cyclically connected.
By Definition \ref{D_gammaconn}, we thus conclude that these sets are $(\tilde \mu,\tilde{\Gamma},\mathtt c_{\tilde{\mathbf C}})$-cyclically connected, and then applying the same reasoning to any transference plan $\tilde \pi$, we get that the subcollection of sets $\{\bar Z^k_{\mathfrak a,\mathfrak t}\}_{\nfrac{\mathfrak a \in \mathfrak{A}}{\mathfrak t\in\mathfrak T^k(\mathfrak a)}}$ defined by
\begin{equation*}
\bar Z^k_{\mathfrak a,t} = \mathrm{int} \bar Z^k_{\mathfrak a,\mathfrak t}, \quad \text{with} \ \tilde\mu_\mathfrak a(\bar\theta^{-1}(\mathfrak a,\mathfrak t)) > 0,
\end{equation*}
is $\Pi^f_{\mathtt c_{\tilde{\mathbf C}^k}}(\tilde \mu,\tilde \nu)$-cyclically connected, thus concluding the proof.
\end{proof}
\begin{remark}
\label{R_why_this_meth}
We observe here that, being the equivalence relation $\bar \preccurlyeq \cap (\bar \preccurlyeq)^{-1}$ constructed in Theorem \ref{T_cfibrcfol} minimal in the family \eqref{E_equivcount}, by Theorem \ref{T_minimal_equival} the equivalence classes cannot be further decomposed by equivalence relations of the form \eqref{E_equivcount}. However, the information on cyclical connectedness of the equivalence classes can be deduced only for the equivalence classes with positive $\tilde \mu_a$-measure, because of the particular choice of the family \eqref{E_equivcount} satisfying Theorem \ref{T_order_gamma}.
\end{remark}
\section{\texorpdfstring{Cone approximation property for linearly ordered $\mathtt c_{\tilde{\mathbf C}}$-Lipschitz foliations}{Cone approximation property for linearly ordered C-Lipschitz foliations}}
\label{S_cone_approx_folia}
In this section we prove the following result.
\begin{theorem}
\label{T_coneappr_fol}
Let $\{\theta^{-1}(\mathfrak a,\mathfrak t)\}_{\mathfrak a,\mathfrak t} \subset \mathtt P(\mathfrak{A} \times \mathbb{R}^k)$, $\mathfrak{A} \subset \mathbb{R}^{d-k}$, be a $\mathtt c_{\tilde{\mathbf C}}$-Lipschitz foliation given by the equivalence classes of a Borel ${\mathtt c_{\tilde{\mathbf C}}}$-compatible linear preorder $\preccurlyeq$ on $\mathbb{R}^{d-k}\times\mathbb{R}^k$ as in \eqref{E_bar_theta_th}.
Then
for all $\mathfrak a \in \mathfrak{A}$, the subdifferential partition of $\mathcal R^{+}\theta(\mathfrak a)$ satisfies the initial forward cone approximation property and the subdifferential partition of $\mathcal R^-\theta(\mathfrak a)$ satisfies the final backward cone approximation property.
\end{theorem}
In particular, we conclude from Theorem \ref{T_FC_no_initial} that the initial and final points $\mathcal I^+\theta(\mathfrak a)$, $\mathcal E^-\theta(\mathfrak a)$ are $\mathcal H^k$-negligible and, by integration w.r.t. $\mathcal H^{d-k}$ on $\mathfrak{A}$, the sets $\mathcal I^+\theta$, $\mathcal E^-\theta$ are also $\mathcal L^d$-negligible. Moreover, by Theorem \ref{T_one_d_slicing_FC}, the disintegration of $\mathcal L^d$ w.r.t. the differential partition of the regular set $\mathcal R\theta$ is \emph{regular}, i.e. it satisfies \eqref{E_disint_regular} of Definition \ref{D_disint_regular}.
Since the disintegration of $\mathcal H^k$ on equivalence classes of $\theta$ with positive $\mathcal H^k$-measure is clearly regular, we thus have the following corollary.
\begin{corollary}
\label{C_infinnegl}
If $\{Z^\ell_{\mathfrak a,\mathfrak b}\}_{\underset{\mathfrak a\in\mathfrak{A},\mathfrak b\in\mathfrak B}{\ell=1,\dots,k}}$ is the partition of a $\mathtt c_{\tilde{\mathbf C}}$-Lipschitz foliation given by the equivalence classes of a Borel ${\mathtt c_{\tilde{\mathbf C}}}$-compatible linear preorder obtained as the union of the differential partition and of the classes of positive $\mathcal H^k$-measure, then the disintegration of the Lebesgue measure $\mathcal L^d$ restricted on the $\mathtt c_{\tilde{\mathbf C}}$-Lipschitz foliation
\[
\mathcal L^d \llcorner_{\mathbf Z} = \int \upsilon^\ell_{\mathfrak a,\mathfrak b} d\eta(\ell,\mathfrak a,\mathfrak b), \qquad \mathbf Z = \bigcup_{\ell,\mathfrak a,\mathfrak b} Z^\ell_{\mathfrak a,\mathfrak b},
\]
satisfies
\[
\upsilon^\ell_{\mathfrak a,\mathfrak b} \simeq \mathcal H^\ell \llcorner_{Z^\ell_{\mathfrak a,\mathfrak b}}, \quad\text{ for $\eta$-a.e. $(\ell,\mathfrak a,\mathfrak b)$}.
\]
\end{corollary}
\begin{remark}
\label{R_contin_quotient}
If the quotient set $\{1,\dots,k\} \times \mathfrak{A} \times \mathfrak B$ is chosen to be a countable union of sets as in \eqref{E_mathfrak_A_k_def}, then the quotient measure
\[
\eta = \sum_{\ell = 1}^k \eta^\ell,\quad\eta^\ell(\{\ell\}\times\mathfrak{A}\times\mathfrak B)=1
\]
satisfies
\[
\eta^\ell \simeq \mathcal H^{d-\ell} \llcorner_{\underset{i\in\mathbb{N}}{\cup} \mathfrak C^\ell_i}
\]
for some $\mathfrak C^\ell_i \subset V^{d-\ell}_i \in \mathcal A(d-\ell,\mathbb{R}^d)$.
\end{remark}
\begin{proof
In the following we identify $\{\mathfrak a\}\times\mathbb{R}^k$ with $\mathbb{R}^k$ and omit the variable $\mathfrak a$ when clear from the context. Unless explicitly stated, for the notions and notations used in the proof we refer to Section \ref{S_disintechnique}.
Since the proof of the initial forward cone approximation property is the same as the forward cone approximation property up to the Borel selection given by Lemma \ref{L_reg_tt_d_h_pm}, for simplicity we prove the forward cone approximation property.
The proof will be given in three steps, and we will restrict to the case of $\ell < k$, due to the structure of equivalence classes of positive $\mathcal H^k$-measure given in Proposition \ref{P_fol_char} and the existence of at most two degenerate equivalence classes (see Definition \ref{D_non_dege}), which are clearly of positive measure, as observed in Remark \ref{R_nondeg}.
{\it Step 1.} By Definitions \ref{D_coneapprpart} and \ref{D_1_dim_slice_sheaf}, we have to prove the forward cone approximation property for the $1$-dimensional slices $\mathcal D$-cylinder (see Definition \ref{D_mathcal_D_cyl}) of the superdifferential partition of $\mathcal R^{+}\theta(\mathfrak a)$ ($\ell\in\{1,\dots,k-1\}$) given by
\[
\bigl\{ Z^{\ell,+}_{\mathfrak b}, C^{\ell,+}_{\mathfrak b} \bigr\}_{\mathfrak b \in \mathfrak B_{\ell,+}(\mathfrak a)}, \qquad \mathbf Z^{\ell,+} = \bigcup_{\mathfrak b \in \mathfrak B_{\ell,+}(\mathfrak a)} Z^{\ell,+}_{\mathfrak b} \subset \mathbb{R}^k,
\]
with reference plane $V^\ell = \langle \mathtt e^\ell_1,\dots,\mathtt e^\ell_\ell \rangle \in \mathcal G(\ell,\mathbb{R}^k)$ and base rectangle $U(\{\mathtt e^\ell_i\},\mathrm h^-,\mathrm h^+)$. By \eqref{E_mathfrak_A_k_def}, the set $\mathfrak B_{\ell,+}(\mathfrak a)$ is a subset of $(z + (V^\ell)^\perp) \in \mathcal A(k-l,\mathbb{R}^k)$ for some $z \in \inter_{\mathrm{rel}} U(\{\mathtt e^\ell_i\},\mathrm h^-,\mathrm h^+)$.
Let us fix a $1$-dimensional slice of $\mathbf Z^{\ell,+}$ with reference configuration $(\mathtt e,w + h^-(w,\mathtt e)\mathtt e,w+h^+(w,\mathtt e)\mathtt e)$, $w \in \inter_{\mathrm{rel}} U(\{\mathtt e^\ell_i\},\mathrm h^-,\mathrm h^+)$, $\mathtt e\in C(\{\mathtt e^\ell_i\})$ (see Definition \ref{D_1_sim_slice}), i.e.
\begin{equation}
\label{E_1_dim_superdif}
\mathbf Z^{\ell,+}_\mathtt e = \mathbf Z^{\ell,+} \cap \mathtt p^{-1}_{ V^\ell} \Big( w + \big( h^-(w,\mathtt e),h^+(w,\mathtt e) \big) \mathtt e \Big)
\end{equation}
with $\varepsilon>0$ such that the set
\[
\mathbf Z^{\ell,+} \cap \mathtt p^{-1}_{V^\ell} \Big( w + \big( h^-(w,\mathtt e)-\varepsilon,h^+(w,\mathtt e)+\varepsilon \big) \mathtt e \Big)
\]
is still a $1$-dimensional model set.
Let
\begin{equation}
\label{E_tt_d_e+}
\mathtt d_\mathtt e^+ = \mathcal D \cap \mathtt p^{-1}_{V^\ell}\langle \mathtt e\rangle
\end{equation}
be its direction vector field as in \eqref{E_mathtt_d_e}.
As in \eqref{E_tt_d_e_sigma}, by definition of $1$-dimensional slice and since the cones of directions are given by the directions of the superdifferential, for all
\[
z_{\mathfrak b,h^-} \in P_{w+h^-(w,\mathtt e)\mathtt e} = Z^{\ell,+}_\mathfrak b \cap \mathtt p^{-1}_{V^\ell}(w+ h^-(w,\mathtt e)\mathtt e), \quad\mathfrak b\in\mathfrak B_{\ell,+}(\mathfrak a)
\]
we have
\[
\mathtt d_\mathtt e^+(z_{\mathfrak b,h^-}) = \frac{\sigma^{h^-,h^+ + \varepsilon}(z_{\mathfrak b,h^-}) - z_{\mathfrak b,h^-}}{\big| \sigma^{h^-,h^++\varepsilon}(z_{\mathfrak b,h^-}) - z_{\mathfrak b,h^-} \big|},
\]
where
\[
\sigma^{h^-,h^++\varepsilon}(z_{\mathfrak b,h^-}) = z_{\mathfrak b,h^++\varepsilon}
\]
is the unique point of $P_{w + (h^+(w,\mathtt e)+\varepsilon) \mathtt e}$ satisfying
\begin{equation}
\label{E_w-+}
\begin{split}
\big\{ z_{\mathfrak b,h^++\varepsilon} \big\} &= \big( z_{\mathfrak b,h^-} + C^{\ell,+}_\mathfrak b \big) \cap Z^{\ell,+}_{\mathfrak b}\cap\mathtt p_{V^\ell}^{-1} \big( w + (h^+(w,\mathtt e) + \varepsilon) \mathtt e \big) \\
&\overset{\eqref{Point_sudiff_is_partition_in}}{=} \partial^+ \theta(z_{\mathfrak b,h^-}) \cap \mathtt p_{V^\ell}^{-1} \big( w + (h^+(w,\mathtt e) + \varepsilon) \mathtt e \big).
\end{split}
\end{equation}
{\it Step 2.}
Let
\[
\bar\mu \simeq \mathcal H^{k-\ell} \llcorner_{P_{w + \mathtt h^-(w,\mathtt e) \mathtt e}}, \qquad \bar\mu \big( P_{w + \mathtt h^-(w,\mathtt e) \mathtt e} \big) = 1,
\]
and set $\bar\nu=\sigma^{h^-,h^{+}+\varepsilon}_{\#}\bar\mu$. Then clearly
\[
\bar\pi := \big( \mathbb{I}\times\sigma^{h^-,h^{+}+\varepsilon} \big)_{\#} \bar\mu \in \Pi^f_{\mathtt c_{\tilde{\mathbf C}(\mathfrak a)}}(\bar \mu,\bar \nu).
\]
In this step we prove that \eqref{E_w-+} and the fact that $\{\theta^{-1}(\mathfrak a,\mathfrak t)\}_{\mathfrak t \in \mathfrak T}$ is induced by a Borel $\mathtt c_{\tilde{\mathbf C}}$-compatible linear preorder $\preccurlyeq$ imply that
\begin{equation*}
\bar \pi \ \text{\it is the unique transport plan in } \Pi^f_{\mathtt c_{\tilde{\mathbf C}(\mathfrak a)}}(\bar\mu,\bar\nu).
\end{equation*}
First of all observe that, by transversality of $\mathtt p^{-1}_{V^\ell}\langle\mathtt e\rangle$ w.r.t. $\tilde{\mathbf C}(\mathfrak a)$,
\begin{equation}
\label{E_csection}
\Pi^f_{\mathtt c_{\tilde{\mathbf C}(\mathfrak a)}}(\bar\mu,\bar\nu)=\Pi^f_{\mathtt c_{\tilde{\mathbf C}(\mathfrak a)\cap\mathtt p_{V^\ell}^{-1}\langle \mathtt e\rangle}}(\bar\mu,\bar\nu).
\end{equation}
Then, consider the Borel linear preorder
\[
\bar\preccurlyeq =\, \preccurlyeq \cap \; \bar{\mathbf Z}^{\ell,+}_\mathtt e \times \bar{\mathbf Z}^{\ell,+}_\mathtt e, \qquad \bar{\mathbf Z}^{\ell,+}_\mathtt e := \mathbf Z^{\ell,+} \cap \mathtt p^{-1}_{ V^\ell}\Big( w + \big[ h^-(w,\mathtt e),h^+(w,\mathtt e) \big] \mathtt e \Big).
\]
Its equivalence classes are given by the (closed) segments
\[
\bar Z^{\ell,+}_{\mathfrak b,\mathtt e} = Z^{\ell,+}_\mathfrak b \cap \mathtt p^{-1}_{ V^\ell} \Big( w + \big[ h^-(w,\mathtt e),h^+(w,\mathtt e) \big] \mathtt e \Big),
\]
from the transversality of $\mathtt p^{-1}_{V^\ell}(w + [ h^-(w,\mathtt e),h^+(w,\mathtt e) ] \mathtt e)$ w.r.t. $Z^{\ell,+}_{\mathfrak b}$. Moreover, it is clearly $\mathtt c_{\tilde{\mathbf C}(\mathfrak a)\cap\mathtt p_{V^\ell}^{-1}\langle \mathtt e\rangle}$-compatible and, since a carriage of $\bar\pi$ is given by
\[
\bar\Gamma = \mathrm{graph}\,\sigma^{h^-,h^++\varepsilon} \cup \mathrm{graph}\,\mathbb{I} \llcorner_{\bar{\mathbf Z}^{\ell,+}_\mathtt e},
\]
then
\[
\bar \pi \bigg( \bigcup_{\mathfrak b} \bar Z^{\ell,+}_{\mathfrak b,\mathtt e} \times \bar Z^{\ell,+}_{\mathfrak b,\mathtt e} \biggr) = 1.
\]
By Remark \ref{rem_pre} and Theorem \ref{T_A2}, we conclude that $\bar \preccurlyeq$ is
a Borel $(\mathtt c_{\tilde{\mathbf C}(\mathfrak a)\cap\mathtt p_{V^\ell}^{-1}(\langle \mathtt e\rangle)},\bar\mu,\bar\nu)$-compatible linear preorder. In particular, by \eqref{E_csection}, any transport plan $\pi\in\Pi^f_{\mathtt c_{\tilde{\mathbf C}(\mathfrak a)}}(\bar\mu,\bar\nu)$ must satisfy
\[
\pi \bigg( \bigcup_{\mathfrak b} \bar Z^{\ell,+}_{\mathfrak b,\mathtt e} \times \bar Z^{\ell,+}_{\mathfrak b,\mathtt e} \biggr) = 1,
\]
and then since $\bar\mu$ and $\bar\nu$ are supported only on the sections $P_{w + h^-(w,\mathtt e)\mathtt e}$ and $P_{w + (h^+(w,\mathtt e) + \varepsilon) \mathtt e}$, \eqref{E_w-+} implies that $\pi=\bar\pi$.
{\it Step 3.}
In this step we prove the cone approximation property for the $1$-dimensional slice $Z^{\ell,+}_{\mathtt e}$. The sequence of approximating finite union of cone vector fields (see Definition \ref{D_cone_vector}) will be given by the transport rays of transport plans which are optimal w.r.t. the secondary cost
\begin{equation}
\label{E_tt_c_C_''}
\mathtt c_{\tilde{\mathbf C}(\mathfrak a)}''(w,w') =
\begin{cases}
|w-w'| & \text{if $\mathtt c_{\tilde{\mathbf C}(\mathfrak a)}(w,w') < +\infty$}, \\
+\infty & \text{otherwise},
\end{cases}
\end{equation}
and whose first marginal is $\bar\mu$ and second marginal is given by finite sums of Dirac deltas which are weakly converging to $\bar\nu$, as given by the next lemma.
\begin{lemma}
\label{L_appr_meas}
There exists a sequence
\[
\big\{ \bar \nu_n \big\}_{n \in \mathbb{N}} \in \mathcal P \big( P_{w + (h^+(w,\mathtt e) + \varepsilon) \mathtt e} \big)
\]
such that
\begin{enumerate}
\item $\bar \nu_n$ is locally finitely atomic, $\bar \nu_n = \underset{i \in \mathbb{N}}{\sum}\alpha_{n_i}\delta_{z'_{n_i}}$,
\item $\Pi^f_{\mathtt c_{\tilde{\mathbf C}(\mathfrak a)}}(\bar\mu,\bar\nu_n) \neq \emptyset$,
\item $\bar \nu_n$ weakly converges to $\bar \nu$ in $\mathcal P(\mathbb{R}^k)$.
\end{enumerate}
\end{lemma}
Once the lemma is assumed to be valid, the proof of the theorem is concluded as follows.
Let $\bar \pi_n \in \Pi^f_{\mathtt c_{\tilde{\mathbf C}(\mathfrak a)}}(\bar \mu,\bar \nu_n)$ be optimal for $\mathtt c''_{\tilde{\mathbf C}(\mathfrak a)}$ and let $\bar \Gamma''_n$ be a $\mathtt c''_{\tilde{\mathbf C}(\mathfrak a)}$-cyclically monotone carriage. A standard argument based on $|\cdot|$-cyclical monotonicity (see e.g. \cite{Car:strictly}) implies the interiors of the segments $[z,z']$, with $[z,z'] \in \bar \Gamma''_n$, do not intersect. Since $\bar \Gamma''_n$ can be decomposed as
\begin{equation}
\label{E_Gamma_n_def}
\bar \Gamma''_n = \bigcup_{i \in \mathbb{N}} B_i \times \{z'_{n_i}\}, \qquad B_i \subset P_{w + h^-(w,\mathtt e)\mathtt e} \quad \text{Borel and disjoint},
\end{equation}
up to a $\bar\mu$-negligible set, then the approximating sequence of finite unions of cone vector fields (see Definition \ref{D_cone_vector}) is given by
\[
E_{\mathtt d_i} := \big\{ (1-t) z + t z'_{n_i}: t \in [0,1], z \in B_i \big\}
\]
Let
\[
\sigma^{h^-,h^++\varepsilon}_n : P_{w + h^-(w,\mathtt e)\mathtt e} \rightarrow P_{w + h^+(w,\mathtt e)\mathtt e}, \qquad \sigma^{h^-,h^++\varepsilon}_n(B_i) = z'_{n_i},
\]
be the Borel function such that $\bar \pi_n = (\mathbb{I}\times\sigma^{h^-,h^++\varepsilon}_n) _\# \bar \mu$, as it follows from \eqref{E_Gamma_n_def}. The uniqueness result proved in Step 2 implies that the measure $\bar \pi_n$ converges weakly to the measure $\bar \pi$, which equivalently means that
\[
\sigma^{h^-,h^++\varepsilon}_n \rightarrow \sigma^{h^-,h^++\varepsilon} \quad \mathcal H^{k-\ell}\text{-a.e.}.
\]
Then, the last condition for the forward cone approximation property required in Definition \ref{D_forw_back} is satisfied and the theorem is proved.
\end{proof}
We are left with the proof of Lemma \ref{L_appr_meas}.
\begin{proof}[Proof of Lemma \ref{L_appr_meas}]
W.l.o.g. assume that $\exists\,C'\in \mathcal C(k-\ell,\mathbb{R}^{k-\ell})$ such that, for all $z_{\mathfrak b,h^-} \in P_{w + h^-(w,\mathtt e)\mathtt e}$,
\begin{equation*}
\big\{ \mathtt c_{\tilde{\mathbf C}(\mathfrak a)}(z_{\mathfrak b,h^-},\cdot) < +\infty \big\} \cap \mathtt p_{V^\ell}^{-1} \big( w + (h^+(w,\mathtt e) + \varepsilon)\mathtt e \big) \supset z_{\mathfrak b,h^++\varepsilon} + C'
\end{equation*}
and
\[
C' \supset \bigcap_{i=1}^{k-\ell} \big\{ w : w \cdot \mathrm e'_i \geq 0 \big\}
\]
for some fixed system of coordinates $\{\mathrm e'_1,\dots,\mathrm e'_{k-\ell}\}\subset\mathbb{R}^{k-\ell}$
Let
\[
\bar Q = \prod_{i=1}^{k-\ell} \bigg[ - \frac{1}{2}, \frac{1}{2} \bigg] \mathrm e'_j, \qquad \bar Q(z',r) : = z' + r \bar Q.
\]
For $n \in \mathbb{N}$, choose $r_n>0$ sufficiently small such that
\[
\bar Q \bigg( z' + \frac{1}{n} \sum_{i = 1}^{k-\ell} \mathrm e'_{i}, 2 r_n \bigg) \subset z ' + C',
\]
for all $z' \in \mathtt p_{V^\ell}^{-1} ( w + (h^+(w,\mathtt e) + \varepsilon)\mathtt e )$ (clearly any $r_n \leq \frac{1}{n}$ suffices). Let
\[
\big\{ \bar Q(z'_{n_i},r_n) \big\}_{i\in\mathbb{N}}
\]
be a locally finite covering of $\mathtt p_{V^\ell}^{-1}( w + (h^+(w,\mathtt e) + \varepsilon)\mathtt e )$. Then, define the map
\[
\mathtt T_n : \mathtt p_{V^\ell}^{-1} \big( w + (h^+(w,\mathtt e) + \varepsilon) \mathtt e \big) \rightarrow \mathtt p_{V^\ell}^{-1} \big( w + (h^+(w,\mathtt e) + \varepsilon) \mathtt e \big)
\]
by
\begin{equation}
\label{E_tt_T_n}
\mathtt T_n(z') := z_{n_i}' \qquad \text{where} \ i = \min \bigg\{ j : z'_{n_j} \in \bar Q \bigg( z' + \frac{1}{n} \sum_{i=1}^{k-\ell} \mathrm e'_{i}, 2r_n \bigg) \bigg\}.
\end{equation}
The measures
\[
\bar \nu_n := {\mathtt T_n}_\#\bar \nu, \qquad \bar \pi_n := (\mathbb{I}\times\mathtt T_n \circ \sigma^{h^-,h^++\varepsilon})_\# \bar\mu.
\]
satisfy
\[
\bar\pi_n\in\Pi^f_{\mathtt c_{\tilde{\mathbf C}(\mathfrak a)}}(\bar\mu,\bar\nu_n) \qquad \text{and} \qquad \bar\nu_n\rightharpoonup\nu\in\mathcal P(\mathbb{R}^k),
\]
proving the lemma.
\end{proof}
\begin{remark}
\label{R_not_gener_potential}
Since the level sets of the function $\theta$ form a $\mathtt c_{\tilde{\mathbf C}(\mathfrak a)}$-Lipschitz foliation, then from Proposition \ref{P_fol_char} the equivalence classes of negligible $\mathcal H^k$-measure are complete $\mathtt c_{\tilde{\mathbf C}(\mathfrak a)}$-Lipschitz graph. One could then suspect that the proof of Theorem \ref{T_coneappr_fol} can be deduced from the proof of Theorem \ref{T_cone_graph}.
This would be the case if the disintegration of the $\mathcal H^k$-measure on the level sets of negligible $\mathcal H^k$-measure of the generating $\theta$ were absolutely continuous w.r.t. $\mathcal H^{k-1}$. Here we show that in general this is not the case. In fact, we will construct a Borel function $\Theta : [0,1]^2 \rightarrow [0,1]$ whose level sets are subsets of $C^\infty$-functions on $[0,1]$ such that there exists a Cantor set $C$ of positive $\mathcal L^2$-measure on which $\Theta$ is injective. This clearly implies that the disintegration of $\mathcal L^2 \llcorner_C$ w.r.t. $\Theta$ has conditional probabilities made of a single Dirac $\delta$-mass.
Let $\varrho : [0,1] \rightarrow [0,1]$ be a strictly increasing $C^\infty$ function such that
\[
\varrho(0) = 0, \ \varrho(1) = 1, \qquad \varrho(x) = 1 - \varrho(1-x), \qquad \frac{d^k\varrho}{dx^k}(0) = \frac{d^k\varrho}{dx^k}(1) = 0 \quad \forall k \in \mathbb{N}.
\]
Consider the sequence of numbers
\[
c_k = 2^{-2k-1}, \qquad a_0 = 1, \ a_k = \frac{a_{k-1}-c_k}{2} = 3 \cdot 2^{-2-k} + 2^{-2-2k}, \quad b_k = 2^{-2 k^2}.
\]
{\it Step 1.}
If
\[
Q^1 = [0,a_1] \times [0,b_1] = \bigg[ 0,\frac{7}{16} \bigg] \times \bigg[ 0,\frac{1}{4} \bigg]
\]
and having fixed the points
\[
x_0 = 0, \ x_1 = a_1 + c_1 = \frac{9}{16}, \quad y_0 = 0, \ y_1 = b_1 + \frac{b_1}{3} = \frac{1}{3}, \ y_2 = 2 \bigg( b_1 + \frac{b_1}{3} \bigg) = \frac{2}{3},
\]
consider the squares
\[
Q^1_{0,j} = \bigg( x_0,y_j+\frac{1}{24} \bigg) + Q^1, \ Q^1_{1,j} = (x_1,y_j) + Q^1, \qquad j=0,1,2.
\]
We define the level sets of $\Theta$ outside the squares $Q^1_{i,j}$ as follows. First, in the strips
\[
[0,1] \times \bigg[ \frac{7}{24}, \frac{1}{3} \bigg], \quad [0,1] \times \bigg[ \frac{15}{24}, \frac{2}{3} \bigg], \quad [0,1] \times \bigg[ \frac{23}{24}, 1 \bigg]
\]
the level sets are horizontal segments $y = \mathrm{constant}$. In the remaining strips, the construction is completely similar so that we show only the case $[0,1] \times [0,7/24]$.
In the strip $[9/16,1] \times [1/4,7/24]$ the level lines are again $y = \mathrm{constant}$. In the remaining strip $[7/16,9/16] \times [0,7/24]$ we show how to use the function $\varrho$ to connect the points of the vertical segment $\{7/16\} \times [0,1/24]$ to the points of the vertical segment $\{9/16\} \times [0,1/4]$: the construction in the remaining part is symmetrical. In this case, define the level set of $\Theta$ as the curves
\[
\Big\{ y + 5 y \varrho \big( 8 (x - 7/16) \big), x \in [7/16,9/16] \Big\}, \quad y \in [0,1/24].
\]
Due to the regularity of $\varrho$, these curves are $C^\infty$ in the set of definition.
\begin{figure}
\centering{\resizebox{12cm}{12cm}{\input{sudakovcounterex2bis.pdf_t}}}
\caption{The first steps of the construction done in Remark \ref{R_not_gener_potential}.}
\label{Fi_counterex2}
\end{figure}
{\it Step 2.}
In this step we show how to repeat the above construction in a square of the form $[0,a_k] \times [0,b_k]$, in order to define the level sets of $\Theta$ outside finitely many squares of size $[0,a_{k+1}] \times [0,b_{k+1}]$.
Define the points
\[
x_0 = 0, \ x_1 = a_{k+1} + c_{k+1} = \frac{3}{8} \cdot 2^{-k} + \frac{3}{16} \cdot 2^{-2k}, \quad y_j = j \cdot \bigg( b_{k+1} + \frac{b_{k+1}}{\frac{b_k}{b_{k+1}} - 1} \bigg), \ j=0,\dots,\frac{b_k}{b_{k+1}} - 2.
\]
The new squares which will be used in the next step are given by
\[
Q^k_{0,j} = \bigg( x_0,y_j + \frac{1}{2} \cdot \frac{b_{k+1}}{\frac{b_k}{b_{k+1}} - 1} \bigg) + [0,a_{k+1}] \times [0,b_{k+1}], \quad Q^k_{1,j} = (x_1,y_j) + [0,a_{k+1}] \times [0,b_{k+1}],
\]
with $j=0,\dots,\frac{b_k}{b_{k+1}} - 2$.
As before, in the strips
\[
[0,a_k] \times \bigg[ y_j - \frac{1}{2} \cdot \frac{b_{k+1}}{\frac{b_k}{b_{k+1}} - 1},y_j \bigg], \quad j = 1,\dots,\frac{b_k}{b_{k+1}} - 2,
\]
the level sets of $\Theta$ are straight lines $y = \mathrm{constant}$, and the same for the strips
\[
[0,a_{k+1}] \times \bigg[y_j,y_j + \frac{1}{2} \cdot \frac{b_{k+1}}{\frac{b_k}{b_{k+1}} - 1} \bigg], \quad j = 0,\dots,\frac{b_k}{b_{k+1}} - 2,
\]
\[
[a_{k+1}+c_{k+1},a_k] \times \bigg[y_j - \frac{b_{k+1}}{\frac{b_k}{b_{k+1}} - 1}, y_j - \frac{1}{2} \cdot \frac{(b_{k+1})^2}{\frac{b_k}{b_{k+1}} - 1} \bigg], \quad j = 1,\dots,\frac{b_k}{b_{k+1}} - 1.
\]
Similarly as done in Step 1, we just show how to define the level sets connecting the segments
\[
\{a_{k+1}\} \times \bigg[ 0,\frac{1}{2} \cdot \frac{b_{k+1}}{\frac{b_k}{b_{k+1}} - 1} \bigg] \quad \text{and} \quad \{a_{k+1} + c_{k+1}\} \times [0,b_{k+1}].
\]
One just defines the level sets of $\Theta$ to be
\[
\bigg\{ y + \bigg( 2 \frac{b_k}{b_{k+1}} - 3 \bigg) y \varrho \bigg( \frac{x - a_{k+1}}{c_{k+1}} \bigg), x \in [a_{k+1},a_{k+1}+c_{k+1}] \bigg\}, \quad y \in \bigg[ 0,\frac{1}{2} \cdot \frac{b_{k+1}}{\frac{b_k}{b_{k+1}} - 1} \bigg].
\]
{\it Step 3.}
We show that the level sets are $C^\infty$. In fact, by the estimate
\[
\frac{d^\ell}{dx^\ell} \big( b_k \varrho(x/c_k) \big) = \mathcal O(1) \frac{b_k}{c_k^\ell} = \mathcal O(1) 2^{-k^2 + 2 k \ell},
\]
it follows that the curves have a uniform bound in $C^\ell$, for every fixed $\ell$, and thus they belongs to $C^\infty$.
Moreover, it is fairly easy to see that the intersection of each level set with the line $\{1/2\} \times [0,1]$ determines completely the curve, so that the function $\Theta$ can be defined as the $y$-coordinate of this intersection. With a slight variation of this construction one can obtain $\Theta$ to be regular.
{\it Step 4.}
Let $C$ be the compact Cantor set obtained by intersecting all the squares $Q^k_{i,j}$:
\[
C = \bigcap_{k \in \mathbb{N}} \bigcup_{i,j} Q^k_{i,j}.
\]
It is standard to see that
the function $\Theta$ is single valued on $C$. Moreover, a simple area estimate yields
\[
\mathcal L^2 \bigg( \bigcup_{i,j} Q^k_{i,j} \bigg) = 2^{k} a_{k} \cdot \bigg( 1 - \sum_{\ell = 1}^k b_k \bigg) \geq \frac{3}{4} \cdot \frac{2}{3} = \frac{1}{2},
\]
and this concludes the example.
\end{remark}
\section{Proof of Theorems \ref{T_subpart_final}-\ref{T_Monge_final}.}
\label{S_proff_main_Th}
In this final section we collect the proofs of the remaining main theorems stated in Section \ref{Ss_main_Sud}.
\subsection{Proof of Theorems \ref{T_subpart_step} and \ref{T_final_nu}}
\label{Ss_proof_Th_2}
In this section we prove Theorem \ref{T_subpart_step}, that we recall below.
\begin{theorem2}
\label{T_subpart_step2}
Let $\{Z^k_\a, C^k_\a\}_{\nfrac{k = 0,\dots,d}{\mathfrak a \in \mathfrak A^k}}$ be a Lebesgue-regular directed locally affine partition in $\mathbb{R}^d$ and let $\mu$, $\nu$ be probability measures in $\mathcal P(\mathbb{R}^d)$ such that $\mu\ll\mathcal L^d$ and $\Pi^f_{\mathtt c_{\mathbf D}}(\mu,\nu)\neq\emptyset$.
Then, for all fixed $\check \pi \in \Pi^ f_{\mathtt c_{\mathbf D}}(\mu,\nu)$, there exists a directed locally affine subpartition $\{\check Z^\ell_{\mathfrak b}, \check C^\ell_{\mathfrak b}\}_{\nfrac{\ell = 0,\dots,d}{\mathfrak b \in \mathfrak B^\ell}}$ of $\{Z^k_\a, C^k_\a\}_{k,\mathfrak a}$, up to a $\mu$-negligible set $N_{\check \pi}$, such that
\begin{equation*}
\big\{ \check Z^\ell_\mathfrak b,\check C^\ell_\mathfrak b \big\}_{\ell,\mathfrak b} \quad \text{is Lebesgue-regular},
\end{equation*}
and setting $\check \nu^\ell_\mathfrak b := (\mathtt p_2)_\#\check \pi^\ell_\mathfrak b$, where $\check \pi^\ell_\mathfrak b$ is the conditional probability on the partition $\{\check Z^{\ell}_\mathfrak b \times \mathbb{R}^d\}_{\ell,\mathfrak b}$, then the sets
\begin{equation}
\label{E_subp_ell_cycl}
\Bigl\{ \check Z^\ell_\mathfrak b:\,\check Z^\ell_\mathfrak b\subset Z^\ell_\mathfrak a\text{ for some $\mathfrak a\in\mathfrak{A}^\ell$}, \ell=1,\dots,d \Bigr\}
\end{equation}
form a $\Pi^ f_{\mathtt c_{\mathbf D}}(\mu,\{\check \nu^\ell_\mathfrak b\})$-cyclically connected partition.
\end{theorem2}
\begin{proof}
Let $\{Z^k_\a, C^k_\a\}_{\nfrac{k = 0,\dots,d}{\mathfrak a \in \mathfrak A^k}}$ be a Lebesgue-regular directed locally affine partition in $\mathbb{R}^d$ and let $\mu \ll \mathcal L^d$, $\nu$ be probability measures such that $\Pi^f_{\mathtt c_{\mathbf D}}(\mu,\nu) \neq \emptyset$.
{\it Step 1.}
By Proposition \ref{P_countable_partition_in_reference_directed_planes}, we can restrict the proof of the theorem to a fixed $\ell$-directed sheaf set, which we will denote again by $\{Z^\ell_\mathfrak a, C^\ell_\mathfrak a\}_{\mathfrak a\in\mathfrak{A}^\ell}$.
Moreover, by Proposition \ref{P_sub_sheaf_fol} it is enough to prove the existence of subpartitions as in Theorem \ref{T_subpart_step} for the fibration $\{\tilde Z^\ell_\mathfrak a,\tilde C^\ell_\mathfrak a\}_{\mathfrak a\in\mathfrak{A}^\ell}$, $\tilde{\mathbf C}^\ell(\mathfrak a)=C^\ell_\mathfrak a$, given by Proposition \ref{P_map_sheaf_set_into_fibration}. It is indeed clear that $(\mu,\check \Gamma,\mathtt c_{\mathbf D})$-cyclically connected sets, where $\check \pi(\check \Gamma)=1$ are mapped into $(\tilde\mu,\tilde\Gamma,\mathtt c_{\tilde{\mathbf C}})$-cyclically connected sets and viceversa, being $\tilde\mu$ obtained through \eqref{E_rxr}. Since the map $\mathtt r$ defined in \eqref{E_r_map} is not a bijection of $\mathbb{R}^d$ into $\mathfrak{A}^\ell \times \mathbb{R}^\ell$, then the transport problem on the fibration $\{\tilde Z^\ell_\mathfrak a,\tilde C^\ell_\mathfrak a\}_{\mathfrak a\in\mathfrak{A}^\ell}$ depends on the conditional second marginals $\{\check \nu^\ell_\mathfrak a\}$ of $\check \pi$ w.r.t. the partition $\{Z^\ell_\mathfrak a\}_{\ell,\mathfrak a}$. Let $\tilde\nu=\int\check\nu^\ell_\mathfrak a\,dm(\ell,\mathfrak a)$.
{\it Step 2.}
Let $\bar \theta$ be the equivalence relation given by Theorem \ref{T_cfibrcfol} for the transport problem $\Pi^f_{\mathtt c_{\tilde{\mathbf C}}}(\tilde\mu,\tilde \nu)$. In particular, \eqref{E_subcollection} gives the sets that, when mapped back through the map $\mathbf r$ on the sheaf set, satisfy \eqref{E_subp_ell_cycl} and are $\Pi^f_{\mathtt c_{\mathbf D}}(\mu,\{\check \nu^\ell_\mathfrak b\})$-cyclically connected.
By Propositions \ref{P_ex_fol} and \ref{P_fol_char}, the remaining sets form a $\mathtt c_{\tilde{\mathbf C}}$-foliation into graphs of cone-Lipschitz functions. Let $\{\check Z^\ell_{\mathfrak a,\mathfrak b}\}_{\ell'<\ell}$ be the $\mathtt c_{\tilde{\mathbf C}}$-differential partition given by Corollary \ref{C_v}. By Theorems \ref{T_coneappr_fol} and \ref{T_FC_no_initial} its complementary $N_{\check \pi}$ is $\mu$-negligible, and by Corollary \ref{C_infinnegl} the partition is Lebesgue-regular.
\end{proof}
\subsection{Proof of Theorem \ref{T_final_nu}}
\label{Ss_proof_T_final}
The only missing point is to prove that the conditional second marginals $\{\check \nu^\ell_\mathfrak b\}_{\ell,\mathfrak b}$ are independent of the particular transference plan $\check \pi \in \Pi^{\mathrm{opt}}_{\d{\cdot}}(\mu,\nu)$ chosen.
From Corollary \ref{C_infinnegl} and Theorem \ref{T_FC_no_initial} it follows that
\[
\nu = \int \check \nu^\ell_\mathfrak b\,dm(\ell,\mathfrak b)
\]
is a disintegration, and Corollary \ref{C_transp_graph} implies that $(\mathtt p_2)_\#\check \pi^\ell_\mathfrak b (\check Z^\ell_\mathfrak b) = 1$. Hence, from the uniqueness of strongly consistent disintegrations, it follows that
\[
\check \nu^\ell_\mathfrak b = (\mathtt p_2)_\#\check \pi^\ell_\mathfrak b \qquad m\text{-a.e.}\ \ell,\mathfrak b,
\]
yielding the improved version of Conditions (4')-(5') in Theorem \ref{T_final_nu}.
\newpage
|
\section{Introduction}
\label{intro}
The task of data analysis in particle physics often deals with data sets comprising collision events that contain the signature of a scattering of interest as well as background events that correspond to uninteresting processes mimicking the signal. When estimating the number of signal events, background probability density functions (PDFs) are often extrapolated from control regions to a predefined signal window and are subsequently used in template fits. However, this can only provide an estimate of the expected, or average, number of background events under the signal, and cannot take into account the effect of fluctuations. In practice, when the number of signal events, $S$, is much higher than the size of the typical fluctuations on the number of background events, $\sigma_B=\sqrt{\left<B\right>}$, $\left<B\right>$ being the expected number of background events in the signal window, the discrepancy between $B$ and $\left<B\right>$ can be neglected.
Nonetheless, when the number of signal events is low enough, the difference can be sizable. This can occur in the early stages of data analysis following the discovery of a new particle, or in the analysis of low-cross section processes. In such cases, the expected number of background events in the signal window can be a biased estimate of the actual number.
We report on the development of a data-driven technique that aims to estimate the actual number of background events under an observed signal, as opposed to the expected number. Our algorithm makes it possible to decompose an input mixture of signal and background events, e.g. a collection of events that pass all selection criteria corresponding to the end-point of a given analysis. This allows the shape of the background PDF to be estimated from the data, thereby taking into account the effect of statistical fluctuations. The development of this technique was influenced by a number of statistical methods, most notably the Gibbs Sampler \cite{geman} for mixture model decomposition, Expectation Maximisation \cite{EM}, and Data Augmentation \cite{DA}.
\section{The algorithm}
\label{algo}
The algorithm that we use to decompose the input mixture of signal and background events is related to a method that we have proposed with reference to a different application to data analysis at high-luminosity hadron colliders \cite{gibbshep2, gibbshep}.
The PDF of the underlying statistical model has the form $F=\alpha_0 f_0(x) + \alpha_1 f_1(x)$, where $\alpha_0$ and $\alpha_1$ are the fractions of background and signal events in the input data set, respectively, with $\alpha_0 + \alpha_1=1$, and where $f_0$ ($f_1$) is the background (signal) PDF. In the context of this study, the variable $x$ is interpreted as the invariant mass of a set of final state particles.
A notable feature of our approach, when compared to classical mixture models where predefined subpopulation PDF shapes are typically enforced a priori, is the nonparametric definition of the subpopulation PDFs, $f_j$. At every iteration of the algorithm, individual events are mapped to signal or background on a probabilistic basis, and the estimate $\varphi_j$ of the subpopulation PDF $f_j$ at that iteration is obtained by means of spline interpolation\footnote{
We have used the alglib C++ library \cite{alglib} with this implementation of the algorithm.
} of the histograms of $x$ corresponding to those events that are mapped to signal or background at that iteration.
This allows the algorithm to estimate generic deviations of the PDF shapes from the corresponding control sample templates due to fluctuations in the data. The shapes of the signal and background distributions in the data set analysed are ultimately estimated as splined histograms averaged over a predefined number of iterations.
The pseudocode of the algorithm is given below, subscripts ``sig'' and ``bkg'' relating to signal and background, respectively. The value of quantity $v$ at iteration $t$ is denoted by $v^{(t)}$ throughout.
\begin{enumerate}
\item {\bf Initialization:} Set $\alpha_{bkg}=\alpha^{(0)}_{bkg}=\alpha_{sig}=\alpha^{(0)}_{sig}=0.5$, where $\alpha_{bkg}=\alpha_0$ and $\alpha_{sig}=\alpha_1=1-\alpha_{bkg}$. Initial estimates $\varphi^{(0)}_j$ of the subpopulation PDFs $f_j$, $j=0,1$, are given by splined one-dimensional histograms of $x$ obtained from high-statistics control samples.
\item {\bf Iteration $t$:}
\begin{enumerate}
\item Generate $z_{ij}^{(t)}$ for all events $i$ and distributions $j$ according to $P(z_{ij}^{(t)}=1 | \alpha_j^{(t-1)}, \varphi_j^{(0)},x_i) = \frac{\alpha_j^{(t-1)}\varphi_j^{(0)}(x_i)}{\alpha_0^{(t-1)}\varphi_0^{(0)}(x_i)+\alpha_1^{(t-1)}\varphi_1^{(0)}(x_i)}$. Both the nonparametric treatment of the PDFs and the use of $\varphi_j^{(0)}$ instead of $\varphi_j^{(t-1)}$ to map individual events to signal or background distinguish this implementation from the classical Gibbs sampler for mixture models.
\item Set $\alpha_j^{(t)}=\sum_{i=1}^N z_{ij}^{(t-1)}/N$, $j=0,1$.
\end{enumerate}
\end{enumerate}
We used a total number of 6,000 iterations, and averaged the PDF estimates, $\varphi_j$, over the last 4,000. These settings allowed the algorithm to reach convergence in all runs performed in this study, and no significant difference in the results was observed by changing them.
A more detailed description of this implementation of the algorithm can be found in \cite{gibbshep5}. The execution time was $\sim50~\mbox{s}$ per run on the data sets analysed using a 2~GHz Intel Processor with 1~GB RAM, which we consider reasonable for offline use.
\section{Results}
\label{results}
We illustrate this technique on a toy Monte Carlo data set obtained superimposing a gaussian signal with a first-order polyomial background. In the following, we will interpret the signal distribution as an invariant mass distribution corresponding to
the decay of a particle with mass $m=125~\mbox{GeV/c}^2$ and width $1~\mbox{GeV/c}^2$.
We superimposed $S=200$ signal events to a total of 4,200 background events in the region $115~\mbox{GeV/c}^2<m<135~\mbox{GeV/
c}^2$, corresponding to an average of $\left<B\right>=1,600$ background events in the signal region, which is defined by $120~\mbox{GeV/c}
^2<m<130~\mbox{GeV/c}^2$.
Due to statistical fluctuations in the data, different samples correspond to different numbers of background events in the signal window. In this study, the standard deviation on the number of background events with $120~\mbox{GeV/c}^2<m<130~\mbox{GeV/c}^2$ is $\sigma_B=\sqrt{\left<B\right>}=40$ events, which is sizable when compared to the number of signal events generated, $S=200$. This illustrative scenario is not dissimilar from the early stages of data analysis following the observation of a Higgs boson in the $\gamma\gamma$ final state at the Large Hadron Collider (LHC) at CERN \cite{higgs_LHC_2012_ATLAS, higgs_LHC_2012_CMS}.
High-statistics control samples were generated corresponding to 30,000 signal and 30,000 background events, and were used to obtain initial conditions on the signal and background PDF shapes. The function of the algorithm is essentially to iteratively refine those initial conditions based on the data, thereby taking into account the effect of statistical fluctuations. As a consistency check, the estimated fraction of background events in the input data set, $\hat{\alpha}_0$, was found to be in agreement with the true value within 2\% in all runs used in this study.
\begin{figure*}
\centering
\subfloat[]{
\includegraphics[scale=0.48]{hs_var1_bkg_truth_cs.eps}
}
\subfloat[]{
\includegraphics[scale=0.48]{hs_var1_bkg_truth_avg.eps}
}\\
\subfloat[]{
\includegraphics[scale=0.48]{comp_var1_bkg_cs_truth.eps}
}
\subfloat[]{
\includegraphics[scale=0.48]{comp_var1_bkg_avg_truth.eps}
}
\caption[]{
(a) True background distribution (points) superimposed with the PDF obtained from the high-statistics control sample (curve). (b) The same true background distribution (points) superimposed with the background PDF estimated from the data using the algorithm (curve). (c) Ratio between the background PDF obtained from the control sample and the true distribution. (d) Ratio between the background PDF estimated using the algorithm and the true distribution.
}
\label{fig:PDF_avg}
\end{figure*}
The performance of the algorithm in terms of estimating the shape of the background PDF in the data set analysed is illustrated in figure \ref{fig:PDF_avg}. Figure \ref{fig:PDF_avg} (a) displays the true background distribution (points) superimposed with the PDF obtained from the high-statistics control sample (curve). The discrepancies due to statistical fluctuations in the data are apparent. The points in figure \ref{fig:PDF_avg} (b) show the same true background distribution as in figure \ref{fig:PDF_avg} (a), but in this case the superimposed curve is the PDF estimated from the data using the algorithm, averaged over the last 4,000 iterations from a total of 6,000.
The ratio between the background control sample PDF and the true PDF is displayed in figure \ref{fig:PDF_avg} (c), which again highlights the effect of fluctuations. The corresponding ratio between estimated and true PDF is shown in figure \ref{fig:PDF_avg} (d), and shows a significantly-improved agreement.
It is worth recalling that, for the purpose of this study, what we are interested in is the shape of the background PDF. In fact, our objective is to estimate the actual number of background events under the signal as opposed to the expected number. The signal-related plots corresponding to figure \ref{fig:PDF_avg} showed good agreement between the estimated and the true distribution, and were used together with the estimated fraction of background events in the data in order to check the consistency of the results obtained using the algorithm.
The plots in figure \ref{fig:PDF_avg} refer to a run of the algorithm on a data set with $B=1,571$ background events in the signal region $120~\mbox{GeV/c}^2<m<130~\mbox{GeV/c}^2$. The corresponding number of events estimated with that run of the algorithm was $\hat{B}=1586.5$.
The algorithm was also run on multiple toy Monte Carlo data sets, corresponding to different numbers of background events in the signal window. Our preliminary estimate of the uncertainty on $\hat{B}$, i.e. on the estimated number of background events under the signal, is $\sim50$ events. Work is underway to reduce this uncertainty below the size of typical background fluctuations in the data, $\sigma_B=\sqrt{\left<B\right>} = 40$ events. Our studies suggest that the uncertainty on $\hat{B}$ is dominated by the uncertainty on the estimated fraction of background events in the data set, $\hat{\alpha}_0$. In fact, when the algorithm is run with $\alpha_0$ kept fixed at the corresponding true value, the uncertainty on $\hat{B}$ drops from 50 to 12 events.
The results obtained running the algorithm on the different input data sets are summarised in table \ref{tab:syst_toy}, where $B_{gen}$ denotes the true number of background events in the signal window at generation, $\hat{B}$ is the corresponding number estimated using the algorithm, and $\Delta B = \hat{B}-B_{gen}$. The quantities $\hat{B}^*$ and $\Delta B^*$ in the table have a similar meaning as $\hat{B}$ and $\Delta B$, but the values were obtained running the algorithm with $\alpha_0$ kept fixed at its true value. The average and standard deviation of $\hat{B}$ across the runs are referred to as $\left<B\right>$ ( $\left<B\right>^*$) and $\sigma_B$ ($\sigma_B^*$), respectively.
\end{multicols}
\begin{table}[ht]
\centering
\begin{tabular}{c c c c c c}
\hline\hline
Run & $B_{gen}$ & $\hat{B}$ & $\Delta B$ & $\hat{B^*}$ & $\Delta B^*$ \\ [0.5ex]
\hline
1 & 1536 & 1618.2 & 82.2 & 1549.3 & 13.3 \\% [1ex]
2 & 1569 & 1645.0 & 76.0 & 1592.3 & 23.3 \\% [1ex]
3 & 1579 & 1615.2 & 36.2 & 1584.7 & 5.7 \\% [1ex]
4 & 1625 & 1637.7 & 12.7 & 1630.2 & 5.2 \\% [1ex]
5 & 1558 & 1579.7 & 21.7 & 1548.0 & -10.0 \\% [1ex]
6 & 1576 & 1602.5 & 26.5 & 1588.2 & 12.2 \\% [1ex]
7 & 1571 & 1586.5 & 15.5 & 1579.1 & 8.1 \\% [1ex]
8 & 1584 & 1628.6 & 44.6 & 1584.8 & 0.8 \\% [1ex]
9 & 1597 & 1664.1 & 67.1 & 1604.4 & 7.4 \\% [1ex]
10 & 1644 & 1621.9 & -22.1 & 1640.4 & -3.7 \\% [1ex]
11 & 1631 & 1688.9 & 57.9 & 1636.6 & 5.6 \\% [1ex]
12 & 1573 & 1661.4 & 88.4 & 1586.9 & 13.9 \\% [1ex]
13 & 1626 & 1655.6 & 29.6 & 1616.3 & -9.7 \\% [1ex]
14 & 1583 & 1641.2 & 58.2 & 1592.8 & 9.8 \\% [1ex]
15 & 1613 & 1663.2 & 50.2 & 1635.8 & 22.8 \\% [1ex]
16 & 1593 & 1663.7 & 70.7 & 1606.1 & 13.1 \\% [1ex]
17 & 1583 & 1604.9 & 21.9 & 1585.2 & 2.2 \\% [1ex]
18 & 1603 & 1646.8 & 43.8 & 1586.8 & -16.2 \\% [1ex]
19 & 1624 & 1667.8 & 43.8 & 1630.4 & 6.4 \\% [1ex]
20 & 1580 & 1604.5 & 24.5 & 1575.4 & -4.6 \\% [1ex]
\hline
& $\left<\Delta B\right>$ = 42.5 & $\sigma_B$ = 27.3 & $\left<\Delta B^*\right>$ = 5.3 & $\sigma_B^*$ = 10.4 & \\
\hline
\end{tabular}
\caption{
Results obtained running the algorithm on different toy Monte Carlo data sets. The quantities $B_{gen}$ and $\hat{B}$ refer to the true and to the estimated number of background events in the signal region, respectively, and $\Delta B = \hat{B}-B_{gen}$. The quantities $\hat{B}^*$ and $\Delta B^*$ have a similar meaning as $\hat{B}$ and $\Delta B$, but the values were obtained keeping $\alpha_0$ fixed at its true value. The average and standard deviation of $\hat{B}$ across the runs are represented by $\left<B\right>$ ( $\left<B\right>^*$) and $\sigma_B$ ($\sigma_B^*$), respectively.
}
\label{tab:syst_toy}
\end{table}
\begin{multicols}{2}
\section{Conclusions and outlook}
\label{concl}
We have reported on the development of a data-driven technique that aims to estimate the actual, as opposed to the expected, number of background events under an observed signal in particle physics. Established methods that rely on the extrapolation of background distributions from control regions to a predefined signal window allow a precise estimation of the expected, or average, number of background events under the signal. However, the actual number of background events in the signal window can deviate from the average due to statistical fluctuations in the data. Although the discrepancy is often negligible when compared to the number of signal events, it is not necessarily so in the early stages of data analysis following the discovery of a new particle, or more generally in the analysis of low-cross section processes.
We have described an algorithm that uses the data to estimate the shape of the background distribution in a predefined signal window, e.g. using the end-point of a given analysis i.e. a collection of events that pass all selection criteria. Control samples are used only to provide initial conditions for the background PDF, but the PDF shape is otherwise estimated directly from the same data set that contains the observed excess of signal events. We have discussed results on toy Monte Carlo data, with reference to an illustrative scenario that is not dissimilar from the early stages of data analysis following the discovery of a Higgs boson in the $\gamma\gamma$ channel.
We have provided a preliminary estimate of the uncertainty associated with the estimated number of background events in the signal window at the level of 50 events, out of a total average number $\left<B\right> = 1,600$. Although we consider these results encouraging, the uncertainty is still larger than the size of the typical background fluctuations in the data, which is given by $\sigma_B = \sqrt{\left<B\right>} = 40$ events. Work is underway to improve the performance of the algorithm in this respect. It should also be emphasised that, since the above uncertainty is expected to depend significantly on $B$, the assessment of the performance of this method will have to take into account the specifics of the analysis in question.
\section{Acknowledgments}
The author wishes to thank the High Energy Physics Group at Brunel University for a stimulating environment, and particularly Prof. Akram Khan, Prof. Peter Hobson and Dr. Paul Kyberd for fruitful conversations, as well as Dr. Ivan Reid for help on technical issues. Particular gratitude also goes to the High Energy Physics Group at University College London, especially to Prof. Jonathan Butterworth for his valuable comments. The author also wishes to thank Prof. Trevor Sweeting and Dr. Alexandros Beskos at the UCL Department of Statistical Science for fruitful discussions. Finally, particular gratitude goes to Prof. Carsten Peterson and to Prof. Leif Lönnblad at the Department of Theoretical Physics, Lund University.
|
\section{Introduction}
The understanding of high energy scattering in gauge theories is both very important
and interesting. Of particular interest is the specific case of the Regge limit, namely
the limit of high energy and fixed momentum transfer. The relevant dynamics in QCD
depends on the magnitude of the momentum transfer in comparision with $\Lambda_{QCD}$.
When it is small, we are dealing with the physics of the `soft pomeron' which
is nonperturbative and still remains mysterious from the theoretical point of view.
When the momentum transfer is large enough (but fixed), perturbative computations
are applicable and the description is in terms of the `hard pomeron:'
the BFKL pomeron~\cite{BFKL}.
This regime is relevant for applications to Deep Inelastic Scattering processes.
The BFKL pomeron leads to power law behaviour of scattering amplitudes with energy $s^j$,
where $j$ is the pomeron intercept.
Its value depends on quantum numbers of the particular state of the pomeron being exchanged.
In the Regge limit, there is
a clear separation between the longitudinal plane defined by the collision and
the transverse plane. The chief characteristic of the BFKL equation is its invariance
under the \ensuremath{\mathfrak{sl}(2,\mathbb{C})}{} symmetry group of the transverse plane. Therefore the intercept
is always a function of the relevant principal continuous series representation
of \ensuremath{\mathfrak{sl}(2,\mathbb{C})}{} parametrized by\footnote{In case of compound states of more than two reggeized
gluons, additional discrete quantum numbers will appear.}
\begin{equation}
h=\f{1+n}{2}+i \nu \quad\quad \tilde{h}=\f{1-n}{2}+i\nu
\end{equation}
Currently, the BFKL pomeron intercept is known at leading and next-to-leading order both in QCD
and in ${\cal N}=4$ SYM \cite{NLO}. In fact at leading order the results exactly coincide in the two theories.
At NLO level, differences appear but still there are many similarities in the structure
of the expressions,
hence insight into the properties of the BFKL pomeron in ${\cal N}=4$ SYM at any coupling
would be also significant for QCD physics.
The reason why it is interesting to study the BFKL pomeron in ${\cal N}=4$ SYM is that due to
the AdS/CFT correspondence we can access nonperturbative strong coupling physics.
In particular, now we know the BFKL pomeron properties at strong coupling,
where the intercept starts off at 2 \cite{JP} (corresponding to graviton exchange in the dual
picture) and decreases as we lower the coupling \cite{BPST,Costa,KL}.
Even more so, using the methods of integrability in the AdS/CFT correspondence,
we may hope to obtain an exact expression for the intercept valid at any coupling which
interpolates between the known LO and NLO BFKL and the strong coupling
behaviour. However, we are still away from realizing this goal.
Currently the dominant approach in tackling the problem of the BFKL pomeron using
integrability is
to employ a relation between the analytical continuation of the anomalous dimensions
of twist-2 operators and the BFKL intercept. This relation has been initially used
to derive constraints on the higher loop anomalous dimensions \cite{KLRSV,TWIST2L4,TWIST2L5}
and to show the neccessity of wrapping corrections. However now, as our knowledge
of the anomalous dimensions is much more complete, it has been exploited
in the opposite direction to
gain knowledge about the strong coupling BFKL intercept \cite{Costa,KL}.
The drawback of this way of approach is that we are led to study
\emph{the analytical continuation} of the anomalous dimensions
at nonphysical values of the spin. This poses significant challenges for
the current exact formulations of the ${\cal N}=4$ SYM spectral problem which typically
provide just numerical values (in case of generic coupling) of the anomalous
dimensions for \emph{integer} spin.
However, some work in this direction has already been carried out \cite{RJBAXTER}.
In this paper we would like to adopt a different, more direct approach,
where we would like to identify string configurations relevant for BFKL
without assuming \emph{a~priori} any analytical continuation properties.
This is especially interesting in view of a more general situation,
e.g.\ with nonzero values of the conformal spin, where we lack a clear picture
of what would be the relevant class of analytically continued operators.
Similarly, apart from the BFKL pomeron, in perturbative gauge theory there is
a whole range of compound states of more than two reggeized gluons which
also give contributions to scattering amplitudes. It is not known what would
be their behaviour/counterparts at strong coupling. We hope that the direct
approach advocated in this paper might shed some light on these issues.
In this paper we will restrict ourselves to the strong coupling regime.
Let us first review the setting of the spectral problem in this context.
The spectral problem amounts to expressing the energy (equivalently, the
dimension) as a function of all other conserved charges of the theory, i.e.\
two spins and three R-charges, and other quantum numbers labelling different
states with the same conserved charges. At strong coupling, an important
class of states corresponds to classical spinning string solutions
in $AdS_5 \times S^5$. These solutions have all conserved charges proportional
to $\sqrt{\lambda}$ and a specific solution to the equations of motion
of the $AdS_5 \times S^5$ $\sigma$-model leads to an explicit
relation
\begin{equation}
\f{E}{\sqrt{\lambda}} = F\left( \f{S_1}{\sqrt{\lambda}}, \f{S_2}{\sqrt{\lambda}};
\f{J_1}{\sqrt{\lambda}},\f{J_2}{\sqrt{\lambda}},\f{J_3}{\sqrt{\lambda}} \right)
\end{equation}
We would like to adopt the same philosophy and find string configurations
which would similarly provide a relation between the strong coupling intercept
and the relevant quantum numbers. Again, in the classical limit we will
consider the case that all the quantum numbers scale like $\sqrt{\lambda}$,
hence we are looking for a relation of the form
\begin{equation}
\f{j}{\sqrt{\lambda}} =\tilde{F} \left( \f{\nu}{\sqrt{\lambda}}, \f{n}{\sqrt{\lambda}}, \ldots \right)
\end{equation}
This relation should follow from exhibiting a specific solution to the equations of motion
of the $AdS_5 \times S^5$ $\sigma$-model.
One of the goals of the present paper is to identify these solutions and to analyze
them from the point of view of integrability of strings in $AdS_5 \times S^5$. Hence we
will identify the properties of the corresponding algebraic curve.
In particular we would like to analyze the particle interpretation of these solutions.
In this paper we will restrict ourselves to the case of zero conformal spin, in order
to have a direct counterpart of the more standard link with twist-2 operators.
However our methods are straightforward to apply for the general case of nonzero conformal
spin going beyond the twist-2 sector (using e.g.\ analogs of classical solutions
in \cite{FT}).
The plan of this paper is as follows. In the next two sections we will review some basic facts
about weak and strong coupling BFKL and the identification of the pomeron quantum
numbers within the complexified conformal group.
Then we will move on to describe the relevant classical string solution and, after
reviewing for completeness the particle interpretation of algebraic curve constructions,
we will analyze this solution within the algebraic curve framework. We close the paper with
a summary and outlook.
\section{Scattering in the Regge limit and BFKL solution}
As stated in the introduction, the BFKL intercept is expressed as a function
of the \ensuremath{\mathfrak{sl}(2,\mathbb{C})}{} representations
\begin{align} h=\frac{1+n}2+i\nu \qquad \tilde{h}=\frac{1-n}2+i\nu \label{eq:sl2labels}\end{align}
where $n$ is an integer and $\nu$ is a continuous real parameter.
Often it is convenient to consider the case of vanishing conformal spin $n=0$ and
then use the notation $\gamma=h=\tilde{h}=\f{1}{2}+i\nu$.
At weak coupling, we know the first two terms in the expression for the intercept:
\begin{equation}
j=1+\f{\lambda}{4\pi^2} \left[ \chi_{LO}(\gamma)+\f{\lambda}{16\pi^2} \chi_\textit{NLO}(\gamma)
+\ldots \right]
\end{equation}
Namely
\begin{align}
\chi_{LO}(\gamma) &= 2\Psi(1)-\Psi(\gamma)-\Psi(1-\gamma) \\
\chi_\textit{NLO}(\gamma) &= \Psi''(\gamma)+\Psi''(1-\gamma)+6\zeta_3 -2\zeta_2 \chi_{LO}(\gamma)-2\Phi(\gamma)-2\Psi(1-\gamma)
\end{align}
with
\begin{equation}
\Phi(\gamma)=2\sum_{k=0}^\infty \f{(-1)^{k+1}}{k+\gamma} \beta'(k+1)
\quad\quad
\beta'(z) =\f{1}{4} \left[ \Psi'\left(\f{1+z}{2}\right) -\Psi'\left(\f{z}{2}\right)
\right]
\end{equation}
A general form of the strong coupling intercept was postulated to be
\begin{equation}
j(\nu)=2-\f{2+2\nu^2}{\sqrt{\lambda}} \left(1+\sum_{n=2}^\infty
\f{\tilde{j}_n(\nu^2)}{\lambda^{(n-1)/2}}\right)
\end{equation}
with $\tilde{j}_n(\nu^2)$ being a polynomial of order $n-2$ \cite{Cornalba}.
Subsequently, it was argued in \cite{Costa}, that $\tilde{j}_n(\nu^2)$ should be
a polynomial of order at most $n/2$.
This observation is in fact a direct consequence of the existence of an underlying
classical string solution. Namely, for a classical string solution both $j$ and $\nu$
should scale at most like $\sqrt{\lambda}$. If there would be terms of higher order than
$n/2$ this would clearly be impossible.
At strong coupling we know the expression up to $\lambda^{-2}$ terms, as well as the leading-$\nu$ term at $\lambda^{-5/2}$~\cite{Costa,KL}
\begin{align}\label{e.strong}
j(\nu)&=2-\f{2+2\nu^2}{\sqrt{\lambda}}\times\\\notag&\times\left(1+\frac{\tfrac12}{\sqrt\lambda}+\frac{-\tfrac18+\tfrac32\nu^2}{\lambda}+\frac{-1-3\zeta_3+(\tfrac{21}8-3\zeta_3)\nu^2}{\lambda^{3/2}}+\frac{\frac{21}4\nu^4+O(\nu^2)}{\lambda^2}\right)
\end{align}
Let us now obtain the classical part of the strong coupling pomeron intercept
result (\ref{e.strong}). We will introduce
\begin{equation}
j_{cl} \equiv \f{j}{\sqrt{\lambda}} \quad\quad \nu_{cl} \equiv \f{\nu}{\sqrt{\lambda}}
\end{equation}
Then the formula (\ref{e.strong}) gives
\begin{equation}
j_{cl}=-2\nu_{cl}^2 -3\nu_{cl}^4- \f{21}{2} \nu_{cl}^6 + \ldots \label{eq:jcl}
\end{equation}
In the next two sections we will show how to reproduce this formula directly from classical string solutions and how to generate many more terms beyond those following from (\ref{e.strong}).
\section{The \ensuremath{\mathfrak{sl}(2,\mathbb{C})}{} subalgebra}
A crucial step in identifying solutions relevant to the BFKL Hamiltonian is the identification of an \ensuremath{\mathfrak{sl}(2,\mathbb{C})}{} subalgebra of the (complexified) conformal group~\cite{BST07,BC}:
\begin{align} J_0&=\tfrac12(-iD+M_{12}) & J_+&=\tfrac12(P_1-iP_2) & J_-&=\tfrac12(K_1+iK_2) \\
\bar J_0&=\tfrac12(-iD-M_{12}) & \bar J_+&=\tfrac12(P_1+iP_2) & \bar J_-&=\tfrac12(K_1-iK_2)\end{align}
where the directions 12 indicate the transverse plane.
One should now look for eigenstates of $J_0,\bar J_0$, which should correspond to the eigenvalues $h+m,\bar h+m'$, respectively, for some integer $m,m'$. By \eqref{eq:sl2labels}, this gives the following values for the charges:
\begin{align} -iD=1+2i\nu+m+m' \qquad M_{12}=n \end{align}
Another crucial identification, following from the relation between the BFKL Hamiltonian and the boost operator in the longitudinal plane, reads~\cite{BST07}
\begin{align} j=-iM_{+-}=-iM_{03}\end{align}
Therefore, one should look for solutions with nonzero $D,M_{03},M_{12}$. However, for simplicity we will restrict ourselves to the case of vanishing conformal spin $n=M_{12}=0$.
\section{Folded string solutions in $AdS_3$}
\label{s.class}
The simplest example of a classical string solution with only two nonzero conserved charges
is the GKP folded string~\cite{GKP} which we briefly review below.
\subsubsection*{The GKP folded string}
We start from the solution for the GKP folded string in $AdS_3$ subspace as parameterized in \cite{Ts10}, which in embedding coordinates reads
\begin{align} Y_0 &= \cosh\rho\sin\kappa t & Y_1 &= \sinh\rho\cos\omega t & Y_3&=0 \nonumber\\
Y_5 &= \cosh\rho\cos\kappa t & Y_2 &= \sinh\rho\sin\omega t & Y_4&=0
\label{e.gkp}
\end{align}
and the Virasoro constraints relate the parameters $\kappa,\omega$ and function $\rho$ as
\begin{align} \omega = \frac2\pi K(k^2) \qquad \kappa = k\omega \qquad \sinh\rho = \frac k {\sqrt{1-k^2}} \cn \left(\omega\sigma+K(k^2) \middle\vert k^2 \right)
\end{align}
The only nonzero charges are equal to
\begin{align} E \equiv S_{50} &= \sqrt\lambda\frac2\pi \frac k {1-k^2} E(k^2) \label{eq:gkpE}\\
S \equiv S_{12} &= \sqrt\lambda\frac2\pi \left( \frac1{1-k^2}E(k^2) - K(k^2) \right) \label{eq:gkpS}\end{align}
The physical range of $k$ is the interval $k \in(0,1)$ which interpolates between the limits of very short ($k\to0$) (small spin) and infintely long ($k\to1$) string (large spin).
\subsubsection*{The BFKL folded string}
As discussed above, we are interested in working with a configuration that has nonzero $S_{03},S_{54}$, and we can obtain one by reshuffling and complexifing the embedding
coordinates with respect to (\ref{e.gkp})
\begin{align} Y_0 &= -i \sinh\rho\sin\omega t & Y_3 &= \phantom{-i}\sinh\rho\cos\omega t & Y_1&=0 \nonumber\\
Y_5 &= \phantom{-i}\cosh\rho\cos\kappa t & Y_4 &= -i \cosh\rho\sin\kappa t & Y_2&=0\end{align}
In the above, the embedding condition ($\eta_{AB} Y^A Y^B=-1$) is preserved, and the Virasoro constraints take an identical form to the GKP case. We dub this solution the \emph{BFKL folded string}.
Now, the nonzero charges are
\begin{align} 2\nu = S_{54} &= - i \sqrt\lambda\frac2\pi \frac k {1-k^2} E(k^2) \label{eq:S54}\\
ij = S_{03} &= \phantom{-}i \sqrt\lambda\frac2\pi \left( \frac1{1-k^2}E(k^2) - K(k^2) \right) \label{eq:S03}\end{align}
where $k=\frac\kappa\omega$ just as before, but we no longer have the physical motivation to restrict this parameter to the real axis. Note that the form of the $S_{54}$ charge is consistent with the relation
\begin{align} L\cdot\tfrac12(P_0+K_0)\cdot L^{-1} = -iD \end{align}
in this case realised by $SO(4,2)$ rotations by $\frac\pi2$ in the embedding space in the planes 13, 02, and 04, or, in an unambiguous way, by
\begin{align} L = \exp \frac\pi2 \left(\frac4{3\sqrt3}(M_{42}+M_{20}+M_{40})+M_{31}\right)\end{align}
with $M_{ab}$ generators of the respective rotations.
The formulas \eqref{eq:S54}-\eqref{eq:S03} allow us to write out the classical intercept to a~priori arbitrary order, by expanding $k$ as a series in $\nu_{cl}$ and substituting it into the formula for $j_{cl}$. The result
\begin{align}j_{cl}=-2\nu_{cl}^2-3\nu_{cl}^4-\frac{21}2\nu_{cl}^6-\frac{391}8\nu_{cl}^8-\frac{8439}{32}\nu_{cl}^{10}+\cdots\end{align}
notably agrees with the expansion \eqref{eq:jcl}.
\begin{figure}[t]\includegraphics[width=\textwidth]{im_s54_paper.png}
\caption{Plot of $\IM S_{54}$ \eqref{eq:S54} with marked points at which the function vanishes.\label{fig:imS54}}\end{figure}
To determine the locus of `physical' $k$ on the complex plane, we need to turn to the reality conditions following from the \ensuremath{\mathfrak{sl}(2,\mathbb{C})}\ representation theory, that is
\begin{align}
S_{54}=2\nu\in\mathbb R
\end{align}
This condition is met for purely imaginary $k$, which is obvious when one notes that then
$K(k^2),E(k^2)\in\mathbb R$. This is, however, not the most general condition, as evidenced in a plot of $\IM S_{54}$ (see \figref{fig:imS54}). Indeed there exist curved paths of physical $k$ that branch from the purely imaginary axis into the complex plane and terminate at the singularities at $k=\pm1$.
Let us now describe in more detail the behaviour of $\nu$. On the imaginary axis in the
upper half plane $\nu$ initialy rises from 0, reaches a maximum value $\nu_*$ and then
falls off to a constant ($\sqrt{\lambda}/\pi$) when $k \to +i\infty$. The intercept is
real and negative, has a minimum at $\nu=\nu_*$ and approches 0 as $k \to +i\infty$.
Of course, from the point of view of \ensuremath{\mathfrak{sl}(2,\mathbb{C})}\ representations we do not expect to have any
bound on $\nu$. And indeed, as shown in \figref{fig:imS54}, we find that real $\nu>\nu_*$
are realized on a curved path of complex $k$. Surprisingly, the intercept acquires
an imaginary part there. Nevertheless, in what follows we will work
only with purely imaginary $k$ as we are mainly interested in smaller values
of the intercept. However we find the whole structure and behaviour of the classical
strong coupling BFKL intercept unexpectedly subtle.
\section{Algebraic curves and particle interpretation}
The current paradigm of integrability for the spectral problem can be understood
as a progression of three basic stages. Firstly, one can analyze a particular operator/string state in terms of the Asymptotic Bethe Ansatz. At this stage one can identify the
momentum carrying Bethe roots with specific particle content of the particular
excited state of the string. In the second stage, one can add the contribution
of leading wrapping corrections through L{\"u}scher formulas. In these formulas,
the particle content is still explicit. Finally, in the exact TBA/FiNLIE formulations
which resum all multiple wrapping terms, the notion of particle content still
plays a fundamental role as specifying driving terms in the integral equations
(or equivalently some specific analytical properties of the relevant functions entering
the equations).
For this reason, the question of particle content of a particular string solution
becomes very important. However this is especially mysterious if one analyses
BFKL physics. There, there are two continuous charges which appear naturally -- $j$ and
$\nu$. It is very difficult to imagine how this can be realized in terms of
a multi-particle state with a definite (integer) number of excitations characteristic
of an integrable worldsheet theory.
In fact, the above puzzle was the main motivation for undertaking the present investigation.
Of course, the notion of deducing the multiparticle state corresponding to a given
classical solution is far from trivial. For the question even to make sense,
we have to deal with an integrable QFT. Fortunately this is the case here.
However, since the theory is nonlinear, extracting the particle picture is still
not trivial. The way this can be done is to use the algebraic curve construction
of classical string solutions which can be viewed as some kind of `nonlinear Fourier
transform' of the classical solution.
Within the algebraic curve construction (also known as finite-gap construction),
the branch cuts of the algebraic curve correspond to the location of Bethe roots of the
Asymptotic Bethe Ansatz, which in turn correspond to excitations of the string.
Moreover, we can identify the momenta and energies of elementary string excitations forming the classical string solution. In the next section we will show how the `BFKL folded string'
solution found earlier can be embedded into the algebraic curve construction.
We will then identify the branch cuts and energies and momenta of the constituents.
Before turning to the concrete examples, let us make here some general comments. First, we will recall --- and adapt to our setting --- the introduction of the particle density.
This is of course very well known (see \cite{KMMZ,KZ,AFS}), but due to some subtle
differences (like lack of any charge on the $S^5$) it is convenient for us to review
the formulas here.
Let us start with the Bethe ansatz equations, given here for the \ensuremath{\mathfrak{sl}(2)}\ sector \cite{BS05}:
\begin{align}
\left(\frac{x_k^+}{x_k^-}\right)^J=\prod^M_{\substack{j=1\\j\neq k}}\left(\frac{x_k^+-x_j^-}{x_k^--x_j^+}\right)^{-1}\frac{1-\frac{g^2}{x_k^+x_j^-}}{1-\frac{g^2}{x_k^-x_j^+}}\sigma^2(x_k,x_j)
\end{align}
with the usual shorthands of $x_j=x(u_j),x_j^\pm=x(u_j\pm\frac i2)$, where $\sigma$ is the dressing phase and the spectral variable $x$ is related to the Bethe roots by the Zhukovsky transformation
\begin{align}
u(x)=x+\frac{g^2}{x}
\end{align}
In the typical setting, the roots are rescaled with $J$, but in our case this charge equals 0 from the outset (so the left-hand side is trivial). Instead, as we work in the strong coupling limit, the variables are rescaled by $g$. The leading order of the logarithm of the Bethe equation reads
\begin{align}
2i\pi n=\frac ig\sum^K_{\substack{j=1\\j\neq k}}\frac{-2x_kx_j(x_kx_j-1)}{(x_k^2-1)(x_j^2-1)(x_k-x_j)}
\end{align}
or, in the continuum limit, integrating along all cuts,
\begin{align}
\pi n=\frac1g\Xint-\frac{-xx'(xx'-1)\rho(u')du'}{(x^2-1)(x'^2-1)(x-x')}
\label{eq:BethePV}
\end{align}
where $du'=(1-\frac1{(x')^2})dx'$ and we have introduced a density of roots in the $u$ variable.
For cuts symmetric under $x \to -x$, the above equation simplifies to give
\begin{equation}
\pi n = -\f{1}{g} \Xint- \f{2 \rho(u') x dx'}{x^2-{x'}^2}
\label{eq:BethePVsym}
\end{equation}
In the above formulas, the density $\rho(u)$ is normalized in the following way:
\begin{equation}
\label{e.rhonorm}
\int \rho(u) du=K
\end{equation}
where the number of roots $K$ coincides with the spin. Care is necessary when the
roots are complex. Then obviously $\rho(u)du$ is real (and positive) along the cut.
Let us now compare the preceding formulas to the ones following from the algebraic curve
construction. The basic algebraic curve approach does not determine the exact location of the cuts, except of course for their endpoints. The unimodularity condition
\begin{align}
p(x-\epsilon)+p(x+\epsilon)=2n\pi
\end{align}
does not fix it either: we are free to flip the sign of the square root of $dp$ over some area, i.e.\ partially bordered by the cut. In order to determine the position of the cut, one identifies
the algebraic curve counterpart of (\ref{e.rhonorm}) and imposes the condition that
the resulting $\rho(u)du$ is \emph{real} along the cut.
Using the asymptotics of the pseudomomentum \cite{KZ}
\begin{align}
p(x\approx0) = \frac{2\pi(S-E)}{\sqrt\lambda} x + \dotsb \quad\quad\quad\quad p(x\approx\infty) = \frac{2\pi(S+E)}{\sqrt\lambda x} + \dotsb
\label{eq:pAsympCharges}
\end{align}
and integrating $p(x)du$ anticlockwise around the cuts
\begin{equation}
\oint_C p(x) du =\oint_C p(x) \left(1-\f{1}{x^2}\right) dx
\end{equation}
and deforming the contour $C$ to encircle $x=0$ and $x=\infty$ one obtains
\begin{align}
\int\disc p\,du&=\left(\oint_0+\oint_\infty\right)p(x)\left(1-\frac1{x^2}\right)dx\notag\\&=2\pi i(-2\pi(S-E)-2\pi(S+E))/\sqrt\lambda=-8\pi^2iS/\sqrt\lambda
\end{align}
Comparing the above equation with (\ref{e.rhonorm}), one can identify the discontinuity of $p$
with the particle density in the $u$ variable:
\begin{align}
\rho=\frac{\sqrt\lambda}{-8\pi^2i}\disc p
\label{eq:densitydef}
\end{align}
The location of the cuts will then follow from the reality of $\rho(u)du$ defined in the above
way.
For completeness, let us recall how to determine the energies and momenta of the constituent
particles. Under assumption that the momenta scale as $p=P/g$, the leading $g\to\infty$ term gives the following relation:
\begin{align}
u=\frac{\sqrt{1+4P^2}}P
\end{align}
leading to the expression for the momentum
\begin{equation}
P(x)=\f{x}{x^2-1}
\end{equation}
Inserting the result in the dispersion relation
\begin{align}
\epsilon(p)=\sqrt{1+16g^2\sin^2\tfrac p2}
\end{align}
gives the magnon energy in the above limit
\begin{align}
\epsilon(x)=\frac{x^2-1}{x^2+1}
\end{align}
Indeed, using the asymptotic properties (\ref{eq:pAsympCharges}) one easily verifies that
\begin{align}
\int \rho(x) \epsilon(x)\,du=\int\rho(x)\frac{x^2+1}{x^2}dx=E
\label{eq:energynorm}
\end{align}
again integrating along all cuts.
\section{Algebraic curve analysis}
The aim of this section is to identify the particle content of the BFKL folded string
from the algebraic curve construction. In particular we would like to find the location
of the branch cuts for physical values of $k$ determined in section~\ref{s.class} (at
least for $\nu<\nu_*$). We will start by discussing the case of the
GKP folded string, where the answer is well known, before proceeding to study the BFKL
folded string.
\subsection{The GKP folded string}
As argued in \cite{JLg12}, the GKP folded string is described by the following
quasi-momentum\footnote{Note that here $J=0$. Although the conserved charges are smooth in $J$, the pseudomomentum $dp$ cannot be obtained
directly as a limit $J \to 0$ of the known $J>0$ expression.}:
\begin{align}
dp = \frac{Ax^2+B}{(x^2-1)\sqrt{(x^2-1)(x^2-a^2)}}dx
\end{align}
The cuts between the sheets extends between 1 and (real) $a$, as well as between $-1,-a$ (see \figref{f:gkpcuts}), and the energy and spin can be extracted from the asymptotic behaviour described in \cite{KZ},
which in this case amounts to
\begin{align}
p(x\approx0) = -\frac Ba x + \dotsb \quad\quad\quad\quad p(x\approx\infty) = \frac Ax + \dotsb
\label{eq:pAsympSigns}
\end{align}
The parameters $A,B$ can be fixed by computing the A-cycle, B-cycle, and $\Gamma$-contour (see \cite{KZ}) integrals of $dp$, which, for any algebraic curve, are respectively defined along contours encircling any given cut, linking two cuts, and linking a cut with infinity. In our case, which consists of just two cuts, the A-cycle can be deformed to a contour along the imaginary axis closed off by an infinite semi-circle at either side. The integral corresponding to the latter part of the contour vanishes, therefore the A-period is equal
to the integral of $dp$ along the imaginary axis that evaluates to
\begin{align}
\oint_A dp = \frac{2i}{a^2-1}\left(-(A+B)E(1-a^2)+(a^2A+B)K(1-a^2)\right)
\label{eq:Acycle}
\end{align}
This expression is required to vanish, yielding an expression for one of the coefficients.
The B-cycle can be deformed to pass through infinity, therefore splitting to a sum of two
$\Gamma$-contours. They in turn will be identical due to evenness of $dp$. We need to evaluate just one of them, and we start with the result of symbolic integration of $dp$:
\begin{align}
p(x) &= -\frac{(A+B)x}{a^2-1}\cdot\frac{\sqrt{x^2-a^2}}{\sqrt{x^2-1}} \notag\\
&+ \tfrac1{a(a^2-1)}\left(a^2(A+B)E(\arcsin x\vert\tfrac1{a^2})+(1-a^2)BF(\arcsin x\vert\tfrac1{a^2})\right)
\label{eq:explicitP}
\end{align}
The $\Gamma$-contour is now equal to the value of this function at any point of the cut, and it is quite easy to manipulate at $x=a$, where the first term vanishes. After applying several identities for elliptic integrals, most notably the Legendre relation (see \cite{HMF10}), as well as the condition obtained from vanishing of \eqref{eq:Acycle}, we arrive at the result
\begin{align}
\int_\Gamma dp = p(a) = \frac\pi2\cdot\frac A {E(1-a^2)-K(1-a^2)} = \pi n
\end{align}
This equation, and the vanishing of the A-period, allows us to fix both coefficients $A$ and $B$ in terms of the position
of the branch point $a$. We obtain
\begin{align}
E &=
\sqrt\lambda\frac n{2\pi} \frac{a-1}a \left(aE(1-\tfrac1{a^2})+K(1-\tfrac1{a^2})\right) \\
S &=
\sqrt\lambda\frac n{2\pi} \frac{a+1}a \left(aE(1-\tfrac1{a^2})-K(1-\tfrac1{a^2})\right)
\end{align}
reproducing a special case of the expressions deduced in \cite{GS$^+$11}. These formulas
allow us to link the algebraic curve parameter $a$ (location of the branch point)
with the parameter $k$ appearing in the classical solution of section~\ref{s.class}, in terms
of which we have expressed the physicality conditions.
Indeed these formulas can be matched (for $n=1$) with \eqref{eq:gkpE}-\eqref{eq:gkpS}
using a (descending) Landen transformation
\begin{align}
a=\frac{1+k}{1-k}
\end{align}
Consequently, since in this case $k\in (0,1)$, $a$ is a real number greater than~1. It is
straightforward to check that the reality condition for $\rho(u) du$ are satisfied
for the cuts being the interval $(1,a)$ and its mirror image.
\begin{figure}\begin{center}
\begin{tikzpicture}[scale=4/3]
\begin{scope}[xslant=1,yscale=0.5]
\draw (-3,-3) rectangle (3,3);
\begin{scope}[every node/.style={fill,circle,inner sep=0.75}]
\draw[thick] (0:1) node{} -- (0:2.2) node{};
\draw[thick] (180:1) node{} -- (180:2.2) node{};
\end{scope}
\draw (0:1) node[left]{$1$};
\draw (0:2.2) node[right]{$a$};
\end{scope}
\end{tikzpicture}
\end{center}\caption{Position of cuts for the GKP regime, $a$ is real and larger than 1.}\label{f:gkpcuts}\end{figure}
\subsection{The BFKL folded string}
Since the BFKL folded string amounts just to a reshuffling and complexification of
the embedding coordinates of the GKP folded string, the preceeding discussion
will follow through, together with the relation for the branch point
\begin{align}
a=\frac{1+k}{1-k}
\end{align}
Now, however, the physical values of $k$ are (for $0\leq\nu\leq \nu_*$) purely imaginary and thus
$a$ will have unit modulus\footnote{A convenient parametrization is $k=i\tan \alpha/2$, which leads to $a=e^{i\alpha}$.}, allowing a guess that in this case the cuts will lie on the unit circle, as shown in \figref{f:bfklcuts}.
To verify this, we need to impose reality conditions on the density $\rho(u) du$. Indeed
one can check that on the unit circle $\disc p$ is purely imaginary, while $du$ is
obviously real (parametrizing the unit circle as $x=e^{i\phi}$) thus satisfying
the reality condition. Note that this location of the branch cut leads to purely
imaginary energies and momenta of the constituent magnons.
\begin{figure}\begin{center}
\begin{tikzpicture}[scale=2]
\begin{scope}[xslant=1,yscale=0.5]
\draw (-2,-2) rectangle (2,2);
\draw[loosely dashed] (0,0) circle (1);
\begin{scope}[every node/.style={fill,circle,inner sep=0.75}]
\draw[thick] (0:1) node{} arc (0:65:1) node{};
\draw[thick] (180:1) node{} arc (180:245:1) node{};
\end{scope}
\draw (0:1) node[below right]{$1$};
\draw (65:1) node[above right]{$a=e^{i\alpha}$};
\end{scope}
\end{tikzpicture}
\end{center}\caption{Position of cuts for the BFKL regime, $a$ lies on the unit circle.}\label{f:bfklcuts}\end{figure}
\begin{figure}[h]\begin{center}
\includegraphics[width=0.7\textwidth]{integrandJ.pdf}
\end{center}\caption{The integrand of \eqref{eq:BFKLdensitynorm}, i.e.\ $\rho(e^{i\phi})\frac{du}{d\phi}/\sqrt\lambda$, for $\alpha=\frac32$ and $\phi\in[0,\alpha]$.}
\label{f:integrandJ}\end{figure}
Let us finally express $\nu$ and the intercept $j$ in terms of integrals of some real density.
By comparing pairwise \eqref{eq:gkpE}-\eqref{eq:gkpS} and \eqref{eq:S54}-\eqref{eq:S03}, we can reformulate the $p$ asymptotics as follows:
\begin{align}p(x\approx0)=\frac{2\pi(j-2i\nu)}{\sqrt\lambda}x+\dotsb\qquad p(x\approx\infty)=\frac{2\pi(j+2i\nu)}{\sqrt\lambda x}+\dotsb\end{align}
The density integrals evaluate thus to
\begin{align}\frac{\sqrt\lambda}{-8\pi^2i}\int\disc p\,du=-\frac{\sqrt\lambda}{2\pi^2}\int_0^\alpha\IM p(e^{i\phi})(1-e^{-2i\phi})ie^{i\phi}d\phi&=j\in\mathbb R\label{eq:BFKLdensitynorm}\\
\frac{\sqrt\lambda}{-8\pi^2i}\int\disc p\,\epsilon\,du=-\frac{\sqrt\lambda}{2\pi^2}\int_0^\alpha\IM p(e^{i\phi})(1+e^{-2i\phi})ie^{i\phi}d\phi&=2i\nu\in i\mathbb R\end{align}
The integrand of the former is plotted in \figref{f:integrandJ}.
One can also numerically verify the continuum Bethe equations \eqref{eq:BethePVsym}.
Thus on the classical level there does not seem to be a problem with a particle
description of the BFKL folded string solution. There is just one subtlety.
The normalization of the density to the spin as in \eqref{e.rhonorm} suggests that the particle density is in this case negative definite.
This might indicate a need for some kind of continuum Baxter equation interpretation,
however the difference seems so minor that it could be undone by redefining
the quantum numbers carried by elementary magnons (recall that we are not in the
standard \ensuremath{\mathfrak{sl}(2)}\ sector now but in a reshuffled/complexified version).
Classical considerations are not enough to unambigously settle this issue.
\section{Conclusions}
In this paper we have used the identification between quantum numbers of
the BFKL pomeron and specific charges of the complexified conformal group
to bypass the neccessity of performing analytical continuation of
anomalous dimensions of specific operators in order to study high energy
scattering physics.
In the strong coupling regime, we have identified classical string solutions
with the relevant conserved charges. We have restricted ourselves, for simplicity,
to the case of zero conformal spin where the results could be cross checked with
the results obtained using analytical continuation of anomalous dimensions
of twist-2 operators. However the same methods could be applied to the case
of nonzero conformal spin, where it is not known \emph{a~priori}
what is the precise class of operators which would be linked by analytical
continuation. In addition, other non-leading string solutions might be
the counterparts of compound multi-reggeized gluon states. For these
other states the link with anomalous dimensions of specific operators
is unknown as well.
The relevant classical solution for vanishing conformal spin -- the BFKL
folded string -- is a very close cousin of the well known GKP folded string.
It immediately reproduces the classical part of the strong coupling pomeron
intercept.
Our main focus, however, was on investigating the integrable properties of this
solution and its particle-like description. We have analyzed the algebraic
curve construction of this solution and identified the locations of the cuts
which, for a wide range of parameters, lie on two arcs on the unit circle.
The solution is consistent with a real density of particles (at the classical
level, its sign could be perhaps absorbed in a redefinition of the charges of the
elementary magnons), so at the classical level the integrable description
of this solution does not exibit any marked pathologies.
This is in contrast to the weak coupling case, where the Baxter equation
description is very nonstandard~\cite{RJBAXTER}.
There are numerous avenues for further research. Firstly, one could
investigate more general solutions with nonzero conformal spin and find
the classical intercept in these cases. Secondly, it would be very interesting
to perform a short string limit along the lines of \cite{GS$^+$11,RT}.
Finally, the direct algebraic curve description of a classical solution
relevant for BFKL might be useful in formulating a Y-system approach.
\paragraph{Acknowledgments}
This work is supported by the International PhD Proj\-ects Programme of the Foundation for Polish Science within the European Regional Development Fund of the European Union, agreement no.\ MPD\hspace{0pt}/\hspace{0pt}2009\hspace{0pt}/\hspace{0pt}6. RJ was supported by NCN grant 2012/06/A/ST2/00396. PLG thanks the Laboratory of Theoretical Physics of École Normale Superieure in Paris for hospitality during the period when a part of this work has been performed. RJ thanks IPhT Saclay for hospitality when this work
was finalized.
|
\section{Conclusions}
\label{sec:topconclusions}
This is the concluding section for top quark snowmass 2013 studies. We have discussed six topics --
the top quark mass, top quark couplings to other SM particles, kinematics of top-like final states,
rare decays of top quarks and top quark physics beyond the Standard Model. We will describe
our conclusions for each of these topics.
We have argued that a theoretically clean measurement of the top mass to about 300~MeV
is sufficient for many of the physics goals that are currently discussed, in particular
electroweak precision fits. If no new physics is
found at the LHC, it will be important to address the vacuum stability issue of the SM. To
address this, a top mass measurement with a precision of 100~MeV is required, given the
expected precision of the Higgs mass measurement.
The top quark mass can be measured with an accuracy of about 500~MeV in individual measurements
at the LHC, and their combination might reduce the uncertainty further. We note that both novel
methods and the high-luminosity option are required for achieving this accuracy.
The top mass can be measured with an accuracy of about 100~MeV
(dominated by theoretical uncertainties) at a lepton collider,
which matches well with the precision on the $W$~mass achievable at such a facility.
While the LHC and a future linear collider provide complementary information on top quark
couplings, there is no doubt that the LHC, especially the high-luminosity option,
will probe a majority of top quark
couplings to gluons, photons, $Z$'s, $W$'s and the Higgs boson with precision that should
allow us to detect deviations caused by generic BSM physics at the TeV scale.
The much higher precision achievable at a linear collider should then either allow us to study
these deviations or exclude the existence of generic BSM physics at even higher scales, in
particular for the $\gamma$ and $Z$~couplings.
The top Yukawa coupling, one of the most
important top couplings, will be measured to roughly equal precision at the LHC and the 500~GeV
ILC and to better precision at a high-energy linear collider.
Understanding how top quarks are produced and decay is an integral part of top physics at any collider.
Kinematic distributions and differential cross sections are the key to achieving this goal.
The measurement of basic top observables will help improve modeling of top quark events.
The large top event samples available in the future will allow the study of new observables
such as angular correlations or asymmetries that can uncover subtle new physics effects which
may not be accessible otherwise. We expect the LHC may be able to resolve the Tevatron $A_{FB}$ discrepancy.
The LHC and a future linear collider are complementary in probing rare decays of the top quark.
The LHC is better at probing flavor-changing couplings involving gluons, with about a factor two
improvement in the branching ratio limits expected from the high-luminosity option. A linear
collider is better for processes involving $\gamma$'s and $Z$'s. If rare decays are found, a
linear collider also is able to probe the spin structure of the couplings involved.
Top quarks play a very important role in searches for physics beyond the SM. In particular,
solutions to the hierarchy problem require new particles decaying to top-like final states,
such as stops in SUSY or top partners in other models.
The LHC is able to cover the region of interest up to a few TeV in mass for stops, top-partners and
resonances decaying into top quarks. The high-luminosity option extends the mass reach for these
particles by roughly 50\%.
Given the current limits, only a multi-TeV lepton collider will be able to produce top partners and
resonances directly.
We note that there are stop models that might be difficult to discover at the LHC but can be probed
at a linear collider, for example stealth stops.
The 14~TeV LHC is a complex environment, especially the high pileup of the high-luminosity
option which makes precision measurements of top mass, couplings and kinematic
distributions challenging. Moreover, the 14~TeV LHC provides a large sample of boosted top quarks
for the first time whose decay products can no longer be individually identified using traditional techniques.
Our studies indicate that both of these challenges can be mitigated with algorithm developments
and other improvements, many of which have not been deployed yet for these Snowmass studies in the
high-luminosity scenario.
In particular, many analyses will need to rely on these algorithms in future data collection periods, to maintain sensitivity to new physics processes in the high-mass regime
The experimental environment at a lepton collider does not suffer from
these problems and instead offers an ideal environment for precision top physics;
there are few or no additional interactions per crossing and the detectors are
more fine-grained and have better resolution.
In summary, the LHC and the HL-LHC will in two stages dramatically improve our knowledge of the top quark and extends the
reach for new physics to interesting and relevant regions. A future lepton collider will be able
to study the top quark in even more detail, in particular its mass and couplings.
We are confident that the predictions in this report are conservative and that the experiments
will do better with actual data than predicted here.
\section{Top quark couplings}
\label{sec:topcouplings}
~
The couplings of the top quark to the $W$~and $Z$~bosons, photon, gluon,
and the Higgs boson are explored in this section. It is particularly important to make a direct
measurement of the top quark-Higgs boson Yukawa coupling. Simple estimates suggest that
typical BSM physics at the TeV scale modifies the top quark couplings to gauge bosons
at the percent level~\cite{Juste:2006sv} but, at the same time, larger
${\cal O}(10 \%)$ shifts are still possible (see also discussion in Section~\ref{sec:newphysics}).
Also, our knowledge of the top quark Yukawa
coupling is poor at the moment and the direct measurement of this coupling with
any precision is very important. Modifications of top quark couplings typically
lead to a more complex structure of the interaction vertices, going well beyond simple-minded
re-scaling of SM couplings. This creates additional complications
and requires us to understand how all the different couplings can be disentangled.
We note that most of the top quark couplings are measured by comparing observed {\it rates} of
relevant processes with SM expectations. This puts stringent requirements on theoretical
predictions and experimental control of systematics, making couplings measurements a difficult
endeavor at the LHC.
This section compares the precision reach of couplings measurements
at low-and high-luminosity LHC to that expected at lepton colliders (mainly ILC and CLIC).
Higher-energy hadron colliders are not expected to improve the measurements much beyond the LHC
sensitivity (except possibly for the $t\bar{t}Z$ coupling)
and are thus not studied here. The muon collider allows for the same studies as done
at the ILC, but with smaller beam-related uncertainties and higher luminosity.
TLEP provides larger data samples than the ILC, though only near the $t\bar{t}$ treshold,
and it has insufficient energy to measure
Yukawa coupling through direct $t \bar t H$ production though it should be able to reach a
sensitivity of ${\cal O}(30 \%)$ to the $ttH$ coupling from a threshold scan.
The top quark couplings sensitivity
is compared here using the anomalous coupling notation; a related discussion in terms
of effective operators can be found in Refs.~\cite{AguilarSaavedra:2008zc,Zhang:2010px}.
\subsection{Strong interaction}
The strong coupling constant of the top quark is fixed in the Standard Model by the requirement of
$SU(3)$ color gauge-invariance. The modifications of this coupling can be expected through radiative
corrections which may introduce additional structures, such as chromoelectric and chromomagnetic
dipole operators in the $g t \bar t$ vertex. These modifications occur both in the Standard Model
and in models of new physics.
For example, the Higgs exchange between top quarks modifies the strength of gluon-top quark
interaction in top pair production by ${\cal O}(0.5\%)$ while it does not affect the interaction
of light quarks to gluons.
Strong interactions of the top quark are studied in top quark pair production,
including the $t \bar t +{\rm jets}$ processes, both at the Tevatron and the LHC. A
summary of the current prediction and measurements is shown in Table~\ref{tab:topxs}.
An experimental uncertainty of about $5 \%$ on $\sigma(pp \to t \bar t)$ has been achieved
at the $8~{\rm TeV}$ LHC and it is not expected to significantly improve beyond that during further
LHC operations. The theory prediction for the total cross-section through NNLO QCD is available
\cite{Czakon:2013goa,Baernreuther:2012ws,Czakon:2012zr}; it shows a residual scale uncertainty
of about $3.5\%$, comparable to experimental precision. Note that, at this level
of precision, electroweak corrections may be important; indeed, as shown in a recent
update~\cite{Kuhn:2013zoa}, the weak corrections to $t \bar t$ production at the LHC
are close to $-2.5 \%$. We conclude that, at a few percent level, there is no indication
that strong interactions of top quarks are significantly different from that of light quarks.
More exotic types of modifications of top quark strong interactions, such as chromoelectric $d_t$
and chromomagnetic $\mu_t$
dipole moments of top quarks, are better constrained from changes in kinematic distributions.
We will discuss this in Section~\ref{sec:kinematics}. Ref.~\cite{Baumgart:2012ay} finds that
constraints of 1\% or below are possible with
$100~\mathrm{fb}^{-1}$ at 13~TeV.
Exchanges of axigluons or Kaluza-Klein excitations of gluons not only modify couplings
of top quarks to gluons, but also generate
four-fermion operators that involve light and heavy quarks
$ \left ( \bar q T^{a} q \right ) \; \left ( \bar t T^{a} t \right )$.
These operators can be directly probed at the LHC, where the sensitivity to scales
between $1.2~{\rm TeV}$ and $3~{\rm TeV}$ can be expected~\cite{TopCouplWhitePaper2}.
Finally, top quark coupling to gluons can be probed at a linear collider through a threshold scan.
The peak cross-section
at threshold is proportional to $\sigma_{\rm peak} \sim \alpha_s^3/(m_t \Gamma_t)$. Using the total
cross-section and other measurements at threshold, one can determine the strong coupling
constant with better than one percent precision and the total width of the top quark $\Gamma_t$
with the precision of a few percent~\cite{Martinez:2002st,Seidel:2013sqa}.
~
\begin{table}[!h!tbp]
\begin{center}
\begin{tabular}{|l|c|c||c|c|}
\hline
&\multicolumn{2}{c||}{Theory prediction}& \multicolumn{2}{c|}{LHC Measurement} \\
CM Energy [TeV] & 7 & 8 & 7 & 8 \\
Luminosity [fb$^{-1}$]& & & 1-5 & 2-15 \\
\hline
Top pairs $\sigma(t\bar{t})$ [pb]& $172\pm7$~\cite{Czakon:2013goa}&$246\pm10$~\cite{Czakon:2013goa}&$173\pm10$&$241\pm32$ (ATLAS)~\cite{ATLAS-CONF-2012-149} \\
& & &(LHC comb.)~\cite{ATLAS-CONF-2012-134} &$227\pm15$ (CMS)~\cite{CMS-PAS-TOP-12-007} \\
Single top $\sigma$(t-chan) [pb]& $66\pm2$~\cite{Kidonakis:2012rm} & $87\pm3$~\cite{Kidonakis:2012rm} &$83\pm20$ (ATLAS)~\cite{Aad:2012ux}& $95\pm18$ (ATLAS)~\cite{ATLAS-CONF-2012-132} \\
& & &$67\pm6$ (CMS)~\cite{Chatrchyan:2012ep}&$80\pm13$ (CMS)~\cite{CMS-PAS-TOP-12-011} \\
Single top $\sigma(Wt)$ [pb]&$15.6\pm1.2$~\cite{Kidonakis:2012rm}& $22.2\pm1.5$~\cite{Kidonakis:2012rm}& $16.8\pm5.7$ (ATLAS)~\cite{Aad:2012xca}& $27.2\pm5.8$(ATLAS)~\cite{ATLAS-CONF-2013-100} \\
& & &$16\pm4$ (CMS)~\cite{Chatrchyan:2012zca}& $23.4\pm5.4$ (CMS)~\cite{CMS-PAS-TOP-12-040}\\
\hline
\end{tabular}
\vspace*{0.5cm}
\caption{LHC single top and top pair production cross-section measurements.
}
\label{tab:topxs}
\end{center}
\end{table}
\subsection{Weak interactions: W boson}
~
The coupling of the top quark to the $W$~boson is studied in top quark decays
and in single top quark production at the LHC and the Tevatron, and in top quark
decays at the linear collider. The effective Lagrangian describing the $Wtb$
interaction including operators up to dimension five is~\cite{AguilarSaavedra:2008zc}
\begin{eqnarray}
\mathcal{L}&=&-\frac{g}{\sqrt{2}}\bar{b} \gamma^{\mu}
(V_L P_L + V_R P_R) t W_{\mu}^{-}
-\frac{g}{\sqrt{2}} \bar{b} \frac{i\sigma^{\mu\nu} q_{\nu}}{M_W} (g_L P_L + g_R P_R) t W_{\mu}^{-}
+ h.c. \, ,
\label{eq:Wtbcoupling}
\end{eqnarray}
where $M_W$ is the mass of the $W$~boson, $q_{\nu}$ is its four-momentum,
$P_{L,R}=(1 \mp \gamma_5)/2$ are the left- (right-) handed projection operators, and $V_L$
is the left-handed coupling, which in the SM is equal to the
Cabibbo-Kobayashi-Maskawa matrix element $V_{tb}$~\cite{Cabibbo:1963yz}.
The right-handed vector coupling $V_R$ and the
left-and right-handed tensor couplings $g_L$ and $g_R$ may only appear in the SM through radiative
corrections.
The measurement of helicity fractions of $W$~bosons through lepton angular distributions
in top quark decays can distinguish SM-like left-handed vector couplings from right-handed vector
and from left-or right-handed tensor couplings. With the data
collected at the 8~TeV LHC, $V_R, g_L$ and $g_R$ can be constrained to be smaller than $0.1$.
We note that theoretical predictions for $W$-boson helicity fractions in the SM have been
extended to NNLO QCD \cite{Czarnecki:2010gb,Gao:2012ja,Brucherseifer:2013iv} and, therefore, theory
uncertainties on helicity fractions are about one order of magnitude smaller than experimental
one. Measuring the helicity fraction to a similar level at the high-luminosity LHC and beyond
is therefore necessary to obtain the best sensitivity to new physics.
Single top quark production involves the $tWb$ vertex in top quark production and
thus also provides information on the magnitude of the $tWb$ coupling and the
CKM matrix element $|V_{tb}|$. Single top quarks are produced in three different modes: the
``$t$-channel'' mode where a $W$~boson is exchanged between a light quark line and a heavy
quark line, which has the largest cross-section; the ``$Wt$ associated production'' mode where
either the decay or the exchange of a virtual $b$~quark leads to the final state of a top quark
and a $W$~boson, with the next-to-largest cross-section; and the ``$s$-channel'' production
and decay of a virtual $W$~boson, which has a very small cross-section. The LHC
cross-section measurements for $t$-channel and $Wt$ together with the corresponding prediction
are shown in Table~\ref{tab:topxs}.
The three modes have different sensitivities to new physics and anomalous couplings.
LHC measurements of single top quark production, in particular in the $t$-channel
mode, are also sensitive to off-diagonal CKM matrix elements~\cite{Lacker:2012ek}. The
single top production cross-section measurement already is dominated by systematic
uncertainties~\cite{Aad:2012ux,Chatrchyan:2012ep,Aad:2012xca}, and
the situation is not expected to improve much at higher energies or with larger
datasets. The ultimate cross-section uncertainty will likely be around 5\%, similar to
top pair production, so that uncertainties on $tWb$ coupling and $|V_{tb}|$ will be close to
2.5\%~\cite{TchanWhitePaper}. Searches for anomalous couplings in the $tWb$ vertex depend on the
ability to separate the signal from both SM single top and from large backgrounds and are less
limited by systematic uncertainties.
A search for CP violation through an anomalous coupling gives a limit on
$Im(g_R)$~\cite{ATLAS-CONF-2013-032}. Finally, an extrapolation of the sensitivity to anomalous
couplings from single top quark production and decay shows that with 300~fb$^{-1}$ the anomalous
couplings as small as $0.01$ can be probed.
Electron-positron colliders are expected to do a comparable job in exploring the strength of
$tWb$ interaction vertex by considering the cross-section scan of
$\sigma_{t b W}$ cross-section at CM energies between $m_t$ and $2m_t$. It was estimated
in Ref.~\cite{Juste:2006sv} that $g_{tWb}$ can be measured with the precision of about two percent.
Among more exotic options is the possibility to study $tWb$ interaction
at a $\gamma e$ collider, with a reach of
$10^{-1}$ to $10^{-2}$~\cite{Boos:2001sj}. The reach is about $10^{-3}$ to
$10^{-2}$ for a LHC-based electron-proton collider with a CM energy of
1.3~TeV~\cite{Dutta:2013mva}.
Knowledge of $tWb$ interaction can be used to compute the top quark decay width $\Gamma_t$.
This can be compared to direct measurements of $\Gamma_t$, discussed in Section~\ref{sec:topmass}.
\subsection{Electroweak interaction: Z boson and photon}
The interaction of the top quark with neutral electroweak gauge bosons has not been studied in
detail so far. Indeed, although both the charge of the top quark~\cite{Aad:2013uza}
and the production cross-section of top pair in association with a photon were measured
experimentally~\cite{ATLAS-CONF-2011-153}, this does not give us all the information
required to fully constrain the $t \bar t \gamma $ vertex. The interaction of top quarks with
the $Z$~boson has not been measured yet. Similarly to other coupling, a measurement with
${\cal O}(10\%)$ precision will
be useful for constraining models of physics beyond the Standard Model. As an example,
Section~\ref{sec:newphysics} discusses compositeness, which would be constrained by a measurement
at this precision.
It is challenging, but perhaps not impossible, to probe $t\bar t Z$ and $t \bar t \gamma$
couplings at the LHC with 10\% precision. A lepton collider would provide even higher precision.
A general expression for $t \bar t V$, $V=\gamma, Z$ interaction vertex is~\cite{Juste:2006sv}
\begin{equation}
\Gamma_\mu^{ttX}=ie\left\{-\gamma_{\mu}\left((F_{1V}^X+F_{2V}^X) + \gamma_5 F_{1A}^X\right)
+ \frac{(q-\overline{q})_\mu}{2m_t}\left(
F_{2V}^X-i\gamma_5 F_{2A}^X\right) \right\}\,,
\label{eq:tZg}
\end{equation}
where $X$ is either a photon ($X=\gamma$) or $Z$~boson ($X=Z$). The couplings
$F_{1V}^\gamma$, $F_{1V}^Z$ and $F_{1A}^Z$ have tree-level SM values.
The LHC experiments have measured the production of photons in association with
top quark pairs, and will measure both the $\gamma+t\overline{t}$ and
$Z+t\overline{t}$ cross-sections. However, in both cases, significant kinematic
cuts on final state particles are required either to suppress the backgrounds or, in case
of photons, to select events where photons are emitted from top quarks rather than from their
decay products \cite{Baur:2005wi,Baur:2004uw,Juste:2006sv}.
Therefore, extracting the top-photon or top-$Z$ coupling from the associated production
is difficult;
it relies on a detailed theoretical understanding of the production
process. This is becoming available thanks to recent studies of
$pp \to t \bar t \gamma$ and $pp \to t \bar t Z$ processes in next-to-leading order in
QCD~\cite{Lazopoulos:2008de,Melnikov:2011ta,Garzelli:2012bn,Garzelli:2011is}.
Single top quark production in association with a $Z$~boson can also
be used to study the $tZ$ coupling~\cite{Campbell:2013yla}.
Measurements of the $t \bar t \gamma$ and $t \bar t Z$ couplings with the
highest precision can be performed at a linear collider~\cite{Baer:2013cma}.
The two couplings are entangled in the top pair production process. Separating the two couplings
requires polarized beams. For the projections
in Table~\ref{tab:topZxs}, electron and positron polarizations of 80\% and 30\%,
respectively, are assumed. It follows from Table~\ref{tab:topZxs} that
most of the top quark couplings to the photon and the $Z$~boson can be measured
at a linear collider (ILC/CLIC) to a precision that is typically an order of magnitude better than
at the LHC. Despite the lack of detailed studies,
the precision on the combined coupling accessible at TLEP is expected to be even better than
that at the linear collider due to the higher integrated luminosity. However, the lower energy
and lack of beam polarization make it impossible to disentangle the $\gamma$ and $Z$~couplings
and the different couplings in Eq.~\ref{eq:tZg}. Similarly,
detailed studies also have not been performed for a muon collider, which provides larger
integrated luminosity and smaller beam uncertainties but also
challenging backgrounds; thus it is not clear if it will be able to improve on the linear collider
measurements.
In summary, although a linear collider will achieve the highest precision in the $t \bar t Z$
and $t \bar t \gamma$ coupling measurements, it is clear that the LHC -- and in particular its
high-luminosity
phase -- will be able to probe these couplings in an interesting precision range where deviations
due to generic BSM physics are expected.
~
\begin{table}[!h!tbp]
\begin{center}
\begin{tabular}{|l|c|c|c|c|c|c|c|c|}
\hline
Collider & \multicolumn{2}{c|}{LHC}& ILC/CLIC \\
CM Energy [TeV] & 14 & 14 & 0.5 \\
Luminosity [fb$^{-1}$] & 300 & 3000 & 500 \\
\hline
SM Couplings & & & \\
~~photon, $F_{1V}^\gamma$ (0.666) & 0.042& 0.014 & 0.002 \\
~~$Z$ boson, $F_{1V}^Z$ ( 0.24) & 0.50 & 0.17 & 0.003 \\
~~$Z$ boson, $F_{1A}^Z$ (0.6) & 0.058& -- & 0.005 \\
Non-SM couplings & & & \\
~~photon, $F_{1A}^\gamma$ & 0.05 & -- & -- \\
~~photon, $F_{2V}^\gamma$ & 0.037& 0.025 & 0.003 \\
~~photon, $F_{2A}^\gamma$ & 0.017& 0.011 & 0.007 \\
~~$Z$ boson, $F_{2V}^Z$ & 0.25 & 0.17 & 0.006 \\
~~$Z$ boson, $ReF_{2A}^Z $& 0.35 & 0.25 & 0.008 \\
~~$Z$ boson, $ImF_{2A}^Z $& 0.035& 0.025 & 0.015 \\
\hline
\end{tabular}
\vspace*{0.5cm}
\caption{Expected precision of the
top quark coupling measurements to the photon and the $Z$~boson at the
LHC~\cite{TopCouplWhitePaper1,TopCouplWhitePaper2} and the linear
collider~\cite{Baer:2013cma}. Expected magnitude of such couplings
in the SM is shown in brackets. Note that the ``non-standard model'' couplings appear
in the Standard Model through radiative corrections; their expected magnitude, therefore,
is $10^{-2}$.
}
\label{tab:topZxs}
\end{center}
\end{table}
\subsection{Yukawa coupling}
The coupling of the top quark to the Higgs boson is of great interest. Since the
top quark provides one of the largest contributions to the mass shift of the
Higgs boson, any deviation in the $t\bar{t}H$ coupling from its Standard Model value may have
far-reaching consequences for the naturalness problem.
The coupling of the top quark to the Higgs boson can be measured at the LHC
in different final states. It will also be studied in detail at lepton colliders. More
details on the top Yukawa coupling measurements can be found in the Higgs working
group report~\cite{HiggsWhitepaper}.
The process $pp \to t \bar t H$ can be studied in a variety of final states,
depending on the top quark decay mode (lepton+jets or dilepton or all-jets) and the Higgs decay
mode ($b\overline{b}$, $\gamma\gamma$, $WW$ {\rm etc.}).
Each final state has a its own, typically large, background, mainly from top
quark pair production in association with jets or electroweak bosons. The coupling of the top
quark to the Higgs boson is extracted from these measurements with relatively large
uncertainties of about 20\% initially, with an improvement to 10\% at the high-luminosity
LHC~\cite{PhysRevD.86.073009,Onyisi:2013gta,CMS-NOTE-2012-006,ATLAS-collaboration:1484890,TopCouplWhitePaper1,TopCouplWhitePaper2}.
At the high-luminosity LHC, the $t\overline{t}H$ final state is also a promising channel
to measure the muon coupling of the Higgs boson~\cite{TtHmumuWhitePaper}, though it is still
statistics-limited. Production of $t\overline{t}H$ with Higgs decay to photons is observable
at the LHC~\cite{TopCouplWhitePaper2}, which allows for a study of the CP structure of the
top-Higgs vertex~\cite{Gunion:1996xu}.
Better precision in the top-Higgs coupling can be achieved at lepton colliders running
at a sufficiently high CM energy and collecting large integrated luminosity.
Initial studies focused on a CM energy of 800~GeV where the $t\overline{t}H$ cross-section
is largest, however a measurement at 500~GeV is also possible. For the
projections in Tab.~\ref{tab:topHiggs}, electron and positron polarizations of 80\% and
30\%, respectively, are used. For the ILC/CLIC, the nominal luminosity for 1~TeV running
is assumed, which corresponds to twice the ILC luminosity at 500~GeV.
A comparison of the top Yukawa coupling precision expected at different colliders is
shown in Table~\ref{tab:topHiggs}, from where it follows that a linear collider provides
improvements compared to the high-luminosity LHC. It is also possible to
measure the Yukawa coupling in a threshold scan that is sensitive to the modification of the
$t \bar t$ production cross-section through a Higgs exchange. A precision
of ${\cal O}(30)\%$ can, perhaps, be achieved in this case. Note that this is the only way
to get information on the top Yukawa coupling at TLEP.
\begin{table}[!h!tbp]
\begin{center}
\begin{tabular}{|l|c|c|c|c|c|c|c|c|c|}
\hline
Collider & \multicolumn{2}{c|}{LHC}& ILC & ILC & CLIC \\
CM Energy [TeV] & 14 & 14 & 0.5 & 1.0 & 1.4 \\
Luminosity [fb$^{-1}$] & 300 & 3000 & 1000 & 1000 & 1500 \\
\hline
Top Yukawa coupling $\kappa_t$& ($14-15$)\%& ($7-10$)\% & 10\% & 4\% & 4\%\\
\hline
\end{tabular}
\vspace*{0.5cm}
\caption{Expected precision of the top quark Yukawa coupling measurement
expected at the LHC and the linear collider~\cite{HiggsWhitepaper}. The range
for the LHC precision corresponds to an optimistic scenario where systematic
uncertainties are scaled by 1/2 and a conservative
scenario where systematic uncertainties remain at the 2013
level~\cite{CMS-NOTE-2012-006,ATLAS-collaboration:1484890,Onyisi:2013gta}.
The ILC~\cite{Baer:2013cma,Yonamine:2011jg}
and CLIC~\cite{Abramowicz:2013tzc} projections assume polarized beams and nominal integrated
luminosities.
}
\label{tab:topHiggs}
\end{center}
\end{table}
\subsection{Methods particular to boosted top quarks}
\label{sec:boostedtop}
~
As we explained in Section~\ref{sec:newphysics},
top quarks play a very important role in
many searches for new particles at the highest energies.
We find that current algorithms for top quark identification at high-$p_T$ can lead to
performance that is similar to what is achieved
in current experiments, provided that some modifications to the reconstruction methods are implemented or
detectors upgrades are performed.
\begin{figure}[tb]
\begin{center}
\subfigure[Anti-kt R=0.5]{
\includegraphics[scale=0.32, angle=0]{TopQuark/detectors/figs/z_Reweighted_tt1_MassPt_AK5_Reweighted}
}
\subfigure[Cambridge Aachen R=0.8]{
\includegraphics[scale=0.32, angle=0]{TopQuark/detectors/figs/z_Reweighted_tt1_MassPt_CA8_Reweighted}
}
\end{center}
\caption{
Jet mass vs jet transverse momenta in the {\sc DELPHES} fast simulation for pp collisions at
$14$ TeV for different jet algorithms. The jet transverse spectrum has been reweighed to be flat.
}
\label{fig:jetmass_pileups}
\end{figure}
The decay products of a top quark with high $p_T$ are sufficiently collimated
to be reconstructed within a single jet. This happens above $\sim 400$ GeV for jets with
$R=0.8$. Figure~\ref{fig:jetmass_pileups} shows the evolution of jet mass with the jet
transverse momentum for the $t\bar{t}$ process. Because all of the top decay products
fall within a single jet, specialized techniques involving jet substructure are
required~\cite{Abdesselam:2010pt,Altheimer:2012mn}.
Semileptonic top decay reconstructions must introduce modified isolation criteria when the
lepton starts to overlap with the $b$~quark jet from the top decay.
This reconstruction of the top mass within a single jet itself is a good discriminant
between boosted top quarks and the overwhelming background from QCD jet production. For
example, a recent study~\cite{Auerbach:2013by} has shown that a signal of boosted hadronic
top quarks from a $Z'$~boson decay can be observed in the jet mass distribution alone for jets
with $p_T > 800$~GeV. Discrimination can possibly be improved further with the addition of
$b$-tagging.
The reconstruction of the top jet through its proximity to the mass of the top is the
basic idea behind the boosted top studies. In addition, further signal/background
separation is achieved by using specialized algorithms
that split the top jet into sub-jets, and then examine those to determine if observed jet substructure
is consistent with soft and collinear QCD radiation or with the decay of a heavy object into jets through a
point-like interaction vertex.
{\bf Jet grooming. } Boosted jets are affected by pileup just like the unboosted ones
discussed in Section~\ref{sec:unboosted}. Several algorithms, collectively known as jet
grooming algorithms, attempt to mitigate the effect of pileup on jet observables, such as jet
mass, by removing soft and wide-angle constituents of jets.
The effect of three different jet grooming algorithms have been studied:
pruning~\cite{Ellis:2009me, Ellis:2009su}, trimming~\cite{Krohn:2009th}, and
filtering~\cite{butterworth-2008-100}. The application of these jet grooming algorithms results
in a jet mass distribution that is relatively stable as the number of pileup events
increases. Additionally, the jet grooming procedures significantly reduce the
masses of QCD jets, enhancing signal/background discrimination significantly.
{\bf Substructure and jet shapes.}
Jet substructure and jet shapes are often discussed as a useful tool for the identification
of top quarks and for reduction of the overwhelming rate from conventional QCD processes
\cite{Agashe:2006hk,Lillie:2007yh,Butterworth:2007ke,Almeida:2008tp,Almeida:2008yp,
Kaplan:2008ie,Brooijmans:2008,Butterworth:2009qa,Ellis:2009su,ATL-PHYS-PUB-2009-081,CMS-PAS-JME-09-001,Almeida:2010pa,Hackstein:2010wk,Chekanov:2010vc,Chekanov:2010gv,ATL-PHYS-PUB-2010-008}.
For example, the $N$-subjettiness algorithm~\cite{Nsubj} aims to determine the consistency of a
jet with a hypothesized number of subjets.
Such tools can give good discrimination between top quark jets and QCD jets, however, such
discrimination degrades somewhat with the additional pileup activity.
It is also beneficial to identify the two subjets corresponding to the $W$~boson produced in
the top quark decay. Using trimming, a $W$ mass peak can be extracted which is relatively
stable even with 140 additional pileup events added.
{\bf Top tagging.} In addition to the substructure quantities described above, there are
several algorithms (top taggers) which combine multiple jet observables to identify top jets
and provide additional discrimination from QCD jets.
Two top-tagging algorithms which are currently in use by experimental efforts include the CMS
Top Tagger~\cite{CMS-PAS-JME-09-001,Kaplan:2008ie} and the HEP Top
Tagger~\cite{Plehn:2010st,Aad:2013gja,ATLAS:2012am,Aad:2012dpa,Aad:2012raa}.
The CMS top tagger decomposes a jet into up to 4 subjets.
Then requirements on the jet mass ($140 < m_j < 250$ GeV), number of subjets (3 or more) and
a quantity which is a proxy for the mass of the $W$~boson within the jet (minimum pairwise
subjet mass $>50$ GeV) , are imposed to isolate boosted top quarks.
We have studied the effect of pileup on the efficiency of the CMS top tagger.
With no additional pileup events, the efficiency of the algorithm maintains its maximum value
of $\sim 40$\% up to jet $p_T$ values of 1.2-1.3 TeV, at which point the efficiency begins to
fall to 10\% or lower for jets with $p_T > 1.5$ TeV. This efficiency drop at high $p_T$ can
be ameliorated by increasing calorimeter granularity, but extra radiation off of the top quark
also affects the algorithm in the very high $p_T$ regime. With additional pileup events (and no
correction applied to the subjets), the efficiency degradation happens at much lower $p_T$
values.
The rate of QCD jets passing the algorithm is also affected. With no additional pileup events,
this mistag rate remains below 5\% over the entire range of jet $p_T$. After adding 140
pileup interactions to the simulated events, the mistag rate from QCD jets increases to a
maximum of 45\% at a $p_T$ of 500~GeV, though this can be reduced through additional algorithm
improvements. However, above approximately $p_T > 1$ TeV we expect that there will be minimal
impact after pile-up corrections.
{\bf Detector effects.} At large values of the top quark $p_T$, such as the region above
1.5~TeV at the LHC, QCD radiation as well as the size of the detector elements become
a limiting factor. In this regime, top quarks will have hard radiations that may be
identified as subjets and the top quark decay products become so highly collimated that they
cannot be individually resolved due to calorimeter detector segmentation and tracking failures.
The effects mentioned above cause a degradation in the top quark jet resolution at large $p_T$. For example, the width of the top quark jet mass peak increases by a factor of two when
comparing top quarks with $p_T > 1.6$ TeV to those with $p_T > 0.8$ TeV, see
Fig.~\ref{fig:jetmass_pileupsFit}. Algorithmic improvements extend the $p_T$ range where
top jets can be reconstructed, but ultimately the granularity of individual calorimeter cells
must be increased to maintain a good top jet reconstruction.
\begin{figure}[tb]
\begin{center}
\subfigure[$p_T(jet)>0.8$ TeV]{
\includegraphics[scale=0.32, angle=0]{TopQuark/detectors/figs/ttbar_cball_mu140_p1}
}
\subfigure[$p_T(jet)>1.6$ TeV]{
\includegraphics[scale=0.32, angle=0]{TopQuark/detectors/figs/ttbar_cball_mu140_p3}
}
\end{center}
\caption{
Jet mass for $t\bar{t}$ events for different $p_{T}(\mathrm{jet})$ and $\langle \mu \rangle=140$.
The core of the peak was fitted using a
Crystal Ball function~\cite{Oreglia}.
All histograms are normalized to 1000 events.
}
\label{fig:jetmass_pileupsFit}
\end{figure}
The reconstruction of top jets and substructure within large cone-size jets is a relatively
new field that has made tremendous progress in only a few years. More improvements are likely
to come, especially as sizable top quark event samples at the highest momenta become available
at the LHC. The ultimate limit is expected to come from the detector resolution, particularly
from calorimeter granularity, and future
detectors such as for CLIC or VLHC machines will need account for this.
\section{Top quarks and detectors}
\label{sec:detector}
\include{unboosted_summary}
\include{boosted_summary}
\subsection{Lepton colliders}
~
A lepton collider (linear $e^+e^-$ colliders ILC and CLIC and circular $e^+e^-$ collider TLEP
and the $\mu^+\mu^-$ collider) will allow for the study of electroweak production of
$t\overline{t}$ pairs with no concurring QCD background. Linear colliders can use polarized
beams, giving samples enriched in top quarks of left- or right-handed helicities. This can
allow one to probe new physics scenarios predicting anomalous production rates of
right-handed $t$~quarks compared to the SM, and to disentangle the $t\gamma$ and $tZ$
couplings, see Section~\ref{sec:topcouplings}.
Due to the electroweak production mechanism, all interesting processes occur at roughly the
same rate, and backgrounds can easily be reduced to a negligible level. After applying selection
cuts, it is possible to retain a signal sample of approximately 10$^5$ events at the
500~GeV linear collider with 500~fb$^{-1}$ of integrated luminosity.
Unlike at the LHC, there are no or few additional interactions (pileup) per beam crossing,
especially for the ILC. Additional activity may come from $\gamma \gamma$ interactions.
Ongoing studies show that this residual pile-up can be controlled when applying the invariant
$k_t$ jet algorithm~\cite{Catani:1993hr,Ellis:1993tq} for background
suppression~\cite{WEUSTE:1499132}.
The lepton collider detectors can be designed to be more fine-grained and to have better resolution than the
LHC detectors. The charge of the $b$~quark will be measured at a purity of 60\% and
better~\cite{Devetak:2010na}. This is indispensable for the measurement of $A^t_{FB}$ in fully
hadronic decays, see Section~\ref{sec:kinematics}. The jet energy resolution for LHC detectors
is between 10\% and 15\% for jets below 100~GeV~\cite{Aad:2012ag} whereas it is below 4\%
at the linear collider~\cite{Baer:2013cma}. This results in a clean top quark sample with
a narrow reconstructed mass as shown in Fig.~\ref{fig:TopMass}.
Using $A^t_{FB}$, the top-Higgs coupling $\lambda_t$ and the $t\overline{t}$ production
cross-section, electroweak couplings can be determined at the percent level. It is important that
experimental and theoretical errors are kept at the same level. This requires a precise
measurement of the luminosity and the beam polarization. Currently, both
parameters are expected to be controlled to better than 0.5\% at the linear collider.
In general the realization of machine and detectors must not compromise the precision physics.
This may be the biggest challenge in the coming years.
\subsection{Top quark identification at low transverse momentum}
\label{sec:unboosted}
~
Many of the top quarks produced at the LHC have low transverse momenta, where
$p_\perp$ is in the range $25-50$~GeV. Measurements of
the total and differential $t\bar{t}$ cross-sections (Sections~\ref{sec:kinematics}
and~\ref{sec:topcouplings}), of the top-quark mass (Section~\ref{sec:topmass}),
charge asymmetry (Section~\ref{sec:kinematics}), and single-top measurements
(Section~\ref{sec:topcouplings})
all require precise and efficient reconstruction of top quarks at low transverse momenta.
Top-quark reconstruction at low transverse energies is limited by a
number of factors that determine total systematic uncertainty, including:
a) jet-energy scale uncertainty which typically accounts for
$50\%$ of the overall uncertainty in traditional top-quark measurements based on jets;
b) jet-energy resolution uncertainty;
c) $b$-tagging efficiency uncertainty and mistag rates; and
d) uncertainty on missing transverse-energy reconstruction.
This indicates that progress in precision top measurements that involve jet
reconstruction at low $p_\perp$ will require a better
understanding of low-$p_{T}$ jets and $b-$tagging.
The high-luminosity upgrade of the LHC will have an important impact on low-$p_\perp$ top
physics. In the current design, we expect more than $100$ pileup events per bunch crossing, which
will have a negative impact on many final-state observables, particularly on low-$p_T$ jets and
$b$-tagged jets due to their large associated systematics.
Studies of this scenario \cite{TopAlgWhitePaper} were performed
for $pp$ collisions at a center-of-mass energy of $\sqrt{s}=14$~TeV using
a fast detector simulation
based on the {\sc Delphes} 3.08 framework~\cite{Ovyn:2009tx}.
Jets are reconstructed at the LHC using the anti-$k_T$ algorithm~\cite{Cacciari:2008gp}
with distance parameter $R=0.4$ (ATLAS) and $R=0.5$ (CMS and Snowmass-specific studies).
These high-luminosity MC simulation studies showed that, in general, pileup events
deposit energy in many calorimeter cells and hence shift the raw jet transverse
energies by approximately $50~(120)$~GeV for $50~(140)$ pileup events,
adding about one additional GeV for each pileup event. This energy needs to be subtracted
jet-by-jet using average energies deposited elsewhere in the calorimeter. Tracking
in jet reconstruction is also useful, not only to refine the jet energy measurement but
also to mitigate the impact of pileup.
Nevertheless, the subtraction of pileup results in smeared jet transverse momenta.
In addition, there will be many
low-$p_T$ fake jets created from pileup events. While tracking can be used to address
some of these issues as well, pileup also creates many additional tracks that need to be separated
from the tracks belonging to each jet in an event.
Figure~\ref{pileupjets}(a) shows the effect of different pileup scenarios on the jet $p_T$
distribution.
One consequence of the energy shift is that for the selection of top quark signal jets,
pileup subtraction techniques will likely correct energies of the signal jets by 200-400\%, leading
to larger uncertainties compared to previous analyses.
Uncertainties due to pileup will become dominant, and are expected to increase by a factor
of two or more at the highest LHC luminosity.
As an example, a $2\%$ jet-energy scale uncertainty
for a jet measurement without pileup
translates to a $3\%$($5\%)$ uncertainty in case of 50 (140) pileup events scenario.
\begin{figure}
\begin{center}
\subfigure[]{
\includegraphics[scale=0.35, angle=0]{TopQuark/detectors/figs/jetpt_mu0_pileup}
}
\subfigure[]{
\includegraphics[scale=0.33, angle=0]{TopQuark/detectors/figs/top_massmass}
}
\caption{(a) Plots of jet $p_T$ distributions for different pileup scenarios using the {\sc Delphes} simulation. Also
shown are only the jets matched to the top quarks in the event for each pileup scenario, demonstrating
the large effect of additional pileup events on top quark reconstruction. (b) Reconstructed top quark masses from trijets by requiring at least four jets with $p_T >25$~GeV and $|\eta|<$ 2.5, and at least one of the jets must be tagged by a $b$-tagging algorithm.}
\label{pileupjets}
\end{center}
\end{figure}
Since uncertainties in jet resolution, jet energy scales,
and $b$-tagging are dominant uncertainties in many measurements related to top quarks,
it is to be expected that precision of such measurements will not improve at higher
luminosities and will deteriorate unless new jet energy calibration methods are adopted.
While data-driven techniques are likely to improve the assessment of the jet energy scale,
it is unlikely that this can make a significant difference to the above conclusion. As
the result, the standard top mass measurements do not improve at the high-luminosity
phase of the LHC. It should be noted, however,
that new techniques which are less reliant on precise knowledge of jet energies
will be able to take advantage
of the high statistics of the high-luminosity LHC, as we discuss in Section~\ref{sec:topmass}.
The reconstruction of the top quark mass that is used in many other top quark analyses
will also be degraded by the high pileup in high-luminosity runs.
A {\sc Delphes} MC study shows that using the trijet mass for top-reconstruction is strongly
affected by pileup events even when particle-flow methods and pile-up subtraction
techniques are used to mitigate the problem~\cite{TopAlgWhitePaper}. Figure~\ref{pileupjets}(b) shows
the reconstructed top mass using a procedure similar to the one discussed
in~\cite{ATLAS-CONF-2011-120}.
It was also observed~\cite{TopAlgWhitePaper} that
the trijet mass for top-reconstruction strongly depends on top transverse momentum $p_{T}$
due to large jet multiplicity from ISR/FSR.
For $p_{T}>$700~GeV, the peak position is at 400~GeV, assuming the same transverse
momentum cuts as for low-$p_T$ measurements. This may limit our ability to identify top quarks
at such large $p_{T}$ using the traditional low-energy approaches.
Runs at high pileup will also affect other top physics measurements,
such as $t(\bar{t})$+jets and associated top production (such as $Ht\bar{t}$),
discussed in Section~\ref{sec:topcouplings},
as well as searches for new physics that require a good understanding of
low-$p_{T}$ top quarks, for example searches for rare top decays
(Section~\ref{sec:raredecays}).
Indeed, low-$p_T$ top quarks require the reconstruction of jets
with transverse momentum $30-100$~GeV, which are exactly the jets that are difficult to correct for pileup
effects. However, for rare decays or other counting measurements, it may not be
so necessary to determine jet energy scales as precisely as measurements using kinematic
shapes, so the advantage of the large statistics may very well dominate. Nevertheless,
these measurements will be affected by the reduced performance of $b$-tagging
at high pileup, so work will be needed in this area compared to existing algorithms.
In conclusion, we find that
the high-luminosity $pp$ collision runs at 14~TeV with more than hundred pile-up events
are unfavorable for high-precision top quark measurements based on jets with
transverse momenta below 100~GeV.
This conclusion will affect $t\bar{t}$ measurements with top quarks produced
near threshold that rely on precise
knowledge of jet energies, but will affect rate-dependent measurements to a lesser extent,
especially with improvements in $b$-tagging algorithms.
We believe that a combination of multiple measurements by CMS and ATLAS may lead to a
reduction of systematic uncertainties even in the high pile-up environment.
As will be discussed below, the high-luminosity LHC runs will affect studies of
high-$p_{T}(\mathrm{top})$ top quarks to a lesser extent.
It is therefore important to discuss the future of boosted top measurements, where additional reconstruction
techniques can be utilized.
\section{Introduction}
\label{sec:topintro}
The top quark was discovered in 1995~\cite{Abe:1995hr,Abachi:1995iq} and it is still the heaviest
elementary particle known today. Thanks to its large mass, and the related strength of its
coupling to the Higgs boson, the top quark may be a key player in
understanding the details of electroweak symmetry breaking.
Studies of the top quark properties at the Tevatron and Run I of the LHC have given
us a detailed understanding of many properties of this particle, including its mass,
production and decay mechanisms, electric charge and more.
With the exception of the large forward-backward asymmetry in $t \bar t$ production that
has been observed at the Tevatron, all results on top quark pairs and single top production
obtained so far have been consistent with the Standard Model. We note that in this context,
the anomaly in the $b$ quark forward-backward asymmetry observed at LEP might get amplified for the much heavier top quark.
In the short and mid-term future, top quark studies will be mainly driven by the LHC experiments.
Exploration of top quarks will, however, be an integral part of particle physics studies at any
future facility. Future lepton colliders will have a rich top quark physics program which would
add to our understanding of this interesting quark.
Detailed simulation studies have been carried out for linear electron-positron machines
(ILC and CLIC). First attempts have been made to extrapolate these to the case of a circular
machine (TLEP). In this report we describe what can be achieved based on projection studies
for the LHC and for future lepton colliders. The report is organized along six topics:
\begin{itemize}
\item Measurement of the top quark mass;
\item Studies of kinematic distributions of top-like final states;
\item Measurements of top quark couplings;
\item Searches for rare decays of top quarks;
\item Probing physics beyond the Standard Model with top quarks;
\item Algorithms and detectors for top quark identification at future facilities.
\end{itemize}
Main conclusions for each topic are presented in Sect.\ref{sec:topconclusions}.
\section{Kinematics of top-like final states}
\label{sec:kinematics}
~
Working with top quarks requires us to understand how they are produced
and how they decay. In this section, we discuss what we know about that and what we can learn
in the future. While such a discussion is interesting in its own right, it also allows us to understand
to what extent deviations from expected behavior of various top quark distributions
in different kinematic regimes can be probed at existing and future facilities. In general, after the run I
of the LHC and the studies of top quark pair production at the Tevatron, it is fair to say that
dynamics of $t \bar t$ production is well-understood. The only, but significant,
discrepancy that exists is the disagreement between forward-backward asymmetry for top quarks
expected in the Standard Model and the measured value of this asymmetry at the Tevatron. Is it
possible to clarify the situation with forward-backward asymmetry at the LHC or other future
facilities? This is a data-motivated question that we address in this section.
\subsection{Kinematic distributions in top quark pair production}
Our current understanding of top quark pair production in hadron collisions is based
on next-to-leading order computations for the fully-differential process
$pp \to t \bar t \to W^+W^- b \bar b$ both within and beyond the narrow width
approximation~\cite{Denner:2012yc,Bevilacqua:2010qb,Melnikov:2009dn,Campbell:2012uf}.
The comparison of these
computations ensures that the narrow width approximation works very well at the LHC unless one
moves to extreme kinematic regimes where production of two on-shell top quarks becomes
kinematically unfavorable. The success of the narrow width approximation in $t \bar t$ production
allows us to claim its validity for more complicated processes, such us production of top quark pairs
in association with jets \cite{Melnikov:2010iu,Dittmaier:2008uj,Kardos:2011qa}
or with gauge bosons, that we will discuss in the next Section.
Existing theoretical results on top quark pair production
will be further improved by extending available results for differential
quantities to next-to-next-to-leading order in perturbative QCD. We note that
such results for the total cross-section $pp \to t \bar t $ were recently obtained
\cite{Czakon:2013goa,Baernreuther:2012ws,Czakon:2012zr}.
We will now take a closer look at the quality of theoretical description of various kinematic
distributions. To this end, we show distributions in the top quark transverse momentum
$p_\perp$ in $pp \to t \bar t$ at the 14~TeV LHC in Fig.~\ref{fig:topkin-basicdistr}
and indicate the uncertainties in the predictions caused by imperfect knowledge of parton
distribution functions and missing higher-order corrections. We estimate the latter by varying
renormalization and factorization scales by a factor of two around the fixed value $\mu = m_t$.
The computations are performed with MCFM~\cite{Campbell:2010ff}.
We
see that scale uncertainties dominate and that uncertainties in theory predictions
are at the level of 20\%.
\begin{figure}[!h!tbp]
\centering
\label{fig:topkin-basicdistr}
\includegraphics[width=0.7\hsize]{TopQuark/kinematics/LHC14_NLO_notboosted_pT}
\includegraphics[width=0.7\hsize]{TopQuark/kinematics/LHC14_NLO_boosted_pT}
\caption{NLO QCD predictions \cite{Campbell:2010ff} for the transverse momentum of the top
quark at the $14$~TeV LHC.
Blue error bars correspond to the central MSTW pdf set and scale variation by a factor of two around $\mu = m_t$.
Dark red error bands correspond to 1 SD variation of MSTW pdf error sets for fixed renormalization and factorization scale
at $\mu = m_t$. Note that the red and blue bars can be off-set because at NLO the central scale does not necessarily corresponds to the center of the blue bar. In this case, it seems that it is towards the upper value of the blue bar.
}
\end{figure}
Another interesting kinematic regime is when each top quark is produced with a high Lorentz boost (>2), resulting in collimated
decay products that may be clustered into a single jet (``boosted'' topology). As we will see, it is
more difficult to understand the uncertainty in the theoretical prediction for this quantity.
Indeed, a MCFM-based computation shows that for $p_\perp > 800~{\rm GeV}$, the uncertainties on rapidity
and $p_\perp$ distributions roughly double compared to the non-boosted
regime~\cite{TopKine:Snowmass}. However, these uncertainties may be underestimated.
Indeed, resummation computations, either traditional or Soft-Collinear Effective Field Theory (SCET).
point towards additional positive contributions to $p_\perp$ distributions
at high values of the top quark momentum \cite{Ahrens:2011mw,Auerbach:2013by}.
Forthcoming NNLO computations will be required to resolve this issue.
All kinematic distributions in top quark pair production are routinely checked
for signs of new physics. Prominent among them is the distribution in the invariant mass
of a $t \bar t$ pair, which may be significantly modified by the presence of resonances that decay
to top pairs. Theoretical predictions for such distributions exist both in fixed order
QCD and in SCET~\cite{Ahrens:2011mw}; they show theoretical errors between ten and fifteen percent,
depending on $m_{t \bar t}$ {\it and}. Similarly to the $p_\perp$ distribution, these show significant differences
between fixed order and resummed results at large values of $m_{t \bar t}$.
Other kinematic distributions, such as angular correlations between either top quarks
or their decay products, did not lead to studies at the Tevatron because
of low statistics. However, such studies at the LHC will become increasingly important
as the tool to analyze various subtle features of top quark interactions with with both SM
and, hopefully, BSM particles. In the following
subsections, we discuss examples of this, considering top quark spin correlations and the forward-backward $t \bar t$ asymmetry.
\subsection{Top quark spin correlations}
\label{sec:topkin-spincorrel}
~
Spin correlations between $t$ and $ \bar t$ are an interesting feature of
top quark physics, related to the fact that top quark lifetime
is so short that $t(\bar t)$ spin information is transferred to their decay products without being affected by non-perturbative
hadronization effects.
Observable spin correlations are affected by the structure of $g \bar t t$
and $tWb$ interaction vertices.
After the observation of top quark spin correlations at the
Tevatron \cite{Abazov:2011gi} and recently at the LHC
\cite{ATLAS:2012ao,CMS-PAS-TOP-12-004}, experimental analyses will soon
be able to probe spin correlations in detail. Perhaps, they will use spin correlations
as an analysis tool to find and constrain physics beyond the Standard Model.
The cleanest $t\bar t$ samples for the study of spin correlations are the ones with two opposite-sign
leptons in the final state. Spin correlations in this dilepton
mode manifest themselves most prominently in the distribution
of the relative azimuthal angle between the two leptons \cite{Mahlon:2010gw}.
This distribution is robust under higher order corrections and parton showering
effects~\cite{Melnikov:2009dn,Bernreuther:2010ny,Frixione:2007zp}. For standard
acceptance cuts, NLO QCD effects introduce shape changes of at most 20\%.
If additional cuts are applied that enhance spin
correlations, NLO corrections increase the correlation even further.
Electroweak corrections have negligible effect and scale variations
are small because distributions are typically normalized.
On the experimental side, the
reconstruction of the lepton opening angle in the laboratory frame
is straightforward and can be done with small systematic uncertainties. The normalized
azimuthal opening angle distribution is therefore an ideal observable
for studying top quark spin correlations. Other
observables such as helicity angles, double differential
distributions and asymmetries can also be explored.
The utility of top quark spin correlations to search for physics beyond the Standard Model
stems from the vector coupling of top quarks
to gluons, from the fermion nature of the top quark, and from the couplings of a top quark to
a $W$~boson and a $b$~quark through a left-handed vector current. Any changes in that
list must lead to an observable change in the spin correlation pattern.
For example, it has been shown that top quark spin correlations can be used to
distinguish SM top quarks from scalar partners (stops) even if tops and stop
are degenerate in mass \cite{Han:2012fw}. The potential of spin correlations
to distinguish SM top pair production and
stop ( $m_{\tilde t} = 200~{\rm GeV}$ ) pair production is illustrated in
Fig.~\ref{fig:spincor}~\cite{TopKine:Snowmass}.
~
\begin{figure}[!h!tbp]
\centering
\includegraphics[width=0.7\hsize]{TopQuark/kinematics/1205_5808_Fig6}
\caption{Top quark spin correlation angle for top quark production in the SM and without
spin correlation and for stop quark production with different couplings~\cite{TopKine:Snowmass}.
}
\label{fig:spincor}
\end{figure}
Modifications of the $g \bar t t$ vertex, that can be parametrized in
terms of top quark chromomagnetic
$\hat\mu_t$ and electric $\hat d_t$ dipole moments, can be exposed through
spin correlations in the dileptonic and in the semileptonic
channels~\cite{Baumgart:2012ay,Bernreuther:2013aga}.
Here, the magnetic and electric dipole moments $\hat \mu_t$ and $\hat d_t$ correspond to the Lagrangian
$\frac{1}{2} G^a_{\mu\nu} \left( \tilde\mu_t \bar t \sigma^{\mu\nu} T^a t + \tilde d_t \bar t i \sigma^{\mu\nu} \gamma_5 t \right)$
where we write
$ \tilde \mu_t = \frac{g_s}{m_t} \hat \mu_t$ and $\tilde d_t = \frac{g_s}{m_t} \hat d_t$.
Indeed, using dilepton events sample of the $20~\mathrm{fb}^{-1}$
run at 8~TeV, it should be possible to constrain
$\mathrm{Re}(\hat\mu_t)$ and $\mathrm{Re}(\hat d_t)$ at the few percent
level. The imaginary parts $\mathrm{Im}(\hat\mu_t)$ and
$\mathrm{Im}(\hat d_t)$ can be constrained with 15-20\% precision from lepton-top helicity
angles in the semileptonic channel where a full reconstruction of the
$t\bar t$ system is possible, using the same dataset.
Ref.~\cite{Baumgart:2012ay} finds that constraints
at the level of one percent or even below are possible with
$100~\mathrm{fb}^{-1}$ at 13~TeV. Finally, in case of the discovery of a new
resonances which decays into $t\bar t$ pairs, top quark spin
correlations can also be used to analyze couplings of this new
particle~\cite{Baumgart:2011wk,Caola:2012rs}.
\subsection{Top quark pair forward-backward asymmetry}
Top quark pair production in $q \bar q$ collisions exhibits forward-backward asymmetry that
arises in higher orders in perturbative QCD~\cite{Kuhn:1998jr,Kuhn:1998kw,
PhysRevD.78.014008,PhysRevD.77.014003,PhysRevD.73.014008}. As the result, the top quark is
preferentially emitted in the direction of the incoming
quark, while the anti-top quark follows the direction of the incoming antiquark. At the Tevatron,
the direction of the incoming quark corresponds to the direction of the incoming proton, while
the incoming anti-quark most likely comes from an anti-proton. Since LHC is a proton-proton collider, the
$t \bar t$ asymmetry observation becomes difficult because directions of quark and anti-quark
are not correlated with directions of initial hadrons and, in addition, there is a large
gluon flux that reduces the asymmetry. The forward-backward asymmetry is measured at the LHC through
the difference in rapidity distributions of $t$ and $\bar t$. The harder spectrum of valence quarks in the
proton and the correlation of top quark direction with the direction of the incoming quark make the top
rapidity distribution
broader than the rapidity distribution of the anti-top. The corresponding asymmetry is referred to as the charge
asymmetry. It can be written as
\begin{equation}
A_{C}^{\eta} = \frac{N(\Delta|\eta| > 0)-N(\Delta|\eta| < 0)}{N(\Delta|\eta| > 0)+N(\Delta|\eta| < 0)}
\end{equation}
where $\Delta|\eta|\equiv |\eta_t |-|\eta_{\bar t}| $ tells us whether
the reconstructed top or anti-top is more central according to
lab-frame {\em pseudo-rapidity}.
Inclusive forward-backward asymmetries measured at the Tevatron exceed SM predictions by almost three standard
deviations~\cite{PhysRevD.87.092002,PhysRevD.87.011103}, with stronger dependence on $t \bar t$ invariant mass and rapidity
than predicted by the SM.
At the LHC, the ATLAS and CMS Collaborations have performed measurements of
the charge asymmetry
$A_{C}$~\cite{ATLAS:2012an, Chatrchyan201228} and found agreement with SM predictions. However, these measurements
have large errors that makes them not conclusive.
Given that the forward-backward asymmetry is the {\it only} measurement in top physics that shows profound disagreement
with the Standard Model prediction, we feel it is important to understand wether this problem can
be resolved. Our estimates for the LHC are presented below. At a linear collider, it is not possible to address
this problem directly unless the asymmetry mediator is light and can be directly studied in $e^+e^- \to t \bar t jj$.
The higher energy of the $14~{\rm TeV}$ LHC
increases the fraction of $t\bar t$ events that arise from gluon
fusion, relative to $7$ and $8$~TeV LHC. Since $gg \to t \bar t$ does not
produce an asymmetry,
the asymmetric signal decreases with increased center of
mass energy of the collider. Already at $7$~TeV LHC measurements of the top charge asymmetry
are limited by systematic uncertainties. The situation will not change at a higher-energy machine, but
higher luminosity can eventually allow us to improve the systematic errors.
SM predictions for the $14$~{\rm TeV} LHC as a function of cuts on the minimum invariant mass of the top pair
$m_{t\bar t}$
are calculated in Ref.~\cite{PhysRevD.86.034026}.
Cutting on either $t \bar t$ invariant mass
or center-of-mass rapidity increases the proportion of $q \bar q$-initiated top
pair events relative to gluon-initiated events, and thus enhances the
signal. However, even with kinematic cuts, the
size of the signal at the 14 TeV LHC is comparable to the systematic
uncertainties on the current measurements.
The dominant contributions to the systematic errors are
jet energy scale, lepton identification,
background modeling ($t\bar t$, $W+$ jets, multijets), the model dependence of signal generation
and the unfolding procedure. Several contributions to systematic errors, such as jet energy scale
and lepton identification, can be reduced with increased luminosity. Possible improvements in
background modeling are less clear.
The dilepton channel can also be used, usually by defining
a lepton-based asymmetry rather than the top quark based $A_{C}$, with a sensitivity similar to
the lepton+jets one~\cite{ATLAS-CONF-2012-057, CMS-PAS-TOP-12-010}.
Our estimates of the ultimate LHC sensitivity~\cite{TopKine:Snowmass} show
that, with sufficient luminosity, the 14~TEV LHC will be able to {\it conclusively} measure
the SM asymmetry provided that largest systematic errors
(background modeling ($40\%$), lepton identification ($30\%$) and $W+{\rm jets}$ modeling ($13\%$)~\cite{Chatrchyan201228})
scale with luminosity. If the asymmetry is enhanced
due to BSM effects -- as indicated by the Tevatron
data -- the prospects for observing the asymmetry by CMS and ATLAS become event brighter.
An internal study of the LHCb collaboration \cite{lhcb} concludes that a
measurement of the SM $t \bar t$ asymmetry by LHCb experiment is possible
at the $14~{\rm TeV}$ LHC with sufficient luminosity, as suggested earlier in~\cite{Kagan:2011yx}.
This will provide a measurement of $A_c$ at the LHC which is complementary to the measurement of $A_c$ by ATLAS and CMS
collaborations. Combining all of the measurements, one can probably achieve a significant improvement
in the precision compared to individual experiments and hopefully solve the
forward-backward asymmetry puzzle.
To this end, note that
out of the vast zoology of proposed BSM explanations for the Tevatron
anomaly in the top forward-backward asymmetry, axigluons
\cite{Frampton:2009rk, Bai:2011ed, Tavares:2011zg}
are left looking least constrained after the low-energy LHC run has been completed.
Detailed discussions of experimental constraints on axigluon models can be
found in \cite{Gresham:2012kv} for ``light'' ($M_{G'}<450$ GeV)
axigluons and in \cite{Haisch:2011up} for heavy axigluons.
The high-luminosity LHC should be able to rule out axigluon models currently under consideration,
though it is possible to come up with models that explain the Tevatron asymmetry and are difficult
to probe at the LHC.
\subsection{Other kinematic observables related to $A_{FB}$ at the LHC}
\label{toc:topkin-newObs}
The $A_{FB}$ asymmetry is only one of many angular variables whose
distributions can be measured in hadron collisions.
Indeed, if we consider $t \bar t$ production in parton collisions in semileptonic mode,
the full kinematics of the event is characterized by 12 angles and the center-of-mass
partonic collision energy. In principle, kinematic distributions in these angles describe all
kinematic correlations in $t \bar t$ events and therefore are sensitive to potential deviations of
top couplings to $q \bar q$ or $gg$ initial states from their Standard Model values. The
forward-backward asymmetry provides an example of this more general framework.
It will be worthwhile to pursue full angular analysis of the $t \bar t$ events to understand subtle aspects of
top quark pair production or even processes with additional radiation, e.g. $t \bar t j$,
especially in the context of studying top quark couplings to other Standard Model
particles, discussed in Section~\ref{sec:topcouplings}. Unfortunately, this general analysis has
not been attempted so far. Here, we illustrate this general idea by mentioning
additional kinematic observables that can be explored. For example,
Refs.~\cite{Berge:2013xsa,Berge:2013csa} introduce two type of additional asymmetries in
$t \bar t j$ events that can be used to either probe the charge asymmetry or
energy asymmetry in a complementary way or to
provide additional tools to measure the $q g$ contribution to $t \bar t$ production.
\subsection{Kinematics at the linear collider}
At a linear collider, observables such as $A^t_{FB}$ or the slope of the helicity angle $\lambda_t$~\cite{Berger:2012nw} are sensitive to the chiral structure of the $t\overline{t} X$ vertex. A result of a full simulation study of semileptonic $t\overline{t}$ decays~\cite{bib:topcoupl-lalific} is shown in Fig.~\ref{fig:afb_hel}.
\begin{figure}
\begin{center}
\includegraphics[width=0.47\textwidth]{TopQuark/detectors/figs/AFB_wbkg_chi2cut}
\includegraphics[width=0.49\textwidth]{TopQuark/detectors/figs/Helicity_wbkg}
\caption{\sl \underline {Left:} Reconstructed forward backward asymmetry compared with the prediction by the event generator WHIZARD~\cite{Kilian:2007gr,Moretti:2001zz}. \underline{Right:} Polar angle of the decay lepton in the rest frame of the $t$~quark.}
\label{fig:afb_hel}
\end{center}
\end{figure}
The figure demonstrates that it will be possible to measure both the production angle $\theta_{top}$ of the $t$~quark and the helicity angle $\theta_{hel}$ to great precision over a large range, leading to measurements of $A^t_{FB}$ and $\lambda_t$ with a precision of about 2\%. Additionally, the $A^t_{FB}$ and other measurements of the $t\overline{t}$ system, will benefit from a $>60$\% pure sample \cite{Devetak:2010na} in which to measure the $b$ quark charge. The chiral structures of couplings can be possibly be probed in this way.
Since a significant fraction of top studies will be around the $t \bar t$ threshold, understanding kinematic distributions
of top quark decay products in this region is important. This is a non-trivial
problem that is affected by the need to account for
QCD Coulomb interactions to all orders. While results for the total threshold
cross-section $e^+e^- \to t \bar t$ are currently known through NNLO in QCD \cite{Hoang:2000yr},
similar accuracy for kinematic distributions has not been achieved and it is an interesting and important problem
to pursue in the future, if the potential of the threshold scan at the LHC is to be fully exploited.
\section{Searches for new particles associated with the top quark}
\label{sec:newphysics}
The top quark
provides a sensitive probe for physics beyond the Standard Model, based on the following
argument:
the presence of new physics at the TeV scale is very well-motivated by its role in solving the
Planck-weak hierarchy problem of the SM. Namely, such
new particles
can prevent
quantum corrections from dragging the Higgs boson mass (and hence its vev, i.e., the weak scale) all the way up to Planck scale.
Such new particles
must then necessarily couple to the Higgs boson.
However,
because the top quark has the largest coupling (among SM particles) to the Higgs boson, quantum
corrections due to the top quark
are the
dominant source of
destabilization
of the weak scale.
Thus,
such new particles typically also couple
preferentially to
the top quark (among the other SM particles).
In this section, we focus on the {\em direct} production of such new particles, followed by their decay into top-like final states.
In most solutions to the Planck-weak problem, there are actually
charge
$+2/3$, colored NP which accomplish the job of canceling the divergence from top quark loop
in the Higgs mass (and thus directly stabilizing
the weak scale). These can be scalar/spin-0, i.e.,
stops in supersymmetry (SUSY: see review in \cite{Martin:1997ns}).
Alternatively, they can be fermionic
(often denoted by ``top-partners''), as realized in little Higgs
(see reviews in \cite{Schmaltz:2005ky, Perelstein:2005ka}) and composite Higgs (accompanied by composite top) models.
The latter are conjectured to be dual to the framework of a warped extra dimension, following the
AdS/CFT correspondence: see reviews in~\cite{Davoudiasl:2009cd, Contino:2010rs}).
The composite Higgs and top (or extra dimensional) models often also contain bosonic
$t \bar{t}$ resonances.
With the above motivation, the studies performed for the Snowmass process can be grouped into the following three categories:
searches for stops, top-partners and $t \bar{t}$ resonances. These are described in turn below.
Note that virtual/indirect effects of such new particles also lead to rare/flavor changing neutral current decays of the
top quark which are discussed in section \ref{sec:raredecays} of this report.
In addition, there can be shifts in couplings of the top quark which already exist in the SM (for example, flavor-preserving ones),
as discussed in section \ref{sec:topcouplings} of this report.
Finally, these studies overlap with the work of the Snowmass Beyond Standard Model group, where further
examples of New Physics related to top quarks can be found \cite{bsmreport}.
To illustrate an impact that top physics studies can have on discovering and understanding physics
beyond the Standard Model, we now describe a {\bf discovery story}. Here the
tell-tale signs for top and Higgs compositeness at the TeV scale is provided
by a multitude of measurements and observations in top
physics~\cite{Davoudiasl:2009cd, Contino:2010rs}.
In particular, suppose there is an observation of a shift in the $t \bar{t} Z$ cross-section of the order of
$10\%$ at a linear collider and, possibly even at the HL-LHC (see Section~\ref{sec:topcouplings}). It
could be a ``smoking-gun'' signal for this scenario, since this model predicts such a size for this shift (as compared to
weakly coupled theories such as
SUSY, where this effect is signficiantly suppressed and gives no signal at the LHC).
At the same time, in composite models rare decays $t \to c Z$ or $t \to c h$ can occur with a
branching fraction of up to $10^{-4}$ which would then be accessible at the LHC and a lepton collider,
see Section~\ref{sec:raredecays}.
Finally, both top-partners and $t \bar{t}$ resonances with $O(\hbox{TeV})$ masses are
omnipresent in the compositeness scenario and would thus be accessible at the LHC, and especially
its high-luminosity option, via direct production.
Therefore, as our story illustrates, top quark physics at the LHC and at a linear collider may
be a crucial element of discovering and elucidating physics beyond the Standard Model.
\subsection{Stops}
\label{stops}
\def\tilde{t}{\tilde{t}}
\def\tilde{\chi}^0{\tilde{\chi}^0}
\def\tilde{g}{\tilde{g}}
\def\tilde{\chi}^\pm{\tilde{\chi}^\pm}
SUSY is perhaps the most popular solution to the Planck-weak hierarchy problem of the SM. It involves
addition of a {\em superpartner} for every particle of the SM, with a spin differing by $1/2$-unit from that of the corresponding
SM particle.
While, in general, superpartner masses in SUSY models are very model-dependent, naturalness strongly suggests that the scalar partners of the top quark, or {\it stops}, should have masses around the week scale.
The reason is that (as mentioned above) the stops
cancel the largest divergence in the Higgs mass squared parameter, coming from the SM top loop.
This
makes stops a prime target for LHC searches.
The results of such searches are typically presented in terms of the ``vanilla stop'' simplified model, which contains two particles, a stop $\tilde{t}$ and a neutralino LSP $\tilde{\chi}^0$ (i.e., superpartner of photon and $Z$ or Higgs boson). The stop is assumed to decay via $\tilde{t}\to t\tilde{\chi}^0$ with a 100\% branching ratio. Within this model, the current ``generic'' bound on the stop mass is about 700 GeV~\cite{ATLAS8,CMS8}.
One of the tasks of future experiments is to improve the reach on $m(\tilde{t})$ for generic spectra.
In fact, both ATLAS and CMS have presented estimates of the discovery reach of LHC-14 and HL-LHC in the vanilla stop model, extrapolating the present 1-lepton search~\cite{ATLAS:2013hta,CMS:2013xfa}. For a ``generic'' spectrum, stops up to approximately 800 (900) GeV can be discovered, at a 5-$\sigma$ level, with 300 fb$^{-1}$ (3 ab$^{-1}$) integrated luminosity.
It is interesting to determine if the reach at LHC 14 TeV for this generic case can be extended beyond
the above ATLAS/CMS projections using {\em special} techniques developed recently and so far applied only to the LHC 7/8 TeV.
The first study (as part of the Snowmass process) mentioned below is along these lines.
Moreover, it must be emphasized that lighter stops are still allowed by LHC 7/8 TeV.
In particular: \\
\vspace{0.01in} \\
(a) If $m(\tilde{\chi}^0)>250$ GeV, stops of any mass are allowed; \\
\vspace{0.01in} \\
(b) in the ``off-shell top'' region, $m_t > m(\tilde{t})-m(\tilde{\chi}^0) > m_W$, stops above 300 GeV are allowed; \\
\vspace{0.01in} \\
(c) in the ``compressed'' region, $m(\tilde{t})\approx m(\tilde{\chi}^0) +m_t$, stops of any mass are allowed (this includes the particularly challenging ``stealthy'' region, $m(\tilde{t})\approx m(t) \gg m(\tilde{\chi}^0)$); and \\
\vspace{0.01in} \\
(d) in the ``squeezed'' region, $m(\tilde{t})-m(\tilde{\chi}^0) < m_W$, stops of any mass are allowed. \\
\noindent In all of these regions, the kinematics of stop production and decay yields events with little missing transverse energy (MET), reducing the efficiency of LHC searches.
Thus, another goal of future experiments should be to explore the special regions listed above.
A couple of studies to cover the stealth
stops of case (c) above were done
as part of Snowmass process and are outlined below.
Although LHC will clearly play a leading role in the generic case\footnote{direct production of stops
at the ILC in this region is not possible, given the current bounds},
it should be emphasized that in any of the special regions, stops can still be within the kinematic reach of the ILC/CLIC, at $\sqrt{s}=500$ GeV or 1 TeV. In this case, the ILC could play a crucial role in discovering the stops and precisely determining/confirming their properties, {\it e.g.} spin and masses.
Finally, {\em addition} of particles (such as gluino or chargino, i.e., superpartners of SM gluon or $W$)
to the above simplified model is well-motivated.
Adding such particles to the model generally weakens the exclusion limits.
Studies along these lines
were also performed for the Snowmass process and are described below.
\subsubsection{Vanilla stops}
Here fully hadronic decays using strategies inspired by~\cite{Plehn:2010st,Plehn:2012pr,Kaplan:2012gd,Dutta:2012kx}
are considered.
The fully hadronic channel has two advantages over leptonic searches. The first is that it has the largest branching fraction for the top decays. The second is that it has no inherent missing energy from neutrinos, so all the missing energy comes from the neutralinos. This allows many backgrounds to be reduced by vetoing events with leptons.
Jet-substructure based
top tagging (see section \ref{sec:detector} of this report) is used to distinguish signal from background.
%
The results are summarized in table~\ref{tab:vanilla}: for more details, see reference \cite{Stolarski:2013msa}.
\begin{table}[t]
\centering
\begin{tabular}{|c|c|c|c|c|}
\hline
Collider & Energy & Luminosity & Cross Section & Mass \\ \hline \hline
LHC8 & 8 TeV & 20.5 fb$^{-1}$ & 10 fb & 650 GeV \\ \hline
LHC & 14 TeV & 300 fb$^{-1}$ & 3.5 fb & 1 TeV \\
HL LHC & 14 TeV & 3 ab$^{-1}$ & 1.1 fb & 1.2 TeV \\
HE LHC & 33 TeV & 3 ab$^{-1}$ & 91 ab & 3.0 TeV \\
VLHC & 100 TeV & 1 ab$^{-1}$ & 200 ab & 5.7 TeV \\
\hline
\end{tabular} \hspace{-0.138cm}\vline
\vspace{0.3cm}
\caption{Stop production mass limits. The first line gives the current bound on stops.
The remaining lines give the estimated 5 $\sigma$ reach, based on a study for the Snowmass process, in stop pair production cross-section
and mass for different future hadron collider runs. }
\label{tab:vanilla}
\end{table}%
\subsubsection{Stealth stops}
\label{sec:stops}
In the above-mentioned ATLAS/CMS projections of reach for stops at LHC 14 TeV, significant gaps in the coverage remain: for example, no discovery is possible for the LSP mass above 500 GeV, as well as in the compressed and stealthy regions, even at HL-LHC. It is clear that novel search strategies will be needed to cover these regions.
Two studies of such strategies were contributed to our working group (see table \ref{tab:stealth} for summary of results).
Reference
~\cite{Han:2013lna} focused on the stealthy stop region, which is particularly challenging since, unlike the region with a heavy neutralino, no significant MET is generated even in the presence of ISR jets. The challenge is to distinguish $\tilde{t}\t^*$ events from a much larger $t\bar{t}$ background. Two methods to achieve this task have been studied: (a) using spin correlations, which are present in $t\bar{t}$ but not in $\tilde{t}\t^*$ events, due to $\tilde{t}$ being a scalar particle~\cite{Han:2012fw} (see also section \ref{sec:topkin-spincorrel} of this report); and (b) using an $m_{T2}$ cut in dileptonic event sample~\cite{Kilic:2012kw}.
It was found that, using spin correlations, LHC-14 with 100 fb$^{-1}$ of data will be able to discover the stealthy
stop at the $5\sigma$ level, assuming the stop mass of up to 200 GeV
and considering statistical errors only.
Assuming a 15\% systematic error,
the $m_{T2}$ method will be able to discover right-handed stealthy stops {\em except} in the $(185, 195)$ GeV window.
The sensitivity to the
left-handed stop is poor, since there is no $m_{T2}$ tail in the signal in this case.
The second study \cite{wpS1} analyzed the possibility of using the vector boson fusion stop production channel, which provides additional jets that could be used to tag events with stealthy, compressed, or light stops. For compressed stops with $\Delta M = m(\tilde{t}) - m(\tilde{\chi}^0) = 10$ GeV, it was found that the LHC-14 with 300 fb$^{-1}$ of data will be able to probe $m(\tilde{t}) = 400$ GeV at a $5 \sigma$ level. The mass reach increases for larger $\Delta M$. Studies for stops in the stealthy, ``off-shell top'', and ``squeezed'' regions using vector boson fusion are ongoing.
\begin{table}[t]
\centering
\begin{tabular}{|c|c|c|c|}
\hline
Collider & Luminosity & Technique & Mass reach
%
%
\\ \hline \hline
LHC 14 TeV & 100 fb$^{-1}$ & spin-correlations & 175--200 GeV
%
%
(for stealth stops, \\
& & & statistical uncertainty only) \\ \hline
LHC 14 TeV & 100 fb$^{-1}$ & dileptonic $m_{ T2 }$ &
175--185~GeV and 195-200~GeV
%
%
(for stealth tops) \\ \hline
LHC 14 TeV & 300 fb$^{-1}$ & VBF & 400 GeV
%
%
(for compressed case, \\
& & & with $m(\tilde{t}) - m(\tilde{\chi}^0) = 10$ GeV) \\ \hline
\end{tabular} \hspace{-0.138cm}
\vspace{0.3cm}
\caption{Estimated 5 $\sigma$ discovery reach for stealth/compressed stops, based on various studies for the Snowmass process. See the text for explanation of these concepts.}
\label{tab:stealth}
\end{table}%
\subsubsection{Gluino-initiated stop production}
In addition to stops, naturalness also strongly motivates a light gluino, constraining its mass through the one-loop QCD correction to stop mass. A rough naturalness bound is $m(\tilde{g})<2m(\tilde{t})$~\cite{Brust:2011tb}. This motivates considering a simplified model with gluino, stop and an LSP, with a decay $\tilde{g}\to t\bar{t}+$MET. Assuming that this decay proceeds via an off-shell stop and has a 100\% branching ratio, LHC-8 searches rule out gluino masses up to about $1.3$ TeV, provided that the LSP mass is below 500 GeV~\cite{ATLAS8go,CMS8go}. Extrapolating the search in the all-hadronic channel, CMS estimates a $5\sigma$ discovery reach of $1.7$ TeV at LHC-14 with 300 fb$^{-1}$ of data~\cite{CMS:2013xfa}.
For gluino masses above TeV, boosted tops become increasingly common in $\tilde{g}$ decays, and can be used to tag SUSY events~\cite{Berger:2011af} (see section \ref{sec:detector} of this report for techniques to identify boosted tops).
A preliminary study (with no detector simulation)
in reference~\cite{wpmaxim1} suggests that a search using top tags, in combination with more traditional kinematic cuts in all-hadronic channel, at the LHC-14 with 300 fb$^{-1}$ (3 ab$^{-1}$) of data will be able to discover gluinos up to $1.8$ ($2.1$) TeV, provided that the stop mass is below $1.1$ ($1.4$) TeV.
\begin{table}[t]
\centering
\begin{tabular}{|c|c|c|}
\hline
Collider & Luminosity & Mass
%
%
\\ \hline \hline
LHC 14 TeV & 300 fb$^{-1}$ & 1.8 TeV \\ \hline
LHC 14 TeV & 3000 fb$^{-1}$ & 2.1 TeV \\ \hline
\end{tabular} \hspace{-0.138cm}
\vspace{0.3cm}
\caption{Estimated 5 $\sigma$ discovery reach for gluino decaying into stops, with $R$-parity conservation, based on a study for the Snowmass process.}
\label{tab:}
\end{table}%
\subsubsection{Including more electroweak particles}
Another well-motivated extension of the vanilla stop simplified model is to add a chargino $\tilde{\chi}^\pm$, with $m(\tilde{\chi}^\pm)<m(\tilde{t})$. This is also motivated by naturalness, since the charged Higgsino mass is controlled by the $\mu$ parameter which cannot be far above 100 GeV in natural SUSY
models \cite{Brust:2011tb}. This simplified model has the possibility of {\it asymmetric} stop events: {\it e.g.} $pp\to \tilde{t}\t^*, \tilde{t}\to t\tilde{\chi}^0, \tilde{t}^* \to b\tilde{\chi}^\pm$. A study of the LHC sensitivity to this signal was performed: for details, see reference \cite{wpG}. The proposed search uses the 1-lepton+MET channel, and relies crucially on the ``topness'' variable, introduced in~\cite{Graesser:2012qy} as a general tool to suppress the $t\bar{t}$ background in this channel.
It was found that $5\sigma$ discovery is possible at LHC-14 with $300$ fb$^{-1}$ for stop masses up to about 1 TeV, if $m(\tilde{\chi}^0)$ is below about 400 GeV. With $3000$ fb$^{-1}$, the discovery reach extends to stop masses about 1.3 TeV for similar $\tilde{\chi}^0$.
A related simplified model was used in the study in reference \cite{wpS2}. Motivated by the ``well-tempered neutralino'' dark matter scenario~\cite{ArkaniHamed:2006mb}, this study considered a spectrum with light bino and Higgsino, leading to three neutralino and one chargino states at the bottom of the SUSY spectrum. It was assumed that all these states are lighter than the stop. The analysis focused on the dilepton signature, where the leptons can come either from top decays or from $\chi^0_{2,3}\to Z\chi^0_1$. It was found that the reach is about 700 GeV.
\begin{table}[t]
\centering
\begin{tabular}{|c|c|c|c|}
\hline
Collider & Luminosity & Technique/channel & Mass
%
%
\\ \hline \hline
LHC 14 TeV & 300 fb$^{-1}$ & topness, asymmetric & 1 TeV \\ \hline
LHC 14 TeV & 3000 fb$^{-1}$ & topness, asymmetric & 1.3 TeV \\ \hline
LHC 14 TeV & 300 fb$^{-1}$ & dilepton, well-tempered neutralino &
700 GeV$^{\rm preliminary}$ \\ \hline
\end{tabular} \hspace{-0.138cm}
\vspace{0.3cm}
\caption{Estimated 5 $\sigma$ discovery reach for stops decaying into chargino, based on various studies for the Snowmass process. See the text for explanation of these concepts.}
\label{tab:chargino}
\end{table}%
\subsubsection{$R$-parity violation}
Yet another interesting scenario is $R$-parity violating (RPV) supersymmetry, where
decay modes are modified relative to the above cases of $R$-parity conservation.
For example, a stop can decay via $\tilde{t}\to \bar{b}\bar{s}$ induced by the $UDD$ superpotential operator. This scenario emerges naturally from models with minimal flavor violation~\cite{Nikolidakis:2007fc,Csaki:2011ge}.
Direct stop production in this case yields all-hadronic final states,
but it might still be possible to search in this channel: see, for example, the Snowmass study \cite{Duggan:2013yna}.
However, just as in conventional SUSY, naturalness strongly suggests the presence of relatively light gluinos. Gluino decays via cascades involving stops, $\tilde{g}\to \tilde{t} t, \tilde{t} \to 2j$, may be observable, even though they do not produce large MET. If $\tilde{g}$ is Majorana, as in simplest SUSY models, such decays can provide a striking same-sign dilepton (SSDL) signature. Current SSDL searches already rule out gluinos up to 800 GeV, independent of the stop mass, in the RPV scenario~\cite{Berger:2013sir}.
At LHC-14 with 300 fb$^{-1}$ (3 ab$^{-1}$) of data, the projected 5 $\sigma$ reach of this search in gluino mass is $1.4$ ($1.6-1.75$) TeV, depending on the stop mass~\cite{Saelim:2013gea}.
These estimates include fast detector simulation with Delphes~\cite{Delphes3}, but no pile-up.
It also
includes an improvement in sensitivity due to an additional requirement of one or two massive jets. The massive jets can be either due to boosted stop decays, or to accidental mergers of neighboring jets in a high jet multiplicity signal event.
An alternative approach is a search in a single-lepton channel, which has a higher rate and applies to both Majorana and Dirac gluinos~\cite{Han:2012cu}. In this case, the requirement of stop mass reconstruction from jet pairs can be used as an additional handle to suppress backgrounds. At the 14~TeV LHC, this search will be sensitive to {\em Dirac} gluino masses up to about $2$ TeV~\cite{wpK}.
\begin{table}[t]
\centering
\begin{tabular}{|c|c|c|c|}
\hline
Collider & Luminosity & Technique & Reach \\ \hline \hline
LHC 14 TeV & 300 fb$^{-1}$ & same-sign dilepton & 1.4 TeV \\ \hline
LHC 14 TeV & 3000 fb$^{-1}$ & same-sign dilepton & 1.6--1.75 TeV \\ \hline
LHC 14 TeV & 300 fb$^{-1}$ & single-lepton, reconstruct mass & 2 TeV $^{\rm preliminary}$
\\
\hline
\end{tabular} \hspace{-0.138cm}
\vspace{0.3cm}
\caption{Estimated 5 $\sigma$ reach for gluino decaying into stops, with $R$-parity violation, based on various studies for the Snowmass process. See the text for explanation of these concepts.
}
\label{tab:RPV}
\end{table}%
\subsection{Top-partners}
As mentioned above, in non-SUSY solutions to the Planck-weak hierarchy problem,
the divergence in Higgs mass squared parameter from SM top loop is canceled
by new {\em fermions} which are vector-like under the SM gauge symmetries. Typically, they are color triplets with electric charge 2/3 (i.e., same as the SM top and hence these new particles
are dubbed top-partners.
Such
particles can also arise in other extensions of the SM, so it is useful to follow a model-independent, simplified approach
in studying their signals.
The top-partners can be produced via QCD interactions in pairs or singly \cite{Mrazek:2009yu},
the latter resulting from coupling of
top-partner to SM top/bottom, as needed to cancel the SM top divergence in Higgs mass squared parameter.
Based on the $SU(2)_L$ gauge symmetry of the SM, the top-partners are often accompanied by ``bottom-partners''.
Finally, in some composite Higgs models, an extension of the EW symmetry group (from that in the SM) is motivated by
the goal of avoiding constrains from $Z b \bar{b}$ \cite{Agashe:2006at}: this results in the appearance of color triplet, but charge $5/3$ particles
(in addition to the above top/bottom partners).
In short, there are three types of vector-like quarks which are well-motivated extensions
of the SM, namely, top and bottom-partners and charge-$5/3$ fermions.
Once produced, these vector-like quarks can decay into a
top-like final state.
All of these cases were studied for various LHC scenarios as part of the Snowmass process (including
both single and pair production of top-partners mentioned above) and are discussed below.
Note that the current (LHC 7/8 TeV) bound on these quarks is at least 800 GeV \cite{CMS-PAS-B2G-12-015,
ATLAS-CONF-2013-060} so that
their direct production at the ILC is not possible.
The single production mechanisms have larger rates than those of pair productions for
{\em heavier} top/bottom partners. Moreover, analyses of single-production channels
might permit the measurement of the
the
couplings of top-partners to the SM top/bottom quark (which were mentioned towards the end of the first paragraph of this section).
Note that the
top-partner single-production, that proceeds via the intermediate exchange of a bottom quark,
has a rate significantly higher than those for a single bottom partner or a charge-$5/3$
partner, which are mediated by the exchange of a top. Hence, in the following, for bottom partners
and charge $5/3$ quarks, only pair production is considered.
All samples used for the top-partner studies were processed with the {\sc Delphes}~\cite{Delphes3} fast
detector simulation, using the generic Snowmass detector parameters~\cite{SnowmassDetSim}. The
background samples were generated in bins of $H_T$, as described in~\cite{SnowmassBackgrounds, SnowmassOSG}.
{\bf Pair production of top-partners:}
The top-partner has three possible decay modes:
$bW$, $ t H$ and $ Z t $. The interesting feature is that, in the limit of a heavy top-partner,
the decay modes are equally shared by these three modes
(following the principle of Goldstone equivalence). A general analysis
such that the three branching fractions $bW$, $ t H$, and $ Z t $ are
free parameters, subject to the constraint that they add up to unity
and span a ``triangle'' of branching fractions, has been performed recently~\cite{CMS-PAS-B2G-12-015}.
However, it is useful to consider a nominal set of branching fractions,
one that is motivated by naturalness with BF($T\rightarrow Wb$) = 0.5,
BF($T\rightarrow th$) = BF($T\rightarrow Zt$) = 0.25.
The results from an analysis~\cite{CMS-PAS-B2G-12-015} based on lepton+jets and multi-lepton signatures for all
decay modes $bW$, $ t H$ and $ Z t $, and utilizing presence of multiple $b$-jets, large number of jets,
are given in table \ref{tab:topbPartner}.
Due to the large mass of the top-partner,
jet substructure techniques are utilized to identify the $W$-tagged and top-tagged jets and to keep the
signal yield high and fully optimize the signal-to-background discrimination.
Reference \cite{Bhattacharya:2013iea} contains more details of the analysis.
{\bf Single production of top/bottom-partners:}
As mentioned above, the top-partner
can decay into one of three possible final states: $h t$, $Zt$ and $W b$. Since the $W +$ jet backgrounds are considerable for the third mode, here the focus is on the first two decay modes.
The basic idea is to reconstruct the top-partner mass using the full event information in the decays
$ht\rightarrow b\overline{b}\ell\nu b$ and $Zt\rightarrow \ell\ell\ell\nu b$.
The results are summarized in Table~\ref{tab:topbPartner},
for more details, see reference \cite{Andeen:2013zca}.
{\bf Pair production of bottom-partners:}
The decays of bottom-partners
can be into $W^- t$, $Zb$, or $Hb$.
Thus, pair production of bottom partners can lead to interesting signal of
same-sign dileptons via $W^- t W^+ \bar{t} \rightarrow b \bar{b} \; 2 \; W^+ \; 2 W^-$,
followed by leptonic decays of both $W^+$ (or $W^-$).
More details of this study can be found in reference \cite{Varnes:2013pxa}; here,
only the final results, obtained using the nominal branching fractions
BF$(B \rightarrow Wt)$ = 0.5, BF$(B \rightarrow Zb)$ = BF$(B \rightarrow Hb)$= 0.25, are shown in Table~\ref{tab:topbPartner}.
{\bf Pair production of Charge-$5/3$ fermion:}
The charge-$5/3$ vector-like fermions
($T_{5/3}$)~\cite{Contino:2008hi} decay via $T_{5/3} \rightarrow t W^+ \rightarrow b W^+ W^+$
and thus the pair production of these particles results in same-sign dileptons
with a branching fraction of approximately 0.2. These final states can be distinguished from
Standard Model backgrounds with same-sign dileptons by the presence of
jets or leptons from the second $T_{5/3}$ and by the magnitude of the
scalar sum of transverse momenta of the decay products.
At $T_{5/3}$ masses of interest, hadronically $W$ bosons and top quarks from the $T_{5/3}$ decay
are often highly boosted and can be identified using the tagging methods described in
Section~\ref{sec:boostedtop} which enhance the background
discrimination. Furthermore, it is possible to fully reconstruct the second $T_{5/3}$,
in case of fully hadronic decay, and compute its mass.
Table \ref{tab:topbPartner} displays the reach for
these exotic quarks; for more details, see reference\cite{Avetisyan:2013rca}.
\begin{table}[t]
\centering
\begin{tabular}{|c|c|c|c|c|c|}
\hline
Collider & Luminosity & Pileup & 3$\sigma$ evidence & 5$\sigma$ discovery & 95\% CL \\ \hline \hline
\multicolumn{6}{|l|}{\bf top-partner pair production} \\
LHC 14 TeV & 300 fb$^{-1}$ & 50 & 1340~GeV & 1200~GeV& 1450~GeV \\ \hline
LHC 14 TeV & 3 ab$^{-1}$ & 140 & 1580~GeV & 1450~GeV& 1740~GeV \\ \hline
LHC 33 TeV & 3 ab$^{-1}$ & 140 & 2750~GeV & 2400~GeV & 3200~GeV\\ \hline \hline
\multicolumn{6}{|l|}{\bf top-partner single production} \\
LHC 14 TeV & 300 fb$^{-1}$ & 50 & 1275~GeV & 1150~GeV & \\ \hline
LHC 14 TeV & 3 ab$^{-1}$ & 140 & 1130~GeV & 1000~GeV & \\ \hline
LHC 33 TeV & 3 ab$^{-1}$ & 140 & 1350~GeV & 1220~GeV & \\ \hline
LHC 100 TeV & 3 ab$^{-1}$ & 50 & 1750~GeV & 1600~GeV & \\ \hline
LHC 100 TeV & 3 ab$^{-1}$ & 140 & 1750~GeV & 1575~GeV & \\ \hline\hline
\multicolumn{6}{|l|}{\bf bottom-partner pair production}\\
LHC 14 TeV & 300 fb$^{-1}$ & 50 & 1210~GeV & 1080~GeV & 1330~GeV \\ \hline
LHC 14 TeV & 3 ab$^{-1}$ & 140 & 1490~GeV & 1330~GeV & $>$1500~GeV \\ \hline
LHC 33 TeV & 300 fb$^{-1}$ & 50 & $>$ 1500~GeV & $>$ 1500~GeV & $>$ 1500~GeV \\ \hline\hline
\multicolumn{6}{|l|}{\bf Charge $5/3$ fermion pair production}\\
LHC 14 TeV & 300 fb$^{-1}$ & 50 & 1.51 TeV & 1.39 TeV & 1.57 TeV \\ \hline
LHC 14 TeV & 3 ab$^{-1}$ & 140 & 1.66 TeV & 1.55 TeV & 1.76 TeV \\ \hline
LHC 33 TeV & 3 ab$^{-1}$ & 140 & 2.50 TeV & 2.35 TeV & 2.69 TeV \\ \hline
\end{tabular}
\vspace{0.3cm}
\caption{Expected mass sensitivity for heavy top and bottom partners, based on various studies for the Snowmass process.}
\label{tab:topbPartner}
\end{table}%
\subsection{$t \bar{t}$ resonances}
As mentioned earlier, in
non-supersymmetric solutions to the Planck-weak
hierarchy problem, there are typically
bosonic new particles which decay dominantly into $t \bar{t}$.
Examples are leptophobic $Z^\prime$'s in topcolor models ~\cite{Harris:1999ya}
or KK gluons in warped extra dimensional frameworks (conjectured to be dual 4D
composite Higgs models: see reviews in \cite{Davoudiasl:2009cd, Contino:2010rs}).
Moreover, such $t \bar{t}$ resonances are favored to be rather
heavy (a few TeV) due to the constraints from various precision tests.
and/or by the current direct bounds from LHC 7/8 TeV \cite{Aad:2013nca, CMSttbar1, Aad:2012raa, CMSttbar2}.
Thus, the
top quarks resulting from their decays are boosted so that the top decay products
can be quite collimated, requiring special identification techniques which have
been developed recently (for more details, see section \ref{sec:detector} of this report).
In some models, these $t \bar{t}$ resonances can also be broad, thereby adding to the challenge of
searching for them.
Three such studies of discovery of $t \bar{t}$ resonances were done as part of the Snowmass process and are discussed in what follows.
Of course, post-discovery, the focus will shift to determination of the quantum numbers
of these $t \bar{t}$ resonances.
For example, the spin and chiral structure of couplings of these resonances
can be measured via angular distribution and polarization of the resulting top quarks: see, for example, references \cite{Agashe:2006hk, Barger:2006hm, Baumgart:2011wk}.
Finally, note that given the mass range of these $t \bar{t}$ resonances, ILC/CLIC would not play a role
in a direct search.
\subsubsection{Dileptonic}
One of the studies (reference \cite{Iashvili:2013ada}) focused on $W$'s from both top quarks decaying into lepton (called ``dileptonic'' $t \bar{t}$).
One expects hadronic activity near the leptons, since the boost of the top quark can put a lepton and a $b$-jet into the same cone.
So, SM $t \bar{t}$ background can be suppressed by in fact requiring {\em smaller} separation between lepton and the
closest jet.
The results are summarized in Table \ref{tab:dileptonic}.
\begin{table}[t]
\centering
\begin{tabular}{|c|c|c|c|c|}
\hline
Collider & Luminosity & Pileup & 95 \% exclusion & 5 $\sigma$ discovery \\ \hline \hline
LHC 14 TeV & 300 fb$^{-1}$ & 50 & 4.4 TeV & 2.8 TeV \\ \hline
LHC 14 TeV & 3 ab$^{-1}$ & 140 & 4.7 TeV & 4.1 TeV \\ \hline
\end{tabular} \hspace{-0.138cm}
\vspace{0.3cm}
\caption{Expected mass sensitivity for a $Z^{ \prime }$ decaying into {\em di}leptonic $t \bar{t}$, based on a study for the Snowmass
process.}
\label{tab:dileptonic}
\end{table}%
\subsubsection{Semileptonic and fully hadronic}
Alternatively, one of the two $W$'s from the decays of the top quarks can give a lepton, while
the other one decays into hadrons (semileptonic $t \bar{t}$) or none of the two $W$'s
decays into leptons (fully hadronic $t \bar{t}$).
A study for the Snowmass process of the fully hadronic channel
utilized $b$-tagging and large-$R$ jet substructure to distinguish jets from boosted top quarks from jets from QCD
multijet production, and evaluated the prospects for a search for narrow resonances.
The results are expressed in terms of both a leptophobic $Z^{ \prime }$ and KK graviton in warped extra dimensional models: see
tables \ref{tab:yale1} and for details, see sections 3.2 and 3.3.2 of reference \cite{Agashe:2013fda}.
The study for Snowmass process of the semileptonic channel is still to be completed and so the results
shown in \ref{tab:yale2} (for leptophobic $Z^{ \prime }$ and KK gluon) are from the ATLAS Snowmass
whitepaper~\cite{ATLAS:2013hta}.
\begin{table}[t]
\centering
\begin{tabular}{|c|c|c|c|c|}
\hline
Collider & Luminosity & Pileup & 95 \% exclusion for $Z^{ \prime }$ & 95 \% exclusion for KK graviton\footnote{The numbers shown are for right-handed top quark being localized near TeV brane and with
the parameter $k / M_{ \rm Pl } = 1$.}
\\ \hline \hline
LHC 14 TeV & 300 fb$^{-1}$ & 50 & 3.7 TeV & 2 TeV \\ \hline
LHC 14 TeV & 3 ab$^{-1}$ & 140 & 4.1 TeV & 2.8 TeV \\ \hline
\end{tabular} \hspace{-0.138cm}
\vspace{0.3cm}
\caption{Expected mass sensitivity for a leptophbic $Z^{ \prime }$ and KK graviton
decaying into
fully hadronic $t \bar{t}$, based on a study for the Snowmass process.
}
\label{tab:yale1}
\end{table}%
\begin{table}[t]
\centering
\begin{tabular}{|c|c|c|c|c|}
\hline
Collider & Luminosity & Pileup & 95 \% exclusion for $Z^{ \prime }$
%
& 95 \% exclusion for KK gluon
%
\\ \hline \hline
LHC 14 TeV & 300 fb$^{-1}$ & 50 & 3.3 TeV
%
& 4.3 TeV
%
\\ \hline
LHC 14 TeV & 3 ab$^{-1}$ & 140 & 5.5 TeV
%
& 6.7 TeV
%
\\ \hline
\end{tabular} \hspace{-0.138cm}
\vspace{0.3cm}
\caption{Expected mass sensitivity for a leptophobic $Z^{ \prime }$
and KK {\em gluon}
decaying into
{\em semi}leptonic
$t \bar{t}$~\cite{ATLAS:2013hta}.
}
\label{tab:yale2}
\end{table}%
Another study focused on the KK gluon (which is typically a {\em broad} resonance) in warped extra dimensional models.
In order to identify boosted top quarks, it used
the Template Overlap Method (TOM) \cite{Almeida:2010pa}.
TOM has been extensively studied in the past in the context of theoretical studies of boosted tops and boosted Higgs
decays \cite{Backovic:2012jj}, as well as used by the ATLAS collaboration for a boosted resonance search \cite{Aad:2012raa}. The method is designed to match the energy distribution of a boosted jet to the parton-level configuration of a boosted top decay, with all kinematic constraints taken into account.
Low susceptibility to intermediate levels of pileup (i.e. 20 interactions per bunch crossing), makes TOM particularly attractive for boosted top analyses at the LHC.
For more details about how the TOM is used in this study, see section 3.4 of reference \cite{Agashe:2013fda}: the results are
shown in Table \ref{tab:TOM}.
\begin{table}[t]
\centering
\begin{tabular}{|c|c|c|c|c|}
\hline
Collider & Luminosity & Pileup & 3 $\sigma$ evidence &
%
%
5 $\sigma$ discovery \\ \hline \hline
LHC 14 TeV & 300 fb$^{-1}$ & 50 & 3.8 TeV & 3.2 TeV \\ \hline
LHC 14 TeV & 3 ab$^{-1}$ & 50 & 4.4 TeV & 3.5 TeV \\ \hline
\end{tabular} \hspace{-0.138cm}
\vspace{0.3cm}
\caption{Expected mass sensitivity for a KK gluon decaying into semileptonic
$t \bar{t}$, based on a study for the Snowmass process using
the template overlap method.}
\label{tab:TOM}
\end{table}%
\subsubsection{Single-top resonance}
~
Resonances can appear not only in top pair production, but also in single top quark production. This
final state is particularly sensitive to a high-mass $W'$~boson that couples primarily to
quarks. Current limits for $W'$ production are around
1.8~TeV~\cite{Aad:2012ej,ATLAS-CONF-2013-050,Chatrchyan:2012gqa}. A Snowmass study shows that the reach
for $W'$ can be extended to 5~TeV (6~TeV) with 300~fb$^{-1}$ (3000~fb$^{-1}$) at the 14~TeV
LHC~\cite{WKKgWhitePaper}.
In warped extra dimensional models, the KK gluon discussed in the previous section can also have a
sub-dominant decay into $ t \bar{c}$ (and $\bar{t} c$)~\cite{Aquino:2006vp}. This process is also
relevant for the flavor sector, see the chapter on Flavor working group report~\cite{flavorWG}.
The final state has a single top quark, just like $W'\rightarrow tb$, but now the other quark jet
is from a charm quark rather than a bottom quark. This has consequences for the $b$-tag multiplicity
and background suppression. The Snowmass study finds a mass limit on the KK gluon of about 3.5~TeV if the
branching ratio to $tc$ is 20\%. If this branching ratio is less than 5\%, the signal is buried
below backgrounds and no limit can be set.
A fourth-generation quark with chromomagnetic couplings will be visible in the single top plus $W$~boson
final state~\cite{Nutter:2012an,Aad:2013rna}. Due to the strong nature of the
fourth-generation bottom quark
production process,
the reach for this particle at the high-luminosity LHC should be multi-TeV, similar to the $W'$.
\section{Rare decays}
\label{sec:raredecays}
\subsection{Introduction}
Extensions of the SM often induce
sizable
flavor-violating couplings between the top quark and other Standard Model particles, typically through new physics
in loops. In contrast, flavor-changing neutral couplings of the top are highly suppressed in the SM, so that the measurement of anomalous or flavor-violating couplings of the top quark provides a sensitive probe of physics beyond the Standard Model. Since
the top quark decays before hadronizing, top flavor violation is ideally probed through direct flavor-changing neutral current (FCNC) production and decays of the top quark in experiments at the energy frontier.
Although flavor-violating couplings of the top may arise from many sources, if the responsible new physics is heavier than the top, it can be integrated out and its effects described by an effective Lagrangian: for details, see, for example, \cite{rarewp}.
In Section \ref{sec:theory} we summarize predictions for the size of flavor-changing top decays in the Standard Model and in various motivated models for new physics. In Section \ref{sec:limits} we collect the current best limits on top FCNC decays from direct searches.
In Section \ref{sec:future} we investigate the potential for future measurements at the LHC and ILC to constrain top FCNCs.
\subsection{Flavor-violating Top Decays} \label{sec:theory}
The
branching ratio (BR) of a flavor-violating decay of the top quark is
given
by the
ratio
of the flavor-violating partial width relative to the dominant top quark partial width, $\Gamma(t \to bW)$. In Table \ref{tab:theory} we summarize predictions for top FCNC
BRs
in the Standard Model and various motivated new physics models. In the case of new physics models, the listed BR is intended as an approximate maximal value given ancillary direct and indirect constraints.
\begin{table}
\caption{SM and new physics model predictions for branching ratios of top FCNC decays. The SM predictions are taken from~\cite{AguilarSaavedra:2004wm}, on 2HDM with flavor violating Yukawa couplings~\cite{AguilarSaavedra:2004wm, Atwood:1996vj} (2HDM (FV) column), the 2HDM flavor conserving (FC) case from~\cite{Bejar:2006ww}, the MSSM with $1$TeV squarks and gluinos from~\cite{Cao:2007dk}, the MSSM for the R-parity violating case from~\cite{Yang:1997dk, Eilam:2001dh}, and warped extra dimensions (RS) from~\cite{Agashe:2006wa, Agashe:2009di} . }
\begin{center}
\begin{tabular}{ccccccc}
\hline
Process & SM
& 2HDM(FV)
& 2HDM(FC)
& MSSM
& RPV
& RS
\\ \hline\hline
$t \to Z u$ & $7 \times 10^{-17}$ & -- & -- & $\leq 10^{-7}$ & $\leq 10^{-6}$ & -- \\
$t \to Z c$ & $1 \times 10^{-14}$ & $\leq 10^{-6}$ & $\leq 10^{-10}$ & $\leq 10^{-7}$ & $\leq 10^{-6}$& $\leq 10^{-5}$ \\
$t \to g u$ & $4 \times 10^{-14}$ & -- & -- & $\leq 10^{-7}$ & $\leq 10^{-6}$& -- \\
$t \to g c$ & $5 \times 10^{-12}$ & $\leq 10^{-4}$ & $\leq 10^{-8}$ & $\leq 10^{-7}$& $\leq 10^{-6}$& $\leq 10^{-10}$\\
$t \to \gamma u$ & $4 \times 10^{-16}$& -- & -- & $\leq 10^{-8}$ & $\leq 10^{-9}$ & --\\
$t \to \gamma c$ & $5 \times 10^{-14}$ & $\leq 10^{-7}$ & $\leq 10^{-9}$ & $\leq 10^{-8}$ & $\leq 10^{-9}$ & $\leq 10^{-9}$\\
$t \to h u$ & $2 \times 10^{-17}$ & $6 \times 10^{-6}$ & -- & $\leq 10^{-5}$ & $\leq 10^{-9}$ & -- \\
$t \to h c$ & $3 \times 10^{-15}$ & $2 \times 10^{-3}$ & $\leq 10^{-5}$ & $\leq 10^{-5}$ & $\leq 10^{-9}$& $\leq 10^{-4}$ \\
\hline\hline
\end{tabular}
\end{center}
\label{tab:theory}
\end{table}%
\subsubsection{SM top FCNCs}
SM contributions to top FCNCs are necessarily small, suppressed by both the GIM mechanism
and by the large total width of the top quark due to the dominant mode $t \to bW$ \cite{Eilam:1990zc, Mele:1998ag}. This essentially guarantees that any measurable branching ratio for top FCNC decays is an indication of new physics. The values in Table \ref{tab:theory} are from the updated numerical evaluation in
reference
\cite{AguilarSaavedra:2004wm}. Note that the results are very sensitive to the value of $m_b$, since they scale as $m_b(m_t)^4$. The difference between decays involving $u$ quark and $c$ quarks arises from the relative factor $|V_{ub}/V_{cb}|^2$.
\subsubsection{BSM top FCNCs}
Many models for new physics predict new contributions to top FCNCs that are orders of magnitude in excess of SM expectations. Extended electroweak symmetry breaking sectors with two Higgs doublets (2HDM) lead to potentially measurable FCNCs. Parametric expectations are particularly large for 2HDM with tree-level flavor violation, for which flavor-violating couplings between Standard Model fermions and the heavy scalar Higgs $H$ or pseudoscalar $A$
are
typically
assumed
to
scale with quark masses, as
$\sqrt{m_q m_t / m_W^2}$, in order to remain consistent with limits on light quark FCNCs. The estimates in Table \ref{tab:theory} are taken from references \cite{Luke:1993cy, Atwood:1996vj}. The flavor-violating decays arise at one loop due to the exchange of $H, A$, and the charged Higgs scalar $H^\pm$, with the rate that depends on both the tree-level flavor-violating couplings between fermions and the heavy Higgs bosons and the masses of the heavy Higgs bosons themselves.
Even when tree-level flavor conservation is guaranteed in the 2HDM by discrete symmetries, the model
predicts measurable top FCNCs due to loop processes that involve the additional charged Higgs bosons. In this case the rate for flavor-violating processes depends on the mass of the charged Higgs and the angle $\tan \beta$ parameterizing the distribution of vacuum expectation values between the two Higgs doublets. In the Type-I 2HDM, the branching ratios are typically small; the most promising candidate is $t \to g c \sim 10^{-8}$, with rates for $t \to h q$ several orders of magnitude smaller. In the Type-II 2HDM, the leading contribution to $t \to h q$ is enhanced by $\mathcal{O}(\tan^4 \beta)$ and may be considerable at large $\tan \beta$. The most optimistic cases are presented in Table \ref{tab:theory}, taken from \cite{Bejar:2006ww} for Type I and Type II 2HDM. However, given that Higgs coupling measurements now constrain the allowed range of mixing angles in these 2HDM, the maximal rates for $t \to h q$ consistent with ancillary measurements are likely smaller.
In the MSSM, top FCNCs arise at one loop in the presence of flavor-violating mixing in the soft SUSY-breaking mass matrices. Flavor violation involving the stops is much more weakly constrained by indirect measurements than flavor violation involving light squarks (particularly in the down-squark sector), allowing for potentially large mixing. However, rapidly-advancing limits on direct sparticle production have pushed the mass scale of squarks and gluinos to $\geq 1$ TeV, suppressing loop-induced branching ratios. To obtain realistic estimates, in Table \ref{tab:theory} we extrapolate the results of \cite{Cao:2007dk} to the case of $m_{\tilde g} \sim m_{\tilde q} = 1$ TeV. If $R$-parity is violated in the MSSM, top decays may also be induced at one loop by
baryon
($B$)
or lepton
($L$)
number-violating RPV couplings. The effects of $B$-violating couplings are larger
by an order of magnitude or more. For the estimates in Table \ref{tab:theory}, we extrapolate the results of \cite{Yang:1997dk, Eilam:2001dh} to $m_{\tilde q} = 1$ TeV; for \cite{Yang:1997dk} we take their coupling parameter $\Lambda = 1$.
In models of warped extra dimensions,
top FCNCs arise when Standard Model fermions propagate in the extra dimension with profiles governed by the corresponding Yukawa couplings. These non-trivial profiles lead to flavor-violating couplings between SM fermions and the Kaluza-Klein (KK) excitations of the SM gauge bosons. Such couplings are largest for the top quark, whose profile typically has the most significant
overlap with the gauge KK modes, and lead to flavor-violating couplings that depend on 5D Yukawa couplings and the mass scale of the gauge KK modes. Appreciable flavor-violating couplings involving the top quark and Higgs boson arise from analogous processes involving loops of fermion KK modes.
{\bf A possible ``Discovery story''}: it is conceivable that the sensitivity of the LHC and the
ILC/CLIC to top FCNCs could lead to the discovery and identification of physics beyond the
Standard Model. An intriguing scenario is the observation of the flavor-violating decay $t \to Zc$ at the LHC with a branching ratio on the order of $10^{-5}$, at the limit of the projected high-luminosity reach. Such a branching ratio would be some nine orders of magnitude larger than the Standard Model expectation and a clear indication of new physics. At the LHC the primary backgrounds to this channel are Standard Model diboson $ZZ$ and $WZ$ production with additional jets, with a lesser component from $Z$+jets and rarer SM top processes $ttW$ and $ttZ$. The diboson backgrounds are fairly well understood and are in excellent agreement with simulations, and even such rare contributions as $ttW$ and $ttZ$ will be well-characterized by the end of the high-luminosity LHC run, making the observation of $t \to Zc$ fairly reliable.
A $t \to Zc$ signal described above is consistent with new physics arising from a variety of models,
such as warped extra dimensions, a composite Higgs, or a flavor-violating two-Higgs-doublet model. Ancillary probes of FCNC processes become crucial for validating the signal and identifying its origin. Some of the most important probes that allow
differentiation between these options are the rare decays $t \to g c, t \to \gamma c$, and $t \to h c$, which have similar reach at the high-luminosity LHC. In the case of warped extra dimensions or a composite Higgs, the corresponding branching ratios for $t \to g c$ and $t \to \gamma c$ are orders of magnitude below the sensitivity of the LHC, but the branching $t \to h c$ may be as large as $10^{-4}$, within the reach of high-luminosity LHC. Thus a signal in $t \to Z c$ with a tentative signal in $t \to h c$ but no other channels would be indicative of warped extra dimensions or a pseudo-Goldstone composite Higgs, see Section~\ref{sec:newphysics}.
Such rates
would also suggest a relatively low KK scale, so that complementary direct searches for heavy
resonances (see Section~\ref{sec:newphysics})
would play a crucial role in testing the consistency of this possibility.
In addition, such a KK scale would also lead to up to a
$10$ \% shift in $\bar{t} t Z$ coupling which can be probed at the
LHC or the ILC/CLIC (see section \ref{sec:topcouplings} of this report).
In contrast, in flavor-violating two-Higgs-doublet models, a visible $t \to Zc$ signal can be accompanied by comparable signals in $t \to g c$ and $t \to h c$, allowing this scenario to be similarly differentiated.
Complementary information can be provided by the ILC. Projections of the $\sqrt{s} = 500$ GeV ILC with 500 fb$^{-1}$ place its sensitivity to $t \to Z q$ coming from a $\gamma^\mu$ spin structure at the level of $10^{-4}$, but sensitivity to $t \to Z q$ in single top production from a $\sigma^{\mu \nu}$ structure at $\sim 10^{-5}$. The observation of comparable $t \to Zc$ signals at the LHC and ILC could then favor a $\sigma^{\mu \nu}$ coupling and rule out candidate explanations such as warped extra dimensions.
\subsection{Current Limits} \label{sec:limits}
Limits on various top FCNC decays have progressed rapidly in the LHC era. We summarize the current best limits from direct searches in Table \ref{tab:current}. CMS places the strongest limit on the decay $t \to Zq$ in the trilepton final state
\cite{Chatrchyan:2012hqa} using the full 8 TeV data set. ATLAS sets a sub-leading limit on $t \to Zq$ using a portion of the 7 TeV data set, but also sets the leading limits on $t \to gq$ via a search for $s$-channel top production \cite{TheATLAScollaboration:2013vha} using 8 TeV data. The Tevatron still maintains the best limits on some rare processes, in particular $t \to \gamma c$ from Run I \cite{Abe:1997fz} and $t \to {\rm invisible}$ from Run II at CDF \cite{CDFinvis}. ZEUS maintains the best inferred limit on $t \to \gamma u$ \cite{Aaron:2009vv}. The Tevatron and HERA limits on $t \to \gamma q$ are expected to be superseded by LHC limits using the $7 \oplus 8$ TeV data set, but to date no official results are available.
The recent discovery of the Higgs allows for limits to be set on $t \to hq$. The ATLAS collaboration sets the current best limit on $t \to hq$ using the diphoton decay of the Higgs with the full $7 \oplus 8$ TeV data set \cite{TheATLAScollaboration:2013nia}. In \cite{Craig:2012vj} a limit was obtained on $t \to h q$ using the 7 TeV CMS multilepton search with 5 fb$^{-1}$ of data, assuming Standard Model branching ratios for a Higgs boson with $m_h = 125$ GeV. Similar limits may be set using the CMS same-sign dilepton search. The CMS multilepton search has recently been updated to $5 \oplus 20$ fb$^{-1}$ of $7 \oplus 8$ TeV data, and now includes $b$-tagged categories useful for constraining $t \to hq$; an official CMS search for $t \to hq$ using multi-leptons is ongoing.
\begin{table}
\caption{Current direct limits on top FCNCs. $(^*)$ denotes unofficial limits obtained from public results. The $q$ in the final state denotes sum over $q=u,c$.}
\begin{center}
\begin{tabular}{ccccc}
\hline\hline
Process & Br Limit & Search & Dataset & Reference \\ \hline
$t \to Z q$ & $7 \times 10^{-4}$ & CMS $t \bar t \to W b + Z q \to \ell \nu b + \ell \ell q$ & 19.5 fb$^{-1}$, 8 TeV & \cite{Chatrchyan:2012hqa} \\
$t \to Z q$ & $7.3 \times 10^{-3}$ & ATLAS $t \bar t \to W b + Z q \to \ell \nu b + \ell \ell q$ & 2.1 fb$^{-1}$, 7 TeV & \cite{Aad:2012ij} \\
$t \to g u$ & $3.1 \times 10^{-5}$ & ATLAS $q g \to t \to Wb$ & 14.2 fb$^{-1}$, 8 TeV & \cite{TheATLAScollaboration:2013vha} \\
$t \to g c$ & $1.6 \times 10^{-4}$ & ATLAS $q g \to t \to Wb$ & 14.2 fb$^{-1}$, 8 TeV &\cite{TheATLAScollaboration:2013vha} \\
$t \to \gamma u$ & $6.4 \times 10^{-3}$ & ZEUS $e^\pm p \to (t {\rm \; or \;} \bar t) + X$ & 474 pb$^{-1}$, 300 GeV & \cite{Aaron:2009vv} \\
$t \to \gamma q$ & $3.2 \times 10^{-2}$ & CDF $t \bar t \to Wb + \gamma q$ & 110 pb$^{-1}$, 1.8 TeV & \cite{Abe:1997fz} \\
$t \to h q$ & $8.3 \times 10^{-3}$ & ATLAS $t \bar t \to Wb + h q \to \ell \nu b +\gamma \gamma q$ & 20 fb$^{-1}$, 8 TeV &\cite{TheATLAScollaboration:2013nia} \\
$t \to h q$ & $2.7 \times 10^{-2}$ & CMS$^*$ $t \bar t \to Wb + h q \to \ell \nu b + \ell \ell q X$ & 5 fb$^{-1}$, 7 TeV &\cite{Craig:2012vj} \\
$t \to$ invis. & $9 \times 10^{-2}$ & CDF $t \bar t \to Wb$ & 1.9 fb$^{-1}$, 1.96 TeV & \cite{CDFinvis} \\ \hline
\end{tabular}
\end{center}
\label{tab:current}
\end{table}%
Indirect limits on top FCNCs may also be set through single top production, $D^0$ oscillations, and neutron EDM limits. At present these limits are not competitive with direct searches at the LHC for final states involving
photons and $Z$ bosons
\cite{Fox:2007in}, though they are comparable for final states involving $h$ \cite{Harnik:2012pb}.
\subsection{Projected Limits} \label{sec:future}
Although current direct limits on flavor-violating top couplings do not appreciably encroach on the parameter space of motivated theories
(compare tables \ref{tab:theory} and \ref{tab:current}),
future colliders should attain meaningful sensitivity
(see table \ref{tab:LHCILC}). Here we will focus on the sensitivity of the $\sqrt{s} = 14$ TeV LHC after 300 and 3000 fb$^{-1}$ of integrated luminosity, as well as the ILC operating at $\sqrt{s} = 250$ and the ILC/CLIC at $500$ GeV, with 500 fb$^{-1}$ of integrated luminosity. The case of the $\sqrt{s} = 250$ GeV ILC is particularly interesting, since it possesses sensitivity to top FCNCs through single-top production via a photon or $Z$ boson.
\subsubsection{LHC projections}
At present, estimates of future LHC sensitivity to top FCNCs arise from two sources: official projections from the European Strategy Group (ESG) report
\cite{ATLAS:2013hta}
and approximate extrapolation from current searches at the 7 and 8 TeV LHC based on changes in luminosity, energy, and trigger thresholds. Table
\ref{tab:LHCILC}
provides a summary of the projected limits at the 14 TeV LHC with 300 and 3000 fb$^{-1}$ integrated luminosity.
\begin{table}
\caption{Projected limits on top FCNCs at the LHC and ILC. ``Extrap.'' denotes estimates based on extrapolation as described in the text.
For the ILC/CLIC, limits for various tensor couplings (i.e., with $\sigma_{ \mu \nu }$ structure) are shown inside ().
}
\begin{center}
\begin{tabular}{ccccc}
\hline\hline
Process & Br Limit & Search & Dataset & Reference \\ \hline
$t \to Z q$ & $2.2\times 10^{-4}$ & ATLAS $t \bar t \to W b + Z q \to \ell \nu b + \ell \ell q$ & 300 fb$^{-1}$, 14 TeV & \cite{ATLAS:2013hta} \\
$t \to Z q$ & $ 7 \times 10^{-5}$ & ATLAS $t \bar t \to W b + Z q \to \ell \nu b + \ell \ell q$ & 3000 fb$^{-1}$, 14 TeV & \cite{ATLAS:2013hta} \\
$t \to Z q$ & $ 5 \, (2) \times 10^{-4 }$ & ILC single top, $\gamma_\mu$ ($\sigma_{\mu \nu}$) & 500 fb$^{-1}$, 250 GeV & Extrap.\\
$t \to Z q$ & $ 1.5 \, (1.1) \times 10^{-4 \, (-5)}$ & ILC single top, $\gamma_\mu$ ($\sigma_{\mu \nu}$) & 500 fb$^{-1}$, 500 GeV & \cite{AguilarSaavedra:2001ab} \\
$t \to Z q$ & $ 1.6 \, (1.7) \times 10^{-3}$ & ILC $t \bar t$, $\gamma_\mu$ ($\sigma_{\mu \nu}$) & 500 fb$^{-1}$, 500 GeV & \cite{AguilarSaavedra:2001ab} \\
\hline
$t \to \gamma q$ & $8\times 10^{-5}$ & ATLAS $t \bar t \to W b + \gamma q$ & 300 fb$^{-1}$, 14 TeV &
\cite{ATLAS:2013hta} \\
$t \to \gamma q$ & $ 2.5 \times 10^{-5}$ & ATLAS $t \bar t \to W b + \gamma q$ & 3000 fb$^{-1}$, 14 TeV &
\cite{ATLAS:2013hta} \\
$t \to \gamma q$ & $ 6 \times 10^{-5}$ & ILC single top & 500 fb$^{-1}$, 250 GeV & Extrap. \\
$t \to \gamma q$ & $ 6.4 \times 10^{-6}$ & ILC single top & 500 fb$^{-1}$, 500 GeV &\cite{AguilarSaavedra:2001ab} \\
$t \to \gamma q$ & $ 1.0 \times 10^{-4}$ & ILC $t \bar t$ & 500 fb$^{-1}$, 500 GeV & \cite{AguilarSaavedra:2001ab} \\
\hline
$t \to g u$ & $4 \times 10^{-6}$ & ATLAS $q g \to t \to Wb$ &300 fb$^{-1}$, 14 TeV & Extrap. \\
$t \to g u$ & $1 \times 10^{-6}$ & ATLAS $q g \to t \to Wb$ & 3000 fb$^{-1}$, 14 TeV & Extrap. \\
$t \to g c$ & $1 \times 10^{-5}$ & ATLAS $q g \to t \to Wb$ & 300 fb$^{-1}$, 14 TeV & Extrap. \\
$t \to g c$ & $4 \times 10^{-6}$ & ATLAS $q g \to t \to Wb$ & 3000 fb$^{-1}$, 14 TeV & Extrap. \\
$t \to h q$ & $2 \times 10^{-3}$ & LHC $t \bar t \to Wb + h q \to \ell \nu b + \ell \ell q X$ & 300 fb$^{-1}$, 14 TeV & Extrap.\\
$t \to h q$ & $5 \times 10^{-4}$ & LHC $t \bar t \to Wb + h q \to \ell \nu b + \ell \ell q X$ & 3000 fb$^{-1}$, 14 TeV & Extrap. \\
$t \to h q$ & $5 \times 10^{-4}$ & LHC $t \bar t \to Wb + h q \to \ell \nu b +\gamma \gamma q$ & 300 fb$^{-1}$, 14 TeV & Extrap. \\
$t \to h q$ & $2 \times 10^{-4}$ & LHC $t \bar t \to Wb + h q \to \ell \nu b +\gamma \gamma q$ & 3000 fb$^{-1}$, 14 TeV & Extrap. \\
\hline
\end{tabular}
\end{center}
\label{tab:LHCILC}
\end{table}%
The ATLAS projections for $t \to q Z, \gamma$ are shown in the table.
At present there is no public document from CMS with projections for 14 TeV sensitivity, nor are there official projections from either collaboration for $t \to gq$ or $t \to hq$.
Estimates for LHC sensitivity to $t \to gq$ and $t \to hq$ are obtained by an approximate extrapolation from current searches accounting for changes in luminosity, energy, and trigger thresholds. While crude, when applied to $t \to Zq$ this procedure agrees to within $\mathcal{O}(10\%)$ with the official ATLAS ESG projections and so provides a useful benchmark in lieu of detailed study.
The $t \rightarrow h q$ entries in the table for the multilepton final state are derived from those in
\cite{Craig:2012vj} by scaling with the luminosity and $t \bar t$ production cross-section.
This implies a 95\% CL limit ${\rm Br}(t \to h q) < 2 \times 10^{-3} (5 \times 10^{-4})$ with 300 (3000) fb$^{-1}$ at 14 TeV.
Similarly, estimates based on \cite{TheATLAScollaboration:2013nia} give a sensitivity (95\% CL limit) in the $\ell \nu b + \gamma \gamma q$ final state of ${\rm Br}(t \to h q) < 5 \times 10^{-4} (2 \times 10^{-4})$ with 300 (3000) fb$^{-1}$ at 14 TeV.
The extrapolation of $t \to gq$ is more delicate, since the process under study involves the $t gq$ anomalous coupling in the production mode.
Using the results from \cite{Gao:2011fx} to extrapolate the observed 7 TeV limit \cite{Aad:2012gd} to 14 TeV, we find
${\rm Br}(t \to g u) < 4 \times 10^{-6} (1 \times 10^{-6})$ with 300 (3000) fb$^{-1}$ at 14 TeV and
${\rm Br}(t \to g c) < 1 \times 10^{-5} (4 \times 10^{-6})$ with 300 (3000) fb$^{-1}$ at 14 TeV.
\subsubsection{Linear collider (ILC/CLIC) projections}
At the ILC/CLIC, sensitivity studies have focused on operation at $\sqrt{s} \geq 500$ GeV in order to probe both $e^+ e^- \to t \bar t, t \to Xq$ as well as the single top process $e^+ e^- \to t q$ due to, e.g., $t Zq$ or $t \gamma q$ anomalous vertices\footnote{As mentioned
in section \ref{sec:topcouplings}, TLEP has larger $t \bar{t}$ samples, but no polarization so that
separating couplings to $\gamma$ from those to $Z$ will be difficult.}. Linear collider performance at $\sqrt{s} \geq 500$ GeV is studied in some detail in \cite{AguilarSaavedra:2001ab}, which forms the basis for sensitivity estimates quoted here. The study \cite{AguilarSaavedra:2001ab} includes 95\% CL estimates for various polarization options, including 80\% $e^-$ polarization and 45\% $e^+$ polarization, which are close to the polarization parameters
planned
for the ILC.
In what follows we quote the 80\%/45\% polarization sensitivity, with the difference between 45\% $e^+$ polarization and 30\% $e^+$ polarization expected to lead to a small effect. We rescale the results of \cite{AguilarSaavedra:2001ab} to 500 fb$^{-1}$ to match the anticipated ILC/CLIC integrated luminosity; the results are presented in Table~\ref{tab:LHCILC}. Based on these estimates, ILC/CLIC sensitivity at $\sqrt{s} = 500$ GeV should be comparable to LHC sensitivity with 3 ab$^{-1}$ for $t \to Zq$ and $t \to \gamma q$. Since much of the sensitivity comes from single top production, the ILC/CLIC is less likely to provide comparable sensitivity to $t \to h q$ and $t \to g q$.
The ILC also provides sensitivity to $t Zq$ and $t \gamma q$ anomalous couplings at $\sqrt{s} = 250$ GeV through single top production via the $s$-channel exchange of a photon or $Z$ boson, $e^+ e^- \to t \bar c + \bar t c$. In fact, production via $Z$ exchange through the $\gamma_\mu$ vertex reaches its maximal cross-section
around 250 GeV and falls with increasing center-of-mass energy. Single top production cross-sections through $\gamma$ exchange or $Z$ exchange through the $\sigma_{\mu \nu}$ coupling grow with increasing energy but are still appreciable at $\sqrt{s} = 250$ GeV . The disadvantage of $\sqrt{s} = 250$ GeV relative to higher center-of-mass energies is primarily the larger SM backgrounds to the single-top final state. {\em In any event, this provides an intriguing opportunity for the ILC to probe new physics in the top sector even when operating below the $t \bar t$ threshold.}
The prospects for constraining $t Zq$ and $t \gamma q$ anomalous couplings at $\sqrt{s} = 250$ GeV have not been extensively studied, but we may extrapolate sensitivity reasonably well based on the results of \cite{Han:1998yr}. To obtain an estimate, we rescale the signal cross-section
after cuts for $e^+ e^- \to t \bar c + \bar t c$ via anomalous couplings at $\sqrt{s} = 192$ GeV in \cite{Han:1998yr} to $\sqrt{s} = 250$ GeV and conservatively assume the background cross-sections are similar between $\sqrt{s} = 192$ GeV and $\sqrt{s} = 250$ GeV. In actuality, the backgrounds should decrease with increasing center-of-mass energy. We assume a 60\% $b$-tag efficiency and arrive at 95\% CL estimates in Table~\ref{tab:LHCILC}.
\subsection{Vts and Vtd}
~
The measurement of the ratio of top decays with $b$-tagging to all top decays is sensitive to the
off-diagonal CKM matrix elements $V_{ts}$ and $V_{td}$~\cite{CMS-PAS-TOP-12-035}. A measurement of this
ratio at the sub-percent level should be possible at the high-luminosity LHC. The rapidity of
the top quark in $t$-channel single top quark production is also sensitive to $V_{ts}$ and
$V_{td}$~\cite{AguilarSaavedra:2010wf}. The
ultimate precision in $V_{ts}$ and $V_{td}$ will come from a combination of the
two methods~\cite{Lacker:2012ek}.
Systematic uncertainties
and their correlations between different measurements will be a limiting factor, but a precision of
better than 0.05 in $|V_{ts}|$ and $|V_{td}|$ should be achievable based on current studies.
\subsection{Summary}
Various well-motivated models predict branching ratios for top FCNC decays starting at $\sim 10^{-4}-10^{-5}$, with the most promising signals arising in two-Higgs-doublet models and various theories with warped extra dimensions. At present the LHC sensitivity to top FCNC decays is somewhat below the level predicted by motivated theories, with the notable exception of $t \to gu$ where searches for resonant single top production yield a limit $\mathcal{O}(10^{-4})$.
However, future colliders, such as the 14 TeV LHC and $\sqrt{s} = 250$ ILC or 500 ILC/CLIC, provide meaningful sensitivity to flavor-violating couplings of the top quark, of the same order as the largest rates predicted in motivated theories.
The LHC and the
ILC/CLIC can be complementary in this regard: while the sensitivities in $t q Z/\gamma$ are (roughly) comparable
for the two colliders, the LHC is better for gluon couplings, but the ILC/CLIC is the way to go for
probing the spin-structure of couplings.
Intriguingly, even at $\sqrt{s} = 250$ GeV the ILC should provide sensitivity to $t \to Zq, \gamma q$ that is
comparable to that
of the high-luminosity LHC.
Finally, going from LHC to HL-LHC can improve reach by roughly a factor of two (in rates).
\section{The top quark mass}
\label{sec:topmass}
~
The top quark mass is a parameter whose precise value
is essential for testing the overall consistency of the Standard Model or
models of new physics through precision electroweak fits.
The exact value of the top quark mass is also crucial for understanding whether the Standard Model
{\it without further extensions} can be continued to energies compared to the Planck scale,
without running into problems with the stability of electroweak vacuum \cite{Degrassi:2012ry}.
To put both of these statements into perspective, we note that
the value of the top quark mass, as quoted by the Particle Data Group, is $m_t = 173.5 \pm 0.6 \pm 0.8~{\rm GeV}$.
The total uncertainty on $m_t$ is therefore close to $1~{\rm GeV}$; this is the best relative
precision available for {\it any} of the quark masses.
Nevertheless, we know that for precision electroweak fits, a $0.9~{\rm GeV}$ uncertainty
in the top quark mass corresponds to a $5.4~{\rm MeV}$ uncertainty in the $W$-mass
(see e.g. Refs.~\cite{pew} and~\cite{Baak:2012kk}). Since the $W$-mass is expected to be
measured with this precision at the LHC, and significant improvements in $\delta M_W$
beyond this are not likely, we conclude that the future of
precision electroweak physics requires the measurement of the top quark mass
to at least a precision of less than $0.5~{\rm GeV}$, and desireably to 0.3~GeV so that the
top sector is not limiting in EW precision fits.
In addition, the vacuum stability issue depends
strongly on the value of the top quark mass. Indeed, as shown in Ref.~\cite{Degrassi:2012ry},
changing $m_t$ by $2.1~{\rm GeV}$ around the central value $m_t = 173.1~{\rm GeV}$,
the energy scale where the Higgs potential becomes unstable changes
by {\it six} orders of magnitude, from $\mu_{\rm neg} \sim 10^8~{\rm GeV}$
to $\mu_{\rm neg} \sim 10^{14}~{\rm GeV}$!
It is easy to estimate that if $m_t$ is known
with $0.3-0.5~{\rm GeV}$ uncertainty, as required by the electroweak fit, the scale can
be estimated much more precisely, to within a factor of five.
We conclude that the knowledge of the top quark mass with the $0.5~{\rm GeV}$ uncertainty will have an
important impact on our understanding of particle physics.
Furthermore, it has recently been suggested~\cite{tlep} that a much more precise
measurement of the $W$ mass can be performed at a
circular $e^+e^-$ collider such as TLEP, where $\delta M_W \le 1.5~{\rm MeV}$ can probably be achieved.
For the purpose of precision electroweak
fits, such high precision can be only utilized if the top quark mass is measured with the matching precision of about $0.1~{\rm GeV}$.
As we explain below, this can be accomplished at an $e^+e^-$ collider such as the ILC, CLIC,
or TLEP itself. Knowing
$m_t$ with such a precision will also allow for a much more decisive tests of the vacuum
stability problem in the Standard Model.
The interest in testing this scenario
may increase greatly if no new physics at the TeV scale is found in the Run II of the LHC.
Note that for these purposes, a numerical value for
theoretically well-defined top quark mass parameter, for example $m_t^{\overline {\rm MS}}$, is required.
\subsection{Linear Colliders}
A $e^+e^-$ collider will allow us to study electroweak production of $t\overline{t}$
pairs with no concurring QCD background.
Therefore, precise measurements of top quark properties become possible.
The top quark mass can be measured at $e^+e^-$ machines using two complementary methods.
First, one can use the invariant mass of the reconstructed $bW$ system from the top decay.
The result of a full simulation study at a 500~GeV linear collider~\cite{Seidel:2013sqa} (CLIC,
with similar results for ILC) is shown in Fig.~\ref{fig:TopMass}.
\begin{figure}
\centering
\includegraphics[width=0.49\textwidth]{TopQuark/detectors/figs/TopMassFH}
\includegraphics[width=0.49\textwidth]{TopQuark/detectors/figs/TopMassSL}
\caption{Distribution of reconstructed top mass for events classified as fully-hadronic (left) and semileptonic (right). The data points include signal and background for an integrated luminosity of 100\,fb$^{-1}$~\cite{Seidel:2013sqa}. The pure background contribution contained in the global distribution is shown by the green solid histogram. The top mass is determined with an unbinned likelihood fit of this distribution, which is shown by the solid line.}
\label{fig:TopMass}
\end{figure}
The figure demonstrates also the small residual background expected for top quark studies at
any $e^+e^-$ machines. In the second
method the top mass is determined in a threshold scan, an option unique to an $e^+e^-$ machine.
In the threshold scan the so-called 1S top quark mass can be measured to an experimental precision of better than
40~MeV where studies have shown that the statistical error is dominant. Expressing the measurement in terms of
the theoretically well defined \mbox{$\overline{\rm MS}$} mass will inflate the uncertainty
to $\sim100$~MeV, as shown in detailed simulations~\cite{Martinez:2002st, Seidel:2013sqa,Asner:2013hla}
and advanced theoretical computations ( see e.g. Ref.~\cite{Beneke:2008ec} and references therein).
We note that with respect to the top quark mass determination, all lepton colliders that were
suggested so far perform similarly\footnote{We note that some improvements in the $m_t$
determination can be expected at the muon collider and at TLEP thanks to reduced beamstrahlung,
although this still has to be demonstrated by detailed simulations.}
and that an additional attraction of measuring $m_t$ at a lepton collider is a clean theoretical
interpretation of the result of the measurement. As we explain below, the situation is
more confusing at a hadron collider, although new methods
for $m_t$ measurements developed at the LHC help to mitigate this difference.
\subsection{Top quark mass at the LHC}
As previously noted, a precision of $0.5~{\rm GeV}$ or better in the top quark mass
is motivated by the future of precision electroweak fits. It is an interesting question
whether $m_t$ measurements with such
a precision can be accomplished
at the LHC. To answer it, we will first make some general remarks about measurements of $m_t$.
Existing measurements of the top quark mass
rely on complex techniques required by the difficult hadron collider environment.
The highest accuracy is currently achieved using the so-called matrix-element
method ( for a recent review, see \cite{Gainer:2013iya}).
We will explain a generic measurement of the top quark mass by considering the following example.
Any measurement of the top quark mass is based on fitting a particular piece of data to a theory
prediction where $m_t$ enters as a free parameter. Hence, we write
\begin{equation}
D = T(m_t,\alpha_s,\Lambda_{\rm QCD} ) = T^{(0)}(m_t) + \frac{\alpha_s}{\pi} T^{(1)}(m_t) + {\cal O}(\Lambda_{\rm QCD}/m_t,\alpha_s^2),
\label{eq0}
\end{equation}
where $D$ on the left hand side is a particular kinematic distribution measured in experiment and $T$ on the right-hand side is a theoretical
prediction, expanded in power series in the strong coupling constant.
We have indicated in Eq.(\ref{eq0}) that the selected distribution should be minimally affected by
non-perturbative corrections; we will return to this point below. We also note that inclusion of QCD corrections
necessitates a clear definition of the renormalization scheme, which then fixes the mass parameter extracted from
the fit. Since the two popular choices of the renormalized mass parameter, the pole mass and the $\overline {\rm MS}$ mass,
differ by almost $7~{\rm GeV}$, the specification of the renormalization scheme in the extraction of the top quark
mass is an important issue. Solving Eq.(\ref{eq0}), we find the top quark mass $m_t$. In general, the quality
of such a solution
depends on the {\it accuracy} of the theoretical prediction that we have in the right hand side, which is
controlled by the order in perturbation theory included there. The majority
of current analyses are performed with
leading order theoretical tools. This amounts to setting $T^{(1)} \to 0$ in the above equation.
The expected error on $m_t$ is then
\begin{equation}
\delta m_t \sim \frac{\alpha_s}{\pi} \frac{T^{(1)}}{T^{(0)'} }
\sim \frac{\alpha_s m_t}{\pi} \frac{T^{(1)}}{T^{(0)}}
\sim \frac{\alpha_s}{\pi} m_t \sim 6~{\rm GeV},
\label{eq1}
\end{equation}
where $T^{(0)'} = {\rm d} T^{(0)}/{\rm d} m_t$ and we used $T^{(0)'} \approx T^{(0)}/m_t$. It is obvious
from Eq.(\ref{eq1}) that the estimated error in Eq.(\ref{eq1})
is {\it significantly larger} than the current ${\cal O}(1)~{\rm GeV}$ error on $m_t$.
We conclude that if $m_t$ is obtained from a generic distribution at leading order, one can
not, in general, expect the accuracy that is better than few GeV. Fortunately, there are two ways to get around this problem.
The first one requires inclusion of NLO QCD corrections into a theory prediction; effectively, this pushes the error to
$m_t (\alpha_s/\pi)^2 \sim 0.3 ~{\rm GeV}$ which is acceptable. The second one amounts to
finding a kinematic distribution which has a {\it strong} dependence
on $m_t$; in this case, ${\rm d}T^{(0)}/{\rm d} m \gg T^{(0)}/m_t$ and the estimate in Eq.(\ref{eq1})
receives an additional suppression.
As we show below, {\it new}
experimental techniques that address the question of the
top quark mass determination follow the two approaches described above.
The above discussion can be used to argue
that {\it well-established} methods for the top quark mass determination may have additional
systematic errors which are not accounted for in their error budgets.
Indeed, the matrix element method\footnote{The template method \cite{ATLAS:2012aj} is subject to similar arguments.}
is designed to maximize probabilities for
kinematics of observed events by
adjusting values of the top quark mass on an event-by-event basis.
It can be thought therefore as an attempt to fit a very large
number of kinematic distributions for the best value of $m_t$.
An unsatisfactory feature of this method is its ``black-box'' nature that
does not allow one to understand which kinematic features of the top quark pair production process drive this sensitivity.
While such methods -- by design -- should find distributions that show strong dependence on $m_t$,
it is not clear if the relevant distributions are sensitive to non-perturbative effects whose description
from first principles is not possible. Moreover, such approaches routinely rely on the use of parton
shower event generators instead of proper QCD theory. This means that Eq.(\ref{eq0}) becomes
\begin{equation}
D = T(m_t,\alpha_s,\Lambda_{\rm QCD}) \approx T^{(0)}_{\rm MC}(m_t,\alpha_s,\Lambda_{\rm QCD},
{\rm tunes}),
\label{eq2}
\end{equation}
where, as indicated in the last step, additional approximations, including parton shower tunes, are performed on the ``theory'' side.
While the quality of this approximation
{\it for the purpose of determining $m_t$} may be good,
it is simply not clear how to assign the error
to the parameter $m_t$ which is extracted following this procedure.
To make this problem explicit, the top quark mass extracted from Eq.(\ref{eq2}) should be properly referred to
as the ``Monte-Carlo
mass '', whose relation to short-distance definition
of $m_t$ is not understood.
\begin{table}[t]
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|}
\hline
& {\rm Ref.}\cite{Chatrchyan:2012cz} & \multicolumn{5}{c|} {\rm Projections}\\
\hline
{\rm CM Energy} & {\rm 7 TeV } & \multicolumn{5}{c|}{\rm 14 TeV } \\
\hline
{\rm Luminosity} & {$5 fb^{-1}$} & \multicolumn{2}{c|}{$100 fb^{-1}$} & \multicolumn{2}{c|}{$300 fb^{-1}$} & {$3000 fb^{-1}$} \\
\hline
{\rm Pileup} & 9.3 & 19 & 30 & 19 & 30 & 95 \\
\hline\hline
{\rm Syst. (GeV)} & 0.95 & 0.7 & 0.7 & 0.6 & 0.6 & 0.6 \\
\hline
{\rm Stat. (GeV)} & 0.43 & 0.04 & 0.04 & 0.03 & 0.03 & 0.01 \\
\hline\hline
{\bf Total, GeV } & {\bf 1.04} & {\bf 0.7} & {\bf 0.7} & {\bf 0.6} & {\bf 0.6} & {\bf 0.6} \\
\hline
\end{tabular}
\vspace*{0.5cm}
\caption{Precision of the
top quark mass measurements that can be expected using conventional (likelihood-type) methods.
Extrapolations are based on the published CMS lepton-plus-jets analysis. An additional $0.3~{\rm GeV}$
systematic error was added to all extrapolated results.}
\label{Table_massconv}
\end{center}
\end{table}
In spite of the caveats with the top quark mass determination that are inherent to conventional methods,
it is interesting to estimate precision in $m_t$ that can be achieved at the LHC.
We do that using extrapolations of what has been accomplished at the Tevatron and during the run I of the LHC.
In Table~\ref{Table_massconv} we show such projections for conventional methods assuming that the
mass is measured in the lepton + jet channel for the $14~{\rm TeV}$ LHC for different integrated luminosities
and pile-up scenarios. We assume the $t \bar t$ production
cross-section to be $\sigma_{pp \to t \bar t} = 167(951)~{\rm pb}$ at $7$ and $14$ TeV LHC, respectively.
It follows from Table~\ref{Table_massconv} that conventional
methods may, eventually, lead to the measurement of the top quark mass with an error of about $0.6~{\rm GeV}$ and
that this error is totally dominated by systematic uncertainties.
It is interesting to point out that precision in $m_t$ saturates for the integrated luminosity of $300~{\rm fb}^{-1}$
and that there is no benefit of using yet higher luminosity for the top quark mass measurement.
The reason for this is the increased pile-up and related degradation
of the jet energy scale determination in the high-luminosity environment, see a detailed
discussion in Section~\ref{sec:unboosted}. Note, however, that the systematic
error estimate in Table~\ref{Table_massconv} includes $0.3~{\rm GeV}$ that was added to all extrapolated
results to account for unforeseen sources of systematics; if we omit this $0.3~{\rm GeV}$ uncertainty, the
uncertainty on the top quark mass measurement becomes very small.
Conceptual problems with conventional methods can be mitigated be measuring the top quark mass from
well-defined kinematic distributions which, on the one hand,
are sufficiently sensitive to $m_t$ and, on the other hand, can be cleanly interpreted in terms of a particular
type of the top quark mass. The latter requirement forces us to select kinematic distributions that are infra-red
safe, so that their computations in higher-orders of QCD perturbation theory can be performed.
In addition, methods for measuring the top quark mass should, ideally, be immune
to contamination from beyond the Standard Model physics -- a scenario that is conceivable if there is top-like BSM physics
at the energy scale close to $2 m_t$. For example, if
$m_t$ is determined from the total cross-section $\sigma_{pp \to t \bar t}$ and if the measurement
of $pp \to t \bar t$
receives unknown contributions from top-like BSM physics,
the extracted value of the top quark mass will be smaller than the true $m_t$. This scenario can
occur for example in SUSY models with light stop squarks $m_{\tilde t} \sim m_t$ that are still
not excluded experimentally (cf. discussion in Section~\ref{sec:stops}).
\begin{table}[t]
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|}
\hline
& {\rm Ref.}\cite{Chatrchyan:2013boa} & \multicolumn{3}{c|} {\rm Projections}\\
\hline
{\rm CM Energy} & {\rm 7 TeV } & \multicolumn{3}{c|}{\rm 14 TeV } \\
\hline
{\rm Luminosity} & {$5 fb^{-1}$} & {$100 fb^{-1}$} & {$300 fb^{-1}$} & {$3000 fb^{-1}$} \\
\hline\hline
{\rm Syst. (GeV)} & 1.8 & 1.0 & 0.7 & 0.5 \\
\hline
{\rm Stat. (GeV)} & 0.90 & 0.10 & 0.05 & 0.02 \\
\hline\hline
{\bf Total } & {\bf 2.0} & {\bf 1.0} & {\bf 0.7} & {\bf 0.5} \\
\hline
\end{tabular}
\vspace*{0.5cm}
\caption{Projections for the uncertainty in $m_t$ determined using the CMS end-point
method \cite{Chatrchyan:2013boa}. Extrapolations are based on the published CMS analysis.}
\label{endpoint}
\end{center}
\end{table}
Methods for top quark mass determination that are based on the
analysis of kinematic distributions of top quark decay products are as close to an ideal method as possible.
The main reason is that, up to small effects related to selection cuts and combinatorial backgrounds,
the kinematic variables involved in the analysis can often be chosen to be Lorentz invariant in which
case they decouple the production stage
from the decay stage. This minimizes impact of any physics, BSM or SM, related to $t \bar t$ production on the top
quark mass measurement. Some of these methods are also insensitive to the physics of top quark decay
and are entirely driven by energy-momentum conservation.
We will describe two
of the methods that belong to this category -- the ``end-point'' method developed recently by the CMS
collaboration \cite{Chatrchyan:2013boa}
and the $``J/\psi''$ method suggested long ago in Ref.~\cite{Kharchilava:1999yj}.
The idea of the end-point method is based on the observation that the invariant mass distribution of a lepton
and a $b$-jet contains a relatively
sharp edge whose position is correlated with $m_t$. Therefore, by measuring the position of this end-point,
one can determine the top quark mass. The number of events close to the end-point
is fitted to a linear combination of a flat background and a linear function
$N_{lb} \sim N_{\rm bck} + S ( m_{\rm lb} - m_{0} ) $, where $m_0$ gives the position of the end-point.
The attractive feature of this method is that it is (almost) independent of any assumption about the matrix element
and that it clearly measures either the pole mass {\it or} some ``kinematic'' mass which is close to it.
At the small expense of being more model-dependent, one can actually improve on this method by utilizing
not {\it only} the position of the end-point but also the shape of the $m_{\rm lb}$ distribution.
Note that away from the kinematic end-point the shape of $m_{lb}$ distribution is accurately predicted through NLO QCD including
off-resonance contributions and signal-background interferences \cite{Denner:2012yc,Bevilacqua:2010qb}. Close
to the end-point re-summed predictions are probably required and are not available at present.
Nevertheless, even without potential improvements, the
end-point method offers an interesting alternative to conventional methods. Uncertainties in $m_t$ that
one may hope to achieve are estimated in Table~\ref{endpoint}.
We note that by using the end-point method we {\it do gain in precision
by going to high-luminosity LHC}. Our projections show that the error as small as $0.5~{\rm GeV}$ can be reached.
The dominant contributions to systematic uncertainty for
each of these studies are the jet-energy scale and hadronization uncertainties. Similar to estimates of
$\delta m_t$ that can be achieved using conventional methods, we add $300~{\rm MeV}$ to the systematic uncertainty
in Table~\ref{endpoint}, to account for unforeseen sources of the systematics.
Another approach to measuring the top quark mass that is very different from conventional ones
is the so-called $J/\psi$ method~\cite{Kharchilava:1999yj}.
Here the top quark mass is obtained from fits to the invariant mass distribution of three leptons
from the {\it exclusive} decays
of the top quark $t \to e B \to e J/\psi X \to eee X$, where $X$ denotes light hadrons.
The extrapolations for the $J/\psi$-method are shown in Table~\ref{JPsi}.
The attractive feature of this approach is its absolute complementarity to more traditional methods discussed
above. The uncertainties in case of the $J/\psi$ method are dominated by statistical uncertainties for
luminosities below $100~{\rm fb}^{-1}$ and by
theory uncertainties for higher luminosities. The theory uncertainties in $m_t$
are estimated to be of the order of $1~{\rm GeV}$; they are caused
by scale and parton distribution functions
uncertainties and by uncertainties in $b \to B$ fragmentation function. Some reduction of theory
uncertainties can be expected, although dramatic improvements in our knowledge of the fragmentation function
are not very likely. This is reflected in the change of the theory error shown in Table~\ref{JPsi}
for $14~{\rm TeV}$ LHC with $3000~{\rm fb}^{-1}$ where it is assumed that NNLO QCD computation
of the exclusive production of $J/\psi$ in $t \bar t$ events will become available and that the scale
uncertainty will be reduced by a factor of two.
\begin{table}[t]
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|}
\hline
& {\rm Ref. analysis} & \multicolumn{5}{|c|}{\rm Projections} \\
\hline
{\rm CM Energy} & {\rm 8 TeV } & \multicolumn{3}{c|}{\rm 14 TeV } & {\rm 33 TeV} & {\rm 100 TeV}\\
\hline
{\rm Luminosity} & {$20 fb^{-1}$} & {$100 fb^{-1}$} & {$300 fb^{-1}$} & {$3000 fb^{-1}$} & {$3000 fb^{-1}$} & {$3000 fb^{-1}$} \\
\hline\hline
{\rm Theory (GeV)} & - & 1.5 & 1.5 & 1.0 & 1.0 & 0.6 \\
\hline
{\rm Stat. (GeV)} & 7.00 & 1.8 & 1.0 & 0.3 & 0.1 & 0.1 \\
\hline
{\bf Total} & {\bf -} & {\bf 2.3} & {\bf 1.8} & {\bf 1.1} & {\bf 1.0} & {\bf 0.6}\\
\hline\hline
\end{tabular}
\vspace*{0.5cm}
\caption{Extrapolations of uncertainties in top quark mass measurements that can be
obtained with the $J/\Psi$ method.}
\label{JPsi}
\end{center}
\end{table}
We note that other methods of measuring $m_t$ with relatively high precision are possible and
were, in fact, discussed in the literature. On the experimental side, the three-dimensional
template fit method
was recently presented by the ATLAS collaboration \cite{ATLAS:2013coa}. The key idea here
is to determine the top quark mass, the light-quark jet energy scale and the $b$-quark jet energy
scale from a simultaneous fit to data, thereby transforming a large part of the systematic uncertainty
related to jet energy scales to a statistical one. While this measurement determines the ``Monte Carlo''
mass and the error on this measurement is not competitive
with other $m_t$-determinations at the moment, its key idea can be applied in conjunction
with other methods and will, hopefully, help to reduce systematic uncertainties.
Another potentially interesting
opportunity is provided by the top quark mass measurements based on
exploiting $m_t$-dependence of lepton kinematic distributions.
Although such studies were not actively pursued experimentally, they may offer
an interesting avenue for the top quark mass measurement in the high-pile-up scenario given
their independence of jet energy scale uncertainties. Theoretical studies of some lepton distributions
and their sensitivity to $m_t$ were performed through NLO QCD in Ref.~\cite{Biswas:2010sa} with
the conclusion that an ${\cal O}(1.5)~{\rm GeV}$ error on $m_t$ can be achieved; further studies
that include more realistic estimates of uncertainties are clearly warranted. Finally, it was proposed
recently to employ $t \bar t j$ events to constrain the top quark mass
\cite{Alioli:2013mxa}. This method is clean theoretically and appears
to be feasibly experimentally; as shown in Ref.~\cite{Alioli:2013mxa},
an ${\cal O}(1)~{\rm GeV}$ uncertainty in $m_t$ can be achieved.
The top quark width of 1.4~GeV is too narrow to be measured directly at the LHC. It can be probed
indirectly through single top quark production~\cite{Abazov:2012vd}, which can be determined
to about 5\% at high-luminosity LHC,
see Section~\ref{sec:topcouplings}. The width can be measured directly to a few percent
through a top pair threshold scan at a lepton collider~\cite{Baer:2013cma,Martinez:2002st}.
We conclude by making a general remark about the future of the top quark measurements at a hadron collider.
While hadron collider measurements of the top quark mass {\it cannot} compete with $e^+e^-$ colliders,
our discussion shows that it is possible to have a number of top quark measurements at the LHC,
including the high-luminosity option, which are clean theoretically and show high sensitivity to $m_t$.
It is also important to stress that these measurements are typically limited by different types of uncertainties, so that
combining their results under the assumption that errors are uncorrelated is a reasonable thing to do. A combination
of the results of different measurements, that determine theoretically well-defined top quark mass,
can lead to further reduction in the error on $m_t$ that
is achievable at the LHC,
pushing it into a $0.3 - 0.4~{\rm GeV}$ range. Further reduction of the uncertainty
in the top quark mass determination is possible at suggested $e^+e^-$ machines (ILC, CLIC, TLEP). Such measurements
are important for testing if the Standard Model {\it without further extensions} can be consistently extrapolated
to Planckian energy scales. Interest in such studies should increase if no new physics is found at the Run 2 at the LHC.
\chapter{Top quark working group report}
\label{chap:top}
\begin{center}\begin{boldmath}
\input TopQuark/authorlist.tex
\end{boldmath}\end{center}
\input TopQuark/intro.tex
\input TopQuark/topmass/topmass.tex
\input TopQuark/couplings/couplings.tex
\input TopQuark/kinematics/kinematics.tex
\input TopQuark/raredecays/raredecays.tex
\input TopQuark/newphysics/newphysics.tex
\section{Top Algorithms and Detectors}
\label{sec:detector}
\input{TopQuark/detectors/unboosted_summary}
\input{TopQuark/detectors/boosted_summary}
\input{TopQuark/detectors/lc_summary}
\input TopQuark/conclusions.tex
|
\section{Introduction}
Consider the following optimal affine control problem (P):
\begin{align*}
\mbox{minimize } & \int_{0}^{T_f(u)} f_0(x)+\sum_{i=0}^{m}g_{0i}(x)u_i \; dt \\
\mbox{subject to } & \begin{cases}
\dot{x} = f(x)+\sum_{i=0}^{m}g_i(x)u_i \\
u=(u_1,\dots,u_m):[0,T_f(u)]\rightarrow\R{m} \mbox{ such that }\\
|u_i(t)|\leq K(t),\; \forall t\in [0,T_f(u)]&, i=1, \dots, m \\
x(0) = A \\
x(T_f(u)) = B
\end{cases}
\end{align*}
where:
\begin{enumerate}
\item $x=(x_1,\dots,x_n)\in\R{n}$ and $u=(u_1,\dots,u_m)\in\R{m}$
\item $f$, $g_i$, $i=1, \dots, m$, are analytic vector fields in $\R{n}$
\item $f_0$, $g_{0i}$, $i=1, \dots, m$, are analytic maps from $\R{n}$ to $\R{}$
\item $K$ is analytic and strictly positive
\item $u_i \in \mes{\R{}}$, $i=1, \dots, m$
\end{enumerate}
From the Pontryagin's Maximum Principle{}, if one defines the Hamiltonian function:
\begin{align*}
H_\lambda:T^*\R{n} \times \R{m} &\longrightarrow \R{} \\
(x,p,u) &\longmapsto \innerprod{p}{f(x)+\sum_{i=0}^{m}g_i(x)u_i} - \lambda \left( f_0(x)+\sum_{i=0}^{m}g_{0i}(x)u_i \right)
\end{align*}
where $(x,p) \in T_x\R{n}$ and $\lambda \in \{0,1\}$, then each optimal trajectory $(\bar{x},\bar{u}):[0,\bar{T}] \rightarrow \R{n} \times \R{m}$ has a lift to the cotangent space such that $H_\lambda(\bar{x}(\bar{T}),\bar{p}(\bar{T}),\bar{u}(\bar{T}))=0$ and
\begin{equation*}
\mbox{(Adj) }
\begin{cases}
\frac{d\bar{x}}{dt}(t)=\frac{\partial H_\lambda}{\partial p}(\bar{x}(t),\bar{p}(t),\bar{u}(t)) \\
\frac{d\bar{p}}{dt}(t)=-\frac{\partial H_\lambda}{\partial x}(\bar{x}(t),\bar{p}(t),\bar{u}(t)) \\
H_\lambda(\bar{x}(t),\bar{p}(t),\bar{u}(t))=\sup\left\{ H_\lambda(\bar{x}(t),\bar{p}(t),v) | v \in U \right\}
\end{cases}
\end{equation*}
for almost all $t \in [0,\bar{T}]$ and $\lambda \in \{0,1\}$. Also $(\lambda,\bar{p}(t))\neq 0$, for almost all $t\in [0,\bar{T}]$. The solutions of (Adj) are called extremals and might not be an optimal solution of the original problem. The new variable $p$ is known as the \emph{adjoint variable}.
Setting
\begin{align*}
\bar{f}&=(f_0,f) & \bar{g}_i&=(g_{0i},g_i) \\
\bar{x}&=(x_0,x) & \bar{p}&=(\lambda,p)
\end{align*}
where $x_0$ satisfies:
\begin{equation*}
\dot{x}_0=f_0(x)+\sum_{i=0}^{m}g_{0i}(x)u_i,
\end{equation*}
the Hamiltonian function become $H_\lambda=\innerprod{\bar{p}}{\bar{f}} + \sum_{i=i}^{m} \innerprod{\bar{p}}{\bar{g}_i}u_i$ and one has:
\begin{equation} \label{e:adj_simples}
\begin{aligned}
\dot{\bar{x}}&=\frac{\partial H_\lambda}{\partial \bar{p}} = \bar{f}+\sum_{i=0}^{m}\bar{g}_i(x)u_i\\
\dot{\bar{p}}&=-\frac{\partial H_\lambda}{\partial \bar{x}} = -\bar{p}\frac{\partial f}{\partial x} - \sum_{i=1}^{m} \bar{p}\frac{\partial g_i}{\partial x}u_i.
\end{aligned}
\end{equation}
Henceforth, the simplified notation $f, g_i, x$ and $p$ stands for $\bar{f}, \bar{g}_i, \bar{x}$, respectively. Note that the dimension of the new problem is $n+1$.
Since the Hamiltonian $H_\lambda$ is linear in $u$, one has by the Pontryagin's Maximum Principle{} that $u_i(t)=\sign{\innerprod{p(t)}{g_i(x(t))}} K(t)$ on the intervals where $\innerprod{p(t)}{g_i(x(t))}$ is non zero almost everywhere. These are known as the \emph{nonsingular intervals}. In this case the control is said nonsingular on these intervals.
Analogously, the intervals where $\innerprod{p(t)}{g_i(x(t))}$ vanishes almost everywhere are known as \emph{singular intervals} and the controls are called singular on these intervals. Along a singular interval the control can not be designed by the Pontryagin's Maximum Principle{}, but one could consider the time derivatives of $\innerprod{p(t)}{g_i(x(t))}$.
Indeed, in a singular interval, the time derivatives of the column vector $\left[\innerprod{p}{g_i}\right]_{i=1,\dots,m}$ could be evaluated until a relation that depends on $u$ explicitly is obtained. In other words, for $l \in \N{}$, one has the $m \times m$ matrix:
\begin{equation*}
B_l = \frac{\partial}{\partial u} \left( \frac{d^l}{dt^l} \left[\innerprod{p}{g_i}\right]_{i=1,\dots,m} \right).
\end{equation*}
Consider the first one that is not identically zero, say the $k$-th derivative $B_k$. Then one has:
\begin{equation*}
0 = \frac{d^k}{dt^k} \left[\innerprod{p}{g_i}\right]_{i=1,\dots,m} = A_k(x,p) + B_k(x,p)u.
\end{equation*}
If $B_{k}$ is nonsingular, all the controls can be evaluated. Otherwise, it is necessary to find some controls, reducing the problem and restarting the procedure. It's now clear the central role of the functions $\innerprod{p(t)}{g_i(x(t))}$ and its derivatives on the design of optimal controls.
The number $q=k/2$ is called \emph{problem order} or \emph{intrinsic order}. Note that even if $B_k$ is not identically zero it can became singular, or even identically zero, along an specific extremal. This fact lead us to another concept of order, know as \emph{arc order} or \emph{local order}. There were a lot of confusion around these concepts of order, first noticed by Lewis in 1980 \cite{LEWIS:1980}. There is one issue, although, that was not mentioned by Lewis, which will be treated in this paper.
When Robbins, in 1967 \cite{ROBBINS:1967}, has enunciated and proved the Generalized Legendre-Clebsch (GLC) Condition, he was not aware of these order concepts, so he did not make clear which order he was considering. For that reason, some authors, including Lewis, have thought that the GLC Condition was valid for the problem order. Unfortunately it's not true, since the problem order can be a fraction, i.e., $k$ can be odd. Consider, for instance, the following optimal control problem:
\begin{align*}
\dot{x} &= v_1 \cos \theta + v_2 \sin \theta &
\dot{y} &= v_2 \cos \theta - v_1 \sin \theta &
\dot{\theta} &= \Omega \\
\dot{v}_1 &= u_1 &
\dot{v}_2 &= u_2 &
\dot{\Omega} &= u_3.
\end{align*}
No matter what functional one wants to optimize, the intrinsic order of this problem is always $\tfrac{3}{2}$. However, if one wants to minimize $\int_O^T x^2+y^2+\theta^2 dt$, it's easy to see that, been at the origin, stay at the origin with a fully singular trajectory is the optimal solution. Fortunately, in this case the matrix $B_k$ is identically zero, so this example does not contradict the GLC Condition, if one considers it along the trajectory, in other words, if we consider the GLC Condition with the local order.
The problem order fail to be an integer mainly because there are several inputs in this example. In the case one has just one input, the problem order is always a positive integer, as will be proved soon.
\section{Evaluating the problem order}
To evaluate the problem order a simple known lemma is necessary.
\begin{lem} \label{l:derivada_de_<p,h>}
Let $h$ be a smooth vector field. Then along an extremal we have:
\begin{equation*}
\frac{d}{dt}\innerprod{p}{h}=\innerprod{p}{\lieprod{f}{h}+\sum_{i=1}^{m}u_i\lieprod{g_i}{h}}.
\end{equation*}
\end{lem}
\begin{proof}
We know that
\begin{equation*}
\frac{d}{dt}\innerprod{p}{h}=\innerprod{\dot{p}}{h} + \innerprod{p}{\frac{\partial h}{\partial x}\dot{x}}.
\end{equation*}
From the equations \eqref{e:adj_simples} and the inner product linearity, the first portion of the sum above is:
\begin{equation*}
\innerprod{\dot{p}}{h}= \innerprod{p}{-\frac{\partial f}{\partial x}h} + \sum_{i=1}^{m}\innerprod{p}{-\frac{\partial g_i}{\partial x}}u_i
\end{equation*}
and the second one is:
\begin{equation*}
\innerprod{p}{\frac{\partial h}{\partial x}\dot{x}}=\innerprod{p}{\frac{\partial h}{\partial x}f}+\sum_{i=1}^{m}\innerprod{p}{\frac{\partial h}{\partial x}g_i}u_i
\end{equation*}
from which follows the result.
\end{proof}
One uses this lemma to evaluate the derivatives of $\phi = \left[\innerprod{p}{g_i}\right]_{i=1,\dots,m}$:
\begin{equation}
\phi^{(1)} =
\left[\begin{array}{c}
\frac{d}{dt}\innerprod{p}{g_1} \\
\vdots \\
\frac{d}{dt}\innerprod{p}{g_m}
\end{array}\right]
=
\left[\begin{array}{c}
\innerprod{p}{ \ad{f}{}{g_1} + \sum_{i=1}^{m} \lieprod{g_i}{g_1}u_i } \\
\vdots \\
\innerprod{p}{ \ad{f}{}{g_m} + \sum_{i=1}^{m} \lieprod{g_i}{g_m}u_i }
\end{array}\right].
\end{equation}
If the first derivative do not depend on $u$, in other words, if $\lieprod{g_i}{g_j}=0$, $\forall i,j$, then the second derivative is evaluated:
\begin{equation}
\phi^{(2)} =
\left[\begin{array}{c}
\frac{d}{dt}\innerprod{p}{ \ad{f}{}{g_1} } \\
\vdots \\
\frac{d}{dt}\innerprod{p}{ \ad{f}{}{g_m} }
\end{array}\right]
=
\left[\begin{array}{c}
\innerprod{p}{ \ad{f}{2}{g_1} + \sum_{i=1}^{m} \lieprod{g_i}{\ad{f}{}{g_1}}u_i } \\
\vdots \\
\innerprod{p}{ \ad{f}{2}{g_m} + \sum_{i=1}^{m} \lieprod{g_i}{\ad{f}{}{g_m}}u_i }
\end{array}\right].
\end{equation}
Proceeding in this way until one finds an expression which explicitly depends on $u$ one gets:
\begin{equation}
\phi^{(k)} =
\left[\begin{array}{c}
\innerprod{p}{ \ad{f}{k}{g_1} + \sum_{i=1}^{m} \lieprod{g_i}{\ad{f}{k-1}{g_1}}u_i } \\
\vdots \\
\innerprod{p}{ \ad{f}{k}{g_m} + \sum_{i=1}^{m} \lieprod{g_i}{\ad{f}{k-1}{g_m}}u_i }
\end{array}\right].
\end{equation}
It is possible to rewrite this expression in the form $\phi^{(k)} = A_k + B_ku$ where:
\begin{equation*}
A_k=\left[\begin{array}{c}
\innerprod{p}{\ad{f}{k}{g_1}} \\
\innerprod{p}{\ad{f}{k}{g_2}} \\
\vdots \\
\innerprod{p}{\ad{f}{k}{g_m}}
\end{array}\right]
\end{equation*}
\begin{equation*}
B_k=\left[\begin{array}{cccc}
\innerprod{p}{\lieprod{g_1}{\ad{f}{k-1}{g_1}}} & \innerprod{p}{\lieprod{g_2}{\ad{f}{k-1}{g_1}}} & \cdots & \innerprod{p}{\lieprod{g_m}{\ad{f}{k-1}{g_1}}} \\
\innerprod{p}{\lieprod{g_1}{\ad{f}{k-1}{g_2}}} & \innerprod{p}{\lieprod{g_2}{\ad{f}{k-1}{g_2}}} & \cdots & \innerprod{p}{\lieprod{g_m}{\ad{f}{k-1}{g_2}}} \\
\vdots & \vdots & \ddots & \vdots \\
\innerprod{p}{\lieprod{g_1}{\ad{f}{k-1}{g_m}}} & \innerprod{p}{\lieprod{g_2}{\ad{f}{k-1}{g_m}}} & \cdots & \innerprod{p}{\lieprod{g_m}{\ad{f}{k-1}{g_m}}}
\end{array}\right].
\end{equation*}
\section{Main result}
In this section one considers nonlinear affine control systems with single input, $u\in\R{}$, and it is proved that the intrinsic order of optimal control problems of these systems is always a positive integer.
\begin{teo} \label{p:q inteiro}
Given an optimal control problem in the form (P) with single input, i.e., $m=1$, then its intrinsic order is a positive integer.
\end{teo}
To simplify the proof of this theorem, some useful results are proved first.
\begin{prop} \label{l:passa j}
If there is $k\in \N{}$ such that $\lieprod{g}{\ad{f}{i}{g}}=0$ for all $x$, where $0 \leq i \leq k-1$ then for all $j \in \N{}$, $1 \leq j \leq k$, one has:
\begin{equation*}
\lieprod{\ad{f}{j}{g}}{\ad{f}{l}{g}}=-\lieprod{\ad{f}{j-1}{g}}{\ad{f}{l+1}{g}}
\end{equation*}
for all $x$, with $0\leq l \leq k-(j+1)$.
\end{prop}
\begin{proof}
The proof uses induction on $j$. One knows that if $0 \leq l \leq k-1$ then $\lieprod{g}{\ad{f}{l}{g}}=0$, therefore $\lieprod{f}{\lieprod{g}{\ad{f}{l}{g}}}=0$. From the Jacobi Identity follows:
\begin{equation*}
\lieprod{\ad{f}{1}{g}}{\ad{f}{l}{g}}=-\lieprod{g}{\ad{f}{l+1}{g}}
\end{equation*}
thus the proposition is true when $j=1$.
Now, suppose that there is a $j_0<k$ such that the thesis is valid for all $j \leq j_0$, i.e.,
\begin{equation*}
\lieprod{\ad{f}{j}{g}}{\ad{f}{l}{g}}=-\lieprod{\ad{f}{j-1}{g}}{\ad{f}{l+1}{g}}
\end{equation*}
for all $x$, with $0\leq l \leq k-(j+1)$.
Then, for all $j<j_0$, if $0\leq l \leq k-(j+1)$, one shows by induction in $m$ that:
\begin{equation} \label{e:inducao em m}
\lieprod{\ad{f}{m}{g}}{\ad{f}{l+j-m}{g}}=0
\end{equation}
for all $x$, with $0 \leq m \leq j$. Indeed, if $m=0$, we have $\lieprod{g}{\ad{f}{l+j}{g}}$, which is null by the proposition hypothesis since $l+j\leq k-1$. If there is a $m_0<j$ such that the relation is valid for all $m \leq m_0$ then, since we have the following inequalities:
\begin{align*}
m+1 &\leq j \\
l+j-(m+1) & \leq k-(m+1),
\end{align*}
then, by the induction hypothesis on $m$, one has:
\begin{equation*}
0=\lieprod{\ad{f}{m}{g}}{\ad{f}{l+j-m}{g}},
\end{equation*}
but, by the induction hypothesis on $j$, one gets:
\begin{equation*}
\lieprod{\ad{f}{m+1}{g}}{\ad{f}{l+j-(m+1)}{g}}=-\lieprod{\ad{f}{m}{g}}{\ad{f}{l+j-m}{g}}
\end{equation*}
which is precisely the relation for $m+1$, ending the induction on $m$:
So, taking $m=j$ in equation \eqref{e:inducao em m}, it follows that $\lieprod{\ad{f}{j}{g}}{\ad{f}{l}{g}}=0$, and this implies that $\lieprod{f}{\lieprod{\ad{f}{j}{g}}{\ad{f}{l}{g}}}=\lieprod{f}{0}=0$. Again, by the Jacobi Identity,
\begin{equation*}
\lieprod{\ad{f}{j+1}{g}}{\ad{f}{l}{g}}=-\lieprod{\ad{f}{(j+1)-1}{g}}{\ad{f}{l+1}{g}},
\end{equation*}
ending the induction on $j$.
\end{proof}
\begin{cor}
With the same hypothesis of Proposition \ref{l:passa j}, if $0 \leq j \leq k$, then
\begin{equation*}
\lieprod{g}{\ad{f}{k}{g}}=(-1)^j\lieprod{\ad{f}{j}{g}}{\ad{f}{k-j}{g}}.
\end{equation*}
\end{cor}
\begin{proof}
Clearly, the assertion is true for $j=0$. Suppose that it's also true for all $j<k$. One has already these inequalities,
\begin{gather*}
1 \leq j+1 \leq k \\
k-j-1 = k-(j+1) \leq k-(j+1)
\end{gather*}
and use the Proposition \ref{l:passa j}. Therefore
\begin{equation*}
\lieprod{g}{\ad{f}{k}{g}}=(-1)^j\lieprod{\ad{f}{j}{g}}{\ad{f}{k-j}{g}}=(-1)^{j+1}\lieprod{\ad{f}{j+1}{g}}{\ad{f}{k-(j+1)}{g}}.
\end{equation*}
\end{proof}
\begin{cor} \label{c:k_par_[g,ad_f^kg]=0}
With the same hypothesis of Proposition \ref{l:passa j}, if $k$ is even then $\lieprod{g}{\ad{f}{k}{g}}=0$.
\end{cor}
\begin{proof}
It's straightforward since $\lieprod{g}{\ad{f}{k}{g}}=(-1)^{\frac{k}{2}}\lieprod{\ad{f}{\frac{k}{2}}{g}}{\ad{f}{\frac{k}{2}}{g}}$.
\end{proof}
Now it is very simple to prove the Theorem \ref{p:q inteiro}.
\begin{proof}[Proof of Theorem \ref{p:q inteiro}]
Suppose that all the derivatives up to $k$ of $\innerprod{p}{g}$ does not depend explicitly of $u$, i.e., the control appears with a identically zero coefficient. So, from Lemma \ref{l:derivada_de_<p,h>}, it is known that $\innerprod{p}{\lieprod{g}{\ad{f}{i}{g}}}=0$, for all $i=0,\ldots,k-1$.
Since $p$ is arbitrary, one concludes also that $\lieprod{g}{\ad{f}{i}{g}}=0$, for all $i=0,\ldots,k-1$.
Note that $k$ is not necessarily the problem order, but $k$ is less than or equal it.
If $k$ is even, one can use the Corollary \ref{c:k_par_[g,ad_f^kg]=0} to conclude that $\lieprod{g}{\ad{f}{k}{g}}=0$, and from that $\innerprod{p}{\lieprod{g}{\ad{f}{k}{g}}}=0$.
Using again the Lemma \ref{l:derivada_de_<p,h>}, it is easy to see that $\innerprod{p}{\lieprod{g}{\ad{f}{k}{g}}}$ is precisely the coefficient of $u$ in the derivative of order $k+1$ of $\innerprod{p}{g}$. Therefore the first derivative of $\innerprod{p}{g}$, that explicitly depends of $u$ can not be odd. Therefore the problem order is always even.
\end{proof}
The problem order, unlike the local order, can be easily evaluated. Thereby, it allows one to develop tools that can be directly applied to solve practical problems, like the design of optimal controls on a singular interval, and to answer theoretical questions, like the qualitative behavior of optimal controls at junctions of singular and nonsingular intervals. This paper has given a proof that the problem order is certainly a positive integer only if the system has a single input.
|
\section{Introduction}
It is now well established that there are many different ways of achieving, within circuit quantum electrodynamics (cQED), the essential primitive operations for quantum information processing. Beyond protocols for achieving highly accurate single-\cite{bench1} and two-qubit\cite{bench2} gate operations, the achievement of fast, flexible, accurate quantum measurements\cite{ssr} is also essential. The current experimental emphasis on reliable feedback of measurement data to control subsequent quantum operations\cite{goodM1, goodM2, goodwithF} fulfils the theoretical hope that such capabilities will find important uses in reliable quantum information processing\cite{LT12}. In particular, the achievement of successful fault tolerant quantum computation relies on the implementation of adaptive gate sequences conditioned on the results of specific kinds of quantum measurements\cite{BT}, namely multi-qubit parity measurements\cite{LT10}. In a parity measurement, the measurement outcome is to be one single bit, regardless of the number of qubits involved. The bit should simply indicate whether the number of ones in the set of measured qubits is even or odd. It is essential for the proper functioning of this measurement that no other information about the qubits be uncovered by the measurement. Furthermore, it is necessary, for applications in quantum error correction, that the measurement be ``quantum non-demolition'' (QND); if the state of the qubits before measurement is an eigenstate of the measurement (i.e., is purely even or purely odd), then the final quantum state of the qubits should be unchanged.\\
In the preferred (topological) error correction code schemes, the parity of four\cite{BT} or three\cite{3par} qubits is needed. It has generally been assumed that this parity measurement would be accomplished with a quantum circuit involving one- and two-qubit gates; the parity is thus computed by elementary logic operations, with the result stored in another ancillary qubit. The parity measurement is then completed by a conventional single-qubit measurement on the ancilla. This measurement does not even have to have QND character on the ancilla -- the net result is a QND measurement on the data qubits. Nevertheless, there are reasons for wishing to replace this quantum circuit with a more direct procedure. First, a direct parity measurement dispenses with the need for extra ancilla qubits. Furthermore, the problem of the accumulation of error is ameliorated. That is, in the circuit approach, the net error will be the sum of the errors of each of the quantum gates and of the single-qubit measurement. This is no fewer than four individual operations; it is known that for achievement of fault tolerance, each of these individual operations needs to have an error rate no larger than about 1\%\cite{goodest}. This means that the ``all in one'' operation studied here is permitted to have a larger error rate, around 4\%, say.\\
Several detailed concepts for direct two-qubit parity measurements have been analyzed recently\cite{mabuchiway,coherent2,LT10,LT12,NG}, with a number of them being promising for applications. The central idea of these approaches is that qubits are off-resonantly (i.e. dispersively) coupled to a cavity mode; the frequency of the mode is shifted by an amount dependent on the state of the qubits, and this shift is then read out by measuring the phase of a microwave tone either transmitted through or reflected from the cavity. \\
For the most part, these earlier proposals have no straightforward generalisation to measurement of the parity of more than two qubits. Refs. \cite{mabuchiway,firat} indicated that a generalisation of previous schemes that would make multiqubit parity measurement possible involves {\em multiple resonant modes}. In particular \cite{firat} showed that, by having just two cavity modes, each of which is subject to qubit-state-dependent dispersive shifts, three qubit parity measurements become possible. The recent proposal of Nigg and Girvin \cite{NG} is clearly extendible to multi-qubit parity measurement; by loading a cavity with a coherent state in a precisely timed way, the state's phase can accumulate information about a particular subset of qubits (with others removed by refocussing), such that the subsequent dispersive measurement of another ancilla qubit can give a readout of any subset parity.\\
In this paper, we take up a detailed analysis of the multi-qubit parity measurement proposal of \cite{firat} using the stochastic master equations used to represent realistic homodyne measurements, as in Refs \cite{LT10,LT12,coherent2}. \cite{firat} developed the two-mode concept using a completely different approach, which involved obtaining scattering parameters from a classical linear electrical circuit analysis, combined with an unrealistic model of measurement in which a hypothetical von-Neumann measurement is performed instantaneously after a coherent-state probing pulse has completed its unitary interaction with the system. As a part of the present work, we provide a new derivation based on input-output theory\cite{MW} of the quantum optics of a cavity, coupled to a continuum, with two closely spaced resonant modes. \\
Many of the broad features of \cite{firat} are confirmed in the present realistic study: there exists a choice of parameters, in which all the relevant parameters of the problem (the dispersive coupling of qubit to cavity, the detuning of the two cavity modes from each other, and the coupling strength of the two modes to the continuum) are of comparable strength, for which a successful three-qubit parity measurement is obtained. In fact, our present analysis provides new, simple formulas for the ideal setting of all these parameters. For these settings, the steady-state statistics of the homodyne measurement are identical for any state in one of the parity subspaces (even or odd). The transient response, however, does distinguish individual states, and thus degrades the fidelity of the parity measurement. While \cite{firat} indicated that a good strategy for dealing with these transient effects is to use a low-intensity, weak measurement of long duration, the details of the present optimisation of the measurement in light of the transient effects were not anticipated by the earlier analysis. Furthermore, in the present study we consider a realistic measurement setting in which qubit decay, determined fundamentally in the cavity setting by the Purcell effect\cite{Pur}, constrains the improvement that can be obtained by prolonging the measurement. Our optimisations indicate that despite the current limitations of superconducting qubit-cavity systems, parity measurements of impressive fidelity (c. 95\%) will be possible, but only if we permit a moderate degree (c. 50\%) of post-selection to retain only those cases where the homodyne measurement is most conclusive. Better performance with the presently-analyzed scheme is not precluded, but would appear to require qubits with longer $T_1$ and $T_2^*$ times.\\
The paper is organised as follows. In \Sec{sec:system}, we present the model for the circuit QED system, containing two resonant modes and three qubits. By tracing out the mode degrees of freedom we derive an effective stochastic master equation for the qubit dynamics. In \Sec{sec:paritymeas}, we discuss the desired properties of a parity measurement and derive the optimal values of circuit parameters to obtain these. We define the measures of measurement fidelity and study the effect of measurement inefficiency in \Sec{sec:measfidelity} and \Sec{sec:eta} respectively. \Sec{sec:transients} is devoted to the study of transient effects and a strategy for mitigating the undesired measurement back-action is presented. In \Sec{sec:results} we give the main results and conclude in \Sec{sec:conclusions}.
\section{The system}\label{sec:system}
The system whose parity we want to measure consists of three (artificial) atoms coupled to two fundamental modes of a cavity (or two different cavities), which couple to a common input and output continuum, as depicted in \Fig{fig:system}.
\begin{figure}[htp]
\centering
\includegraphics[scale=0.87]{Fig1recolored.pdf}
\caption{A possible physical realization of the three-qubit parity measurement analyzed in this paper. This concept uses elements from traditional optics and cavity QED; Ref. \cite{firat} illustrates several possible realizations of the measurement using microwave techniques, i.e., using circuit QED. Three atoms or artificial atoms are held in space, either by trapping techniques or by being embedded in a crystal, in the middle of a crossed-mode double cavity. The two relevant horizontal and vertical modes are to be slightly detuned from one another, and are far detuned from the atomic transitions, so that the cavity-atom interaction is dispersive. The two modes are driven simultaneously with pulsed coherent radiation whose frequency is in between that of the two cavity modes. Parity information is extracted by a homodyne measurement of the reflected field. The ``fiber coupler'' accomplishing the splitting and combining can be a standard three-port component such as a symmetric Y-branch coupler. A modification of this setup is straightforwardly possible in which the output field emerges in transmission rather than reflection.}
\label{fig:system}
\end{figure}
For simplicity, we neglect the possible influence of higher qubit levels and approximate each atom as a two-level system. The system is operated in the dispersive regime, where the transition frequency $\Omega_i$ of qubit $i$ is far detuned from the resonance frequency of either resonator mode $\omega_j$ such that $g_i^j/|\Omega_i - \omega_j| \ll 1$ where $g_i^j$ is the coupling strength between qubit $i$ and mode $j$. In this regime the Hamiltonian, in the rotating frame defined by the measurement-tone frequency $\omega_m$, is given by \cite{blaisPRA2004}
\begin{equation}\label{eq:H}
H = \Big(\Delta^b+\sum_{j=1}^{3}\chi^b_j \sigma_z^{(j)}\Big)b^{\dagger}b+\Big(\Delta^a+\sum_{j=1}^{3}\chi^a_j \sigma_z^{(j)}\Big)a^{\dagger}a+\sum_{j=1}^{3}\frac{\Omega_j+\chi^a_j+\chi^b_j}{2}\sigma_z^{(j)} + \Big[\epsilon(t)(\sqrt{\kappa^a}a^{\dagger}+\sqrt{\kappa^b}b^\dagger)+ \mathrm{h.c.}\Big].
\end{equation}
where $ \Delta^k=\omega^k-\omega_m $ (with $ k=a,b $) are the cavity detunings, and $ \chi_j^k=(g_j^k)^2/\Delta^k_{j} $ ($j=1,2,3$ and $k=a,b$) are the dispersive coupling strengths with $\Delta_j^{k}=\omega^k-\Omega_j $. The amplitude of the measurement signal is given by $\epsilon(t)$. The resonator modes are described by the annihilation (creation) operators $ a(a^{\dagger}) $ and $ b(b^{\dagger}) $. The coupling between resonator mode $i$ and the environment is given by $\kappa_i$. In the absence of measurement, the master equation describing the system evolution is given by
\begin{equation}\label{eq:master}
\dot{\rho} = -i[H,\rho]+\sum_{j=1}^{3}\Big(\gamma_{1j}\textit{D}[\sigma_{-}^{(j)}]\rho+ \frac{\gamma_{\varphi j}}{2}\textit{D}[\sigma_{z}^{(j)}]\rho\Big)+\textit{D}\Big[\sqrt{\kappa^a}a+\sqrt{\kappa^b}b\Big]\rho+
\kappa^a\sum_{j=1}^{3}\textit{D}\Big[\lambda^a_j\sigma_{-}^{(j)}\Big]\rho+
\kappa^b\sum_{j=1}^{3} \textit{D}\Big[\lambda^b_j\sigma_{-}^{(j)}\Big]\rho,
\end{equation}
where $\textit{D}\left[c\right]\rho = c\rho c^\dagger + 1/2(c^\dagger c \rho + \rho c^\dagger c )$ is a dissipation superoperator of Lindblad form \cite{lindbladCMP1976} and $\gamma_{1j}$ and $\gamma_{\varphi j}$ are the relaxation and dephasing rates of qubit $j$ respectively. The last two terms describe the Purcell relaxation \cite{Pur} where $ \lambda_j^k=g_j^k/\Delta^k_{j} $ ($j=1,2,3$ and $k=a,b$) and we have assumed distinct qubit frequencies such that $|\Omega_i - \Omega_j| \gg \kappa^a \lambda^a_j \lambda^a_i$ which allows us to neglect all cross-terms between operators belonging to different qubits. This assumption is also important if we want to suppress the direct coupling between qubits mediated by virtual photons \cite{blaisPRA2004}.
In such a two-mode setting, it would be common to also have terms in the Hamiltonian involving mode-mode coupling, i.e., terms containing $a^\dagger b$. While such terms are indeed typically nonzero, it has been shown that, using the flexibility offered within circuit QED, circuit structures can readily be devised where these interactions are tuned away\cite{tunecQED}. While such terms would not fundamentally alter the parity-measurement scheme that we analyze here, we find that the study of the effects are more transparent with the minimal Hamiltonian Eq. (\ref{eq:H}), which we will henceforth employ throughout this paper.
From the point of view of the cavity degrees of freedom, \Eq{eq:H} and \Eq{eq:master} describe the generation and evolution of coherent states, whose amplitudes are governed by the differential equations \cite{gambettaPRA2006}
\begin{eqnarray}\label{eq:cavity_amplitudes}
\dot{\alpha}_{ijk}&&=-i\sqrt{\kappa^a}\epsilon-i(\Delta^a+\chi^a_{ijk})\alpha_{ijk}-\frac{\kappa^a}{2}\alpha_{ijk}-\frac{\sqrt{\kappa^a\kappa^b}}{2}\beta_{ijk},\nonumber \\
\dot{\beta}_{ijk}&&=-i\sqrt{\kappa^b}\epsilon-i(\Delta^b+\chi^b_{ijk})\beta_{ijk}-\frac{\kappa^b}{2}\beta_{ijk}-\frac{\sqrt{\kappa^a\kappa^b}}{2}\alpha_{ijk}.
\end{eqnarray}
such that the cavity fields are entangled with the qubit states through the coupling $\chi^m_{ijk}=\langle ijk|\sum_{l=1}^3 \chi_l^m \sigma_z^{(l) }|ijk\rangle $, with ($ m=a,b $). In this way, the cavity fields act like pointer states with allows us to indirectly infer the state of the qubit system through a measurement on the field.\\
The unconditional evolution described by \Eq{eq:master} is sufficient if one is interested in calculating average quantities of system operators. When studying the performance of a measurement it is however necessary to calculate properties of the system conditioned on a certain subset of measurement results. For this purpose, we need to describe the system evolution conditioned on a single measurement result. In circuit QED, phase sensitive amplification allows for the equivalent of homodyne detection in optics. The system dynamics including the measurement back action is described by the stochastic master equation \cite{wisemanPRA1993}
\begin{equation}\label{eq:SME}
d\rho= \mathcal{L} \rho dt + \sqrt{\eta}\textit{M}\Big[(\sqrt{\kappa^a}a+\sqrt{\kappa^b}b )e^{-i \phi}\Big]\rho dW(t),
\end{equation}
where $\mathcal{L}$ is the superoperator written in \Eq{eq:master} and $\textit{M}[c]\rho =c\rho+\rho c^{\dagger}-\langle c+c^{\dagger}\rangle \rho$ is the superoperator describing the measurement back-action and $\eta$ is the efficiency of the measurement. The stochastic evolution, fundamentally originating from the collapse of the state, is realized through the Wiener process $dW(t)$ with the defining statistical properties $\mathrm{E}[dW(t)] = 0$ and $\mathrm{E}[dW(t)^2] = dt$ \cite{klebaner}. The measurement signal is given by the homodyne current
\begin{equation}\label{eq:current}
j(t)dt = \sqrt{\eta}\big\langle ( \sqrt{\kappa^a}a+\sqrt{\kappa^b}b )e^{-i \phi} + (\sqrt{\kappa^a}\dagg{a}+\sqrt{\kappa^b}\dagg{b} )e^{i \phi} \big\rangle dt + dW(t).
\end{equation}
\Eq{eq:SME} and \Eq{eq:current} can in principle be used to numerically study the evolution of the system and the performance of the measurement. However, to gain qualitative understanding with the long term goal of achieving a high fidelity measurement, it is necessary to reduce these equations such that they describe the evolution of the qubits' degree of freedom only.
\subsection{Effective stochastic master equation of three-qubit/two-mode circuit QED}
The elimination of the cavity degrees of freedom to obtain an effective SME for the qubits has been treated in detail in \cite{gambettaPRA2008, coherent2}. There the analysis was done for one and two qubits coupled to a single cavity mode. Here we extend this derivation to the case of more cavity modes and qubits. The elimination of cavity degrees of freedom is carried out by moving to a frame of reference which follows the average cavity field, whose state is conditioned on the state of the qubits:
\begin{equation}\label{eq:polaron}
P=\sum_{i,j,k=0,1}D_a(\alpha_{ijk})D_b(\beta_{ijk})\Pi_{ijk},
\end{equation}
where $\alpha$ and $ \beta$ are the coherent amplitudes of cavity modes $ a $ and $ b $, respectively and $D_c(\gamma) = \exp(\gamma c^\dagger - \gamma^*c)$ is the displacement operator with respect to each cavity field \cite{gerryKnight}. $\Pi_{ijk}=|ijk\rangle\langle ijk| $ are projection operators onto the basis states of the three-qubit Hilbert space. The field dynamics in this frame of reference is given by the vacuum fluctuations only and in the limit $\gamma_{1j}\ll \kappa_i$ we can trace out the photonic states, yielding the effective master equation
\begin{eqnarray}\label{eq:SMEeff}
d \rho&&=-i\Big[\sum_{j=1}^{3}\frac{\omega^{atom}_j+\chi^a_j+\chi^b_j}{2}\sigma_z^{(j)},\rho \Big]dt+\Big(\sum_{j=1}^{3}\gamma_{1j}\textit{D}[\sigma_{-}^{(j)}]+ \frac{\gamma_{\varphi j}}{2}\textit{D}[\sigma_{z}^{(j)}]+
\kappa^a\textit{D}\Big[\lambda^a_j\sigma_{-}^{(j)}\Big]+\kappa^b \textit{D}\Big[\lambda^b_j\sigma_{-}^{(j)}\Big]\Big)\rho dt\nonumber\\
&&+\sum_{ijk,lmn}\Big(\chi^a_{lmn,ijk}[\mathrm{Im}(\alpha^{*}_{ijk}\alpha_{lmn})+i\mathrm{Re}(\alpha^{*}_{ijk}\alpha_{lmn})]\Big)\Pi_{ijk}\rho \Pi_{lmn} dt \nonumber\\
&&+\sum_{ijk,lmn}\Big(\chi^b_{lmn,ijk}[\mathrm{Im}(\beta^{*}_{ijk}\beta_{lmn})+i\mathrm{Re}(\beta^{*}_{ijk}\beta_{lmn})]\Big)\Pi_{ijk}\rho \Pi_{lmn} dt \\
&&+i\frac{\sqrt{\kappa^a\kappa^b}}{2}\sum_{i\neq l}\sum_{j\neq m}\sum_{k\neq n}\Big[\mathrm{Im}(\alpha_{ijk}\beta^{*}_{lmn})+\mathrm{Im}(\beta_{ijk}\alpha^{*}_{lmn})\Big]\Pi_{ijk}\rho \Pi_{lmn} dt \nonumber\\
&&+\frac{\sqrt{\kappa^a\kappa^b}}{2}\sum_{i\neq l}\sum_{j\neq m}\sum_{k\neq n }\Big[\mathrm{Re}(\beta_{ijk}\alpha^{*}_{lmn})+\mathrm{Re}(\alpha_{ijk}\beta^{*}_{lmn})-\mathrm{Re}(\alpha^{*}_{ijk}\beta_{ijk})-\mathrm{Re}(\alpha^{*}_{lmn}\beta_{lmn})\Big]\Pi_{ijk}\rho \Pi_{lmn} dt \nonumber\\
&&+\sqrt{\eta}\textit{M}\Big[ \Pi_{\Sigma}e^{-i\phi}\Big]\rho dW(t),\nonumber
\end{eqnarray}
where $ \chi^m_{ijk,lmn}=\chi_{ijk}^m - \chi_{lmn}^m $ (with $ m=a,b $). Here, we introduce the short-hand measurement operator $\Pi_{\Sigma}=\sum_{i,j,k}\Sigma_{ijk} \Pi_{ijk}$, where $\Sigma_{ijk} = \sqrt{\kappa^a}\alpha_{ijk} + \sqrt{\kappa^b}\beta_{ijk}$ is the linear combination of the cavity fields visible through the connection port.
In addition to the system dynamics, the homodyne current is also be expressed in terms of the qubits' degrees of freedom
\begin{equation}\label{eq:currentq}
j(t)dt = \sqrt{\eta}\big\langle \Pi_\Sigma e^{-i \phi} + \dagg{\Pi_\Sigma}e^{i \phi} \big\rangle dt + dW(t).
\end{equation}
\section{Parity measurement}\label{sec:paritymeas}
The goal of an experimental setup as in Fig. \ref{fig:system} is to realize a parity measurement of the joint state of the three qubits, that is, we would like the measurement to distinguish between states belonging to the mutually orthogonal sub-spaces
\begin{eqnarray}
\mathcal{H_+} &=& \mathrm{span}\left(\ket{000}, \ket{011}, \ket{101}, \ket{110}\right), \nonumber \\
\mathcal{H_-} &=& \mathrm{span}\left(\ket{001}, \ket{010}, \ket{100}, \ket{111}\right),
\end{eqnarray}
without distinguishing between different states within $\mathcal{H_+}$ and $\mathcal{H_-}$. In addition to this, the measurement should not cause any back-action on the measured state apart from the necessary state collapse associated with the gain of information. To realize these properties, the dynamics of the pointer states, together with the chosen measurement basis, must reflect these constraints. In this section we therefore study the evolution given in \Eq{eq:cavity_amplitudes} to obtain a choice of system parameter values such that these conditions are fulfilled.
We start by analyzing the steady state solution to \Eq{eq:cavity_amplitudes}, given by
\begin{equation}\label{eq:sigma_ss}
\Sigma_{ijk}^{ss} = -2 \epsilon_{ss} \frac{\Delta^a\kappa^b + \Delta^b\kappa^a + (\kappa^a + \kappa^b) \chi_{ijk}}{\Delta^b(\kappa^a + \kappa^b) + \Delta^a(\kappa^b + 2i(\Delta^b + \chi_{ijk})) + \chi_{ijk}(\kappa^a + \kappa^b + 2i\chi_{ijk})} \equiv C_{ijk} \epsilon_{ss},
\end{equation}
where $\epsilon_{ss}$ is the steady-state amplitude of the drive and $C_{ijk}$ is a constant which only depends on circuit parameters. Here, the subscripts $ijk$ again refer to the qubit eigenstate $\ket{ijk}$ and we will from now on assume that $\chi_{ijk} = \chi$ is the same for all $i,j,k$, which can be achieved by proper choices of the coupling energies $g_i^j$. In the following we assume the LO phase to be $\phi = \pi/2$ corresponding to a measurement of the imaginary part of the output field $\text{Im}(\Sigma_{ijk})$. In order to reduce the complexity of the analysis we limit the number of free parameters by assuming that $\Delta^a = -\Delta^b$ and $\kappa^a = \kappa^b$. In \Fig{fig:kernel_a} we plot $\text{Im}(\Sigma)/\epsilon$ as a function of the remaining free parameters $\kappa^a$ and $\Delta^a$. Each surface corresponds to one of the four distinct values of $\chi_{ijk} = \{-3 , -1, 1, 3\}\chi$ which is set by the three-qubit basis states.
\begin{figure}[ht]
\centering
\subfigure[ The solution to \Eq{eq:sigma_ss} as a function of $\kappa^a = \kappa^b$ and $\Delta^a = -\Delta^b$. The blue (red) surfaces show the solution for the negative (positive) parity subspace. The intersection between the planes is indicated below the solution (see text for details). The (optimal) black point shows $\kappa^a = \kappa^b = 2\chi$ and $\Delta^a = -\Delta^b = \sqrt{3}\chi$.
]{
\includegraphics[scale=0.4]{kernel.pdf}
\label{fig:kernel_a}
}
\subfigure[ The pointer states of the detected field $\Sigma_{ijk}(t)$ for the eigenstates $\ket{000}$ (solid red), $\ket{011}$ (dashed red), $\ket{111}$ (solid blue) and $\ket{001}$ (dashed blue) of the three qubit system. The steady state of the system allows for a perfect parity measurement, while the different transient trajectories result in an undesirable distinguishability within each subspace. The parameters are $\epsilon_0 = \sqrt{\chi}$, $\kappa^a = \kappa^b = 2\chi$, $\Delta^a = -\Delta^b = \sqrt{3}\chi$ and $\sigma = 10\chi$. ]{
\includegraphics[scale=0.4]{pointerstates.pdf}
\label{fig:kernel_b}
}
\caption[Optional caption for list of figures]{The pointer states of measurement. a) shows the steady state solution $\Sigma_{ijk}^{ss}$ of the pointer states for the different parity subspaces and b) shows the corresponding transient time evolution. }
\label{fig:kernel}
\end{figure}
The blue (red) surfaces show the negative (positive) parity solution corresponding to $\chi_{ijk} = \{-1, 3\}(\{1, -3\})\chi$. The intersection between the planes, shown by the blue and red lines on the base of the figure, gives the set of parameter values for which $\text{Im}(\Sigma_{000}) = \text{Im}(\Sigma_{011})=\text{Im}(\Sigma_{101}) = \text{Im}(\Sigma_{110}) = \Sigma_+$ and $\text{Im}(\Sigma_{111}) = \text{Im}(\Sigma_{001})=\text{Im}(\Sigma_{100}) = \text{Im}(\Sigma_{010}) = \Sigma_-$ such that the states within each subspace cannot be distinguished. As a crucial property, the indicated set of solutions has a symmetry relating the positive and negative parity subspaces. At the same time, $\Sigma_+ \neq \Sigma_-$, which allows the measurement to distinguish the two subspaces.
In addition to the conditions imposed by the measurement which are satisfied by the solutions in the blue branch, the solutions in the red branch satisfy the condition that the real parts of the fields also are the same. As discussed in \Sec{sec:eta}, any difference between the pointer states not recorded by the measurement will result in measurement-induced dephasing, so that it is crucial to equate these real parts if we want no additional back-action generated by the measurement. We therefore expect, and numerically find, that deviating from the red, dashed branch decreases the fidelity of the measurement. The indicated point in \Fig{fig:kernel_a} shows the specific choice of parameters used in all numerical simulations which we return to in \Sec{sec:results}. We have numerically verified that changing the values along the red line has only a negligible effect on the fidelity. \\
The above study has allowed us to extract the right parameter values to achieve the desired properties of the pointer states in the steady state. Equipped with this knowledge we now return to the full solution to \Eq{eq:cavity_amplitudes}. In \Fig{fig:kernel_b}, we plot this as trajectories in the IQ-plane with a pulse shape given by
\begin{equation}\label{eq:atanpulse}
\epsilon(t) = \frac{\epsilon_{ss}}{\pi} \left( \arctan\left( \sigma(t - t_{on}) \right) + \frac{\pi}{2} \right).
\end{equation}
As expected, the steady state of the solution satisfies the condition that $\text{Im}(\Sigma_{ijk})$ is the same for states of the same parity whereas $\Sigma_+ \neq \Sigma_-$. Unfortunately this condition is only valid for the steady state while the transient path taken from the vacuum to the steady state is such that $\Sigma_{000}$
and the fields in the set $\{ \Sigma_{011}, \Sigma_{101}, \Sigma_{110}\}$ can be distinguished by the measurement (as with the pointer states corresponding to states in $\mathcal{H_-}$). This will cause a partial measurement within $\mathcal{H}_\pm$ during this period and therefore decrease the fidelity of the post-measurement state. This is the main source of infidelity of the proposed measurement scheme and the remaining part of the paper is devoted to the study of how to best mitigate this.
\subsection{Measurement rates}
As discussed above, whenever the measurement is able to distinguish different pointer states from each other, it will give rise to back-action on the qubits. We can quantify the strength of this back-action by considering all the measurement rates and how they correspond to the magnitude of the difference between pointer states. For the specific choice of LO phase made above ($\phi=\pi/2$), the measurement superoperator in \Eq{eq:SMEeff} can be separated into four parts
\begin{eqnarray}\label{eq:explicit_measoperator}
\mathcal{M}[-i\Pi_\Sigma] \rho
&=&
\frac{\text{Im}(\xi(t))}{2} \mathcal{M}[\Pi_+ - \Pi_-] \rho +
\frac{\text{Im}(\delta(t))}{2} \mathcal{M}[\Pi_{000} - \Pi_{011} - \Pi_{101} - \Pi_{110}] \rho \nonumber \\
&-&
\frac{\text{Im}(\delta(t))}{2} \mathcal{M}[\Pi_{111} - \Pi_{001} - \Pi_{010} - \Pi_{100}] \rho \nonumber \\
&-&
i\frac{\text{Re}(\delta(t))}{2} \left[ \Pi_{000} + \Pi_{111} - \Pi_{001} - \Pi_{010} - \Pi_{011} - \Pi_{100} - \Pi_{101} - \Pi_{110} , \rho \right],
\end{eqnarray}
where we have defined the sum and difference fields $\xi(t) \equiv \Sigma_{000}(t) + \Sigma_{011}(t)$ and $\delta(t) \equiv \Sigma_{000}(t) - \Sigma_{011}(t)$ and used the specific symmetries of the pointer states e.g. $\text{Im}(\Sigma_{000}) = -\text{Im}(\Sigma_{111})$. The operators $\Pi_\pm = \sum_{ijk \in \mathcal{H}_\pm} \ket{ijk}\bra{ijk}$ are projection operators on $\mathcal{H}_\pm$. The first term in \Eq{eq:explicit_measoperator} represents the gain of information about the parity of the state as expected from the measurement. Hence, we define a parity measurement rate
\begin{equation}
\Gamma_m^\mathrm{P}(t) = \eta \text{Im}(\xi(t))^2.
\end{equation}
The next two terms arise from the fact that the pointer states within each parity subspace are not identical, resulting in an undesired measurement within each subspace. This intra-parity subspace measurement rate is given by
\begin{equation}
\Gamma_m^\mathrm{IP}(t) = \eta \text{Im}(\delta(t))^2.\label{IP}
\end{equation}
The last term gives a stochastic rotation of the phase in the coherences between the subspaces $\mathrm{Span}(\left(\ket{000}, \ket{111}\right)$ and $\mathrm{Span}(\left(\ket{001}, \ket{010}, \ket{100}, \ket{011}, \ket{101}, \ket{110}\right)$, an effect that does not affect the purity of the post-measurement state for a single measurement. It will however affect the purity of the average state. This effect could be cancelled by the use of feedback and poses no fundamental limitation on the measurement fidelity. This cannot, however, be said for the unwanted, intra-parity measurements. Since the two effects are both $\propto \delta(t)^2$, it is clear that we need to make $\delta(t)$ as small as possible to get a measurement with high fidelity. We return to this issue in \Sec{sec:transients}.
\subsection{Measurement fidelity -- two-outcome vs. three-outcome measurement}\label{sec:measfidelity}
In this section we introduce two measures used to assess the fidelity of the measurement; the signal to noise ratio, SNR, and the overlap fidelity of the post-measurement state relative to the pre-measurement state. The SNR quantifies the distinguishability between signals conditioned on states with different parity, while the overlap fidelity measures the undesired back-action the measurement has by comparing the real post-measurement state to the ideal one. \\\\
To convert the time-dependent current into a single measurement outcome we use the integrated current
\begin{equation}
s_{j(t)}(\tau) = \int_0^\tau j(t) dt
\end{equation}
as our single (real-valued) measurement result. Here the measurement time is given by $\tau$.
We will consider two possible approaches to further interpreting this real-valued outcome as a discrete-valued measurement. Ideally, the integrated measurement outcome has an unambiguous sign; for some of the measurement parameters considered below, this is in fact the case. Under these circumstances, it is satisfactory to infer a parity directly from the measurement: $s>0$ meaning even parity, and $s<0$ meaning odd parity. However, we find that to improve the intra-sector overlap fidelity, it is important to consider measurement parameters that result in a significant number of outcomes with $s\approx 0$. In this case, it is natural to introduce a finite ``conclusiveness threshold'' $s_{th}$.
That is, in addition to assigning outcome ``even'' if $s>s_{th}$ and ``odd'' if $s<-s_{th}$, we call the measurement ``inconclusive'' if $|s|<s_{th}$. A high value of $s_{th}$ allows the observer to discard measurement results that would otherwise lead to a corrupted post-measurement state due to mixing of states with different parity. We will see that allowing a moderate percentage of ``inconclusive'' assignments permits the even/odd overlap fidelity to be dramatically improved in the successful cases. Depending on the objective of the measurement different choices of this threshold will be appropriate, as we discuss further in the Conclusions. \\\\
For each state $\ket{ijk}$ the current in \Eq{eq:currentq} is given by
\begin{equation}\label{eq:current_by_state}
j_{ijk}(t)dt = 2\sqrt{\eta}\text{Im}(\Sigma_{ijk}(t))dt + dW(t).
\end{equation}
which, by the linearity of quantum mechanics, gives the current from a general state in $\mathcal{H_\pm}$: $\ket{\psi_\pm} = \sum_{ijk \in \mathcal{H}_\pm }\gamma_{ijk}\ket{ijk}$
\begin{eqnarray}\label{eq:current_pm}
j_\pm(t)dt &\approx& 2 \sqrt{\eta} \sum_{ijk\in\mathcal{H}_\pm} |\gamma_{ijk}|^2 \text{Im}(\Sigma_{ijk}(t))dt + dW(t),
\end{eqnarray}
where we have assumed that the coefficients $\gamma_{ijk}$ are unaltered during the measurement, that is, we assume that the transients have negligible effect on the post-measurement state. This assumption can be justified if we consider most of the signal to be generated in the steady state. In the weak-measurement limit discussed in \Sec{sec:transients}, this is a fair assumption. \\
Given a current, $j_\pm(t)$, conditioned on a state in $\mathcal{H}_\pm,$ we define the SNR to be the ratio between the mean and standard deviation of the difference $s_{j_+}(\tau) - s_{j_-}(\tau)$
\begin{equation}
\mathrm{SNR}(\tau) =
\frac{\mathrm{E}[s_{j_+}(\tau)] - \mathrm{E}[s_{j_-}(\tau)] }{\sqrt{\mathrm{Var}[s_{j_+}(\tau)] + \mathrm{Var}[s_{j_-}(\tau)]}}
= \frac{\mathrm{E}[s_{j_+}(\tau)] - \mathrm{E}[s_{j_-}(\tau)] }{\sqrt{2\tau}},
\end{equation}
where we have dropped the time argument in $j_\pm(t)$ for notational transparency and used the statistical properties of the Wiener process in the second equality.
Within the assumptions made above, the SNR is given by
\begin{equation}
\mathrm{SNR}(\tau) = \sqrt{\frac{2\eta}{\tau}} \int_0^\tau\left( \sum_{ijk\in\mathcal{H}_+} |\gamma_{ijk}|^2 \text{Im}(\Sigma_{ijk}(t)) -
\sum_{ijk\in\mathcal{H}_-} |\gamma_{ijk}|^2 \text{Im}(\Sigma_{ijk}(t)) \right)dt
\end{equation}
which can be further approximated if we assume that the fraction of the measurement time spent in the transient region is negligible $\tau \gg 1/\kappa$, that is we make the replacement $\Sigma_{ijk} \to C_{ijk}\epsilon_{ss}$
\begin{equation}\label{eq:SNR_ideal}
\mathrm{SNR} \approx 2 \sqrt{2} \sqrt{\eta} \text{Im}(C_{111}) \epsilon_{ss} \sqrt{\tau},
\end{equation}
where we recall the definition of $C_{ijk}$ below \Eq{eq:sigma_ss}. As expected, the fact that $\mathrm{SNR} \propto \epsilon_{ss} \sqrt{\tau}$ shows that $\mathrm{SNR} \gg 1$ can be achieved for arbitrarily low value of measurement strength $\epsilon_{ss}$. We will further explore this limit in \Sec{sec:transients} when studying the effect of field transients. \\\\
If one is interested in only measuring the parity of the state it is enough to have $\mathrm{SNR} \gg 1$ for the measurement to be considered high fidelity. A good example of such a standard, high-fidelity quantum measurement is photon detection using high-fidelity avalanche diodes. Here however the photon is completely destroyed in the process. In a quantum informational setting, the post-measurement state is often to be further processed in some algorithm or error correction scheme. In this case it is crucial that the post-measurement state conditioned on the outcome of the detection is the expected one. As a measure of this we consider the overlap fidelity
\begin{equation}\label{eq:fidelity_def}
F_\pm = \sqrt{\bra{\psi_\pm} \mathrm{E}_\pm[\rho]\ket{\psi_\pm} },
\end{equation}
where $\ket{\psi_\pm}$ is the expected post-measurement state and $ \mathrm{E}_\pm[\rho ]$ is the ensemble averaged, post-measurement state where the mean is taken over states assigned to either $\mathcal{H}_\pm$ by the measurement.
\subsection{Measurement efficiency}\label{sec:eta}
\begin{figure}[ht]
\centering
\subfigure[ Measurement results with $\eta = 1$. ]{
\includegraphics[scale=0.4]{compare_eta_1.pdf}
\label{fig:histograms_eta_a}
}
\subfigure[ Measurement results with $\eta = 0.5$. ]{
\includegraphics[scale=0.4]{compare_eta_0p5.pdf}
\label{fig:histograms_eta_b}
}
\caption[Optional caption for list of figures]{The effect of measurement efficiency. Histograms of 1000 measurement results corresponding to the pre-measurement state in \Eq{eq:psi_pre}. The parameters are as in \Fig{fig:kernel} with $\gamma_{1j} = \gamma_{\varphi j} = \gamma_p = 0$. The measurement time and drive strength are chosen to be $\tau = 20/(\eta\chi)$, $\epsilon_{ss} = 2\sqrt{\chi/20}$ corresponding to $\mathrm{SNR} = 4\sqrt{2}$ in both cases. The value of $\eta$ is specified in each subfigure. The red curves are normal distributions with mean and variance defined in \Eq{eq:gaussian}. The values of $F_\pm$ are given above each figure.}
\label{fig:histograms_eta}
\end{figure}
The quantum efficiency, $\eta$, quantifies how much of the information, which is flowing out of the system, is actually measured. Given a pure initial state and a quantum limited measurement, that is no additional back action apart from the necessary state collapse, with $\eta = 1$ the projection postulate ensures that the post-measurement state is pure. For $\eta < 1$ this is in general no longer true since the observer must average over the non-observed measurement results to obtain the post-measurement state. This procedure is the origin of measurement induced dephasing and lowers the fidelity of the measurement in the sense of \Eq{eq:fidelity_def}. In \Fig{fig:histograms_eta}, we plot the histograms corresponding to 1000 measurement results for $\eta = 0.5$ and $1$. In both cases the initial state is given by
\begin{equation}\label{eq:psi_pre}
\ket{\psi}_\mathrm{pre} = \frac{1}{\sqrt{8}}\sum_{ijk}\ket{ijk},
\end{equation}
which for a perfect parity measurement would be projected on to the post-measurement states
\begin{eqnarray}\label{eq:psi_post}
\ket{\psi}_+ = \frac{1}{\sqrt{4}}\left( \ket{000} + \ket{011} + \ket{101} + \ket{110} \right), \nonumber \\
\ket{\psi}_- = \frac{1}{\sqrt{4}}\left( \ket{001} + \ket{010} + \ket{100} + \ket{111} \right),
\end{eqnarray}
with equal probability. While confining our attention to this initial state does not explore all aspects of the measurement superoperator, it is optimally sensitive to any loss of intra-sector coherence during the measurement, and it is a state with a structure, with its equal superposition of qubit basis states, resembling that of the important stabiliser states of quantum error correction\cite{NC}.
The measurement drive strength $\epsilon_{ss}$ is kept the same for the two cases in \Fig{fig:histograms_eta}. The SNR is also held at a constant value by increasing the measurement time to compensate for the lower value of $\eta$. Since we are interested in the effect of lowering $\eta$, we ignore the effect of decoherence, that is, $\gamma_{1j} = \gamma_{\varphi j} = \gamma_p = 0$ where
\begin{equation}
\gamma_ p = (g/\Delta)^2(\kappa^a + \kappa^b) \label{exPurcell}
\end{equation}
is the Purcell decay rate.
The red curves in \Fig{fig:histogram} are normal distributions with mean and variance
\begin{equation}\label{eq:gaussian}
\mathrm{E}[s_\pm(\tau)] = \mp 2\sqrt{\eta}\int_0^\tau \frac{1}{4}\text{Im}\left( \Sigma_{000}(t) + 3\Sigma_{011}(t) \right) dt, \qquad
\mathrm{Var}[s_\pm(\tau)] = \sqrt{\tau}.
\end{equation}
\\
From the overlap fidelity, it is clear that the purity of the state is not affected by the decrease in $\eta$. This robustness comes from the fact that the pointer states corresponding to states within $\mathcal{H}_\pm$ are perfectly indistinguishable in the steady state. Hence there are no unrecorded measurement results to average over and the state remains pure. The fact that $F_\pm < 1$ is an effect of the transient evolution of the pointer states which is not affected by the measurement efficiency. Note however that, in the presence of decoherence, the measurement efficiency will have an indirect effect on $F_\pm$ through the longer measurement times needed to keep SNR high.
\section{Effect of field transients}\label{sec:transients}
It is clear that the integrated rate $\Gamma_m^\mathrm{IP}(t)$ of Eq. (\ref{IP}) (total effect on the state) must be minimized to limit the unwanted effect due to the transient behavior of the pointer states.
To realize this we make the observation that $\delta(t)\to 0$ when $\epsilon_{ss} \to 0$. This is also true for $\xi(t)$ and the measurement will therefore be weak in this sense. The measurement can however still be strong in the sense that the SNR defining product $\epsilon_{ss}\sqrt{\tau}$ can be kept constant by increasing the measurement time such that the value of this product is kept constant as $\epsilon_{ss} \to 0$. In the absence of decoherence mechanisms, we can keep the measurement on for as long as we want, and in this way realize a strong measurement while mitigating the effect of the unwanted back action. In \Fig{fig:histogram}, we plot the histograms of 1000 measurement results for two different values of $\epsilon_{ss}$ and $\sqrt{\tau}$ such that $\epsilon_{ss} \sqrt{\tau} = 2$ but varying $\tau$.
\begin{figure}[ht]
\centering
\subfigure[ Measurement results for short measurement time, $\tau = 10/\chi$ and $\epsilon_{ss} = 2/\sqrt{\tau}$.]{
\includegraphics[scale=0.4]{compare_eps_10.pdf}
\label{fig:histograms_a}
}
\subfigure[ Measurement results for long measurement time, $\tau = 100/\chi$ and $\epsilon_{ss} = 2/\sqrt{\tau}$.]{
\includegraphics[scale=0.4]{compare_eps_100.pdf}
\label{fig:histograms_b}
}
\caption[Optional caption for list of figures]{The effect of transients. Histograms of 1000 measurement results corresponding to the pre-measurement state in \Eq{eq:psi_pre} and $\mathrm{SNR} = 4\sqrt{2}$. The parameters are as in \Fig{fig:kernel} with $\gamma_{1j} = \gamma_{\varphi j} = \gamma_p = 0$. The values if $\epsilon_{ss}$ and $\tau$ are specified in each subfigure. The red curves are normal distributions with mean and variance defined in \Eq{eq:gaussian}. The values of $F_\pm$ are given above each figure.}
\label{fig:histogram}
\end{figure}
The initial state is given in \Eq{eq:psi_pre} and, since we want to single out the detrimental effect of the field transients, we ignore the effect of decoherence, that is, $\gamma_{1j} = \gamma_{\varphi j} = \gamma_p = 0$. The histograms in \Fig{fig:histogram} show comparable overlap for the different measurement strengths, but the overlap fidelity is near unity for the long measurement time $\tau = 100/\chi$ while being significantly lower for the shorter measurement time $\tau = 10/\chi$. This clearly shows that, in the limit of long measurement time, that is, weak measurement pulse, the setup considered here makes for a perfect parity measurement. Although the SNR defining product $\epsilon_{ss}\sqrt{\tau}$ is chosen to be identical for both measurement times, we note that this only estimates the SNR accurately if all of the measurement takes place in the steady state. For $\tau = 10/\chi$, we are approaching the limit where the transient behavior makes up a non-negligible part of the measurement duration, leading to lower SNR. \\%
Another way to possibly enhance the measurement fidelity is to have the measurement pulse $\epsilon(t)$ turned on slowly compared to all the other time scales in the system. This allows the pointer states to approximately follow the paths given by their instantaneous steady state value. We consider the pulse shape given \Eq{eq:atanpulse} and plot $|\text{Im}(\delta(t))|$ in \Fig{fig:field_differences_a} for a large range of rise times $1/\sigma$.
\begin{figure}[ht]
\centering
\subfigure[ The drive pulse $\epsilon(t)$ for different values of rise time $\sigma$ and the corresponding difference $\text{Im}(\delta(t))$. ]{
\includegraphics[scale=0.4]{adiabatic.pdf}
\label{fig:field_differences_a}
}
\subfigure[ The integrated difference for a large range of rise times $\sigma$. The effect is on the order of a few percent. ]{
\includegraphics[scale=0.4]{integrated_delta.pdf}
\label{fig:field_differences_b}
}
\label{fig:field_differences}
\caption[Optional caption for list of figures]{ The difference field $\delta(t)$ and $\int \delta(t) dt$ for a pulse given in \Eq{eq:atanpulse}. }
\end{figure}
As expected, the sharper the onset of $\epsilon(t)$ is, the bigger the difference $\delta(t)$ becomes between the pointer states in the same parity subspace. Making the pulse smoother in time decreases this difference but simultaneously spreads it over a larger time. To quantify the effect of this trade-off, we plot the integrated value of $\text{Im}(\delta(t))$ in \Fig{fig:field_differences_b}. We see that the difference is negligible over a large range of $\sigma$. We can therefore safely say that the measurement time will not be limited by any adiabaticity constraints and, in the remainder of this work, we focus on the effect of long measurement time as described in the beginning of this section.
\section{Results for optimal measurement}\label{sec:results}
So far all the results have been derived without considering the detrimental effect of decoherence, which is inevitablly present due to coupling of the cavity modes to the continuum. From the analysis so far it is however evident that, for the measurement to be high fidelity, we need a long measurement time such that $\epsilon_{ss} \sqrt{\tau} \gg 1$ and $\epsilon_{ss} \to 0$. But increasing $\tau$ indefinitely is not possible in the presence of qubit decay mechanisms. Thus, there exists an optimal measurement time $\tau_\mathrm{opt}$, which we will identify below, for which the two competing effects of transients and qubit relaxation balance one another. In this optimisation for $\tau_\mathrm{opt}$ we will fix the SNR given in \Eq{eq:SNR_ideal} at a desired value and calculate the measurement record and post-measurement state for different measurement times. Note that the actual calculated SNR will differ from the estimate that we used to fix the relationship between time and drive strength. This is due to the fact that the expression in \Eq{eq:SNR_ideal} is an idealization -- decoherence will cause additional dynamics not accounted for in that analysis. \\
In \Fig{fig:no_purcell} we plot the integrated measurement record and the overlap fidelity of the corresponding post-measurement state for a relatively small SNR = $2\sqrt{2}$. The initial state is given in \Eq{eq:psi_pre}. Here we have not included any decoherence effects but the objective is rather to see how good the measurement can be in the absence of imperfections, just taking into account the effect of field transients.
\begin{figure}[htp]
\centering
\includegraphics[scale=0.5]{histo_no_purcell.pdf}
\caption{Measurement results for $\eta=1$ without decoherence, compared with simple Purcell model. Histograms a), b), and c) show the measurement results $s(\tau)$ with a fixed SNR = $2\sqrt{2}$. In d) we plot the overlap fidelity as function of measurement time (overlapping solid blue and red) along with the fraction of accepted results (red stem). The overlapping blue and red dashed lines show $F_\pm$ in the absence of measurement, including the effect of decoherence, with $\gamma_p = \chi/400$ and $\gamma_\varphi=\chi/300$. In e) we plot $F_\pm$ as a function of $s_{th}$ for $\tau = 10/\chi$ and f) shows the corresponding fraction of accepted measurement results.}
\label{fig:no_purcell}
\end{figure}
Each histogram in Fig. \ref{fig:no_purcell}(a), (b), and (c) shows the measurement results for 1000 simulated trajectories. In \Fig{fig:no_purcell}(d), we plot the overlap fidelities $F_\pm$, which are essentially identical for even and odd parities, along with the fraction of conclusive measurement results (even plus odd) for $s_{th} = 5$. As the measurement time is increased, the fidelity approaches unity yielding a perfect parity measurement. The fraction of accepted measurement results also increase with $\tau$ as the SNR approaches that of \Eq{eq:SNR_ideal}.\\
The dashed lines in \Fig{fig:no_purcell}(d) (essentially identical for even and odd) are the overlap fidelities without measurement, only including the effect of decoherence. Here, the initial states are given by $\ket{\psi_\pm}$. These lines give a benchmark for how fast we need to perform the measurement in order to not be limited by decoherence. Here, we choose the Purcell rate $\gamma_ p = \chi/400$ and dephasing $\gamma_\varphi = \chi/300$ such that if $\chi = 1$MHz, the relaxation rate and decoherence rate would be $T_1 = 1/\gamma_p = 400 \mu s$ and $T_2^* = 1/(\gamma_p/2 + \gamma_\varphi) = 218 \mu s$ respectively. Such a value for $\gamma_\varphi$ has been obtained in 3D circuit QED architectures \cite{RigettiPRB2013}, while the value of $\gamma_p$ is $\sim 4$ times smaller than current state-of-the-art experimental values (note that we have included all contributions to relaxation into $\gamma_p$ for simplicity). We believe that, with the ongoing experimental progress in improving these numbers \cite{SandbergAPL2013, ChangAPL2013}, numbers like the above should be possible in the near future. It should be noted that any application with the need for multiple qubits would place similar requirements on longer qubit lifetimes. For this choice of parameter values, we see that the measurement time needs to be $\tau \simeq 10/\chi$ since, for larger times, the measurement is limited by decoherence. \\
In \Fig{fig:no_purcell}(e) we plot $F_\pm$ as a function of $s_{th}$ for a measurement time of $\tau = 10/\chi$. By discarding measurement results (i.e., labelling them inconclusive), we can increase the conditional fidelity of the post-measurement state up to $> 98\%$. Note that for $s_{th} > 10$, the number of accepted measurement results are too few to allow good statistics, hence the increase in variance of $F_\pm$. The decreasing fraction of accepted results as function of $s_{th}$ is plotted in \Fig{fig:no_purcell}(c). Note that, we can get estimates for the physical parameters implied by these parameter settings: ignoring the distinction between different modes and different qubits, we get, using Eq. (\ref{exPurcell}) and the standard dispersive relation $\chi=g^2/\Delta$,
\begin{equation}
\Delta=4\chi^2/\gamma_p,\,\,\,\,g=\sqrt{4 \chi^3/\gamma_p}.
\end{equation}
This gives numerical values $\Delta= 1.6$GHz and $g=40$MHz. We see that a large value of detuning, combined with a moderate value of the qubit-cavity coupling constant $g$, gives the best measurement. Note that in order to avoid direct qubit-qubit coupling, the detuning $\Delta$ should be different from one qubit to the other by, say, hundreds of MHz, with the $g$'s correspondingly adjusted so that the $\chi$ parameters are all equal.
In \Fig{fig:with_Purcell} we plot the same quantities as in \Fig{fig:no_purcell} but with decoherence included.
\begin{figure}[htp]
\centering
\includegraphics[scale=0.5]{histo_with_purcell.pdf}
\caption{Measurement results with decoherence. Histograms a), b), and c) show the measurement results $s(\tau)$ with a fixed SNR = $2\sqrt{2}$. In d) we plot the overlap fidelity as function of measurement time (overlapping solid blue and red) along with the fraction of accepted results (red stem). The overlapping blue and red dashed lines show $F_\pm$ in the absence of measurement, including the effect of decoherence. In e) we plot $F_\pm$ as a function of $s_{th}$ for $\tau = 10/\chi$ and f) shows the corresponding fraction of accepted measurement results. The parameters are as in \Fig{fig:kernel} with $\gamma_p = \chi/400$, $\gamma_\varphi = \chi/300$ and $\eta =1$.}
\label{fig:with_Purcell}
\end{figure}
The peaks in the histograms are less separated than in \Fig{fig:no_purcell} since the Purcell decay mixes the different parity subspaces. The overlap fidelity in \Fig{fig:with_Purcell}(d) follows the fidelity set by the decoherence in absence of measurement (dashed lines). We observe that the fidelity is actually slightly better with the measurement on, which we can understand as a type of Zeno-effect. Since the Purcell relaxation is dominated by single qubit bit-flip errors, and these also change the parity of the state, the measurement partly protects the state from the dominant decay process with higher fidelity as a consequence. From \Fig{fig:with_Purcell}(d) we see that a post-measurement overlap fidelity of $\sim 90 \%$ is possible which can be increased to $\sim 95\%$ provided that $\sim 60\%$ of the measurement results are labelled as inconclusive.
\section{Conclusions}\label{sec:conclusions}
In conclusion, we have performed an analysis of a three qubit parity measurement in a circuit-QED setup where the joint state of two single mode resonances are used as pointer states of the measurement. We find that the measurement fidelity is limited by the transient dynamics of these pointer states and show that this limitation, in the absence of other decoherence mechanisms, can be overcome by the use of a weak probe signal. In this limit we can still obtain a high signal-to-noise ratio due to the fact that the steady state of the pointer states perfectly fulfills the conditions of a parity measurement, and the weak probe can thus be compensated by a longer measurement time. In the presence of additional decay, this strategy breaks down; but we show that, with realistic numbers for the decoherence, we can obtain a state fidelity of $\sim 95\%$ for the post-measurement state provided that we throw away $\sim 60\%$ of the measurement results. \\
The sort of measurement described here has direct application to the implementation of fault tolerant quantum computation using topological error correction codes employing three-qubit checks in the code of \cite{3par}, or the analogous four-qubit parity checks in the surface code\cite{BT}.When such applications are attempted, it is clear that very different statistical considerations would be brought to bear in the interpretation of the measurement results. In such error correction, there is an ``error free'' state of fixed parity, and the measurement is expected to give this outcome a large majority of the time. At a minimum, this would, on account of Bayesian reasoning, move the threshold $s_{th}$ away from its symmetric setting. Furthermore, there would never be any reason to interpret any measurement outcome as 100\% conclusive, since optimal corrective actions will be inferred from a large amount of measurement data of varying degrees of certainty. Finally, the correlation of measurement outcome $s$ with the overlap fidelity $F$ changes the interpretation of subsequent error syndrome measurements, because a departure of $F$ from unity implies a degradation of the multiqubit state which will be expected to show up as an erroneous parity outcome in the conjugate basis, which is needed on overlapping clusters of qubits in the surface code. More research will be needed to determine what measurement SNRs and fidelities are needed for the topological error correction to be successful. \\
Of course, there are further problems that are untouched by the present analysis; most real qubits have more than two quantum levels, which requires an extension of the present analysis, and brings in the possibility of leakage out of the computational space. It is only beginning to be understood how to effectively deal with leakage-type errors in surface-code error correction.\cite{leak} Nevertheless, the fact that there is no fundamental limitation to the fidelity of the proposed measurement scheme, indicates that as qubits with yet longer decoherence times become available, our circuit QED-based measurement schemes will become a prime tool for the preservation and control of complex quantum-computational states. \\
\section*{Acknowlegements}
We thank Firat Solgun and Ben Criger for useful discussions. DDV and SB are grateful for support from the Alexander von Humboldt
foundation. LT acknowledges financial support from the Swedish Research Council, and the EU through the projects SOLID and ScaleQIT.
|
\section{Introduction}
Approximation algorithms and parameterized complexity are two of the most
popular ways of dealing with NP-hard optimization problems. Nevertheless, the
two sets of techniques are usually treated independently. It's therefore a very
natural question whether combining the techniques of both theories can be used
to obtain algorithmic results which are out of reach for each one of them
separately. This has often been identified as a promising research field (see
\cite{Marx08} for a survey), but its development has so far been somewhat
limited. The goal of this paper is to add some results in this area by
designing parameterized approximation schemes for problems which are both
parameterized intractable (W-hard) and hard to approximate in polynomial time
(APX-hard).
The problems we will focus on are optimization problems on graphs of bounded
treewidth or clique-width. These two graph widths are of central importance to
parameterized complexity theory. At the same time, they play a significant role
in the design of approximation algorithms, since subroutines employing them are
often used as building blocks of larger algorithms. Therefore, understanding
the extent to which we can efficiently approximate problems which remain W-hard
for these widths is of potentially great importance from several points of
view.
In this paper we want to show that many such hard problems actually turn out
to be well-approximable in FPT time. Perhaps the easiest way to explain our
aim is to state a representative example of the type of results we will
establish.
\begin{theorem}[partial statement]
There exists a randomized $(1+\eps)$-approximation algorithm for \MC\ running
in time $\left(\frac{\log n}{\eps}\right)^{O(w)} n^{O(1)}$, where $w$ is the
input graph's clique-width.
\end{theorem}
\setcounter{theorem}{0}
\MC\ is of course a problem of central importance in the contexts of both
approximability and parameterized complexity. It is APX-hard (so an
approximation ratio of $1+\eps$ is probably impossible in polynomial time) and
W-hard parameterized by clique-width (so the fastest exact algorithm probably
needs time roughly $n^w$). Our main point here is that using a
\emph{parameterized approximation} approach we can evade these lower bounds,
leading to a $(1+\eps)$-approximation running in time only $(\log n)^{O(w)}$,
that is, an FPT approximation scheme. More generally, the goal of this paper
is to provide, in a \emph{uniform} way, similar approximation (or bicriteria
approximation) schemes for a diverse set of W-hard and APX-hard graph problems.
The problems for which we will provide algorithms are the following: \MC, \textsc{Edge Dominating Set},
\textsc{Bounded Degree Deletion}, \textsc{Capacitated Dominating Set}, \textsc{Capacitated Vertex Cover}, \EQC\ and \MMO. For most of these problems we are able to
provide equally efficient algorithms for both treewidth and clique-width (a
detailed description of all results is given further below).
\subsubsection*{Paper overview:} In this paper we adopt a generic technique
that is a variation of the standard dynamic programming used for treewidth and
clique-width. We observe that for a number of problems which are parameterized
intractable for these widths, the hardness intuitively stems from the fact that
large \emph{integers} need to be stored in the dynamic programming table. These
integers are usually calculated simply by \emph{adding} previously calculated
entries (all the problems listed above fall into this general category, though
for some this is not obvious). We want to shrink the table, and thus speed up
the algorithms, by storing these integers approximately.
The basic idea we use is very natural. We fix a parameter $\delta>0$ and
represent all integers in $\{1,\ldots,n\}$ by rounding them to the closest
integer power of $(1+\delta)$. If $\delta$ is not too small
($\delta=\Omega(\frac{1}{\log^c n})$) the natural dynamic programming table's
size is dramatically reduced from $n^w$ to $(\log n )^{O(w)}$. The obvious
obstacle to this approach, however, is that during the process of running a
dynamic programming algorithm on the approximate values the rounding errors
will propagate and potentially pile up to a large error. How can we keep the
errors under control?
In this paper we suggest a very natural randomized rounding approach to this
problem. The main contribution is to show that this rounding idea can be
seamlessly incorporated into the standard dynamic programming techniques of
treewidth and clique-width to give efficient approximation schemes. In order to
make the transition from exact to approximate algorithms as cleanly as possible
we separate the analysis into two parts. First, we introduce an abstract model
of computation, called Approximate Addition Trees, which captures the essence
of the rounding ideas we described. We fully analyze the approximation
performance of these Trees and prove some general approximation theorems. Then,
relying on this analysis we give a series of approximation algorithms using
clique-width and treewidth. It's worth stressing at this point that all the
algorithms we will present follow the standard dynamic programming mold that
should be very familiar to readers accustomed to graph widths. The important
difference is that their analysis, in addition to standard methods, also
crucially relies on our results on Approximate Addition Trees (which we can use
as a black box). Thus, by abstracting away the Addition Trees, our technique
can be viewed as a natural extension of well-known ideas. The hope is that
this modularization will allow our technique to be easily reused and eventually
become part of the standard graph width toolkit.
Thus, what is left to describe is the workings of Approximate Addition Trees.
This is, expectedly, the most technical part of the paper. As we will see,
there do exist some important special cases where a complicated analysis can be
avoided (notably, when the input tree is balanced) and there is some value in
these cases since they can help make some algorithms deterministic. However,
in order to obtain the more interesting results of this paper we need an
analysis of full generality. In other words, we need to establish an
approximation theorem that works for all Addition Trees without making any
special assumptions about their structure. Our main technical contribution is
that we do establish such a result and this allows us to analyze all the
algorithms of this paper in terms of Addition Trees. We thus present a robust,
unified technique that works for both treewidth and clique-width (and
potentially other similar graph widths), without relying on any non-trivial
width-specific properties.
\subsubsection*{Summary of results:} Let us now formally state the algorithmic
results presented in this paper. Full problem definitions are given further
below.
In the following theorems, $n^{O(1)}$ factors are omitted from the running
times.
\begin{theorem} \label{thm:cw-main}
Given an $n$-vertex graph $G(V,E)$, a clique-width expression with $w$ labels,
and an error parameter $\eps>0$, there exist randomized algorithms which, with
high probability, achieve the following:
\begin{itemize}
\item Produce an approximate solution to \MC\ with size at least
$\frac{\mathrm{OPT}}{1+\eps}$ in time $(\log n/\eps)^{O(w)}$.
\item Produce an approximate solution to \textsc{Edge Dominating Set}\ with size at most $(1+\eps)\mathrm{OPT}$
in time $(\log n/\eps)^{O(w)}$.
\item Given an integer $k$, either decide (correctly) that $G$ does not admit
an \EQC\ with $k$ colors or produce a valid $k$-coloring where the ratio of the
sizes of any two color classes is at most $(1+\eps)$ in time $(\log
n/\eps)^{O(k)} 2^{kw} $.
\item Given an integer $\Delta$ find a set of vertices that is at most as large
as the optimal solution for \textsc{Bounded Degree Deletion}\ to degree $\Delta$ and whose deletion makes
the maximum degree at most $(1+\eps)\Delta$, in time $(\log n/\eps)^{O(w)}$.
\item Given a capacity for each vertex, find a \textsc{Capacitated Dominating Set}\ of size at most $\mathrm{OPT}$,
such that all but at most $\eps n$ vertices are dominated, in time $(\log
n/\eps)^{O(w)}$.
\end{itemize}
In addition, if instead of a clique-width expression we are given a tree
decomposition of width $w$, there exist deterministic algorithms, with the same
running times, achieving all the above.
\end{theorem}
Furthermore, we also have the following results for treewidth.
\begin{theorem} \label{thm:tw-main}
Given an $n$-vertex graph $G(V,E)$, a tree decomposition of width $w$, and an
error parameter $\eps>0$, there exist deterministic algorithms which achieve
the following:
\begin{itemize}
\item Produce an approximate solution with cost at most $(1+\eps)\mathrm{OPT}$ for
\MMO, in time $(\log n)^{O(w)}$.
\item Given a capacity for each vertex, find a \textsc{Capacitated Dominating Set}\ (or \textsc{Capacitated Vertex Cover}) of size at most
$\mathrm{OPT}$, such that no capacity is violated by a factor of $(1+\eps)$ or more, in
time $(\log n)^{O(w)}$.
\end{itemize}
\end{theorem}
The algorithms achieving the results of Theorems \ref{thm:cw-main} and
\ref{thm:tw-main} are given in Section \ref{sec:algs}.
\subsubsection*{Previous work:}
Let us first review previous work for the specific problems we will focus on.
In \MC\ we want to partition the vertices of a graph into two sets so that the
number of edges with one endpoint in each set is maximized. \MC\ was shown to
be W-hard when parameterized by clique-width in \cite{FominGLS10}. The problem
is known to be APX-hard in general \cite{papadimitriou1991optimization}. In
\textsc{Edge Dominating Set}\ we want to select the smallest possible set of edges such that all edges
share an endpoint with a selected edge. This problem is also APX-hard and
W-hard for clique-width \cite{FominGLS10}. Both problems are FPT parameterized
by treewidth. Let us remark that, in addition to these two problems, very few
other problems are known to be W-hard for clique-width but FPT for treewidth.
The set of problems that behave this way are sometimes considered part of ``the
price of generality'' that clique-width affords, compared to treewidth.
Investigating this price has been an interesting research topic in
parameterized complexity. One interpretation of the results of this paper is
therefore that this ``price'' is not as high as previously believed, since two
of the most prominent problems from this family can be well-approximated for
clique-width.
In \EQC\ we want to find a proper $k$-coloring of a graph so that all color
classes have the same size. \EQC\ is known to be W-hard by the results of
Fellows et al. \cite{fellows2011complexity} even when parameterized by both
the treewidth of the input graph $w$ and the number of colors $k$. An exact XP
algorithm parameterized by $w+k$ can be easily obtained with standard
techniques, but a more general XP algorithm parameterized by $w$ only is shown
in \cite{bodlaender2005equitable}. The problem is easily shown to generalize
\textsc{Graph Coloring} (if we add a sufficiently large independent set to a
graph $G$ then $G$ admits an equitable coloring with $k$ colors if and only if
its chromatic number is at most $k$).
In \MMO\ we are given an edge-weighted graph and need to find an orientation of
the edges so that the maximum weighted out-degree of any vertex is minimized.
\MMO\ is sometimes also called \textsc{Minimum Maximum Outdegree} in the
literature. One motivation for the study of this problem is that it is a
special case of min-makespan scheduling (if vertices represent machines and
edges represent jobs that can be processed only by their endpoints). \MMO\ was
shown to be W[1]-hard parameterized by treewidth by Szeider \cite{szeider11}.
In \cite{AsahiroMO08} it was shown to be in P when all edge weights are equal,
while in \cite{Szeider11MFCS} an XP algorithm was shown when the problem is
parameterized by treewidth. Regarding polynomial-time approximations, in
\cite{EbenlendrKS08} the problem was shown to be hard to approximate with a
ratio better than 1.5 even if all weights are 1 or 2. In the same paper a
1.75-approximation was given.
In \textsc{Bounded Degree Deletion}\ we want to delete as few vertices as possible to make the maximum
degree of a graph $\Delta$. \textsc{Bounded Degree Deletion}\ was shown to be W[1]-hard parameterized by
treewidth by Betzler et al.~\cite{betzler2012bounded}. The problem generalizes
\textsc{Vertex Cover} and is therefore APX-hard to solve for general graphs.
In \textsc{Capacitated Dominating Set}\ (\textsc{Capacitated Vertex Cover}) we are given a graph where each vertex $v$ has a capacity
$c(v)$ and we want to find a dominating set (resp.~vertex cover) such that each
selected vertex $v$ is used to dominate at most $c(v)$ other vertices
(resp.~cover $c(v)$ edges). \textsc{Capacitated Dominating Set}\ and \textsc{Capacitated Vertex Cover}\ were shown W[1]-hard when
parameterized by the size of a minimum feedback vertex set (and therefore, also
when parameterized by treewidth) in \cite{DomLSV08}. Both problems are at least
APX-hard to approximate in polynomial time, since they generalize
\textsc{Dominating Set} and \textsc{Vertex Cover}. Since the algorithms we give
approximate the capacity (or degree) constraints they can be described as
$(1,1+\eps)$-bicriteria approximations for these problems.
Let us also recall some more general related work. Very few FPT approximation
schemes are currently known. For an overview of the most important results see
the survey by Marx \cite{Marx08}. The same paper gives an FPT approximation
scheme for \textsc{Max Vertex Cover} parameterized by the size of the cover.
This is extended in \cite{SkowronF13} to an FPT approximation scheme for
\textsc{Max Cover}. See also \cite{BonnetP13a} for FPT approximation schemes
for related covering problems. \textsc{Sum Edge Multicoloring} is a rare
example of a problem currently known to admit an FPT approximation scheme
parameterized by treewidth \cite{Marx04}.
Let us also mention that the notion of FPT approximation also makes sense when
one is trying to obtain constant factor approximations (instead of
approximation schemes) (see e.g. \cite{FratiGGM13}). It's also interesting to
approximate problems which are FPT, if the approximation algorithm can run in
significantly improved time (see \cite{BrankovicF10} or \cite{Bodlaender13}).
The complexity of parameterized approximations for naturally parameterized
problems (that is, parameterized by the size of the solution) has also been
considered. Unfortunately, the evidence so far seems to suggest that standard
problems, such as \textsc{Clique} and \textsc{Dominating Set} are hard to
approximate even in a parameterized setting
\cite{Escoffier12},\cite{ChitnisHK13}.
In this paper we focus on problems parameterized by treewidth or clique-width.
For an introduction to these notions see
\cite{bodlaender2008combinatorial,courcelle2000linear,EspelageGW01}. It was
initially believed that problems solvable on trees are almost always FPT
parameterized by treewidth. Gradually, many exceptions were discovered. This
includes problems such as \textsc{Edge-disjoint Paths} and
\textsc{L(2,1)-coloring}, which are NP-hard for graphs of treewidth 2
\cite{N01,FialaGK05} and \textsc{Steiner Forest} which is NP-hard for graphs of
treewidth 3 (\cite{BateniHM10} gives a PTAS for this problem using treewidth).
More relevant to our purposes are problems which are solvable in polynomial
time for constant treewidth, but not FPT. Some examples of such problems when
parameterized by treewidth (in addition to the problems we consider in this
paper) are the following: \textsc{Target Set Selection} \cite{Ben-ZwiHLN11},
\textsc{Maximum Path Coloring} \cite{Lampis11}, \textsc{List Hamiltonian Cycle}
\cite{Meeks11}, \textsc{List Coloring} (even if parameterized by vertex cover)
\cite{fellows2011complexity}, \textsc{General Factor} \cite{SamerS08},
\textsc{Generalized Satisfiability} parameterized by the treewidth of the dual
or incidence graph \cite{SamerS10}, \textsc{Generalized Domination}
\cite{Chapelle10}, \textsc{Bounded Edge-Disjoint Paths} \cite{GolovachT11}.
All problems which are W-hard for treewidth are of course also W-hard for the
more general clique-width. Additionally, in \cite{FominGLS09} it is shown that
\textsc{Graph Coloring, Hamiltonicity,} \MC\ and \textsc{Edge Dominating Set}\ are W-hard parameterized
by clique-width.
\section{Definitions and Preliminaries}
We assume that we are using the real-word RAM model. We use $\log(n)$ to denote
the base-2 logarithm of $n$ and $\ln(n)$ to denote the natural logarithm. In
addition $\log_{(1+\delta)}(n)$ is the logarithm base-$(1+\delta)$, for
$\delta>0$. We have $\log_{(1+\delta)}(n) = \ln(n)/\ln(1+\delta)$. The function
$\lfloor x \rfloor$, for $x\in\mathbb{R}$ denotes the maximum integer that is
not larger than $x$.
We use boldface to denote vectors, for example $\vec{d}$. Sometimes a vector
$\vec{s}\in S^k$ for $S$ a set and $k\in\mathbb{N}$ will also be viewed as a
function from $\{1,\ldots,k\}$ (or some other convenient set of size $k$) to
$S$, and vice-versa. For a function $f:S_1\to S_2$ we use $f^{-1}(u), u\in S_2$
to denote the set $\{ v\in S_1\ |\ f(v)=u \}$.
We will also use the following standard facts.
\begin{lemma} \label{lem:facts}
Let $x,\delta\in\mathbb{R}$. Then the following hold:
\begin{enumerate}
\item \label{item:a} $1+x \le e^x$
\item \label{item:b} If $x\in (0,\frac{1}{2})$ then $e^{x/2} \le 1+x$
\item \label{item:c} If $x\in (-\frac{1}{2},\frac{1}{2})$ then $e^{x} \le
1+x+x^2$
\item \label{item:d} If $x \in (-\frac{1}{2},\frac{1}{2})$ then
$|\ln (1+x)| \ge \frac{|x|}{2}$
\item \label{item:e} If $\delta \in (0,\frac{1}{2})$ and $\delta |x| \le
\frac{1}{2}$ then $|(1+\delta)^x-1| \ge \frac{1}{4} \delta |x|$
\end{enumerate}
\end{lemma}
\begin{proof}
For item \ref{item:a} one can consider the function $f(x)=e^x-x-1$. This
function has a global minimum at $x=0$ (this can be established by looking at
its derivative), thus $f(x)\ge f(0)=0$. For item \ref{item:b} we can use the
Taylor expansion $e^{x/2}=\sum_{i=0}^{\infty} \frac{(x/2)^i}{i!} \le 1 +
\frac{x}{2} + \sum_{i=2}^\infty \frac{x^i}{8} \le 1 + \frac{x}{2} +
\frac{x^2}{4} \le 1+x$, where we have used the fact that $x\le \frac{1}{2}$.
For item \ref{item:c} we again use the Taylor expansion $e^x = 1 + x +
\sum_{i=2}^\infty \frac{x^i}{i!} \le 1 + x + \sum_{i=2}^\infty \frac{|x|^i}{2}
\le 1 + x + x^2$, where in the last inequality we used the fact that $|x|\le
\frac{1}{2}$.
For item \ref{item:d} if $x\ge 0$ this follows from item \ref{item:b} by taking
the natural logarithm of both sides. For $x\le 0$ we can use the fact that
since $|x| < 1$ we have $\ln(1-|x|) = - \sum_{i=1}^{\infty} \frac{|x|^i}{i}$,
therefore $|\ln(1+x)| = |\ln(1-|x|)| = \sum_{i=1}^{\infty} \frac{|x|^i}{i} \ge
|x|$.
For item \ref{item:e} first consider the case $x\ge 0$. We have $\ln(1+\delta)
\ge \frac{\delta}{2}$ (by item \ref{item:b}) so since $x\ge 0$ we have
$x\ln(1+\delta) \ge \frac{\delta x}{2} \Rightarrow (1+\delta)^x \ge
e^{\frac{\delta x}{2}} \ge 1+\frac{\delta x}{2}$, where we used item
\ref{item:a}. Thus, $(1+\delta)^x-1 \ge \frac{\delta x}{2} \Rightarrow
|(1+\delta)^x-1| \ge \frac{\delta |x|}{2}$, since $x\ge 0$ and the item is
proved in this case. For $x<0$ we have $|(1+\delta)^x-1| = 1 - (1+\delta)^x =
\frac{(1+\delta)^{|x|}-1}{(1+\delta)^{|x|}}$. Using our calculations from the
$x\ge 0$ case we have $(1+\delta)^{|x|}-1 \ge \frac{\delta|x|}{2}$ so we get
$|(1+\delta)^x-1| \ge \frac{\delta|x|}{2(1+\delta)^{|x|}} \ge
\frac{\delta|x|}{2e^{\delta |x|}} \ge \frac{\delta |x|}{2 \sqrt{e}} \ge
\frac{\delta|x|}{4}$ where we have used that $1+\delta \le e^\delta$ (item
\ref{item:a}) and $\delta|x|\le \frac{1}{2}$. \qed
\end{proof}
We assume the reader has some familiarity with standard definitions from
parameterized complexity, such as the classes FPT and XP (see
\cite{flum2006parameterized}). For a parameterized problem with input size $n$
and parameter $k$ an FPT Approximation Scheme (FPT-AS) is an algorithm which,
given an error parameter $\eps>0$ runs in time $f(k,\frac{1}{\eps})$ (that is,
FPT time when parameterized by $k+\frac{1}{\eps}$) and produces a solution that
is at most a multiplicative factor $(1+\eps)$ from the optimal (see
\cite{Marx08}). The problems we consider in this paper are parameterized by
some graph width. We design (randomized) algorithms running in time
$\left(\frac{\log n}{\eps}\right)^{O(k)} n^{O(1)}$. By standard facts in
parameterized complexity theory such running times imply FPT algorithms.
\subsubsection*{Graph widths}
We use standard graph theoretic notation. For an undirected graph $G(V,E)$ and
$S\subseteq V$ we denote by $G[X]$ the graph induced by $X$. We will use the
standard notion of tree decomposition (for an introduction to this notion see
the survey by Bodlaender and Koster \cite{bodlaender2008combinatorial}). Given
a graph $G(V,E)$ a tree decomposition of $G$ is a tree $T(I,F)$ such that every
node $i\in I$ has associated with it a set $X_i\subseteq V$, called the bag of
$i$. In addition, the following are satisfied: $\bigcup_{i\in I} X_i = V$; for
all $(u,v)\in E$ there exists $i\in I$ such that $\{u,v\}\subseteq X_i$; and
finally for all $u\in V$ the set $\{ i\in I\ |\ u\in X_i\}$ is a connected
sub-tree of $T$. The width of a tree decomposition is defined as $\max_{i\in I}
|X_i| - 1$. The treewidth of a graph $G$ is the minimum treewidth of a tree
decomposition of $G$.
As is standard, when dealing with problems on graphs of bounded treewidth we
will assume that a ``nice'' tree decomposition of the input graph is supplied
with the input. In a nice tree decomposition the tree $T$ is a rooted binary
tree and each node $i$ of the tree is of one four special types (see
\cite{bodlaender2008combinatorial} for a definition).
We will also use the notion of clique-width (see
\cite{courcelle2000linear,EspelageGW01}). The set of graphs of clique-width
$w$ is the set of vertex-labelled graphs which can be constructed inductively
using the following operations:
\begin{enumerate}
\item Introduce: $i(l)$, for $l\in\{1,\ldots,w\}$ is the graph consisting of a
single vertex with label $l$.
\item Union: $\cup(G_1,G_2)$, for $G_1, G_2$ having clique-width $w$ is the
disjoint union of these two graphs.
\item Join: $\sigma(G,l_1,l_2)$, for $G$ having clique-width $w$ and
$l_1,l_2\in\{1,\ldots,w\}$ is the graph obtained from $G$ by adding all
possible edges from vertices with label $l_1$ to vertices with label $l_2$.
\item Rename: $\rho(G,l_1,l_2)$, for $G$ having clique-width $w$ and
$l_1,l_2\in\{1,\ldots,w\}$ is the graph obtained from $G$ by changing the label
of all vertices having label $l_1$ to $l_2$.
\end{enumerate}
A clique-width expression of width $w$ for $G(V,E)$ is a recipe for
constructing a $w$-labelled graph isomorphic to $G$. More formally, a
clique-width expression is a rooted binary tree such that each node has one of
four possible types, corresponding to the operations described above. In
addition, all leaves are Introduce nodes, each Introduce node has a label
$\in\{1,\ldots,w\}$ associated with it, and each Join or Rename node has two
labels in $\{1,\ldots,w\}$ associated with it. For each node $i$ the graph
$G_i$ is defined as the graph obtained by applying the operation of node $i$ to
the graph (or graphs) associated with its child (or children). All graphs $G_i$
are subgraphs of $G$ and for all leaves with label $l$ we define their
associated graph to be $\eta(l)$.
Again, as is customary, when dealing with a problem parameterized by
clique-width we will assume that a clique-width expression of the input graph
is supplied with the input. We can also assume without loss of generality that
the given clique-width expression has some nice properties. In particular,
whenever the operation $\sigma(G_i,l_1,l_2)$ is used we can assume that there
are no edges between vertices with labels $l_1,l_2$, since otherwise we can
edit the clique-width expression up to this point to remove the Join operations
that produced such edges and this does not affect the final graph.
\section{Approximate Addition Trees} \label{sec:trees}
In this section we describe an abstract model of computation which one may
naturally call Addition Trees. In such a Tree each node calculates a value that
is the sum of the values of its children. We also define an Approximate version
of these trees, where each node \emph{probabilistically} rounds calculated
values to integer powers of $(1+\delta)$, for some parameter $\delta>0$. These
trees capture the rounding scheme that will be the heart of the algorithms of
the next section. Our goal is to prove that the values of Approximate and
Exact Addition Trees are almost always very close, even if $\delta$ is not too
small (we want $\delta=\Omega(1/\log^cn)$). We require $\delta$ to be in this
range, because in the end the algorithms of the next section will run in time
roughly $(\log n/\delta)^w$. Thus, if we allow $\delta$ to become polynomial in
$n$ (which would make this section easy), we will get algorithms as slow as the
trivial exact ones.
Intuitively, there are two extreme cases to consider here. First, if a tree is
balanced (that is, it has logarithmic height), it is not hard to establish that
rounding errors cannot pile up too badly (Theorem \ref{thm:easy}). Somewhat
surprisingly, this easy case is already sufficient to obtain several
non-trivial algorithmic results, because tree decompositions can always be
balanced reasonably well (more details are given in the next section). However,
to obtain the more interesting results of this paper we need to deal with
clique-width, where the input decomposition cannot in general be balanced.
Therefore, we have to deal with general Addition Trees.
Our proof strategy then is to move on to a second extreme case: caterpillars.
Here the height of the tree is large, but we know that one operand of each
addition has no previously accumulated error. Despite this, this is actually a
pretty hard case. To see why, intuitively one can think of the accumulated
error at each level of the tree as a random variable, since the rounding
performed on each step is randomized. The error has some probability of
increasing and some of decreasing, depending on how we round, but it changes
by at most a factor of $(1+\delta)$ in each step. So, if we look at its
logarithm (with base $(1+\delta)$) it can (randomly) increase or decrease by at
most 1. Thus, the process we are trying to analyze is akin to a memoryless
random walk on the real line. We want to prove that the end result of the walk
after $n$ steps is with high probability contained in an area of size only
roughly $1/\delta = \mathrm{poly}(\log n)$. Such a statement would be,
however, false if the walk was completely unbiased, so we cannot rely on
standard tools, such as Chernoff bounds or the Azuma inequality, because the
result they give is too weak (they give concentration in an area of size
roughly $\sqrt{n}$). Instead, we need to use moment-generating functions to
derive a problem-specific concentration bound that takes into account our
algorithm's tendency to ``self-correct''. Because of this special tendency
(Lemma \ref{lem:self-correct}), our random walk is much more strongly
concentrated around its expectation than usual random walks.
Thus, eventually we establish (in Lemma \ref{lem:paths}) that the approximation
error is small in the caterpillar case, with high probability. Once this has
been shown we can extend the same ideas to prove a sufficiently good
approximation theorem for general trees by performing induction on the
``balanced height'' of the tree (Theorem \ref{thm:trees}). Roughly speaking,
the idea is that in any node of an arbitrary tree either both children have
roughly the same accumulated error (in which case the node is balanced) or one
has potentially much higher error (in which case the proof is similar to that
for caterpillars). So, combining the ideas of the two cases we can handle
arbitrary trees.
We remark that the only parts of this section needed to follow the results of
the next one are Definitions \ref{def:AT},\ref{def:AAT} and Theorems
\ref{thm:easy},\ref{thm:trees}. Let us now proceed to give full details.
For the following definition recall that a rooted binary tree is full if all
non-leaf nodes have exactly two children.
\begin{definition} \label{def:AT}
An Addition Tree is a full rooted binary tree $T$ where we associate with each
leaf $l$ a non-negative integer \emph{input} $x_l$ and with each node $v$ a
non-negative integer \emph{value} $y_v$. The inputs are given with $T$. The
value of each node is calculated as follows:
\begin{enumerate}
\item For each leaf $l$ we set $y_l:=x_l$
\item For each internal node $v$ with two children $u_1, u_2$ whose values have
already been calculated we set $y_v:= y_{u_1} + y_{u_2}$.
\end{enumerate}
\end{definition}
Let us also define an approximate version of this model.
\begin{definition} \label{def:AAT}
An Approximate Addition Tree with parameter $\delta$ is a full rooted binary
tree $T$ where we associate with each leaf $l$ a non-negative integer
\emph{input} $x_l$ and with each node $v$ a non-negative \emph{approximate
value} $z_v$. The approximate value of each node is calculated as follows:
\begin{enumerate}
\item For each leaf $l$ we set $z_l:=x_l$
\item For each internal node $v$ with two children $u_1,u_2$ we set $z_v :=
z_{u_1} \oplus z_{u_2}$, where the $\oplus$ operation is defined below.
\end{enumerate}
Let $a_v:=z_{u_1}+z_{u_2}$. We will call $a_v$ the \emph{initial approximate
value} of $v$.
We use $\oplus$ to denote the following operation: for two non-negative numbers
$x_1,x_2$ we define $x_1\oplus x_2:=0$ if $x_1=x_2=0$. Otherwise, select a real
number $r\in(0,1)$ uniformly at random and set $x_{1}\oplus x_{2}:=
(1+\delta)^{\lfloor \log_{(1+\delta)}(x_1+x_2) + r \rfloor}$.
\end{definition}
The motivation behind this definition is that the number of possible values
that can be stored in a node is much smaller than in an exact tree. In
particular, note that whenever $z_v$ is non-zero, it must be an integer power
of $(1+\delta)$. If the maximum value calculated by the exact tree is at most
polynomial in $n$ and $\delta=\Omega(1/\log^c n)$ then there are at most
$\mathrm{poly}(\log n)$ many different values that may appear in an approximate
tree. Using this we will be able to obtain smaller dynamic programming tables
in the next section. First though, we have to show that the approximate values
are close to the correct ones.
Let us now give the definition of approximation error by which we will measure
the progress of our algorithm. Since it is not hard to see that for any node
$v$ for which $y_v=0$ an Approximate Addition Tree will also have $z_v=0$, we
are only concerned with the approximation error for nodes where $y_v\neq 0$.
Therefore, in the remainder we will implicitly assume that we are talking about
a tree where for all $v$, $y_v>0$, because sub-trees with value 0 can be
ignored.
\begin{definition}
Let $v$ be a node of an Addition Tree, $y_v$ its (positive) value and $z_v$ its
approximate value calculated if we view the tree as an Approximate Addition
Tree. Then the error $\lambda_v$ is defined as
$\lambda_v:=\log_{(1+\delta)}\frac{z_v}{y_v}$.
\end{definition}
Note that $\lambda_v$ and $z_v$ are random variables (they depend on the random
rounding choices made during the computation), while $y_v$ is fully specified
once the inputs of the tree are fixed. Informally, $\lambda_v$ measures how
many ``$(1+\delta)$ intervals'' off our approximation is from the correct
interval.
Before we go on, let us make an easy observation that will be sufficient to
handle an important special case.
\begin{theorem} \label{thm:easy}
If an Approximate Addition Tree has maximum depth $h$ then for all nodes $v$ we
always have $|\lambda_v|\le h+1$. Therefore, if $\delta<\frac{\eps}{2h}$ then
for all $v$ we have $\max\{\frac{z_v}{y_v},\frac{y_v}{z_v}\} < 1+\eps$.
\end{theorem}
\begin{proof}
The proof is simple and proceeds by induction on $h$. For trees of height 0
(that is, isolated nodes), $z_v$ has the correct interval, so $|\lambda_v|\le
1$. For the inductive step observe that when adding two values the maximum
absolute relative error cannot increase by more than a factor of $1+\delta$.
\qed
\end{proof}
As a consequence of Theorem \ref{thm:easy} we get that in trees of height
$O(\log n)$ we can set $\delta = \Theta(1/\log n)$ and get error at most
$1+\eps$ everywhere \emph{with probability 1}. Thus, such balanced trees are an
easy case which is already solved, even without the use of randomization. As
mentioned though, we also need to handle the much more complicated general
case.
The main intuitive observation that we will use to give an approximation
guarantee can be summarized as follows: the process by which the initial
approximation $a_v$ is calculated is ``self-correcting'', while the rounding
step that follows it is unbiased. Therefore, the whole process tends to
self-correct. More precisely, we will show that if
$\lambda_{u_1},\lambda_{u_2}$ are the errors of the two children of a node,
then the error for $a_v$ is somewhere between the two. This means that unless
the two errors are exactly equal, the maximum of the two errors will have a
tendency to decrease. This intuition will be made more precise in Lemma
\ref{lem:self-correct}.
First, we need to introduce a definition and an auxiliary lemma.
\begin{definition}
Let $v$ be an internal node of an Approximate Addition Tree, and $a_v$ its
initial approximate value as in Definition \ref{def:AAT}. We define $p_v:=
\log_{(1+\delta)}(a_v)-\lfloor \log_{(1+\delta)}(a_v)\rfloor$.
\end{definition}
Of course $p_v$ is again a random variable, since it depends on $a_v$.
However, note that $p_v$, as defined, is exactly equal to the probability of
``rounding up'', that is, the probability (over the choice of $r\in(0,1)$) that
$\lfloor \log_{(1+\delta)}(a_v) + r \rfloor > \lfloor \log_{(1+\delta)}(a_v)
\rfloor$. To see this, observe that we round up if and only if
$\log_{(1+\delta)}(a_v) + r \ge \lfloor \log_{(1+\delta)}(a_v)\rfloor +1
\Leftrightarrow r\ge 1-p_v$ which is an event that has probability $p_v$, since
$r$ is selected uniformly at random and $p_v\in[0,1]$.
\begin{lemma} \label{lem:y2big}
Consider an Approximate Addition Tree with parameter $\delta<\frac{1}{2}$ and
let $v$ be an internal node with two children $u_1,u_2$. Then $z_{u_2} \ge
\frac{1}{2} \delta p_v z_{u_1}$.
\end{lemma}
\begin{proof}
We have $p_v = \log_{(1+\delta)}(a_v)-\lfloor \log_{(1+\delta)}(a_v)\rfloor$,
but $a_v = z_{u_1}+z_{u_2}$ therefore $\lfloor \log_{(1+\delta)}(a_v)\rfloor
\ge \lfloor \log_{(1+\delta)}(z_{u_1})\rfloor = \log_{(1+\delta)}(z_{u_1})$
where we used the fact that $z_{u_1}$ is an integer power of $(1+\delta)$.
We now have $p_v \le
\log_{(1+\delta)}(z_{u_1}+z_{u_2})-\log_{(1+\delta)}(z_{u_1}) =
\log_{(1+\delta)}\left( 1 + \frac{z_{u_2}}{z_{u_1}} \right)$. Therefore, to
establish the lemma it is sufficient to show that $\frac{z_{u_2}}{z_{u_1}} \ge
\frac{1}{2}\delta \log_{(1+\delta)}\left( 1 + \frac{z_{u_2}}{z_{u_1}}\right)$.
To ease notation let $\gamma = \frac{z_{u_2}}{z_{u_1}}$.
We have $\gamma \ge \frac{\delta}{2} \log_{(1+\delta)}(1+\gamma)
\Leftrightarrow \gamma \ln(1+\delta) \ge \frac{\delta}{2} \ln(1+\gamma)
\Leftrightarrow (1+\delta)^\gamma \ge (1+\gamma)^{\frac{\delta}{2}}$. Now, from
Lemma \ref{lem:facts}-(\ref{item:b}) and $\delta < \frac{1}{2}$ we have
$(1+\delta)^\gamma \ge e^{\frac{\delta\gamma}{2}}$, while from Lemma
\ref{lem:facts}-(\ref{item:a}) we have $(1+\gamma)^{\frac{\delta}{2}} \le
e^{\frac{\delta\gamma}{2}}$. The result follows. \qed
\end{proof}
We are now ready to give the main self-correction lemma.
\begin{lemma} \label{lem:self-correct}
Consider an Approximate Addition Tree with parameter $\delta<\frac{1}{2}$. Let
$v$ be an internal node with two children $u_1,u_2$. If
$\max\{|\lambda_{u_1}|,|\lambda_{u_2}|\}< \frac{1}{4\delta}$ then
$|\log_{(1+\delta)}\frac{a_v}{y_{u_1}+y_{u_2}}| \le
\max\{|\lambda_{u_1}|,|\lambda_{u_2}|\}
- \frac{1}{20} \delta p_v |\lambda_{u_1}-\lambda_{u_2}|$.
\end{lemma}
\begin{proof}
Informally, what this lemma states is that, before the rounding step is
applied, the absolute value of the error will decrease when compared with the
maximum error up to this point. In fact, the decrease will be proportional to
the absolute difference of the two previous errors.
Since we have not assumed an ordering on the children we can assume without
loss of generality that $|\lambda_{u_1}|\ge|\lambda_{u_2}|$. From the
definitions we have $a_v=(1+\delta)^{\lambda_{u_1}}y_{u_1} +
(1+\delta)^{\lambda_{u_2}}y_{u_2}$. If $\lambda_{u_1}=\lambda_{u_2}$ then the
inequality of the lemma is true, so assume that
$|\lambda_{u_1}|>|\lambda_{u_2}|$.
We now have $\log_{(1+\delta)}\frac{a_v}{y_{u_1}+y_{u_2}} = \lambda_{u_1} +
\log_{(1+\delta)}\frac{y_{u_1}+(1+\delta)^{\lambda_{u_2}-\lambda_{u_1}}y_{u_2}}{y_{u_1}+y_{u_2}}
$. Taking into account that $|\lambda_{u_1}|>|\lambda_{u_2}|$ it is not hard to
see that the second term is positive if and only if the first term is negative.
Therefore, $\left|\log_{(1+\delta)}\frac{a_v}{y_{u_1}+y_{u_2}}\right| =
|\lambda_{u_1}| - \left|
\log_{(1+\delta)}\frac{y_{u_1}+(1+\delta)^{\lambda_{u_2}-\lambda_{u_1}}y_{u_2}}{y_{u_1}+y_{u_2}}\right|$.
Now consider the term $\left|
\log_{(1+\delta)}\frac{y_{u_1}+(1+\delta)^{\lambda_{u_2}-\lambda_{u_1}}y_{u_2}}{y_{u_1}+y_{u_2}}\right|$.
We only need to show that this term is at least as large as $\frac{1}{20}\delta
p_v |\lambda_{u_1}-\lambda_{u_2}|$ to establish the lemma. Observe that this
term is an increasing function of $y_{u_2}$. Therefore, we can use a lower
bound on $y_{u_2}$ to give a lower bound on this term. But by Lemma
\ref{lem:y2big} we have $z_{u_2} \ge \frac{1}{2} \delta p_v z_{u_1}
\Leftrightarrow y_{u_2} \ge \frac{1}{2} \delta p_v
(1+\delta)^{\lambda_{u_1}-\lambda_{u_2}}y_{u_1}$. We thus get:
\begin{eqnarray}
\left| \log_{(1+\delta)}\frac{y_{u_1}+(1+\delta)^{\lambda_{u_2}-
\lambda_{u_1}}y_{u_2}}{y_{u_1}+y_{u_2}}\right|& \ge&
\left|\log_{(1+\delta)}\frac{y_{u_1}+\frac{1}{2}\delta p_v y_{u_1}}{y_{u_1}+\frac{1}{2}\delta
p_v (1+\delta)^{\lambda_{u_1}-\lambda_{u_2}}y_{u_1}}\right| = \label{eq:1}\\
&& = \left| \log_{(1+\delta)} \frac{1+\frac{1}{2}(1+\delta)^{\lambda_{u_1}-\lambda_{u_2}}\delta p_v}
{1+\frac{1}{2}\delta p_v} \right| = \label{eq:2}\\
&& = \frac{\left|\ln\left(1+\frac{\frac{1}{2}\delta p_v\left( (1+\delta)^{\lambda_{u_1}-\lambda_{u_2}}-1\right)}{1+\frac{1}{2}\delta p_v}\right)\right|}{\ln(1+\delta)} \label{eq:3} \ge\\
&& \ge \frac{1}{\delta}\left|\ln\left(1+\frac{\frac{1}{2}\delta p_v\left( (1+\delta)^{\lambda_{u_1}-\lambda_{u_2}}-1\right)}{1+\frac{1}{2}\delta p_v}\right)\right| \ge \label{eq:4}\\
&& \ge \frac{p_v \left| (1+\delta)^{\lambda_{u_1}-\lambda_{u_2}}-1\right|}{4+2\delta p_v} \ge \label{eq:5}\\
&& \ge \frac{1}{4} \frac{ \delta p_v|\lambda_{u_1}-\lambda_{u_2}|}{4+2\delta p_v} \ge\label{eq:6}\\
&& \ge \frac{1}{20} \delta p_v |\lambda_{u_1}-\lambda_{u_2}|\label{eq:7}
\end{eqnarray}
To go from (\ref{eq:1}) to (\ref{eq:2}) we used the fact that
$|\log(x)|=|\log(1/x)|$, while for (\ref{eq:3}) we simply changed the base of
the logarithm and performed some calculations. For (\ref{eq:4}) we use the fact
that $\ln(1+\delta)\le \delta$ which can be inferred from Lemma
\ref{lem:facts}-(\ref{item:a}). To get (\ref{eq:5}) we use Lemma
\ref{lem:facts}-(\ref{item:d}). To see that Lemma
\ref{lem:facts}-(\ref{item:d}) applies observe that
$(1+\delta)^{\lambda_{u_1}-\lambda_{u_2}} \le (1+\delta)^{\frac{1}{2\delta}}
\le \sqrt{e} \le 2$. To go from (\ref{eq:5}) to (\ref{eq:6}) we use Lemma
\ref{lem:facts}-(\ref{item:e}) and the fact that $|\lambda_{u_1}-\lambda_{u_2}|
\in (-\frac{1}{2\delta},\frac{1}{2\delta})$. Finally, (\ref{eq:7}) follows
from the fact that $\delta<\frac{1}{2}$ and $p_v\le 1$. \qed
\end{proof}
Lemma \ref{lem:self-correct} will be our main tool in proving that with high
probability the values of an Approximate Addition Tree are not too far from
those of the corresponding exact tree. We will proceed by induction, starting
from a simple case of path-like trees (caterpillars). We need the following
definition:
\begin{definition}
Let $T$ be a rooted full binary tree. The \emph{balanced height} of a node $v$,
denoted $\bh(v)$ is defined as follows:
\begin{enumerate}
\item If $v$ is a leaf then $\bh(v)=0$.
\item If $v$ is an internal node with children $u_1,u_2$ and
$\bh(u_1)>\bh(u_2)$ then $\bh(v)=\bh(u_1)$.
\item If $v$ is an internal node with children $u_1,u_2$ and
$\bh(u_1)=\bh(u_2)$ then $\bh(v)=\bh(v_1)+1$.
\end{enumerate}
\end{definition}
\begin{lemma} \label{lem:bheight}
In any full binary tree with $n$ nodes and root $r$ we have $n\ge 2^{\bh(r)}$.
\end{lemma}
\begin{proof}
The proof proceeds by induction on $n$. The lemma is trivial for $n=1$. For a
larger tree, consider the two trees rooted at children of the root. If they
have the same balanced height then by induction they both have at least
$2^{\bh(r)-1}$ nodes, which means the whole tree has at least $2^{\bh(r)}$
nodes. Otherwise, one has balanced height $\bh(r)$ and by induction that
sub-tree contains at least $2^{\bh(r)}$ nodes. \qed
\end{proof}
We will proceed inductively to prove a bound on the probability of a large
error occurring during the computation of an Approximate Addition Tree. Our
bound will depend on $\delta,n$ and the balanced height of the root of the
tree.
To begin, observe that the case $\bh(r)=0$ is trivial, as this can only occur
if the tree contains only one node. So the first interesting case is
$\bh(r)=1$, which happens if the rooted tree is made up of a single path with a
leaf attached to each internal node and the root.
In the remainder we will assume we are dealing with an Approximate Addition
Tree with $n$ nodes, root node $r$ and parameter $\delta\in(0,\frac{1}{2})$. We
first establish the following lemma:
\begin{lemma} \label{lem:paths}
If $\bh(r)\le 1$ then for all
$\lambda\in(\frac{2}{\sqrt{\delta}},\frac{1}{4\delta})$ we have $\PR{\exists v
\in T:\ |\lambda_v| > \lambda} \le 2n^2 e^{-\frac{\lambda\sqrt{\delta}}{20}}$.
\end{lemma}
\begin{proof}
We will first bound the probability that a certain node is the first to have
absolute error greater than $\lambda$, that is, the probability that it has
such an error even though all its descendants did not. Then we will take a
union bound over all nodes. Observe that it is sufficient to consider internal
nodes, since the errors for leaves are by definition 0.
Consider an arbitrary internal node $v$. We will bound the conditional
probability that its absolute error is larger than $\lambda$, given the fact
that all its descendants have smaller absolute error. (To ease presentation we
will omit this conditioning, which should be read implicitly in the remainder
of this proof). Let $t\in(0,\frac{1}{2})$ be a parameter we will set later. We
have
$$\PR{ |\lambda_v| > \lambda} = \PR{ e^{t|\lambda_v|}
> e^{t\lambda}} \le
\frac{\E{e^{t|\lambda_v|}}}{e^{t\lambda}}$$
where for the last step we applied Markov's inequality, since we are dealing
with a non-negative random variable.
Let us now try to bound the expectation which appears on the right-hand side.
The node $v$ has two children $u_1,u_2$, one of which must be a leaf, since the
tree has balanced height 1. Without loss of generality let $u_2$ be the leaf,
and therefore $\lambda_{u_2}=0$. By applying Lemma \ref{lem:self-correct} we
get that $|\log_{(1+\delta)}(a_v/y_v)|\le |\lambda_{u_1}| - \frac{1}{10}\delta
p_v |\lambda_{u_1}|$, where we are using the fact that we are conditioning
under the event that $|\lambda_{u_1}|<\lambda \le \frac{1}{4\delta}$.
Let us now look at the rounding step. Recall that we denote by $p_v$ the
probability of ``rounding up''. We have that $\log_{(1+\delta)}(z_v) =
\log_{(1+\delta)}(a_v) + r_v$, where $r_v$ is a random variable that depends on
the choice of $r\in(0,1)$. In particular, $r_v$ has the following distribution:
with probability $p_v$ we have $\log_{(1+\delta)}(z_v) = \lfloor
\log_{(1+\delta)}(a_v) \rfloor + 1$ and therefore $ r_v = 1 - p_v$. On the
other hand, with probability $1-p_v$ we have $\log_{(1+\delta)}(z_v) = \lfloor
\log_{(1+\delta)}(a_v) \rfloor$ and therefore $ r_v = -p_v$.
Rewriting we get $\lambda_v = \log_{(1+\delta)}\frac{z_v}{y_v} =
\log_{(1+\delta)}\frac{a_v}{y_v} + r_v \Rightarrow |\lambda_v| =
|\log_{(1+\delta)}\frac{a_v}{y_v} + r_v| = |\log_{(1+\delta)}\frac{a_v}{y_v}| +
r'_v \le |\lambda_{u_1}| - \frac{1}{10}\delta p_v |\lambda_{u_1}| + r'_v$ where
we define $r'_v$ as follows: $r'_v:= r_v$ if $a_v\ge y_v$ and $r'_v:=-r_v$
otherwise. Let us now bound the change caused in the expectation by the
rounding step.
\begin{claim}
For $t>0$ we have $\E{e^{t r'_v}\ |\ p_v}\le e^{t^2p_v }$
\end{claim}
\begin{proof}
In case $r'_v=r_v$ we have $\E{e^{t r_v}|p_v} = p_v e^{t(1-p_v)} + (1-p_v)
e^{-tp_v} = e^{-tp_v}(1-p_v+p_v e^t) \le e^{-tp_v} e^{-p_v+p_ve^t} \le e^{-tp_v
- p_v + p_v + tp_v + t^2p_v} = e^{t^2p_v }$, where we used Lemma
\ref{lem:facts}-(\ref{item:a}) and Lemma \ref{lem:facts}-(\ref{item:c}).
In case $r'_v=-r_v$ we have $\E{e^{-t r_v}|p_v} = p_v e^{-t(1-p_v)} + (1-p_v)
e^{tp_v} = e^{tp_v}(1-p_v+p_v e^{-t}) \le e^{tp_v} e^{-p_v+p_ve^{-t}} \le
e^{tp_v -p_v + p_v -tp_v +t^2p_v} = e^{t^2p_v}$ where again we used Lemma
\ref{lem:facts}. \qed
\end{proof}
We now have:
\begin{eqnarray}
\E{e^{t|\lambda_v|}} &=& \E{e^{t|\lambda_v|}\ |\ |\lambda_{u_1}|\ge \frac{1}{\sqrt{\delta}}} \PR{|\lambda_{u_1}|\ge \frac{1}{\sqrt{\delta}} }
+ \E{e^{t|\lambda_v|}\ |\ |\lambda_{u_1}|< \frac{1}{\sqrt{\delta}}} \PR{|\lambda_{u_1}|<
\frac{1}{\sqrt{\delta}}} \label{eq:8}
\end{eqnarray}
The second term of the right-hand-side of (\ref{eq:8}) can be upper-bounded by
$e^{t+t\sqrt{1/\delta}}$, since $|\lambda_v|\le |\lambda_{u_1}|+1$. We now seek
to find a value of $t$ such that the first term is upper-bounded by
$\E{e^{t|\lambda_{u_1}|}}$.
\begin{eqnarray}
\E{e^{t|\lambda_v|}\ |\ |\lambda_{u_1}|\ge \sqrt{1/\delta}} &\le&
\E{e^{t|\lambda_{u_1}| - t\frac{\delta}{20} p_v |\lambda_{u_1}| + t r'_v}\ |\ |\lambda_{u_1}|\ge \sqrt{1/\delta}} \le
\label{eq:9}\\
&\le& \E{\E{e^{t|\lambda_{u_1}| - t\frac{\sqrt{\delta}}{20} p_v + t
r'_v}\ |\ p_v}\ |\ |\lambda_{u_1}| \ge \sqrt{1/\delta}} \le \label{eq:10}\\
&\le& \E{e^{t|\lambda_{u_1}| - t\frac{\sqrt{\delta}}{20} p_v
+ t^2p_v}\ |\ |\lambda_{u_1}| \ge \sqrt{1/\delta}} \label{eq:11}
\end{eqnarray}
To obtain (\ref{eq:9}) we used Lemma \ref{lem:self-correct}. Then to obtain
(\ref{eq:10}) we used the fact that we are conditioning on
$|\lambda_{u_1}|\ge\sqrt{1/\delta}$ to replace $\lambda_{u_1}$ in one term of the
exponent. We also used standard properties of conditional expectations to
remove the dependence on $r'_v$, which using the claim gives (\ref{eq:11}). We
can now set $t=\frac{\sqrt{\delta}}{20}$ (so we do indeed have
$t\in(0,\frac{1}{2})$ as promised) and we get
\begin{eqnarray*}
\E{e^{\frac{\sqrt{\delta}}{20}|\lambda_v|}} &\le&
\E{e^{\frac{\sqrt{\delta}}{20}|\lambda_{u_1}|}\ |\ |\lambda_{u_1}|\ge \sqrt{1/\delta}}\PR{|\lambda_{u_1}|\ge \sqrt{1/\delta}}
+ e^{\frac{\sqrt{\delta}}{20}+\frac{1}{20} } \le \\
&\le& \E{e^{\frac{\sqrt{\delta}}{20}|\lambda_{u_1}| }} + 2
\end{eqnarray*}
Since we have established the above for any internal node $v$, it now follows
by induction than for any node $v$ we have
$$ \E{e^{\frac{\sqrt{\delta}}{20}|\lambda_v|}} \le 2n $$
Using this fact, we can now conclude that the probability that an arbitrary
internal node $v$ is the first to have absolute error greater than $\lambda$ is
at most $2ne^{-\frac{\lambda\sqrt{\delta}}{20}}$. Therefore, by union bound,
the probability that some node has absolute error at least $\lambda$ is at most
$2n^2e^{-\frac{\lambda\sqrt{\delta}}{20}}$, as claimed. \qed
\end{proof}
We are now ready to extend the previous lemma to trees of arbitrary balanced
height.
\begin{lemma} \label{lem:trees}
Let $T$ be an Approximate Addition Tree with root $r$ and $\bh(r)= h$. Then for
all $\lambda\in(\frac{2}{\sqrt{\delta}},\frac{1}{4\delta})$ we have $\PR{
\exists v\in T\ :\ |\lambda_v| > \lambda h} \le (h+1)n^{h+1}
e^{-\frac{\lambda\sqrt{\delta}}{20}}$.
\end{lemma}
\begin{proof}
The proof proceeds by induction on $h$. The base case for $h=1$ is exactly
Lemma \ref{lem:paths}, which we have already established. Now, assume that the
result has been proved for trees of balanced height up to $h-1$. Let $B_{h-1}$
be the event that there exists some node $u$ with $\bh(u)\le h-1$ such that
$|\lambda_u|>(h-1)\lambda$. Then we have:
$$ \PR{ \exists v\in T\ :\ |\lambda_v| > \lambda h} \le
\PR{ B_{h-1} } +
\PR{ \exists v\in T\ :\ |\lambda_v| > \lambda h\ |\ \neg B_{h-1}} $$
Let us bound the two terms separately. For the first term, for every node $u\in
T$ with $\bh(u)\le h-1$ we consider the sub-tree rooted at $u$ and containing
all its descendants. By the inductive hypothesis in a rooted tree of balanced
height at most $h-1$ the probability that there exists a vertex $u$ with
$|\lambda_u|>(h-1)\lambda$ is at most
$hn^he^{-\frac{\lambda\sqrt{\delta}}{20}}$. There are at most $n$ such trees
considered, so by union bound $\PR{ B_{h-1} } \le
hn^{h+1}e^{-\frac{\lambda\sqrt{\delta}}{20}}$.
For the second term, the argument is similar to that of the proof of Lemma
\ref{lem:paths}. Consider an arbitrary node $v$ of the tree such that
$\bh(v)=h$. We will bound the probability that this is the first node with
absolute error greater than $\lambda h$, given that the absolute errors of all
its descendants with the same balanced height are strictly smaller than
$\lambda h$. This time, we are also conditioning on the event $\neg B_{h-1}$,
so its descendants with smaller balanced height have error at most
$(h-1)\lambda$.
Observe that any node $v$ of balanced height $h$ has at most one child of
balanced height $h$ (otherwise the balanced height of the tree would be at
least $h+1$). If both children of $v$ have balanced height $h-1$ and the event
$\neg B_{h-1}$ is true, then $|\lambda_v|\le (h-1)\lambda+1 < \lambda h$ with
probability 1. So the interesting case is when exactly one child of $v$ has
balanced height $h$.
Let $u_1,u_2$ be the two children of $v$ with $\bh(u_1)=h$ and $\bh(u_2)<h$.
Again, $t\in(0,\frac{1}{2})$ is a parameter to be fixed later and we are
conditioning over the events that all descendants of $v$ have absolute error at
most $\lambda h$ and $\neg B_{h+1}$ is true.
$$\PR{ |\lambda_v| > \lambda h} = \PR{ e^{t|\lambda_v|}
> e^{t\lambda h}} \le
\frac{\E{e^{t|\lambda_v|}}}{e^{t\lambda h}}$$
We now need to bound the expectation which appears on the right-hand side.
Similarly to Equation (\ref{eq:8}) of Lemma \ref{lem:paths} we have:
\begin{eqnarray}
\E{e^{t|\lambda_v|}} &=& \E{e^{t|\lambda_v|}\ |\ |\lambda_{u_1}|\ge (h-1)\lambda + \sqrt{1/\delta}} \PR{|\lambda_{u_1}|\ge (h-1)\lambda + \sqrt{1/\delta}} \nonumber \\
&& + \E{e^{t|\lambda_v|}\ |\ |\lambda_{u_1}|< (h-1)\lambda + \sqrt{1/\delta}} \PR{|\lambda_{u_1}|< (h-1)\lambda+
\sqrt{1/\delta}} \label{eq:12}
\end{eqnarray}
By using the fact that $\neg B_{h-1}$ is true we know that if
$|\lambda_{u_1}|>(h-1)\lambda + \frac{1}{\sqrt{\delta}}$ then the difference
$|\lambda_{u_1}-\lambda_{u_2}|\ge\frac{1}{\sqrt{\delta}}$. Thus, we can
upper-bound the first term by $\E{e^{t|\lambda_{u_1}|}}$ in exactly the same
way as in Lemma \ref{lem:paths} by setting $t=\frac{\sqrt{\delta}}{20}$. The
second term is at most $e^{t((h-1)\lambda+\frac{1}{\sqrt{\delta}}+1)}$. We can
now use the same argument as in Lemma \ref{lem:paths} to get
\begin{eqnarray*}
\E{e^{t|\lambda_v|}} \le n e^{\frac{\sqrt{\delta}}{20}((h-1)\lambda+\frac{1}{\sqrt{\delta}}+1)} &\Rightarrow&
\PR{ |\lambda_v| > \lambda h} \le 2n e^{-\frac{\lambda\sqrt{\delta}}{20}}
\end{eqnarray*}
The above is an upper bound on the probability that an arbitrary node $v$ of
balanced height $h$ is the first to have absolute error more than $\lambda h$,
assuming all nodes of balanced height at most $h-1$ have error at most
$(h-1)\lambda$. By union bound, the probability that any node of balanced
height $h$ has error more than $\lambda h$ (again conditioned on $\neg
B_{h-1}$) is at most $2n^2e^{-\frac{\lambda\sqrt{\delta}}{20}}$. Summing with
the upper bound we have on the probability of $B_{h-1}$ we have that the
probability of any node having absolute error more than $\lambda h$ is at most
$ hn^{h+1}e^{-\frac{\lambda\sqrt{\delta}}{20}} +
2n^2e^{-\frac{\lambda\sqrt{\delta}}{20}} <
(h+1)n^{h+1}e^{-\frac{\lambda\sqrt{\delta}}{20}}$ and the result follows. \qed
\end{proof}
We are now ready to prove the main theorem of this section
\begin{theorem} \label{thm:trees}
Let $T$ be an Approximate Addition Tree on $n$ nodes with parameter
$\delta\in(0,\frac{1}{2})$. There exists a fixed constant $C>0$ such that for
all $\eps\in(0,\frac{1}{8})$ and sufficiently large $n$, if
$\delta<\frac{\eps^2}{C\log^6n}$ we have
$$ \PR{\exists v\in T: \max\{\frac{z_v}{y_v},\frac{y_v}{z_v}\} > 1+\eps} \le
n^{-\log n}$$
\end{theorem}
\begin{proof}
The theorem follows from Lemma \ref{lem:trees} as follows. First, note that
$|\lambda_v|>\lambda h \Leftrightarrow \max\{\frac{z_v}{y_v},\frac{y_v}{z_v}\}
> (1+\delta)^{\lambda h}$. By Lemma \ref{lem:facts} $(1+\delta)^{\lambda h} \ge
e^{\frac{\delta \lambda h}{2}} \ge 1 + \delta \lambda h/2$.
Now we set $\lambda=\frac{2\eps}{\delta h}$. Notice that $\frac{2\eps}{\delta
h}< \frac{1}{4\delta}$ because $h\ge 1$ (otherwise the tree is trivial). Also,
$\frac{2\eps}{\delta h} > \frac{2}{\sqrt{\delta}} \Leftrightarrow \eps > h
\sqrt{\delta}$ which holds because by Lemma \ref{lem:bheight} we have $h\le
\log n$. Therefore, we can apply Lemma \ref{lem:trees} for the chosen
$\lambda$. We get
$$ \PR{ \exists v\in T: \max\{\frac{z_v}{y_v},\frac{y_v}{z_v}\} > 1 + \eps} \le
(h+1) n^{h+1} e^{-\frac{\eps}{10h\sqrt{\delta}}} $$
The result follows by noting that the bound on the right is an increasing
function of $h$ and $h\le \log n$. \qed
\end{proof}
\section{Approximation Schemes} \label{sec:algs}
We are now ready to use the results of the previous section to design some
approximation algorithms. As mentioned, we only need Definitions
\ref{def:AT},\ref{def:AAT} (including the definition of the $\oplus$ operation)
and Theorems \ref{thm:easy}, \ref{thm:trees}. Let us first describe the
general technique.
In all problems we will assume we are supplied either a tree decomposition of
width $w$ or a clique-width expression with $w$ labels. An important property
of all the problems we consider is that they admit an exact dynamic program
that takes time (roughly) $n^w$. These dynamic programs calculate a set of $w$
integers on each node of the decomposition by adding previously calculated
values or values read from the graph. The idea is to reuse this dynamic
program, but round all values to integer powers of $(1+\delta)$ and perform all
additions with the $\oplus$ operation. Note that, we are in fact able to also
handle some other auxiliary operations besides addition (such as comparisons).
It will, however, be important that our dynamic programs avoid using
subtractions (because then we lose the approximation guarantee) and part of our
effort is devoted to achieving this.
How do we analyze the approximation ratio of such an algorithm? As usual, each
node of the tree decomposition or the clique-width expression represents a
subgraph of the original graph. For each problem, we have a notion of partial
solution confined to a subgraph. We want to show that for each partial solution
there exists (with high probability) an entry in the approximate dynamic
programming table that matches its value and vice-versa. This is where Addition
Trees become useful. For each partial solution value we define inductively an
Addition Tree $T$ whose root calculates that value. We then inductively
establish that the approximate dynamic programming table contains some value
that \emph{follows the same distribution} as the result of $T$, if $T$ is
viewed as an Approximate Addition Tree. Thus, the approximate dynamic
programming table contains (with high probability) an almost correct value. To
simplify things, by tweaking the parameters appropriately we can make the
probability high enough that \emph{all} the values of the approximate table are
close to being correct, assuming $w$ is not too large. Observe that this
allows the analysis to be performed using essentially the same inductive
reasoning as for standard (exact) dynamic programming algorithms.
What remains to say is what values we select for the parameter $\delta$. Here
we have two choices. In the general case, we set $\delta$ to the value dictated
by Theorem \ref{thm:trees}. This works quite well essentially all the time, as
we can guarantee that with high probability the Addition Trees we consider
during the analysis of our algorithm will have almost correct values. We can
thus obtain randomized approximation schemes for both clique-width and
treewidth with the promised running time of roughly $(\log n/\eps)^{O(w)}$.
This is the general technique we consider to be the main contribution of this
paper.
Nevertheless, as mentioned there exists an important special case where things
can become simpler. It is known that for any graph of treewidth $w$ there
exists a tree decomposition of width $O(w)$ and only logarithmic (in $n$)
height \cite{BodlaenderH98}.
Thus, another approach available to us is to use this theorem to first balance
the decomposition and then rely on Theorem \ref{thm:easy}, instead of the more
general Theorem \ref{thm:trees}. Unfortunately, this does not speed up our
algorithms significantly (because the larger $\delta$ given in Theorem
\ref{thm:easy} is almost out-weighed by the blow-up in the width of the
decomposition). Importantly, it is impossible to apply this approach to
clique-width, because similar balancing results blow up the number of labels
used exponentially (\cite{CourcelleK07}), which would result in an unreasonable
running time of roughly $(\log n)^{2^w}$ (and this is likely to be
unavoidable). More generally, relying on a treewidth-specific balancing theorem
detracts from the bigger point of developing a general technique, since there
are many graph widths for which balancing would not even make sense
(e.g.~pathwidth). Despite these shortcomings, we still believe there may be
some value in this treewidth-specific balancing trick and mention it as an
alternative approach because it allows us to get rid of randomization in this
case. To the best of our knowledge, this is the first application of tree
decomposition balancing theorems to approximation algorithms (the main
motivation for their study so far was parallel and distributed algorithms). It
would be interesting to see if in the future a combination of a balancing
preprocessing step with the ideas of Theorem \ref{thm:trees} could lead to
better running times or more derandomized algorithms.
In the rest of this section we first, for the sake of completeness, give a
complete proof of one algorithmic result, namely the approximation scheme for
\MC. We then simply sketch the dynamic programs for all other problems, since
the analysis is similar, highlighting some interesting details. We also give
some brief remarks explaining why some algorithms cannot (easily) be extended
to clique-width and why for some problems we only get bi-criteria
approximations.
\subsection{Max Cut}
In this section we attack the \MC\ problem parameterized by clique-width. In
\MC\ we are given a graph $G(V,E)$ and are asked to find a partition of $V$
into $L,R\subseteq V,\ L=V\setminus R$ such that the number of edges with
exactly one endpoint in $L$ is maximized. We assume that a clique-width
expression for $G$ with $w$ labels is supplied as part of the input.
We will follow the straightforward dynamic program for this problem, as used
for example in \cite{FominGLS10}. The idea is to describe a partial solution by
keeping track of the intersection of each label set with $L$. The first
difference is of course that we will round all values to save time. The second
(somewhat counter-intuitive) difference is that for each label set we will keep
track of the size of its intersection with \emph{both} $L$ and $R$. In the
exact program this would be wasteful, since one of these two numbers can be
calculated from the other, using the (known) size of the whole label set.
However, here we are trying to implement a dynamic program that relies only on
additions. As a side effect, our dynamic program will likely include solutions
which are clearly incorrect (the sum of the sizes of the two intersections
being different from the size of the set) but this will not be a major problem
if we can guarantee that all values are at most $(1+\eps)$ far from a valid
solution.
\subsubsection*{Dynamic Program}
As mentioned, we view the clique-width expression as a rooted binary tree.
First, define the set $B:=\{0\}\cup \{ (1+\delta)^j\ |\ j\in\mathbb{N}\}$.
Informally, $B$ is the set of rounded values that may appear in our table. Even
though $B$ is infinite, whenever the algorithm produces an entry with value
larger than $(1+\eps)n^2$ we simply drop that entry. It will not be hard to see
that this will not affect the analysis if all entries are within a $(1+\eps)$
factor of being correct (then nothing will be dropped). Observe that the size
of $B$ then becomes $\log_{(1+\delta)}(n^2) = \mathrm{poly}(\log n/\eps)$, if
we set $\delta$ according to Theorem \ref{thm:trees}.
The dynamic programming table $D_i$ for a node $i$ is a subset of $B^w\times
B^w\times B$. The informal meaning is that an entry $(\vec{l},\vec{r},c)\in
D_i$ if and only if there exists a partition of the subgraph of $G$ represented
by the sub-tree rooted at $i$ into $L_i,R_i$ such that:
\begin{enumerate}
\item for all $l\in\{1,\ldots,w\}$ the number of vertices in $L_i$
(respectively $R_i$) with label $l$ is roughly $\vec{l}(l)$ (respectively
$\vec{r}(l)$)
\item the number of edges with exactly one endpoint in $L_i$ is roughly $c$
\end{enumerate}
A dynamic programming algorithm can now be formulated in a straightforward way.
It is easy to fill the table for initial nodes. For Rename and Union nodes the
exact dynamic program would perform some addition, which we now replace with
the $\oplus$ operation. For example, consider a Rename node with labels
$l_1\to l_2$. For each entry $(\vec{l},\vec{r},c)$ in the child's table we
create an entry $(\vec{l'},\vec{r'},c)$ in the current node as follows: we set
$\vec{l'}(l_1):=0$, $\vec{l'}(l_2):= \vec{l}(l_1)\oplus \vec{l}(l_2)$ and
$\vec{l'}$ the same as $\vec{l}$ for other labels to make the vector $\vec{l'}$
of a new entry (similarly for $\vec{r'}$). In the same way, for a Union node,
for each entry $(\vec{l_1},\vec{r_1},c_1)$ in the first child's table and for
each entry $(\vec{l_2},\vec{r_2},c_2)$ in the second child's table we construct
an entry $(\vec{l_1}\oplus \vec{l_2}, \vec{r_1}\oplus\vec{r_2}, c_1\oplus
c_2)$, where $\oplus$ is applied component-wise for vectors.
For Join nodes with labels $l_1,l_2$ we do something similar. For each entry
$(\vec{l},\vec{r},c)$ in the child's table construct a new vector with the same
$\vec{l},\vec{r}$ and $c':= c \oplus (\vec{l}(l_1)\cdot \vec{r}(l_2) +
\vec{l}(l_2)\cdot \vec{r}(l_1))$. Note that if the elements of
$\vec{l},\vec{r}$ are known with error at most $(1+\eps)$ then the second term
of this addition is known with error at most $(1+\eps)^2\approx 1+2\eps$.
\subsubsection*{Analysis}
First, observe that because the running time of the algorithm is polynomial in
the size of the tables, the algorithm clearly achieves the running time stated
in Theorem \ref{thm:cw-main}. We only have to prove its approximation
guarantee.
Any node $i$ of the clique-width expression defines a subgraph $G_i$ of $G$
(the graph produced by the sub-expression rooted at $i$). A partial solution is
simply a restriction of a solution $L,R$ for $G$ to the vertices of $G_i$. The
\emph{signature} of a partial solution is a vector of $2w+1$ integers that
would represent this solution in an exact dynamic programming table. In other
words, the signature is the entry we would expect to see in the table $D_i$ if
we were not rounding. In this case, it contains the exact size of the
intersections of $L,R$ with each label set and the size of the cut in $G_i$. To
keep things simpler, we will only be concerned with the $2w$ values that
represent the intersection. As mentioned, if we can get these right, the
algorithm also gets the size of the cut right within a factor of roughly
$(1+2\eps)$.
Consider a partial solution and its signature. Our strategy is to inductively
define a mapping that gives for each such signature a collection of $2w$
Addition Trees. The exact results of these trees will be the values of the
signature (that is, the sizes of the intersections of the two sets with each
label). We will then establish (by induction) that there exists an entry in our
algorithm's table whose values follow the same distribution as those Trees, if
they are viewed as Approximate Addition Trees. It will follow that for every
partial solution there exists, with high probability, an entry with roughly the
same values. The converse statement can be shown with essentially the same
ideas.
The above can clearly be done for initial nodes, where all values stored are 0
or 1, so our algorithm stores them exactly. The relevant Addition Trees are of
height 0. Assume inductively that we have established the correspondence up to
a certain height in the clique-width expression. Let us then consider a Rename
node $i$ with labels $l_1\to l_2$ and a child $j$. Consider a partial solution
to $G_i$. This partial solution has some corresponding partial solution in
$G_j$. By the inductive hypothesis, there exists an entry in
$(\vec{l},\vec{r},c)\in D_j$ corresponding to this solution. In particular,
there exist Approximate Addition Trees whose results have the same distribution
as $\vec{l}(l_1)$ and $\vec{l}(l_2)$ and whose exact values are the same as the
corresponding values in the partial solution to $G_j$. Consider the Tree
obtained from these by adding a new root and making the old roots its children.
This Tree now follows the same distribution as the value $\vec{l'}(l_2)$
calculated by our algorithm. Its exact value is the value we would get in the
exact signature (since the same was true for the sub-trees). Reasoning in the
same way about the other coordinates we have completed the inductive step in
this case.
The cases of Join and Union nodes can be handled with similar inductive
arguments as above. We can thus establish that for any valid partial solution
signature there exists an entry of the table that approximately corresponds to
it, in the sense that their respective values are connected through Addition
Trees. It is also not hard to use inductive reasoning to also establish the
converse (for every approximate entry there exists a corresponding partial
solution). We now select the entry in the root's table that has maximum $c$.
With high probability, it must correspond to a solution with approximately the
same cut size (otherwise we can find an Approximate Addition Tree with large
error). Retracing the steps of the dynamic programming we can turn this entry
into an actual cut.
\subsection{Edge Dominating Set}
Let us now give a dynamic programming algorithm for \textsc{Edge Dominating Set}. Here, things are a
little trickier, because it's not immediately obvious that using subtractions
can be avoided. We will use the following equivalent version of the problem: we
are looking for a minimum-size set of vertices $S$ such that $S$ is a vertex
cover of $G$ and $G[S]$ has a perfect matching. It is not hard to see that this
is the same problem (intuitively, $S$ is the set of vertices incident on an
edge of the edge dominating set) because an optimal edge dominating set is also
a maximal matching.
We define the dynamic programming table $D_i$ for a node $i$ as a subset of
$(B\cup\{F\})^w\times B^w$, where $F$ is a special symbol. Let $G_i$ be the
corresponding subgraph. Fix a solution to the problem in $G$, that is a vertex
cover $S$ and a matching of its vertices $M$. The intended meaning is that an
entry $(\vec{s},\vec{c})\in D_i$ represents this solution if
\begin{enumerate}
\item for all $l\in\{1,\ldots,w\}$ the number of vertices of $G_i$ with label
$l$ included in $S$ is roughly $\vec{s}(l)$. If $\vec{s}(l)=F$ then \emph{all}
vertices with label $l$ are in $S$.
\item for all $l\in\{1,\ldots,w\}$ the number of vertices of $G_i$ with label
$l$ incident to an edge of $M\cap G_i$ is roughly $\vec{c}(l)$
\end{enumerate}
Informally, $\vec{s}$ tells us how many vertices we have selected in $S$ from
each label set, while $\vec{c}$ tells us how many of the selected vertices have
already been matched. Clearly, the intended meaning implies that $\vec{c}(j)\le
\vec{s}(j)$ for all $j$, but this may not always be the case. We will still be
happy if this is true up to an error factor $(1+\eps)$. In the calculations
below, when we perform arithmetic operations on a value $\vec{s}(l)$ that is
equal to $F$ we substitute it with the size of the set $V_l$ of vertices with
label $l$.
Let us now describe the algorithm. Initial nodes are easy ($\vec{c}$ is
all-zero, $\vec{s}$ can have a single coordinate that is 0 or F). For Rename
and Union nodes we just have to add (component-wise) appropriate entries,
taking care to maintain $F$ values if possible. The interesting case is Join
nodes.
Let $i$ be a join node with labels $l_1 \leftrightarrow l_2$ and child $j$.
For each entry $(\vec{s},\vec{c})\in D_j$ do the following. Let $V_{l_1},
V_{l_2}$ be the set of vertices with label $l_1,l_2$ respectively in $G_i$. If
$\vec{s}(l_1) \neq F$ and $\vec{s}(l_2) \neq F$ then ignore this entry because
this partial solution is not a vertex cover of $G_i$. Otherwise, for each
$m\in\{0,\ldots,\min(|V_{l_1}|, |V_{l_2}|)\}$ calculate a vector $\vec{c_m}$
which is identical to $\vec{c}$ except that $\vec{c_m}(l_1)=\vec{c}(l_1)\oplus
m$ and $\vec{c_m}(l_2)=\vec{c}(l_2)\oplus m$. Informally, we are selecting how
many of the join edges will eventually be used in the matching $M$. If we have
$\vec{c_m}(l_1)>(1+\eps)\vec{s}(l_1)$ or $\vec{c_m}(l_2)>(1+\eps)\vec{s}(l_2)$
ignore $\vec{c_m}$. Otherwise, add $(\vec{s},\vec{c_m})$ to $D_i$.
Once the root's table has been calculated we select the entry
$(\vec{s},\vec{c})$ such that $\sum_l \vec{s}(l)$ is minimized, among entries
where $\vec{c}(l) \ge \vec{s}(l)/(1+\eps)$. Retracing the steps of the dynamic
programming we then obtain a vertex cover $S$. In polynomial time we can
calculate a maximum matching in $G[S]$. This is our initial candidate
solution. If some vertex of $S$ is unmatched we add one of the edges connecting
it to $V\setminus S$. We now have an edge dominating set.
We are now faced with two problems. First, we need to prove that $S$ has
roughly the same number of vertices as an optimal solution. Second, we need to
prove that $G[S]$ contains an almost perfect matching (more precisely, it
contains a set of edges $M$ such that almost all vertices are incident on one
edge of $M$). We can again rely on an inductive analysis using Approximate
Addition Trees, as in the case of \MC. The main difference is that our
algorithm now occasionally drops some partial solutions. This happens in the
case of Join nodes when we have not selected enough vertices of
$V_{l_1},V_{l_2}$ to produce a proper vertex cover. This step is easily seen
to be correct, as there is no approximation involved. Alternatively, solutions
are dropped when we are trying to select too many of the new edges in $M$. In
this case, we leave enough ``slack'' in our comparisons so that if a solution
is erroneously dropped we can extract an Approximate Addition Tree with high
error, something that can only happen with low probability. Barring this, we
can assume that all partial solutions are represented in the root's table and
each table entry approximately corresponds to a solution. Thus, the solution
we select in the end will have an almost optimal size for $S$ and contain an
almost perfect matching.
\subsection{Equitable Coloring}
This problem admits a very simple additive dynamic program. Let us briefly
describe it. In the case of clique-width, for each node the table is a subset
of $(B^k)^w$. Informally, the signature of a partial coloring is the number of
vertices of each color contained in each label set. We can guarantee to produce
a valid coloring if we make sure that in all Join nodes we drop partial
solutions that use the same color somewhere in both label sets. This can be
done without a problem because $0$ values are stored exactly in our table. In
the end, we pick the table entry appearing to give the most equitable coloring
and extract a coloring by retracing the dynamic program. The running time is
$(\log n/\eps)^{O(kw)}$.
Let us also note that we can give a faster algorithm for treewidth. Here the
dynamic program needs to remember the size of each color class and the coloring
of the $w$ vertices of a bag. Thus, the table is a subset of $B^w\times
\{1,\ldots,k\}^w$, leading to a running time of $(k\log n/\eps)^{O(w)}$.
\subsection{Capacitated Dominating Set and Vertex Cover}
Let us first describe the algorithm for \textsc{Capacitated Dominating Set}\ on clique-width promised in
Theorem \ref{thm:cw-main}. The dynamic programming table $D_i$ is a subset of
$B^w\times B^w\times B^w\times\{0,\ldots,n\}$. Informally,
$(\vec{a},\vec{u},\vec{d},c)\in D_i$ if there exists a set $S$ of
\emph{exactly} $c$ vertices and a partial mapping of vertices of $V\setminus S$
to $S$ such that
\begin{enumerate}
\item The total capacity of vertices of $S$ with label $l$ is roughly
$\vec{a}(l)$
\item The number of vertices with label $l$ in $V\setminus S$ mapped to $S$
(i.e.~dominated) is roughly $\vec{d}(l)$
\item The number of vertices of $V\setminus S$ mapped to vertices in $S$ with
label $l$ is roughly $\vec{u}(l)$.
\end{enumerate}
More simply, $\vec{a}$ is the total available capacity in a label set,
$\vec{u}$ is the capacity already used in our partial solution and $\vec{d}$ is
the number of vertices of each label set that we have already covered.
It is not hard to see how to fill this table for Initial, Rename and Union
nodes (only additions are needed). For Join nodes, the interesting point is
that we need to decide how many of the new edges are used. Similarly to the
case of \textsc{Edge Dominating Set}, for each possible number $m$ we add $m$ to the \emph{used}
capacity $\vec{u}$ on one side and the number of \emph{dominated} vertices
$\vec{d}$ on the other. We drop solutions which are clearly (by more than a
factor $1+\eps$) invalid. Note that, because for each label set we only care
about its \emph{total} capacity, not the number of selected vertices, we can
afford to keep track of the total size of the dominating set exactly. This only
adds a polynomial factor to the algorithm's running time.
In the end we extract a solution from the root's table and argue about its
correctness as usual. One complication is that, since we do not know $\vec{a}$
and $\vec{d}$ exactly, the solution may be violating some capacities and it may
not be a complete dominating set. We repair the solution by allowing some more
vertices to be undominated. If the total capacity of each label set was
violated by no more than $(1+\eps)$ this drops at most $\eps n$ vertices.
Similarly, if the solution was not a complete dominating set (and one exists),
at most $\eps n$ vertices are not dominated. We thus get the promised
bi-criteria guarantee.
Let us also discuss the case of treewidth (Theorem \ref{thm:tw-main}). Here the
dynamic program for \textsc{Capacitated Dominating Set}\ is easier. We just need to remember which of the
vertices of each bag have been selected and how much of their capacity has
already been used. The result is a set that dominates the whole graph, has size
at most $\mathrm{OPT}$, but may violate some capacities by $(1+\eps)$. Observe that
this is stronger than the bi-criteria approximation we get for clique-width
since, again, we could drop the coverage for some vertices to fix the
capacities (but in general, it's not clear if we can trade in the other
direction).
It's worthy of note that for \textsc{Capacitated Dominating Set}\ we can probably not hope to obtain anything
better than a bi-criteria approximation, such as the one we gave here. The
reason is that in \cite{DomLSV08} it is shown that \textsc{Capacitated Dominating Set}\ is W-hard even if
parameterized by \emph{both} the treewidth and $\mathrm{OPT}$. Thus, if we could obtain
an FPT approximation scheme for $\mathrm{OPT}$ without violating constraints, we would
be able to set $\eps$ to an appropriate small value and obtain an FPT algorithm
for a W-hard problem.
Finally, let us mention the \textsc{Capacitated Vertex Cover}\ algorithm promised in Theorem
\ref{thm:tw-main}. This can be obtained by reducing \textsc{Capacitated Vertex Cover}\ to a
vertex-weighted version of \textsc{Capacitated Dominating Set}. First, observe that if we associate with each
vertex of an instance of \textsc{Capacitated Dominating Set}\ a non-negative integer cost and ask for a
solution of minimum total cost the algorithm we gave still works with minimal
modification. Consider now a \textsc{Capacitated Vertex Cover}\ instance. First, sub-divide each edge (this
does not increase the treewidth $w$) and set the capacities of new vertices to
$0$. Then, add a new vertex $u$ and connect it to all of the original vertices
of the instance (this increases $w$ by 1). Set the capacity of $u$ to be equal
to its degree. Finally, set the cost of each original vertex to $1$, the cost
of $u$ to $0$ and the cost of vertices constructed in the sub-divisions to $n$.
If we view this as an instance of weighted \textsc{Capacitated Dominating Set}\ it is straightforward that a
solution of cost $s<n$ exists if and only if the original \textsc{Capacitated Vertex Cover}\ instance had a
solution of size $s$.
It would be interesting to see if an algorithm for \textsc{Capacitated Vertex Cover}\ on clique-width can be
given. At the moment this seems more challenging than for \textsc{Capacitated Dominating Set}, because it is
not sufficient to remember the total available capacity of a whole label set.
\subsection{Bounded Degree Deletion}
A dynamic program for clique-width can be given as follows: each table is a
subset of $B^w\times B^w\times \{0,\ldots,n\}$. Informally, a partial solution
is a set of ``active'' vertices which will \emph{not} be deleted. We remember
how many of these we have in each label set. We also remember what is the
maximum degree of any active vertex in each label set. Finally, we remember the
total number of vertices we have deleted. The only interesting case is Join
nodes, where the new maximum degree is calculated by adding to the previous
maximum degree the number of active vertices in the other label set.
A dynamic program for treewidth can be formulated by keeping track of the
active vertices inside a bag. For each one of these we remember (approximately)
the number of active neighbors it has in the set of vertices that appear in
bags lower in the decomposition. The running time for both clique-width and
treewidth is $(\log n/\eps)^{O(w)}$.
Let us remark that, even though we only present a bi-criteria approximation
algorithm here, to the best of our knowledge there is no complexity barrier
ruling out the existence of an approximation scheme for the natural objective
of this problem (the size of the deletion set).
\subsection{Graph Balancing}
We only deal with this problem for treewidth. We assume that edge weights are
written in unary (this is the interesting case that is usually studied to avoid
unnecessary complications). For a bag of the tree decomposition, the signature
of a partial solution is the weighted out-degree each vertex of the bag has
already accumulated due to edges whose other endpoint appears lower in the
decomposition. We keep track of this information and the maximum seen so far,
making the table be a subset of $B^{w+1}$. When a vertex is forgotten (that is,
we move to the first bag that does not contain it) we go through every
orientation of the edges connecting it to the rest of the bag and update the
out-degrees of other vertices accordingly. The running time is $(\log
n/\eps)^{O(w)}$ and we can extract a solution with out-degree at most
$(1+\eps)\mathrm{OPT}$.
Let us remark that for the case of clique-width, it's not possible to achieve a
similar result. It is known that even for graphs with edge weights 1 and 2 it
is NP-hard to tell if $\mathrm{OPT}$ is 2 or 3 \cite{EbenlendrKS08}. Take an arbitrary
such instance, multiply all edge weights with $n^2$ and then replace every
non-edge with an edge of weight 1. It is now NP-hard to tell if $\mathrm{OPT}$ is at
most $2n^2+n$ or at least $3n^2$, so a better than $3/2$ approximation is
NP-hard even for cliques.
\section{Conclusions}
We have presented a generic technique which can be applied with minor
modifications to a number of hard problems. The question now becomes to which
other problems we can apply this technique. The most prominent next target is
\textsc{Hamiltonicity} parameterized by clique-width. This is another example
that separates clique-width from treewidth in term of parameterized
tractability, and its natural dynamic program uses integers. Unfortunately, the
natural program also uses subtractions, so it does not immediately yield to our
methods. In several of the problems of this paper we managed to work around
such obstacles, rewriting the natural program to only use additions. Is this
possible here, or are completely new ideas needed?
Let us also mention that in this paper we have focused on establishing that our
algorithms run in $(\log n/\eps)^{O(w)}$, but we have not been paying much
attention to the constant hidden in the exponent. Another good direction would
therefore be to improve the analysis of Approximate Addition Trees in order to
obtain better running times for our schemes. Perhaps combining this with the
idea of (partially) balancing the given decomposition could also help speed up
the overall algorithms.
\subsubsection*{Acknowledgment} I am grateful to an anonumous reviewer for
pointing out that treewidth balancing theorems can be used to obtain some of
the results of this paper in a simpler way.
\bibliographystyle{plain}
|
\section{Introduction}
The charmonium system covers a valuable energy region, where both the perturbative and non-perturbative QCD dynamics can be studied.
Potential models have been successful in describing the charmonium spectrum, particularly below the open-charm threshold (for an overview, see~\cite{charmonium_overview}).
However, in some cases experimental results differ significantly from the results of potential model calculations,
and the possible explanation of these discrepancies are non-perturbative effects, such as effects of intermediate charmed-meson loops~\cite{quark_mass_hanhart}.
The importance of these loops can be studied by the BESIII experiment.
A potential model that is based on a non-relativistic framework and for which the potential is taken in the form of Coulomb plus linear terms, predicts a $2S$ hyperfine splitting of
$M(\psi') - M(\eta_{c}') = 67.0$~MeV~\cite{eichten2004}. This is about 20~MeV larger than the experimental value of 48.5$\pm$3.3~MeV~\cite{hyperfine_bes3}.
Including coupled-channels effects into the theoretical model improves the agreement between the theory and experiment, predicting a splitting with a size
$M(\psi') - M(\eta_{c}') = 46.1$~MeV. It is tempting to conclude that the $\psi'-\eta_{c}'$ splitting reflects the influence of virtual decay channels~\cite{eichten2004}.
Further hints to the importance of intermediate charmed-meson loops (ICML) can be in radiative transitions in charmonium.
For instance, the prediction of a non-relativistic potential model for the width of the M1 radiative transition $\psi' \rightarrow \gamma \eta_{c}$ is 9.7~keV \cite{m1_theory}.
This result is nearly one order of magnitude higher than the experimental value from the CLEO-c data, which gave a width of $\Gamma(\psi' \rightarrow \gamma \eta_{c}) = 0.97 \pm 0.14$~keV~\cite{m1_cleo}.
Such a large discrepancy suggests that one cannot disregard ICML in theoretical calculations and that quenched quark models cannot be applied.
Recently, a new unquenched theoretical model based on an effective Lagrangian approach was developed~\cite{m1_theory}.
The prediction of this model is $\Gamma(\psi' \rightarrow \gamma \eta_{c}) = 2.05^{+2.65}_{-1.75}$~keV, which is in reasonable agreement with the measured value.
Moreover, quenched lattice calculations predict a value $\Gamma(\psi' \rightarrow \gamma \eta_{c}) = 0.4\pm0.8$~keV \cite{m1_lattice}
that overlaps within one standard deviation with the experimental result.
Note that the large uncertainty in theoretical calculations complicates the physics interpretation and a further tuning of the existing models is needed.
From the experimental side, an improvement of the precision of the partial widths and, therefore, more measurements of radiative transitions in the charmonium mass region are needed.
The BESIII experiment~\cite{bes3} at the BEPCII $e^{+}e^{-}$ collider in Beijing obtained the world's largest data sets
of vector charmonium states both below and above the open-charm threshold.
The precision of various resonance parameters and decay properties of charmonium states has been improved significantly by BESIII
with respect to previously published results.
Furthermore, rare charmonium transitions have been successfully observed for the first time~\cite{hc_bes3}.
In this paper, we present a fraction of the recently obtained results with BESIII, namely
systematic studies of isospin-violating transitions in the charmonium system, such as $\psi' \rightarrow \pi^{0} J/\psi$ and
$\psi' \rightarrow \pi^{0} h_{c}$.
The relevant transitions in the charmonium system are depicted in Fig.~\ref{fig:charm2013_levels}.
\begin{figure}[ht!]
\begin{center}
\includegraphics[scale=0.4]{charm2013_levels.pdf}
\caption{An overview of the studied isospin-violating transitions in the charmonium-mass region measured by BESIII.
The gray dashed arrow indicates the open-charm threshold.
The solid arrow indicates the transition $\psi'\rightarrow \pi^{0} J/\psi$.
The fine-dashed arrow shows the transition $\psi'\rightarrow \pi^{0} h_{c}$.
The dashed arrow shows the isospin-violating process $e^{+}e^{-} \rightarrow \pi^{0} J/\psi$ at a center-of-mass energy of $\sqrt s = 4010$~MeV.}
\label{fig:charm2013_levels}
\end{center}
\end{figure}
\section{Measurements of branching fractions of \\ $\psi'\rightarrow \pi^{0}J/\psi$}
The hadronic transitions $\psi'\rightarrow \pi^{0}(\eta)J/\psi$ in the charmonium system were considered to be a reliable source for the extraction of the ratio between up and down quark masses~\cite{quark_mass}.
There are two possible sources of the isospin symmetry breaking, namely the electromagnetic processes and the difference between the masses of up and down quarks.
Particularly for the isospin violating transition $\psi'\rightarrow \pi^{0}J/\psi$ (indicated by the solid arrow in Fig.~\ref{fig:charm2013_levels}), the
contribution of electromagnetic processes was shown to be negligible ~\cite{em_effects}. Thus, for a long time, there was hope to
access the bare masses of light quarks using this transition.
Based on a leading order QCD multipole expansion, the relation between the masses of the up and down quarks ($m_{u}, m_{d}$) and
the ratio of the branching fractions of charmonium transitions $R_{\pi^{0}/\eta} = \frac{B(\psi'\rightarrow \pi^{0} J/\psi)}{B(\psi' \rightarrow \eta J/\psi)}$ can be written as~\cite{quark_mass}:
\begin{equation}
R_{\pi^{0}/\eta} = 3\left(\frac{m_{d} - m_{u}}{m_{d} + m_{u}}\right)^{2} \frac{F^{2}_{\pi^{0}}}{F^{2}_{\eta}} \frac{M^{4}_{\pi^{0}}}{M^{4}_{\eta}} \left |
\frac{\overrightarrow{q}_{\pi^{0}}}{\overrightarrow{q}_{\eta}} \right | ^{3},
\label{eq:eq1}
\end{equation}
where $F_{\pi^{0}(\eta)}$ and $M_{\pi^{0}(\eta)}$ are the decay constants and masses of the $\pi^{0}$ and $\eta$ mesons, respectively,
and ${\overrightarrow{q}_{\pi^{0}(\eta)}}$ stands for the momentum of the corresponding particles produced in the transition $\psi' \rightarrow \pi^{0}(\eta)J/\psi$ in the $\psi'$ rest frame.
The up-down quark-mass ratio obtained using Eq.~\ref{eq:eq1} combined with the experimental results from the CLEO-c collaboration ~\cite{quark_mass_cleo} is
$m_{u}/m_{d} = 0.40 \pm 0.01$, which is in discrepancy with other experimental methods, for example using the masses of the lightest scalar mesons ~\cite{quark_mass_light}.
It was suggested by various theoretical groups that the main source of discrepancy is the contribution of ICML~\cite{quark_mass_hanhart}.
In Ref.~\cite{quark_mass_hanhart} it was shown that in the charmonium transitions $\psi' \rightarrow \pi^{0}(\eta) J/\psi$ the contributions of ICML are enhanced by a factor of two
compared to the tree-level contributions. Thus, the light-quark mass ratio can only be extracted from these decays after establishing a complete effective field theory up to next-to-leading order.
Recently, the BESIII collaboration performed a new measurement of the ratio of branching fractions
$R_{\pi^{0}/\eta}$, with $\pi^{0}(\eta) \rightarrow \gamma\gamma$ and the leptonic decay mode of $J/\psi$ ($J/\psi \rightarrow e^{+}e^{-}/\mu^{+}\mu^{-}$) ~\cite{pi0jpsi_bes3}.
The reconstructed two-photon invariant-mass distributions of $\pi^{0}(\eta)$ together with the background contributions are shown in Fig.~\ref{fig:mggfit}.
Clear signals from the $\pi^{0}$ and $\eta$ decays can be observed. Points with error bars represent data, the red solid lines represent total fit results, the dashed lines show the fitted background shapes and
the filled histograms are the background shapes obtained from inclusive MC simulations of $\psi'$ decays and from a side-band analysis.
The value of the ratio $R_{\pi^{0}/\eta}$ was found to be
$R_{\pi^{0}/\eta} = (3.74 \pm 0.06 \pm 0.04) \%$,
which is in good agreement with the previously reported measurements~\cite{quark_mass_cleo} and its precision is significantly improved.
The current estimate of the ratio of branching fractions
$R_{\pi^{0}/\eta}$ provided by the non-relativistic field theories is $(11 \pm 6)\%$~\cite{quark_mass_hanhart}.
This value falls within two standard deviations from the experimental result.
The theoretical value lacks precision and the existing model needs further improvement.
To this end, systematic experimental studies of isospin-violating transitions in the charmonium system are needed and can be carried out by the BESIII experiment.
Such studies will in general help to improve our knowledge concerning non-perturbative effects in a relatively simple system such as charmonium, thereby testing the dynamics of QCD.
\begin{figure}[!htb]
\begin{center}
\includegraphics[scale=0.6]{fit_all.pdf}
\caption{
Two-photon invariant-mass ($M_{\gamma\gamma}$) distributions and fit results.
(a) $\psi'\to\pi^{0} J/\psi,J/\psi\to e^{+}e^{-}$,
(b)$\psi'\to\pi^{0} J/\psi,J/\psi\to \mu^{+}\mu^{-}$,
(c)$\psi'\to\eta J/\psi,J/\psi\to e^{+}e^{-}$,
(d)$\psi'\to\eta J/\psi,J/\psi\to \mu^{+}\mu^{-}$,
where the points with error bars are data, and the red solid lines are the total fit results,
and dashed lines are the fitted background shapes.
The filled histograms present the shapes of the total background contributions obtained from a complete MC simulation
of all possible $\psi'$ decays excluding the channel of interest, and from the side-bands analysis. \label{fig:mggfit}
}
\end{center}
\end{figure}
\section{Measurements of branching fractions of \\$\psi'\rightarrow \pi^{0}h_{c}$}
Another isospin-violating transition, $\psi' \rightarrow \pi^{0} h_{c}$ (shown with the fine-dashed line in Fig.~\ref{fig:charm2013_levels})
was observed by the BESIII collaboration~\cite{hc_bes3}.
This observation provided the first measurement of the branching fraction of this transition and its value was found to be
$B(\psi'\rightarrow \pi^{0}h_{c}) = (8.4 \pm 1.3 \pm 1.0) \times 10^{-4}$, corresponding to a width of $\Gamma(\psi'\rightarrow \pi^{0} h_{c}) = 0.26 \pm 0.05$~keV.
Non-relativistic effective field theories predict that the contribution of the ICML is in the order of 10\% \cite{hc_hanhart}.
By only accounting for the tree-level contributions and using a dimensional analysis, the width of the isospin-suppressed process is estimated to be
$\Gamma(\psi' \rightarrow \pi^{0} h_{c}) = (0.9 \pm 0.6)C^{2}$~keV, where $C$ is the order of one.
The theoretical value is in reasonable agreement with the experimental data, which supports the model and the dominance of tree-level contributions.
Also this value is consistent with predictions of
single-channel calculations based on a QCD multipole expansion using the Cornell potential model, which gives a partial width of
$\Gamma(\psi' \rightarrow \pi^{0} h_{c}) = (0.4 - 1.3)$~keV~\cite{hc_potential}.
\section{Measurements of branching fractions of \\$e^{+}e^{-} \rightarrow \pi^{0} J/\psi$ at $\sqrt{s}=$~4.009~GeV}
Effects of ICML above the open-charm threshold are expected to be significant according to
existing theoretical models.
For instance, an unquenched effective Lagrangian approach based on heavy-quark symmetry and chiral symmetry~\cite{pi0jpsi4010_theory}
predicts that the isospin-violating process $e^{+}e^{-} \rightarrow \pi^{0} J/\psi$ is strongly affected by the open-charm effects and that
the cross sections of these reactions as functions of energy may provide an opportunity for revealing the significance of ICML contributions.
Authors provide calculations of $\sigma(e^{+}e^{-}\rightarrow \pi^{0}J/\psi)$ for the energy range of 3.65-4.30~GeV.
Recently the isospin-breaking process $e^{+}e^{-} \rightarrow \pi^{0} J/\psi$ at $\sqrt{s} =$~4.009~GeV (
shown as the dashed line in Fig.~\ref{fig:charm2013_levels}) was studied by BESIII~\cite{pi0jpsi4010_bes3}.
An upper limit on the $\pi^{0}J/\psi$ production cross section is set at $\sigma(e^{+}e^{-}\rightarrow \pi^{0}J/\psi) < 1.6$~pb at a 90\% confidence level.
This measurement is in agreement with the upper limit set by the CLEO experiment~\cite{pi0jpsi4010_cleo},
$\sigma(e^{+}e^{-}\rightarrow \pi^{0}J/\psi) < 10$~pb at a 90\% confidence level and
also does not contradict the theoretical prediction ~\cite{pi0jpsi4010_theory}.
The theoretical calculations are dependent on the relative phases among the resonance transition amplitudes
and give an upper limit of around $5\times 10^{-2}$~pb.
However, more systematic measurements are needed to draw more solid conclusions.
\section{Summary}
In summary, we present recent BESIII results on a study of isospin-violating transitions in the charmonium-energy region,
e.g. on the transitions $\psi' \rightarrow \pi^{0} J/\psi$, $\psi' \rightarrow \pi^{0} h_{c}$ and $e^{+}e^{-} \rightarrow \pi^{0} J/\psi$ at $\sqrt{s} =$~4.009~GeV.
These experimental studies give an important input for the construction of an effective field theory that will reveal the effects of
intermediate charmed-meson loops on transitions in charmonium system and would provide us more insight on the dynamics of non-perturbative QCD.
\singlespacing
|
\section*{Abstract}
We use a computational model to propose a physiological mechanism by which transient control of beta oscillations in the indirect pathway of the basal ganglia is orchestrated at the cellular level. Our model includes a simple and robust mechanism by which a cellular switch (from bursting to tonic) almost instantaneously translates into a functional gating switch (from blocking to conducive) in an excitatory-inhibitory network. Applied to the control of beta oscillations in the basal ganglia, the model shows the modulation of beta activity under the action of a transient depolarization, for instance a dopamine signal. The model predicts, by analogy to the thalamocortical circuit, a novel gating function by which the transfer of cortical spikes through the indirect pathway is blocked under the inhibitory drive preceding movement but briefly released at the onset of movement execution.
\section*{Significance Statement}
Connecting the dots between the molecular action of a neurotransmitter and the rapid modulation of a brain network rhythm is an important challenge of neuroscience. We propose a simple, generic, and robust mechanism by which a switch mediated by neurotransmitters at the cellular level can switch off network oscillations. We illustrate the mechanism on the modulation of beta oscillations in the basal ganglia. The model suggests a striking analogy between gating mechanisms in the basal ganglia and in the thalamocortical circuits.
\section*{Introduction}
The beta rhythm of the basal ganglia and its functional role have been the topic of increasing attention in recent years~\cite{Boraud2005, Nambu2008} primarily because of the development of deep-brain stimulation (DBS) as a treatment to alleviate the most debilitating symptoms of Parkinson's disease~\cite{Nambu2008, Hammond2007}. It is well established that the basal ganglia are composed of two parallel pathways, the direct---movements releasing---and indirect---movements inhibiting---pathways. Until two decades ago, movement disorders such as those observed in Parkinson's disease were explained by firing rate changes~\cite{Nambu2008, Albin1989}. Recently, attention has shifted to the physiological and pathological rhythmic activity in specific neuronal populations, particularly in the beta band (8 - 35Hz)~\cite{Boraud2005}. In this paper, we propose a computational model by which a cellular switch controls oscillations at the population level and modulates activity transfer gating through the indirect pathway.
The first part of the paper is purely computational and generic: we start from recent work that models a cellular switch from tonic spiking to bursting, mediated for instance by calcium channels~\cite{Franci2013a}, and explore the consequences of the cellular switch in a network composed of two interacting heterogenous populations, one excitatory and the other inhibitory. The tonic mode at the cellular level translates into an exogenous rhythm at the network level, conducive of synaptic inputs. In sharp contrast, and with no change in the synaptic connections, the bursting mode at the cellular level translates into a strongly endogenous rhythm at the network level, mostly blind to synaptic inputs.
The second part of the paper uses a large amount of recent experimental evidence to demonstrate that the basal ganglia include all the required ingredients to provide a robust instance of the proposed mechanistic link and to support the role of T-type calcium channels as key ionic mediators between the molecular action of a transient depolarization source and the modulation of beta oscillations in the basal ganglia. At the cellular level, experimental evidence includes the prominence of T-type calcium channels both in the subthalamic nucleus (STN) and in a fraction of the globus pallidus pars externa (GPe)~\cite{Hallworth2003, Nambu1994, Cooper2000}, electrophysiological recordings of a cellular switch in neurons of the STN induced by hyperpolarization~\cite{Beurrier1999}, and the role of dopamine receptors in the striatum~\cite{Weinberger2011}. At the network level, experimental evidence includes the prominent role of the GPe-STN network in local field potential (LFP) oscillations~\cite{Plenz1999, Tachibana2011}, the modulation of beta oscillations in voluntary movements~\cite{Weinberger2011, Levy2002a, Brown2005, Sharott2005, Mallet2008a}, the link between dopamine and movements~\cite{Jin2010, Jenkinson2011}, and a number of experimentally reported effects of DBS in alleviating the abnormal oscillations~\cite{Hammond2007}.
We are not aware of an earlier computational model connecting the dots between the causal effects of depolarization, T-type calcium channels, cellular bursting, and beta oscillations. The model differs from previously published models of beta oscillations most radically in that a common mechanism simultaneously explains the transient physiological oscillations associated to movement control and the persistent pathological parkinsonian oscillations without any required change in the synaptic connectivity. Emphasis on the specific role of T-type calcium channels is scarce in the literature on network oscillations~\cite{Astori2011}, with the notable exception of the recent study~\cite{Tai2011} that clearly establishes their role in the basal ganglia oscillations. At a higher level, by analogy to thalamocortical circuits, the model suggests a gating of motor activity by basal ganglia circuits. The gating of sensory information has been well studied in the thalamus~\cite{McCormick1990, McCormick1997}.
\section*{Results}
\subsection*{A general mechanism for cellular control of LFP oscillations.}
Our computational model illustrates how a cellular switch in an excitatory-inhibitory (E-I) network robustly induces LFP oscillations at a frequency that can be markedly different from the isolated cellular rhythm (Fig.~\ref{fig:switch_all}).
\textbf{Cellular switch:} The cellular switch (Fig.~\ref{fig:switch_all}a, upper traces) has been the topic of recent modeling work by the authors~\cite{Franci2013a, Drion2012, Franci2012, Franci2013}. Many neuronal models switch from tonic spiking to bursting as the balance of ionic currents in the slow time scale of repolarization evolves from negative feedback (slow restorativity) to positive feedback (slow regenerativity) at the resting potential~\cite{Franci2013}. Slow regenerative channels---most prominently, calcium channels---provide the neuron with short-term memory\cite{Marder1996a}, a key source of the hysteretic nature of bursting firing patterns~\cite{Franci2013a}. The balance can be dynamically regulated in many different ways but one prominent mechanism relevant for this paper is a transient deinactivation of T-type calcium channels by hyperpolarization~\cite{Drion2012}. A reduced low-dimensional model is sufficient to capture the cellular switch in a population computational model (see methods).
\textbf{Network switch:} In an E-I network, the cellular switch triggers a change at the population level. Fig.~\ref{fig:switch_all}a, middle traces, illustrates the rhythmic switch in a cartoon two-neuron network composed of an inhibitory neuron (I~neuron) and an excitatory neuron (E~neuron), connected reciprocally: the I~neuron is connected to the E~neuron via a GABAergic connection and the E~neuron is connected to the I~neuron with a glutamatergic connection (see methods).
In isolation, the two neurons have distinct intra and interburst frequencies and generate distinct rhythms (Fig.~\ref{fig:switch_all}a, upper traces). In the network configuration, the cellular switch triggers a clear switch in the network. In their tonic mode, the two neurons in the network barely affect each other because the effect of a single spike on the other neuron is weak(Fig.~\ref{fig:twoeffects}a). The two neurons, separately or connected, have almost the same firing pattern (Fig.~\ref{fig:switch_all}a, upper and middle traces, Tonic mode). The two neurons fire asynchronously. In sharp contrast, the influence of one neuron on the other becomes strong in their bursting mode. The network bursting is not a simple reflection of the unicellular bursting, but a combination of two concomitant effects: the endogenous effect and the network effect (Fig.~\ref{fig:twoeffects}b)\footnote{This is in contrast to previous work such as~\cite{Frohlich2006}.}.
\begin{figure}[t]
\centerline{\includegraphics[width=8.7cm] {switch_all4.pdf}}
\caption{Regenerative excitability at the cell level robustly induces LFP oscillations at the network level. The transitions between modes are controlled by the applied current (see methods). \textbf{a}:~Cell level. Upper traces - single cells: excitability switch in two neurons with distinct intrinsic properties. Middles traces - two-neuron E-I network: in the burst mode, the two neurons synchronize and generate a network rhythm. In the tonic mode, the two neurons do not synchronize. Lower traces - E-I populations (spike rasters): the neurons fire in synchrony in the burst mode and in a decorrelated manner in the tonic mode. \textbf{b}:~Population level. First rows: LFP time-frequency plots. Oscillations at the 30 Hz frequency band develop in the burst mode but no clear oscillatory activity develops in the tonic mode. Second rows: cross-correlations for all the possible two-neuron combinations in the different modes. The burst patterns are correlated in the burst mode but not in the tonic mode. The color code (from blue to red) indicates the spectral power (from low to high).\label{fig:switch_all}}
\end{figure}
As illustrated in Fig.~\ref{fig:twoeffects}, the novel network rhythm depends both on the cellular switch and on the E-I interconnection. Intrinsic properties of each neuron, not interconnection, determine the onset (through E~neuron) and the termination (through I~neuron) of the network oscillations. But the interconnection influences the duration of the network oscillation: the E burst initiates the I burst through its excitatory drive while the inhibitory drive of the I~neuron terminates the E burst earlier than in isolation. The result is that the cellular switch transmits to the network its exogenous (tonic) or endogenous (bursting) nature, but that the network interconnection regulates the frequency of the endogenous network rhythm, which can markedly differ from the frequency of the bursting neurons. It should also be emphasized that the mechanism is robust to many variants: the I burster does not need to burst in isolation but could instead be `burst excitable' (Fig.~S1b), that is, capable of bursting in response to an excitatory post-synaptic potential (EPSP) and of terminating its burst endogenously. Excitatory synapses are robust to the synapse type (see methods). Finally, the cellular control of the network switch can be through the E or I~neuron only rather than through both. In short, the network configuration does increase the robustness of the network rhythm compared to the cellular rhythm in isolation against variability of the intrinsic properties of the neuron.
\textbf{Population switch:} Fig.~\ref{fig:switch_all}a, lower traces, and Fig.~\ref{fig:switch_all}b illustrate how the network rhythm translates into LFP oscillations at the population level. Raster plots illustrate that the population switches from asynchronous behavior in tonic mode to synchronous behavior in bursting mode. The cellular switch is therefore amplified at the population level because only the endogenous bursting mode can synchronize a heterogenous population. Cross-correlograms and time-frequency plots confirm a high spectral power at 30~Hz confined to the bursting periods of the neurons.
We simulated a network with sixteen neurons in each population, random connections with gaussian white noise (each E~neuron is connected to a fixed number of random I~neurons and vice versa), and heterogenous neurons (intrinsic parameters vary in a fixed range). With identical neurons, the connection ratio between the two nuclei could be as low as $1/16$ (each neuron is connected to a single neuron in the other population which is composed of sixteen neurons) and the two populations would still synchronize and generate oscillations. With heterogenous neurons (10\% variability range in this case), the connection ratio could be as low as $1/4$ (each neuron is connected to four neurons in the other population) and the populations still synchronize. Remarkably, the parameter variability has a strong impact on the unicellular rhythm (Fig.~\ref{fig:twoeffects}b) but not on the network interactions. Therefore, a high variability in the burst onset of E~neurons is compatible with the network rhythm. The random connections, on the contrary, act on the network effects and on the burst onsets of the I~neurons and on the burst termination of the E~neurons. The cellular variability is averaged due to the presence of multiple recurrent connections. The burst duration in I~neurons is also function of the intrinsic parameters and is therefore affected by the parameters variability among neurons. Here also both the cellular switch and the network interconnection contribute to the robustness of the collective phenomenon. Many variants of the mechanism are possible depending on the network connectivity within each population. The switch control can be achieved through a fraction of the populations only provided that the heterogeneity in the cellular switch is compensated for by a sufficient level of connectivity within the population. Our experience with the computational model suggests a versatile mechanism whose robustness is only increased in larger populations.
\begin{figure}[t]
\centerline{\includegraphics[width=8.7cm]{twoeffects.pdf}}
\caption{The network rhythm results from the combination of the cellular switch and the network coupling. The GABAergic and the glutamatergic currents are represented in dark and light grey, respectively. \textbf{a}:~In the tonic mode, there is no network rhythm because the synaptic current of a single spike is not strong enough to initiate a spike in the other neuron. \textbf{b}:~In the burst mode, the network rhythm is due to the interplay of two effects: endogenous effect due to the intrinsic properties of the neurons and network effect due to the synaptic connections.\label{fig:twoeffects}}
\end{figure}
\textbf{Functional switch:} A plausible function of the population switch is to gate activity transfer of incoming spikes through the E (or I) population~\cite{McCormick1990, McCormick1997, Sherman2001}. Fig.~\ref{fig:infotransfer}a illustrates that the endogenous (burst) rhythm of the E population carries little information about an input spike train. This is in sharp contrast with the exogenous (tonic) rhythm, where the individual pattern of action potentials is an almost faithful representation of incoming EPSPs or depolarizing inputs (Fig.~\ref{fig:infotransfer}b). The action potential pattern is function of the intensity, duration, and frequency of input spikes~\cite{McCormick1990}. In tonic firing, larger EPSPs elicit higher firing rates, in contrast to the burst mode~\cite{Sherman2001}. The tonic mode is therefore called a `transfer' (or exogenous) mode and the burst mode an `oscillatory' mode~\cite{McCormick1990}.
\subsection*{Hyperpolarization controls oscillations in the GPe-STN network.}\
\textbf{Cellular switch:} STN neurons have been shown to switch from tonic firing to bursting in isolation~\cite{Hallworth2003}. The switch is controlled by the transient deinactivation of low threshold T-type calcium currents, a slow regenerative ion channel. T-type calcium channels are inactivated in the tonic mode. However, when the neuron is hyperpolarized, the channels deinactivate and act as a strong source of regenerativity, switching the neuron to burst. Fig.~S1a reproduces, in our reduced model, a switching pattern that has been observed \textit{in vitro} (for instance, see Fig.~4~\cite{Beurrier1999}). The polarization of the membrane sets the level of inactivation of calcium channels. Some neurons in the GPe also exhibit two distinct modes of firing~\cite{Nambu1994, Cooper2000}. Although no experimental recording has shown the switch in isolation, many GPe neurons do possess the regenerative low threshold T-type calcium current~\cite{Nambu1994, Cooper2000}, and the switch can therefore be activated in models of GPe neurons as in STN neurons by modulating the balance of restorative and regenerative slow channels through hyperpolarization.
\textbf{Network switch:} Oscillations have been recorded in the GPe-STN network~\cite{Plenz1999, Tachibana2011}. The most striking evidence comes from the experimental study of~\cite{Plenz1999} showing that GPe and STN are able to generate synchronized oscillatory bursting activity \textit{in vitro}. It has also been shown that reciprocal GPe-STN interconnections, on top of glutamatergic inputs to the STN, are primordial for the generation and amplification of the oscillatory activity of STN neurons~\cite{Tachibana2011}. These considerations led to the STN/GPe pacemaker hypothesis~\cite{Bevan2002} which was implemented in numerous models~\cite{Gillies2002, Terman2002, Humphries2006, NevadoHolgado2010, Pavlides2012, Tsirogiannis2010, Kumar2011} to generate beta oscillations in the basal ganglia. The GPe-STN network fits the general mechanism of the previous section. The GPe is the I~population whereas the STN is the E~population. In this network model, the oscillatory state is controlled by the tonic input on the GPe neuron. If the GPe neuron is depolarized, both neurons are in the tonic mode and no oscillatory activity develops at the network level. By contrast, if the GPe neuron is hyperpolarized, it switches to the burst firing mode. The increased instantaneous firing rate during the burst discharge, greater than during the tonic mode (even though the global firing rate might be lower or equal to the one of the tonic mode) sets the STN neuron to an hyperpolarized level and the STN neuron crosses the excitability switch. The network oscillations start.
\textbf{Population switch:} LFP oscillations have been recorded in the GPe-STN network~\cite{Mallet2008a, Goldberg2004}. In normal conditions, single-unit recordings display no oscillatory behavior and LFP cross-correlations show no correlated behavior, just as in the tonic mode in our model. In animal models of Parkinson's disease, the single-unit recordings reveal strong beta activity and the cross-correlations exhibit strong periodic correlated behaviors, similar to the burst mode.
\subsection*{Transient modulation of basal ganglia beta oscillations.}
The basal ganglia is a set of subcortical nuclei involved in the generation of voluntary movements. Our model targets the behavior of three of the nuclei of the indirect pathway: the striatum, the GPe, and the STN. We study the influence of dopamine modulation on the beta rhythm through the striatal D2-type dopamine receptor~\cite{Weinberger2011}.
Our basal ganglia network simulates the dopamine modulation of the strength and coherence of the beta oscillations (see discussion). A transient increase in the nigrostriatal dopamine level from neurons in the substantia nigra pars compacta (SNc), or `kick of dopamine', triggers the decrease of the beta-band activity coherence and power~\cite{Jenkinson2011}. The unicellular switch mechanism described above is a perfect candidate due to the rapid time scales in play in movement onset (on the order of hundreds of milliseconds)~\cite{McCarthy2012, Canolty2012}. The simulated network is composed of four neurons in the striatum (only the D2-type dopamine receptors are considered) and sixteen neurons in each of the GPe and STN nucleus. The connections are as in Fig.~\ref{fig:kickDA}a. The inputs are the nigrostriatal dopaminergic level which modulates the firing rate of the striatal neurons, and cortical inputs to STN neurons. Results of the simulation show the expected behaviors (Fig.~\ref{fig:kickDA}): a kick of dopamine (Fig.~\ref{fig:kickDA}b) triggers a striatum-mediated switch in the neuronal firing rate in GPe and STN neurons (Fig.~\ref{fig:kickDA}, c, d, and e) which induces an event-related desynchronization (ERD) (Fig.~\ref{fig:kickDA}, f and g). After the movement onset, the dopamine goes back to its basal level and the striatum recovers its basal firing frequency. GPe and STN neurons undergo a reversal switch and start bursting in synchrony, generating strong network oscillations.
\begin{figure*}
\begin{center}
\centerline{\includegraphics[width=17.8cm]{kickDA.pdf}}
\caption{The indirect pathway of the basal ganglia and the dopamine control of the cellular and population behaviors in movement-related switches. \textbf{a}:~Striatal neurons with D2-type dopamine receptors, modulated by the dopamine from the SNc, make indirect contact with the output nuclei (the GPi) via relays (the GPe and STN). The output of the basal ganglia is a selectivity signal which acts on the thalamocortical activity. Blue, red, and green arrows represent dopaminergic modulatory, GABAergic inhibitory, and glutamatergic excitatory connections, respectively. SNc-substantia nigra pars compacta, GPe-globus pallidus pars externa, STN-subthalamic nucleus, GPi-globus pallidus pars interna. \textbf{b}:~Nigrostriatal kick of dopamine. \textbf{c}, \textbf{d}, and \textbf{e}:~Membrane potential of a single neuron out of the population in the striatum, GPe, and STN, respectively. The kick of dopamine decreases the firing rate of the striatal neurons which switches the GPe and the STN neurons to tonic firing. \textbf{f} and \textbf{g}:~Time-frequency plots for the GPe and STN populations, respectively. Clear beta activities exist in the GPe and STN prior to the kick of dopamine. The kick triggers an ERD before resuming to the basal activity in the beta band. The color code (from blue to red) indicates the spectral power (from low to high).\label{fig:kickDA}}
\end{center}
\end{figure*}
\subsection*{Network oscillations gate cortical activity transfer.}
The two oscillation states are opposite in their transfer of incoming spikes. Fig.~\ref{fig:infotransfer}, c and d display cross-correlations between the STN membrane potentials and cortical input spikes. The cross-correlations before the dopamine increase, when the neurons are in the burst `oscillatory' mode, possess a low maximum with no sharp transition. Input spikes from the cortex are not transmitted directly to the output firing patterns which is rather function of the internal variables. The incoming spikes might have a small impact on the interburst duration which explains the smooth increase in the cross-correlations. In sharp contrast, during the dopamine kick, when the neurons are in the tonic `transfer' mode, the cross-correlations display a sharp increase around 0 and reach a maximum of high value. Incoming spikes are transmitted to the output firing patterns, generating output spikes in STN neurons after a time delay corresponding to the synaptic transmission.
\begin{figure}
\centerline{\includegraphics[width=.45\textwidth]{infotransfer2.pdf}}
\caption{The two firing modes are opposite in the processing of incoming spikes. \textbf{a}~and~\textbf{b}:~A neuron membrane potential in the burst `oscillatory' filtering mode and in the tonic `transfer' mode with incoming spikes. \textbf{c}~and~\textbf{d}:~Cross-correlations (xcorr) between the membrane potential of the STN neurons and the cortical inputs. \textbf{c}:~Cross-correlations in the oscillatory network. The correlation is weak and cortical spikes are filtered. \textbf{d}:~Cross-correlations in the non-oscillatory network. The correlation is high and cortical spikes are transmitted.\label{fig:infotransfer}}
\end{figure}
\section*{Discussion}
\subsection*{Exogenous inputs can switch the network activity from exogenous to endogenous.}
An E-I network of neurons with sufficient expression of T-type calcium channels has the property of robustly switching from an endogenous oscillatory mode to an exogenous tonic mode under the modulation of its external drive. Hyperpolarization of the neurons deinactivates the T-type calcium channels, causing the neurons to burst and the network to oscillate, whereas depolarization inactivates the channels, causing the neurons to fire tonically and the network to stop oscillating. Both the neuron intrinsic excitability and the network interconnections participate in making the network rhythm robust and not a mere image of the isolated neuronal rhythm.
The switch is dynamic because it is controlled at the cellular level. It is robust because the E-I network oscillation is a collective synchronous phenomenon allowing for high variability at the cellular level. Because the cellular switch can be efficiently modeled in reduced neural models~\cite{Franci2013a, Drion2012, Franci2012, Franci2013}, the model is amenable to realistic simulations in large populations. The mechanism of hyperpolarization-induced bursting has been described in single-neuron electrophysiology, but we are not aware of earlier research investigating how the cellular mechanism translates to a network effect in E-I networks.
\subsection*{Beta oscillations in the GPe-STN network can be transiently switched off.}
The structure of the indirect pathway (Fig.~\ref{fig:kickDA}a) of the basal ganglia suggests that the described oscillation mechanism is consistent with the generation of beta activity in the GPe-STN network. External inputs can modulate the exogenous drive and switch the network state. For instance, an increase of cortico-striatal activity corresponds to a gain increase in the inhibitory drive from the striatum to the GPe, which might set beta activity in the GPe-STN network. However, any transient depolarization source can instantaneously switch off the beta activity. The depolarization sources are multiple, dopamine being one of them, through D2-receptors in the striatum for instance. Excitatory inputs on the STN could also be a source of depolarization.
Dopamine has been shown to correlate with the power and coherence of the beta oscillations in the basal ganglia and the cortex (in animals~\cite{Sharott2005, Mallet2008a} and in humans~\cite{Weinberger2011, Levy2002a, Brown2005}). A dopamine depletion or disrupted transmission correlates with an increase in beta oscillation power in the STN~\cite{Sharott2005, Mallet2008a}. On the opposite, dopamine therapies are associated with a decrease in the beta oscillation power and dopamine is known to suppress beta oscillations in the basal ganglia~\cite{Levy2002a, Brown2005}.
A number of experimental data are consistent with the model mechanism. At rest, the level of nigrostriatal dopamine is almost constant but differs from zero (Fig.~2e of~\cite{Jin2010}). An increase in cortico-striatal activity can set the GPe-STN network in its oscillatory state. When a kick of dopamine is produced by nigrostriatal dopaminergic neurons (Fig.~2e of~\cite{Jin2010}), the firing rate of the striatal neurons decrease drastically (Fig.~2b of~\cite{Jin2010}). The tonic inhibition on the GPe neurons is decreased and allows the neurons to switch from their burst mode to their tonic firing mode. The same phenomenon takes place in the STN neurons. GPe and STN neurons fire asynchronously and no global oscillatory activity is detected at the network level~\cite{Weinberger2011, Levy2002a, Brown2005, Engel2010, Kuhn2006, Kuhn2004}. After a few hundreds of milliseconds, the kick of dopamine ends and the GPe and STN neurons resume to their initial bursting state.
\subsection*{Relationship to previous basal ganglia models.}
The combination of a unicellular switch in individual neurons and its emulation at the population level in an E-I network is a central and novel feature of the proposed model, which, to our knowledge, is the first to investigate the dynamic modulation of beta oscillations in physiological conditions. Several earlier models have investigated GPe-STN oscillations in the pathological framework of Parkinson's disease~\cite{Gillies2002, NevadoHolgado2010, Pavlides2012, Tsirogiannis2010, Terman2002, Humphries2006, Kumar2011}. In firing-rate models~\cite{Gillies2002, NevadoHolgado2010, Pavlides2012, Tsirogiannis2010}, the modulation of oscillatory activity in an E-I network involves changes in the connectivity strength. Such changes cannot be the result of synaptic plasticity in the rapid modulation of beta activity but they could be regarded as a mean-field abstraction of the cellular modulation described here. Spiking neural models have emerged in the recent years to model network oscillatory activity~\cite{Terman2002, Humphries2006, Kumar2011}. In these models, individual neurons are usually of the leaky integrate-and-fire type, which prevents to capture an excitability switch at the neuronal level. As a consequence, network oscillations in these models are also modulated indirectly via a change of connectivity. The closest such effort to our model is described in~\cite{Kumar2011}, where modulation of dopamine levels triggers network oscillation in a GPe-STN network.
\subsection*{Transient beta oscillations and relation to movements.}
The basal ganglia play a major role in the planning and initiation of voluntary movements. In particular, the coherence and strength of the beta oscillations in the basal ganglia are associated with voluntary movements initiation, both in animals~\cite{Engel2010, Leventhal2012} and humans~\cite{Levy2002a, Brown2005, Engel2010, Kuhn2006, Amirnovin2004, Kuhn2004}. Experimental studies show a reduction or abolishment of oscillatory activity at the single-cell level during movements~\cite{Amirnovin2004}. At the population level, an ERD of oscillatory beta activity in the STN region is observed at the movement initiation~\cite{Levy2002a, Kuhn2006}. Additionally, the ERD onset latency strongly correlates with the mean reaction time in reaction-time tasks~\cite{Kuhn2004}. After the movement onset, an event-related synchronization (ERS) is observed in the beta band~\cite{Kuhn2006}.
Initiation of voluntary movements is also linked to an increase in dopamine and, in particular, to a transient increase in the activity of nigrostriatal circuits in response to cues in reaction-time tasks or prior to voluntary movements~\cite{Nambu2008, Korchounov2010, Jin2010, Jenkinson2011}. The increase in the firing rate of SNc dopaminergic neurons lasts on the order of hundreds of milliseconds and happens prior to the movement execution (Fig.~2e of~\cite{Jin2010},~\cite{Jenkinson2011}). Striatal neurons with D2-type dopamine receptors follow the inverse trend and decrease their firing rate for a few hundreds of milliseconds, covering the movement duration (Fig.~2b of~\cite{Jin2010}).
\subsection*{Pathological oscillations.}
The hallmark of Parkinson's disease is a dopaminergic denervation of the striatum, input stage of the basal ganglia, altering information patterns along movement-related ganglia-mediated pathways in the brain. Severe motor symptoms result~\cite{Hammond2007, Weinberger2009}. In Parkinson's disease, the baseline level of synchrony within and between basal ganglia nuclei in the beta band is elevated: increased beta-band activity has been reported in the STN, GPe, and GPi nuclei in both single-unit activity and LFPs~\cite{Boraud2005}. Similarly, animal models of the disease have shown increases in basal ganglia structures in burst discharges, oscillatory firing, and synchronous firing patterns~\cite{Rubin2012}. The excessive synchrony in the beta frequency band correlates with the motor symptoms and is believed to limit the information coding of neurons, the neurons being locked to the population rhythm~\cite{Hammond2007, Brown2005}. Beta synchrony diminution under dopaminergic therapies, ablative surgeries, or during DBS is associated with an amelioration of the motor impairments~\cite{Hammond2007, Weinberger2009}. It has been suggested that the degree of suppression of beta oscillations in the STN by dopaminergic medications can predict the level of improvement in bradykinesia and rigidity~\cite{Weinberger2009}.
Parkinson's disease corresponds to a pathological depleted dopaminergic state. Normally, at the onset and during movement, the synchronization in basal ganglia structures is drastically reduced. But in parkinsonian conditions, baseline levels of synchrony are higher and more resistant to suppression~\cite{Hammond2007}. In this configuration, the transient depolarization from the dopamine kick is not sufficient to trigger a network switch and strong beta oscillations remain~\cite{Jenkinson2011}, impairing cortical spikes transfer through the basal ganglia and motor movements. Our model is fully consistent with these observations and offers further insights: a decrease in the tonic nigrostriatal dopamine level projecting on neurons in the striatum results in a higher striatal firing rate. GPe neurons have a higher tonic inhibition and are maintained in the burst mode. The same observation holds for neurons in the STN. A kick of dopamine is not sufficient to bring GPe neurons to a depolarized state and to a tonic firing mode. The burst filtering mode is maintained and cortical activity cannot be processed faithfully through the basal ganglia. Prolonged synchronization among neurons in the basal ganglia, particularly between the GPe and STN, can generate a change in connection strength via synaptic plasticity, stabilizing the pathological dynamics in the network, which coincides with experimental observations~\cite{Albin1989}.
Experimental evidences show that T-type calcium channels, as employed in our model, are strongly involved in the disease. Indeed, pharmacological blockade of T-type calcium channels decreases burst activity in STN neurons both \textit{in vitro} and \textit{in vivo} and reduces motor deficits in animal models~\cite{Tai2011, Xiang2011}. In addition, low-frequency DBS, with long depolarization pulses, improve motor impairments in parkinsonian rats presumably by inactivating T-type channels~\cite{Tai2011}. In our model, this corresponds to bringing STN neurons to a depolarized level and tonic firing mode.
\subsection*{Analogy to thalamocortical oscillations and gating functions.}
The oscillation mechanism proposed in the present paper can also be relevant to network oscillations in the thalamocortical circuit, in particular in the thalamocortical relay-thalamic reticular network. At the individual level, the cellular switch has been described in both types of cell and is also controlled through deinactivation of T-type calcium channels~\cite{McCormick1997}. At the network level, the two populations switch from a tonic exogenous mode to an oscillatory endogenous mode and the resulting network oscillations have been proposed as the neural correlates of sleep-spindles~\cite{McCormick1997}. In contrast to our approach, however, existing computational models of the thalamocortical network~\cite{Destexhe1993a, Timofeev2005, Wang1995} do not provide a mechanism to explain the population switch. Instead, they focus on the oscillation mechanism solely in the oscillating state. In thalamic cells, the activation loss comes from a series of neuromodulators (acetylcholine, norepinephrine, and serotonin) involved in slow-wave sleep and anesthesia~\cite{McCormick1997, Llinas2006}.
The gating function of those thalamocortical network oscillations has been studied throughly~\cite{McCormick1990, McCormick1997}. The analogy suggests a similar gating function in the basal ganglia, consistent with the modulation of beta activity during movement preparation and execution. The oscillatory network state functions as an open or closed gate towards the transfer of incoming cortical inputs through the indirect pathway. The two opposite activity processing states give a functional significance to the ERD in the basal ganglia prior to voluntary movements. During movement preparation, the power of the beta band is high and the basal ganglia-thalamocortical circuit is unresponsive to motor stimuli. The indirect pathway provides a strong inhibitory drive to the thalamus as a result of the inhibitory synchronized bursts and cortical spikes are not transferred through the indirect pathway. Right before movement onset, the sharp decrease in the beta band sets the basal ganglia-thalamocortical circuit in a responsive state, both suppressing the inhibitory drive of the indirect pathway and allowing for faithful transfer of cortical spikes. The potential functional role of beta-band oscillations could therefore be to maintain the current sensorimotor state during movement preparation~\cite{Engel2010}, until dopamine signals the onset of movement execution~\cite{Jenkinson2011}. It will be of interest to further investigate the generality of the proposed mechanism in gating population switches.
\section*{Materials}
All the numerical simulations and analyses were performed with MATLAB, MathWorks. More details on the model and analyses are given in the \textit{SI}.
\subsection*{Computational model} The neuron model used in this paper is a modified version of the model described previously~\cite{Franci2013a, Drion2012, Franci2012, Franci2013}. The effect of T-type calcium channels on the neuron excitability is accounted for by a voltage dependence of the bifurcation parameter, $w_0$, which reflects the ultraslow deinactivation of T-type calcium channels at low threshold (see \textit{SI}). A gaussian white noise is added to the system (see \textit{SI}). The synaptic connections are made with exponential synapses, the GABAergic connections being of the GABA$_{\textrm{A}}$ type, the glutamatergic of the AMPA type (see \textit{SI}) for the E-I network, and of the NMDA type for the basal ganglia network(see \textit{SI}). The generic E-I network shows robustness toward the glutamatergic synapse type (AMPA or NMDA) provided that the excitatory current ends before the burst onset of the I neuron.
The variability of the intrinsic parameters concerns the slow time-scale, $\epsilon$, the ultraslow time-scale, $\epsilon_z$, and the ultraslow equivalent gain, $k_z$. Parameters were drawn from a uniform distribution covering a 10\% range around the initial parameter value. The robustness of the switch control was tested by varying the value of $w_{0,min}$.
In the generic E-I network (Fig.~\ref{fig:switch_all}), the cellular and network switches are controlled by the applied current, $I_{app}$, in both populations, which controls the value of $w_0$ by depolarizing or hyperpolarizing the neuron.
In the basal ganglia network, the GPe-STN connections are made in an all-to-all manner. The synaptic current, $I_{syn}$, results from the synaptic connections and controls the value of $w_0$. The striatal neuron model includes the effect of dopamine on D2-receptors via a scaling coefficient for the striatal applied current (see \textit{SI}). Striatal neurons connect to GPe neurons in an all-to-all manner with GABA$_{\textrm{A}}$ synapses. Cortical inputs on the STN neurons are spikes with spike intervals selected from a Poisson distribution, connected with AMPA synaptic connections (see \textit{SI}). We emphasize that our objective in the present paper is not a fine tuned quantitative modeling of the neuron firing pattern. Rather, we attempt to provide a qualitative picture of the switch behavior of neurons and how this switch behavior impacts neural activity processing properties. For this reason, the firing rates, burst durations, and interburst durations are not fitted to the ones observed experimentally.
\subsection*{Analyses} LFPs dynamics are modeled by the low-pass filtered (<100Hz) normalized sum of the collective synaptic activities (see \textit{SI}). The time-frequency plots result from the logarithmic representation of the LFP spectrogram obtained via a short-time Fourier transform. The cross-correlations are obtained from the MATLAB cross-correlation function estimate, for all the combination of two neurons' membrane potential (Fig.~\ref{fig:switch_all}) or the input cortical spike train and all the STN neurons' membrane potential (Fig.~\ref{fig:infotransfer}).
\section*{Acknowledgments}
The authors would like to thank T. Boraud, P. Maquet, and V. Seutin for their helpful comments. This paper presents research results of the Belgian Network DYSCO (Dynamical Systems, Control, and Optimization), funded by the Interuniversity Attraction Poles Programme, initiated by the Belgian State, Science Policy Office. The scientific responsibility rests with its authors. J. Dethier is supported by the F.R.S.-FNRS (Belgian Fund for Scientific Research).
|
\section{Introduction}
The Main Injector Particle Production (MIPP) experiment is a fixed target hadron production experiment at Fermilab which was operated from December 2004 to February 2006. It used a primary beam of 120 GeV/c protons from the Main Injector and these protons impinged on a copper target to produce secondary beams of charged pions, kaons, protons and anti-protons from 5 to 90 GeV/c~\cite{proc}. The experiment was designed to measure the total charged particle production of $\pi^{\pm}$, K$^{\pm}$, p and $\bar{p}$ off various nuclei including liquid hydrogen (LH$_{2}$), NuMI target and thin targets of beryllium, carbon, bismuth and uranium. \\
The MIPP is a high acceptance spectrometer and it provides excellent charged particle identification using Time Projection Chamber (TPC), Time of Flight (ToF), multicell Cherenkov (CKOV), Ring Imaging Cherenkov (RICH) detectors, and Calorimeter for neutrons. The lay-out of the experiment is shown in Figure 1.
\section{Motivation}
Monte Carlo (MC) simulation programs such as Geant4, MARS, Fluka, etc. model hadronic interactions based on available data. MIPP is a high acceptance spectrometer and has high statistics data with 6 beam species. These data could be used to improve the hadronic shower simulations.
\section{Inelastic cross section measurements}
Inelastic cross section measurements have been done for 58 and 85 GeV/c proton interactions with LH$_{2}$ target, and 58 and 120 GeV/c proton interactions with carbon target. The LH$_{2}$ target is 14 cm long having 1.5\% interaction length and carbon target is in the form of a disc having 1 cm thickness and 2 inch diameter, and 2\% interaction length. Good events are selected using a cut on transverse positions of incident beam to select good beam tracks, primary vertex selection, rejection of elastics (if any) etc. The interactions are selected using interacton trigger. MIPP used a scintillator-based interaction trigger which requires at least 3 charged tracks for the scintillator to fire. This causes inefficiencies at the low multiplicities and these have to be corrected for. The formula used for cross section calculation is given below:
\begin{figure}
\begin{center}
\includegraphics[width=10cm,height=6cm]{mippspecpic.eps}
\caption{The lay-out of the MIPP spectrometer.}
\end{center}
\end{figure}
\begin{center}
$\sigma$ = $\frac{N_{int} \times 10000}{N_{beam} \times n_{t} \times \epsilon}$ mb, n$_{t}$ = $\frac{N_{A} \times density \times thickness}{Atomic weight}$ cm$^{-2}$
\end{center}
where N$_{int}$ is the number of interactions, N$_{beam}$ is the number of beam particles, n$_{t}$ is the number of target particles per cm$^{2}$ and $\epsilon$ is the product of efficiencies which include the trigger efficiency, cut efficiency and acceptance. The MIPP data are compared with the DPMJET and FLUKA Monte Carlos, and previous measurements for LH$_{2}$ and carbon target in Tables 1 and 2 respectively. The numbers given here are after applying all the corrections which are calculated from the MC. Both the statistical and systematic uncertainties are given. The statistical uncertainty is calculated based on the number of events, and for the systematic uncertainty, we have taken into account the contributions from the beam flux and corrections applied. The systematics from the beam flux is the dominated one, and is $\sim$5\%.\\
\begin{table}[t]
\small
\begin{center}
\scalebox{0.9}{
\begin{tabular}{l|ccc}
Energy (GeV) & PDG (mb) & DPMJET (mb) & MIPP (mb) \\ \hline
58 & 31.13$\pm0.42\rm(stat+syst)$ & 31.6 & 30.33$\pm0.92\rm(stat)^{+2.39}_{-2.48}$(syst) \\
85 & 31.42$\pm0.52\rm(stat+syst)$ & 31.8 & 30.65$\pm0.53\rm(stat)^{+2.42}_{-2.52}$(syst) \\ \hline
\end{tabular}}
\caption{Comparison of the MIPP data with the MC predictions and PDG for p-H interactions at 58 and 85 GeV.}
\end{center}
\end{table}
\begin{table}[t]
\small
\begin{center}
\scalebox{0.9}{
\begin{tabular}{l|ccc}
Energy (GeV) & Previous measurements (mb) & FLUKA (mb) & MIPP (mb) \\ \hline
58 & 252$\pm4.73\rm(stat+syst)$~\cite{publ1} & 239.1 & 267.6$\pm12.5\rm(stat)^{+17.6}_{-18.6}$(syst) \\
& 222$\pm7\rm(stat+syst)$~\cite{publ2} & & \\
120 & & 240.2 & 264.6$\pm2.94\rm(stat)^{+17.5}_{-18.5}$(syst) \\ \hline
\end{tabular}}
\caption{Comparison of the MIPP data with the MC predictions and previous measurements for p-C interactions at 58 and 120 GeV.}
\end{center}
\end{table}
The comparison is also shown in the form of graph in Figures 2a and 2b for LH$_{2}$ and carbon target respectively. The hydrogen data at both energies are consistent, within error bars, with the PDG and DPMJET. The carbon data at 58 GeV is consistent, within error bars, with the measurement of S. P. Denisov et al.~\cite{publ1} and $\sim$20\% higher than the measurement of A. S. Carroll et al.~\cite{publ2}. FLUKA is $\sim$11\% lower than the 58 GeV/c p-C data. There is no previous measurement for 120 GeV/c p-C interactions. FLUKA is $\sim$9\% lower than the 120 GeV/c p-C data.
\begin{figure*}[ht]
\begin{center}
\subfigure[]{
\includegraphics[width=2.5in,height=2.3in]{pHinelx.eps}
}
\subfigure[]{
\includegraphics[width=2.5in,height=2.3in]{pCinelx.eps}
}
\caption{Comparison of the MIPP data with the MC predictions and previous measurements for p-H (a) and p-C (b) interactions. Both statistical and systematic uncertainties are shown.}
\end{center}
\end{figure*}
\section{KNO-based technique to get trigger efficiency}
A KNO-based technique has been developed to calculate the trigger efficiency. KNO stands for Koba, Nielsen and Olesen, the three authors who put forward the hypothesis that the probability distributions P$_{n}$(s) of producing n charged particles in a certain collision process should exhibit the scaling relation
\begin{center}
P$_{n}$(s) = $\frac{1}{<n(s)>}\psi(\frac{n}{<n(s)>})$, $\frac{n}{<n(s)>}$ = Z
\end{center}
as s $\rightarrow \infty$ with $<$n(s)$>$ being the average multiplicity of secondaries at collision energy s. $\psi$(Z) is called the KNO scaling function. The KNO scaling states that the function $\psi$(Z) is independent of the energy and only dependent on the variable Z. \\
In this method, we use a K-matrix K(n$_{o}|n_{t}$) from the MC which denotes the probability of obtaining observed multiplicity n$_{o}$, given a true multiplicity n$_{t}$. The trigger is not required in the formation of this matrix. The K-matrix is multiplied by true probabilities from the KNO function~\cite{kno} to get the predicted distribution. The observed distribution where the trigger is applied, is then fitted to the predicted distribution to extract the trigger efficiencies. The trigger efficiencies are the parameters fitted here. \\
The comparison of the observed and predicted distributions at the minimum and the comparison of KNO-based and MC trigger efficiencies as a function of number of tracks passing through the scintillator for 58 GeV/c p-H interactions are shown in Figures 3a and 3b respectively. The observed and predicted distributions agree very well. A drop in the trigger efficiency is observed at n=4 and n=6 which is not reasonable as the trigger efficiency should monotonically increase with the number of tracks through the scintillator. From the studies done, it has been found that this drop occurs due to the discrepancies between the MC and data multiplicity shapes. We have crosschecked the inelastic cross sections using this method and they are within 10\%.
\begin{figure*}[ht]
\begin{center}
\subfigure[]{
\includegraphics[width=2.5in,height=2.3in]{knofit_pH_jun5.eps}
}
\subfigure[]{
\includegraphics[width=2.5in,height=2.3in]{compeff_pH_jun5.eps}
}
\caption{(a) Comparison of observed and predicted distribution at the minimum. (b) Comparison of KNO-based and MC trigger efficiencies as a function of number of tracks passing through the scintillator for 58 GeV/c p-H interactions.}
\end{center}
\end{figure*}
\section{LH$_{2}$ and carbon multiplicities}
We know that charged multiplicities should be even in case of hydrogen target. Both even and odd multiplicities are observed in the data because of acceptances and reconstruction inefficiencies. We can unfold the MC K-matrix to go back to the data truth. As discrepancies have been observed between the data and MC multiplicity shapes, it has now been decided to use the KNO scaling function to get the data true multiplicity probabilities. The $<$n$>$ from our data is used. Probabilities are multiplied by the average inelastic cross section to get the cross sections as a function of multiplicity. Similarly cross sections are calculated for carbon target where multiplicities are both odd and even.\\
The comparison of the MIPP data with the MC predictions and previous measurements for 58 GeV/c p-H and p-C interactions is shown in Figures 4a and 4b respectively. For LH$_{2}$ target, discrepancies are found between the data and PDG at the lower end and tails. For carbon target, the data is consistent, within error bars, with the measurement of S. P. Denisov et al.~\cite{publ2} and consistent with the measurement of A. S. Carroll et al.~\cite{publ2} for multiplicity$>$15 only. The DPMJET and FLUKA shapes are not consistent with the data.
\begin{figure*}[ht]
\begin{center}
\subfigure[]{
\includegraphics[width=2.5in,height=2.3in]{inelxvsn_pH58_aug6.eps}
}
\subfigure[]{
\includegraphics[width=2.5in,height=2.3in]{pCinelxvsmult.eps}
}
\caption{Comparison of the MIPP data with the MC predictions and the previous measurements for 58 GeV/c p-H (a) and p-C (b) interactions.}
\end{center}
\end{figure*}
\section{Charged particle production cross sections}
Data and MC charged particle production cross sections have also been measured in bins of momentum, p$_{T}^{2}$ and x$_{F}$ for 58 and 120 GeV/c p-C interactions, and compared. For this, all charged particles are selected from the primary interaction. \\
Figures 5a, 5b and 5c show the comparison of the data and MC production cross sections for 58 GeV/c p-C interactions in bins of momentum, p$_{T}^{2}$ and x$_{F}$ respectively. The average production cross sections are calculated by summing over all the bins for both the data and MC for 58 and 120 GeV/c p-C interactions, and shown in Table 3. The MC is $\sim$12\% lower than the data at 58 GeV and $\sim$11\% lower than the data at 120 GeV. The average production cross section should be equal to average multiplicity times the average inelastic cross section. The average multiplicity is $\sim$7 and $\sim$8.5 in case of 58 and 120 GeV/c p-C interactions respectively. If we multiply it by the average inelastic cross section for both the data and MC, we get the numbers very close to the numbers shown in Table 3. \\
A drop of $\sim$8\% is observed in the cross section when it is binned in x$_{F}$. The x$_{F}$ of a particle is dependent on its mass and mass of the pion has been used for all the charged particles assuming that all of them are pions. It can be the reason for the observed drop in the cross section binned in x$_{F}$.
\begin{figure*}[ht]
\begin{center}
\subfigure[]{
\includegraphics[width=1.8in,height=1.8in]{comp_datamc_momX_pC58_jun10.eps}
}
\subfigure[]{
\includegraphics[width=1.8in,height=1.8in]{comp_datamc_pt2X_pC58_jun10.eps}
}
\subfigure[]{
\includegraphics[width=1.8in,height=1.8in]{comp_datamc_xfX_pC58_jun10.eps}
}
\caption{Comparison of the data and MC charged particle production cross sections in bins of momentum (a), p$_{T}^{2}$ (b) and x$_{F}$ (c) for 58 GeV/c p-C interactions.}
\end{center}
\end{figure*}
\section{Particle identification}
Four hypotheses are considered for particle identification (PID) --e/$\pi$/K/p (denoted by H). TPC, ToF and RICH detectors are used and PID quantities dE/dx, time of flight and ring radii are measured. Maximum likelihood technique has been used to determine the spectra of each particle type in data. A weight is formed by using the likelihood of a PID quantity for a particular hypothesis and the sum of the likelihoods of that quantity for all the hypotheses and is calculated as:
\begin{center}
Weight = $\frac{Likelihood}{\Sigma_{H}Likelihoods}$
\end{center}
A global weight is formed using the total likelihood i.e. product of all the detector likelihoods. This is determined for each track and used to weight the track for each hypothesis. Each particle enters all hypothesis dependent plots with its hypothesis dependent weight. The aim is to determine the momentum spectrum for each particle type. \\
Figure 6 shows the Bethe Bloche dE/dx curves superimposed on the TPC data. The data is described well by the predictions. The global weighted mass squared distributions for the pion and kaon hypotheses from the ToF and RICH detectors using 120 GeV/c p-C data are shown in Figure 7. The means of these distributions are at exactly the masses of the pion and kaon. Small contamination from other particles can be seen in the kaon m$^{2}$ distribution from the RICH. We are working on the algorithm to remove this contamination. Figure 8 shows the momentum spectra for the negatively and positively charged pions and kaons from 120 GeV/c p-C data. The contamination can be seen at $\sim$120 GeV/c in the momentum spectrum for K$^{+}$ particles.\\
\begin{table}[t]
\small
\begin{center}
\scalebox{0.9}{
\begin{tabular}{l|cc}
Energy (GeV) & Data (mb) & MC (mb) \\ \hline
58 & 1924.1$\pm55\rm(stat)^{+89.5}_{-94.2}$(syst) & 1697.3$\pm1.9\rm(stat)\pm51.1$(syst) \\
120 & 2224.5$\pm20.4\rm(stat)^{+92.5}_{-97.1}$(syst) & 1986.5$\pm1.7\rm(stat)\pm51.5$(syst) \\ \hline
\end{tabular}}
\caption{Comparison of the data and MC average charged particle production cross section for p-C interactions at 58 and 120 GeV.}
\end{center}
\end{table}
\begin{figure}
\centering
\includegraphics[width=8cm,height=6cm]{NuMI_dEdxPRPlots.eps}
\caption{Comparison of TPC data with the dE/dx predictions.}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=4in,height=3in]{m2tofandrich_oct10.eps}
\caption{Global weighted m$^{2}$ distributions for pions and kaons from the ToF (top) and RICH (bottom) detectors using 120 GeV/c p-C data.}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=4in,height=3in]{pikplmin_momspec_pC120data_oct10th13.eps}
\caption{Momentum spectra for $\pi^{\pm}$ (top) and K$^{\pm}$ (bottom) particles from 120 GeV/c p-C interactions.}
\end{figure}
We have to work out the inclusive pion and kaon production cross sections as a function of momentum for both the data and MC, and compare them.
|
\section{Proof that the full conditional densities of $\gamma_{st}$ are log-concave}
\label{sec:log_concave}
For the sake of convenience we re-arrange the factors of the full conditional
of $\gamma_{s^*t^*}$ as follows:
\begin{align}
[\gamma_{\ell,s^*t^*}\mid Z,C,\boldsymbol{\gamma}_{-s^*,-t^*},k]&\propto
\prod_{\ell=1}^k\frac{\Gamma\left(\sum_{i:z_i=j}\sum_{j:c_j=\ell}N_{i,s^*t^*}+\gamma_{s^*t^*}\right)}
{\Gamma\left(\gamma_{s^*t^*}\right)}\notag\\
&\times
\left(\frac{\Gamma\left(\sum_{t=1}^K\gamma_{s^*t}\right)}
{\Gamma\left(\sum_{t=1}^K\sum_{i:z_i=j}\sum_{j:c_j=\ell}N_{i,s^*t}+\sum_{t=1}^K\gamma_{s^*t}\right)}\notag\\
&\times \gamma^{a_{jk}-1}_{s^*t^*}\exp\left(-b_{jk}\gamma_{s^*t^*}\right)
\label{eq:fullcond_gamma2}
\end{align}
\section{Proof that ${F^L}_i$ and ${F^U}_i$ are distribution functions}
\label{sec:distribution_function}
Letting $X_{-i}$ denote all unknown variables other than $z_i$
we need to show that for almost all $X_{-i}$ the following holds:
\begin{itemize}
\item[(i)] $\lim_{h\rightarrow -\infty}F^L_i(h)=\lim_{h\rightarrow -\infty}F^U_i(h)=0$.
\item[(ii)] $\lim_{h\rightarrow \infty}F^L_i(h)=\lim_{h\rightarrow \infty}F^U_i(h)=1$.
\item[(iii)] For any $x_1\geq x_2$, $F^L_i(x_1)\geq F^L_i(x_2)$ and $F^U_i(x_1)\geq F^U_i(x_2)$.
\item[(iv)] $\lim_{h\rightarrow x+}F^L_i(h)=F^L_i(x)$ and $\lim_{h\rightarrow x+}F^U_i(h)=F^U_i(x)$.
\end{itemize}
{\bf Proof:}
Let $X_{-i}$ denote all unknown variables other than $z_i$. To prove (i), note that for all $h<1$,
$F_i(h\mid X_{-i})=0$ for almost all $X_{-i}$. Hence, by (\ref{eq:full_cond_z}) of MB
and by definition,
both $F^L_i(h)$ and $F^U_i(h)$ are 0 with probability 1. Hence, $\lim_{h\rightarrow -\infty}F^L_i(h)=\lim_{h\rightarrow -\infty}F^U_i(h)=0$
almost surely.
To prove (ii) note that for all $h>p$, $F_i(h\mid X_{-i})=1$ for almost all $X_{-i}$. Hence, for $h>p$, $F^L_i(h)=F^U_i(h)=1$,
that is, $\lim_{h\rightarrow \infty}F^L_i(h)=\lim_{h\rightarrow \infty}F^U_i(h)=1$ for almost all $X_{-i}$.
To show (iii), let $h_1>h_2$. Then, since $F_i(\cdot\mid X_{-i})$ is a distribution function satisfying monotonicity, it holds that
$F^L_i(h_2)=\inf_{X_{-i}}F_i(h_2\mid X_{-i})\leq F_i(h_2\mid X_{-i})\leq F_i(h_1\mid X_{-i})$ for almost all $X_{-i}$.
Hence, $F^L_i(h_2)\leq \inf_{X_{-i}}F_i(h_1\mid X_{-i})=F^L_i(h_1)$.
Similarly, $F^U_i(h_1)=\sup_{X_{-i}}F_i(h_1\mid X_{-i})\geq F_i(h_1\mid X_{-i})\geq F_i(h_2\mid X_{-i})$ for almost all $X_{-i}$. Hence,
$F^U_i(h_1)\geq\sup_{X_{-i}}F_i(h_2\mid X_{-i})=F^U_i(h_2)$.
To prove (iv), first observe that due to the monotonicity property (iii), the following hold for any $x$:
\begin{eqnarray}
\lim_{h\rightarrow x+}F^L_i(h)&\geq& F^L_i(x)\label{eq:mono1}\\
\lim_{h\rightarrow x+}F^U_i(h)&\geq& F^U_i(x)\label{eq:mono2}
\end{eqnarray}
Then observe that, due to discreteness, $F_i(\cdot\mid X_{-i})$ is constant in the interval $[x,x+\delta)$ for some $\delta>0$.
Since the supports of $F^L_i$, $F^U_i$ and $F_i(\cdot\mid X_{-i})$ for almost all $X_{-i}$ are same, $F^L_i$ and $F^U_i$
must also be constants in $[x,x+\delta)$. This implies that equality holds in (\ref{eq:mono1}) and (\ref{eq:mono2}).
Hence, both $F^L_i$ and $F^U_i$ satisfy all the properties of distribution functions.
\\[2mm]
{\bf Remark:} The right continuity property formalized by (iv) not be true for continuous variables.
Suppose $X\sim U(0, \theta)$, $\theta>0$. Here the distribution function is
$F(x\mid \theta) = \frac{x}{\theta}$, $0<x<\theta<\infty$. But
\[
\lim_{x\rightarrow 0+} \sup_{\theta}\frac{x}{\theta}\nonumber\\
= \lim_{x\rightarrow 0+} 1\nonumber\\
= 1\nonumber\\
\]
and,
\[
\sup_{\theta} \lim_{x\rightarrow 0+} \frac{x}{\theta}\nonumber\\
= \sup_{\theta} 0\nonumber\\
= 0\nonumber\\
\]
As a consequence of the above problem, attempts to construct suitable stochastic bounds for the
continuous parameters $(\Pi_p,\Theta_p)$ may not be fruitful. In our case such problem does not arise
since we only need to construct bounds for the discrete random variables to achieve our goal.
\section{Proof of validity of our CFTP algorithm}
\label{sec:validity}
{\bf Theorem:} The terminal chains coalesce almost surely in finite time and the value obtained at
time $t=0$ is a a realization from the target distribution.
\\[2mm]
{\bf Proof:}
Let $z^L_{it}$ denote the realization obtained at time $t$ by inverting
$F^U_i$, that is, $z^L_{it}={F^U_i}^{-}(R_{z_i,t})$, where
$\{R_{z_i,t}; i=1,\ldots,n;t=1,2,\ldots\}$ is a common set of $U(0,1)$ random numbers
which are $iid$ with respect to both $i$ and $t$.
used to simulate
$Z=(z_1,\ldots,z_n)'$ at time $t$ for Markov chains starting at all possible initial values.
Similarly, let $z^U_{it}={F^L_i}^{-}(R_{z_i,t})$.
Clearly, for any $z_{it}=F^{-}_i(R_{z_i,t}\mid X_{-i})$ started with any initial value and for any
$X_{-i}$, $z^L_{it}\leq z_{it}\leq z^U_{it}$ for all $i$ and $t$.
For $i=1,\ldots,n$ and for $j=1,2,\ldots$, we denote by $S^j_i$ the event
\[ z^L_{i,-2^j}(-2^{j-1})=z^U_{i,-2^j}(-2^{j-1}), \] which signifies that
the terminal chains and hence the individual chains started at
$t=-2^j$ will coalesce at $t=-2^{j-1}$.
It is important to note that both $F^L_i$ and $F^U_i$ are irreducible
which has the consequence that the probability of $S^j_i$, $P(S^j_i)>\epsilon_i>0$,
for some positive $\epsilon_i$. Since, for fixed $i$, $\{S^j_i;j=1,2,\ldots\}$ depends only upon the random numbers
$\{R_{z_i,t};t= -2^j,\ldots,-2^{j-1}\}$, $\{S^j_i;j=1,2,\ldots\}$ are independent with respect to $j$. Moreover, for fixed $j$, $S^j_i$
depends only upon the $iid$ random numbers $\{R_{z_i,-2^j};i=1,\ldots,n\}$. Hence, $\{S^j_i;i=1,\ldots,n;j=1,2,\ldots\}$ are
independent with respect to both $i$ and $j$.
Let $\epsilon=\min\{\epsilon_1,\ldots,\epsilon_n\}$. Then
due to independence of $\{S^j_i;i=1,\ldots,n\}$, it follows that for $j=1,2,\ldots$,
$\bar S^j=\cap_{i=1}^nS^j_i$ are independent, and
\begin{equation}
P\left(\bar S^j\right)\geq \epsilon^n
\label{eq:indep1}
\end{equation}
The rest of the proof resembles the proof of Theorem 2 of \ctn{Casella01}. In other words,
\begin{eqnarray}
P(\mbox{No coalescence after T iterations} )&\leq & \prod_{j=1}^T\left\{1-P(\bar S^j)\right\}\label{eq:indep2}\\
&=& \left\{(1-\epsilon^n)\right\}^T\rightarrow 0\ \ \mbox{as} \ \ T\rightarrow\infty\label{eq:indep3}.
\end{eqnarray}
Thus, the probability of coalescence is 1. That the time to coalesce is almost surely finite follows
from the Borel-Cantelli lemma, exactly as in \ctn{Casella01}.
The realization obtained at time $t=0$ after occurrence of the coalescence
event $\bar S_j$ for some $j$ yields $Z=Z_0$ exactly from its marginal posterior
distribution. Given this $Z_0$, drawing $\Pi_{p0}$ from the full conditional distribution
(\ref{eq:full_cond_pi}) of MB
and then drawing $\Theta_{p0}$ sequentially from (\ref{eq:full_cond_lambda})
and (\ref{eq:full_cond_mu}) of MB
given $Z_0$ and $\Pi_{p0}$, yields a realization $(Z_0,\Pi_{p0},\Theta_{p0})$ exactly from
the target posterior. The proof of this exactness follows readily from the general proof (see, for example,
\ctn{Prop96}, \ctn{Casella01})
that if convergent Markov chains colasece in a CFTP algorithm during time $t\leq 0$, then the realization obtained
at time $t=0$ is exactly from the stationary distribution.
\section{Uniform ergodicity}
\label{sec:uniform_ergodicity}
Let $P(\cdot,\cdot)$ denote a Markov transition kernel where
$P(x,A)$ denotes transition from the state $x$ to the set $A\in\mathcal B$, $\mathcal B$
being the associated Borel $\sigma$-algebra.
If we can show that for all $x$ in the state space the following minorization holds:
\[ P(x, A)\geq\epsilon Q(A),\hspace{2mm}A\in {\mathcal{B}}, \]
for some $0<\epsilon\leq 1$ and for some probability measure $Q(\cdot)$,
then $P(\cdot,\cdot)$ is uniformly ergodic.
In our mixture model situation the Gibbs sampling transition kernel is
\begin{eqnarray}
&&\left[Z^{(t)},\Pi^{(t)}_p,\Theta^{(t)}_p\mid Z^{(t-1)},\Pi^{(t-1)}_p,\Theta^{(t-1)}_p\right]\nonumber\\
&& \ \ = \left[Z^{(t)}\mid\Pi^{(t-1)}_p,\Theta^{(t-1)}_p,Y\right]\left[\Pi^{(t)}_p\mid Z^{(t)},Y\right]\left[\Theta^{(t)}_p\mid Z^{(t)},\Pi^{(t)}_p,Y\right]\nonumber\\
&& \ \ \geq \left\{\inf_{\Pi^{(t-1)}_p,\Theta^{(t-1)}_p}\left[Z^{(t)}\mid\Pi^{(t-1)}_p,\Theta^{(t-1)}_p,Y\right]\right\}
\left[\Pi^{(t)}_p\mid Z^{(t)},Y\right]\left[\Theta^{(t)}_p\mid Z^{(t)},\Pi^{(t)}_p,Y\right]
\label{eq:minorization}
\end{eqnarray}
The infimum in inequality (\ref{eq:minorization}) is finite since both $\Pi^{(t-1)}_p$ and $\Theta^{(t-1)}_p$ are bounded.
Denoting the right hand side of inequality (\ref{eq:minorization}) by
$g(Z^{(t)},\Pi^{(t)}_p,\Theta^{(t)}_p)$, we put
\begin{equation}
\epsilon = \sum_{Z} \int_{\Pi_p}\int_{\Theta_p} g(Z,\Pi_p,\Theta_p)d\Pi_pd\Theta_p>0.
\label{eq:epsilon}
\end{equation}
Since $g(\cdot)$ is bounded above by the Gibbs transition kernel which integrates to 1, it follows
from (\ref{eq:epsilon}) that $0<\epsilon\leq 1$.
Hence, identifying the density of the $Q$-measure as $g(\cdot)/ \epsilon$, the minorization
condition required for establishment of uniform ergodicity of our Gibbs sampling chain is seen to hold.
\section{Proof that coalescence of $C$ implies the coalescence of $S$}
\label{sec:s_coalescence}
Let $C=(c_1,\ldots,c_M)'$ be coalescent.
For convenience of illustration assume that after simulating each $c_j$, followed by drawing $\theta_j$ depending
upon the simulated value of $c_j$, the entire set $S$ is obtained from the updated set of parameters $\Theta_M$.
Note that in practice, only $s_j$ will be obtained immediately after updating $c_j$ and $\theta_j$. Let $S_{-j}=\{s_1,\ldots,s_{j-1},s_{j+1},\ldots,s_M\}$.
Then $c_{j+1}=\ell$ denotes the $\ell$-th distinct element of $S_{-j}$. If $\{1,\ldots,d_j\}$ are the distinct components in $S_{-j}$,
$d_j$ being the number of distinct components, and $\ell\leq s_j$, then $s_{j+1}=\ell$. On the other hand, if $\ell<c_{j+1}\leq d_j+1$, then $s_{j+1}=s_j+1$.
Now note that $s_1=1$, which is always coalescent. If $c_2>1$, then $s_2=2$, else $s_2=1$, for all Markov chains. Hence, $s_2$
is coalescent. If $c_3>s_2$, then $s_3=s_2+1$, else $s_3=c_3$. Since $s_2$ is coalescent,
then so is $s_3$. In general, if $c_{j+1}>s_j$, then $s_{j+1}=s_j+1$, else $s_{j+1}=c_{j+1}$. Since $s_1,\ldots,s_j$ are coalescent, hence
so is $s_{j+1}$, for $j=1,\ldots,M-1$. In other words, $S$ must coalesce if $C$ coalesces.
\section{Introduction}
\label{sec:introduction}
We consider the problem of clustering a panel of categorical time series $\boldsymbol{y}_i;i=1,\ldots,N$
into several classes (components), assuming that the number of classes is unknown.
The known number of components situation has been recently handled by \ctn{Pamminger10}, who
consider a Bayesian mixture model based approach with a fixed number of components.
However, the authors reported serious difficulties in reliably determining the appropriate
number of components using the traditional approaches like Bayes Information Criterion (BIC).
We completely avoid the difficulties of the fixed components approach by adopting the
assuming that the number of components is unknown, but is bounded above by a number
specified by experts, the upper bound signifying that the number of possible clusters
of the time series can not exceed the specified upper limit. Such a model has been
proposed by \ctn{Bhattacharya08}; see also \ctn{Sabya10} and \ctn{Sabya10a}.
We develop a perfect simulation method for sampling exactly from the underlying posterior
distribution. Perfect simulation for mixtures with unknown number of components has been
developed by \ctn{Sabya10b}, but in the time series context there are some additional complications,
to be explained in due course.
Indeed, these additional difficulties led us to develop
a general perfect simulation methodology in the case of joint distributions with log-concave full conditional
distributions, which is of independent interest.
With our new developments related to pefect sampling, we then proceed to analyze the Austrian
wage mobility data, obtaining the modes of the posterior distribution of clustering as well
as the desired highest posterior distribution credible regions, using the methods detailed
in \ctn{Sabya10a} and \ctn{Sabya10}. In particular, we demonstrate that ignoring uncertainty
in the number of clusters in the approach of \ctn{Pamminger10} seriously affects inference.
The rest of our paper is structured as follows. In Section \ref{sec:categorical_mixture},
adopting the mixture model of \ctn{Bhattacharya08} we model the time series as mixtures of
unknown number of components, and in Section \ref{sec:fullcond} provide the full conditional
distributions to be used for perfect simulation, along with the need for perfect simulation
from joint distributions of continuous parameters with log-concave full conditionals in our problem.
Some more details are presented in the supplement, the sections of which we refer to by
using the prefix ``S-".
In Section \ref{sec:perfect_logconcave} we introduce our perfect simulation idea in the case
of continuous joint distributions having log-concave full conditionals.
Using this development, and adopting the perfect simulation idea for mixtures with unknown
number of components proposed by \ctn{Sabya10b} we then present the relevant perfect
simulation methodology for our categorical time series problem.
\section{Mixtures of categorical time series with unknown number of components}
\label{sec:categorical_mixture}
In this work we confine ourselves to mixtures of categorical time series with Markov chain
clustering, which has also been the main aspect of study in \ctn{Pamminger10}, albeit
the latter consider only fixed number of components. In what follows we shall borrow
some notation already described in \ctn{Pamminger10}.
We consider the mixture model of the following form:
for $i=1,\ldots,N$,
\begin{equation}
f(\boldsymbol{y}_i\mid\boldsymbol{\Theta})=\frac{1}{M}\sum_{h=1}^M\prod_{r=1}^{T_i}f(y_{ir}\mid y_{i,r-1},\boldsymbol{\theta}_h)
=\frac{1}{M}\sum_{h=1}^M\prod_{s=1}^K\prod_{t=1}^K\theta^{N_{i,st}}_{h,st},
\label{eq:categorical_mixture}
\end{equation}
where $N_{i,st}=\#\{y_{ir}=t,y_{i,r-1}=s\}$ is the number of transitions from state $s$ to state $t$
observed in time series $i$, and, for each $h\in\{1,\ldots,M\}$, $\boldsymbol{\theta}_h=((\theta_{h,st}));s,t=1,\ldots,K$,
is the transition matrix of the underlying Markov chain model consisting of $K$ states.
The latter satisfies $\sum_{t=1}^K\theta_{h,st}=1$ $\forall h,s$.
In (\ref{eq:categorical_mixture}), $M$ is the maximum number of components, specified, perhaps,
by some expert; however, \ctn{Sabya11} show how $M$ can be obtained objectively and optimally from a
Bayesian asymptotics perspective.
We next consider the following Dirichlet process (DP) prior for $\boldsymbol{\Theta}$:
for $h=1,\ldots,M$,
\begin{align}
\boldsymbol{\theta}_h&\stackrel{iid}{\sim}\boldsymbol{G}\label{eq:dp1}\\
\boldsymbol{G}&\sim DP(\alpha\boldsymbol{G}_0)\label{eq:dp2}
\end{align}
Under $\boldsymbol{G}_0$, for $h=1,\ldots,M$ and $s=1,\ldots,K$,
\begin{align}
(\theta_{h,s1},\ldots,\theta_{h,sK})&\sim Dirichlet(\gamma_{s1},\ldots,\gamma_{sK})
\label{eq:dp3}
\end{align}
In addition, we assume that
\begin{align}
\gamma_{st}&\sim Gamma(a_{st},b_{st});\ \ \ s=1,\ldots,K; t=1,\ldots,K,
\label{eq:gamma_prior}
\end{align}
where $Gamma(a,b)$ denotes the gamma distribution of the form
$\frac{b^a}{\Gamma(a)}x^{a-1}\exp(-bx)$, having mean $a/b$ and variance $a/b^2$.
We remark that \ctn{Pamminger10} assumed a discrete prior distribution on $\{\gamma_{st};s,t=1,\ldots,K\}$,
namely, the negative multinomial distribution. However, continous priors, such as ours,
are perhaps more appropriate and more natural.
The Dirichlet process prior assumption entails that mixture
model (\ref{eq:categorical_mixture}) reduces to the following form:
\begin{equation}
f(\boldsymbol{y}_i\mid\boldsymbol{\Theta}_M)=\sum_{\ell=1}^p\pi_{\ell}\prod_{r=1}^{T_i}f(y_{ir}\mid y_{i,r-1},\boldsymbol{\phi}_{\ell})
=\sum_{\ell=1}^p\pi_{\ell}\prod_{s=1}^K\prod_{t=1}^K\phi^{N_{i,st}}_{\ell,st},
\label{eq:categorical_mixture2}
\end{equation}
where $\boldsymbol{\phi}_{\ell}$ denotes the $\ell$-th distinct component among $\boldsymbol{\Theta}_M=\{\boldsymbol{\theta}_1,\ldots,\boldsymbol{\theta}_M\}$,
and $\pi_{\ell}=M_{\ell}/M$, with $M_{\ell}=\#\{h:\boldsymbol{\theta}_h=\boldsymbol{\phi}_{\ell}\}$. In (\ref{eq:categorical_mixture2}),
$p~(1\leq p\leq M)$ denotes the {\it random} number of distinct mixture components.
\section{Full conditional distributions}
\label{sec:fullcond}
Let $\boldsymbol{Y}=\{\boldsymbol{y}_1,\ldots,\boldsymbol{y}_n\}$ denote the data set.
We define the set of allocation variables $Z=\{z_1,\ldots,z_n\}$, where $z_i=j$ if $\boldsymbol{y}_i$
arises from the $j$-th component.
Letting $\boldsymbol{\Phi}=\{\boldsymbol{\phi}_1,\ldots,\boldsymbol{\phi}_k\}$ denote the distinct components in $\Theta_M$,
the element $c_j$ of the configuration vector $C=\{c_1,\ldots,c_M\}$ is defined as $c_j=\ell$ if and only if
$\boldsymbol{\theta}_j=\boldsymbol{\phi}_{\ell}$; $j=1,\ldots,M$, $\ell=1,\ldots,k$. Thus, $(Z,\Theta_M)$ is reparameterized
to $(Z,C,k,\boldsymbol{\Phi})$, $k$ denoting the number of distinct components in $\Theta_M$.
\subsection{Full conditionals of $\{z_1,\ldots,z_n\}$}
\label{subsec:fullcond_z}
For $i=1,\ldots,n$, let $Z_{-i}=\{z_1,\ldots,z_{i-1},z_{i+1},\ldots,z_n\}$, and
let $C$ consist of $k$ distinct components. Then, denoting the set $\{\gamma_{st};s,t=1,\ldots,K\}$
by $\boldsymbol{\gamma}$, the full conditional distribution of $z_i$ is given by
\begin{align}
[z_i=r\mid \boldsymbol{Y}, Z_{-i},C,\boldsymbol{\Phi},\boldsymbol{\gamma},k]&\propto
\prod_{s=1}^K\prod_{t=1}^K\theta^{N_{i,st}}_{r,st}
\label{eq:fullcond_z}
\end{align}
\subsection{Full conditionals of $\{c_1,\ldots,c_M\}$}
\label{subsec:fullcond_c}
To obtain the full conditional of $c_r;r=1,\ldots,M$, first let $k_{r}$ denote the number of distinct values in
$\boldsymbol{\Theta}_{-rM}=\{\boldsymbol{\theta}_1,\ldots,\boldsymbol{\theta}_{r-1},\boldsymbol{\theta}_{r+1},\ldots,\boldsymbol{\theta}_M\}$, and let
$\boldsymbol{\phi}^{(r^*)}_{\ell}$; $\ell=1,\ldots,k_{r}$ denote
the distinct values.
Also suppose that $\boldsymbol{\phi}^{(r^*)}_{\ell}$ occurs $M_{\ell r}$ times.
Then the conditional distribution of $c_r$ is given by
\begin{equation}
[c_r=\ell\mid \boldsymbol{Y},Z,C_{-r},\boldsymbol{\Phi},\boldsymbol{\gamma},k_r]=\left\{\begin{array}{c}\kappa q_{\ell r}\hspace{2mm}\mbox{if}\hspace{2mm}
\ell=1,\ldots,k_r\\ \kappa q_{0r}\hspace{2mm}\mbox{if}\hspace{2mm}\ell=k_r+1\end{array}\right.
\label{eq:fullcond_c}
\end{equation}
where
\begin{align}
q_{\ell r}&\propto M_{\ell r}\times
\prod_{s=1}^K\prod_{t=1}^K\phi^{\sum_{i:z_i=r}N_{i,st}}_{\ell,st}
\label{eq:fullcond_c1}
\end{align}
and
\begin{align}
q_{0 r}&\propto \alpha\times
\prod_{s=1}^K
\frac{\Gamma\left(\sum_{t=1}^K\gamma_{st}\right)}
{\prod_{t=1}^K\Gamma\left(\gamma_{st}\right)}
\times\prod_{s=1}^K
\frac{\prod_{t=1}^K\Gamma\left(\sum_{i:z_i=r}N_{i,st}+\gamma_{st}\right)}
{\Gamma\left(\sum_{t=1}^K\sum_{i:z_i=r}N_{i,st}+\sum_{t=1}^K\gamma_{st}\right)}
\label{eq:fullcond_c2}
\end{align}
\subsection{Full conditionals of $\{\boldsymbol{\phi}_{\ell};\ell=1,\ldots,k\}$}
\label{subsec:fullcond_phi}
Assuming that there are $k$ distinct components in $C$, the full conditional distribution
of $\boldsymbol{\phi}_{\ell};\ell=1,\ldots,k$, is given by
\begin{align}
[\boldsymbol{\phi}_{\ell}\mid \boldsymbol{Y}, Z, C,\boldsymbol{\Phi}_{-\ell},\boldsymbol{\gamma},k]
&= \prod_{s=1}^K\prod_{t=1}^K\phi^{\sum_{i:z_i=j}\sum_{j:c_j=\ell}N_{i,st}+\gamma_{st}-1}_{\ell,st}\notag\\
& \ \ \ \ \times\prod_{s=1}^K\frac{\Gamma\left(\sum_{t=1}^K\sum_{i:z_i=j}\sum_{j:c_j=\ell}N_{i,st}+\sum_{t=1}^K\gamma_{st}\right)}
{\prod_{t=1}^K\Gamma\left(\sum_{i:z_i=j}\sum_{j:c_j=\ell}N_{i,st}+\gamma_{st}\right)},
\label{eq:fullcond_phi}
\end{align}
which are conditionally independent of $\boldsymbol{\Phi}_{-\ell}$.
The conditional mean and variance of $\phi_{\ell,s^*t^*}$ are given, respectively, by
\begin{equation}
\zeta_{\ell,s^*t^*}=E\left[\phi_{\ell,s^*t^*}\mid \boldsymbol{Y}, Z, C,\boldsymbol{\gamma},k\right]=
\frac{\gamma_{s^*t^*}+\sum_{i:z_i=j;j:c_j=\ell} N_{i,s^*t^*}}
{\sum_{t=1}^K\left(\gamma_{s^*t}+\sum_{i:z_i=j;j:c_j=\ell} N_{i,s^*t}\right)},
\label{eq:phi4}
\end{equation}
and
\\[2mm]
$\varphi_{\ell,s^*t^*}=Var\left[\phi_{\ell,s^*t^*}\mid \boldsymbol{Y}, Z, C,\boldsymbol{\gamma},k\right]$
\begin{equation}
=\frac{\left(\gamma_{s^*t^*}+\sum_{i:z_i=j;j:c_j=\ell} N_{i,s^*t^*}\right)
\left\{\sum_{t\neq t^*}\left (\gamma_{s^*t}+\sum_{i:z_i=j;j:c_j=\ell} N_{i,s^*t}\right)\right\}}
{\left\{\sum_{t=1}^K\left(\gamma_{s^*t}+\sum_{i:z_i=j;j:c_j=\ell} N_{i,s^*t}\right)\right\}^2
\left\{1+\sum_{t=1}^K\left(\gamma_{s^*t}+\sum_{i:z_i=j;j:c_j=\ell} N_{i,s^*t}\right)\right\}},
\label{eq:phi5}
\end{equation}
It follows that
\begin{equation}
\frac{\varphi_{\ell,s^*t^*}}
{\zeta_{\ell,s^*t^*}
\left(1-\zeta_{\ell,s^*t^*}\right)}
=\frac{1}{1+\sum_{t=1}^K\left(\gamma_{s^*t}+\sum_{i:z_i=j;j:c_j=\ell} N_{i,s^*t}\right)}.
\label{eq:cluster1}
\end{equation}
Accordingly, as in \ctn{Pamminger10}, but somewhat differently, we can interpret
$\Sigma_{\ell,s^*}=$$\sum_{t=1}^K(\gamma_{s^*t}$$+\sum_{i:z_i=j;j:c_j=\ell} N_{i,s^*t})$
as a {\it conditional} measure of heterogeneity in the corresponding rows of $\boldsymbol{\phi}_{\ell}$
of the $\ell$-th cluster. Small values of $\Sigma_{\ell,s^*}$ implies high degree of variability
of the individual transition probabilities $\phi_{\ell,s^*t^*}$ and large deviations of
$\boldsymbol{\phi}_{\ell,s^*}=\left(\phi_{\ell,s^*1},\ldots,\phi_{\ell,s^*K}\right)$ from the group mean
$\boldsymbol{\zeta}_{\ell,s^*}=\left(\zeta_{\ell,s^*1},\ldots,\zeta_{\ell,s^*K}\right)$.
Large values of $\Sigma_{\ell,s^*}$ indicate small variability in the $s^*$-th row, implying
that the individual transition probabilities $\phi_{\ell,s^*t^*}$ are nearly the same as
as the group means $\zeta_{\ell,s^*t^*}$.
Interestingly, for the purpose of perfect simulation, the full or marginal conditional
distributions of $\phi_{\ell,s^*t^*}$, given below, will be shown to be more
important than those of $\boldsymbol{\phi}_{\ell}$,
even though the latter is just the standard Dirichlet distribution and straightforward to simulate from.
\subsubsection{Full and marginal conditionals of $\phi_{\ell,s^*t^*}$}
The full conditional of $\phi_{\ell,s^*t^*}$ is given by
\\[2mm]
$[\phi_{\ell,s^*t^*}\mid \boldsymbol{Y}, Z, C,\boldsymbol{\Phi}_{-\ell,-s^*,-t^*},\boldsymbol{\gamma},k]$
\begin{align}
&\propto \phi^{\sum_{i:z_i=j}\sum_{j:c_j=\ell}N_{i,s^*t^*}+\gamma_{s^*t^*}-1}_{\ell,s^*t^*}\notag\\
&\ \ \ \ \times\left(1-\sum_{t=1}^K\phi_{\ell,s^*t}\right)^{\sum_{i:z_i=j}\sum_{j:c_j=\ell}N_{i,s^*K}+\gamma_{s^*K}-1}\notag\\
\label{eq:fullcond_phi2}
\end{align}
In the above, $\boldsymbol{\Phi}_{-\ell,-s^*,-t^*}$ denotes $\boldsymbol{\Phi}$ without $\phi_{\ell,s^*t^*}$.
The marginal conditional of $\phi_{\ell,s^*t^*}$ is given by
\\[2mm]
$[\phi_{\ell,s^*t^*}\mid \boldsymbol{Y}, Z, C,\boldsymbol{\Phi}_{-\ell,-s^*,-t^*},\boldsymbol{\gamma},k]$
\begin{align}
&\propto \phi^{\sum_{i:z_i=j;j:c_j=\ell}N_{i,s^*t^*}+\gamma_{s^*t^*}-1}_{\ell,s^*t^*}\notag\\
&\ \ \ \ \times\left(1-\phi_{\ell,s^*t^*}\right)^{\sum_{t\neq t^*}
\left(\sum_{i:z_i=j;j:c_j=\ell}N_{i,s^*t}+\gamma_{s^*t}\right)-1},
\label{eq:marginal_cond_phi2}
\end{align}
which is a $Beta$ distribution with parameters $\sum_{i:z_i=j;j:c_j=\ell}N_{i,s^*t^*}+\gamma_{s^*t^*}$
and $\sum_{t\neq t^*}\left(\sum_{i:z_i=j;j:c_j=\ell}N_{i,s^*t}+\gamma_{s^*t}\right)$.
\subsection{Full conditionals of $\{\gamma_{st};s,t=1,\ldots,K\}$}
\label{subsec:fullcond_gamma}
Assuming that $C$ consists of $k$ distinct components, the full
conditional distribution of $\gamma_{s^*t^*}$, for
$s^*=1,\ldots,K$, and $t^*=1,\ldots,K$, is given by
\begin{align}
[\gamma_{\ell,s^*t^*}\mid \boldsymbol{Y}, Z,C,\boldsymbol{\Phi},\boldsymbol{\gamma}_{-s^*,-t^*},k]&\propto
\left(\prod_{\ell=1}^k\phi^{\gamma_{s^*t^*}-1}_{\ell,s^*t^*}\right)
\times
\left(\frac{\Gamma\left(\sum_{t=1}^K\gamma_{s^*t}\right)}
{\Gamma\left(\gamma_{s^*t^*}\right)}\right)^k\notag\\
&\times \gamma^{a_{s^*t^*}-1}_{s^*t^*}\exp\left(-b_{s^*t^*}\gamma_{s^*t^*}\right)
\label{eq:fullcond_gamma}
\end{align}
In the above, $\boldsymbol{\gamma}_{-s^*,-t^*}$ denotes all elements of the $\boldsymbol{\gamma}$-parameters
except $\gamma_{s^*t^*}$.
\subsection{Relabeling $C$}
\label{subsec:relabeling}
Simulation of $C$ by successively simulating from the full conditional distributions (\ref{eq:fullcond_c}) incurs
a labeling problem. For instance, it is possible that all $c_j$ are equal even though each of them corresponds to
a distinct $\boldsymbol{\theta}_j$. For an example, suppose that $\boldsymbol{\Phi}$ consists of $M$ distinct elements, and $c_j=M$ $\forall j$.
Then although there are actually $M$ distinct components, one ends up obtaining just one distinct component.
For perfect sampling \ctn{Sabya10b} created a labeling method which relabels $C$ such that the relabeled version,
denoted by $S=(s_1,\ldots,s_M)'$, coalesces if $C$ coalesces.
To construct $S$ we first simulate $c_j$ from (\ref{eq:fullcond_c}); if $c_j\in\{1,\ldots,k_j\}$, then we set $\boldsymbol{\theta}_j=\boldsymbol{\phi}_{c_j}$
and if $c_j=k_j+1$, we draw $\boldsymbol{\theta}_j=\boldsymbol{\phi}_{c_j}\sim G_j$. The elements of $S$ are obtained from
the following definition of $s_j$:
$s_j=\ell$ if and only if $\boldsymbol{\theta}_j=\boldsymbol{\phi}_{\ell}$. Note that $s_1=1$ and $1\leq s_j\leq s_{j-1}+1$.
\ctn{Sabya10b} proved that coalescence of $C$ implies the coalescence of $S$, irrespective of the value of $\boldsymbol{\Phi}$.
\subsection{Full conditionals using $S$}
\label{subsec:fullcond_s}
With the introduction of $S$ it is now required to modify some of the full conditionals of the unknown random variables, in addition
to introduction of the full conditional distribution of $S$.
The form of the full conditional $[z_i\mid \boldsymbol{Y},S,k,\boldsymbol{\Phi},\boldsymbol{\gamma}]$ remains the same as
(\ref{eq:fullcond_z}), but $\Theta_M$ involved in the right hand side
of (\ref{eq:fullcond_z}) is now obtained from $S$ and $\boldsymbol{\Phi}$. The modified full conditional of $c_j$, which we denote
by $[c_j\mid \boldsymbol{Y},Z,S_{-j},k_j,\boldsymbol{\Phi}]$, now depends upon $S_{-j}$, rather than $C_{-j}$, the notation being clear from the context.
The form of this full conditional remains the same as (\ref{eq:fullcond_c}) but now the distinct components
$\boldsymbol{\phi}^{j^*}_{\ell}$; $\ell=1,\ldots,k_{j}$ are associated with the corresponding components of $S$ rather than $C$. The form of the modified full
conditional distribution of $\boldsymbol{\phi}_{\ell}$, which we now denote by $\left[\boldsymbol{\phi}_{\ell}\mid \boldsymbol{Y},Z,S,k\right]$, remains the same as
(\ref{eq:fullcond_phi}), only $C$ must be replaced with $S$.
Also $k$ and $k_j$ are now assumed to be associated with $S$.
The conditional posterior $[S\mid \boldsymbol{Y},C,\boldsymbol{\Phi},\boldsymbol{\gamma},k]$ gives point mass to $S^*$, where
$S^*=\{s^*_1,\ldots,s^*_M\}$ is the relabeling obtained from $C$ and $\Theta_M$
following the method described in Section \ref{subsec:relabeling}.
For the construction of bounds, the individual full conditionals $[s_j\mid Y, S_{-j},C,\boldsymbol{\Phi},\boldsymbol{\gamma},k]$,
giving full mass to $s^*_j$, will be considered due to
convenience of dealing with distribution functions of one variable.
It follows that once $Z$ and $C$ coalesces, $S$ and $\boldsymbol{\Phi}$ must also coalesce.
In the next section we describe how to construct efficient bounding chains for $Z$, $C$ and $S$.
Bounding chains for $S$ are not strictly necessary as it is possible to optimize the bounds for $Z$ and $C$ with respect to $S$, but the efficiency of the other bounding chains
is improved, leading to an improved perfect sampling algorithm,
if we also construct bounding chains for $S$.
\subsection{Need for perfect simulation of $(\boldsymbol{\Phi},\boldsymbol{\gamma})$ given the rest}
\label{subsec:perfect_need}
The perfect sampling methodology for mixtures of unknown number of components
developed in \ctn{Sabya10b} can be envisaged for simulating exactly from the posterior
in this categorical time series problem.
Their method requires simulation of the discrete parameters $(Z,C,S)$ only and not the continuous parameters
$\boldsymbol{\Phi}$ and $\boldsymbol{\gamma}$ until coalescence
of the discrete parameters. Simulation of the continuous parameters is necessary only after
the discrete parameters have coalesced. In our example, however, simulation of $\boldsymbol{\Phi}$ and $\boldsymbol{\gamma}$
given $Z$ and $C$, even after coalescence of the latter, is not straightforward. This is because
there does not seem to exist any method of directly simulating from the (joint) full conditional of $(\boldsymbol{\Phi},\boldsymbol{\gamma})$
and so it is required to simulate from the component-wise full conditionals of $\boldsymbol{\phi}_{\ell}$ given $\boldsymbol{\gamma}$, and
from the (non-standard) component-wise full conditionals of $\gamma_{st}$, given $Z$, $S$, and $\boldsymbol{\gamma}_{-s,-t}$ and $\boldsymbol{\Phi}$,
and although the initial values of $Z$ and $S$ are the coalesced values of the respective
bounding chains, the initial values of $\boldsymbol{\gamma}$ for simulating $\boldsymbol{\phi}_{\ell}$
or the initial values of $\boldsymbol{\Phi}$ and $\boldsymbol{\gamma}_{-s,-t}$ for simulating $\gamma_{st}$, are not available. The non-availability
of starting values is due to the fact that before coalescence of $(Z,C,S)$,
$\boldsymbol{\Phi}$ and $\boldsymbol{\gamma}$ are not simulated at all.
The above problem calls for the need for perfect simulation of $\boldsymbol{\Phi}$ and $\boldsymbol{\gamma}$ given $(Z,S)$, using the
full conditionals of $\boldsymbol{\Phi}_{\ell}$ and $\gamma_{st}$, given by (\ref{eq:fullcond_phi}) and
(\ref{eq:fullcond_gamma}), respectively. Thus, our main perfect
simulation methodology must proceed via incorporation of another perfect sampling method involving
the full conditionals of $\gamma_{st}$. But Gibbs sampling-based perfect simulation in the case
of continuous parameters is not developed in the literature. In this paper, we propose and develop
a novel and general perfect simulation methodology using full conditional distributions of continuous
parameters. All we require is that the full conditionals are log-concave. We then specialize our
general methodology to the problem of perfectly simulating $\boldsymbol{\Phi}$ and $\boldsymbol{\gamma}$ given $Z,S$ within the perfect
sampling methodology of \ctn{Sabya10b}.
However, as mentioned already, for perfect sampling, we shall need to utilize the full conditional
of $\phi_{\ell,st}$, given by (\ref{eq:fullcond_phi2}) rather than that of $\boldsymbol{\phi}_{\ell}$, given by
(\ref{eq:fullcond_phi}).
Indeed, it is easy to see that the full conditionals of $\phi_{\ell,st}$ and $\gamma_{\ell,st}$ satisfy
\begin{align}
&\frac{d^2}{d\phi^2_{\ell,st}}[\phi_{\ell,st}\mid\boldsymbol{Y},Z,S,\boldsymbol{\Phi}_{-\ell,-s,-t},\boldsymbol{\gamma},k]<0,\label{eq:lc1}\\
&\frac{d^2}{d\gamma^2_{st}}[\gamma_{st}\mid\boldsymbol{Y},Z,S,\boldsymbol{\Phi},\boldsymbol{\gamma}_{-s,-t},k]<0,\label{eq:lc2}
\end{align}
provided that $\sum_{i:z_i=j}\sum_{j:c_j=\ell}N_{i,st}+\gamma_{st}>1$ and $a_{st}>1$.
The proof of (\ref{eq:lc1}) follows by simple differentiation, while the proof of (\ref{eq:lc2})
also requires the formula (see \ctn{Bowman88}):
\begin{equation}
\frac{d^2}{dx^2}\log\left[\Gamma(x)\right]=\frac{1}{x}+\frac{1}{2x^2}
+\frac{2\pi}{x}\int_0^{\infty}\frac{y\sqrt{t}}{(x^2+t)(y-1)^2}dt,
\label{eq:lc3}
\end{equation}
where $y=\exp(2\pi\sqrt{t})$.
Using the above formula the proof of (\ref{eq:lc2}) follows in similar lines
as the proof of Proposition 2 of \ctn{He98}.
Note that although it is possible to integrate out $\boldsymbol{\Phi}$ thanks to conjugacy
and obtain the marginalized full conditionals of $(Z,C,\boldsymbol{\gamma})$ (see Section \ref{sec:marginalized_fullcond}
of the supplement), it can be
easily verified that the resulting expression
for the full conditional of $\gamma_{st}$ need not admit log-concavity;
see Section \ref{subsec:gamma_logconcavity} of the supplement for details.
This lack of log-concavity makes it difficult to generate perfect samples from the full conditional of $\boldsymbol{\gamma}$.
\section{Perfect simulation in posteriors with log-concave full conditionals}
\label{sec:perfect_logconcave}
Before introducing our perfect simulation idea
in Gibbs sampling for continuous, log-concave, full conditionals, we first provide a brief description of
adaptive rejection sampling (ARS) following \ctn{Gilks92a}.
\subsection{Overview of ARS}
\label{subsec:ars}
Assuming that it is required to sample from a log-concave density $g(\cdot)$ with
$g(\cdot)$ continuous and differentiable everywhere on a set $D$, let us suppose that
$h(x)=\log g(x)$ and $h'(x)$, the first differential of $h(\cdot)$ has been evaluated
at $m$ abscissae in $D:x_1\leq x_2\leq\cdots x_m$. For $j=1,\ldots,m-1$, define
\begin{equation}
v_j=\frac{h(x_{j+1})-h(x_j)-x_{j+1}h'(x_{j+1})+x_jh'(x_j)}{h'(x_j)-h'(x_{j+1})}
\label{eq:v}
\end{equation}
For $x\in[v_{j-1},v_j]$; $j=1,\ldots,m$, define
\begin{equation}
u_m(x)=h(x_j)+(x-x_j)h'(x_j),
\label{eq:u}
\end{equation}
Here $v_0$ is the lower bound of $D$ (or $-\infty$ if $D$ is not bounded below)
and $v_m$ is the upper bound of $D$ (or $\infty$ if $D$ is not bounded above).
Also define
\begin{equation}
s_m(x)=\frac{\exp\{u_m(x)\}}{\int_D\exp\{u_m(x')\}dx'}
\label{eq:s}
\end{equation}
Also define, for $x\in[x_j,x_{j+1}]$; $j=1,\ldots,m-1$,
\begin{equation}
\ell_m(x)=\frac{(x_{j+1}-x)h(x_j)+(x-x_j)h(x_{j+1})}{x_{j+1}-x_j},
\label{eq:ell}
\end{equation}
and for $x<x_1$ or $x>x_m$, $l_m(x)=-\infty$. Thus, for all
$x\in D$, we have, due to concavity of $h(\cdot)$,
\begin{equation}
\ell_m(x)\leq h(x)\leq u_m(x)
\label{eq:squeeze}
\end{equation}
To sample using ARS, draw $x^*\sim s_m$ and $w\sim Uniform(0,1)$ independently
and accept $x^*$ if $w\leq \exp\{\ell_m(x^*)-u_m(x^*)\}$. Else accept $x^*$
if $w\leq \exp\{h(x^*)-u_m(x^*)\}$. If the sampling is to be continued then the
accepted values may be included in the set of abscissae (the latter to be re-arranged
in ascending order) to adaptively make the bounds (\ref{eq:squeeze}) finer and finer;
this enhances efficiency as the sampling progresses.
For our purpose of perfect sampling using the log-concave full conditionals
we shall need to represent the Gibbs transition kernel in a special form using the lower bound of the form
given in (\ref{eq:squeeze}), while using ARS in conjunction for sampling.
We introduce our perfect sampling idea in the next section.
\subsection{Construction of perfect simulation methodology in posteriors with log-concave full conditionals}
\label{subsec:perfect_logconcave}
For the sake of generality, we consider full conditionals of the form
$\pi_i(\xi_i)=\pi(\xi_i\mid \boldsymbol{\xi}_{-i});i=1,\ldots,p$, where it is necessary to simulate
perfectly from the joint distribution of $\boldsymbol{\xi}=\{\xi_1,\ldots,\xi_p\}$;
here $\boldsymbol{\xi}_{-i}=\boldsymbol{\xi}\backslash \xi_i$. We
assume that each $\pi_i(\xi_i)$ is log-concave.
It then follows from (\ref{eq:squeeze}) that
\begin{equation}
\pi_i(\xi)\geq\exp\{\ell_{m,\boldsymbol{\xi}_{-i}}(\xi)\},
\label{eq:lb1}
\end{equation}
where $\ell_{m,\boldsymbol{\xi}_{-i}}(\cdot)$ may depend upon $\boldsymbol{\xi}_{-i}$. Taking the infimum over $\boldsymbol{\xi}_{-i}$ yields
\begin{equation}
\pi_i(\xi)\geq\exp\{\ell_{m,\boldsymbol{\xi}_{-i}}(\xi)\}\geq\inf_{\boldsymbol{\xi}_{-i}}\exp\{\ell_{m,\boldsymbol{\xi}_{-i}}(\xi)\}=\exp\{\ell_{m,i}(\xi)\},
\label{eq:lb2}
\end{equation}
where $\exp\{\ell_{m,i}(\xi)\}=\inf_{\boldsymbol{\xi}_{-i}}\exp\{\ell_{m,\boldsymbol{\xi}_{-i}}(\xi)\}$ is independent of $\boldsymbol{\xi}_{-i}$.
However, the right hand side of (\ref{eq:lb2}) need not be a density in that it need not integrate to one.
Firstly, finiteness of the integral can be ensured
at least if $\boldsymbol{\xi}$ is restricted to a compact set. That restriction of the support of the
parameters to some judicously constructed compact set is not unrealistic is discussed in detail in \ctn{Sabya10b}.
Let $D_i$ denote a compact interval to which $\xi_i$ is restricted. Let $\epsilon_i=\int_{D_i}\exp\{\ell_{m,i}(\xi)\}$,
and let $g_{m,i}(\xi)=\epsilon^{-1}_i\exp\{\ell_{m,i}(\xi)\}$ denote the density corresponding to $\exp\{\ell_{m,i}(\xi)\}$.
Then, we have, for each $i=1,\ldots,p$,
\begin{equation}
\pi_i(\xi)\geq\epsilon_ig_{m,i}(\xi),
\label{eq:lb3}
\end{equation}
which implies that
\begin{equation}
\prod_{i=1}^p\pi_i(\xi_i)\geq\prod_{i=1}^p\epsilon_ig_{m,i}(\xi_i)=\epsilon g_m(\boldsymbol{\xi}),
\label{eq:lb4}
\end{equation}
where $\epsilon=\prod_{i=1}^p\epsilon_i$ and $g_m(\boldsymbol{\xi})=\prod_{i=1}^pg_{m,i}(\xi_i)$.
That $0<\epsilon\leq 1$ is clear since for each $i=1,\ldots,p$, $0<\epsilon_i\leq 1$, the latter following
by integrating both sides of (\ref{eq:lb3}) over the support $D_i$.
This then implies that the Gibbs transition kernel, given by
\begin{equation}
P(\boldsymbol{\xi}^{(t+1)}\mid\boldsymbol{\xi}^{(t)})=\prod_{i=1}^p
\pi(\xi^{(t+1)}_i\mid\xi^{(t+1)}_1,\ldots,\xi^{(t+1)}_{i-1},\xi^{(t)}_{i+1},\ldots,\xi^{(t)}_p),
\label{eq:gibbs_kernel}
\end{equation}
can be represented as
\begin{equation}
P(\boldsymbol{\xi}^{(t+1)}\mid\boldsymbol{\xi}^{(t)})=\epsilon g_m(\boldsymbol{\xi}^{(t+1)})+(1-\epsilon)R_m(\boldsymbol{\xi}^{(t+1)}\mid\boldsymbol{\xi}^{(t)}),
\label{eq:rep1}
\end{equation}
where
\begin{equation}
R_m(\boldsymbol{\xi}^{(t+1)}\mid\boldsymbol{\xi}^{(t)})=\frac{P(\boldsymbol{\xi}^{(t+1)}\mid\boldsymbol{\xi}^{(t)})-\epsilon g_m(\boldsymbol{\xi}^{(t+1)})}{1-\epsilon}
\label{eq:residual}
\end{equation}
is the residual density.
Hence, in order to simulate the Gibbs chain $P(\boldsymbol{\xi}^{(t+1)}\mid\boldsymbol{\xi}^{(t)})$ one can first draw $\delta^{(t+1)}\sim Bernoulli(\epsilon)$;
if $\delta^{(t+1)}=1$, then $\boldsymbol{\xi}^{(t+1)}$ is drawn from $g_m(\cdot)$, and if $\delta^{(t+1)}=0$, $\boldsymbol{\xi}^{(t+1)}\sim R_m(\cdot\mid\boldsymbol{\xi}^{(t)})$.
Thus, if $\delta^{(t+1)}=1$, then $\boldsymbol{\xi}^{(t+1)}$ is drawn from $g_m(\cdot)$ which does not depend upon the previous iteration $\boldsymbol{\xi}^{(t)}$.
We shall exploit this fact for our perfect sampling algorithm. Indeed, we constructed the mixture representation (\ref{eq:rep1})
just to achieve this independence of $\boldsymbol{\xi}^{(t)}$ which happens with positive probability $\epsilon$. This implies that once
$\delta^{(t)}=1$ for some $t<0$ in the associated coupling from the past algorithm (CFTP), all possible chains initialised
at all possible values of the state-space, will coalesce!
But since drawing directly from $R_m(\cdot\mid\boldsymbol{\xi}^{(t)})$ (necessary when $\delta^{(t)}=0$) is not straightforward, we consider
a rejection sampling method which we now describe. Note that
\begin{align}
R_m(\boldsymbol{\xi}^{(t+1)}\mid\boldsymbol{\xi}^{(t)})&=\frac{P(\boldsymbol{\xi}^{(t+1)}\mid\boldsymbol{\xi}^{(t)})-\epsilon g_m(\boldsymbol{\xi}^{(t+1)})}{1-\epsilon}\label{eq:rs1}\\
&\leq \frac{P(\boldsymbol{\xi}^{(t+1)}\mid\boldsymbol{\xi}^{(t)})}{1-\epsilon}\label{eq:rs2}
\end{align}
Hence, we consider the rejection sampling scheme provided in Algorithm \ref{algo:rs1}.
\begin{algo}\label{algo:rs1}\hrule height 1pt width \textwidth \vspace*{2pt}
Rejection sampling from $R_m(\cdot\mid\boldsymbol{\xi}^{(t)})$ \vspace*{2pt}\hrule height 1pt width \textwidth \vspace*{4pt} \normalfont \ttfamily
\begin{itemize}
\item[(1)] Draw $\boldsymbol{\xi}\sim P(\cdot\mid\boldsymbol{\xi}^{(t)})$ by sampling from the full conditionals,
and independently draw $U\sim Uniform(0,1)$.
\item[(2)] Accept $\boldsymbol{\xi}$ as a realization from $R_m(\cdot\mid\boldsymbol{\xi}^{(t)})$ if
$$U<(1-\epsilon)\frac{R_m(\boldsymbol{\xi}\mid\boldsymbol{\xi}^{(t)})}{P(\boldsymbol{\xi}\mid\boldsymbol{\xi}^{(t)})}.$$
\end{itemize}
\rmfamily
\vspace*{2pt}\hrule height 1pt width \textwidth \vspace*{4pt}
\end{algo}
Note that sampling from $P(\cdot\mid\boldsymbol{\xi}^{(t)})$ may require ARS from the full conditionals.
To avoid ARS one may further upper bound $P(\boldsymbol{\xi}\mid\boldsymbol{\xi}^{(t)})$ using the upper bounds
available for log-concave densities as follows.
\begin{equation}
P(\boldsymbol{\xi}\mid\boldsymbol{\xi}^{(t)})\leq\eta f_m(\boldsymbol{\xi}),
\label{eq:ub1}
\end{equation}
where $\eta=\prod_{i=1}^p\eta_i$,
$\eta_i=\int_{D_i}\exp\{u_{m,i}(\xi)\}d\xi$,
$u_{m,i}(\xi)=\sup_{\boldsymbol{\xi}_{-i}}u_{m,\boldsymbol{\xi}_{-i}}(\xi)$,
and
\begin{equation}
f_m(\boldsymbol{\xi})=\prod_{i=1}^pf_{m,i}(\xi_i),
\label{eq:ub2}
\end{equation}
with $f_{m,i}(\xi)=\eta^{-1}_i\exp\{u_{m,i}(\xi)\}$.
Since we also have the lower bound $P(\boldsymbol{\xi}\mid\boldsymbol{\xi}^{(t)})\geq\epsilon g_m(\boldsymbol{\xi})$, the following
rejection sampling method given by Algorithm (\ref{algo:rs2}) can be employed to sample from $P(\cdot\mid\boldsymbol{\xi}^{(t)})$.
\begin{algo}\label{algo:rs2}\hrule height 1pt width \textwidth \vspace*{2pt}
Rejection sampling from $P(\cdot\mid\boldsymbol{\xi}^{(t)})$ \vspace*{2pt}\hrule height 1pt width \textwidth \vspace*{4pt} \normalfont \ttfamily
\begin{itemize}
\item[(1)] Draw $\boldsymbol{\xi}\sim f_m(\cdot)$,
and independently draw $U\sim Uniform(0,1)$.
\item[(2)] Accept $\boldsymbol{\xi}$ as a realization from $P(\cdot\mid\boldsymbol{\xi}^{(t)})$ if
$$U<\frac{\epsilon g_m(\boldsymbol{\xi})}{\eta f_m(\boldsymbol{\xi})}.$$
\item[(3)] Else accept $\boldsymbol{\xi}$ as a realization from $P(\cdot\mid\boldsymbol{\xi}^{(t)})$ if
$$U<\frac{P(\boldsymbol{\xi}\mid\boldsymbol{\xi}^{(t)})}{\eta f_m(\boldsymbol{\xi})}.$$
\end{itemize}
\rmfamily
\vspace*{2pt}\hrule height 1pt width \textwidth \vspace*{4pt}
\end{algo}
It is important to remark that whenever simulation from $P(\cdot\mid\boldsymbol{\xi}^{(t)})$ is straightforward,
that is, whenever the full conditionals $\pi_i(\xi_i)$ are of standard forms, rejection sampling
or ARS will not be used for sampling from the Gibbs kernel.
Our mixture Gibbs kernel (\ref{eq:rep1}) resembles that associated with the ``multigamma coupler"
of \ctn{Murdoch98}, but the latter is a representation of one-dimensional cases only. Moreover, such
mixture representation is very rarely achievable in reality for densities that are not log-concave. Perfect
simulation of high-dimensional variables, using the one-dimensional multigamma coupler for each univariate full conditional
densitiy, is possible in principle, but is likely to be extremely inefficient, particularly as
the dimension of the random variable tends to be large.
As is evident from our construction, we completely
bypass such difficulties by representing the Gibbs kernel of the entire high-dimensional random variable
$\boldsymbol{\xi}$ as a mixture of two (high-dimensional) densities, obtained using properties of log-concavity
of the full conditionals. We have also shown how to sample from the two high-dimensional densities.
In particular, we have provided an explicit rejection sampling method for simulating from the
residual density of the mixture representation, whatever the dimensionality. We remark that explicit
methods of simulating from the residual density has not been provided in \ctn{Murdoch98}
or \ctn{Green99}. Although \ctn{Mykland95} proposed a trick to completely avoid simulation
from the residual density in the context of regenerative simulation, such trick is not
applicable in perfect simulation.
For perfect simulation we exploit the following idea first presented in \ctn{Murdoch98}.
Note that there is a fixed probability $\epsilon$ that at any given time $T=t$, $\boldsymbol{\xi}$ will be drawn
from $g_m(\cdot)$. Hence, $T$ follows a geometric distribution given by
$P(T=t)=\epsilon (1-\epsilon)^t$; $t=0,1,2,\ldots$. As a result, it is possible to simulate
$T$ from the geometric distribution and then draw $\boldsymbol{\xi}^{(-T)}\sim g_m(\cdot)$. Then the chain only
need to be carried forward in time till time $t=0$, using $\boldsymbol{\xi}^{(t+1)}=\psi(\boldsymbol{\xi}^{(t)},\boldsymbol{U}^{(t+1)})$,
where $\psi(\boldsymbol{\xi}^{(t)},\boldsymbol{U}^{(t+1)})$ is the deterministic function corresponding to the simulation
of $\boldsymbol{\xi}^{(t+1)}$ from $R_m(\cdot\mid\boldsymbol{\xi}^{(t)})$ using the set of appropriate random numbers $\boldsymbol{U}^{(t+1)}$;
the sequence $\{\boldsymbol{U}^{(t)};t=0,-1,-2,\ldots\}$ being assumed to be available before beginning the perfect
sampling simulation. The resulting draw $\boldsymbol{\xi}^{(0)}$ sampled at time $t=0$ is a perfect sample
from $\pi(\boldsymbol{\xi})$. For subsequent reference we present this in an algorithmic way in Algorithm \ref{algo:perfect_logconcave}.
\begin{algo}\label{algo:perfect_logconcave}\hrule height 1pt width \textwidth \vspace*{2pt}
Perfect simulation from $\pi(\boldsymbol{\xi})$ \vspace*{2pt}\hrule height 1pt width \textwidth \vspace*{4pt} \normalfont \ttfamily
\begin{itemize}
\item[(1)] Draw $T\sim Geometric(\epsilon)$.
\item[(2)] Draw $\boldsymbol{\xi}^{(-T)}\sim g_m(\cdot)$.
\item[(3)] Carry the chain forward till time $t=0$ using the deterministic functional
relationship $\boldsymbol{\xi}^{(t+1)}=\psi(\boldsymbol{\xi}^{(t)},\boldsymbol{U}^{(t+1)})$ and the available sequence
$\{\boldsymbol{U}^{(t)};t=0,1,2,\ldots\}$.
\item[(4)] Report $\boldsymbol{\xi}^{(0)}$ as a perfect sample from $\pi(\boldsymbol{\xi})$.
\end{itemize}
\rmfamily
\vspace*{2pt}\hrule height 1pt width \textwidth \vspace*{4pt}
\end{algo}
The above perfect sampling algorithm is to be embedded in the perfect sampling algorithm
for mixture simulation in the context of categorical time series. This we do in the next section.
\section{Perfect simulation for mixtures of categorical time series with unknown number of components}
\label{sec:perfect_categorical}
We first note that coalescence of $(Z,C,S)$ (equivalently, coalescence of $(Z,C)$ since
coalescence of $C$ implies coalescence of $S$) implies coalescence of $(\boldsymbol{\Phi},\boldsymbol{\gamma})$.
We exploit the bounding chains construction approach of \ctn{Sabya10b} for facilitating coalescence.
The idea is to obtain stochastic lower and upper bounds for the discrete parts of the Gibbs sampler, namely
for $(Z,C,S)$, by maximizing and minimizing their respective distribution functions
with respect to the continuous parameters, simulating only from the lower and the upper bounding chains thus
created, and noting their coalescence.
Remarkably, there is no need to simulate the continuous parameters $\boldsymbol{\Phi}$ and $\boldsymbol{\gamma}$ before coalescence
of $(Z,C,S)$. Simulation of $(\boldsymbol{\Phi},\boldsymbol{\gamma})$, conditional on $(Z,S)$, is necessary only after the coalescence of
the latter. However, as already discussed in Section \ref{subsec:perfect_need},
methods for directly simulating $(\boldsymbol{\Phi},\boldsymbol{\gamma})$ given $(Z,S)$ are not available, and we must resort
to the perfect simulation method introduced in Section \ref{sec:perfect_logconcave} using the available
full conditionals which are, thankfully, log-concave.
We now proceed to construction of appropriate bounding chains for the discrete parameters $(Z,C,S)$.
\subsection{Bounding chains}
\label{subsec:sb_bounding_chains}
\subsubsection{Bounds for $Z$}
\label{subsubsec:z_bound}
Let $F_{z_i}(\cdot\mid \boldsymbol{Y},S,k,\boldsymbol{\Theta}_M)$ denote
the distribution function of the full conditional of $z_i$, and let $F_{c_j}(\cdot\mid \boldsymbol{Y}, S_{-j},k_j,\boldsymbol{\Phi})$ and
$F_{s_j}(\cdot\mid \boldsymbol{Y}, S_{-j},C,\boldsymbol{\Theta}_M)$ stand for those of $c_j$ and $s_j$, respectively.
In addition, when required, we shall assume that
$\{\gamma_{st}\}$ have compact supports not containing zero. This assumption entails multiplication
of a constant to the prior to take care of the truncation, but clearly this constant does not destroy
the log-concavity of the full conditional of $\gamma_{st}$. On the other hand, truncation of $\phi_{\ell,st}$
would involve a factor that depends upon $\gamma_{st}$, which might affect log-concavity of $\gamma_{st}$.
However, we did not find truncation of $\phi_{\ell,st}$ to be necessary in our simulations.
Letting $\bar S$ denote the set consisting of only those $s_j$ that have coalesced, and let $S^-=S\backslash\bar S$ consist of the
remaining $s_j$. Then
\begin{eqnarray}
F^L_{z_i}\left(\cdot\mid \boldsymbol{Y},\bar S\right)&=&\inf_{S^-,k,\boldsymbol{\Phi}}F_{z_i}(\cdot\mid \boldsymbol{Y},\bar S,S^-,k,\boldsymbol{\Phi})\label{eq:inf_z}\\
F^U_{z_i}\left(\cdot\mid \boldsymbol{Y},\bar S\right)&=&\sup_{S^-,k,\boldsymbol{\Phi}}F_{z_i}(\cdot\mid \boldsymbol{Y},\bar S,S^-,k,\boldsymbol{\Phi})\label{eq:sup_z}
\end{eqnarray}
Fixing $\bar S$ helps reduce the gap between (\ref{eq:inf_z}) and (\ref{eq:sup_z}).
As in \ctn{Sabya10b} we calculate the infimum and the supremum above by simulated annealing.
For further details, wee \ctn{Sabya10b}.
\subsubsection{Bounds for $C$}
\label{subsubsec:c_bound}
Let $\bar Z$ denote the set of coalesced $z_i$, and let $Z^-=Z\backslash\bar Z$ consist of those $z_j$ that did not yet coalesce.
Then
\begin{eqnarray}
F^L_{c_j}\left(\cdot\mid \boldsymbol{Y},\bar S, \bar Z\right)&=&\inf_{S^-,k_j,Z^-,\boldsymbol{\Phi}}F_{c_j}(\cdot\mid \boldsymbol{Y},\bar S,S^-,k_j,\bar Z,Z^-,\boldsymbol{\Phi})\label{eq:inf_c}\\
F^U_{c_j}\left(\cdot\mid \boldsymbol{Y},\bar S, \bar Z\right)&=&\sup_{S^-,k_j,Z^-,\boldsymbol{\Phi}}F_{c_j}(\cdot\mid \boldsymbol{Y},\bar S,S^-,k_j,\bar Z,Z^-,\boldsymbol{\Phi})\label{eq:sup_c}
\end{eqnarray}
As noted in \ctn{Sabya10b}, the supremum corresponds to $k_j=1$ and the infimum corresponds to $k_j=M-1$.
For details on optimization using simulated annealing, see \ctn{Sabya10b}.
\subsubsection{Bounds for $S$}
\label{subsubsec:s_bound}
Letting $\bar C$ and $C^-=C\backslash\bar C$ denote the sets of coalesced and the non-coalesced $c_j$, the lower
and the upper bounds for the distribution function of $s_j$ are
\begin{eqnarray}
F^L_{s_j}\left(\cdot\mid \boldsymbol{Y},\bar C\right)&=&\inf_{C^-,\boldsymbol{\Phi}}F_{s_j}(\cdot\mid \boldsymbol{Y},\bar C,C^-,\boldsymbol{\Phi})\label{eq:inf_s}\\
F^U_{s_j}\left(\cdot\mid \boldsymbol{Y},\bar C \right)&=&\sup_{C^-,\boldsymbol{\Phi}}F_{s_j}(\cdot\mid \boldsymbol{Y},\bar C,C^-,\boldsymbol{\Phi})\label{eq:sup_s}
\end{eqnarray}
Optimization in this case requires careful attention; see \ctn{Sabya10b} for details.
\begin{algo}\label{algo:cftp_unknown}\hrule height 1pt width \textwidth \vspace*{2pt}
CFTP for mixtures with unknown number of components \vspace*{2pt}\hrule height 1pt width \textwidth \vspace*{4pt} \normalfont \ttfamily
\begin{itemize}
\item[(1)] For $j=1\ldots$, until coalescence of $(Z,C)$, repeat steps (2) and (3) below.
\item[(2)] Define $\mathcal S_j=\{-2^j+1,\ldots,-2^{j-1}\}$ for $j\geq 2$,
and let $\mathcal S_1=\{-1,0\}$.
For each $m\in\mathcal S_j$,
generate random numbers $R_{Z,m}$, $R_{C,m}$, $R_{S,m}$, $R_{\boldsymbol{\Theta}_M,m}$, and $R_{\boldsymbol{\gamma},m}$, meant
for simulating $Z$, $C$, $S$, $\boldsymbol{\Theta}_M$, and $\boldsymbol{\gamma}$ respectively.
Note that for each $m$, $R_{\boldsymbol{\Theta},m}$ and $R_{\boldsymbol{\gamma},m}$ are random numbers
corresponding to the perfect simulation algorithm given by Algorithm \ref{algo:perfect_logconcave};
$(\boldsymbol{\Phi},\boldsymbol{\gamma})$ in this problem corresponds to the random vector $\boldsymbol{\xi}$ in that algorithm.
We need to generate and fix these random numbers even though we won't actually simulate $(\boldsymbol{\Phi},\boldsymbol{\gamma})$
before coalescence of $(Z,C)$.
Once generated, treat the random numbers as fixed thereafter for all iterations.
Since step $-2^j$ is the initializing step, no random number generation is required
at this step.
\item[(3)] For $t=-2^j+1,\ldots,-1,0$, implement steps (3) (i), (3) (ii) and (3) (iii):
\begin{itemize}
\item[(i)] For $i=1,\ldots,n$,
\begin{itemize}
\item[(a)] For $\ell=1,\ldots,M$, calculate $F^L_{z_i}(\ell\mid \boldsymbol{Y},\bar S)$ and
$F^U_{z_i}(\ell\mid \boldsymbol{Y},\bar S)$ using the simulated annealing techniques detailed
in \ctn{Sabya10b}.
\item[(b)] Determine $z^L_{it}=F^{U-}_{z_i}(R_{z{i,t}}\mid \boldsymbol{Y},\bar S)$ and
$z^U_{it}=F^{L-}_{z_i}(R_{z{i,t}}\mid \boldsymbol{Y},\bar S)$.
\end{itemize}
\item[(ii)] For $i=1,\ldots,M$,
\begin{itemize}
\item[(a)] For $\ell=1,\ldots,k_i+1$, calculate $F^L_{c_i}(\ell\mid \boldsymbol{Y},\bar S,\bar Z)$ and
$F^U_{c_i}(\ell\mid \boldsymbol{Y},\bar S,\bar Z)$, using the simulated annealing techniques
of \ctn{Sabya10b}. The supremum corresponds to $k_i=\#\bar S\backslash\{s_i\}$, when $S^-$ contains
a single distinct element, and the infimum corresponds to the case where
$k_i=\# \left(\bar S\cup S^-\right)\backslash\{s_i\}$, when all elements of $S^-$ are distinct, and so
the set $S^-$ will be set manually to have a single distinct element or all distinct elements.
\item[(b)] Set $c^L_{it}=F^{U-}_{c_i}(R_{c_{i,t}}\mid \boldsymbol{Y},\bar S,\bar Z)$ and
$c^L_{it}=F^{U-}_{c_i}(R_{c_{i,t}}\mid \boldsymbol{Y},\bar S,\bar Z)$.
\end{itemize}
\item[(iii)] For $i=1,\ldots,M$,
\begin{itemize}
\item[(a)] For $\ell=1,\ldots,M$, calculate $F^L_{s_i}(\ell\mid \boldsymbol{Y},\bar C)$ and
$F^U_{s_i}(\ell\mid \boldsymbol{Y},\bar \boldsymbol{C})$, using simulated annealing techniques detailed in \ctn{Sabya10b}.
\item[(b)]
Since, for some $\ell^*\in\{1,\ldots,M\}$,
$F^L_{s_i}(\ell\mid \boldsymbol{Y},\bar C)=0$ for $\ell<\ell^*$ and 1 for $\ell\geq \ell^*$,
it follows that $s^L_{it}=\ell^*$. Similarly, $s^U_{it}$ can be determined.
\end{itemize}
\end{itemize}
\item[(4)] If, for some $t^*<0$, $z^L_{it^*}=z^U_{it^*}$ $\forall i=1,\ldots,n$, and $c^L_{it^*}=c^U_{it^*}$
$\forall i=1,\ldots,M$, then
run the following Gibbs sampling steps from $t=t^*$ to $t=0$:
\begin{itemize}
\item[(a)] Let $Z^*=\{z^*_1,\ldots,z^*_n\}$ and $C^*=\{c^*_1,\ldots,c^*_M\}$ denote the coalesced values of $Z$
and $C$ respectively, at time $t^*$.
Given $(Z^*,C^*)$, arbitrarily choose any value of $\boldsymbol{\Theta}_M$ which is compatible with $C^*$
(one way to ensure compatibility is to choose any $\boldsymbol{\Theta}_M$ having $M$ distinct elements); then
obtain $S^*$ from $[S\mid\boldsymbol{Y},C,\boldsymbol{\Theta}_M]$ using the algorithm given in Section \ref{subsec:relabeling}.
\item[(b)] Finally, generate $(\boldsymbol{\Phi},\boldsymbol{\gamma})$ using the perfect simulation algorithm
described in Algorithm \ref{algo:perfect_logconcave}, using the random numbers
already generated.
This yields the coalesced value $(Z^*,C^*,S^*,\boldsymbol{\Phi}^*,\boldsymbol{\gamma}^*)$ at time $t=t^*$.
\item[(b)] Using the random numbers already generated, carry forward the above Gibbs sampling
chain started at $t=t^*$ till $t=0$,
simulating, in order, from the full conditionals of the individual components of $(Z,C,S)$, provided in
Sections \ref{subsec:fullcond_z}, \ref{subsec:fullcond_c}, and \ref{subsec:fullcond_s}, and by perfectly simulating
$(\boldsymbol{\Phi},\boldsymbol{\gamma})$ using Algorithm \ref{algo:perfect_logconcave}. Note that, once $(\boldsymbol{\Phi}^*,\boldsymbol{\gamma}^*)$
are generated by perfect sampling at time $t=t^*$, further perfect sampling of $(\boldsymbol{\Phi},\boldsymbol{\gamma})$ for time $t>t^*$
does not seem necessary since now Gibbs sampling can be employed. But somewhat ironically, we are forced
to continue perfect sampling since changing the simulation method in the midway is not legitimate.
\item[(c)]Then, the output of the Gibbs sampler obtained at $t=0$,
which we denote by $(Z_0,C_0,S_0,\boldsymbol{\Phi}_0,\boldsymbol{\gamma}_0)$, is a perfect sample
from the true target posterior distribution.
\end{itemize}
\end{itemize}
\rmfamily
\vspace*{2pt}\hrule height 1pt width \textwidth \vspace*{4pt}
\end{algo}
\input{catsupp}
\section{Marginalized full conditional distributions}
\label{sec:marginalized_fullcond}
\subsection{Marginalized full conditionals of $\{z_1,\ldots,z_n\}$}
\label{subsec:marginalized_fullcond_z}
For $i^*=1,\ldots,n$, let $Z_{i^*}=\{z_1,\ldots,z_{i^*-1},z_{i^*+1},\ldots,z_n\}$, and
let $C$ consist of $k$ distinct components. Then, denoting the set $\{\gamma_{st};s,t=1,\ldots,K\}$
by $\boldsymbol{\gamma}$, the full conditional distribution of $z_{i^*}$ is given by
\begin{align}
[z_{i^*}=r\mid Z_{-i^*},C,\boldsymbol{\gamma},k]&\propto \prod_{\ell=1}^k\prod_{s=1}^K
\frac{\prod_{t=1}^K\Gamma\left(\sum_{i:z_i=j}\sum_{j:c_j=\ell}N_{i,st}+\gamma_{st}\right)}
{\Gamma\left(\sum_{t=1}^K\sum_{i:z_i=j}\sum_{j:c_j=\ell}N_{i,st}+\sum_{t=1}^K\gamma_{st}\right)}
\label{eq:marginalized_fullcond_z}
\end{align}
In the right hand side of (\ref{eq:fullcond_z}), $z_{i^*}$ must be replaced with $r$.
\subsection{Marginalized full conditionals of $\{c_1,\ldots,c_M\}$}
\label{subsec:marginalized_fullcond_c}
To obtain the full conditional of $c_r;r=1,\ldots,M$, first let $k_{r}$ denote the number of distinct values in
$\boldsymbol{\Theta}_{-rM}=\{\boldsymbol{\theta}_1,\ldots,\boldsymbol{\theta}_{r-1},\boldsymbol{\theta}_{r+1},\ldots,\boldsymbol{\theta}_M\}$, and let
$\boldsymbol{\phi}^{(r^*)}_{\ell}$; $\ell=1,\ldots,k_{r}$ denote
the distinct values.
Also suppose that $\boldsymbol{\phi}^{(r^*)}_{\ell}$ occurs $M_{\ell r}$ times.
Then the conditional distribution of $c_r$ is given by
\begin{equation}
[c_r=\ell\mid Y,Z,C_{-r},\boldsymbol{\gamma},k_r]=\left\{\begin{array}{c}\kappa q_{\ell r}\hspace{2mm}\mbox{if}\hspace{2mm}
\ell=1,\ldots,k_r\\ \kappa q_{0r}\hspace{2mm}\mbox{if}\hspace{2mm}\ell=k_r+1\end{array}\right.
\label{eq:marginalized_fullcond_c}
\end{equation}
where
\begin{align}
q_{\ell r}&\propto M_{\ell r}\times
\prod_{s=1}^K
\frac{\Gamma\left(\sum_{t=1}^K\sum_{i:z_i=j,j:c_j=\ell,j\neq r}N_{i,st}+\sum_{t=1}^K\gamma_{st}\right)}
{\prod_{t=1}^K\Gamma\left(\sum_{i:z_i=j,j:c_j=\ell,j\neq r}N_{i,st}+\gamma_{st}\right)}\notag\\
&\times \prod_{s=1}^K
\frac{\prod_{t=1}^K\Gamma\left(\sum_{i:z_i=r}N_{i,st}+\sum_{i:z_i=j,j:c_j=\ell,j\neq r}N_{i,st}+\gamma_{st}\right)}
{\Gamma\left(\sum_{t=1}^K\sum_{i:z_i=r}N_{i,st}+\sum_{t=1}^K\sum_{i:z_i=j,j:c_j=\ell,j\neq r}N_{i,st}+\sum_{t=1}^K\gamma_{st}\right)}
\label{eq:marginalized_fullcond_c1}
\end{align}
and
\begin{align}
q_{0 r}&\propto \alpha\times
\prod_{s=1}^K
\frac{\Gamma\left(\sum_{t=1}^K\gamma_{st}\right)}
{\prod_{t=1}^K\Gamma\left(\gamma_{st}\right)}
\times\prod_{s=1}^K
\frac{\prod_{t=1}^K\Gamma\left(\sum_{i:z_i=r}N_{i,st}+\gamma_{st}\right)}
{\Gamma\left(\sum_{t=1}^K\sum_{i:z_i=r}N_{i,st}+\sum_{t=1}^K\gamma_{st}\right)}
\label{eq:marginalized_fullcond_c2}
\end{align}
\subsection{Marginalized full conditionals of $\{\gamma_{st};s,t=1,\ldots,K\}$}
\label{subsec:marginalized_fullcond_gamma}
Assuming that $C$ consists of $k$ distinct components, the full
conditional distribution of $\gamma_{s^*t^*}$, for
$s^*=1,\ldots,K$, and $t^*=1,\ldots,K$, is given by
\\[2mm]
$[\gamma_{\ell,s^*t^*}\mid Z,C,\boldsymbol{\gamma}_{-s^*,-t^*},k]$
\begin{align}
&\propto
\left\{\prod_{\ell=1}^k\frac{\Gamma\left(\sum_{i:z_i=j}\sum_{j:c_j=\ell}N_{i,s^*t^*}+\gamma_{s^*t^*}\right)}
{\Gamma\left(\sum_{t=1}^K\sum_{i:z_i=j}\sum_{j:c_j=\ell}N_{i,s^*t}+\sum_{t=1}^K\gamma_{s^*t}\right)}\right\}
\times
\left(\frac{\Gamma\left(\sum_{t=1}^K\gamma_{s^*t}\right)}
{\Gamma\left(\gamma_{s^*t^*}\right)}\right)^k\notag\\
&\times \gamma^{a_{jk}-1}_{s^*t^*}\exp\left(-b_{jk}\gamma_{s^*t^*}\right)\notag\\
&= \prod_{\ell=1}^k\left\{\frac{\Gamma\left(\sum_{i:z_i=j}\sum_{j:c_j=\ell}N_{i,s^*t^*}+\gamma_{s^*t^*}\right)}
{\Gamma\left(\sum_{t=1}^K\sum_{i:z_i=j}\sum_{j:c_j=\ell}N_{i,s^*t}+\sum_{t=1}^K\gamma_{s^*t}\right)}
\times
\frac{\Gamma\left(\sum_{t=1}^K\gamma_{s^*t}\right)}
{\Gamma\left(\gamma_{s^*t^*}\right)}\right\}\label{eq:gamma1}\\
&\times \gamma^{a_{jk}-1}_{s^*t^*}\exp\left(-b_{jk}\gamma_{s^*t^*}\right)\notag
\end{align}
\subsection{Discussion on log-concavity of the full conditionals of $\{\gamma_{st};s,t=1,\ldots,K\}$}
\label{subsec:gamma_logconcavity}
Note that each factor in the product (\ref{eq:gamma1}) is of the form
\begin{equation}
h(\gamma_{s^*t^*})=\frac{\Gamma(\gamma_{s*t^*}+a_{t^*})}{\Gamma(\gamma_{s^*t^*})}
\times\frac{\Gamma(\gamma_{s^*t^*}+y_{\ell,t^*})}{\Gamma(\gamma_{s^*t^*}+y_{\ell,t^*}+a_{t^*}+b_{\ell,t^*})},
\label{eq:gamma2}
\end{equation}
$a_{t^*}=\sum_{t=1,t\neq t^*}^K\gamma_{s^*t}$, $y_{\ell,t}=\sum_{i:z_i=j}\sum_{j:c_j=\ell}N_{i,s*t}$ $\forall t=1,\ldots,K$,
and $b_{\ell,t^*}=\sum_{t=1,t\neq t^*}^Ky_{\ell,t}$. Clearly, all the terms are non-negative, with $y_{\ell,t}$ and $b_{\ell,t^*}$
being integers.
As a result, $h(\gamma_{s^*t^*})$ admits the following simple form:
\begin{equation}
h(\gamma_{s^*t^*})=\frac{\prod_{i=1}^{y_{t^*}}(\gamma_{s^*t^*}+y_{t^*}-i)}{\prod_{i=1}^{y_{t^*}+b_{\ell,t^*}}
(\gamma_{s^*t^*}+a_{t^*}+y_{t^*}+b_{\ell,t^*}-i)}.
\label{eq:gamma3}
\end{equation}
Thus,
\begin{align}
\frac{d^2\log h(\gamma_{s^*t^*})}{d\gamma^2_{s^*t^*}}
&=-\sum_{i=1}^{y_{t^*}}\left\{\frac{1}{\left(\gamma_{s^*t^*}+y_{t^*}-i\right)^2}
-\frac{1}{\left(\gamma_{s^*t^*}+a_{t^*}+y_{t^*}-i\right)^2}\right\}\notag\\%\label{eq:gamma4}\\
& \ \ \ \ +\sum_{i=y_{t^*}+1}^{y_{t^*}+b_{\ell,t^*}}\frac{1}{\left(\gamma_{s^*t^*}+a_{t^*}+y_{t^*}-i\right)^2}.
\label{eq:gamma5}
\end{align}
Unless $b_{\ell,t^*}=0$ (that is, $y_{\ell,t}=0$ $\forall t\neq t^*$), (\ref{eq:gamma5}) need not be negative for all
$\gamma_{s^*t^*}$ and $\boldsymbol{\gamma}_{-s^*,-t^*}$.
|
\section{Introduction}
The well-known problem $k$-SAT is NP-complete for $k\geq 3$. If P$\not=$NP, $k$-SAT does not have a polynomial time algorithm. For a CNF formula $F$ over $n$ variables, the naive approach of trying all satisfying assignments takes time $O(2^n\cdot \operatorname{poly}(|F|))$.
Especially for $k=3$ much work has been put into finding so-called ``moderately exponential time'' algorithms running in time $O(2^{cn})$ for some $c<1$. In 1998, Paturi, Pudl\'ak, Saks, and Zane presented a randomized algorithm for 3-SAT that runs in time $O(1.364^n)$. Given the promise that the formula has at most one satisfying assignment (that problem is called Unique 3-SAT), a running time of $O(1.308^n)$ was shown. Both bounds were the best known when published. The running time of general $3$-SAT has been improved repeatedly (e.g.~\cite{schoning1999,it04}), until PPSZ was shown to run in time $O(1.308^n)$ for general 3-SAT~\cite{hertli11}.
Any further improvement of 3-SAT further also improves Unique 3-SAT; however that bound has not been improved upon since publication of the PPSZ algorithm. In this paper, we present a randomized algorithm for Unique 3-SAT with exponentially better bounds than what could be shown for PPSZ. Our algorithm builds on PPSZ and improves it by treating sparse and dense formulas differently.
A key concept of the PPSZ analysis is the so-called critical clause: We call a clause critical for a variable $x$ if exactly one literal is satisfied by this unique satisfying assignment, and that literal is over $x$. It is not hard to see that the uniqueness of the satisfying assignment implies that every variable has at least one critical clause. If some variables have strictly \emph{more} than one critical clause, then we will give a straightforward proof that PPSZ by itself is faster already. Hence the bottleneck of PPSZ is when every variable has \emph{exactly} one critical clause, and in total there are exactly $n$ critical clauses.
Given a formula with exactly $n$ critical clauses, consider how many other (non-crtical) clauses there are. If there are few, we use an algorithm by Wahlstr\"om~\cite{wahlstroem05} that is faster than PPSZ for formulas with few clauses allover. If there are many non-critical clauses we use the following fact: A non-critical clause has two or more satisfied literals (w.r.t.\ unique satisfying assignment); so after removing a literal, the remaining 2-clause is still satisfied. We will exploit this to improve PPSZ.
An remaining problem is the case if only very few (i.e.\ sublinearly many) variables have more than one critical clause or appear in many (non-critical) clauses. In this case, we would get only a subexponential improvement. A significant part of our algorithm deals with this problem.
\subsection{Notation}
We use the notational framework introduced in \cite{welzl05}.
Let $V$ be a finite set of propositional \emph{variables}. A \emph{literal} $u$ over $x\in V$ is a variable $x$ or a negated variable $\ol{x}$. If $u=\ol{x}$, then $\ol{u}$, the negation of $u$, is defined as $x$. We mostly use $x,y,z$ for variables and $u,v,w$ for literals.
We assume that all lite\-rals are distinct. A \emph{clause} over $V$ is a finite set of lite\-rals over pairwise distinct variables from $V$. By $\mathrm{vbl}(C)$ we denote the set of variables that occur in $C$, i.e.\ $\{x\in V\mid x\in C \vee \ol{x}\in C\}$. $C$ is a $k$-clause if $|C|=k$ and it is a $(\leq k)$-clause if $|C|\leq k$.
A formula in \emph{CNF} (Conjunctive Normal Form) $F$ over $V$ is a finite set of clauses over $V$. We define $\mathrm{vbl}(F):=\bigcup_{C\in F}\mathrm{vbl}(C)$.
$F$ is a k-CNF formula (a $(\leq k)$-CNF formula) if all clauses of $F$ are $k$-clauses ($(\leq k)$-clauses).
A (truth) \emph{assignment} on $V$ is a function $\alpha : V \rightarrow
\{0,1\}$ which assigns a Boolean value to each variable.
$\alpha$ extends to negated variables by letting $\alpha(\ol{x}):=1-\alpha(x)$.
A literal $u$ is \emph{satisfied by} $\alpha$ if $\alpha(u)=1$. A clause is \emph{satisfied by} $\alpha$ if it
contains a satisfied literal and a formula is \emph{satisfied by}
$\alpha$ if all of its clauses are. A formula is \emph{satisfiable} if
there exists a satisfying truth assignment to its variables. A formula that is not satisfiable is called \emph{unsatisfiable}.
Given a CNF formula $F$, we denote by $\mathrm{sat}(F)$ the set of assignments on $\mathrm{vbl}(F)$ that
satisfy $F$. $k$\emph{-SAT} is the decision problem of deciding if a $(\leq k)$-CNF formula has a satisfying assignment.
If $F$ is a CNF formula and $x \in \mathrm{vbl}(F)$, we write $F^{[x\mapsto 1]}$ (analogously $F^{[x\mapsto 0]}$) for the formula arising from removing all clauses containing $x$ and truncating all clauses containing $\ol{x}$ to their remaining literals. This corresponds to assigning $x$ to $1$ (or $0$) in $F$ and removing trivially satifsied clauses. We call assignments $\alpha$ on $V$ and $\beta$ and $W$ \emph{consistent} if $\alpha(x)=\beta(x)$ for all $x\in V\cap W$. If $\alpha$ is an assignment on $V$ and $W\subseteq V$, we denote by $\alpha|_W$ the assignment on $W$ with $\alpha|_W(x)=\alpha(x)$ for $x\in W$.
If $\gamma=\{x\mapsto 0,y\mapsto 1,\dots\}$, we write $F^{[\gamma]}$ as a shorthand for $F^{[x\mapsto 0][y\mapsto 1]\dots}$, the \emph{restriction} of F to $\gamma$.
For a set $W$, we denote by $x\gets_{\text{u.a.r.}} W$ choosing an element $x$ u.a.r. (uniformly at random).
Unless otherwise stated, all random
choices are mutually independent.
We denote by $\log$ the logarithm to the base 2. For the logarithm to
the base $e$, we write $\ln$.
By $\operatorname{poly}(n)$ we denote a polynomial factor depending on $n$.
We use the following convention if no confusion arises: When $F$ is a CNF formula, we denote by $V$ its variables and by $n$ the number of variables of $F$, i.e.\ $V:=\mathrm{vbl}(F)$ and $n:=|\mathrm{vbl}(F)|$. By $o(1)$ we denote a quantity dependent on $n$ going to $0$ with $n\to\infty$.
\subsection{Previous Work}
\begin{definition}
(Promise) Unique 3-SAT is the following promise problem: Given a {{$(\leq 3)$-CNF}} with at most one satisfying assignment, decide if it is satisfiable or unsatisfiable.
A randomized algorithm for Unique 3-SAT is an algorithm that, for a uniquely satisfiable {{$(\leq 3)$-CNF}} formula returns the satisfying assignment with probability $\frac{1}{2}$.
\end{definition}
Note that if the formula is not satisfiable, there is no satisfying assignment, and the algorithm cannot erroneously find one. Hence the error is one-sided and we don't have to care about unsatisfiable formulas.
The PPSZ algorithm~\cite{ppsz} is a randomized algorithm for Unique 3-SAT running in time $O(1.308^n)$. The precise bound is as follows:
\begin{definition}
Let $S:=\int_{0}^{1}\left(1-\min\{1,\frac{r^2}{(1-r)^2}\}\right)dr=2\ln 2 - 1$.
\end{definition}
\begin{theorem}[\cite{ppsz}]
\label{thm.ppsz}
There exists a randomized algorithm (called $\textsc{PPSZ}$) for Unique 3-SAT running in time $2^{(S+o(1))n}$.
\end{theorem}
Note that $0.3862<S<0.3863$ and $2^{S}<1.308$.
\subsection{Our Contribution}
For Unqiue 3-SAT, we get time bounds exponentially better than PPSZ:
\begin{theorem}
\label{thm.main}
There exists a randomized algorithm for Unique 3-SAT running in time $2^{(S-{\epsilon_2}+o(1))n}$ where ${\epsilon_2}=10^{-24}$.
\end{theorem}
In Section~\ref{sec.ppsz}, we review the PPSZ algorithm. In Section~\ref{sec.1cc}, we show that the worst case for PPSZ occurs when every variable has exactly one critical 3-clause; this case we improve in Section~\ref{sec.i2c}. In Section~\ref{sec.con}, we pose open problems that arise.
\section{The PPSZ Algorithm}
\label{sec.ppsz}
In this section we review the PPSZ algorithm~\cite{ppsz}, summarized in Algorithm~\ref{alg.ppsz}. We need to adapt some statements slightly. For the straightforward but technical proofs we refer the reader to the appendix. The following two definitions are used to state the PPSZ algorithm.
\begin{definition}
A CNF formula $F$ \emph{$D$-implies} a literal $u$ if there exists a subformula $G\subseteq F$ with $|G|\leq D$ and all satisfying assignments of $G$ set $u$ to $1$.
\end{definition}
In a random permutation, the positions of two elements are not independent. To overcome this, placements were defined. They can be seen as continuous permutations with the nice property that the places of different elements are independent.
\begin{definition}[\cite{ppsz}]
A placement on $V$ is a mapping $V\to[0,1]$. A random placement is obtained by choosing for every $x\in V$ $\pi(x)$ uniformly at random from $[0,1]$, independently.
\end{definition}
\begin{observation}
By symmetry and as ties happen with probability $0$, ordering $V$ according to a random placement gives a permutation distributed the same as a permutation drawn uniformly at random from the set of all permutations on $V$.
\end{observation}
\begin{algorithm}
\caption{$\textsc{PPSZ}($CNF formula $F)$}
\label{alg.ppsz}
\begin{algorithmic}
\STATE $V\gets \mathrm{vbl}(F)$; $n\gets |V|$
\STATE Choose $\beta$ u.a.r.\ from all assignments on $V$
\STATE Choose $\pi: V\to [0,1]$ as a random placement of $V$
\STATE Let $\alpha$ be a partial assignment on $V$, initially empty
\FOR {$x\in V$, in ascending order of $\pi(x)$}
\STATE \textbf{if} $F$ $(\log n)$-implies $x$ or $\bar{x}$, set $\alpha(x)$ to satisfy this literal
\STATE \textbf{otherwise} $\alpha(x)\gets\beta(x)$
\COMMENT {guess $\alpha(x)$ u.a.r.}
\STATE $F\gets F^{[x\mapsto \alpha(x)]}$
\ENDFOR
\RETURN $\alpha$
\end{algorithmic}
\end{algorithm}
The analysis of PPSZ builds on the concept of forced and guessed variables:
\begin{definition}
If in $\textsc{PPSZ}$, $\alpha(x)$ is assigned $0$ or $1$ because of $D$-implication, we call $x$ \emph{forced}. Otherwise (if $\alpha(x)$ is set to $\beta(x)$), we call $x$ \emph{guessed}.
\end{definition}
The following lemma from~\cite{ppsz} relates the expected number of guessed variables to the success probability (the proof is by an induction argument and Jensen's inequality).
\begin{lemma}[\cite{ppsz}]
\label{lem.ppsz}
Let $F$ be a satisfiable {{$(\leq 3)$-CNF}}, let $\alpha^*$ be a satisfying assignment. Let $G(\pi)$ be the expected number of guessed variables conditioned on $\beta=\alpha^*$ depending on $\pi$. Then $\textsc{PPSZ}(F)$ returns $\alpha^*$ with probability at least $\ensuremath{\mathbf{E}}_{\pi}[2^{-G(\pi)}]\geq 2^{\ensuremath{\mathbf{E}}_{\pi}[-G(\pi)]}$.
\end{lemma}
Remember that $S:=\int_{0}^{1}\left(1-\min\{1,\frac{r^2}{(1-r)^2}\}\right)dr=2\ln 2 - 1$, which corresponds to the probability that a variable is guessed. We define $S_p$ where the integral starts from $p$ instead of $0$; this corresponds to the probability that a variable has place at least $p$ and is guessed.
\begin{definition}
Let $S_p:=\int_{p}^{1}\left(1-\min\{1,\frac{r^2}{(1-r)^2}\}\right)dr$.
\end{definition}
\begin{observation}
\label{obs.sp}
For $p\leq \frac{1}{2}$, $S_p=S-p+\int_{0}^{p}\frac{r^2}{(1-r)^2}dr$.
\end{observation}
In the appendix, we derive from~\cite{ppsz} the following:
\begingroup
\def\ref{cor.expguessed}{\ref{cor.expguessed}}
\begin{corollary}
\sloppypar{
Let $F$ a {{$(\leq 3)$-CNF}} with unique satisfying assignment $\alpha$. Then in PPSZ($F$) conditioned on $\beta=\alpha$, the expected number of guessed variables is at most $(S+o(1))n$.
Furthermore, suppose we pick every variable of $F$ with probability $p$, independently, and let $V_p$ be the resulting set. Then in PPSZ($F$) conditioned on $\beta=\alpha$, the expected number of guessed variables is at most $(S_p+o(1))n$.
}
\end{corollary}
\addtocounter{theorem}{-1}
\endgroup
By Lemma~\ref{lem.ppsz}, we have the following corollary:
\begin{corollary}
\label{cor.sp}
Let $F$ a {{$(\leq 3)$-CNF}} with unique satisfying assignment $\alpha$. Then the probability that $\textsc{PPSZ}(F)$ returns $\alpha$ is at least $2^{(-S-o(1))n}$.
Furthermore, suppose we pick every variable of $F$ with probability $p$, independently, and let $V_p$ be the resulting set. Then the expected $\log$ of the probability (over the choice of $V_p$) that $\textsc{PPSZ}(F^{[\alpha|_{V_p}]})$ returns $\alpha|_{V\setminus V_p}$ is at least
$(-S_p-o(1))n$.
\end{corollary}
The first statement is actually what is shown in~\cite{ppsz}, and the second statement is a direct consequence. We need this later when we replace PPSZ by a different algorithm on variables with place at most $p$. It is easily seen that for a $(\leq 3)$-CNF $F$, $\textsc{PPSZ}(F)$ runs in time $2^{o(n)}$. Hence by a standard repetition argument, $\textsc{PPSZ}$ gives us an algorithm finding an assignment in time $2^{(S+o(1))n}$ and we (re-)proved Theorem~\ref{thm.ppsz}.
\section{Reducing to One Critical Clause per Variable}
\label{sec.1cc}
In this section we show that an exponential improvement for the case where every variable has exactly one critical clause gives an exponential improvement for Unique 3-SAT.
\begin{definition}[\cite{ppsz}]
Let $F$ be a CNF formula satisfied by $\alpha$.
We call a clause $C$ \emph{critical} for $x$ (w.r.t.\ $\alpha$) if $\alpha$ satisfies exactly one literal of $C$, and this literal is over $x$.
\end{definition}
\begin{definition}
A {{1C-Unique $(\leq 3)$-CNF}} is a uniquely satisfiable {{$(\leq 3)$-CNF}} where every variable has at most one critical clause. Call the corresponding promise problem 1C-Unique 3-SAT.
\end{definition}
All formulas we consider have a unique satisfying assignment; critical clauses will be always w.r.t.\ that. First we show that a variables with more than one critical clause are guessed less often; giving an exponential improvement for formulas with a linear number of such variables. A similar statement for shorter critical clauses is required in the next section.
\begin{lemma}
\label{lem.morethanone}
Let $F$ be a {{$(\leq 3)$-CNF}} uniquely satisfied by $\alpha$. A variable $x$ with at least two critical clauses (w.r.t.\ $\alpha$) is guessed given $\beta=\alpha$ with probability at most $S - 0.0014 + o(1)$. Furthermore, a variable $x$ with a critical $(\leq 2)$-clause is guessed with probability at most $S-0.035 + o(1)$
\end{lemma}
\begin{proof}
Suppose $\pi(x)=r$.
Let $C_1$ and $C_2$ be two critical clauses of $x$. If $C_1$ and $C_2$ share no variable besides $x$, then the probability that $x$ is forced is at least $2r^2-r^4$ by the inclusion-exclusion principle. If $C_1$ and $C_2$ share one variable besides $x$, then the probability that $x$ is forced is at least $2r^2-r^3$ (which is smaller than $2r^2-r^4$. $C_1$ and $C_2$ cannot share two varibles besides $x$: in that case $C_1=C_2$, as being a critical clause for $x$ w.r.t.\ $\alpha$ predetermines the polarity of the literals. Intutiviely, if $r$ is small, then $2r^2-r^3$ is almost twice as large as $\frac{r^2}{(1-r)^2}$; therefore in this area the additional clause helps us and the overall forcing probability increases. For a critical $(\leq 2)$-clause the argument is analogous. Here, the probability that $x$ is forced given place $r$ is at least $r$. The statement follows now by integration using the dominated convergence theorem, see appendix~\ref{subs.integrals}.
\end{proof}
\begin{corollary}
\label{cor.morecc}
\label{cor.2cc}
Let $F$ be a {{$(\leq 3)$-CNF}} formula uniquely satisfied by $\alpha$.
If $\Delta n$ variables of $F$ have two critical clause, PPSZ finds $\alpha$ with probability at least
$2^{-(S-0.0014\Delta + o(1))n}.$
\sloppy{
If $\Delta n$ variables of $F$ have a critical $(\leq 2)$-clause clause, PPSZ finds $\alpha$ with probability at least
$2^{-(S-0.035\Delta + o(1))n}.$
}
\end{corollary}
If there are only few variables (less than ${\Delta_1} n$) with one critical clause, we can find and guess them by brute-force. If we choose ${\Delta_1}$ small enough, any exponential improvement for 1C-Unique 3-SAT gives a (diminished) exponential improvement to Unique 3-SAT.
To bound the number of subsets of size ${\Delta_1} n$, we define the binary entropy and use a well-known upper bound to the binomial coefficient.
\begin{definition}
For $p\in [0,1]$,
$H(p):=-p\log p -(1-p)\log (1-p)$ ($0\log 0:=0$).
\end{definition}
\begin{lemma}[Chapter 10, Corollary 9 of~\cite{ms77}]
If $pn$ is an integer, then
\label{lem.entropy}
$\binom{n}{pn}\leq 2^{H(p)n}.$
\end{lemma}
We will manily prove that we have \emph{some} exponential improvement. The claimed numbers are straightforward to check by inserting the values from the following table.
\begin{center}
\begin{tabular}{|c|c|p{0.72\textwidth}|}
\hline
name&value&description\\
\hline
${\epsilon_1}$&$10^{-19}$&improvement in 1C-Unique 3-SAT\\
\hline
${\epsilon_2}$&$10^{-24}$&improvement in Unique 3-SAT\\
\hline
${\Delta_1}$&$10^{-21}$&threshold fraction of vars. with more than 1 crit. clause\\
\hline
${\Delta_2}$&$6\cdot 10^{-5}$&${\Delta_2} n$ is the amount of variables for ${\Delta_2}$-sparse and ${\Delta_2}$-dense\\
\hline
${\epsilon_3}$&$10^{-3}$&exponential savings on repetitions if $F$ is ${\Delta_2}$-sparse\\
\hline
${p^{*}}$&$8\cdot 10^{-7}$&prob. that a var. is assigned using indep. $2$-clauses instead of $\textsc{PPSZ}$\\
\hline
\end{tabular}
\end{center}
\begin{lemma}
\label{lem.red}
If there is a randomized algorithm $\textsc{OneCC}(F)$ solving 1C-Unique 3-SAT in time $2^{(S-{\epsilon_1}+o(1))n}$ for ${\epsilon_1}>0$, then there is a randomized algorithm (Algorithm~\ref{alg.ppszimp}) solving Unique 3-SAT in time $2^{(S-{\epsilon_2}+o(1))n}$ for some ${\epsilon_2}>0$.
\end{lemma}
\begin{proof}
Let $F$ be a {{$(\leq 3)$-CNF}} uniquely satisfied by $\alpha$. Let $c(F)$ be the number of variables of $F$ with more than one critical clause.
If $c(F)\geq {\Delta_1} n$, PPSZ is faster by Corollary~\ref{cor.morecc}. If $c(F)=0$, we can use $\textsc{OneCC}(F)$.
However, what if $0<c(F)<{\Delta_1} n$? In that case, we get rid of these variables by brute-force: For all $\lfloor{\Delta_1} n\rfloor$-subsets $W$ of variables and for all $2^{\lfloor{\Delta_1} n\rfloor}$ possible assignments $\alpha'$ on $W$, we try $\textsc{OneCC}(F^{[\alpha']})$.
For one such $\alpha'$, we have $F^{[\alpha']}$ satisfiable and $c(F)=0$; namely if $W$ includes all variables with multiple critical clauses and $\alpha'$ is compatible with $\alpha$. This is because fixing variables according to $\alpha$ does not produce new critical clauses w.r.t.\ $\alpha$.
There are $\binom{n}{\lfloor{\Delta_1} n\rfloor}$ subsets of size $\lfloor{\Delta_1} n\rfloor$ of the variables of $F$, each with $2^{\lfloor{\Delta_1} n\rfloor}$ possible assignments. As $\binom{n}{\lfloor{\Delta_1} n\rfloor}\leq 2^{\H({\Delta_1}) n}$ (Lemma~\ref{lem.entropy}), we invoke $\textsc{OneCC}(F^{[\alpha']})$ at most $2^{({\Delta_1} + \H({\Delta_1}))n}$ times. Setting ${\Delta_1}$ small enough such that ${\Delta_1}+\H({\Delta_1})<{\epsilon_1}$ retains an improvement for Unique 3-SAT.
\end{proof}
\begin{algorithm}
\caption{$\textsc{PPSZImproved}($CNF formula $F)$}
\label{alg.ppszimp}
\begin{algorithmic}
\STATE repeat $\textsc{PPSZ}(F)$ $2^{(S-{\epsilon_2})n}$ times, return if a satisfying assignment has been found
\STATE for all subsets $W$ of size $\lfloor{\Delta_1} n\rfloor$ and all assignments $\alpha'$ on $W$, try $\textsc{OneCC}(F^{[\alpha']})$
\end{algorithmic}
\end{algorithm}
\section{Using One Critical Clause per Variable}
\label{sec.i2c}
In this section we give an exponential improvement for 1C-Unique 3-SAT.
\begin{theorem}
\label{thm.onecc}
Given a {{1C-Unique $(\leq 3)$-CNF}} on $n$ variables, \textsc{OneCC}($F$) runs in expected time $2^{(S-{\epsilon_1}+o(1))n}$ and finds the satisfying assignment with probability $2^{-o(n)}$.
\end{theorem}
Obtaining a randomized algorithm using $2^{o(n)}$ independent repetitions and Markov's inequality is straightforward.
\begin{corollary}
\sloppypar{
There exists a randomized algorithm for 1C-Unique 3-SAT running in time $2^{(S-{\epsilon_1}+o(1))n}$.}
\end{corollary}
Together with Lemma~\ref{lem.red} this immediately implies Theorem~\ref{thm.main}.
We obtain the improvement by doing a case distinction into sparse and dense formulas, as defined now:
\begin{definition}
For a CNF formula $F$ and a variable $x$, the \emph{degree} of $x$ in $F$, $\deg(F,x)$ is defined to be the number of clauses in $F$ that contain the variable $x$. The \emph{3-clause degree} of $x$ in $F$, $\deg_3(F,x)$ is defined to be the number of 3-clauses in $F$ that contain the variable $x$.
For a set of variables $W$, denote by $F\setminus W$ the part of $F$ \emph{independent} of $W$ that consists of the clauses of $F$ that do not contain variables of $W$.
We say that $F$ is \emph{$\Delta$-sparse} if there exists a set $W$ of at most $\Delta n$ variables such that $F\setminus W$ has maximum 3-clause degree $4$. We say that $F$ is \emph{$\Delta$-dense} otherwise.
\end{definition}
We will show that for ${\Delta_2}$ small enough, we get an improvement for ${\Delta_2}$-sparse {{1C-Unique $(\leq 3)$-CNF}} formulas. On the other hand, for any ${\Delta_2}$ we will get an improvement for ${\Delta_2}$-dense {{1C-Unique $(\leq 3)$-CNF}} formulas. In the sparse case we can fix by brute force a small set of variables to obtain a formula with few 3-clauses. We need to deal with the $(\leq 2)$-clauses and then use an algorithm from Wahlstr\"om for CNF formulas with few clauses.
\begin{algorithm}
\caption{$\textsc{OneCC}(${{$(\leq 3)$-CNF}} $F)$}
\label{alg.onecc}
\begin{algorithmic}
\STATE try $\textsc{Dense}(F)$
\STATE try $\textsc{Sparse}(F)$
\end{algorithmic}
\end{algorithm}
\begin{algorithm}
\caption{\textsc{GetInd2Clauses}({$(\leq 3)$-CNF} $F$)}
\begin{algorithmic}
\STATE {}
\COMMENT {for the analysis, $F$ is considered to be ${\Delta_2}$-dense; the procedure might fail otherwise}
\STATE $F_3\gets \{C\in F\mid |C|=3\}$, $F_2\gets \{\}$
\FOR {$\lceil {\Delta_2} n \rceil$ times}
\STATE let $x$ be a variable with $\deg_3(F,x)\geq 5$ (return failure if no such variable exists)
\STATE Choose $C$ u.a.r.\ from all of $F$ with $x\in\mathrm{vbl}(C)$.
\STATE $l\gets$ literal of $C$ over $x$; $C_2\gets C\setminus l$
\STATE $F_2\gets F_2\cup C_2$
\STATE {}
\COMMENT {remove all clauses of $F_3$ sharing variables with $C_2$}
\STATE $F_3\gets \{C_3\in F_3\mid \mathrm{vbl}(C')\cap \mathrm{vbl}(C_2)=\emptyset\}$
\ENDFOR
\RETURN $F_2$
\end{algorithmic}
\end{algorithm}
\begin{algorithm}
\caption{$\textsc{Dense}(${{$(\leq 3)$-CNF}} $F)$}
\label{alg.densealg}
\begin{algorithmic}
\STATE $F_2\gets $\textsc{GetInd2Clauses}($F$)
\FOR {$2^{(S-{\epsilon_1})n}$ times}
\STATE $V_{{p^{*}}}\gets $ pick each $x\in\mathrm{vbl}(F)$ with probability ${p^{*}}$
\STATE $\alpha'\gets\{\}$
\FOR {$C_2\in F_2$}
\IF {$\mathrm{vbl}(C_2)\subseteq V_p$}
\STATE Let $\{u,v\}=C_2$
\STATE $(\alpha'(u),\alpha'(v))\gets\begin{cases}
(0,0),&\text{ with probability }\frac{3}{15}\\
(0,1),(1,0),(1,1),&\text{ with probability }\frac{4}{15}\text{ each}\\
\end{cases}$
\ENDIF
\ENDFOR
\STATE \textbf{for all} $x\in V_p$, if $\alpha'(x)$ is not defined yet let $\alpha'(x)\gets_{\text{u.a.r.}} \{0,1\}$
\STATE $\textsc{PPSZ}(F^{[\alpha']})$; if a satisfying assignment $\alpha$ has been found, return $\alpha\cup\alpha'$
\ENDFOR
\end{algorithmic}
\end{algorithm}
\subsection{Dense Case}
First we show the improvement for any ${\Delta_2}$-dense {{1C-Unique $(\leq 3)$-CNF}}. ${\Delta_2}$-density means that even after ignoring all clauses over any $ {\Delta_2} n$ variables, a variable with 3-clause degree of at least 5 remains. The crucial idea is that for a variable $x$ with 3-clause degree of at least 5, picking one occurence of $x$ u.a.r.\ and removing it gives a 2-clause satisfied (by the unique satisfying assignment) with probability at least $\frac{4}{5}$. The only way a non-satisfied $2$-clause can arise is if the 3-clause $x$ was deleted from was critical for $x$. However we assumed that there is at most one critical clause for $x$.
Repeating such deletions and ignoring all 3-clauses sharing variables with the produced 2-clauses, as in listed in \textsc{GetInd2Clauses}($F$), gives us the following:
\begin{observation}
\label{obs.ind2}
For a ${\Delta_2}$-dense {{1C-Unique $(\leq 3)$-CNF}} $F$, \textsc{GetInd2Clauses}($F$) returns a set of $\lceil \frac{1}{2}{\Delta_2} n \rceil$ independent $2$-clauses, each satisfied (by the unique satisfying assignment of $F$) independently with probability $\frac{4}{5}$.
\end{observation}
As a random 2-clause is satisfied with probability $\frac{3}{4}$ by a specific assignment, this set of 2-clauses gives us nontrivial information about the unique satisfying asignment. Now we show how to use these 2-clauses to improve PPSZ:
\begin{lemma}
\label{lem.manyok}
\sloppy{
Let $F$ be a ${\Delta_2}$-dense {{1C-Unique $(\leq 3)$-CNF}} for some ${\Delta_2}>0$. Then there exists an algorithm ($\textsc{Dense}(F)$) runing in time $2^{(S-{\epsilon_1}+o(1))n}$ for ${\epsilon_1}>0$ and returning the satisfying assignment $\alpha$ of $F$ with probability $2^{-o(n)}.$
}
\end{lemma}
\begin{proof}
First we give some intuition. For variables that occur late in PPSZ, the probability of being forced is large (being almost $1$ in the second half). However for variables that come at the beginning, the probability is very small; a variable $x$ at place $p$ is forced (in the worst case) with probability $\Theta(p^2)$ for $p\to 0$, hence we expect $\Theta(p^3 n)$ forced variables among the first $pn$ variables in total.
However, a $2$-clause that is satisfied by $\alpha$ with probability $\frac{4}{5}$ can be used to guess both variables in a better way than uniform, giving constant savings in random bits required. For $\Theta(n)$ such $2$-clauses, we expect $\Theta(p^2 n)$ of them to have both variables among the first $pn$ variables. For each 2-clause we have some nontrivial information; intuitively we save around $0.01$ bits. In total we save $\Theta(p^2 n)$ bits among the first $pn$ variables, which is better than PPSZ for small enough $p$.
Formally, let $V_{p^{*}}$ be a random set of variables, where each variable of $V$ is added to $V_{p^{*}}$ with probability ${p^{*}}$. On $V_{p^{*}}$, we replace PPSZ by our improved guessing; on the remaining variables $V\setminus V_{p^{*}}$ we run PPSZ as usual. Let $\textrm{E}_{\textrm{guess}}$ be the event that the guessing on $V_{p^{*}}$ (to be defined later) finds $\alpha|_{V_{p^{*}}}$. Let $\textrm{E}_\textsc{PPSZ}$ be the event that PPSZ($F^{[\alpha|_{V_{p^{*}}}]}$) finds $\alpha|_{V\setminus V_{p^{*}}}$. Observe that for a fixed $V_{p^{*}}$, $\textrm{E}_{\textrm{guess}}$ and $\textrm{E}_\textsc{PPSZ}$ are independent. Hence we can write the overall probability to find $\alpha$ (call it $p_s$) as an expectation over $V_{p^{*}}$:
\begin{align*}
p_s&=E_{V_{p^{*}}}[\Pr(\textrm{E}_{\textrm{guess}}\cap\textrm{E}_\textsc{PPSZ}\vert V_{p^{*}})]\\
&=E_{V_{p^{*}}}[\Pr(\textrm{E}_{\textrm{guess}}\vert V_{p^{*}})\Pr(\textrm{E}_\textsc{PPSZ}\vert V_{p^{*}})]\\
&=E_{V_{p^{*}}}[2^{\log\Pr(\textrm{E}_{\textrm{guess}}\vert V_{p^{*}})+\log\Pr(\textrm{E}_\textsc{PPSZ}\vert V_{p^{*}})}]\\
&\geq 2^{E_{V_{p^{*}}}[\log\Pr(\textrm{E}_{\textrm{guess}}\vert V_{p^{*}})+\log\Pr(\textrm{E}_\textsc{PPSZ}\vert V_{p^{*}})]}\\
&=2^{E_{V_{p^{*}}}[\log\Pr(\textrm{E}_{\textrm{guess}}\vert V_{p^{*}})]+E_{V_{p^{*}}}[\log\Pr(\textrm{E}_\textsc{PPSZ}\vert V_{p^{*}})]},
\end{align*}
where in the last two steps we used Jensen's inequality and linearity of expectation.
\sloppy{
By Corollary~\ref{cor.sp}, $E_{V_{p^{*}}}[\log\Pr(\textrm{E}_\textsc{PPSZ})]=(-S_p+o(1))n$. We now define the guessing and analyze $E_{V_{p^{*}}}[\log\Pr(\textrm{E}_{\textrm{guess}})]$ (see Algorithm~\ref{alg.densealg} as a reference):
}
By Observation~\ref{obs.ind2} we obtain a set of $\lceil \frac{1}{2}{\Delta_2} n \rceil$ independent $2$-clauses $F_2$, each satisfied (by $\alpha$) independently with probability $\frac{4}{5}$. In the following we assume that $F_2$ has at least a $\frac{4}{5}$-fraction of satisfied 2-clauses as this happens with constant probability (for a proof, see e.g.~\cite{hamza99}) and we only need to show subexponential success probability.
Using the set of 2-clauses $F_2$, we choose an assignment $\alpha'$ on $V_{p^{*}}$ as follows: For every clause $C_2$ in $F_2$ completely over $V_{p^{*}}$ choose an assignment on both of its variables: with probability $\frac{1}{5}$ such that $C_2$ is violated, and with probability $\frac{4}{15}$ each on of the three assignments that satisfy $C_2$. Afterwards, guess any remaining variable of $V_{p^{*}}$ u.a.r.\ from $\{0,1\}$.
Given $V_{p^{*}}$, let $m_0$ be the number of clauses of $F_2$ completely over $V_{p^{*}}$ \emph{not satisfied} by $\alpha$. Let $m_1$ be the number of clauses of $F_2$ completely over $V_{p^{*}}$ \emph{satisfied} by $\alpha$. Then
\[\Pr(\textrm{E}_{\textrm{guess}}\vert V_{p^{*}})=\left(\frac{1}{2}\right)^{|V_{p^{*}}|-2m_0-2m_1}\left(\frac{1}{5}\right)^{m_0}\left(\frac{4}{15}\right)^{m_1}.\]
This is seen as follows: For any variable for which no clause in $C_2$ is completely over $V_{p^{*}}$, we guess uniformly at random and so correctly with probability $\frac{1}{2}$. For any clause $C_2$ which is completely over $V_{p^{*}}$, we violate the clause with probability $\frac{1}{5}$, and choose a non-violating assignment with probability $\frac{4}{5}$. For any clause not satisfied by $\alpha$, we hence set both variables according to $\alpha$ with probability $\frac{1}{5}$. For any clause satisfied by $\alpha$, we set both variables according to $\alpha$ with probability $\frac{4}{15}$, as we have to pick the right one of the three assignments that satisfy $C_2$. As $E[V_{p^{*}}]={p^{*}} n$, $E[m_0]\leq\frac{1}{5}{p^{*}}^2\lceil{\Delta_2} n\rceil$, $E[m_1]\geq\frac{4}{5}{p^{*}}^2\lceil{\Delta_2} n\rceil$, $E[m_0+m_1]={p^{*}}^2\lceil{\Delta_2} n\rceil$, we have
\begin{align*}E[\log \Pr(\textrm{E}_{\textrm{guess}}\vert V_{p^{*}})]&=-E[V_{p^{*}}-2m_0-2m_1]+\log\left(\frac{1}{5}\right)E[m_0]+\log\left(\frac{4}{15}\right)E[m_1]\\
&\geq -{p^{*}} n+{p^{*}}^2\lceil{\Delta_2} n\rceil\left(2+\log\left(\frac{1}{5}\right)\frac{1}{5}+ \log\left(\frac{4}{15}\right)\frac{4}{5}\right).
\end{align*}
The inequality follows from the observations and $\log\left(\frac{4}{15}\right)\geq\log\left(\frac{1}{5}\right)$.
One can calculate $2+\log\left(\frac{1}{5}\right)\frac{1}{5}+ \log\left(\frac{4}{15}\right)\frac{4}{5}\geq 0.01$, corresponding to the fact that a four-valued random variable where one value occurs with probability at most $\frac{1}{5}$ has entropy at most $1.99$.
Hence by our calculations and Observation~\ref{obs.sp} (to evaluate $S_p$), we have
\begin{align*}\frac{1}{n}\log p_s
&\geq -S+{p^{*}}-\int_{0}^{p^{*}} \frac{r^2}{(1-r)^2}dr + o(1) - {p^{*}} + {\Delta_2} {p^{*}}^2 \cdot 0.01\\
&=-S-\int_{0}^{p^{*}} \frac{r^2}{(1-r)^2}dr+{\Delta_2} {p^{*}}^2 \cdot 0.01+o(1).
\end{align*}
This gives an improvement over $\textsc{PPSZ}$ of $-\int_{0}^{p^{*}} \frac{r^2}{(1-r)^2}dr+{\Delta_2} {p^{*}}^2\cdot 0.01+o(1)$. The first term corresponds to the savings $\textsc{PPSZ}$ would have, the second term corresponds to the savings we have in our modified guessing.
Observe that for small ${p^{*}}$, the integral evaluates to $\Theta({p^{*}}^3)$, but the second term is $\Theta({p^{*}}^2)$. Hence choosing ${p^{*}}$ small enough gives an improvement.
\end{proof}
\begin{algorithm}
\caption{$\textsc{Sparse}(${{$(\leq 3)$-CNF}} $F)$}
\begin{algorithmic}
\STATE repeat the following $2^{(S-{\epsilon_3})n}$ times:
\FOR {all subsets $W$ of size $\lfloor{\Delta_2} n\rfloor$ and all assignments $\alpha'$ on $W$}
\STATE $F'\gets F^{[\alpha']}$
\WHILE {no satisfying assignment found}
\STATE try $\textsc{PPSZ}(F^{[\alpha']})$
\STATE $F'_2\gets \{C\in F'\mid |C|\leq 2\}$
\IF {$|F'_2|\leq \frac{1}{10} |\mathrm{vbl}(F')|$}
\STATE with probability $2^{-(S-0.015)|\mathrm{vbl}(F')|}$, run \textsc{Wahlstroem}($F'$)
\ENDIF
\STATE {}
\COMMENT {set all literals in a uniform $(\leq 2)$-clause to $1$}
\STATE $C'\gets_{\text{u.a.r.}} F'_2$, if $F'_2=\{\}$ return failure
\FOR {$l\in C'$}
\STATE $F'\gets F'^{[l\mapsto 1]}$
\ENDFOR
\ENDWHILE
\ENDFOR
\end{algorithmic}
\end{algorithm}
\subsection{Sparse Case}
Now we show that if ${\Delta_2}>0$ is small enough we get an improvement for a ${\Delta_2}$-sparse {{1C-Unique $(\leq 3)$-CNF}}. For this, we need the following theorem by Wahlstr\"om:
\begin{theorem}[\cite{wahlstroem05}]
\label{thm.ws}
Let $F$ be a CNF formula with average degree at most $4.2$ where we count degree $1$ as $2$ instead. Then satisfiability of $F$ can be decided in time $O(2^{0.371n})\leq 2^{(S-0.015+o(1))n}$. Denote this algorithm by $\textsc{Wahlstroem}(F)$.
\end{theorem}
\begin{lemma}
\label{lem.fewok}
\sloppy{
Let $F$ be a ${\Delta_2}$-sparse {{1C-Unique $(\leq 3)$-CNF}}. For ${\Delta_2}$ small enough, there exists an algorithm ($\textsc{Sparse}(F)$) running in expected time $2^{(S-{\epsilon_3}+o(1))n}$ for ${\epsilon_3}>0$ and finding the satisfying assignment $\alpha$ of $F$ with probability $2^{-o(n)}$.
}
\end{lemma}
\begin{proof}
Similar to Section~\ref{sec.1cc}, we first check by brute-force all subsets $W$ of $\lfloor{\Delta_2} n\rfloor$ variables and all possible assignments $\alpha'$ of $W$; by definition of ${\Delta_2}$-sparse for some $W$, the part of $F$ independent of $W$ (i.e.\ $F\setminus W$) has maximum 3-clause degree $4$. If furthermore $\alpha'$ is compatible with $\alpha$, $F':=F^{[\alpha']}$ is a {{1C-Unique $(\leq 3)$-CNF}} with maximum 3-clause degree $4$: We observed that critical clauses cannot appear in the process of assigning variables according to $\alpha$; furthermore any clause of $F$ not independent of $W$ must either disappear in $F'$ or become a $(\leq 2)$-clause.
As earlier, there are at most $2^{({\Delta_2}+H({\Delta_2}))n}$ cases of choosing $W$ and $\alpha'$. We now analyze what happens for the correct choice of $F'$:
We would like to use $\textsc{Wahlstroem}$ on $F'$; however $F'$ might contain an arbitrary amount of $(\leq 2)$-clauses. The plan is to use the fact that either there are many critical $(\leq 2)$-clauses, in which case PPSZ is better, or few critical $(\leq 2)$-clauses, in which case all other $(\leq 2)$-clauses are non-critical and have only satisfied literals.
The algorithm works as follows: We have a {{1C-Unique $(\leq 3)$-CNF}} on $F'$ on $n':=|\mathrm{vbl}(F')|$ variables; the maximum degree in the 3-clauses is at most 4. First we try PPSZ: if there are $\frac{1}{30} n'$ critical $(\leq 2)$-clauses, this gives a satisfying assignment with probability $2^{(S-0.035\frac{1}{30})n'}$. Otherwise, if there are less than $\frac{1}{10}n'$ $(\leq 2)$-clauses, the criterion of Theorem~\ref{thm.ws} applies: We invoke $\textsc{Wahlstroem}(F')$ with probability $2^{-(S-0.015)n'}$; this runs in expected time $2^{-o(n)}$ and finds a satisfying assignment with probability $2^{-(S-0.015)n'}$.
If both approaches fail, we know that $F'$ has at less than $\frac{1}{30} n'$ critical $(\leq 2)$-clauses clauses, but also more than $\frac{1}{10} n'$ $(\leq 2)$-clauses overall. Hence at most one third of the $(\leq 2)$-clauses is critical. However a non-critical $(\leq 2)$-clause must be a $2$-clause with both literals
satisfied. Hence choosing a $(\leq 2)$-clause of $F'$ uniformly at random and setting all its literals to $1$ sets two variables correctly with probability at least $\frac{2}{3}>2^{-0.371\cdot 2}>2^{-(S-0.015)\cdot 2}$. That is we reduce the number of variables by 2 with a better probability than PPSZ overall; and we can
repeat the process with the reduced formula. This shows that for the correct $F'$, we have expected running time $2^{o(n)}$ and success probability $2^{(-S+{\epsilon_3}-o(1))n}$
for ${\epsilon_3}>0$. It is important to see that ${\epsilon_3}$ does \emph{not} depend on ${\Delta_2}$. Repeating this process $2^{(-S+{\epsilon_3}-o(1))n}$ times gives success probability $2^{o(n)}$.
Together with the brute-froce choice of $W$ and $\alpha'$, we have expected running time of $2^{(S-{\epsilon_3}+{\Delta_2}+H({\Delta_2})+o(1))n}$. By choosing ${\Delta_2}$ small enough we are better than PPSZ.
\end{proof}
\section{Open Problems}
\label{sec.con}
Can we also obtain an improvement for general $3$-SAT? In general $3$-SAT, there might be even fewer critical clauses and critical clauses for some assignments are not always critical for others. We need to fit our improvement into the framework of~\cite{hertli11}. As there is some leeway for multiple assignments, this seems possible, but nontrivial and likley to become very complex.
Another question is whether we can improve (Unique) $k$-SAT. PPSZ becomes slower as $k$ increases, which makes an improvement easier. However the guessing in $\textsc{Sparse}$ relied on the fact that non-critical $(\leq 2)$-clauses have all literals satisfied, which is not true for larger clauses.
Suppose Wahlstr\"om's algorithm is improved so that it runs in time $O(c^n)$ on 3-CNF formulas with average degree $D$. The sparsification lemma~\cite{ipz01} shows that for $c\to 1$ and $D\to\infty$, we obtain an algorithm for 3-SAT running in time $O(b^n)$ for $b\to 1$. Can our approach be extended to a similar sparsification result?
\paragraph{\bf{Acknowledgements}}
I am very grateful to my advisor Emo Welzl for his support in realizing this paper.
|
\section{Introduction}\label{sec:introduction}
In this article, we try to develop an analog of the tropical algebra applied to
the case of real varieties. Our point of view is to define real tropical
varieties as \emph{non-archimedean amoebas plus signs}. They are the
restriction to the reals of the complex tropical curves of G. Mikhalkin in
\cite{Mik05} and a natural extension of combinatorial patchworking. Different
approaches to real tropical geometry appear, for instance, in the study of
logarithmic limits of semialgebraic sets in \cite{Alessandrini-logarithmic} or
the study of initial real radical ideals in \cite{real-radical-tropical}.
Another approach with nice combinatorial properties is the study of the
positive part of a tropical variety \cite{trop-pos-grassman}.
We show that, contrary to the complex case, Kapranov's theorem or, more
generally, the fundamental theorem of tropical geometry does not hold in the
real case and there may not be tropical bases. cf. \cite{Computing_trop_var},
\cite{Kapranov}, \cite{Lifting-Constr}, \cite{Payne-fibers},
\cite{Kapranov-EACA}.
However, for sufficiently simple yet interesting varieties, we can compute real
tropical basis. We include here real radical zero dimensional ideals, linear
varieties and hypersurfaces constructed using combinatorial patchworking. As an
application, we are able to describe the singular locus of a real tropical
hypersurface using an analogue of the techniques introduced in
\cite{singular-tropical-hypersurfaces}.
The presentation is done over the field of real Puiseux series $\mathbb{K}$. In
principle, we could work over any real closed field $\mathbb{F}$ provided with
a nontrivial valuation and most of our results can be translated to this more
general setting without trouble. However, some results depending on the residue
field of $\mathbb{F}$ being archimedean or not, see for instance
Proposition~\ref{prop:fallaennoarquimediano} and
Example~\ref{ex:puiseuxdepuiseux}.
The paper is structured as follows. In Section~\ref{sec:preliminares} we
introduce the notation and basic definitions in real tropical geometry. In
Section~\ref{sec:basis} we show the main results concerning the existence and
computation of real tropical basis. Finally, in Section~\ref{sec:singular} we
apply our results to the computation of the singular locus of a real tropical
hypersurface.
\section{Preliminaries}\label{sec:preliminares}
Let us introduce some notation. Let $\mathbb{K}$ be the valued field of real Puiseux
series,
\[\mathbb{K}=\bigcup_{n\geq 1} \mathbb{R}((t^{1/n}))\quad v:\mathbb{K}^*\rightarrow
\mathbb{Q}\subseteq \mathbb{R}.\]
Every element is a power series with real coefficients and rational exponents
with bounded denominator
\[p=\sum_{i\geq r_0} a_i t^{i/n},\quad a_{r_0}\neq 0\]
The valuation $v(p)$ of a nonzero power series $p$ is the least exponent $r_0$
appearing in the development of the series and the principal coefficient is
$a_{r_0}$. This principal coefficient is the residue of the series $pt^{-v(p)}$
in the residue field $\mathbb{R}$. $\mathbb{K}$ is an ordered field with the order
given
by the relation $p>0$ is positive if and only if its principal coefficient
$a_{r_0}$ is positive. The valuation is compatible with the order in the sense
that if $0<p<q$ then $v(q)\leq v(p)$. Moreover, since $\mathbb{K}[i]$ is the
algebraically closed fields of Puiseux series, $\mathbb{K}$ is a real closed field.
\begin{definition}\label{def:real_tropicalization}
The \textbf{real tropicalization} is the valuation taking into account the sign
of the series:
\[\begin{matrix}trop:&\mathbb{K}^* &\longrightarrow &\mathbb{TR}=\{1,-1\}\times
\mathbb{R}\\
& x & \mapsto& (s(x),v(x))\end{matrix}\]
where $s(x)$ is the sign function $s:\mathbb{K}^*\rightarrow \{1,-1\}$, $s(x)=1$
if $x>0$ and $s(x)=-1$ if $x<0$. We will sometimes denote by $a^+=(1,a)$ and
$a^-=(-1,a)\in \mathbb{TR}$. If $x=(p,a)\in \mathbb{TR}$, then $p=s(x)\in\{1,-1\}$ is the
\textbf{sign} of $x$ and $a=|x|\in\mathbb{R}$ is the \textbf{modulus} of $x$.
If
$a=((s_1,a_1),\ldots,(s_n,a_n))\in \mathbb{TR}^n$, we denote by
$s(a)=(s_1,\ldots,s_n)\in \{1,-1\}^n$, $|a|=(a_1,\ldots,a_n)\in \mathbb{R}^n$, the
\textbf{sign}
and \textbf{modulus} taken component-wise.
\end{definition}
\begin{remark}
Note that, while $\mathbb{TR}$ is a group with the \textbf{tropical multiplication}
$(i,a)\odot(j,b) = (i\cdot j, a+b)$, it is not a tropical semiring, since there
is not a reasonable definition of addition for $a^+\oplus
a^-$.
\end{remark}
We now define our main geometric objects, real tropical varieties, as the image
of an algebraic set under the $trop$ map.
\begin{definition}
Let $V\subseteq (\mathbb{K}^*)^n$ be a real variety in the torus. The \textbf{real
tropicalization} of $V$ is the closure (in $\mathbb{R}^n$) of
the image $trop(V)\subseteq \mathbb{TR}^n$ of the real points of $V$ under the
tropicalization map applied component-wise. If $V=\mathcal{V}(I)$ and $I$ is
generated by polynomials in $\mathbb{R}[x_1,\ldots,x_n]$, then we say that we are
dealing with the \textbf{constant coefficient case}.
\end{definition}
This definition is related with taking the non-archimedean amoeba and co-amoeba
of the set of real points of a real variety (cf. the complex tropical curves
presented in \cite{Mik05} section 6).
\begin{figure}
\begin{tikzpicture}[xscale=0.55,yscale=0.55]
\draw (-8,0) -- (8,0);
\draw (0,-8) -- (0,8);
\draw[dashed] (-8,8) -- (-1,8);
\draw[dashed] (-8,6) -- (-1,6);
\draw[dashed] (-8,4) -- (-1,4);
\draw[dashed] (-8,2) -- (-1,2);
\draw[dashed] (1,8) -- (8,8);
\draw[dashed] (1,6) -- (8,6);
\draw[dashed] (1,4) -- (8,4);
\draw[dashed] (1,2) -- (8,2);
\draw[dashed] (-8,-8) -- (-1,-8);
\draw[dashed] (-8,-6) -- (-1,-6);
\draw[dashed] (-8,-4) -- (-1,-4);
\draw[dashed] (-8,-2) -- (-1,-2);
\draw[dashed] (1,-8) -- (8,-8);
\draw[dashed] (1,-6) -- (8,-6);
\draw[dashed] (1,-4) -- (8,-4);
\draw[dashed] (1,-2) -- (8,-2);
\draw[dashed] (-8,8) -- (-8,1);
\draw[dashed] (-6,8) -- (-6,1);
\draw[dashed] (-4,8) -- (-4,1);
\draw[dashed] (-2,8) -- (-2,1);
\draw[dashed] (8,8) -- (8,1);
\draw[dashed] (6,8) -- (6,1);
\draw[dashed] (4,8) -- (4,1);
\draw[dashed] (2,8) -- (2,1);
\draw[dashed] (-8,-8) -- (-8,-1);
\draw[dashed] (-6,-8) -- (-6,-1);
\draw[dashed] (-4,-8) -- (-4,-1);
\draw[dashed] (-2,-8) -- (-2,-1);
\draw[dashed] (8,-8) -- (8,-1);
\draw[dashed] (6,-8) -- (6,-1);
\draw[dashed] (4,-8) -- (4,-1);
\draw[dashed] (2,-8) -- (2,-1);
\draw[fill] (4,4) circle(4pt) (-4,4) circle(4pt) (-6,6) circle(4pt);
\draw[fill] (-4,-4) circle(4pt) (-6,-6) circle(4pt) (6,-6) circle(4pt);
\draw[fill] (4,-4) circle(4pt);
\draw[thick] (8,4) -- (4,4) -- (4,8);
\draw[thick] (1.5,1.5) -- (4,4);
\draw[thick] (-6,8) -- (-6,6) -- (-4,4) --(-8,4);
\draw[thick] (8,-6) -- (6,-6) -- (4,-4) --(4,-8);
\draw[thick] (-8,-6) -- (-6,-6) -- (-6,-8);
\draw[thick] (-4,-4) -- (-1.5,-1.5);
\draw (5,4.5) node {\tiny $(0^+,0^+)$};
\draw (-3,4.5) node {\tiny $(0^-,0^+)$};
\draw (-5,6.5) node {\tiny $(1^-,1^+)$};
\draw (-5,-3.5) node {\tiny $(0^-,0^-)$};
\draw (-5,-5.5) node {\tiny $(1^-,1^-)$};
\draw (7,-5.5) node {\tiny $(1^+,1^-)$};
\draw (5,-3.5) node {\tiny $(0^+,0^-)$};
\end{tikzpicture}
\caption{$\mathcal{T}_\mathbb{R}(f)$ with $f = 1^+ \oplus 0^+ v \oplus 0^+ w \oplus 0^-
v^2 \oplus 0^+ vw \oplus 0^-w^2$}\label{fig:conica_no_triangulation}
\end{figure}
In the algebraically closed case, the tropicalization of a variety $V$
can be computed using Kapranov's theorem, also know as the fundamental theorem
of tropical geometry. The real case seems more involved. The tropicalization in
the constant coefficient case has been studied in \cite{real-radical-tropical}
but there is a difference in the tropicalization in the non constant
coefficient case we are studying here that has its root in P\'olya's Theorem
\cite{Polyazeros} (See also \cite{positive_polynomial} for a more general
result).
\begin{theorem}[\cite{Polyazeros}]\label{teo:Polya}
Let $F\in \mathbb{R}[x_1,\ldots, x_n]$ be a homogeneous polynomial that is
positive in the set $\{x\in \mathbb{R}^n| x_i\geq 0, x_1+\ldots+x_n \neq 0\}$,
then for $N$ sufficiently large $F\cdot(x_1+\ldots+x_n)^N$ has only positive
coefficients.
\end{theorem}
However, it is well known that this theorem does not hold if the ground field
is not archimedean (See Example~\ref{ex:univariate_not_Polya}). This translates
to the fact that tropicalization of polynomials is not so useful in the
non-constant coefficient case.
We now introduce a different way of defining tropical varieties in terms of a
real tropical polynomial. The motivation of this alternative approach is to get
rid of algebraic polynomials and deal only with tropical polynomials, this is a
common approach in the tropical case over complex Puiseux series.
\begin{definition}
A \textbf{real polynomial} $f\in \mathbb{TR}[w_1,\ldots, w_n]$ is just a formal sum of
tropical monomials of the form $f=\oplus_{\ell\in A}a_\ell w^\ell\in
\mathbb{TR}[w_1,\ldots,
w_n]$, $A\subseteq \mathbb{N}^n$. Every real tropical polynomial defines a
piecewise affine function $f:\mathbb{TR}^n\rightarrow \mathbb{R}$.
$f(p)=\min\{|a_\ell\ |\ +\langle |p|,\ell\rangle \ |\ \ell\in A\}$.
We define the \textbf{real tropical
hypersurface} defined by $f$, $\mathcal{T}_\mathbb{R}(f)$ by $p=(p_1,$ $\ldots,$
$p_n)\in \mathcal{T_\mathbb{R}}(f)$ if there are two monomials $i\neq j$ such that
\[s(a_i)\prod_{l=1}^ns(p_l)^{l_i}=1,\quad s(a_j)\prod_{l=1}^ns(p_l)^{l_j}=-1\]
and
\[|a_i|+\langle i,|p|\rangle=|a_j|+\langle j,|p|\rangle\ \leq\ |a_k|+\langle
k,|p|\rangle\] for all $k\neq i,j$.
That is, if the minimum of $f(p)$ is attained at two different monomial $i,j$
where the evaluation at the point $p$ yields two different signs.
\end{definition}
\begin{example}\label{ex:raices-trop}
Let $f=0^+\oplus 1^+w\oplus 0^+w^2\oplus 1^+w^3\oplus2^-w^4$. We have
associated the piecewise-linear map $p\mapsto \min\{0, 1+p,0+2p, 1+3p, 2+4p\}$.
The minimum of this piecewise-linear is attained at least twice for $p=0,-1$.
The value and sign attained in the candidates of roots are:
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|}
&$0$ &$1+p$&$0+2p$&$1+3p$&$2+4p$\\\hline
$0^+$&$0^+$&$1^+$&$0^+$&$1^+$&$2^-$\\\hline
$0^-$&$0^+$&$1^-$&$0^+$&$1^-$&$2^-$\\\hline
$-1^+$&$0^+$&$0^+$&$-2^+$&$-2^+$&$-2^-$\\\hline
$-1^+$&$0^+$&$0^-$&$-2^+$&$-2^-$&$-2^-$\\\hline
\end{tabular}
\end{center}
Neither $0^+$ nor $0^-$ are real tropical roots of $f$, because the minimum is
attained at $0^+$ so we do not have two different signs. On the other hand,
$-1^+$ and $-1^-$ are both real tropical roots, since the minimum is attained
at different signs $2^+$ and $2^-$. Hence $\mathcal{T}_\mathbb{R}(f)=\{-1^+,-1^-\}$.
\end{example}
In Figure~\ref{fig:conica_no_triangulation}, we show the four tropical quadrants
of the conic given by the polynomial $f = 1^+ \oplus 0^+ v \oplus 0^+ w \oplus
0^- v^2 \oplus 0^+ vw \oplus 0^-w^2$. Note that we are working with the $\min$
and that the subdivision induced by $f$ is not a triangulation, so the picture
is different from the usual ones in patchworking.
\begin{definition}
If $F = \sum_{\ell\in A} a_\ell x^\ell\in \mathbb{K}[x_1,\ldots, x_n]$ is a real
polynomial, write \[trop(F)=f=\bigoplus_{\ell\in A}
(s(a_\ell),v(a_\ell))w^\ell.\]
By abuse of notation, we will write
\[\mathcal{T}_\mathbb{R}(F)= \mathcal{T}_\mathbb{R}(f)\subseteq \mathbb{TR}^n\]
\end{definition}
Clearly $trop(\mathcal{V}(F))\subseteq \mathcal{T_\mathbb{R}}(F)$
(Example~\ref{ex:univariate_Polya}), but contrary to the algebraically closed
set, the inclusion may be strict. Next lemma shows further discrepancies
between the real and usual tropical setting.
\begin{lemma}\label{lem:factorization}
Let $F$, $G\in \mathbb{K}[x_1,\ldots,x_n]$, then
\[\mathcal{T}_\mathbb{R}(F\cdot G) \subseteq \mathcal{T}_\mathbb{R}(F) \cup \mathcal{T}_\mathbb{R}(G)\]
and the inclusion may be strict.
\end{lemma}
\begin{proof}
Let $p\in \mathbb{TR}^n$ such that $p\notin (\mathcal{T}_\mathbb{R}(F) \cup
\mathcal{T}_\mathbb{R}(G))$. Without lost of generality, assume that
$p=(0^+,\ldots,0^+)$. Let $A_1$ (resp $A_2$) be the monomials of $F$ (resp
$G$) where the minimum of $trop(F)$ (resp. $trop(G))$ is attained at $p$. Then,
the tropicalization of the monomials of $A_1$ attain the same sign at $p$ and
so do the monomials in $A_2$. We may also suppose that all these monomials have
positive coefficients. Then, the monomials of $FG$ where the minimum is
attained at $p$ are all of the form $ca_1a_2$, with $c\in \mathbb{K}$, $c>0$, $a_1\in
A_1$, $a_2\in A_2$. It follows that all monomials yield the same sign at $p$
and $p\notin \mathcal{T}_\mathbb{R}(FG)$.
To see that the inclusion may be strict, let $F=x^{2} y + x^{2} - x y - x + y
+ 1$, $G=x y^{2} - x y + y^{2} + x - y + 1$. Then $FG=x^{3} y^{3} + x^{3} +
y^{3} + 1$. So $p=(0^+,0^+)$ belongs to both $\mathcal{T}_\mathbb{R}(F)$ and
$\mathcal{T}_\mathbb{R}(G)$, but not to $\mathcal{T}_\mathbb{R}(FG)$.
\end{proof}
We end this section with an elementary result concerning the number of real
roots that an
univariate polynomial admit.
\begin{definition}\label{def:realmultiplicityalaDescartes}
Let $f=\oplus_{\ell\in A} a_\ell w^\ell\in \mathbb{TR}[w]$ be an univariate real
tropical polynomial. Let $f'=\oplus_{\ell\in A} |a_\ell| w^\ell\in
\mathbb{T}[w]$ be the usual tropical polynomial obtained from $f$ by forgetting
signs.
Let $p\in \mathcal{TR}$ with $|p|$ a tropical root of $f'$. Let
$i_1<\ldots<i_r$ be
the monomials where $|a_\ell| + \langle |p|, \ell\rangle$ attains its minimum.
Take the
sequence of signs
\[S_p=(s(a_{i_1})s(p)^{i_1}, \ldots, s(a_{i_n})s(p)^{i_n})\]
then, the \textbf{complex multiplicity} of $|p|$ in $f'$ is $i_n-i_1$ and the
\textbf{real multiplicity} of $p$ in $f$ is the number of changes of signs in
$S_p$.
\end{definition}
\begin{lemma}\label{lem:descartesmanda}
Let $f = \oplus_{\ell\in A}a_\ell w^\ell\in \mathbb{TR}[w]$ be an univariate real
tropical polynomial of degree $n$. Let $f' = \oplus_{\ell \in A}|a_\ell|w^\ell
\in \mathbb{T}[w]$ be the (usual) tropical polynomial obtained from $f$ by
forgetting the signs. Let $p\in \mathbb{T}$ be a (usual) tropical root of $f'$
of (usual) multiplicity $m$. Let $m^+$ (resp. $m^-$) be the real tropical
multiplicity of $p^+$ (resp. $p^-$) as a real tropical root of $f$. Then $m\leq
m^+ + m^-$ and the difference $m-m^+-m^-$ is an even number.
\end{lemma}
\begin{proof}
Both multiplicities only depend on the support and signs of the coefficients
where the minimum is attained when evaluating $f$ at $p$. Hence, we may assume,
without loss of generality, that all coefficients of $f$ and the tropical root
have modulus 0. consider the polynomial
\[F=\sum_{\ell\in A}s(a_\ell)\cdot t^{\ell^2}x^\ell\in \mathbb{R}[t][x].\]
By combinatorial patchworking \cite{patchworking}, for any sufficiently small
evaluation $t_0$ of $t$, $0<t_0<<1$, $F(t=t_0)$ will have exactly $m^+$
positive roots and $m^-$ negative roots, $m$ equals the number of nonzero roots
of $F$, so $m^++m^-\leq m$ and its difference is the number of non-real roots
of $F$ is even.
\end{proof}
\begin{example}
Let $f=0^+\oplus 1^+w\oplus 0^+w^2\oplus 1^+w^3\oplus2^-w^4$ be the real
tropical polynomial from Example~\ref{ex:raices-trop}, the usual tropical roots
of $f'$ are $0$ and $-1$ with multiplicity 2 each root. Now, for the root zero,
we know that $f$ does not admit a real tropical root. So for $|p|=0$ we have
that
$m_+=m_-=0$, $m=2$ and $m-m_+-m_-=2$. On the other hand, the sequence of signs
$S_{1^+}=(1,1,-1)$ so $m^+=1$, $S_{1^-}=(1,-1,-1)$ and $m_-=1$, $m-m_+-m_-=0$
is also even.
\end{example}
\section{Real Tropical Basis}\label{sec:basis}
We start with the definition of real tropical basis, cf.
\cite{Computing_trop_var}
\begin{definition}
Let $I\subseteq \mathbb{K}[x_1,\ldots,x_n]$. A \textbf{real tropical basis} of $I$ is a
finite set $\{F_1,\ldots, F_r\}$ that generates $I$ such that
$trop(\mathcal{V}(I)) = \cap_{i=1}^r\mathcal{T_\mathbb{R}}(F_i)$.
\end{definition}
\begin{example}\label{ex:univariate_Polya}
Let $F=x^2-x+1$, $\mathcal{V}(F)=\emptyset$, there are no real solutions. But
$trop(F)=f=0^+w^2\oplus 0^-w\oplus 0^+$ and $\mathcal{T_\mathbb{R}}(f)=\{0^+\}$. This
means
that the generator $\{x^2-x+1\}$ is not a real tropical basis of the ideal
$(x^2-x+1)$. Let $x^3+1=(x^2-x+1)(x+1)$. $\mathcal{T_\mathbb{R}}(0^+w^3\oplus
0^+)=\{0^-\}$, hence $\mathcal{T_\mathbb{R}}(f)\cap \mathcal{T_\mathbb{R}}(0^+w^3\oplus
0^+)=\emptyset$ and $I=(x^2-x+1, x^3+1)$ is a real tropical basis of $I$. Note
that, in this case, $I$ is not real radical.
\end{example}
The first result is that real-radical zero-dimensional ideals admit a tropical
basis.
\begin{definition}
Let $I\subseteq \mathbb{K}[\underline{x}]=\mathbb{K}[x_1,\ldots, x_n]$. The \textbf{real
radical}
of $I$, $\sqrt[\mathbb{R}]{I}$ is the ideal
\[\sqrt[\mathbb{R}]{I}=\{F\in \mathbb{K}[\underline{x}]\ |\ \exists G_1,\ldots,G_r\in
\mathbb{K}[\underline{x}], \exists m>0, F^{2m}+\sum_{i=1}^r G_i^2\in I\}.\]
By the real theorem of zeros \cite{real-alg-geo}, \[\sqrt[\mathbb{R}]{I}=\{F\in
\mathbb{K}[x_1,\ldots,x_n]\ |\
\forall a\in \mathcal{V}_\mathbb{K}(I), F(a)=0\}\]
An ideal $I$ is \textbf{real radical} if $I=\sqrt[\mathbb{R}]{I}$.
\end{definition}
\begin{theorem}
Let $I$ be a real radical zero dimensional ideal, with $\mathcal{V}(I)\subseteq
(\mathbb{K}^*)^n$ then $I$ admits a tropical basis.
\end{theorem}
\begin{proof}
Let $V=\mathcal{V}(I)=\{p_1,\ldots, p_r\}\subseteq (\mathbb{K}^*)^n$. Let
$p_i=(p_{i1},\ldots, p_{in})$ and $trop(p_i)=a_i=(a_{i1},\ldots,a_{in})$,
$1\leq i\leq n$. So $trop(V)=\{a_1,\ldots,a_r\}$. Let $F_j$ be the squarefree
part of $\prod_{i=1}^r(x_j-p_{ij})$, since $I$ is real radical $F_j\in I, 1\leq
j\leq n$. The tropical roots of $trop(F_j)$ are precisely $a_{1j},\ldots,
a_{rj}$
and the number of tropical roots of each sign and valuation $a_{ij}$ can be
recovered using Lemma~\ref{lem:descartesmanda} and taking into account that all
the roots of $F_j$ are real. The set of tropical polynomials $\left\{trop(F_j),
1\leq j\leq n\right\}$ describe a finite set $S$ of points containing $trop(V)$.
Let $L=b_1x_1+\ldots +b_nx_n$ be a linear function with integer coefficients
such that $L$ is injective in the set of modulus of $S$, $|S|=\{|a|\ |\ a\in
S\}$. Let $F_0$ be the squarefree part of the numerator of the polynomial
$\prod_{i=1}^r (x_1^{b_1}\cdots x_n^{b_n}-p_{i1}^{b_1}\cdots p_{in}^{b_n})$.
Since this is a polynomial in $I$ whose tropicalization is injective in $S$, it
allows us to discriminate which modulus of points of $S$ belong to $trop(V)$ or
not.
Finally, suppose that $c=(c_1,\ldots,c_n)$ is a tropical point in $S\cap
\mathcal{T}_\mathbb{R}(F_0)$ but not in $trop(\mathcal{V}(I))$. This can only happen
if there
is a point in $a\in trop(V)$ with $b\neq a$, $|b|=|a|$ and for every index
$j$ there is a point in $trop(V)$ with $j$-th coordinate $b_j$.
For every point $a_i\in trop(V)$ with $|a_i|=|c|$, there is a coordinate $h(i)$
with $s(a_{i,h(i)})\neq s(c_{h(i)})$, Let $G_c$ be the squarefree part of
\begin{equation}\label{eq:gc}
\prod_{i, |a_i|=|c|} (x_{h(i)}-p_{i,h(i)}) \times
\prod_{i,|a_i|\neq|c|}(x_1^{b_1}\cdots x_n^{b_n}-p_{i1}^{b_1}\cdots
p_{in}^{b_n}) \in I
\end{equation}
This polynomial vanishes by construction on every element of $trop(V)$, but $c$
is not a real tropical root of $trop(G_c)$, because, for each factor of $G_c$
in
(\ref{eq:gc}), $c$ either has the wrong modulus or the wrong distribution of
signs. By Lemma~\ref{lem:factorization}, $c\notin \mathcal{T}_\mathbb{R}(G_c)$.
If $H$ is any generator set of $I$, then $H\cup \{F_j, 1\leq j\leq
n\}\cup\{F_0\}\cup \{G_c| c\in S-trop(V)\}$ is a real tropical basis of $I$.
\end{proof}
\begin{example}
Let $V=\{(-1,2)$, $(2,3)$, $(-3,-t)$, $(1,-4)$, $(2t,4t)\}\subseteq
(\mathbb{K}^*)^2$, $trop(V) = \{(0^-,0^+)$, $(0^+,0^+)$, $(0^-,1^-)$,
$(0^+,0^-)$, $(1^+,1^+)\}$. Let $I=\mathcal{I}(V)$ be the ideal of $V$. We are
computing a real tropical basis of $I$. First, define
\[F_1(x)=x^{5} + \left(-2 t + 1\right) x^{4} + \left(-2 t - 7\right) x^{3} +
\left(14 t - 1\right) x^{2} + \left(2 t + 6\right) x - 12 t,\]
\[F_2(y)=y^{5} + \left(-3 t - 1\right) y^{4} + \left(-4 t^{2} + 3 t - 14\right)
y^{3} + \left(4 t^{2} + 42 t + 24\right) y^{2} + \]\[\left(56 t^{2} - 72
t\right) y - 96 t^{2}.\]
$trop(F_1)$ and $trop(F_2)$ define the finite set
\[\begin{matrix}
S=&\{(0^+,0^+),&(0^+,1^+), &(0^+,1^-), &(0^+,0^-),\\
&\ (0^-,0^+), &(0^-,1^+), &(0^-,1^-), &(0^-,0^-),\\
&\ (1^+,0^+), &(1^+,1^+), &(1^+,1^-),&(1^+,0^-)\}
\end{matrix}\]
$|S|=\{(0,0), (0,1), (1,0), (1,1)\}$. Note
that $(1,0)\notin |trop(V)|$. Let $L=$ $x+2y$ by a linear function injective in
$|S|$. Then
\[F_0=(xy^{2} - 18) \cdot (x y^{2} - 16) \cdot (x y^{2} + 4) \cdot
(x y^{2} + 3 t^{2}) \cdot (x y^{2} - 32 t^{3})\]
In usual tropical geometry $\{F_1,F_2,F_0\}$ would suffice to provide a
tropical basis, See \cite{TR_bas_reg_proy}, \cite{Resultantes-trop}. But these
polynomials are not enough in the real case. The elements of $S$ that are also
in $\mathcal{T}_\mathbb{R}(F_0)$ are
$\{(0^+,0^+)$, $(0^+,0^-)$, $(0^-,0^+)$, $(0^-,1^+)$, $(0^-,1^-)$, $(0^-,0^-)$,
$(1^+,1^-)$, $(1^+, 1^+)\}$. We have to compute polynomials to discard the
points in $S$ not in $trop(V)$. Following the theorem:\\
$G_{(0^-,0^-)}=(x-2)(x-1)(y-2)(xy^2-32t^3)(xy^2+3t^2)$, and $(0^-,0^-) \notin
\mathcal{T}_\mathbb{R}(G_{(0^+,0^+)})$. Define also
$G_{(0^-,1^+)}=(y+t)(xy^2 + 4)(xy^2 - 18)(xy^2 - 16)(xy^2 - 32t^3)$ and
$G_{(1^+,1^-)}=(y-4t)(xy^2 + 4)(xy^2 - 18)(xy^2 - 16)(xy^2 + 3t^2)$. Then, if
we add to any finite set of generators of $I$ the set $\{F_1$, $F_2$, $F_0$,
$G_{(0^-,0^-)}$, $G_{(0^-,1^+)}$, $G_{(1^+,1^-)}\}$ we get a real tropical
basis of $I$.
\end{example}
The hypothesis of being real radical cannot be avoided as the following example
shows.
\begin{example}\label{ex:univariate_not_Polya}
Let $F=x^2-(2+t)x+1$. The discriminant of $F$ is $t^2 + 4t>0$, so there are no
real roots of $F$ and $\sqrt[\mathbb{R}]{(F)}=(1)$. However, $I=(F)$ has no real
tropical
basis.
To see that, $\{0^+\}$ is a root of $trop(x^2-(2+t)x+1)=0^+x^2\oplus 0^-x\oplus
0^+$. Now $v(F(1))=v(-t)=1>0$, $1$ is a real root of $x^2-2x+1$. If $G\in
\mathbb{K}[x]$, assume without loss of
generality that the minimum of the valuation of the coefficients of $G$ is
zero.
Then $1$ will be a root of the residue polynomial of $F\cdot G$. This means
that
$0^+$ will also be a tropical root of $trop(FG)$ and $\bigcap_{H\in
I}\mathcal{T}_\mathbb{R}(H)=\{0^+\}$.
\end{example}
The reader may think that the failure here is due to the fact that the
\emph{residue} polynomial $x^2-2x+1$ obtained by substituting $t=0$ has a
double root. This is true for the field of real Puiseux series but it may fail
over more complicated fields.
\begin{definition}
Let $F=\sum_{\ell\in A}p_\ell x^\ell\in \mathbb{K}[x_1,\ldots,x_n]$ be a polynomial,
let $w\in \mathbb{R}^n$. Let $i_1,\ldots,i_r$ be the monomials where $trop(F)$ attains
its
minimum at $w$. We define the \textbf{residue polynomial} of $F$ with respect
to
$w$ to the polynomial
\[F_w=\sum_{j=1}^r Pc(p_{i_j})x^{i_j}\in \mathbb{R}[x_1,\ldots,x_n],\]
where $Pc(p_{i_j})\in \mathbb{R}$ is the principal coefficient of the power series
$p_{i_j}$.
\end{definition}
\begin{proposition}\label{prop:fallaennoarquimediano}
Let $F\in \mathbb{K}[x]$ be an univariate polynomial such that, for every real root
$p\in \mathbb{K}^*$ we have that $F_{trop(p)}$ has only simple roots, then $(F)$ admits
a tropical basis.
\end{proposition}
\begin{proof}
Let $\{p_1,\ldots,p_r\}=\mathcal{T}_\mathbb{R}(F)$ be the real tropical roots of
$trop(F)$. If $p_i\notin trop(\{F=0\})$ we have to find a certificate of this
fact. Without loss of generality, we may assume that $p_i=0^+$. $F_0\in \mathbb{R}[x]$
is an univariate polynomial that has no positive real root. From
Theorem~\ref{teo:Polya}, there exists a natural $N$ such that $F_0\cdot(1+x)^N$
has all its coefficients positive. It follows that $0^+ \notin
\mathcal{T}_\mathbb{R}(F\cdot (1+x)^N)$. Iterating over every element in
$\mathcal{T}_\mathbb{R}(F)-trop(\{F=0\})$ we can construct a real tropical basis of
$(F)$.
\end{proof}
We now show that Proposition~\ref{prop:fallaennoarquimediano} cannot be
extended to any real closed valued field.
\begin{example}\label{ex:puiseuxdepuiseux}
Let $\mathbb{F}=\mathbb{K}\{\{s\}\}$ be the field of Puiseux series is $s$ whose
coefficients are Puiseux series in $t$ over the reals. Take again
$F=x^2-(2+t)x+1\mathbb{F}[x]$. Now, $trop(F)$ still has a tropical root $0^+$
but
$F$ has no root in $\mathbb{F}$. By the same argument as in
Example~\ref{ex:univariate_not_Polya} $(F)$ does not admit a tropical basis.
Note that in this new context $F=F_0$ has no multiple root.
\end{example}
Let us consider now other ideals that also admit tropical basis. We can revisit
combinatorial patchworking (See \cite{patchworking} for the details) from the
point of view of tropical basis.
\begin{theorem}\label{teo:T-construction}
Let $F\in \mathbb{K}[x_1,\ldots, x_n]$ be a real polynomial and let $f =
trop(F)$ be the corresponding tropical polynomial. $f$ defines a mixed
subdivision in $Supp(F)$ by duality. Assume that this subdivision is a
triangulation that contains all the monomials in $Supp(F)$ as vertices. Then
$\{F\}$ is a real tropical basis of $(F)$. That is
$\mathcal{T}_\mathbb{R}(F)=trop(\mathcal{V}(F))$.
\end{theorem}
\begin{proof}
We are going to prove that if $p\in \mathcal{T}_\mathbb{R}(F)\cap \mathbb{Q}^n$ then
$p\in trop(\mathcal{V}(F))$. Without loss of generality, we may assume that
$p=(0^+,\ldots, 0^+)$.
Let $S=Supp(F)$, the set of monomials of $S$ where $f$ attains its minimum at
$p$ is a simplex $C$ in $S$ that is not a point. Since $p\in
\mathcal{T}_\mathbb{R}(f)$, then there is a positive
monomial $a\in C$ and a negative monomial $b\in C$. Let
$L=r_0+r_1x_1+\ldots+r_nx_n$
be an affine function with integer coefficients such that for all monomials
$c\in S-\{a,b\}$, $L(a)<L(c)<L(b)$. Let $c_ax^a$, $-c_bx^b$ be the
corresponding
monomials of $F$. Consider the points of the form $(s^{r_1},\ldots, s^{r_n})$.
If $0<s_-<<1$ is a valuation zero positive element of $\mathbb{K}$ small enough, then
$F(s_-^{r_1},\ldots, s_-^{r_n})\sim c_as^{L(a)}>0$. If $1<<s_+$ is a valuation
zero element big enough, $F(s_+^{r_1},\ldots, s_+^{r_n})\sim -c_bs^{L(b)}<0$.
Hence, by the intermediate value theorem, the polynomial $F(s^{r_1},\ldots,
s^{r_n})$ has a root $s_0$ in the
interval $[s_-,s_+]$, and hence $s_0$ is of valuation zero. By construction,
$(s_0^{r_1},\ldots, s_0^{r_n})$ is a zero of $F$ with tropicalization
$(0^+,\ldots, 0^+)$.
\end{proof}
\begin{corollary}\label{cor:Kapranov_hyperplane}
Real vector hyperplanes have tropical basis. More precisely, let
$H=V(a_1x_{i_1}+\ldots+a_rx_{i_r}+a_0)\in \mathbb{K}[x_1,\ldots, x_n]$ be a hyperplane
in $(\mathbb{K}^*)^n$.
Then $trop(H)=\mathcal{T}(trop(a_1)w_{i_1}\oplus\ldots\oplus
trop(a_n)w_{i_r}\oplus trop(a_0))$.
\end{corollary}
\begin{proof}
We are in the hypothesis of Theorem~\ref{teo:T-construction} since $Supp(H)$ is
a
simplex.
\end{proof}
The last goal of this section is to prove that linear ideals have a tropical
basis (cf. \cite{Bergman-complex, Computing_trop_var} for the algebraically
closed case). We start with a Lemma that states that there are
infinite tropical basis.
\begin{lemma}\label{lem:trop-basis-linear}
Let $V \subseteq (\mathbb{K}^*)^n$ be a linear space, then \[trop(V) =
\bigcap_{\substack{\subseteq H\\H\in \mathcal{H}}} trop(H),\] where
$\mathcal{H}$ is the set of hyperplanes of $(\mathbb{K}^*)^n$.
\end{lemma}
\begin{proof}
Clearly $trop(V)\subseteq \bigcap_{V\subseteq H} trop(H)$. We prove the
equality by induction in the dimension $n$ of the ambient space. For $n=2$, $V$
can only be a line, so the result holds by
Corollary~\ref{cor:Kapranov_hyperplane}. Assume it is true for ambient
dimension up to $n-1$. We prove that the result is true in ambient dimension
$n$ by induction in the codimension of $V$. If $codim(V)=1$, then $V$ is a
hyperplane and this is again Corollary~\ref{cor:Kapranov_hyperplane}. Assume
now that the result is true for all codimension $r$ spaces and that $V$ has
codimension $r+1$. If $V\subseteq \{x_i=0\}$ for some $i$ then
$trop(V)=trop(\{x_i=0\})=\emptyset$. Let $p=(p_1,\ldots,p_n)\in
\bigcap_{V\subseteq H} trop(H)$. Without loss of generality
$p=(0^+,\ldots,0^+)$. If $e_i\in V$ is the $i$-th vector of the canonical
basis, the $i$-th coordinate of the points of $V$ may take any value and for
all $V\subseteq H$, the $i$-th coefficient of the defining linear equation of
$H$ is zero. It follows that we can project along the $i$-th coordinate and $p
\in trop(V)$ iff $\pi_i(p)\in trop(\pi_i(V))$ and clearly $\pi_i(p)\in
\bigcap_{V \subseteq H} trop(\pi_i(H)) =\bigcap_{\pi_i(V)\subseteq H} trop(H)$
where $H$ is any hyperplane in dimension $n-1$ containing $\pi_i(V)$ and we are
done by induction.
Hence, we may assume that $e_i\notin V$ for $1\leq i\leq n$. Consider the linear
spaces $V+<e_i>$, $1\leq i\leq n$. Then $p\in \bigcap_{V+<e_i>\subseteq H}
trop(H)=trop(V+<e_i>)$ by induction hypothesis in the codimension of $V$.
Hence, for each $i$, $1\leq
i\leq n$ there is a point $a_i=(a_{i1},\ldots, a_{in}) \in V$ such that
for all $i\neq j$, $a_{ij}>0$ and $v(a_{ij})=0$. We can write the matrix
\[A=\begin{pmatrix}a_{11} & a_{12}&\ldots & a_{1n}\\&&\ldots\\alpha_{n1} &
a_{n2}&\ldots & a_{nn}\end{pmatrix}\ trop(A)=\begin{pmatrix}q_1 & 0^+&\ldots &
0^+\\0^+&q_2&\ldots &0^+\\ 0^+&0^+&\ldots &q_n\end{pmatrix}\]
Since $codim(V)>1$, $rank(A)\leq n-2$ and $A$ is singular, so it cannot happen
that for all $i$ $|q_i|< 0$. Assume that $|q_1| \geq 0$. Since the minor
$A_{11}$ is also singular, there must be another element $|q_i| \geq 0$. Assume
$|q_2|\geq 0$. If for some $i$ is $|q_i| > 0$ let $\lambda$ be an element of
valuation zero such that its residue coefficient $b$ is $0<b<<1$ then $\lambda
a_j +a_i$, $j\in {1,2}, j\neq i$, $trop(\lambda a_j +a_i)=p$. Thus, we may
assume that $|q_i|\leq 0$ for all $i$. But again, since $A$ is singular,
$|q_i|=0$ for all $i$. If for some $i$ is $a_{ii}>0$ we are done and $a_i$ is a
lift of $p$. Last, we have the case that $q_i=0^-$ for all $i$. In this case,
we perform Gauss reduction on the matrix $A$. The coefficients we have to
multiply $a_1$ in order to make zeros below the first column are $0<\lambda_i$,
$v(\lambda_i)=0$. For all the elements $a_{ij}$, $i\neq 1$ $a_{ij}+\lambda_i
a_{1j}$ is a positive element of valuation $0$. On the other hand
$v(a_{ii}+\lambda a_{1i})\geq 0$.
Now, if for some $i\neq 1$, $v(a_{ii}+\lambda_i a_{1i})>0$, let $k\notin
\{1,i\}$, let $0<\eta$ with, $v(\eta)=0$, $0<Pc(\eta)<<1$ then $b=\eta a_k +
(a_i+\lambda_ia_1)$ is a vector $b\in V$ such that $b_i>0$ and $v(b_i)=0$ and
we
are done.
Next case is if for some $i\neq 1$, $a_{ii}+\lambda_ia_{1i}$ is a positive
element of valuation $0$, then for $0<\mu$ of valuation 0, $1<<Pc(\mu)$,
$-a_1+\mu(a_i+\lambda_i a_1)$ is the desired point.
Finally, if for all $i$, $a_{ii}+\lambda_ia_{1i}$ is a
negative element of valuation $0$, then we perform Gauss reduction on the
second column and repeat the process. Since $rank(A)\leq n-2$, after a finite
number of steps, we must arrive to a matrix (after reordering the indices) of
the form
\[\begin{pmatrix}a_{11} & a_{12} & \ldots & a_{1k} & a_{1,k+1} & \ldots &
a_{1n}\\
0 & b_{22} & \ldots & b_{2k} & b_{2,k+1} & \ldots &
b_{2n}\\
\multicolumn{7}{c}{\ldots}\\
0 & 0 & \ldots & b_{k-1,k-1} & b_{k-1,k} & \ldots & b_{k-1,n}\\
0 & 0 & \ldots & 0 & b_{k,k} & \ldots & b_{k,n}\\
\multicolumn{7}{c}{\ldots}\\
0 & 0 & \ldots & 0 & b_{n,k} & \ldots & b_{nn}\\
\end{pmatrix}\]
with real tropicalization
\[\begin{pmatrix}0^- & 0^+ & \ldots & 0^+ & 0^+ & \ldots & 0^+\\
\infty & 0^- & \ldots & 0^+ & 0^+ & \ldots & 0^+\\
\multicolumn{7}{c}{\ldots}\\
\infty & \infty & \ldots & 0^- & 0^+ & \ldots & 0^+\\
\infty & \infty & \ldots & \infty & c_{k,k}^{+} & \ldots & c_{k,n}^+\\
\multicolumn{7}{c}{\ldots}\\
\infty & \infty & \ldots & \infty & c_{n,k}^{+} & \ldots & c_{nn}^+\\
\end{pmatrix}\]
Such that $k=dim(V)\leq n-2$; $b_{kk}\neq 0$; $b_{ij}=0$ if $i>j<k$;
$trop(b_{ij})=0^+$,
if $i<j$ or $i>j\geq k$; $trop(b_{ii})=0^-$ for
$i<k$; vectors $b_{k+1},\ldots b_{n}$ are all multiple of $b_k$. Hence, it must
happen that $trop(b_{kk})=0^+$. Then, for $\eta_2,\ldots, \eta_k$
of valuation zero and positive such that $1<<Pc(\eta_2) << Pc(\eta_3)<<\ldots <<
Pc(\eta_k)$ the vector $-b_1-\eta_2b_2-\ldots -\eta_{k-1}b_{k-1}+\eta_kb_k$ is a
lift of $p$ in $V$.
\end{proof}
\begin{theorem}\label{teo:minimal_support_is_tropical basis}
Let $I$ be the ideal of an affine space $V$ in $(\mathbb{K}^*)$. The affine
polynomials in $I$ with minimal support form a real tropical basis of $I$.
\end{theorem}
\begin{proof}
For each minimal support there is only one linear polynomial in $I$ up to
multiplication by a constant and they generate $I$. Call $\{F_1,\ldots, F_r\}$
this generator set of $I$. Clearly, $trop(V)\subseteq \cap_{i=1}^r
\mathcal{T}_\mathbb{R}(F_i)$. To see the equality, let $p\notin trop(V)$. Homogenizing
if necessary and performing a monomial change of coordinates, we may assume,
without loss of generality that $I$ is homogeneous and that $p=(0^+,\ldots,
0^+)$. By Lemma~\ref{lem:trop-basis-linear}, there is a linear function $F\in
V$ such that $p\notin \mathcal{T}_\mathbb{R}(F)$. It suffices to check that we can take
$F$ with minimal support. If $F$ has no minimal support, then there exists
another linear $G\in I$ whose support is contained in the support of $F$.
Multiply $F$ and $G$ by appropriate constants so that the minimum valuation of
the coefficients of $F$ and of $G$ is zero. By substituting $G$ with a linear
combination of $F$ and $G$, we may also assume that the residue polynomials at
$p$ of $F$ and $G$ are linearly independent forms in $\mathbb{R}[x_1, \ldots, x_n]$.
Let $S_F$ (resp. $S_G$) be the variables where $F$ (resp. $G$) attains the
minimum at $p$. If there is a monomial $g_mx_m$ in $S_G$ that is not in $S_F$,
we can use $G$ to make zero the monomial $x_m$ in $F$ without modifying the
residue polynomial of $F$ at $p$ and proceed recursively.
Assume then that $S_G\subseteq S_F$, after reordering, we may assume that
$S_F=\{x_{i_1},\ldots, x_{i_r}\}$. Let $a_i$ (resp. $b_i$) be the principal
coefficient $F$ (resp. $G$) at the monomial $x_i$. Let $i$ be an index in $S_F$
such that $|b_i/a_i|$ is maximum of all the indices. If $b_i>0$ substitute $G$
by $-G$ so that $b_i<0$. Then, the polynomial $H=F\cdot(-bi/ai)+G$,
$Supp(H)\subsetneq Supp(F)$ and the residue polynomial of $H$ at $p$ has only
positive coefficients. We proceed recursively until we arrive to a linear
polynomial $H'$ with minimal support in $I$ such that $p\notin
\mathcal{T}_\mathbb{R}(H')$.
\end{proof}
We finish this section showing a real radical ideal that has no tropical basis.
This \textbf{dissonance} has been studied in \cite{real-radical-tropical}, but
the example showed there is not real radical.
\begin{example}\label{ex:real_radical_mala_trop}
Consider the cubic defined by:
\[F=x^3+y^3-x^2y-xy^2+2x^2+2y^2+4xy-8x-8y+8\]
The only point with positive coefficients is the singular point $(1,1)$ of
valuation $(0,0)$.
The corresponding tropical polynomial is
\[f = 0^+v^3\oplus 0^+w^3\oplus 0^-v^2w\oplus 0^-vw^2\oplus 0^+2v^2\oplus
0^+w^2\oplus 0^+vw\oplus 0^-v\oplus 0^-w\oplus 0^+\]
In the positive tropical quadrant the tropical polynomial defines a line-like
tropical curve.
\begin{center}
\begin{tikzpicture}
\draw (1,0)--(0,0)--(0,1);
\draw (0,0) -- (-0.7,-0.7);
\draw[fill] (0,0) circle(2pt);
\draw (4,0) -- (3.3,-0.7);
\draw[fill] (4,0) circle(2pt);
\end{tikzpicture}\\
Positive part of a tropical singular cubic (left)\\ and the set of points in
every tropicalization of a polynomial in $(F)$ (right).
\end{center}
We claim that the curve $C=\{F=0\}$ has no real tropical basis. Since the only
positive point of the curve is the singular point $(1,1)$, then the only
tropical positive point is the tropical point $(0^+,0^+)$. The ideal of $C$ is
$(F)$ and this is a real radical ideal. We are showing that, for every $a<0$,
$(a^+,a^+)\in \mathcal{T}_\mathbb{R}(FG)$ for every $G\neq 0$, $G\in \mathbb{K}[x,y]$.
Let $G$ be any real polynomial and $a<0$. Let $G_a$ be the residue polynomial
of $G$ at the point $(a,a)$. Then,
$(FG)_{(a,a)}=(x^3+y^3-2xy^2-2x^2y)G_a=(x+y)(x-y)^2G_a$. Since $(1,1)$ is a
root
of this polynomial, there must be two monomials in $(FG)_{(a,a)}$ with
different sign
where $(a,a)$ attains its minimum. That is $(a^+,a^+)\in
\mathcal{T}_\mathbb{R}(FG)$.
On the other hand, let us take a look at the points of the form $(a^+,0^+)$ and
$(0^+,a^+)$ with $a>0$. The monomials where the minimum of $(0,a)$ is attained
are $F_{(0,a)}=x^3 + 2x^2 - 8x + 8$. This polynomial have no positive root,
hence, by P\'olya's theorem $F_{(0,a)}(1+x)^n$ has only positive monomials.
$(F(1+x)^{11})_{(0,a)}= x^{14} + 13 x^{13} + 69 x^{12} + 195 x^{11} + 308
x^{10}
+ 242 x^{9} + 66 x^{8} + 198 x^{7} + 825 x^{6} + 1441 x^{5} + 1441 x^{4} + 903
x^{3} + 354 x^{2} + 80 x + 8$.
Hence, in the tropical positive quadrant
\[\bigcap_{G}\mathcal{T}_\mathbb{R}(FG) = \{(a^+,a^+)| a\leq 0\}\neq
trop(C)=\{(0^+,0^+)\}\]
\end{example}
\section{The Real Tropical Discriminant}\label{sec:singular}
In this section we extend the study of tropical singularities and discriminants
\cite{tropical-discriminant}, \cite{singular-tropical-hypersurfaces},
\cite{sing-fixed-point} to the real case.
Consider the algebraic closed field $\mathbb{K}[i]$. Let $A\subseteq \mathbb{Z}^n$.
$|A|=d$ such that $<A>=\mathbb{Z}^n$. Consider the Laurent polynomial ring
$\mathbb{K}[i][x_1^{\pm 1},\ldots, x_n^{\pm 1}]$. In this ring we have the
$\mathbb{K}[i]$-linear space of polynomials with support $A$, $F=\sum_{i\in
A}a_i x^i$. We may identify any such polynomial with the point $(a_i|i\in A)\in
\mathbb{K}[i]^d$. Consider the incidence variety
\[H = \{(F,u)\in (\mathbb{K}[i]^*)^d\times (\mathbb{K}[i]^*)^n\ |\ F \textrm{
is singular at } u\}\]
of singular hypersurfaces of support $A$ and singular points. This is a
$\mathbb{Q}$-defined variety. Consider the projections
$\pi_1:(\mathbb{K}[i]^*)^d \times (\mathbb{K}[i]^*)^n\rightarrow
(\mathbb{K}[i]^*)^d$ and $\pi_2: (\mathbb{K}[i]^*)^d\times (\mathbb{K}[i]^*)^n
\rightarrow (\mathbb{K}[i]^*)^n$. $\pi_1(H)$ is the $A$-discriminant variety of
hypersurfaces with support $A$ and a singular point, while $\pi_2(H) =
(\mathbb{K}^*)^n$, and the fiber over any point is a linear space isomorphic to
$\pi_2^{-1}(1,\ldots, 1)=\{F\ |\ F$ is singular at $(1,\ldots,1)\}$. Our aim is
to describe the real tropicalization of $\pi_1(H)$.
\begin{definition}
Let $f\in \mathbb{TR}[w_1,\ldots, w_n]$ be a tropical polynomial with signs.
Let $p=(p_1,\ldots, p_n)\in \mathcal{T}_\mathbb{R}(f)$. we say that $p$ is a
\textbf{singular point} of $f$ if there exists a pair $F\in
\mathbb{K}[x_1,\ldots, x_n]$ and $P\in \mathbb{K}^n$ such that $P$ is a singular point
of $\{F=0\}$, $trop(F)=f$ and $trop(P)=p$.
\end{definition}
The definition of tropical Euler derivative
\cite{singular-tropical-hypersurfaces} can be easily extended to tropical
polynomials with signs.
\begin{definition}
Let $f = \bigoplus_{\ell\in A} a_\ell w^\ell\in \mathbb{TR}[w_1,\ldots, w_n]$
and $<A>=\mathbb{Z}^n$ be a tropical polynomial with signs. Let
$L=b_0+b_1w_1+\ldots + b_nw_n$ be an affine function with integer coefficients.
The \textbf{Euler derivative} of $f$ with respect to $L$ is
\[\bigoplus_{\substack{\ell \in A\\ L(\ell)\neq 0}} s(L(\ell)) a_\ell w^\ell.\]
We eliminate all the monomials in the support where $L$ vanish and swap signs
of the monomials where $L$ is negative. If $F\in \mathbb{K}[x_1,\ldots, x_n]$ is any
polynomial such that $trop(F)=f$ then
\[\frac{\partial f}{\partial L} = T\left(\frac{\partial F}{\partial
L}\right)=T\left (b_0F + b_1x_1\frac{\partial
F}{\partial x_1}+\ldots +b_nx_n\frac{\partial F}{\partial x_n}\right )\]
\end{definition}
\begin{example}
Let $f=0^+ \oplus 0^+w_1 \oplus 0^+w_2 \oplus 0^+ w_1^2\oplus 0^+w_1w_2 \oplus
0^+w_2^2$. Then $\frac{\partial f}{\partial w_1-w_2} = 0^+w_1 \oplus 0^-w_2
\oplus 0^+ w_1^2 \oplus 0^-w_2^2$
\end{example}
We have the following consequence of
Theorem~\ref{teo:minimal_support_is_tropical basis}
\begin{theorem}
Let $f$ be a tropical polynomial with signs. The set of tropical singularities
of $\mathcal{T}_\mathbb{R}(f)$ is
\[\bigcap_{L} \mathcal{T}_\mathbb{R}\left(\frac{\partial f }{\partial
L}\right)\]
\end{theorem}
\begin{proof}
If $p$ is a singularity of $\mathcal{T}_\mathbb{R}(f)$, then let $P\in (\mathbb{K}^*)^n$ and
$F\in \mathbb{K}[x_1,\ldots, x_n]$ be a polynomial, with $trop(P)=p$, $trop(F)=f$, and
$P$
is in the singular locus of $\{F=0\}$. Then, for any $L$, $\frac{\partial
F}{\partial L}(P)=0$. So $p\in \mathcal{T}_\mathbb{R}(\frac{\partial f}{\partial L})$.
Conversely, let $p\in \bigcap_{L} \mathcal{T}_\mathbb{R}\left(\frac{\partial f
}{\partial
L}\right)$. Without loss of generality, assume that $p=(0^+,\ldots, 0^+)$.
The set tropical polynomials with signs having a singularity at $p$ and support
$A$ is a linear space that is the tropicalization of the linear system $H_1$ of
polynomials with support $A$ in $\mathbb{K}[x_1,\ldots, x_n]$ having a singularity at
$P_1=(1,\ldots, 1)$. Let $F=\sum_{\ell\in A}a_\ell x^\ell$ be the generic
polynomial with indeterminate coefficients and support $A$. The linear system
$H_1$ is generated by $F(P_1)$, $(x_1\frac{\partial F}{\partial x_1})(P_1)$,
$\ldots$, $(x_n\frac{\partial F}{\partial x_n})(P_1)$. By
Lemma~\ref{lem:trop-basis-linear},
\[trop(H_1)=\bigcap_{L}\mathcal{T}_f(\frac{\partial F}{\partial
L})=\bigcap_{L}\mathcal{T}_f(\frac{\partial f}{\partial L})\]
Hence, there exists $F\in H_1$ with a singularity at $(1,\ldots, 1)$ and $p$ is
a singularity of $f$.
\end{proof}
\begin{definition}[cf.
\cite{Bergman-complex},\cite{tropical-discriminant},\cite{sing-fixed-point},
\cite{singular-tropical-hypersurfaces}]
Let $p$ be a point in $\mathcal{T}_\mathbb{R}(f)$. We define the \textbf{flag} of $f$
with respect to $p$ as the flag of subsets $\mathcal{F}(p)$ of $A$ defined
inductively by:
$\mathcal{F}_{-1}=\emptyset \subsetneq \mathcal{F}_0(p) \subsetneq
\mathcal{F}_1(p) \subsetneq\ldots\subsetneq \mathcal{F}_r(p)$, $dim
<\!\mathcal{F}_r (p)\!> = n$, and for any $\ell$: $\mathcal{F}_{\ell +1} (p) -
\mathcal{F}_\ell (p)$ is
the subset of $A- <\!\mathcal{F}_\ell(p)\!>$ where the tropical polynomial
$\oplus_{j \in A-<\!\mathcal{F}_\ell\!>} a_j w^j$ attains its minimum at
$p$. The weight class of the flag $\mathcal{F}(p)$ are all the points $p'\in
\mathcal{T}_\mathbb{R}(f)$ for which $\mathcal{F}(p)=\mathcal{F}(p')$. If $p$ and $p'$
are in the same weight class and $s(p)=s(p')$ then $p$ is singular if and only
if $p'$ is singular.
\end{definition}
Let $f\in \mathbb{TR}[w_1,\ldots, w_d]$ and $p=(p_1,\ldots, p_d)\in \mathbb{R}$ be a
singular point of the usual tropical variety defined by $f$ forgetting signs
(refer to \cite{singular-tropical-hypersurfaces}. We want to know if
$p^+=(p_1^+,\ldots, p_n^+)$ is a singular point of $\mathcal{T}_\mathbb{R}(f)$. Let
$\mathcal{F}_{-1}= \emptyset \subsetneq \mathcal{F}_0(p)\subsetneq
\mathcal{F}_1(p)\subsetneq \ldots\subsetneq
\mathcal{F}_r(p)$ be the \textbf{weight class} of $p$. For any $\mathcal{F}_i$,
write $\mathcal{F}_i=\mathcal{F}_i^+\cup \mathcal{F}_i^-$, where
$\mathcal{F}_i^+$ (resp. $\mathcal{F}_i^-$) are the monomials whose
corresponding tropical coefficient is positive (resp. negative).
\begin{definition}
Let $A,B$ disjoint sets. Let $L$ be a hyperplane, we say that $L$
\textbf{separates} $A$ and $B$ if $L$ does not contain $A\cup B$ and the sets
are contained in different closed halfspaces defined by $L$. That is $(A\cup
B)\not \subseteq L$, $A\subseteq L_{\geq 0}$ and $B\subseteq L_{\leq 0}$ (or
$A\subseteq L_{\leq 0}$ and $B\subseteq L_{\geq 0}$).
\end{definition}
Note, that, for example, if $A\subseteq L$ and $B\subseteq L_{\geq0}$ then $L$
separates $A$ and $B$. Also, if $A=\emptyset$, any $L$ not containing $B$
separates $A$ and $B$.
\begin{theorem}
With the previous notation, let $p\in \mathbb{TR}^n$ be a point with all its
coordinates
positive. Then, $p$ is a singular point of $\mathcal{T}_\mathbb{R}(f)$ if and only if
for all $i=0,\ldots, r-1$ and for all hyperplane $L$ such that
$\mathcal{F}_{i-1}(p)\subseteq L$, $\mathcal{F}_{i}(p)\not\subseteq L$, $L$
does
not separate $\mathcal{F}_i^+$ and $\mathcal{F}_i^-$.
\end{theorem}
\begin{proof}
Let $L$ be a hyperplane such that $\mathcal{F}_{i-1}(p)\subseteq L$ and $L$
separates $\mathcal{F_i}^+$ and $\mathcal{F_i}^-$. Then, the set of monomials
where $\frac{\partial f}{\partial L}$ attains its minimum at $p$ is
$\mathcal{F}_i(p)$. But, since $L$ separates $\mathcal{F}_i^+$ and
$\mathcal{F}_i^-$, all the monomials in $\mathcal{F}_i(p)$ yield the same sign,
so $p\notin \mathcal{T}_\mathbb{R}(\frac{\partial f}{\partial L})$ and $p$ is not
singular.
Assume now that $p$ is not a singular point. Then, there exists a $L$ such that
$p\notin \mathcal{T}_\mathbb{R}(\frac{\partial f}{\partial L})$. The monomials where
the
minimum of $\frac{\partial f}{\partial L}$ is attained at $p$ is one of the
sets
of the flag $\mathcal{F}_i(p)$. Since $p$ is not in the tropical variety
defined
by this derivative, it follows that $L$ separates $\mathcal{F}_i^+$ and
$\mathcal{F}_i^-$.
\end{proof}
\begin{example}
Consider the real tropical cubic \[f = 0^+v^3\oplus 0^+w^3\oplus 0^-v^2w\oplus
0^-vw^2\oplus 0^+v^2\oplus 0^+w^2\oplus 0^+vw\oplus 0^-v\oplus 0^-w\oplus 0^+\]
The weight classes in the positive orthant are: $\{(0^+,0^+)\}$, $\{(a^+, a^+)|
a<0\}$, $\{(0^+,a^+)| a>0\}$, $\{(a^+,0^+)| a>0\}$. The points in the weight
class $\{(a^+, a^+)| a<0\}$ are not singular, since they do not belong to the
real tropical variety defined by $\frac{\partial f}{\partial v+w-3}$. The
points
of the form $\{(0^+,a^+)| a>0\}$ are all singular,
$\mathcal{F}_0^+=\{1,v^2,v^3\}$, $\mathcal{F}_0^-=\{v\}$. No $L$ separates
$\mathcal{F}_0^+$ and $\mathcal{F}_0^-$ and $(0^+,a^+)$ belongs to the tropical
curve defined by $\frac{\partial f}{\partial w} = 0^+w^3\oplus 0^-v^2w\oplus
0^-vw^2\oplus 0^+w^2\oplus 0^+vw\oplus 0^-w$. Similarly $\{(a^+,0^+)| a>0\}$
are all singular points. Finally the point $(0^+,0^+)$ is also singular, as
shown in Example~\ref{ex:real_radical_mala_trop}. For a point of the form
$(0^+,a^+)$ the polynomial $F=2x^3 + (t^{2a} + t^a - 1)x^2y + (-2t^a - 2)xy^2 +
y^3 + 2x^2 + 2xy + 2y^2 + (-10)x + (-t^a - 1)y + 6$, with $trop(F)=f$ has a
singularity in $(1,t^a)$ and a similar result can be obtained for $(a^+,0^+)$.
\end{example}
\begin{example}
Let $p=(0^+,0^+)\in \mathbb{TR}^2$. We would like to describe the real tropical curves
with a singularity at $p$. We only describe its maximal cones. Such a
description in the complex case appears in \cite{sing-fixed-point} so we just
discuss which are the valid distribution of signs on the coefficient of the
tropical polynomial $f$. Up to symmetry and swap of signs the (maximal)
combinatorial types of polynomials with a singularity at $p$ are.
\begin{enumerate}
\item $\mathcal{F}_0(p)$ is a two-dimensional circuit (four points, each three
affinely independent). Then, the positive and negative monomials of the circuit
cannot be separated by a line.
\item $\mathcal{F}_0(p)$ is a one-dimensional circuit (three collinear points)
contained in a line $L$. The vertices of the circuit are of the same sign and
the interior monomial is of the opposite sign. In this case, $\mathcal{F}_1$
consists on two monomials, we distinguish two cases:
\begin{enumerate}
\item The monomials of $\mathcal{F}_1$ lie on different halfspaces with respect
to $L$. In this case, the two monomials have the same sign.
\item Both monomials of $\mathcal{F}_1$ lie on the same halfspace with respect
to $L$. In this case, the two monomials have different signs.
\end{enumerate}
\end{enumerate}
\end{example}
\begin{center}
\begin{figure}
\begin{tikzpicture}
\draw (-1,0)--(1,0)--(0,1.73)--(-1,0);
\draw[fill] (0,0.58) circle(2pt) (-1,0) circle(2pt) (1,0) circle(2pt) (0,1.73)
circle (2pt);
\draw (-1.1,-0.2) node {$+$} (1.1,-0.2) node {$+$} (0,1.99) node {$+$};
\draw (0.2,0.68) node {$-$};
\draw (3,0) -- (5,0) -- (5,2) -- (3,2) -- (3,0);
\draw[fill] (3,0) circle(2pt) (5,0) circle(2pt) (5,2) circle(2pt) (3,2)
circle(2pt) (3,0);
\draw (2.9,-0.2) node {$+$} (5.1,2.2) node {$+$};
\draw (5.1,-0.2) node {$-$} (2.9,2.2) node {$-$};
\end{tikzpicture}
\caption{Possible distribution of signs on a circuit of dimension
2.}\label{fig:circuit2}
\end{figure}
\end{center}
\begin{center}
\begin{figure}
\begin{tikzpicture}
\draw (0,1) -- (0,0) -- (0,-1);
\draw[fill] (0,1) circle(2pt) (0,0) circle(2pt) (0,-1) circle(2pt);
\draw (-1,0) circle(2pt) (1,0) circle(2pt);
\draw (0.2,0.2) node {$-$} (0.2,1.2) node {$+$} (0.2,-0.8) node {$+$};
\draw (1.2,0.2) node {$\pm$} (-0.8,0.2) node {$\pm$};
\draw (5,1)--(5,0)--(5,-1);
\draw[fill] (5,1) circle(2pt) (5,0) circle(2pt) (5,-1) circle(2pt);
\draw (6,-0.5) circle(2pt) (6,0.5) circle(2pt);
\draw (5.2,0.2) node {$-$} (5.2,1.2) node {$+$} (5.2,-0.8) node {$+$};
\draw (6.2,0.7) node {$\pm$} (6.2,-0.3) node {$\mp$};
\end{tikzpicture}
\caption{Possible distribution of signs around a circuit of dimension 1. The
monomials in $\mathcal{F}_1$ are the white circles.}\label{circuit1}
\end{figure}
\end{center}
|
\section{Introduction}
The behaviour of liquids at interfaces and in confinement is a fascinating and important area of study. For example, the behaviour of a liquid under confinement between two surfaces determines how good a lubricant that liquid is. The nature of the interactions between the liquid and the surfaces is crucial. Consider, for example, the teflon coating on non-stick cooking pans, used because water does not adhere to (wet) the surface. One can approach the problem from a mesoscopic fluid-mechanical point of view, see for example the excellent book by de Gennes, Brochard-Wyart and Quer\'e.\cite{deGennes04} However, if a microscopic approach is required, which relates the fluid properties at an interface to the nature of the molecular interactions, then one must start from statistical mechanics. There are a number of books such as Refs.~\onlinecite{Henderson92,Rowlinson02,Davis96,Hansen06} which provide a good starting point. All of these include a discussion on classical density functional theory (DFT) which is a theory for determining the density profile of a fluid in the presence of an external potential, such as that exerted by the walls of a container.
DFT is a statistical mechanical theory, where the aim is to calculate average properties of the system being studied. In statistical mechanics, the central quantity of interest is the partition function $Z$ and once this is calculated, all thermodynamic quantities are given. However, $Z$ a sum over all the possible configurations of the system, can rarely be evaluated exactly. Instead of focussing on $Z$, in DFT we seek to develop good approximations for the free energy. It can be shown that the free energy is a functional of the fluid density profile $\rho(\ul{r})$ and the equilibrium profile is that which minimises the free energy. Over the years, a great many different approximations for the free energy functionals have been developed, generally by making contact with results from other branches of liquid-state physics. There are now quite a few lecture notes and review articles on the subject.\cite{Evans79, Evans92, Lutsko10, Wu07, Wu06, Tarazona08, Lowen10} This rather large literature can make learning about DFT rather daunting. One of us (AJA) has found in teaching this subject that a good place for students to start learning about the properties of inhomogeneous fluids, is by considering a simple lattice gas (Ising) model. This allows students to avoid much of the liquid-state physics and functional calculus that can be daunting for undergraduates when embarking on studying DFT and its applications.\cite{endnote1} The advantage of starting from a lattice-gas model is that one can quickly develop a simple mean-field DFT (described below) and then proceed to calculate the bulk fluid phase diagram and study the interfacial properties of the model, determining the wetting behaviour, finding wetting transitions and all the other interesting phenomenology of liquids at interfaces. The computer algorithms required to solve these equations are fairly simple. Thus, the threshold for entering the subject and getting to the point where a student can calculate things for themselves is much lower via this route, than most other routes that we can think of.
The aim of this paper is two-fold: (i) to derive the mean-field DFT for an inhomogeneous lattice-gas fluid, whilst explaining the physics of the theory. This presentation assumes the reader has had introductory statistical mechanics and thermodynamics courses, but little else beyond that. (ii) To illustrate the types of quantities that DFT can be used to calculate, such as the surface tension of the liquid-gas interface, to study wetting behaviour or to answer the question ``what is the shape of a drop of liquid on a surface?'' We also give some exercises for students.
This paper is laid out as follows: in \S\ref{sec:continuum} we introduce the statistical mechanics of simple liquids. We set up the DFT model in \S\ref{sec:discrete} and \ref{sec:potentials}. The bulk fluid phase diagram is discussed in \S\ref{sec:phase}. We describe the iterative method for solving the model in \S\ref{sec:solving} before displaying some typical results in \S\ref{sec:results}. Finally, some conclusions are drawn in \S\ref{sec:conclusions}.
\section{Statistical Mechanics of Simple Liquids}\label{sec:continuum}
We consider a fluid composed of $N$ atoms/molecules in a container. What follows is also relevant to colloidal suspensions and so we simply refer to the atoms, molecules, colloids, etc as `particles'. The energy is a function of the set of position and momentum coordinates, $\ul{r}^N \equiv \{ \ul{r}_1,\ul{r}_2,\dots , \ul{r}_N \} $ and $\ul{p}^N \equiv \{ \ul{p}_1,\ul{p}_2,\dots , \ul{p}_N \} $ respectively, and is given by the Hamiltonian \cite{Hansen06}
\begin{equation}\label{eq:hamiltonian}
{\cal H}(\ul{r}^N,\ul{p}^N) = K(\ul{p}^N) + E(\ul{r}^N),
\end{equation}
where $K$ is the kinetic energy
\begin{equation}\label{eq:kinetic}
K=\sum_{i=1}^N \frac{\ul{p}_i^2}{2m},
\end{equation}
and $E$ is the potential energy due to the interactions between the particles and also to any external potentials such as those due to the container walls. When treating the system in the canonical ensemble, which has fixed volume $V$, particle number $N$ and temperature $T$, the probability that the system is in a particular state is\cite{Hansen06, Davis96, Plischke06, Reichl09}
\begin{equation}
f(\ul{r}^N,\ul{p}^N)=\frac{1}{h^{3N}N!}\frac{e^{- \beta {\cal H}}}{Z},
\end{equation}
where
\begin{equation}\label{eq:cPartition}
Z = \frac{1}{h^{3N} N!} \int d \ul{r}^N \int d \ul{p}^N e^{- \beta {\cal H}},
\end{equation}
is the canonical partition function, $h$ is Plank's constant and $\beta = (k_BT)^{-1}$ where $k_B$ is Boltzmann's constant. The partition function allows macroscopic thermodynamic quantities to be related to the microscopic properties of the system which are defined in ${\cal H}$ (see below).
The kinetic energy contribution \eqref{eq:kinetic}, is solely a function of the momenta $\ul{p}^N$, and $E(\ul{r}^N)$, the precise form of which is yet to be defined, only depends on the positions of the particles $\ul{r}^N$. This allows the partition function \eqref{eq:cPartition} to be simplified by performing the Gaussian integrals over the momenta to obtain
\begin{align}
Z & = \frac{1}{h^{3N}} \int d\ul{p}^N e^{- \beta \sum_{i=1}^N \frac{\ul{p}_i^2}{2m}} Q, \nonumber \\
& = \frac{1}{h^{3N}} \int e^{- \beta \frac{\ul{p}_1^2}{2m}} d\ul{p}_1 \dots \int e^{- \beta \frac{\ul{p}_N^2}{2m}} d\ul{p}_N Q, \nonumber \\
& = \frac{1}{h^{3N}} \left(\sqrt{\frac{2m \pi}{ \beta}}\right)^{3} \dots \left(\sqrt{\frac{2m \pi}{ \beta}}\right)^{3} Q, \nonumber \\
& = \left(\sqrt{\frac{2m \pi}{ \beta h^2}}\right)^{3N} Q, \nonumber \\
& = \Lambda^{-3N}Q,
\end{align}
where $ \Lambda$ is the thermal de Broglie wavelength and
\begin{equation}\label{eq:configInt}
Q = \frac{1}{N!}\int d\ul{r}^N e^{- \beta E},
\end{equation}
is the configuration integral.\cite{Plischke06} Thus, the partition function is just the configuration integral multiplied by a factor that depends on $N$, $T$ and $m$ and so the value of $\Lambda$ is irrelevant for determining the state of the system. Changing $ \Lambda$ just adds a constant to the free energy per particle [see Eq.~\eqref{eq:helmholtz1}] and so we safely assume $ \Lambda=1$.
Evaluating $Q$ is the central problem here and, in general, this can not be done and so approximations are required. In the following section we develop a simple lattice model approximation that allows progress. Note that the system described above has been analysed in the canonical ensemble. We discuss below how the system can instead be considered in the grand canonical ensemble.
\section{Discrete Model}\label{sec:discrete}
\subsection{Defining a Lattice}
\begin{figure}[t]
\begin{center}
\begin{subfigure}[!]{0.2\textwidth}
\begin{center}
\includegraphics{adam_hughesFig01a.pdf}
\caption{}\label{subfig:nolattice}
\end{center}
\end{subfigure}
\begin{subfigure}[!]{0.2\textheight}
\begin{center}
\includegraphics{adam_hughesFig01b.pdf}
\caption{}\label{subfig:lattice}
\end{center}
\end{subfigure}
\caption{Illustration of how a free system, (a), may be discretized in space by setting the particles on a lattice, (b).}
\label{fig:lattice}
\end{center}
\end{figure}
We assume that the fluid is two dimensional (2D), to simplify the analysis. However, everything can easily be extended to a three dimensional (3D) system. We imagine a lattice discretises the space occupied by the fluid and so any configuration of particles may be described by a set of lattice occupation numbers $\{n_1,n_2,\dots, n_N\} \equiv \{n_i\}$ which define if the lattice sites are filled ($n_i=1$) or empty ($n_i=0$) with $n_i$ being the occupation number of site $i$. The width of each lattice site is set as $\sigma$, the diameter of a particle, and there are $M$ sites. We set $ \sigma=1$ throughout and use this as our unit of length. The particles are assumed to be spherical, so that their orientation is not important. We now find that the configurational integral in Eq.~\eqref{eq:configInt} becomes a sum over the lattice sites. Note the short hand $i\equiv(k,l)$, where $k$ and $l$ are integer indices defining the 2D lattice.
\subsection{Energy of the System}
To proceed, we must define the potential energy contribution to the Hamiltonian, $E$. We assume the following form
\begin{equation}\label{eq:energy}
E=\sum_{i=1}^M n_iV_i-\sum_{i,j} \epsilon_{ij} n_in_j.
\end{equation}
The first term is the contribution from the external potential $V_i$ and the second term is the energy contribution from pair interactions between particles. We assume that there are no three-body or higher interactions. The interaction energy between particles at two lattice sites $i$ and $j$ is $ \epsilon_{ij}$. This gets smaller as the distance between them increases and so $ \epsilon_{ij}$ has the property that as $|i-j| \rightarrow \infty$, $ \epsilon_{ij} \rightarrow 0$. The term $-\sum_{i,j} \epsilon_{ij} n_in_j $ denotes a sum over all pairs of lattice sites in the system. Considering only pair interactions greatly simplifies the task of evaluating the partition function, but it can still be very arduous to evaluate this sum even for a moderately sized system. The probability of being in a particular configuration, $\{n_i\}$, for a fixed number of particles $N$, is now
\begin{equation}\label{eq:probability}
P(\{n_i\})= \frac{e^{- \beta E(\{n_i\})}}{Z},
\end{equation}
with the partition function defined as
\begin{equation}\label{eq:partition}
Z = \sum_{\textrm{all states}} e^{- \beta E_{\textrm{state}}},
\end{equation}
where `state' is a shorthand for a particular allowed set of occupation numbers $\{n_i\}$. Note the relation to the configuration integral in Eq.~\eqref{eq:configInt}, since the sum over all states approximates the continuum integral $(N!)^{-1}\int d\ul{r}^N (\cdots)$.
\subsection{Helmholtz Free Energy}
The Helmholtz free energy is related to the partition function as follows\cite{Hansen06, Plischke06, Reichl09}
\begin{equation}\label{eq:helmholtz1}
F = - k_BT \ln Z.
\end{equation}
All other thermodynamic quantities are obtained from derivatives of $F$. However, we are still unable to evaluate the sum in Eq.~\eqref{eq:partition} and as a consequence can not calculate $F$. Under certain assumptions we can make some progress: consider the system where there is no external field, i.e.\ $V_i=0$, and that $\epsilon_{ij}=0$, so that the particles do not interact with each other. From Eq.~\eqref{eq:energy}, this gives $E=0$ for all configurations and from Eq.~\eqref{eq:probability} we observe that
\begin{equation}
P(\{n_i\}) = \frac{1}{Z},
\end{equation}
i.e.\ that all configurations are equally likely. From Eq.~\eqref{eq:partition}, we see that $Z$ is just the number of possible states, which, for a system of $M$ lattice sites containing $N$ particles, is
\begin{equation}
Z= \frac{M!}{N!(M-N)!}.
\end{equation}
For large systems, i.e.\ when both $M$ and $N$ are large, this can be simplified using Stirling's approximation, $\ln(N!) \approx N \ln N - N$, which, with Eq.~\eqref{eq:helmholtz1}, gives
\begin{equation}\label{eq:helmholtz2}
F = -k_BT \left[M\ln M - N\ln N - (M-N)\ln(M-N) \right].
\end{equation}
The number density of particles in the system is ${\rho = N/M}$ (recall $\sigma=1$) and so Eq.~\eqref{eq:helmholtz2} gives
\begin{equation}\label{eq:helmholtz3}
F = Mk_BT \left[ \rho \ln \rho + (1- \rho) \ln(1- \rho)\right].
\end{equation}
This homogeneous fluid has a uniform density $\rho$ throughout. However, for an inhomogeneous fluid in the presence of a spatially varying external potential $V_i$ we should expect the density to vary in space. The average density at lattice point $i$ is defined as
\begin{equation}\label{eq:rho_defn}
\rho_i=\langle n_i \rangle,
\end{equation}
i.e.\ it is the average value of the occupation number at site $i$, over all possible configurations: $\langle \cdots \rangle=\sum_{\textrm{all states}} (\cdots) P_{\textrm{state}}$. We now obtain an approximation for the free energy of the inhomogeneous fluid.
\subsection{The Grand Canonical Ensemble}
We previously treated the system in the canonical ensemble with a fixed $N$, $T$ and volume $V$ (strictly, this is an area since the fluid is 2D but we refer to area as `volume' throughout). Now we consider the system in the grand canonical ensemble with fixed $V$ and $T$ but now $N$ can vary by exchanging particles with a reservoir. The reservoir has a fixed chemical potential $\mu$, and as the system is connected to this reservoir it has the same chemical potential (recall that the chemical potential is the energy required to insert a particle into the system). Physically, the easiest way to conceive the grand canonical ensemble is to imagine the system as being a subsystem of a much larger structure, with which it can exchange particles, and where the reservoir fixes $T$ and $ \mu$ in the subsystem.
The probability of a grand canonical system being in a particular state is [cf.\ Eq.~\eqref{eq:probability}]
\begin{equation}\label{eq:gprob}
P(\{n_i\}) = \frac{e^{- \beta(E - \mu N)}}{ \Xi},
\end{equation}
where the number of particles in the system is
\begin{equation*}
N = \sum_{i=1}^M n_i,
\end{equation*}
The normalisation factor $\Xi$ is the grand canonical partition function
\begin{equation}\label{eq:gpartition}
\Xi = \Tr e^{- \beta(E - \mu N)},
\end{equation}
where the trace operator, $\Tr$, is defined as
\begin{equation*}
\Tr x = \sum_{\textrm{all states}} x = \sum_{n_1=0}^1 \sum_{n_2=0}^1 \dots \sum_{n_M=0}^1 x.
\end{equation*}
From the grand canonical partition function we can find the grand potential
\begin{equation}\label{eq:grandpot}
\Omega = -k_BT \ln \Xi,
\end{equation}
in an analogous manner to which the Helmholtz free energy is obtained in the canonical ensemble [cf.\ Eq.~\eqref{eq:helmholtz1}]. The equilibrium state corresponds to the minimum of the grand potential.
\subsection{Gibbs-Bogoliubov Inequality}
We now derive and then use the Gibbs-Bogoliubov inequality to show that there exists an upper bound on the free energy and finding the minimum of this bound gives an approximation to the true free energy.
Eq.~\eqref{eq:grandpot} can be rearranged and equated to Eq.~\eqref{eq:gpartition} to give
\begin{equation}\label{eq:equating}
e^{- \beta \Omega} = \Tr e^{- \beta( E - \mu N)}.
\end{equation}
The energy of a particular state $E$ can be rewritten as
\begin{equation}\label{eq:splitenergy}
E = E_0 + E - E_0 = E_0 + \Delta E,
\end{equation}
where $E_0$ is the energy of a reference system which we choose so as to be able to evaluate the partition function. We choose the system with $\epsilon_{ij}\equiv 0$ and $V_i \neq 0$. From Eq.~\eqref{eq:equating} and \eqref{eq:splitenergy} we obtain
\begin{equation}\label{eq:bbov1}
e^{- \beta \Omega} = \Tr e^{- \beta(E_0 - \mu N)} e^{ - \beta \Delta E}.
\end{equation}
The statistical average value of any quantity $x$ in the reference system is
\begin{equation*}
\langle x \rangle_0 = \Tr \left(\frac{e^{-\beta(E_0-\mu N)}}{\Xi_0}\ x\right),
\end{equation*}
since $P_0=e^{- \beta(E_0 - \mu N)}/ \Xi_0$ [see Eq.~\eqref{eq:gprob}]. So, from \eqref{eq:bbov1} we obtain
\begin{equation}\label{eq:bbov2}
e^{-\beta \Omega} = e^{-\beta \Omega_0} \langle e^{-\beta \Delta E}\rangle_0,
\end{equation}
with $\Xi_0 =e^{-\beta \Omega_0}$ given by Eqs.~\eqref{eq:gpartition} and \eqref{eq:grandpot}. Now, since $e^{-x}$ is a convex function of $x$, then $\langle e^{-x}\rangle \geq e^{-\langle x \rangle}$ and from Eq.~\eqref{eq:bbov2} we obtain the inequality
\begin{equation}
e^{-\beta \Omega} \geq e^{-\beta \Omega_0} e^{-\beta \langle \Delta E\rangle_0}.
\end{equation}
Taking the logarithm of this gives the Gibbs-Bogoliubov inequality\cite{Hansen06}
\begin{equation}\label{eq:bbovFinal}
\Omega \leq \Omega_0 + \langle \Delta E \rangle_0.
\end{equation}
This shows that there is an upper bound to the true grand potential $\Omega$ that depends solely on the properties of the reference system, and, more importantly, it allows us to find a `best' approximation for $\Omega$ by minimising the right hand side of the inequality. We choose $E_0$ to depend upon parameters that may be varied and perform the minimisation with respect to variations in these parameters.
To proceed, we must define $E_0$. We choose
\begin{equation}\label{eq:e0def}
E_0 = \sum_{i=1}^M (V_i +\phi_i)n_i,
\end{equation}
where $V_i$ is the external potential, and $\phi_i$ are the parameters mentioned above, which are yet to be determined. Physically, they are the (mean field) additional effective potentials that incorporate the effect of the interactions between the particles.
The density at a particular lattice site is given by Eq.~\eqref{eq:rho_defn}. Our (mean field) approximation for this quantity is
\begin{widetext}
\begin{align}
\rho_i = \langle n_i \rangle_0 = & \Tr \left( \frac{e^{-\beta E_0- \mu N}}{Z_0} n_i \right), \nonumber\\
= & \frac{1}{Z_0} \left[\sum_{n_1=0}^1e^{-\beta (V_1+\phi_1- \mu)n_1}\right]\dots\left[\sum_{n_i=0}^1n_ie^{-\beta (V_i+\phi_i- \mu)n_i}\right]\dots\left[\sum_{n_M=0}^1e^{-\beta (V_M+\phi_M- \mu)n_M}\right], \nonumber \\
=& \left[ \frac{\sum_{n_1=0}^1e^{-\beta (V_1+\phi_1- \mu)n_1}}{\sum_{n_1=0}^1e^{-\beta (V_1+\phi_1- \mu)n_1}}\right]\dots \left[ \frac{\sum_{n_i=0}^1n_ie^{-\beta (V_i+\phi_i-\mu)n_i}}{\sum_{n_i=0}^1e^{-\beta (V_i+\phi_i-\mu)n_i}}\right] \dots \left[ \frac{\sum_{n_M=0}^1e^{-\beta (V_M+\phi_M-\mu)n_M}}{\sum_{n_M=0}^1e^{-\beta (V_M+\phi_M-\mu)n_M}}\right], \nonumber \\
= & \frac{\sum_{n_i=0}^1n_ie^{-\beta (V_i+\phi_i-\mu)n_i}}{\sum_{n_i=0}^1e^{-\beta (V_i+\phi_i-\mu)n_i}
\hspace{1cm} = \frac{e^{-\beta(V_i+\phi_i-\mu)}}{1+e^{-\beta(V_i+\phi_i-\mu)}}.\label{eq:averageni}
\end{align}
\end{widetext}
Also, the reference system partition function is [cf.\ Eq.~\eqref{eq:gpartition}]:
\begin{align*}
\Xi_0 = & \Tr e^{-\beta(E_0-\mu N)}, \\
= & \Tr e^{-\beta \sum_{i=1}^M(V_i+\phi_i-\mu)n_i}, \\
= & \prod_{i=1}^M (1+ e^{-\beta(V_i+\phi_i-\mu)}).
\end{align*}
This may then be substituted into Eq.~\eqref{eq:grandpot} to obtain the following expression for the grand potential
\begin{equation}\label{eq:omeg0}
\Omega_0 = -k_BT \sum_{i=1}^M \ln (1+ e^{-\beta(V_i+\phi_i-\mu)}).
\end{equation}
Rearranging \eqref{eq:averageni} to give ${1- \rho_i=(1+e^{- \beta(V_i+ \phi_i - \mu)})^{-1}}$ and inserting it into \eqref{eq:omeg0} gives
\begin{equation}
\Omega_0 = k_BT \sum_{i=1}^M \ln(1-\rho_i).
\end{equation}
By rewriting this as
\begin{equation}
\Omega_0 =k_BT \sum_{i=1}^M ( \rho_i + 1 - \rho_i) \ln(1 - \rho_i),
\end{equation}
we can use Eq.~\eqref{eq:averageni} to express $ \Omega_0$ in the following form
\begin{align}\label{eq:omega0def}
\Omega_0 = k_BT & \sum_{i=1}^M \left[ \rho_i \ln \rho_i + (1- \rho_i) \ln( 1-\rho_i) \right] \nonumber \\
+ & \sum_{i=1}^M \left(V_i+ \phi_i - \mu\right) \rho_i.
\end{align}
Note that when $V_i= \phi_i=0$, which corresponds to the case of a uniform fluid with $ \epsilon_{ij}=0$, this reduces to the result we saw earlier in Eq.~\eqref{eq:helmholtz3}, since $ \Omega = F - \mu N$.\cite{Hansen06,Chandler87} Returning to the general case $ \epsilon_{ij} \neq 0$, from the definition of $E_0$ in Eq.~\eqref{eq:e0def}, we find that $ \Delta E = E-E_0$ is
\begin{equation}\label{eq:deltaens}
\Delta E = -\sum_{i,j} \epsilon_{ij} n_i n_j - \sum_{i=1}^M \phi_i n_i.
\end{equation}
From Eq.~\eqref{eq:averageni}, that $ \rho_i = \langle n_i\rangle_0$, with \eqref{eq:deltaens} this gives
\begin{equation}\label{eq:deltaedef}
\langle \Delta E \rangle_0 = -\sum_{i,j} \epsilon_{ij} \rho_i \rho_j - \sum_{i=1}^M \phi_i \rho_i,
\end{equation}
where, because our reference system is non-interacting, we find that
$$\langle n_i n_j \rangle_0 = \langle n_i\rangle_0\langle n_j\rangle_0 = \rho_i \rho_j.$$
Finally, Eqs.~\eqref{eq:omega0def} and \eqref{eq:deltaedef} can be used to obtain
\begin{align}
\hat{\Omega} & = \Omega_0 + \langle \Delta E \rangle_0 , \nonumber\\
& = k_BT \sum_{i=1}^M \left[ \rho_i \ln \rho_i + (1- \rho_i) \ln( 1-\rho_i) \right] \nonumber \\
& \qquad \ -\sum_{i,j} \epsilon_{ij} \rho_i \rho_j + \sum_{i=1}^M (V_i- \mu) \rho_i. \label{eq:fullomega}
\end{align}
As discussed previously, this is an upper bound to the true grand potential $ \Omega$. One should choose the mean field $\{ \phi_i\}$ so as to minimise $\hat{ \Omega}$, in order to generate a best approximation for $ \Omega$. This is equivalent to choosing the set $\{ \rho_i\}$ so as to minimise $\hat{ \Omega}$, since the density $ \rho_i$ is defined by $ \phi_i$ [cf.\ Eq.~\eqref{eq:averageni}]. What we have done here is to derive an approximate DFT for the lattice fluid. For DFT in general, one can prove that the equilibrium fluid density profile is that which minimises the grand potential functional. \cite{Evans79}
\section{Defining the Potentials}\label{sec:potentials}
Up to this point, we have not specified the form of the potentials from the external field, or the particle interactions. We now define $ \epsilon_{ij}$ and $V_i$.
\subsection{Particle Interactions}\label{sec:IVA}
\begin{figure}[t]
\begin{center}
\begin{subfigure}{0.18\textheight}
\centering
\includegraphics{adam_hughesFig02a.pdf}
\caption{}
\label{subfig:nearestNeighbours}
\end{subfigure}
\begin{subfigure}{0.18\textheight}
\centering
\includegraphics{adam_hughesFig02b.pdf}
\caption{}
\label{subfig:nextNearest}
\end{subfigure}
\caption{The distinction between, (a), the nearest neighbors (open circles) to a particle (grey circle), and, (b), the next nearest neighbors.}
\label{fig:nearest}
\end{center}
\end{figure}
The term $-\sum_{i,j} \epsilon_{ij} \rho_i \rho_j$ in Eq.~\eqref{eq:fullomega} represents the contribution to the free energy from the interactions between pairs of particles. A simple example of the continuum fluid we seek to model is made up of particles interacting via a Lennard-Jones pair potential \cite{Hansen06} of the form $v(r) = \epsilon \left[ \left(\frac{r_0}{r}\right)^{12} -2 \left(\frac{r_0}{r}\right)^6\right]$, where $r$ is the distance between pairs of particles and $r_0$ is the distance at the minimum where $v(r_0)=- \epsilon$. Given that $v(2r_0) \approx -0.03 \epsilon$, it is a good approximation to assume that each particle only interacts with the nearest and next nearest neighbouring particles, as illustrated in Fig.~\ref{fig:nearest}, and so we replace the particle interaction term in the free energy with
\begin{equation}\label{eq:internalpotential}
\sum_{i,j} \epsilon_{ij} \rho_i \rho_j \approx \epsilon_{nn} \sum_{i=1}^M \rho_i \sum_{jnni} \rho_j + \epsilon_{nnn} \sum_{i=1}^M \rho_i \sum_{jnnni} \rho_j,
\end{equation}
where $ \epsilon_{nn}$ and $ \epsilon_{nnn}$ are the strengths of the interaction between nearest neighbour and next nearest neighbour particles, respectively. The term $\sum_{jnni} \rho_j$ denotes the sum of densities in lattice sites $j$ which are the nearest neighbours to the site $i$. Similarly, $\sum_{jnnni} \rho_j $ denotes the sum over the next nearest neighbours. We now set $ \epsilon_{nn}= \epsilon$ and $ \epsilon_{nnn}= \epsilon/4$. This ratio $ \epsilon_{nn}/ \epsilon_{nnn}$ is not the value it would have if the Lennard-Jones potential were exactly applied but it is the optimum ratio to obtain circular drops when solved in two dimensions (see \S\ref{sec:drops} below).\cite{Robbins12,Robbins11,Fomel97} Our definition captures the essence of the Lennard-Jones potential: the repulsive core is modelled by the onsite repulsion (one particle per lattice site) and the pair interaction terms crudely model the attractive forces. However, it is worth noting that even though the interaction energy between two well-separated $(r \gg r_0)$ particles can be very small, the net contribution from all such long-range interactions may be significant and neglecting them may result in the theory failing to describe some interesting physics.
\subsection{External Potential}
We assume that the interaction potential between a particle and the particles that form the wall of the container is of the Lennard-Jones form which decays for large $r$ as $v(r) \sim -r^{-6}$. Summing the potential between a single fluid particle with all of the particles in the wall yields a net potential that decays as $V(z)\sim -z^{-3}$ for $z \rightarrow \infty$, where $z$ is the perpendicular distance between the particle and the wall. We therefore assume that the wall exerts a potential of the form
\begin{equation}\label{eq:extPot}
V_i =
\begin{cases}
\infty \quad & \textrm{if } k<1\\
-\epsilon_wk^{-3} \quad & \textrm{if } k \geq 1
\end{cases}
\end{equation}
where $ \epsilon_w$ is the parameter which defines the attractive strength of the confining wall. The integer index $k$ is the distance, in the number of lattice sites, of the particle from the wall.
\section{The Bulk Fluid Phase Diagram}\label{sec:phase}
Before discussing the behaviour of the fluid at this wall, we first calculate the phase diagram of the bulk fluid, away from the influence of any interfaces. When the temperature $T$ is less than the critical temperature $T_c$, the fluid exhibits phase separation into a low density gas phase and a high density liquid phase. The binodal is the line in the phase diagram at which this transition occurs. Along the binodal, the liquid and the gas coexist in thermodynamic equilibrium, i.e.\ where the pressure, chemical potential and temperature of the liquid and gas phases are equal. The lattice gas model has a hole-particle symmetry that is not present in a continuum description, but which is useful for calculating the binodal. This symmetry arises from the fact that if we replace $n_i=1-h_i$, where $h_i$ is the hole occupation number, then the form of Eq.~\eqref{eq:energy} is unchanged. This symmetry leads to the density of the coexisting gas and liquid, $ \rho_g$ and $ \rho_l$ respectively, to be related as
\begin{equation}\label{eq:phase1}
\rho_l= 1-\rho_g.
\end{equation}
From Eq.~\eqref{eq:fullomega} the Helmholtz free energy per lattice site, $f = F/M$, for a uniform fluid with density $ \rho$, is
\begin{equation}\label{eq:phaseextra}
f=k_BT \left[ \rho \ln \rho + (1- \rho) \ln (1- \rho)\right] - \frac{ 5 \epsilon}{2} \rho^2
\end{equation}
where $5 \epsilon/2 = \sum_{i,j} \epsilon_{ij}$ is the sum up to the next nearest neighbours interactions and includes a factor of a half to prevent double counting. The pressure in the system is\cite{Hansen06,Mandl88}
\begin{align}\label{eq:phase2}
P( \rho) & = -\left(\frac{\partial F}{\partial V}\right)_{T,N} \nonumber \\
& = \rho \frac{\partial f}{\partial \rho} - f \nonumber \\
& = -k_BT \ln(1- \rho) - \frac{5}{2} \epsilon \rho^2
\end{align}
The binodal curve can be found by invoking Eq.~\eqref{eq:phase1} and solving $P( \rho) = P(1- \rho)$, giving
\begin{equation}\label{eq:phase4}
\frac{k_BT}{ \epsilon} = \frac{5(2 \rho-1)}{2(\ln \rho - \ln(1- \rho))},
\end{equation}
which is displayed in Fig.~\ref{subfig:densityPhase}. The maximum on the binodal corresponds to the critical point, above which there is no gas-liquid phase separation. From the symmetry \eqref{eq:phase1}, the density at the critical point is $ \rho=1/2$ and the critical temperature is found to be $T_c= 5 \epsilon/4k_B$.
\begin{figure}
\begin{center}
\begin{subfigure}{0.3\textheight}
\includegraphics[width=0.3\textheight]{adam_hughesFig03a.pdf}
\captionsetup{margin={0.65cm,0cm}}
\caption{}
\label{subfig:densityPhase}
\end{subfigure}
\begin{subfigure}{0.3\textheight}
\includegraphics[width=0.3\textheight]{adam_hughesFig03b.pdf}
\captionsetup{margin={0.65cm,0cm}}
\caption{}
\label{subfig:chempotPhase}
\end{subfigure}
\caption{The bulk fluid phase diagram for the 2D lattice fluid. The solid red line is the binodal and the dashed blue line is the spinodal. In (a) we display the phase diagram in the dimensionless temperature $k_BT/ \epsilon$ versus density plane and in (b) as a function of chemical potential.}
\label{fig:phase1}
\end{center}
\end{figure}
The chemical potential can also be calculated from the Helmholtz free energy as\cite{Hansen06,Mandl88}
\begin{align}\label{eq:phase5}
\mu( \rho) & = \left(\frac{\partial F}{\partial N}\right)_{T,V} \nonumber \\
& = \frac{\partial f}{\partial \rho}\nonumber \\
& = k_BT \ln \left( \frac{ \rho}{1- \rho} \right) - 5 \epsilon \rho.
\end{align}
On substituting \eqref{eq:phase4} into \eqref{eq:phase5} we find that the chemical potential at coexistence is
\begin{equation}\label{eq:phase6}
\mu_{coex} = - \frac{ 5}{2}\epsilon,
\end{equation}
which is displayed in Fig.~\ref{subfig:chempotPhase}. The spinodal is also plotted in Fig.~\ref{fig:phase1}. The spinodal denotes the locus in the phase diagram where the compressibility is zero, i.e.\ within this curve the fluid is unstable and spontaneous phase separation occurs. The spinodal is obtained from the following condition
\begin{equation}\label{eq:phase7}
\frac{\partial^2 f}{\partial \rho^2} = 0,
\end{equation}
which, from Eq.~\eqref{eq:phaseextra}, gives the following expression for the density dependence of the temperature along the spinodal,
\begin{equation}\label{eq:phase8}
\frac{k_BT}{ \epsilon} = 5 \rho(1- \rho),
\end{equation}
also plotted in Fig.~\ref{subfig:densityPhase}. The spinodal can also be obtained as a function of $\mu$ from Eqs.\ \eqref{eq:phase5} and \eqref{eq:phase8}. The result is displayed in Fig.~\ref{subfig:chempotPhase}.
\subsection*{Exercise:}
Calculate the binodal for the case when there are only nearest neighbour interactions. What is the critical temperature?
\section{An Iterative Method for Calculating the Density Profile}\label{sec:solving}
We return now to the inhomogeneous fluid in the presence of an external potential. The equilibrium density profile is that which minimises $\hat{\Omega}$ in Eq.~\eqref{eq:fullomega}, i.e.\ it is the set $\{\rho_i\}$ which satisfiy, for all $i$,
\begin{equation*}
\frac{\partial \hat{ \Omega}}{\partial \rho_i}=0.
\end{equation*}
Performing this differentiation and rearranging gives the set of coupled equations,
\begin{equation}\label{eq:iteration}
\rho_i = (1- \rho_i) \exp\left[ \beta \left( \mu + \epsilon \sum_{jnni} \rho_j + \frac{\epsilon}{4}\sum_{jnnni} \rho_j - V_i\right) \right],
\end{equation}
which can be solved iteratively for the profile $\{\rho_i\}$. An initial approximation is required and the closer this is to the true solution, the better. We sometimes use $ \rho_i=\exp(\beta(\mu-V_i))$, which is the exact result in the low density (ideal-gas) limit or we may simply guess a likely profile. We can also use values from previous state points as an initial approximation when calculating at several state points successively, incrementing one parameter each time. With a suitable initial approximation for $\{\rho_i\}$, Eq.~\eqref{eq:iteration} can then be iterated until convergence is achieved.
It is often necessary during each iterative step to mix the result from evaluating the right hand side of Eq.~\eqref{eq:iteration}, $\rho_i^{\textrm{rhs}}$, in a linear combination with the result from the previous iteration $\rho_i^{\textrm{old}}$, i.e.
\begin{equation}\label{eq:mixing}
\rho_i^{\textrm{new}} = \alpha \rho_i^{\textrm{rhs}} + (1-\alpha) \rho_i^{\textrm{old}},
\end{equation}
where $\alpha$ may be small, typically in the range ${0.01<\alpha<0.1}$. This has the effect that only small steps are taken towards the minimum with each iteration. Omitting this mixing (i.e.\ $\alpha=1$) can give a $\rho_i^{\textrm{new}}$ that falls outside of the range (0,1) and once this happens the iterative routine breaks down.
\subsection{Normalising the Density Profile}\label{sec:VIA}
To describe an enclosed (canonical) system with fixed $N$, rather than being coupled to a reservoir which fixes $\mu$, we can think of Eq.~\eqref{eq:fullomega} as a constrained minimisation, i.e.\ as minimising the Helmholtz free energy
\begin{align}
F = k_BT & \sum_{i=1}^M \left[ \rho_i \ln \rho_i + (1- \rho_i) \ln( 1-\rho_i) \right] \nonumber \\
- & \sum_{i,j} \epsilon_{ij} \rho_i \rho_j + \sum_{i=1}^M V_i \rho_i,
\end{align}
subject to the constraint that
\begin{equation}\label{eq:constraint}
N = \sum_{i=1}^M \rho_i.
\end{equation}
The chemical potential $ \mu$ is then the Lagrange multiplier. To achieve this when iteratively calculating the density profile $\{\rho_i\}$, we modify the method described above and at each iteration following \eqref{eq:mixing} the profile is renormalised: $\rho_i^{\textrm{norm}} = A\rho_i^{\textrm{new}}$, with
\begin{equation*}
A = N \left(\sum_{i=1}^M \rho_i^{\textrm{new}}\right)^{-1},
\end{equation*}
so that the constraint \eqref{eq:constraint} is satisfied.
\subsection{Boundary Conditions}\label{sec:boundary}
At the wall, the boundary conditions (BC) for the density profile are straight forward: we simply set $\rho_i=0$ for all lattice sites $i$ `inside' the wall -- i.e.\ for $k<1$ in Eq.~\eqref{eq:extPot}. On the boundaries perpendicular to the wall, we normally use periodic BC, where it is assumed that the nearest neighbour of a lattice site on the boundary is the lattice site on the opposite boundary. For the boundary opposite the wall, periodic BC in this situation creates an artificial substrate (i.e.\ so that the fluid is confined in a capillary, between two walls). This does not cause a problem in sufficiently large systems. However, a more efficient solution is to assume that the fluid is uniform beyond the boundary opposite the wall, with specified density, e.g.\ that of the bulk gas.
\section{Typical Solutions}\label{sec:results}
We now present results using the lattice gas model which are typical of many DFT models for a fluid exhibiting gas-liquid phase separation. After determining the equilibrium density profile using the iterative method described above we may then calculate thermodynamic quantities such as the interfacial tension or the adsorption at the wall, which is defined as
\begin{equation}\label{eq:norm3}
\Gamma=\sum_{i=1}^M( \rho_i- \rho_b),
\end{equation}
where $ \rho_b$ is the bulk density which is obtained by solving Eq.~\eqref{eq:phase5} for $\rho$. Note that $\Gamma$ is an excess number per area; the formula in Eq.~\eqref{eq:norm3} is only true when $\sigma=1$. By calculating results in the grand canonical ensemble we can track how the adsorption changes with $\mu$ (\S\ref{sec:adsorption}). Working in the canonical ensemble we can find drop profiles and calculate the contact angle that the liquid drop makes with a substrate (\S\ref{sec:drops}). From these results we can also determine if the liquid wets the substrate. We characterise a liquid as wetting a substrate when, at liquid-gas coexistence, a macroscopically thick layer of the liquid forms between the gas and the substrate.\cite{Evans89, Schick90, dietrich, BoRo01, BEIMR09, StVe09, Parry, BME87} Grand canonically, where particles are free to enter and leave the system, wetting is characterised by $\Gamma\to\infty$ as coexistence is approached $\mu\to\mu_{coex}^-$. Treating it canonically, the number of particles in the system is fixed $N=\sum_{i=1}^M\rho_i$ (using the normalisation discussed in \S\ref{sec:VIA}) so $\Gamma$ is fixed and we characterise wetting by the contact angle that a liquid drop makes with the substrate. In both cases, wetting only occurs when it is energetically beneficial, i.e.\ the liquid wetting the substrate is the state of least energy.
\subsection{One Dimensional Model}
So far, we have assumed for simplicity that the fluid is in 2D. However, since the density profile is defined as an average over all possible configurations [c.f.\ Eq.~\eqref{eq:rho_defn}], then if the external potential only varies in one direction [such as the potential in Eq.~\eqref{eq:extPot}], then so must the density profile. This is, of course, also the case for the 3D fluid. The equilibrium density profile must have the same symmetry as the external potential and so we may reduce the DFT equations to be solved \eqref{eq:iteration} to a one-dimensional (1D) system, consisting of a line of lattice sites extending perpendicularly away from the wall. We do this by summing over the interactions in the (transverse) direction in which the density does not vary, as illustrated in Fig.\ \ref{fig:2d-1d}. This maps the 2D system onto an effective 1D system with renormalised interactions between lattice sites and also introduces an effective on-site interaction. A similar mapping can also be done for the 3D fluid.
\subsection*{Exercise:}
(i) Implement the procedure described in \S\ref{sec:solving} on a computer for calculating the density profiles for this effective 1D model. (ii) Modify your computer code to solve for the density profile in 2D. (iii) Compare results from the two. Are they the same?
\begin{figure}[t]
\begin{center}
\begin{subfigure}{0.18\textheight}
\centering
\includegraphics{adam_hughesFig04a.pdf}
\caption{}
\label{subfig:2d}
\end{subfigure}
\begin{subfigure}{0.18\textheight}
\centering
\includegraphics{adam_hughesFig04b.pdf}
\caption{}
\label{subfig:1d}
\end{subfigure}
\caption{Illustration of the mapping of the full 2D particle pair interactions (a) onto an effective 1D system (b). The numbers represent the contribution towards the potential (in units of $\epsilon$) from that particular lattice site with reference to the shaded particle in the centre. The 2D case on the left is that discussed above in \S\ref{sec:IVA} and on the right we display the resulting effective potential after mapping this system to 1D.}
\label{fig:2d-1d}
\end{center}
\end{figure}
\subsection{Adsorption at the wall}\label{sec:adsorption}
In Fig.~\ref{fig:adsorptions}(a) we illustrate how $\Gamma$, the adsorption at the wall, changes as the chemical potential is increased $\mu\to\mu_{coex}^-$, to approach the coexistence value in \eqref{eq:phase6}, from below. When $\mu<\mu_{coex}$ the bulk phase (away from the wall) is the gas phase, but for a wall to which the particles are attracted, the density at the substrate can be higher. As $\mu\to\mu_{coex}^-$, the adsorption increases, either diverging $\Gamma\to\infty$, when the liquid wets the wall, or remaining finite, when the liquid does not wet the wall. As $T$ or $\epsilon_w$ are changed, there is often a phase transition from one regime to the other, termed the `wetting transition'.\cite{Evans89, Schick90, dietrich, BoRo01, BEIMR09, StVe09, Parry, BME87}
The adsorption results in Fig.~\ref{fig:adsorptions}(a) are calculated for fixed $\beta\epsilon=1.2$. When the strength of the attraction due to the wall is weak, $\beta\epsilon_w<1.2$, the liquid does not wet the wall and the adsorption remains finite at coexistence, $\mu=\mu_{coex}$. However, for stronger attraction, $\beta\epsilon_w>1.2$, the wetting film thickness diverges as $\mu\to\mu_{coex}^-$. To compute these results a value of $\mu$ is set and the equilibrium profile $\{\rho_i\}$ is found. The value of $\mu$ is then incremented and the previous equilibrium solution used as the initial approximation for the next solution. At each state point the adsorption is calculated via Eq.~\eqref{eq:norm3}.
\begin{figure}
\begin{center}
\begin{subfigure}[!]{0.3\textheight}
\includegraphics[width=0.3\textheight]{adam_hughesFig05a.pdf}
\caption{}
\label{subfig:adsorb1}
\end{subfigure}
\begin{subfigure}[!]{0.3\textheight}
\includegraphics[width=0.3\textheight]{adam_hughesFig05b.pdf}
\caption{}
\label{subfig:adsorb2}
\end{subfigure}
\begin{subfigure}[!]{0.3\textheight}
\includegraphics[width=0.3\textheight]{adam_hughesFig05c.pdf}
\caption{}
\label{subfig:adsorb3}
\end{subfigure}
\caption{(a) The adsorption at the wall as the chemical potential $\mu\to\mu_{coex}^-$ for various different values of the wall attraction strength parameter $\epsilon_w$, as given in the key, for $\beta\epsilon=1.2$. In (b) we display some of the corresponding density profiles for $ \beta \epsilon_w = 1.6$, at $\beta(\mu-\mu_{coex})= -0.2$, -0.108, -0.104, -0.1, -0.04, -0.004 and 0 and in (c) we display density profiles for $ \beta \epsilon_w = 2$, at $\beta(\mu-\mu_{coex})= -0.2$, -0.108, -0.1, -0.04, -0.004 and 0. The points in (a) denote state corresponding to the profiles in (b) and (c), with matching styles and colors (color online).}
\label{fig:adsorptions}
\end{center}
\end{figure}
An interesting thing to note is that for some values of $\epsilon_w$, the adsorption diverges continuously (see e.g.\ the case for $\beta\epsilon_w=2$), but for other values there is a discontinuous jump in $\Gamma$. This jump is a result of crossing the `pre-wetting line'.\cite{Evans89, Schick90, dietrich, BoRo01, BEIMR09, StVe09, Parry, BME87} We see the beginings of this jump as a continuous `shoulder' for $ \beta \epsilon_w=1.7$. As $ \beta \epsilon_w$ decreases the jump becomes larger and occurs closer to $ \mu = \mu_{coex}$. The adsorption for $ \beta \epsilon_w=1.2$ remains very small until almost exactly at $ \mu= \mu_{coex}$ where it jumps to a large value. We also observe some smaller discontinuous changes in $\Gamma$ occurring after the main pre-wetting jump. These smaller jumps are `layering transitions' and are due to an additional layer of particles being discontinuously added to the adsorbed liquid film. Whilst layering transitions are observed in more sophisticated DFT theories, the underlying lattice in the present model leads to an unrealistic amplification of this effect. Figs.\ \ref{subfig:adsorb2} and \ref{subfig:adsorb3} illustrate how the density profile changes as $\mu\to\mu_{coex}^-$, for values of $\epsilon_w$ that lead to wetting of the wall. We see a layer of the liquid phase appearing against the wall, increasing in thickness as coexistence is approached. In Fig.\ \ref{subfig:adsorb2} we also see how the density profiles change discontinuously as the pre-wetting line is crossed.
Tracking the adsorption is useful for understanding how the fluid behaves as coexistence is approached. However, it ought not be used as the sole indicator of the wetting behaviour. One should also calculate the grand potential $\Omega$. It can often arise that a given density profile actually corresponds only to a local minimum of $\Omega$, but in fact the global minimum corresponds to a different density profile (e.g.\ with higher adsorption).
\subsection*{Exercise:}
Set $\beta\epsilon_w=1.3$ and calculate the density profile at coexistence $\mu=\mu_{coex}$, for a range of different `temperatures', $\beta\epsilon$. What do you find?
\subsection{Drop Profiles and Surface Tensions}\label{sec:drops}
We now return to the full 2D model and show typical density profiles corresponding to drops of liquid on a surface acting with the potential in Eq.~\eqref{eq:extPot}. We treat the system canonically, i.e.\ we normalise the system as discussed in \S\ref{sec:VIA}. We also break translational symmetry, placing the centre of mass at the horizontal midpoint.\cite{ReguerraReiss,ArcherMalijevsky}
The initial approximation for initiating our iterative procedure consists of setting the density $\rho_i=\rho_g$ everywhere, apart from in a region in the middle of the system next to the wall, where we set $\rho_i=\rho_l$. The size of this portion defines the size of the final liquid drop. The boundary conditions are as described in \S\ref{sec:boundary}, with the wall at the bottom boundary, the left and right hand sides of the lattice having periodic boundary conditions and we fix $\rho_i=\rho_g$ along the top boundary.
\begin{figure}
\begin{center}
\includegraphics[width=0.32\textheight]{adam_hughesFig06.pdf}
\caption{Drop density profiles for fixed $ \beta \epsilon=1.2$ with varying values of $ \beta \epsilon_w$. The drops spread out with increasing $ \beta \epsilon_w$ until a flat film forms beyond the wetting transition.}
\label{fig:profiles}
\end{center}
\end{figure}
In Fig.~\ref{fig:profiles} we display some typical density profiles for various values of $ \beta\epsilon_w$, calculated on a $100 \times 40$ lattice, for the fluid with temperature $\beta\epsilon=1.2$. The adsorption (i.e.\ particle number) is the same in each. The liquid drop spreads out more on the substrate with the larger value of $\epsilon_w$. The contact angle $\theta$ that the drop makes with the substrate, decreases as $\epsilon_w$ is increased, so that the drop becomes broader, until complete wetting occurs at $\beta \epsilon_w \approx1.2$, when the drop becomes a flat film.
The interfacial tension (or `surface tension' in 3D) is the excess free energy due to the presence of an interface between two phase. In the present system there are three phases: the solid (wall), liquid and gas. Thus, there are three different interfacial tensions, for the wall-liquid, wall-gas and liquid-gas interfaces, $\gamma_{wl}$, $\gamma_{wg}$ and $\gamma_{lg}$, respectively. For just the liquid and gas together, the interfacial tension $\gamma_{lg}$ leads to a liquid drop surrounded by the gas to form a circle (in 2D, or a sphere in 3D), because this shape minimises the interfacial area and therefore its contribution to the free energy. When the wall is present, which can not change, the gas and liquid must arrange themselves so as to minimise the free energy. The resulting configuration depends on the values of the interfacial tensions. The equilibrium value of the contact angle is given by Young's Equation,\cite{Evans89, Schick90, dietrich, BoRo01, BEIMR09, StVe09}
\begin{equation}\label{eq:young}
\gamma_{lg} \cos\theta=\gamma_{wg}-\gamma_{wl},
\end{equation}
which can be understood by considering the balance of the forces due to the interfacial tensions, at the point where the three phases meet.
Within the present microscopic theory, we can calculate the interfacial tensions, enabling a comparison with the macroscopic arguments that lead to to Eq.~\eqref{eq:young}. To determine $\gamma_{lg}$, calculate the density profile through the interface between a semi-infinite slab of the liquid that is adjacent to a semi-infinite slab of the gas. This is obtained in the same manner as the density profiles at the wall in \S\ref{sec:adsorption}, but in this case we remove the wall (setting $V_i=0$ for all $i$) and set the boundary condition that $\rho_i=\rho_l$. At the other end, $\rho_i=\rho_g$, as before. The initial guess for the density profile consists of setting $\rho_i=\rho_l$ in one half of the system and $\rho_i=\rho_g$ in the other half and, of course, we must set $\mu=\mu_{coex}$. From the resulting profile $\{\rho_i\}$ we then calculate the free energy $\Omega$ from Eq.~\eqref{eq:fullomega}. The free energy without the interface (i.e.\ either full of gas or full of just the liquid) is
\begin{equation}
\Omega_0 = -pV,
\end{equation}
where $p$ is the pressure and $V$ is the volume (system size). The interfacial tension is then
\begin{equation}
\gamma_{lg}=\frac{\Omega - \Omega_0}{A}
\end{equation}
where $A$ is the length of the 2D interface. The wall-gas and wall-liquid interfacial tensions are calculated in a similar manner except we retain the wall potential and we initialise the system entirely with either the gas or the liquid density, respectively. Note that above we have solely discussed the interfacial tensions for a straight interfaces. For curved interfaces, the tensions depend on the curvature and the calculations become more involved. A discussion on some of the key issues can be found in Ref.~\onlinecite{stewart05} and references therein.
\begin{table}[t]
\centering
\begin{tabular}{cccc}
$\beta \epsilon_w$ & $\sigma\beta\gamma_{wl}$ & $\sigma\beta\gamma_{wg}$ & $\theta$\\
\hline
0.5 & 0.12 & -0.05 & 115$^\circ$ \\
0.8 & -0.16 & -0.09 & 79$^\circ$ \\
1.0 & -0.36 & -0.13 & 52$^\circ$ \\
1.3 & -0.68 & -0.23 & 0$^\circ$ \\
\hline
\end{tabular}
\caption{Interfacial tensions and contact angle $\theta$ from Eq.~\eqref{eq:young}, for different values of the wall attraction strength $\epsilon_w$.}
\label{tab:tensions}
\end{table}
When $\beta\epsilon=1.2$, the gas-liquid interfacial tension $\gamma_{lg}=0.38k_BT/\sigma$, corresponding to the case for the profiles displayed in Fig.~\ref{fig:profiles}. The other interfacial tensions are given in Table \ref{tab:tensions}, together with the resulting contact angle, from Eq.~\eqref{eq:young}. These are in good agreement with the contact angle one can observe from the density profiles in Fig.~\ref{fig:profiles}. However, these profiles have a diffuse interface, so there is always some uncertainty in the location of the contact line. As $\epsilon_w$ is increased the drop spreads because it is energetically beneficial to do so. Complete spreading (wetting) only occurs when the sum $\gamma_{lg}+\gamma_{wl}<\gamma_{wg}$.
\subsection*{Exercise:}
Calculate one of the density profiles from Fig.\ \ref{fig:profiles} implementing the normalisation procedure introduced in \S\ref{sec:VIA} and then plot the density contour $\rho=(\rho_g+\rho_l)/2$, that corresponds to the mid point of the liquid-gas interface. Where on this curve does the contact angle agrees with the macroscopic result in Eq.~\eqref{eq:young}? Is it where you would expect?
\section{Conclusions}\label{sec:conclusions}
We have presented a derivation of a simple lattice gas model DFT and discussed typical applications. Working with this model gives a good hands-on introduction to many of the important ideas behind DFT and gives a platform to learn about different aspects of inhomogeneous fluids such as phase diagrams, adsorption, wetting and surface tensions. Studying this `toy-model' gives students good insight and a feeling for the physics of inhomogeneous liquids, leaving them in a good position to go on and study the `real thing'.\cite{Henderson92, Rowlinson02, Davis96, Hansen06, Evans79, Lutsko10, Wu07, Wu06, Tarazona08, Lowen10}
\section*{Acknowledgements}
APH acknowledges support through a Loughborough University Graduate School Studentship. AJA thanks all the students who have done projects with him modelling inhomogeneous liquids with this lattice-gas DFT or variants of it. This paper is largely based on informal lectures and many discussions with Blesson Chacko, William Dewey, Mark Robbins and Sen Tian.
|
\section{Introduction}\label{sec:intro}
Wigner crystals~\cite{Wigcrys} composed of trapped and mutually-repelling ions are an outstanding prototype of tailored condensed matter \cite{Dubin_RMP,zig1,Bollinger:1,Drewsen:1,Bloch:2}. The high degree of control they offer makes them an ideal platform for quantum information
processing \cite{CiracZoller,WinelandQC,BlattQC}, quantum communication devices~\cite{Ioncommunicator,Yelin,Drewsen_EIT}
and quantum simulators~\cite{Feynman, Bloch:1,Schneider, Friedenauer,Monroe,Ionsimulator,PorrasCirac,Ultralewen}.
Moreover, they constitute a perfect playground for studying general, distinctive features of condensed-phase systems; above all, phase transitions and critical phenomena~\cite{Mussardo, Sachdev,zig1,Monroe,Bollinger:2,Drewsen:2,Mainz,Morigi:G.5,Shimshoni:2011a,Bermudez}.
One prominent example is the linear ion chain \cite{Dubin_RMP,zig1,Paul1}, which results from the interplay between long-range Coulomb repulsion and a highly anisotropic confinement due to an ion trap \cite{Ghosh}. This quasi-ordered structure can become unstable depending on the trap aspect ratio or on the ion density. Fig.~\ref{fig:design} illustrates the two equilibrium configurations: Here, zigzag order (right panel) becomes energetically favourable at lower transverse confinements \cite{zig1,Paul1,Dubin1993,Schieffer1993,Morigi:G.5}. While it was often argued in the literature that this structural instability is a continuous phase transition \cite{Dubin1993,Schieffer1993}, in Ref. \cite{Morigi:G.5} it was first demonstrated that it can be rigorously mapped to a Landau second-order phase transition in the appropriately defined thermodynamic limit
\cite{Morigi:G.7,DengPorrasCirac}. This structural instability gathered a lot of interest in the latest years as a laboratory system for studying quenches across critical points \cite{DeChiara,Baltrusch,Exp:KZ,Exp:KZ:2}, and also because it is believed to be a promising channel for transport and storage of quantum entanglement \cite{Ioncommunicator,Yelin}.
\begin{figure}
\begin{center}
\begin{overpic}[width = \columnwidth, unit=1pt]{Figure1.jpg}
\put(200, 3){$y$}
\put(167, 40){$x$}
\end{overpic}
\end{center}
\caption{ \label{fig:design} (color online)
Linear-zigzag instability in a chain of interacting atoms: (left) linear and (right) zigzag configurations. The transition is either controlled by changing the longitudinal interparticle spacing $a$ or the frequency $\omega_T$ of the transverse harmonic confinement~\cite{zig1,Morigi:G.5}.
The displacement of a particle from the axis $y = 0$ plays the role of local order parameter, discriminating between
the disordered (linear) phase and the ordered (zigzag) phase.
}
\end{figure}
Quantum effects at the critical point have been theoretically studied in Ref. \cite{Retzker:2008} for small ion chains. In Refs. \cite{Meyer,Shimshoni:2011a} it has been argued that in the thermodynamic limit the linear-zigzag structural instability is a quantum phase transition, which in two ($1+1$) dimensions can be mapped to an Ising model in the transverse field, describing a ferromagnetic transition at zero temperature. This mapping was first proposed for Wigner crystals of electrons in quantum wires \cite{Meyer}, and then derived in Ref. \cite{Shimshoni:2011a,Shimshoni:2011b} using the emergent $\mathbb{Z}_2$ symmetry. In Ref. \cite{Shimshoni:2011b}, in particular, parameter regimes were estimated for which the quantum phase transition could be experimentally measured. It was noticed that, while for ions achieving the quantum critical region can be experimentally challenging, it could be more easily accessed with other kinds of strongly-correlated systems, for instance, dipolar atomic gases\cite{
Menotti} in elongated traps. In this context, we mention that theoretical studies of the linear-zigzag instability with ultracold dipolar systems appeared in Refs. \cite{Morigi:Dipoli,Altman}. Related phenomena were identified in arrays of dipolar tubes \cite{Kollath}.
This framework motivates an accurate characterization of the quantum behavior at criticality, which shall provide ultimate evidence of the universality class of the quantum linear-zigzag instability and allow one to determine the parameter ranges where it can be experimentally measured. Starting from Ref. \cite{Morigi:G.5}, where the long-wavelength behavior at the instability was mapped to a $\phi^4$ model, the question can be posed in more general terms, namely, whether a Landau-Ginzburg model in $1+1$ dimensions belongs to the same universality class of the Ising model in a transverse field \cite{Cardy}. This problem was already numerically approached in Refs. \cite{Barma, MontecarloP4, Sugihara}, but only partial conclusions could be reached.
In this manuscript we address the quantum scenario of the linear-zigzag phase transition by means of Density Matrix Renormalization Group (DMRG)~\cite{White92}, a numerical technique tailored on correlated quantum many-body systems on a one-dimensional lattice~\cite{PrimoMPS}.
With this technique it is possible to address the quantum phase transition problem,
and verify that the linear-zigzag instability belongs to the universality class of the Ising model in a transverse field by determining the relevant critical exponents. Similarly, the simulation can quantify the quantum corrections to the classical linear-zigzag transition.
Here we take a special care in developing in full detail the theoretical framework as well as the numerical architecture employed.
The contribution given by this paper is threefold: first, we
provide for the first time a compact formulation for the mapping from a long-range
into a short-range linear-zigzag model at any order of expansion in the displacement.
Secondly, we give a robust and scientifically sound background to the results some of us previously presented
in Ref.~\cite{NostroANDP}, as well as expand that work by adding previously unreported comments.
Finally, we take a special care in reporting all the numerical expedients we adopt so that all of our
results will be fully reproducible.
The paper is organized as follows. In section \ref{sec:modmod} we introduce the model and we describe a quantitatively robust mapping of the system Hamiltonian into a simpler one, which can be easily tackled numerically. In section \ref{sec:simsim} we review the numerical strategies that we employ to tailor the effective
model into a DMRG architecture, as well as some techniques for quantum state analysis and data-processing. The phase diagram of the quantum linear-zigzag phase transition is sketched in section \ref{sec:phasedia}. In section \ref{sec:uniclass} we determine the relevant critical exponents. The conclusions are drawn in section \ref{sec:conclu}, while the appendix provides details of the mapping presented in section \ref{sec:modmod}.
\section{Quasi-1D Wigner crystal}\label{sec:modmod}
We consider a two-dimensional system of $L$ interacting atoms, trapped by a harmonic potential with frequency $\omega$ along the $y$ direction. The atoms are identical and have mass $M$. Typical distances between atomic ions are of mesoscopic scale \cite{zig1}, so that they are ultimately distinguishable. This allows us to write a first-quantization Hamiltonian despite the many-body character of the dynamics, which reads
\begin{multline} \label{eq:original}
H = \sum_{j = 1}^{L}
\left[ \frac{p_{x,j}^2 + p_{y,j}^2}{2M} + \frac{M \omega^2}{2} y_j^2 + V_{\ell}(x_j)\right]
+ \\ +
\frac{C_\text{int}}{2} \sum_{i \neq j}
\left[ (x_i - x_j)^2 + (y_i - y_j)^2 \right]^{-\alpha/2} \;,
\end{multline}
where the position and canonically conjugated momentum of atom $j$ in the plane are $(x_j,y_j)$ and $(p_{x,j},p_{y,j})$, respectively, while $V_{\ell}(x)$ is a weak confinement along the longitudinal direction, whose shape is conveniently chosen in order to fix the typical interparticle distance. The atom-atom interaction is proportional to the coupling constant $C_\text{int}$, which determines the strength of the interaction, and scales with the interparticle distance $r$ like $\sim 1/r^{\alpha}$, where the exponent $\alpha$ characterizes the nature of the atomic interaction. To provide some examples, for a system of ions we have $\alpha = 1$ and $C_\text{int} = Q^2 / 4 \pi \varepsilon_0$ (with $Q$ charge and $\varepsilon_0$ the vacuum permittivity), while for transversally pinned dipoles~\cite{Kollath,Morigi:Dipoli, Altman} it is $\alpha = 3$, or even $\alpha = 6$ for Rydberg atoms in the induced dipole-dipole interaction regime~\cite{Pohl}. We remark that in general, one should check the conditions under
which the effects of quantum degeneracy on the phase of the gas can be discarded. For atomic ions in typical experimental regimes the particles can be safely considered as distinguishable \cite{Javanainen}.
In the following, starting from the Hamiltonian of Eq. \eqref{eq:original}, we review and critically discuss the basic steps of the mapping onto a lattice $\phi^4$ model. The latter is the basis of the numerical DMRG program, which is described in Sec. \ref{sec:simsim}.
\subsection{Longitudinal-transverse decoupling\\and natural units}
Although the longitudinal and the transverse motion of the atoms are coupled by the dynamics governed by the Hamiltonian in Eq. \eqref{eq:original}, it was argued in Ref.~\cite{Morigi:G.5} that the longitudinal dynamics plays a minor role in the structural transition and can be treated perturbatively. At lowest order in a gradient expansion, it was shown there that the zigzag mode is the soft mode of the transition, which is subject to a Landau-type potential (that possesses a Glodstone mode in three dimensions \cite{Morigi:G.5}). The mapping to a $\phi^4$ model has been explicitly derived in Refs. \cite{DeChiara,Shimshoni:2011a,Shimshoni:2011b}, by assuming that the coupling between longitudinal and transverse vibrations can be neglected sufficiently close to criticality.
Following the arguments reported in Ref. \cite{Shimshoni:2011a}, we employ a model where only the transverse motion of the ions is included, namely
\begin{multline} \label{eq:transverse}
H = \sum_{j = 1}^{L}
\left[ \frac{p_{j}^2}{2M} + \frac{M \omega^2}{2} y_j^2 \right]
+ \\ +
\frac{C_\text{int}}{2} \sum_{i \neq j}
\left[ a^2 (i-j)^2 + (y_i - y_j)^2 \right]^{-\alpha/2} \;,
\end{multline}
where $p_j\equiv p_{y,j}$ and $a$ is the lattice constant. To recast the problem in dimensionless units, we adopt the lattice constant $a$ as the natural length scale and $\mathcal{E}_0 = {C_\text{int}} / a^{\alpha}$ as the energy scale. Then we rescale quantities as follows: $\tilde{y}_j = y_j / a$, $\tilde{H} = H / \mathcal{E}_0$, $\tilde{p}_j = p_{y,j} / \sqrt{ M \mathcal{E}_0}$, and finally $\tilde{\omega} = \omega / \sqrt{ \mathcal{E}_0 / M a^2}$. The Hamiltonian \eqref{eq:transverse} is thus rewritten as
\begin{equation} \label{eq:natural}
H \!= \!\frac{1}{2}\! \sum_{j = 1}^{L}\left(
\tilde{p}_j^2 + \tilde{\omega}^2 \tilde{y}_j^2
+
\sum_{i \neq j}
\frac{1}{
\left[ (i-j)^2 + (\tilde{y}_i - \tilde{y}_j)^2 \right]^{\alpha / 2}
}
\right)\,,
\end{equation}
where the rescaled transverse trap frequency $\tilde{\omega}$ appears explicitly in the expression and is one of the two residual effective parameters (aside from $\alpha$). The other parameter appears in the commutation relation between the displacement $\tilde{y}$ and the transverse momentum $\tilde{p}$:
\begin{equation}
[\tilde{y}_i, \tilde{p}_j] = \imath g \delta_{i,j}\,,
\end{equation}
and reads
\begin{equation} \label{eq:gidef}
g = \sqrt{\frac{\hbar^2}{M a^2 \mathcal{E}_0}} = \hbar \,\sqrt{\frac{a^{\alpha - 2}}{M\, C_\text{int}}}.
\end{equation}
The parameter $g$ is dimensionless and corresponds to the square root of the ratio between the kinetic and the interaction energies. It
plays an analogous role as the Planck constant in the rescaled commutator expression and thus provides a rough estimate of the impact of quantum fluctuations on transverse ordering. For this reason we refer to it as \emph{effective Planck constant}.
Typical values of $g$ depend on the experimental architecture and on the intrinsic nature of the atomic interaction. In particular, $g \propto a^{\alpha/2-1}$. For ions, $g$ increases with the density and takes values in the range between $10^{-5}$ and $10^{-4}$ (for $a \sim 1-10 \mu$m). On the other hand, for $\alpha>2$, which is the case for dipolar gases and Rydberg-dressed gases, $g$ increases as the density decreases, taking values $g > 10^{-2}$, whereas the lattice constant $a$ now spans from fractions of $\mu$m up to a few $\mu$m \cite{Pohl, Morigi:Dipoli, Altman}. Here, the quantum behavior becomes relevant at high densities.
\subsection{Low-energy model}
Close to the transition point the critical behavior is determined by transverse fluctuations whose size is much smaller than the interparticle distance, namely, the lattice constant. This limit corresponds to the inequality $\sqrt{\langle\tilde{y}_j^2}\rangle\ll 1$, where the expectation value is taken over the ground state of the crystal. This condition allows one to substantially simplify the interaction term appearing in Eq. \eqref{eq:natural} by expanding the potential in powers of $(\tilde{y}_j - \tilde{y}_i)$. Taylor-expanding the interaction gives \cite{Morigi:G.5}
\begin{equation} \label{eq:anyord}
\begin{aligned}
\frac{1}{\left[ x_0^2 + \delta \tilde{y}^2 \right]^{\alpha / 2} } &=
\sum_{q = 0}^{t-1} \frac{\prod_{r=0}^{q-1} (\alpha + 2 r)}{\left(-2 \right)^q \; q!}
\frac{\delta \tilde{y}^{2q}}{|x_0|^{\alpha + 2q}}
+ O[\delta \tilde{y}^{2t}] \\
&= \sum_{q = 0}^{t-1} \frac{(-1)^{q} \, \Gamma(q + \textstyle{\frac{\alpha}{2}} )}{q! \;\; \Gamma( \textstyle{\frac{\alpha}{2}} )}
\; \frac{\delta \tilde{y}^{2q}}{|x_0|^{\alpha + 2q}}
+ O[\delta \tilde{y}^{2t}] \\
\end{aligned}
\end{equation}
for arbitrary longitudinal separation $x_0 \geq 1$, where $\Gamma(\cdot)$ is the Euler Gamma function. Substituting this expression into equation \eqref{eq:natural} we obtain
\begin{multline} \label{eq:truexpand}
\tilde{H} = E_0+\sum_{j = 1}^{L}
\left[ \frac{\tilde{p}_j^2}{2} + \frac{\tilde{\omega}^2}{2} \tilde{y}_j^2 \right] \\ +
\frac{1}{2} \sum_{q = 1}^{t-1} (-1)^q \; b_q(\alpha) \sum_{i \neq j}
\frac{ (\tilde{y}_i - \tilde{y}_j)^{2q} }{ |i - j|^{\alpha+2q} }
+ O[\delta \tilde{y}^{2t}]
\;.
\end{multline}
where $E_0$ is a constant and corresponds to the classical ground-state energy of the linear chain \cite{Dubin1997,Morigi:G.7}, while $b_q(\alpha) = \Gamma(q + {\frac{\alpha}{2}} ) / q! \;\Gamma( {\frac{\alpha}{2}} )$ is a positive coupling constant. The previous manipulation makes the problem easier to address by numerical means, since now the parameter $\alpha$ only enters in the coupling coefficients \cite{Shimshoni:2011a}.
We remark that it is important to truncate the expansion at an odd $t$ order, as we did in Eq. \eqref{eq:truexpand}. In fact, this guarantees that the truncated interaction potential in \eqref{eq:anyord} and \eqref{eq:truexpand} is ultimately bounded from below, which is mandatory for avoiding convergence/stability issues of any numerical method we wish to employ.
\subsection{Recasting into a short range theory}
Dealing with a long-range model as in Hamiltonian \eqref{eq:truexpand} is numerically demanding and cumbersome, especially with DMRG, where it leads to slower computational scaling with the system size. For this reason we will adopt an approximation that further simplifies the model.
Based on the arguments introduced in Ref.~\cite{Morigi:G.5, DeChiara:2008}, in the Appendix we show the detailed derivation of a short-range model capable of mimicking quantitatively the linear-zigzag quantum phase transition of the Wigner crystal. We stress that this technique is not based on truncating the interactions at a finite distance: in fact, due to the collective nature of the phononic mode driving the instability (soft mode), truncation would lead to a systematic error in determining the phase diagram. On the contrary, the mapping we adopt reproduces the multi-phonon dispersion bands around the soft mode as faithfully as possible with a nearest-neighbor interacting theory. The starting point is the assumption that at sufficiently low energies, the soft mode, which has quasimomentum $k_0 =\pi$ (in natural lattice units where $a = 1$),
interacts primarily with its neighboring modes in the Brillouin zone $[-\pi,\pi]$. Therefore, for any expansion term $q$ in
Eq.~\eqref{eq:truexpand} we construct the corresponding $q \to q$ phonon scattering function $\Xi[k]$, and approximate it with a short-range interaction matching up to the second order in $\delta k = (k - k_0)$ around $k_0 = \pi$. The algebraic technique employed to achieve this is detailed in appendix \ref{sec:append}. Here we simply report the resulting low-energy Hamiltonian \cite{Shimshoni:2011b}:
\begin{multline} \label{eq:serious}
\tilde{H} \simeq \frac{1}{2} \sum_{j = 1}^{L}
\left[ \tilde{p}_j^2 + \tilde{\omega}^2 \tilde{y}_j^2 + \sum_{q = 1}^{t-1} (-1)^q \times \right.
\\
\left. \times \left( \mathcal{M}_q(\alpha) \,\tilde{y}_j^{2q} - \mathcal{N}_q(\alpha) \left( \tilde{y}_j^{q} - (- \tilde{y}_{j+1} )^{q} \right)^2
\right) \right]
\;.
\end{multline}
The on-site fields $\mathcal{M}$ and coupling constants $\mathcal{N}$ are now functions solely of $q$ and $\alpha$, and not of other physical parameters. They read respectively as follows:
\begin{equation} \label{eq:onsitefield}
\mathcal{M}_{q} =
\frac{\left(2^{\alpha+2q} - 1 \right) \Gamma(q + \textstyle{\frac{\alpha}{2}})}{
q! \;2^{\alpha - 1} \; \Gamma(\textstyle{\frac{\alpha}{2}})}\, \zeta(\alpha+2q)\,,
\end{equation}
\begin{equation} \label{eq:offsitefield}
\begin{aligned}
\mathcal{N}_{1} &=
\left\{ \begin{aligned}
&\ln 2 \qquad &\mbox{for } \alpha = 1 &\\
&\frac{2^\alpha - 2}{2^{\alpha}}\;\alpha \zeta(\alpha) \qquad &\mbox{for } \alpha > 1\,,&
\end{aligned} \right. \\[7pt]
\mathcal{N}_{q > 1} &=\!
\frac{2q-1}{q}\;
\frac{\left(2^{\alpha+2q-2} - 1 \right) \Gamma(q + \textstyle{\frac{\alpha}{2}})}{
q! \; 2^{\alpha - 1} \; \Gamma(\textstyle{\frac{\alpha}{2}})}
\zeta(\alpha+2Q-2)\,.
\end{aligned}
\end{equation}
Let us remark that these couplings capture the collective character of the excitation modes, whose signature is the presence of the Riemann zeta function $\zeta(\cdot)$ in their expression.
We stress that the resulting short-range model \eqref{eq:serious} successfully mimics the dynamics of the original model \eqref{eq:truexpand} only when the excitation energies we are dealing with are small compared with the energy scale associated with the phononic bandwidth. When this is the case, then the modes with quasimomentum $k$ far from $k_0$ (such that $1\lesssim |\delta k| \leq \pi$) play a negligible role in the description of the critical behavior. Since we are interested in the ground state properties, this requirement is satisfied by definition, and thus we can safely accept the approximation.
Moreover, we point out that coefficients $\mathcal{M}_q$ and $\mathcal{N}_q$ grow for large $q$ roughly as $\sim 4^q \; q^{1- \alpha/2}$: this means that the series in Eq.~\eqref{eq:serious} converges only close enough to the linear phase, i.e., $| \langle \tilde{y}^{q} \rangle | < 2^{-q}$. Again, since we are investigating the critical behavior, this requirement is easily achieved.
Let us finally remark that the present
formulation given by Eqs.~\eqref{eq:serious}, \eqref{eq:onsitefield} and \eqref{eq:offsitefield}
of the long-range into short-range mapping was first shown in Refs.~\cite{Morigi:G.5, DeChiara:2008}.
\subsection{Fourth-order expansion}
In this work, we keep up to four total expansion orders in the series described in Eq. \eqref{eq:serious}: we include in the picture the $\tilde{y}_j^4$ local term, as well as the $(\tilde{y}_j + \tilde{y}_{j+1})^2$ interaction, which is second order in the displacement and second order in $\delta k$. We have checked that neglecting further expansion orders generates errors that are compatible with or smaller than errors due to other aspects of the numerical technique we employ afterwards.
The Hamiltonian we simulate, as a function of all the residual parameters, reads
\begin{multline} \label{eq:actualmodel}
\tilde{H}(\alpha, g, \tilde{\omega}) = \frac{1}{2} \sum_{j = 1}^{L}
\left[ - g^2 \frac{\partial^2}{\partial \tilde{y}_j^2} + \left( \tilde{\omega}^2 - \mathcal{M}_1(\alpha) \right) \tilde{y}_j^2 +
\vphantom{\sum} \right.
\\
\left. +\, \mathcal{N}_1(\alpha) \left( \tilde{y}_j + \tilde{y}_{j+1} \right)^2 + \mathcal{M}_2(\alpha)\; \tilde{y}_j^4 \right],
\end{multline}
where we already substituted $\tilde{p}_j = - i g(\partial/\partial \tilde{y}_j)$
to make the parameter $g$ appear explicitly in the expression.
Equation~\eqref{eq:actualmodel} reveals an important physical aspect of the linear-zigzag transition: As long as the coupling with the axial vibrations can be neglected (or just give rise to a renormalization of the parameters of the transverse Hamiltonian), the critical behavior at the phase transition does not depend on the interaction-range scaling exponent $\alpha$. In fact,
given two values $\alpha$ and $\alpha'$
it is possible to map Hamiltonian
$\tilde{H}_0 = \tilde{H}(\alpha, g, \tilde{\omega})$ into
$\tilde{H}' = \tilde{H}(\alpha', g', \tilde{\omega}')$,
by a simple rescaling of energies and length scales \cite{Shimshoni:2011b}.
Precisely, by requiring $\tilde{y}_j' = u \tilde{y}_j$ and $\tilde{H}'= v \tilde{H}_0$ we obtain
\begin{equation}
u = \sqrt{\frac{\mathcal{N}_1 (\alpha') \; \mathcal{M}_2 (\alpha)}{\mathcal{N}_1 (\alpha) \; \mathcal{M}_2 (\alpha')}}
\qquad
v = \frac{\mathcal{N}_1^2 (\alpha') \; \mathcal{M}_2 (\alpha)}{\mathcal{N}_1^2 (\alpha) \; \mathcal{M}_2 (\alpha')}.
\end{equation}
while the other parameters $g$ and $\tilde{\omega}$ must transform as
\begin{align} \label{eq:paramconnect}
\tilde{\omega}'^2
&= \mathcal{M}_1(\alpha') + \frac{\mathcal{N}_1(\alpha')}{\mathcal{N}_1(\alpha)}\left( \tilde{\omega}^2 - \mathcal{M}_1(\alpha) \right)
\\ \label{eq:paradue}
g' &= g \cdot
\frac{\mathcal{M}_2 (\alpha)}{\mathcal{M}_2 (\alpha')}
\left( \frac{\mathcal{N}_1 (\alpha')}{\mathcal{N}_1 (\alpha)} \right)^{\frac{3}{2}}.
\end{align}
Rephrasing, all the linear-zigzag physics formulated according to Eq. \eqref{eq:actualmodel}, for various values of $\alpha$,
are equivalent: the critical behaviour shows the same properties and the phase diagrams in the external parameters space ($g$ and $\tilde{\omega}$) transform into one another according to relations \eqref{eq:paramconnect} and \eqref{eq:paradue}.
We remark that this argument is valid as long as one can safely decouple the axial from the transverse motion, which appears correct for ion Coulomb chains.
We will from now on drop the functional dependence of coefficients $\mathcal{M}$ and $\mathcal{N}$ on $\alpha$. Whenever a specific value of $\alpha$ is implicitly assumed, we will be referring to the ion Wigner crystal scenario i.e.~ $\alpha = 1$. In this setup the coefficients read $\mathcal{M}_1 = 7 \,\zeta(3) / 2 \simeq 4.2072$, $\mathcal{N}_1 = \ln 2 \simeq 0.6931$ and $\mathcal{M}_2 = 93 \,\zeta(5) / 8 \simeq 12.0543$.
\subsection{Connection with the $\phi^4$ model}
Remarkably, the effective model in Eq. \eqref{eq:actualmodel} is closely related to a $\phi^4$ field
theory~\cite{Morigi:G.5,DeChiara,Shimshoni:2011b}: it is basically
an antiferromagnetic formulation of the $\phi^4$ theory on a lattice.
The typical formalism in field theory, where a real
scalar massive field $\phi(x,t)$, undergoing a Klein-Gordon motion, is dressed by a pointwise interaction
of the form $\phi^4$, reads
\begin{equation} \label{eq:phifour}
\mathcal{L}_{\phi^4} = \int \left[ \frac{1}{2} \; \partial^{\mu} \phi \, \partial_{\mu} \phi \; -
\frac{m^2}{2} \, \phi^2 - \frac{\lambda}{4!} \phi^4 \right] dx,
\end{equation}
with flat space-time metric:
$\partial^{\mu} \phi \, \partial_{\mu} \phi = ( \partial_t \phi )^2 - ( \partial_x \phi )^2$,
in units of $\hbar = c = 1$.
We now briefly summarize the steps showing that a Lagrangian of the type \eqref{eq:phifour} can be obtained from
Eq. \eqref{eq:actualmodel}
by means of three simple steps: field staggerization, continuum limit, and canonical rescaling.
Precisely, let $\phi_j = (-1)^j \tilde{y}_j$ be the scalar quantum field. Now, going from a lattice to continuous space
yields
\begin{multline} \label{eq:contimodel}
\tilde{H} = \frac{1}{2} \int
\left[ g^2 \pi^2(x) + \left( \tilde{\omega}^2 - \mathcal{M}_1 \right) \phi^2(x) + \right.
\\
\left. +\, \mathcal{N}_1 \left( \partial_x \phi \right)^2 + \mathcal{M}_2\; \phi^4(x) \right] dx,
\end{multline}
where we performed the substitution $(\phi_{j+1} - \phi_j) \to \partial_x \phi$,
and $\pi(x)$ is the canonically-conjugated field, fulfilling the commutation relation at equal times $[\phi(x),\pi(x')]=i\delta(x-x')$.
In order to obtain an equation of type \eqref{eq:phifour}, we need to rescale
energies and fields ($\tilde{H} \to \tilde{H} / g \sqrt{\mathcal{N}_1}$, $\phi \to \phi\, (g^2 / N_1)^{1/4}$), followed by a standard Legendre transformation. This leads to the Lagrangian
\begin{equation} \label{eq:lege}
\mathcal{L} \!=\! \frac{1}{2}\! \int\! \left[ ( \partial_t \phi )^2 - ( \partial_x \phi )^2 -
\frac{\tilde{\omega}^2 \!-\! \mathcal{M}_1}{\mathcal{N}_1} \, \phi^2 -
\frac{g \mathcal{M}_2}{\mathcal{N}_1^{3/2}} \phi^4 \right] dx,
\end{equation}
which connects to \eqref{eq:phifour} via the relations
\begin{equation}
m = \sqrt{\frac{\omega^2 - \mathcal{M}_1}{\mathcal{N}_1}}
\quad \mbox{and} \quad
\lambda = 12 \,\frac{g \,\mathcal{M}_2}{\mathcal{N}_1^{3/2}}.
\end{equation}
The one-dimensional $\phi^4$ field theory on a lattice has been already addressed by means of numerical simulation, both by Montecarlo methods \cite{Barma, MontecarloP4} and also by DMRG \cite{Sugihara}. In this paper we provide a complete, exhaustive characterization of its quantum criticality,
while exploring the whole phase-diagram boundary, therefore extending and complementing the results we presented in Ref. \cite{NostroANDP}.
\section{DMRG Simulation Details}\label{sec:simsim}
The Density Matrix Renormalization Group (DMRG) is a method developed in the early 90's \cite{White92} which has proven successful for a large variety of one-dimensional many-body quantum problems \cite{PrimoMPS}. The key to its success relies on the fact that the entanglement description capabilities of DMRG match exactly the typical entanglement scaling laws of ground states in one-dimensional quantum systems (and low-lying energy states too) \cite{Arealaws}.
Simulating the effective lattice staggered $\phi^4$ model of Eq.~\eqref{eq:actualmodel} with DMRG is feasible,
but it still requires some additional careful numerical treatment\cite{Localtrunc}.
We will discuss these expedients in the present section.
\begin{figure}
\begin{flushright}
\begin{overpic}[width = 230pt, unit=1pt]{Figure3.jpg}
\put(117, -6){$\tilde{y}$}
\put(-12, 128){$|\psi_q|^2$}
\end{overpic}
\end{flushright}
\caption{ \label{fig:lobas}
Probability densities $|\psi_q(\tilde{y})|^2$ as a function of $\tilde{y}$ for the lowest five quantum energy eigenstates
of the local Hamiltonian of Eq.~\eqref{eq:localham}: $q = 1$ (red line), $q=2$ (green line), $q = 3$ (blue line),
$q = 4$ (yellow line) and $q = 5$ (grey line).
For easy reference we also plot the double well potential profile (black thick line) of the Hamiltonian,
in arbitrary energyscale units.
}
\end{figure}
\subsection{Local basis selection}
Traditional DMRG architectures\cite{White92}
are tailored to models where the local space is a finite-dimensional Hilbert space, like for instance for spin models. In the scenario here considered, however,
the local space is a continuous quantum variable, with the usual Lie algebra $\{\tilde{y}, \tilde{p}\}$ of the quantum particle motion.
To circumvent this obstacle, we adopt a local space truncation approach that was thoroughly discussed in Ref.~\cite{Localtrunc}.
Namely,
we define a related single-particle quantum problem $H_{\text{loc}}$,
then find its $d$ lowest-energy eigenfunctions $|\psi_q\rangle$ and adopt them as local basis
$\{|\psi_q\rangle \}_{q = 1..d}$ for the many-body problem.
If the many-body Hamiltonian with nearest-neighbor interaction is
$\tilde{H} = \sum_ j H_{\text{loc}}^{(j)} + H_{\text{int}}^{(j,j+1)}$, then
the simulation is more accurate for a given $d$ (or it requires smaller $d$ to achieve some target precision),
when the interaction energy $\langle H_{\text{int}} \rangle$ is smaller in modulus; that is, when $H_{\text{int}}$
can be treated as a perturbation.
For the case under study we argue that considering the whole $(\tilde{y}_j + \tilde{y}_{j+1})^2$ term as
the interaction part is a more perturbative approach than just taking the double product $2 \tilde{y}_j \tilde{y}_{j+1}$.
Indeed, while in the linear phase the expectation value on the ground state of
two terms is of roughly the same magnitude, in the zigzag phase the first
is definitely closer to zero, and thus more appropriate to be chosen as interaction part.
According to this scheme, the local Hamiltonian reads
\begin{equation} \label{eq:localham}
H_{\text{loc}} = \frac{1}{2}
\left[ - g^2 \frac{\partial^2}{\partial \tilde{y}^2} + \left( \tilde{\omega}^2 - \mathcal{M}_1\right) \tilde{y}^2 +
\mathcal{M}_2\; \tilde{y}^4 \right],
\end{equation}
describing the motion of a quantum particle in a harmonic-quartic potential. Thanks to the translational invariance
of Eq.~\eqref{eq:actualmodel}, we just have to solve the problem
\eqref{eq:localham} once per given $g$ and $\tilde{\omega}$, and use the resulting basis for every site.
In order to find the low-energy eigenstates of Eq.~\eqref{eq:localham},
we solve $H_{\text{loc}} |\psi_q\rangle = E_q |\psi_q\rangle$ exactly by means of linear algebra numerical methods
to diagonalize tridiagonal matrices.
Afterwards, we express the single-body computational basis as $|q_j\rangle \equiv |\psi_{q}\rangle$, which corresponds to the
atom at site $j$ being in the orbital state $|\psi_q\rangle$.
The resulting many-body computational basis is made of tensor product states of the single-site basis states
$|q_1\,q_2\ldots q_L\rangle = |q_1\rangle \otimes |q_2\rangle \otimes \ldots \otimes |q_L \rangle$.
In this formalism, the global hamiltonian then reads
\begin{equation} \label{eq:dmrgready}
\tilde{H} = \sum_j ( A_j + \mathcal{N}_1 \,W_j ) + \mathcal{N}_1 \,Y_j \otimes Y_{j+1}
\end{equation}
where the matrices $\Theta_j$ ($\Theta = A,W,Y$) are single-site operators acting on site $j$, although their
matrix elements do not explicitely depend on $j$ thanks to translational invariance. Specifically, we have
$A_j = \sum_{q = 1}^d E_q |q_j\rangle \langle q_j|$, then
$Y_j = \sum_{q,q'}^{d} |q_j\rangle \langle \psi_q | \,\tilde{y}\, |\psi_{q'}\rangle \langle q'_j|$ and
$W_j = \sum_{q,q'}^{d} |q_j\rangle \langle \psi_q | \,\tilde{y}^2\, |\psi_{q'}\rangle \langle q'_j|$.
We explicitly express the change of basis $|\psi_q \rangle\langle q_j|$ to stress the fact that we are projecting the
space over the first $d$ local states.
Notice that, as a consequence of this, $W \neq Y^2$.
\begin{figure}
\begin{center}
\begin{overpic}[width = \columnwidth, unit=1pt]{Figure4.jpg}
\put(106, -8){Level index $q$}
\put(123, 86){$p(q)$}
\end{overpic}
\end{center}
\caption{ \label{fig:histo}
Probability distribution $p(q)$, plotted as a function of the local basis number $q$.
$p(q)$ is defined as $p(q) = \langle q | \varrho_j | q \rangle$ where
$\varrho_j = \mbox{Tr}_{j'\neq j}\{|\Psi^{L}_{g,\tilde{\omega}}\rangle\langle\Psi^{L}_{g,\tilde{\omega}}|\}$
is the reduced density matrix of an arbitrary site $j$ in the bulk
of the many-body ground state $|\Psi^{L}_{g,\tilde{\omega}}\rangle$,
simulated by DMRG. Here the system is in the zigzag phase, but close to critical point ($g = 0.03$, $\tilde{\omega}^2 = 1.0$):
The first two levels (symmetric and antisymmetric double peaks respectively)
have equal populations. The third level probability is
two orders of magnitude smaller, and rapidly decaying.
}
\end{figure}
Typical solutions of the local problem defined by Eq.~\eqref{eq:localham}, in the deep quantum regime ($g = 0.1$),
are displayed in Fig.~\ref{fig:lobas}.
Both the second and fourth excited states (blue and grey line, respectively) show a non-negligible probability density at the barrier point ($\tilde{y} = 0$). This reveals that the quantum fluctuations enhance substantial tunneling between the two potential wells, thus ultimately making the linear (disordered) phase energetically favourable.
In order to keep track of the error generated by truncating the basis $|\psi_q\rangle$ to a dimension $d$, we performed several calculations of the same problem (under identical environment parameters) for various values $d \sim 2 \ldots 100$, until we located convergence of the outcomes.
Furthermore, we kept track of the populations of various basis levels on every site: namely
we verify that the occupation probabilities in the one-site reduced density matrix
decrease roughly exponentially with the level index $q$.
This provides a meaningful lower bound for the error generated by the truncation. Figure~\ref{fig:histo} displays the populations $p(q)$ of the first eight local basis levels, obtained after the numerical simulation of the ground state for a given set of parameters (further levels are of order of magnitude below $10^{-5}$ and not shown in the plot). In all the cases we considered, the populations $p(q)$
decay very rapidly with the level index $q$. They can usually be bounded from above by an exponential
decay $p(q \geq 3) \leq e^{-\Lambda q}$. For the case in the figure, $\Lambda \simeq 1.7$.
Hamiltonian \eqref{eq:dmrgready} is ready for simulation, and via standard DMRG architecture \cite{White92} we searched for its ground state for finite system size $L$ with Open Boundary Conditions (OBC). The latter choice is due to a natural tendency of DMRG with respect to OBC: in this scenario it converges faster, and it has enhanced precision and stability.
In the various physical systems considered, a local basis dimension $d \sim 30$ and a
DMRG bondlink $D \sim 50$ were sufficient to make the results converge permanently under our precision
(typically $10^{-10}$ of relative numerical precision).
\subsection{Measured order parameter}
A final remark regards the identification of the phase across the transition. A drawback of working on finite-size quantum samples is the impossibility of achieving a truly spontaneous symmetry-broken phase. In our problem, the two possible zigzag configurations bear a nonzero interference term, which encourages the even superposition of the two as true ground state, thus restoring the $\mathbb{Z}_2$ parity symmetry. A standard technique known in literature\cite{Mussardo}
to circumvent this issue is employing structure factor-based order parameters, which are insensitive to symmetry breaking. We briefly review this strategy for easy reference to the reader. The order parameter we choose is the
\emph{square root of the structure factor density} calculated at the soft mode $k_0 = \pi$, precisely
\begin{equation} \label{eq:strufa}
\xi_L(g, \tilde{\omega}) = \sqrt{\frac{1}{L^2} \sum_{j,j'}^L (-1)^{j-j'} \langle \Psi_{g, \tilde{\omega}}^L |
Y_j \otimes Y_{j'} | \Psi^L_{g, \tilde{\omega}} \rangle
}\;,
\end{equation}
where $ | \Psi_{g, \tilde{\omega}}^L \rangle$ is the many-body ground state calculated via DMRG under parameters
$g$, $\tilde{\omega}$ and size $L$.
This is clearly a non-extensive quantity, and it can be shown to exactly coincide with the standard antiferromagnetic
order parameter $\bar{m} = L^{-1} \sum_j^L (-1)^j \langle Y_j \rangle$ in the thermodynamical limit.
Basically if we assume that correlations can be split into a classical and a quantum contribution, i.e.~
$\langle Y_j Y_{j'} \rangle \simeq \langle Y_j \rangle \langle Y_{j'} \rangle + f_q(j-j')$ respectively,
then the quantum part becomes irrelevant when evaluating $\xi_L$.
In fact, it is either $f_q(j-j') \sim |j-j'|^{-\nu}$ (critical scenario, with $\nu > 0$), or
$f_q(j-j') \sim e^{-|j-j'|/\lambda}$ (noncritical scenario). In both cases it holds
\begin{equation}
\left| \sum_{j,j'}^L \frac{(-1)^{(j-j')}}{L^2} f_q(j-j') \right|
\leq \frac{ f_q(1) + \int_1^{L} \left| f_q(x) \right| dx}{L} \to 0\,.
\end{equation}
Consequently, we obtain
\begin{equation} \label{eq:stru2}
\xi_L = \sqrt{\sum_{j,j'}^L (-1)^{(j-j')}
\frac{ \langle Y_j \rangle \langle Y_{j'} \rangle}{L^2} } \simeq \sqrt{\bar{m}^2} = \bar{m}\,,
\end{equation}
which tells us that $\xi_L$ has all the properties of a local antiferromagnetization without suffering from finite-size symmetry breaking issues, since it is based on two-point correlation measurements and not on local observations.
\section{Phase Diagram}\label{sec:phasedia}
\begin{figure}
\begin{center}
\begin{overpic}[width = \columnwidth, unit=1pt]{Figure5.jpg}
\put(0, 78){$\xi_L$}
\put(121, -10){$1/L$}
\end{overpic}
\end{center}
\caption{ \label{fig:TDPhase} (color online)
Zigzag-order parameter $\xi_L(g, \tilde{\omega})$ as a function of $1/L$.
Here the lattice size $L$ ranges within $100 \leq L \leq 1600$, while
other parameters are $g = 0.08$ and $d = 14$.
The points have been numerically evaluated, the lines
are eye-guides.
Each set of data corresponds to a different value of the square trap frequency
$\tilde{\omega}^2 = 1.30, 1.32, 1.34, \ldots ,1.58, 1.60$ (from top to bottom).
The zigzag order parameter is obtained via DMRG using the square root of structure-factor density
of Eq. \eqref{eq:strufa}.
}
\end{figure}
Some of the results presented in this and the following sections were previously reported in Ref.~\cite{NostroANDP},
in particular Figs.~\ref{fig:phasedia}, \ref{fig:finsiz}, \ref{fig:central} that we report for completeness and comfort to the reader.
Here we describe the complete numerical derivation which allows to process these physical quantities from raw simulation data,
and we give additional comments which clarify technical issues about those results.
First of all,
we characterize the phase of the ground state for a given point in the parameters space ($g, \tilde{\omega}^2$). More precisely, we simulate the same OBC problem for various system sizes $L$, typically up to 3000 sites, with the prescriptions detailed in the previous section. For each simulation, we record the zigzag order parameter $\xi_L(g, \tilde{\omega})$ introduced in Eq. \eqref{eq:strufa} by measuring every two-point correlator. Finally, we fit the thermodynamic limit $\xi_{\infty} = \lim_{L \to \infty} \xi_L$ and we discriminate whether its value is zero, which detects the linear phase, or it is nonzero, thus revealing the zigzag phase. Figure~\ref{fig:TDPhase} displays the order parameter $\xi_L$ as a function of the length $L$ of the sample.
Various data sets are plotted, each one corresponding to a different value of $\tilde{\omega}^2$ (at the same
value of $g$).
Every curve $\xi(L)$ is fitted via various fit functions, polynomial in the inverse length,
and one clearly
sees that the order parameter is typically a very smooth function of the chain size.
\begin{figure}
\begin{center}
\begin{overpic}[width = \columnwidth, unit=1pt]{Figure6.pdf}
\put(-3, 92){\large $\tilde{\omega}_c^2(g)$}
\put(194, 0){\large $g$}
\put(34, 122){$|\Delta \tilde{\omega}_c^2 |$}
\put(170, 152){Linear phase}
\put(85, 127){Zigzag phase}
\put(70, 100){Zigzag}
\put(120, 60){Linear}
\end{overpic}
\end{center}
\caption{ \label{fig:phasedia} (color online)
Phase diagram of an array of ions (i.e.~ $\alpha = 1$) in the $(g, \tilde{\omega}^2)$ parameter space,
at the thermodynamical limit.
Inset: displacement of the critical square frequency $|\Delta \tilde{\omega}^2_c| $ from the classical value,
as a function of $g$. Cyan lines represent a power-law fit of the whole data curve
using $|\Delta \tilde{\omega}^2_c| (g)= u \,g^{v}$ via $u$ and $v$
(see text).}
\end{figure}
Since we are interested in determining the phase diagram, we need to discriminate when $\xi_{\infty}$ is zero and when it is not, regardless of the function used to fit the value. This allows us to detect the phase boundary
in the parameter space ($g, \tilde{\omega}^2$) with an uncertainty
of the order of $10^{-3}$. Figure~\ref{fig:phasedia} displays the phase diagram,
which has been derived after locating the critical trap frequency value $\tilde{\omega}_c(g)$ for several values of $g$
in the range $10^{-5} \leq g \leq 0.2$. As expected, when $g$ increases, the magnitude of quantum fluctuations increases and with it the range of the disordered phase. In accordance to this conjecture, $\tilde{\omega}_c(g)$ is a monotonically decreasing function of $g$.
We determine the shift of the critical point $\tilde{\omega}_c^2(g)$ with respect to the one predicted by the classical theory \cite{Morigi:G.5}, which is given by $\tilde{\omega}_c^2(0) = \mathcal{M} _1$. This corresponds to the quantity $\Delta \tilde{\omega}_c^2(g)\equiv \mathcal{M} _1 - \tilde{\omega}_c^2(g)$, and we obtain that it scales with $g$ according to a power-law behaviour given by
$$\Delta \tilde{\omega}_c^2(g) = (21.91 \pm 0.01) \cdot g^{(0.823 \pm 0.003)}\,.$$
We remark that we also employed other data-processing strategies to draw the critical boundary, which will be detailed in the next section. These are
divergence of the correlation length, entanglement area-law violation,
and finite-size scaling. In all the cases these methods have proven compatible,
albeit less precise, to the structure-factor density technique.
\section{Critical exponents}\label{sec:uniclass}
The largest local symmetry group under which our model $\tilde{H}$ is invariant is parity symmetry. Specifically, let $R$ be the unitary reflection operator along the transverse direction, i.e.~ $R \:\tilde{y}\: R^{\dagger} = -\tilde{y}$. Then $R = R^{\dagger}$, $R^2 = 1$, which is a representation of the $\mathbb{Z}_2$ group, and it is straightforward to check that it is a symmetry of the Hamiltonian, i.e.~$[\tilde{H}, R^{\otimes L}] = 0$. In the thermodynamical limit, the zigzag phase has a two-fold degenerate ground space, which spontaneously breaks the symmetry $R^{\otimes L}$.
Starting from this symmetry argument, conformal field theory predicts that
the continuous model should exhibit the same universality class of the 1D quantum Ising model with transverse field \cite{Sachdev, Mussardo, Cardy}. We are now going to test the validity of this claim in the lattice model by numerically evaluating various critical exponents, and comparing them with the corresponding theoretical predictions.
\begin{figure}
\begin{center}
\begin{overpic}[width = \columnwidth, unit=1pt]{Figure7.pdf}
\put(-4, 134){\large $G(\Delta j)$}
\put(215, 2){\large $\Delta j$}
\put(124, 28){$g$}
\put(50, 110){$\eta$}
\end{overpic}
\end{center}
\caption{ \label{fig:correl} (color online)
Two point correlations $G(\Delta j)$ of the displacement $\tilde{y}$ as a function of the distance
$\Delta j$ (magenta dots), acquired via
numerical ground-state simulation with DMRG.
Data are fitted using
$G_{\rm fit}(\Delta j)$ in Eq. \eqref{G:fit} (black solid line).
The green and orange lines show
fits via an exponential and a power-law respectively.
Inset: fitted anomalous dimension exponents $\eta$ over several points of the critical boundary,
characterized by different values of $g$, the dashed line shows the average value
$\eta = 0.258 \pm 0.012$, the solid line is the value of $1/4$ predicted from the Ising model theory.
}
\end{figure}
\subsection{Anomalous dimension $\eta$}
The transverse coordinate $\tilde{y}$ of the lattice field model plays the same role as the interaction Pauli matrix of the Ising model \cite{Shimshoni:2011a} (namely $\sigma^z$ if one writes $H_{\text{Ising}} = \sum_j \sigma^z_j \sigma^z_{j+1} + B \sigma^x_j$). As a result, studying quantum correlation functions in the $\tilde{y}$ direction, such as
\begin{equation}
Q(j,j') \equiv \langle \tilde{y}_j \tilde{y}_{j'} \rangle - \langle \tilde{y}_j \rangle \langle \tilde{y}_{j'} \rangle\,,
\end{equation}
should reveal the \emph{anomalous dimension} critical exponent $\eta$ \cite{Sachdev}.
To make sure that the scaling of correlations depends only on the distance $\Delta j \equiv |j-j'|$ of the pair and is insensitive to boundary effects, we enlarge our system size up to thousands of sites, and average over the bulk:
the expression
\begin{equation} \label{eq:bulky}
G(\Delta j) = (-1)^{\Delta j}\, \tilde{L}^{-1} \sum_{j}^{\tilde{L}} Q(j,j+\Delta j)
\end{equation}
is chosen in such a way as to take into account staggerization as well.
We calculate $G(\Delta j)$ from the numerical data summing over $j$ sites in \eqref{eq:bulky} such that we disregard sites sitting too
close to the boundary. Namely, we include only pairs $\{j,j+\Delta j\}$ located within the central third of the chain.
Moreover, since we are never exactly simulating the critical point, we must introduce a finite-correlation length $\lambda$ and evaluate it altogether. The function we use to fit $G(\Delta j)$ reads\cite{Mussardo}
\begin{equation}
\label{G:fit}
G_{\rm fit}(\Delta j) = \alpha \,\Delta j^{-\eta} \exp(-\Delta j/\lambda)\,,
\end{equation}
where $\alpha$, $\eta$ and $\lambda$ are fitting parameters. Figure~\ref{fig:correl} displays two-point correlations
functions $G(\Delta j)$ numerically evaluated
at distances up to 300 sites, and the fit according to Eq. \eqref{G:fit}, which provides in all the scenarios considered an impressive match to the numerical data.
After sampling results for several points of the critical boundary, we obtain an average critical exponent of $\eta = 0.258 \pm 0.012$, in good agreement with the predicted $1/4$ value.
\begin{figure}
\begin{center}
\begin{overpic}[width = \columnwidth, unit=1pt]{Figure8.pdf}
\put(20, 114){\large $\xi$}
\put(130, -8){\large $\tilde{\omega}$}
\put(180, 145){$f(x)$}
\end{overpic}
\end{center}
\caption{ \label{fig:finsiz} (color online)
Order parameter $\xi_L(g, \tilde{\omega})$, obtained via DMRG,
as a function of $\tilde{\omega}$, plotted for different system sizes $L = 100, 120 \ldots 300$, at $g=0.12$.
Inset: rescaled data
$\xi_L(\tilde{\omega}_c + \Delta\tilde{\omega}_c L^{\gamma_2}) L^{\gamma_1}$,
with $\gamma_1 = 0.127$ and $\gamma_2 = 1.04$,
characterizing $f(x)$ (see text).
}
\end{figure}
\subsection{Finite-size scaling}
By means of finite-size scaling \cite{FishBarb} it is possible to acquire two critical exponents:
the spontaneous magnetization exponent $\beta$ and the correlation length divergence exponent $\nu$.
In accordance to renormalization group analysis, the order parameter $\xi_L(g, \tilde{\omega})$ in an OBC setup
should obey a precise scaling with the system size $L$ and the parametric distance from the critical point
$(\tilde{\omega} - \tilde{\omega}_c)$. Specifically
\begin{equation} \label{eq:fishbarb}
\xi_L (\tilde{\omega}) \simeq L^{-\beta / \nu} \; f \left( (\tilde{\omega} - \tilde{\omega}_c) \; L^{1 / \nu} \right),
\end{equation}
where $f$ is a non-universal function, depending on the microscopic details of the model.
To exploit this picture, first we tune $\gamma_1 = \beta / \nu$ until we observe a crossing of all the curves
$\xi_L(g, \tilde{\omega}) \; L^{\gamma_1}$
as functions of $\tilde{\omega}$ in a single point, which locates the critical frequency $\tilde{\omega}_c(g)$.
Then, we find the suitable value $\gamma_2 = - 1/\nu$ which makes all the curves
$\xi_L(g, \tilde{\omega}_c + (\tilde{\omega}-\tilde{\omega}_c)L^{\gamma_2}) \, L^{\gamma_1}$ collapse onto one another.
Fig.~\ref{fig:finsiz} shows an example of this procedure applied to our problem:
after plotting the order parameter $\xi_L$ as a function of the trap frequency $\tilde{\omega}$ for a dozen different
systems lengths $100 \leq L \leq 300$, we rescale the coordinates of the data according to this procedure,
and obtain the plot shown in the inset, where all the data collapsed onto a single curve,
which is the $f(x)$ from Eq.~\eqref{eq:fishbarb}.
Critical exponents obtained by employing this strategy
read $\beta = 0.126 \pm 0.011$ and $\nu = 1.03 \pm 0.05$, to be compared with theoretical
predictions of $1/8$ and $1$ respectively.
\subsection{Central charge $c$}
The central charge critical exponent $c$
is related to the scaling of quantum entanglement, under a system bipartition
into two blocks, as a function of the shape and size of the blocks themselves
\cite{Cencharge, VonNeu}.
Indeed, for one-dimensional systems, a logarithmic violation of the area law of entanglement
is a discriminating signature for criticality and a gapless excitation spectrum\cite{Cardy}.
In this framework, the central charge $c$ is the prefactor of the logarithm itself:
at the thermodynamical limit it reads
$\mathcal{S}(\rho_\ell) \simeq \frac{c}{3} \log \ell + c'$, where
$c'$ a non-universal constant\cite{dechiara06}, and
\begin{equation} \label{eq:vonneudef}
\mathcal{S}(\rho_\ell) \equiv - \mbox{Tr} [\rho_\ell \ln \rho_\ell]
\end{equation}
is the Von Neumann entropy of the reduced density matrix
$\rho_\ell$ of $\ell$ adjacent sites.
For a semi-infinite system instead, where the $\ell$ sites contain the single boundary, it is
$\mathcal{S}(\rho_\ell) \simeq \frac{c}{6} \log \ell + c'$. Our scenario is a finite-size OBC system, for which the relation holds \cite{Calacardy}
\begin{equation} \label{eq:cardy}
\mathcal{S}_{\text{th}}(\rho_\ell) = \frac{c}{6} \log \left(L \cdot \sin{\frac{\pi \ell}{L}} \right) + c'\;,
\end{equation}
where $\ell$ is the site at which we are considering a left-right system bipartition, and $\rho_\ell$
is the reduced density matrix of the left (or right) block. Due to the intrinsic nature of
DMRG it is straightforward to evaluate it for every partition $\ell$,
since the Schmidt coefficients are automatically provided from the algorithm.
We fit numerical data using Eq.~\eqref{eq:cardy} via $c$ and $c'$ for various system lengths $L$,
as shown in Fig.~\ref{fig:central}, right inset. At the critical point Eq.~\eqref{eq:cardy} should reproduce
the correct scaling of the entanglement, and thus the fitted $c$ values should be constant as a function of $L$.
This is actually the case, as shown in Fig.~\ref{fig:central}, main plot:
the magenta line corresponds to the critical point and is roughly constant with $L$.
This procedure provides an additional method for locating the critical point,
albeit less precise than the previously discussed ones, and quantifies the critical central charge at the same time.
After averaging over several points of the critical boundary (Fig.~\ref{fig:central}, left inset),
we acquire an estimate of $c = 0.487 \pm 0.015$, in good accordance with the theoretical value of $1/2$.
\begin{figure}
\begin{center}
\begin{overpic}[width = \columnwidth, unit=1pt]{Figure9.pdf}
\put(4, 115){\large $c$}
\put(154, -4){\large $L$}
\put(198, 24){$\ell$}
\put(85, 27){$g$}
\put(60, 48){$c$}
\put(170, 61){\large $\mathcal{S}(\rho_\ell)$}
\end{overpic}
\end{center}
\caption{\label{fig:central}
Right inset: von Neumann entropy $\mathcal{S}(\rho_{\ell})$ (black dots) of the reduced density matrix $\rho_{\ell}$
calculated applying Eq.~\eqref{eq:vonneudef} to the ground state simulation with DMRG,
as a function of the block size $\ell$.
The data shows remarkable agreement with the fit using $\mathcal{S}_{\text{th}}(\rho_{\ell})$ of
Eq.~\eqref{eq:cardy} via $c$ and $c'$ (dashed yellow line). Main: $c$ values as a function of the system size
$L = 100,120 \ldots 300$ for different trap frequencies $\tilde{\omega}^2 = 0.383, 0.384 \ldots 0.390$
(top to bottom) at $g = 0.12$.
The magenta data set detects the critical point
at $\tilde{\omega}_c^2 = 0.385$, and the corresponding estimated $c$ value (dashed black line)
is $c = 0.486$.
Left inset: estimated central charge $c$ values at various points in the critical boundary,
as a function of the parameter $g$. The dashed line corresponds to the average of $c = 0.487$ while
the solid line is the theoretical prediction $c_{\text{th}}=1/2$.
}
\end{figure}
\section{Conclusions}\label{sec:conclu}
In this work we numerically studied the quantum linear-zigzag transition in quasi one-dimensional Wigner crystals by means of extensive numerical simulations based on Density Matrix Renormalization Group.
We first introduced the theoretical framework which lets us describe the phenomenon as a quantum lattice model, as well as the approximations we employed to make it amenable to simulation with DMRG.
According to this picture, we provided an analytical mapping from the original long-range theory to an effective short-range model.
We argued that such a mapping is possible as long as the excitations energies in the many-body state
are small compared to the phononic bandwidth $\mathcal{M}_1$.
Applying this mapping leads to a critical speed-up of the numerical calculation, as we could employ traditional
nearest-neighbor model DMRG techniques to an otherwise difficult problem.
The phase diagram of the phase transition in the reduced external parameter space was determined. We detected the universality class of the criticality by extracting several critical exponents with high precision. These results ultimately show that the linear-zigzag transition is of the same universality class as the quantum Ising model in transverse field.
Our model can be applied to strongly-interacting systems, such as ions in traps, dipolar gases, and Rydberg atoms chains: In fact, as long as the interaction is repulsive and described by a power law, the exponent $\alpha$ determining the strength of the interparticle potential just enters the coefficients of the transverse Hamiltonian, but does not change its form (as long as $\alpha\ge 1$). On the other hand, the study we performed is based on the assumption that transverse and longitudinal vibrations are decoupled. This assumption is correct for ion Coulomb systems, but has to be checked for dipolar and Rydberg systems, where the effect of quantum fluctuations may modify the nature of the transition \cite{Meyer-Rosch}.
\emph{Acknowledgments - } We acknowledge support from EU through PICC, SIQS, and from the German Research Foundation (Heisenberg programme, SFB/TRR21), the BW-grid for computational resources. We warmly thank G. De Chiara, S. Fishman, A. Muramatsu, M.B. Plenio, A. Retzker, A. Rosch, E. Shimshoni, D. Fioretto, and M. Burrello for discussions, and D. Rossini for contributing in the development of the numerical suite.
|
\section{Introduction}
\label{sec:introduction}
A set function $f: 2^V \to \mathbb{R}$ is said to be \emph{submodular}
\cite{fujishige2005submodular} if for all subsets $S, T \subseteq V$,
it holds that $f(S) + f(T) \geq f(S \cup T) + f(S \cap T)$. Defining
$f(j | S) \triangleq f(S \cup j) - f(S)$ as the gain of $j\in V$ in
the context of $S \subseteq V$, then $f$ is submodular if and only
if
$f(j | S) \geq f(j | T)$ for all $S \subseteq T$ and $j \notin T$. The
function $f$ is monotone iff $f(j | S) \geq 0, \forall j \notin S, S
\subseteq V$. For convenience, we assume the ground set is $V = \{1,
2, \cdots, n\}$. While general set function optimization is often
intractable, many forms of submodular function optimization can be
solved near optimally or even optimally in certain cases\arxivalt{,
and hence submodularity is also often called the discrete analog of
convexity~\cite{lovasz1983}.}{.} Submodularity, moreover, is
inherent in a large class of real-world applications, particularly in
machine learning, therefore making them extremely useful in practice.\looseness-1
In this paper, we study a new class of discrete
optimization problems that have the following form:
\begin{align*}
\label{eqn:scsc_scsk_def}
\mbox{Problem 1 (SCSC): } \min \{f(X) \, | \,g(X) \geq c\},\quad
\text{and}
\quad
\mbox{Problem 2 (SCSK): } \max \{ g(X) \,| \,f(X) \leq b\},
\end{align*}
where $f$ and $g$ are monotone non-decreasing submodular functions
that also, w.l.o.g., are normalized ($f(\emptyset) = g(\emptyset) =
0$)\footnote{A monotone non-decreasing normalized ($f(\emptyset) = 0$) submodular function is called a polymatroid function.}, and where $b$ and $c$ refer to budget and cover parameters
respectively. The corresponding constraints are called the submodular cover~\cite{wolsey1982analysis} and submodular knapsack~\cite{atamturk2009submodular}
respectively and hence
we refer to Problem 1 as {\em Submodular Cost Submodular
Cover} (henceforth SCSC) and Problem 2 as {\em Submodular Cost Submodular
Knapsack} (henceforth SCSK). Our motivation stems from an interesting class of problems that
require minimizing a certain submodular function $f$ while
simultaneously maximizing another submodular function $g$. We shall
see that these naturally occur in applications like sensor placement,
data subset selection, and many other machine learning applications. A standard approach used in
literature~\cite{rkiyeruai2012, narasimhanbilmes, kawahara2011prismatic} has been to transform these problems into
minimizing the difference between submodular functions (also called DS
optimization):\looseness-1
\begin{align}
\mbox{Problem 0: } \min_{X \subseteq V} \bigl( f(X) - g(X)\bigr).
\end{align}
While a number of heuristics are available for solving Problem
0\arxiv{, and while these heuristics often work well in
practice~\cite{rkiyeruai2012,narasimhanbilmes}}, in the worst-case
it is NP-hard and inapproximable~\cite{rkiyeruai2012}, even when $f$ and $g$ are monotone. Although an exact
branch and bound algorithm has been provided for this
problem~\cite{kawahara2011prismatic}, its complexity can be
exponential in the worst case. On the other hand, in many
applications, one of the submodular functions naturally serves as part
of a constraint. For example, we might have a budget on a cooperative
cost, in which case Problems 1 and 2 become applicable.
\arxiv{\subsection{Motivation}}
The utility of Problems 1 and 2 become apparent when we consider how
they occur in real-world applications and how they subsume a number of
important optimization problems.\looseness-1
\textbf{Sensor Placement and Feature Selection: }Often,
the problem of choosing sensor locations \arxiv{$A$ from a given set of
possible locations $V$}can be
modeled~\cite{krause2008near, rkiyeruai2012} by
maximizing the mutual information between the chosen variables $A$ and
the unchosen set $V \backslash A$ (i.e.,$f(A) = I(X_A; X_{V
\backslash A})$). \arxiv{Note that, while the symmetric mutual information is not monotone, it can be shown to approximately monotone~\cite{krause2008near}. }Alternatively, we may wish to maximize the mutual
information between a set of chosen sensors $X_A$ and a quantity
of interest $C$ (i.e., $f(A) = I(X_A ; C)$) assuming that
the set of features $X_A$ are conditionally independent given $C$
\cite{krause2008near, rkiyeruai2012}. Both these functions are submodular. Since there are costs involved, we want to simultaneously minimize the cost $g(A)$. Often this cost is submodular~\cite{krause2008near, rkiyeruai2012}. For
example, there is typically a discount when purchasing sensors in bulk
(economies of scale). \arxivalt{Moreover, there may be diminished cost for
placing a sensor in a particular location given placement in certain
other locations (e.g., the additional equipment needed to install a
sensor in, say, a precarious environment could be re-used for multiple
sensor installations in similar environments). Hence this becomes a
form of Problem 2 above. An alternate view of this problem is to find
a set of sensors with minimal cooperative cost, under a constraint
that the sensors cover a certain fraction of the possible locations,
naturally expressed as Problem 1.}{This then becomes a form of either Problem 1 or 2.\looseness-1 }
\textbf{Data subset selection:} A data subset
selection problem in speech and NLP involves finding a limited
vocabulary which simultaneously has a large coverage. This is
particularly useful, for example in speech recognition and machine
translation, where the complexity of the algorithm is determined by
the vocabulary size. The motivation for this problem is to find the
subset of training examples which will facilitate evaluation of
prototype systems~\cite{lin11}. \arxivalt{This problem then occurs
in the form of Problem 1, where we want to find a small vocabulary
subset (which is often submodular~\cite{lin11}), subject to a
constraint that the subset acoustically spans the entire data set
(which is also often submodular~\cite{lin2009select, linbudget}). This can also be phrased as Problem 2, where we ask for maximizing the
acoustic coverage and diversity subject to a bounded vocabulary size
constraint.}{Often the objective functions encouraging small vocabulary subsets and large acoustic spans are submodular~\cite{lin11, lin2009select} and hence this problem can naturally be cast as an instance of Problems 1 and 2.}
\textbf{Privacy Preserving Communication: } Given a set of random
variables $X_1, \cdots, X_n$, denote $\mathfrak I$ as an information
source, and $\mathfrak P$ as private information that should be
filtered out. Then one way of formulating the problem of choosing a
information containing but privacy preserving set of random variables
can be posed as instances of Problems 1 and 2, with $f(A) = H(X_A|
\mathfrak I)$ and $g(A) = H(X_A| \mathfrak P)$, where $H(\cdot|\cdot)$
is the conditional entropy. \arxiv{An alternative strategy would be to
formulate the problem with $f(A) = H(X_A) + H(X_A| \mathfrak I)$ and
$g(A) = H(X_A) + H(X_A| \mathfrak P)$.}
\textbf{Machine Translation}: Another application in machine
translation is to choose a subset of training data that is optimized
for given test data set, a problem previously addressed with
modular functions \cite{moore2010intelligent}. Defining a submodular
function with ground set over the union of training and test sample
inputs $V = V_{\text{tr}} \cup V_{\text{te}}$, we can set $f:
2^{V_\text{tr}} \to \mathbb R_+$ to $f(X) = f(X | V_{\text{te}})$, and
take $g(X) = |X|$, and $b \approx 0$ in Problem 2 to address this
problem. We call this the {\em Submodular Span} problem.
\arxiv{\textbf{Probabilistic Inference:} Many problems in
computer vision and graphical model inference involve finding an
assignment to a set of random variables. The most-probable explanation
(MPE) problem finds the assignment that maximizes the probability. In
computer vision and high-tree-width Markov random fields, \arxivalt{this has
been addressed using graph-cut algorithms
~\cite{boykovJolly01}, which are
applicable when the MRF's energy function is submodular and limited
degree, and more general submodular function minimization in the case
of higher degree.}{the energy functions are often submodular~\cite{boykovJolly01}.} Moreover, many useful non-submodular energy functions have
also recently been phrased as forms of constrained submodular minimization~\cite{jegelka2011-nonsubmod-vision} which still have bounded
approximation guarantees. Some of these non-submodular energy functions can be modeled through Problems 1 and 2\arxivalt{,
where $f$ is still the submodular energy function to be minimized while
$g$ represents a submodular constraint.}{.} For example, in image co-segmentation
\cite{rother2006cosegmentation}, $V=V_1 \cup V_2$ can represent the
set of pixels in two images, $f$ is the submodular energy function of
the two images, while $g$ represents the similarity between the two
histograms of the hypothesized foreground regions in the two images, a
function shown to be submodular in
\cite{rother2006cosegmentation}\footnote{
\cite{rother2006cosegmentation} showed that $-g$ is supermodular.}. \looseness-1
Apart from the real-world applications above, both Problems 1 and 2
generalize a number of well-studied discrete optimization
problems. For example the \emph{Submodular Set Cover} problem
(henceforth SSC)~\cite{wolsey1982analysis} occurs as a special case of Problem 1,
with $f$ being modular and $g$ is submodular. Similarly the
\emph{Submodular Cost Knapsack} problem (henceforth SK)~\cite{sviridenko2004note} is a
special case of problem 2 again when $f$ is modular and $g$
submodular. Both these problems subsume the \emph{Set
Cover} and \emph{Max k-Cover}
problems~\cite{feige1998threshold}. When both $f$ and $g$
are modular, Problems 1 and 2 are called \emph{knapsack
problems}~\cite{kellerer2004knapsack}.\arxiv{Furthermore, Problem 1 also subsumes the cardinality constrained submodular minimization problem~\cite{svitkina2008submodular} and more generally the problem of minimizing a submodular function subject to a knapsack constraints. It also subsumes the problem of minimizing a submodular function subject to a matroid span constraint (by setting $g$ as the rank function of the matroid, and $c$ as the rank of the matroid). This in turn subsumes the minimum submodular spanning tree problem~\cite{goel2009approximability}, when $f$ is monotone submodular.
}\looseness-1
\arxiv{\subsection{Our Contributions}}
The following are some of our contributions. We show that Problems 1 and 2 are intimately connected,
in that any approximation algorithm for either problem can be used to
provide guarantees for the other problem as well. We then provide a framework of
combinatorial algorithms based on optimizing, sometimes iteratively,
subproblems that are easy to solve. These subproblems
are obtained by computing either upper or lower bound approximations
of the cost functions or constraining functions. We also show that
many combinatorial algorithms like the greedy algorithm for SK~\cite{sviridenko2004note} and SSC~\cite{wolsey1982analysis} \arxiv{(both of which seemingly use different
techniques) }also belong to this framework and
provide the first
constant-factor bi-criterion approximation algorithm for
SSC~\cite{wolsey1982analysis} and hence the general set cover
problem~\cite{feige1998threshold}. We then show how with suitable
choices of approximate functions, we can obtain a number of bounded
approximation guarantees and show the hardness for Problems 1 and 2,
which in fact match some of our approximation guarantees\arxiv{ up to
$\log$-factors}. Our guarantees and hardness results depend on the
\emph{curvature} of the submodular
functions~\cite{conforti1984submodular}. We observe a strong
asymmetry in the results that the factors change polynomially based on
the curvature of $f$ but only by a constant-factor with the curvature
of $g$, hence making the SK and SSC much easier compared to SCSK and
SCSC. \arxiv{Finally we empirically evaluate the performance of our
algorithms showing that the typical case behavior is much better than
these worst case bounds.}\looseness-1
\JTR{It's also important here to summarize the result showing how the curvature of $f$ and $g$
has a non-symmetric effect on the hardness of the problem, in that in one case
it only changes a constant, but the other case it changes a function of $n$.
This is the stuff that is discussed briefly at the end of \S\ref{sec:hardness} but
I think this is important and actually should also be added to the abstract (not yet done).
Could
you please add a bit of text here and in the abstract about this?}\RTJ{Added something above. Pl check.}
\section{Background and Main Ideas}
\label{background}
We first introduce several key concepts used throughout the
paper. \arxivalt{}{This paper includes only the main results and we
defer all the proofs and additional discussions to the extended
version~\cite{nipsextendedvsubcons}. }%
Given a submodular function $f$, we define the total curvature,
$\curvf{f}$\arxiv{, and the curvature of $f$ with respect to a set $X$,
$\curvf{f}(X)$,} as\footnote{We can assume, w.l.o.g that $f(j) > 0, g(j) > 0, \forall j \in V$\arxiv{, since if for any $j \in V$, $f(j) = 0$, it follows from submodularity and monotonicity that $f(j | X) = 0, \forall X \subseteq V$. Hence we can effectively remove that element $j$ from the ground set.}}:
\arxivalt{\begin{align}
\curvf{f} = 1 - \min_{j \in V} \frac{f(j | V \backslash j)}{f(j)},\;\;\text{ and }\;\; \curvf{f}(X) = 1 - \min\{\min_{j \in X} \frac{f(j | X \backslash j)}{f(j)}, \min_{j \notin X} \frac{f(j | X)}{f(j)}\}.
\end{align}}{$\curvf{f} = 1 - \min_{j \in V} \frac{f(j | V \backslash j)}{f(j)}$~\cite{conforti1984submodular\arxiv{, vondrak2010submodularity}}.}
\arxiv{The total curvature $\curvf{f}$ is then
$\curvf{f}(V)$~\cite{conforti1984submodular, vondrak2010submodularity}.}
Intuitively, the curvature $0 \leq \curvf{f} \leq 1$ measures the
distance of $f$ from modularity and $\curvf{f} = 0$ if and only if $f$
is modular (or additive, i.e., $f(X) = \sum_{j \in X} f(j)$). \arxiv{Totally normalized \cite{cun82} and
saturated functions like matroid rank have a curvature
$\curvf{f} = 1$. }A number of approximation guarantees in the context of submodular optimization have been refined
via the curvature of the submodular function~\cite{conforti1984submodular, rkiyersemiframework2013, curvaturemin}. \arxiv{For example, when
maximizing a monotone submodular function under cardinality upper
bound constraints, the bound of $1 - e^{-1}$ has been refined to
$\frac{1 - e^{-\curvf{f}}}{\curvf{f}}$
~\cite{conforti1984submodular}. Similar bounds have also been shown in
the context of constrained submodular
minimization~\cite{rkiyersemiframework2013, curvaturemin, iyermirrordescent}.} In this paper, we shall
witness the role of curvature also in determining the
approximations and the hardness of problems 1 and
2.
\narxiv{\begin{wrapfigure}[9]{r}{0.5\textwidth}
\vspace{-4.1ex}
\begin{minipage}{0.5\textwidth}}
\begin{algorithm}[H]
\caption{General algorithmic framework to address both Problems 1 and 2}
\begin{algorithmic}[1]
\FOR{$t = 1, 2, \cdots, T$}
\STATE Choose surrogate functions $\hat{f_t}$ and $\hat{g_t}$
for $f$ and $g$ respectively, tight at $X^{t-1}$.
\STATE Obtain $X^t$ as the optimizer of Problem 1 or 2 with
$\hat{f_t}$ and $\hat{g_t}$ instead of $f$ and $g$.
\ENDFOR
\end{algorithmic}
\label{alg:framework}
\end{algorithm}
\narxiv{\end{minipage}
\end{wrapfigure}}
The main idea of this paper is a framework of algorithms based on
choosing appropriate surrogate functions for $f$ and $g$ to optimize
over. This framework is represented in Algorithm~\ref{alg:framework}. We
would like to choose surrogate functions $\hat{f_t}$ and $\hat{g_t}$
such that using them, Problems 1 and 2 become easier. If the algorithm
is just single stage (not iterative), we represent the surrogates as
$\hat{f}$ and $\hat{g}$. The surrogate functions we consider in this
paper are in the forms of bounds (upper or lower) and
approximations. \arxiv{Our algorithms using upper and lower bounds are
analogous to the majorization/ minimization algorithms proposed
in~\cite{rkiyersemiframework2013, iyermirrordescent}, where we provided a unified
framework of fast algorithms for submodular optimization. We show there in
that this framework subsumes a large class of known combinatorial
algorithms and also providing a generic recipe for different forms
of submodular function optimization. We extend these ideas to the
more general context of problems 1 and 2, to obtain a fast family of
algorithms. The other type of surrogate functions we consider are
those obtained from other approximations of the functions. One such
classical approximation is the ellipsoidal
approximations~\cite{goemans2009approximating}. While computing
this approximation is time consuming, it turns out to provide the
tightest theoretical guarantees.}
\textbf{Modular lower bounds: }
Akin to convex functions, submodular functions have tight modular lower bounds. These bounds are related to the subdifferential $\partial_f(Y)$ of the submodular set function $f$ at a set $Y \subseteq V$\arxivalt{, which is defined
\cite{fujishige2005submodular}
as:
\begin{align}
\partial_f(Y) = \{y \in \mathbb{R}^n: f(X) - y(X) \geq f(Y) - y(Y)\;\text{for all } X \subseteq V\}
\end{align}
For a vector $x \in \mathbb{R}^V$ and $X \subseteq V$, we write $x(X)
= \sum_{j \in X} x(j)$.}{~\cite{fujishige2005submodular}.} Denote a
subgradient at $Y$ by $h_Y \in \partial_f(Y)$. The extreme points of
$\partial_f(Y)$ may be computed via a greedy algorithm: Let $\ensuremath{\pi}$
be a permutation of $V$ that assigns the elements in $Y$ to the first
$|Y|$ positions ($\ensuremath{\pi}(i) \in Y$ if and only if $i \leq |Y|$). Each
such permutation defines a chain with elements $S_0^\ensuremath{\pi} =
\emptyset$, $S^{\ensuremath{\pi}}_i = \{ \ensuremath{\pi}(1), \ensuremath{\pi}(2), \dots, \ensuremath{\pi}(i)
\}$ and $S^{\ensuremath{\pi}}_{|Y|} = Y$. This chain defines an extreme point
$h^{\ensuremath{\pi}}_Y$ of $\partial_f(Y)$ with entries \arxivalt{\begin{align}
h^{\ensuremath{\pi}}_Y(\ensuremath{\pi}(i)) = f(S^{\ensuremath{\pi}}_i) - f(S^{\ensuremath{\pi}}_{i-1}).
\end{align}}{$h^{\ensuremath{\pi}}_Y(\ensuremath{\pi}(i)) = f(S^{\ensuremath{\pi}}_i) -
f(S^{\ensuremath{\pi}}_{i-1})$.} Defined as above, $h^{\ensuremath{\pi}}_Y$
forms a lower bound of $f$, tight at $Y$ --- i.e.,
$h^{\ensuremath{\pi}}_Y(X) = \sum_{j \in X} h^{\ensuremath{\pi}}_Y(j) \leq f(X), \forall X
\subseteq V$ and $h^{\ensuremath{\pi}}_Y(Y) = f(Y)$.\looseness-1
\textbf{Modular upper bounds:}
We can also define superdifferentials $\partial^f(Y)$ of a submodular
function
\cite{jegelka2011-nonsubmod-vision,rkiyersubmodBregman2012}
at
$Y$\arxivalt{:
\begin{align}\label{supdiff-def}
\partial^f(Y) = \{y \in \mathbb{R}^n: f(X) - y(X) \leq f(Y) - y(Y);\text{for all } X \subseteq V\}
\end{align}}{.}
It is possible, moreover, to provide specific supergradients~\cite{rkiyersubmodBregman2012, rkiyersemiframework2013, iyermirrordescent} that define the following two modular upper bounds (when referring either one, we use $m^f_X$):
\notarxiv{\small}
\begin{align*}
m^f_{X, 1}(Y) \triangleq f(X) - \!\!\!\! \sum_{j \in X \backslash Y } f(j| X \backslash j) + \!\!\!\! \sum_{j \in Y \backslash X} f(j| \emptyset)\scalebox{1.3}{,}\;\;\;\arxiv{\\}
m^f_{X, 2}(Y) \triangleq f(X) - \!\!\! \sum_{j \in X \backslash Y } f(j| V \backslash j) + \!\!\!\! \sum_{j \in Y \backslash X} f(j| X). \nonumber
\end{align*}
\normalsize
Then $m^f_{X, 1}(Y) \geq f(Y)$ and $m^f_{X, 2}(Y) \geq f(Y), \forall Y \subseteq V$ and $m^f_{X, 1}(X) = m^f_{X, 2}(X) = f(X)$.
\textbf{MM algorithms using upper/lower bounds: } Using the modular upper
and lower bounds above in Algorithm~\ref{alg:framework}, provide a class of Majorization-Minimization
(MM) algorithms, akin to the algorithms proposed
in~\cite{rkiyersemiframework2013, iyermirrordescent} for submodular optimization and
in~\cite{narasimhanbilmes,rkiyeruai2012} for DS optimization (Problem
0 above). An appropriate choice of the bounds ensures that the
algorithm always improves the objective values for Problems 1 and
2. In particular, choosing $\hat{f_t}$ as a modular upper bound of $f$
tight at $X^t$, or $\hat{g_t}$ as a modular lower bound of $g$ tight
at $X^t$, or both, ensures that the objective value of Problems 1 and
2 always improves at every iteration as long as the corresponding
surrogate problem can be solved exactly. Unfortunately, Problems 1 and
2 are NP-hard even if $f$ or $g$ (or both) are modular~\cite{feige1998threshold}, \JTR{important:
should cite where this is shown}\RTJ{done} and therefore the surrogate
problems themselves cannot be solved exactly. Fortunately, the
surrogate problems are often much easier than the original ones and
can admit $\log$ or constant-factor guarantees. In practice,
moreover, these factors are almost $1$.
\arxivalt{In order to guarantee improvement from a theoretical stand-point however, the iterative schemes can be slightly modified using the following trick. Notice that the only case when the true valuation of
$X^t$ is better than $X^{t+1}$ is when the surrogate valuation of
$X^t$ is better than $X^{t+1}$ (since $X^{t+1}$ is only near-optimal
and not optimal). In such case, we terminate the algorithm at $X^t$.}{Furthermore, with a simple modification of the iterative procedure of Algorithm~\ref{alg:framework}, we can guarantee improvement at every iteration~\cite{nipsextendedvsubcons}.}
What is also fortunate and perhaps surprising, as we show in this
paper below, is that unlike the case of DS optimization (where the problem
is inapproximable in general \cite{rkiyeruai2012}), the constrained
forms of optimization (Problems 1 and 2) do have approximation
guarantees.
\JTR{I reworded this but double check.}\RTJ{I think that’s right}
\textbf{Ellipsoidal Approximation: }
We also consider ellipsoidal approximations (EA) of $f$. The main
result of Goemans et.\ al~\cite{goemans2009approximating} is to
provide an algorithm based on approximating the submodular polyhedron
by an ellipsoid. They show that for any polymatroid function $f$,
one can compute an approximation of the form $\sqrt{w^f(X)}$ for a
certain modular weight vector $w^f \in \mathbb R^V$, such that $\sqrt{w^f(X)} \leq f(X) \leq
O(\sqrt{n}\log{n}) \sqrt{w^f(X)}, \forall X \subseteq V$.
A simple trick then provides a curvature-dependent approximation~\cite{curvaturemin} ---
we define the $\curvf{f}$-\emph{curve-normalized{}} version of $f$ as
follows: \arxivalt{\begin{align}
\label{eq:defineg}
f^{\kappa}(X) \triangleq
\frac{f(X) - {(1 - \curvf{f})} \sum_{j \in X} f(j)}
{\curvf{f}}.
\end{align}}{$f^{\kappa}(X) \triangleq \bigl[f(X) - {(1 - \curvf{f})} \sum_{j \in X} f(j)\bigr]/ \curvf{f}$. Then, the submodular function $f^{\text{ea}}(X) = \curvf{f} \sqrt{w^{f^{\kappa}}(X)} + (1 -
\curvf{f})\sum_{j \in X} f(j)$ satisfies~\cite{curvaturemin}:
\begin{align}\label{maineq}
f^{\text{ea}}(X)
\leq f(X)
\leq O\left(\frac{\sqrt{n} \log{n}}{1 + (\sqrt{n} \log{n} - 1)(1 - \curvf{f})}\right) f^{\text{ea}}(X), \forall X \subseteq V
\end{align}}
\arxiv{The function $f^{\kappa}$ essentially contains the zero curvature component of the submodular function $f$ and the modular upper bound $\sum_{j \in X} f(j)$ contains all the linearity. The main idea is to then approximate only the polymatroidal part and retain the linear component. This simple trick improves many of the approximation bounds.
\JTR{Add something more here about the properties of $f^\kappa$, and why
it is called curve-normalized.}\RTJ{done}
We then have the following lemma.
\begin{lemma} \cite{curvaturemin}\label{cor:learn}
Given a polymatroid function $f$ with a curvature $\curvf{f} < 1$,
the submodular function $f^{\text{ea}}(X) = \curvf{f} \sqrt{w^{f^{\kappa}}(X)} + (1 -
\curvf{f})\sum_{j \in X} f(j)$ satisfies:
\begin{align}\label{maineq}
f^{\text{ea}}(X)
\leq f(X)
\leq O\left(\frac{\sqrt{n} \log{n}}{1 + (\sqrt{n} \log{n} - 1)(1 - \curvf{f})}\right) f^{\text{ea}}(X), \forall X \subseteq V
\end{align}
\end{lemma}}
$f^{\text{ea}}$ is multiplicatively bounded by $f$ by a
factor depending on $\sqrt{n}$ and the curvature. \arxiv{The
dependence on the curvature is evident from the fact that when
$\curvf{f} = 0$, we get a bound of $O(1)$, which is not surprising
since a modular $f$ is exactly represented as $f(X) = \sum_{j \in
X} f(j)$. }We shall use the result above in providing
approximation bounds for Problems 1 and 2. In particular, the
surrogate functions $\hat{f}$ or $\hat{g}$ in
Algorithm~\ref{alg:framework}
can be the ellipsoidal
approximations above, and the multiplicative bounds transform into
approximation guarantees for these problems.
\section{Relation between SCSC and SCSK}
In this section, we show a precise relationship between Problems 1 and
2. From the formulation of Problems 1 and 2, it
is clear that these problems are duals of each other. Indeed, in
this section we show that
the problems are polynomially transformable into each other.
\begin{minipage}{\textwidth}
\notarxiv{\vspace{-1.1ex}}
\centering
\begin{minipage}{.42\textwidth}
\centering
\begin{algorithm}[H]
\caption{Approx.\ algorithm for SCSK using an approximation algorithm for SCSC\arxiv{ using Linear search}.}
\begin{algorithmic}[1]
\STATE \textbf{Input:} An SCSK instance with budget $b$,
an $[\sigma, \rho]$ approx.\ algo.\ for SCSC, \& $\epsilon > 0$.
\STATE \textbf{Output: } $[(1 - \epsilon) \rho, \sigma]$ approx. for SCSK.
\STATE $c \leftarrow g(V), \hat{X_c} \leftarrow V$.
\WHILE{$f(\hat{X_c}) > \sigma b$}
\STATE $c \leftarrow (1 - \epsilon) c$
\STATE $\hat{X_c} \leftarrow [\sigma, \rho]$ approx. for SCSC using $c$.
\ENDWHILE
\arxiv{\STATE Return $\hat{X}_c$}
\end{algorithmic}
\label{alg:alg1}
\end{algorithm}
\end{minipage}
~~~
\begin{minipage}{.42\textwidth}
\centering
\begin{algorithm}[H]
\caption{Approx.\ algorithm for SCSC using an approximation algorithm for SCSK\arxiv{ using Linear search}.}
\begin{algorithmic}[1]
\STATE \textbf{Input:} An SCSC instance with cover $c$, an
$[\rho, \sigma]$ approx.\ algo.\ for SCSK, \& $\epsilon > 0$.
\STATE \textbf{Output: } $[ (1 + \epsilon)\sigma, \rho]$ approx. for SCSC.
\STATE $b \leftarrow \argmin_j f(j), \hat{X_b} \leftarrow \emptyset$.
\WHILE{$g(\hat{X_b}) < \rho c$}
\STATE $b \leftarrow (1 + \epsilon) b$
\STATE $\hat{X_b} \leftarrow [\rho, \sigma]$ approx. for SCSK using $b$.
\ENDWHILE
\arxiv{\STATE Return $\hat{X}_b$.}
\end{algorithmic}
\label{alg:alg2}
\end{algorithm}
\end{minipage}
\arxiv{\captionof{figure}{Bicriterion transformation algorithms for SCSC and SCSK using Linear search.}}
\label{fig:twoalg}
\end{minipage}
\normalsize
\arxiv{\vspace{1.1ex}}
We first introduce the notion of
bicriteria algorithms. An algorithm is a $[\sigma, \rho]$ bi-criterion
algorithm for Problem 1 if it is guaranteed to obtain a set $X$ such
that $f(X) \leq \sigma f(X^*)$ (approximate optimality) and $g(X) \geq
c^{\prime} = \rho c$ (approximate feasibility), where $X^*$ is an optimizer of
Problem 1. Similarly, an algorithm is a $[\rho,\sigma]$ bi-criterion algorithm for Problem 2 if it is guaranteed to
obtain a set $X$ such that $g(X) \geq \rho g(X^*)$ and $f(X) \leq b^{\prime} =
\sigma b$, where $X^*$ is the optimizer of Problem 2. In a
bi-criterion algorithm for Problems 1 and 2, typically $\sigma \geq 1$
and $\rho \leq 1$. \arxiv{We call these type of approximation algorithms, bi-criterion approximation algorithms of type 1.
We can also view the bi-criterion approximations from another angle. We can say that for Problem 1, $\hat{X}$ is a feasible solution (i.e., it satisfies the constraint), and is a $[\sigma, \rho]$ bi-criterion approximation, if $f(\hat{X}) \leq \sigma f(\hat{X}^*)$, where $\hat{X}^*$ is the optimal solution to the problem $\min\{f(X) | g(X) \geq c/\rho\}$. Similarly for Problem 2, we can say that $\hat{X}$ is a feasible solution (i.e., it satisfies the constraint), and is a $[\rho, \sigma]$ bi-criterion approximation, if $g(\hat{X}) \geq \rho g(\hat{X}^*)$, where $\hat{X}^*$ is the optimal solution to the problem $\max\{g(X) | f(X) \leq b \sigma\}$. We call these the bi-criterion approximation algorithms of type 2.
It is easy to see that these algorithms can easily be transformed into each other. For example, for problem 1, a bi-criterion algorithm of type-I can obtain a guarantee of type-II if we run it till a covering constraint of $c^{\prime}/\rho$ (where $c^{\prime}$ in this case, is the approximate covering constraint which the type-I algorithm needs to satisfy -- note that it need not be the actual covering constraint of the problem). Similarly an algorithm of type-II can obtain a guarantee of type-I if run till a covering constraint of $\rho c$ (in this case, $c$ is the actual 'covering' constraint, since a type-II approximate algorithm provides a feasible set). We can similarly transform these guarantees for problem 2. In particular, a bi-criterion algorithm of type-I can be used to obtain a guarantee of type-II if we run it till a budget of $b^{\prime} \sigma$ (again, here $b^{\prime}$ is the approximate budget of the type-I algorithm). Similarly an algorithm of type-II can obtain a guarantee of type-I if run till a covering constraint of $b/\sigma$ (in this case, $b$ is the budget of the original problem).
Though both type-I and type-II guarantees are easily transformable into each other, through the rest of this paper whenever we refer to bi-criterion approximations, we shall consider only the type-I approximations.
}A {\em non-bicriterion} algorithm for Problem 1 is
when $\rho = 1$ and a {\em non-bicriterion} algorithm for Problem 2 is when
$\sigma = 1$.
Algorithms~\ref{alg:alg1} and \ref{alg:alg2} provide the schematics
for using an approximation algorithm for one of the problems
for solving the other.\narxiv{\looseness-1}\arxiv{
The main idea of
Algorithm~\ref{alg:alg1} is to start with the ground set $V$ and
reduce the value of $c$ (which governs SCSC), until the valuation of
$f$ just falls below $\sigma b$. At that point, we are guaranteed to
get a $((1 - \epsilon) \rho, \sigma)$ solution for SCSK. Similarly in
Algorithm~\ref{alg:alg2}, we increase the value of $b$ starting at the
empty set, until the valuation at $g$ falls above $\rho c$. At this
point we are guaranteed a $(( 1 + \epsilon) \sigma, \rho)$ solution for SCSC. In order to avoid degeneracies, we assume that $f(V) \geq b \geq \min_j f(j)$ and $g(V) \geq c \geq \min_j g(j)$, else the solution to the problem is trivial.}
\begin{theorem}
\label{thm1}
Algorithm~\ref{alg:alg1} is guaranteed to find a set $\hat{X_c}$
which is a $[(1 - \epsilon) \rho, \sigma]$ approximation of SCSK in at most
$\log_{1/(1 - \epsilon)} [g(V)/\min_j g(j)]$ calls to the $[\sigma, \rho]$ approximate
algorithm for SCSC. Similarly, Algorithm~\ref{alg:alg2} is
guaranteed to find a set $\hat{X_b}$ which is a $[(1 + \epsilon)
\sigma, \rho]$ approximation of SCSC in $\log_{1 + \epsilon} [f(V)/\min_j f(j)]$ calls
to a $[\rho, \sigma]$ approximate algorithm for SCSK.
\end{theorem}
\arxiv{
\begin{proof}
We start by proving the first part, for Algorithm~\ref{alg:alg1}. Notice that
Algorithm~\ref{alg:alg1} converges when $f(\hat{X_c})$ just falls
below $\sigma b$. Hence $f(\hat{X_c}) \leq \sigma b$ (is approximately feasible) and at
the previous step $c^{\prime} = c/(1 - \epsilon)$, we have that
$f(\hat{X_{c^{\prime}}}) > \sigma b$. Denoting
$X^*_{c^{\prime}}$ as the optimal solution for SCSC at $c^{\prime}$,
we have that $f(X^*_{c^{\prime}}) > b$ (a fact which follows from
the observation that $\hat{X_c}$ is a $[\sigma, \rho]$ approximation
of SCSC at $c$). Hence if $X^*$ is the optimal solution of SCSK, it
follows that $g(X^*) < c^{\prime}$. The reason for this is that,
suppose, $g(X^*) \geq c^{\prime}$. Then it follows that $X^*$ is a
feasible solution for SCSC at $c^{\prime}$ and hence $f(X^*) \geq
f(X^*_{c^{\prime}}) > b$. This contradicts the fact that $X^*$ is an
optimal solution for SCSK (since it is then not even feasible).
Next, notice that $\hat{X_c}$ satisfies that $g(\hat{X_c}) \geq \rho
c$, using the fact that $\hat{X_c}$ is obtained from a $(\sigma,
\rho)$ bi-criterion algorithm for SCSC. Hence,
\begin{align}
g(\hat{X_c})
\geq \rho c
= \rho (1 - \epsilon) c^{\prime}
> \rho (1 - \epsilon) g(X^*)
\end{align}
Hence the Algorithm~\ref{alg:alg1} is a $((1 - \epsilon) \rho,
\sigma)$ approximation for SCSK.
In order to show the converge rate, notice that $c \geq \min_j g(j) > 0$. Since $\sigma \geq 1$ and $b \geq \min_j f(j)$, we can guarantee that this algorithm will stop before $c$ reaches $\min_j g(j)$. The reason is that, when $c = \min_j g(j)$, the minimum value of $f(X)$ such that $g(X) \geq c$ is $\min_j f(j)$, which is smaller than $b$. Moreover, since $\sigma > 1$, it implies that the algorithm would have terminated before this point.
The proof for the second part of the statement, for
Algorithm~\ref{alg:alg2}, is omitted since it is shown using a
symmetric argument.
\end{proof}
}%
\arxiv{\noindent\begin{minipage}{\textwidth}
\centering
\begin{minipage}{.42\textwidth}
\centering
\begin{algorithm}[H]
\caption{Approx.\ algorithm for SCSK using an approximation algorithm for SCSC\arxiv{ using Binary search}.}
\begin{algorithmic}[1]
\STATE \textbf{Input:} An SCSK instance with budget $b$,
an $[\sigma, \rho]$ approx.\ algo.\ for SCSC, \& $\epsilon > 0$.
\STATE \textbf{Output: } $[(1 - \epsilon) \rho, \sigma]$ approx. for SCSK.
\STATE $c_{\min} \leftarrow \min_j g(j), c_{\max} \leftarrow g(V)$
\WHILE{$c_{\max} - c_{\min} \geq \epsilon c_{\max}$ }
\STATE $c \leftarrow [c_{\max} + c_{\min}]/2$
\STATE $\hat{X_c} \leftarrow [\sigma, \rho]$ approx. for SCSC using $c$.
\IF{$f(\hat{X_c}) > \sigma b$}
\STATE $c_{\max} \leftarrow c$
\ELSE
\STATE $c_{\min} \leftarrow c$
\ENDIF
\ENDWHILE
\STATE Return $\hat{X}_{c_{\min}}$
\end{algorithmic}
\arxivalt{\label{alg:alg3}}{\label{alg:alg1}}
\end{algorithm}
\end{minipage}
~~~
\begin{minipage}{.42\textwidth}
\centering
\begin{algorithm}[H]
\caption{Approx.\ algorithm for SCSC using an approximation algorithm for SCSK\arxiv{ using Binary search}.}
\begin{algorithmic}[1]
\STATE \textbf{Input:} An SCSC instance with cover $c$, an
$[\rho, \sigma]$ approx.\ algo.\ for SCSK, \& $\epsilon > 0$.
\STATE \textbf{Output: } $[ (1 + \epsilon)\sigma, \rho]$ approx. for SCSC.
\STATE $b_{\min} \leftarrow \argmin_j f(j), b_{\max} \leftarrow f(V)$.
\WHILE{$b_{\max} - b_{\min} \geq \epsilon b_{\min}$ }
\STATE $b \leftarrow [b_{\max} + b_{\min}]/2$
\STATE $\hat{X_b} \leftarrow [\rho, \sigma]$ approx. for SCSK using $b$
\IF{$g(\hat{X_b}) < \rho c$}
\STATE $b_{\min} \leftarrow b$
\ELSE
\STATE $b_{\max} \leftarrow b$
\ENDIF
\ENDWHILE
\STATE Return $\hat{X}_{b_{\max}}$
\end{algorithmic}
\arxivalt{\label{alg:alg4}}{\label{alg:alg2}}
\end{algorithm}
\end{minipage}
\arxiv{\captionof{figure}{Bicriterion transformation algorithms for SCSC and SCSK using Binary search.}}
\label{fig:twoalg}
\end{minipage}}
\arxiv{\vspace{1.1ex}}
Theorem~\ref{thm1} implies that the complexity of Problems 1 and 2 are
identical, and a solution to one of them provides a solution to the
other. Furthermore, as expected, the hardness of Problems 1 and 2 are
also almost identical. When $f$ and $g$ are polymatroid functions,
moreover, we can provide bounded approximation guarantees for both
problems, as shown in the next section. \narxiv{Alternatively we can also do a binary search instead of a linear search to transform Problems 1 and 2. This essentially turns the factor of $O(1/\epsilon)$ into $O(\log 1/\epsilon)$. Due to lack of space, we defer this discussion to the extended version~\cite{nipsextendedvsubcons}.\looseness-1}
\arxiv{Alternatively we can also do a binary search instead of a linear search to transform Problems 1 and 2. This essentially turns the factor of $O(1/\epsilon)$ into $O(\log 1/\epsilon)$. Algorithms~\ref{alg:alg3} and \ref{alg:alg4} show the transformation algorithms using binary search. While the binary search also ensures the same performance guarantees, it does so more efficiently.
\begin{theorem}
\label{thm2}
Algorithm~\ref{alg:alg3} is guaranteed to find a set $\hat{X_c}$
which is a $[(1 - \epsilon) \rho, \sigma]$ approximation of SCSK in at most
$\log_2 \frac{[g(V)/\min_j g(j)]}{\epsilon}$ calls to the $[\sigma, \rho]$ approximate
algorithm for SCSC. Similarly, Algorithm~\ref{alg:alg4} is
guaranteed to find a set $\hat{X_b}$ which is a $[(1 + \epsilon)
\sigma, \rho]$ approximation of SCSC in $\log_2 \frac{[f(V)/\min_j f(j)]}{\epsilon}$ calls
to a $[\rho, \sigma]$ approximate algorithm for SCSK. When $f$ and $g$ are integral, moreover, Algorithm~\ref{alg:alg3} obtains a $[\rho, \sigma]$ bi-criterion approximate solution for SCSK in $\log_2 g(V)$ iterations, and similarly Algorithm~\ref{alg:alg4} obtains a $[\sigma, \rho]$ bi-criterion approximate solution for SCSC in $\log_2 g(V)$ iterations.
\end{theorem}
\begin{proof}
To show this theorem, we use the result from Theorem~\ref{thm1}. Let $c = c_{\min}$ and $c^{\prime} = c_{\max}$. An important observation is that throughout the algorithm, the values of $c_{\min}$ satisfy $f(\hat{X}_{c_{\min}}) \leq \sigma b$ and $f(\hat{X}_{c_{\max}}) > \sigma b$. Hence, $f(\hat{X_{c}}) \leq \sigma b$ and $f(\hat{X_{c^{\prime}}}) > \sigma b$. Moreover, notice that $c^{\prime}/c = c_{\max}/c_{\min} = 1/(1 - \epsilon)$. Hence using the proof of Theorem~\ref{thm1} the approximation guarantee follows.
In order to show the complexity, notice that $c_{\max} - c_{\min}$ is decreasing throughout the algorithm. At the beginning, $c_{\max} - c_{\min} \leq g(V)$ and at convergence, $c_{\max} - c_{\min} \geq \epsilon c_{\max}/2 \geq \epsilon \min_j g(j)/2$. The bound at convergence holds since, let $c^{\prime}_{max}$ and $c^{\prime}_{min}$ be the values at the previous step. It holds that $c^{\prime}_{\max} - c^{\prime}_{\min} \geq \epsilon c^{\prime}_{\max}$. Moreover, $c_{max} - c_{min} = (c^{\prime}_{\max} - c^{\prime}_{\min})/2 \geq \epsilon c^{\prime}_{\max}/2 \geq \epsilon c_{\max}$. Hence the number of iterations is bounded by $\log_2 \frac{[2g(V)/\min_j g(j)]}{\epsilon}$. Moreover, when $f$ and $g$ are integral, the analysis is much simpler. In particular, notice that once $c_{max} - c_{min} = 1$, the algorithm will stop at the next iteration (this is because at this point, $c = (c_{max} + c_{min})/2$ is equivalent to $c = c_{max}$). Hence, the number of iterations is bounded by $\log_2 g(V)$, and we can exactly obtain a $[\rho, \sigma]$ bi-criterion approximation algorithm.
The proof for the second part of the statement, for
Algorithm~\ref{alg:alg2}, similarly follows using a symmetric argument.
\end{proof}
When $f$ and $g$ are integral, this removes the $\epsilon$ dependence on the factors and could potentially be much faster in practice. We also remark that a specific instance of such a transformation has been used~\cite{krause08robust}, for a specific class of functions $f$ and $g$. We shall show in section~\ref{extentions} that their algorithm is in fact a special case of Algorithm~\ref{alg:alg3} through a specific construction to convert the non-submodular problem into an instance of a submodular one.
Algorithm~\ref{alg:alg1} and Algorithm~\ref{alg:alg2} indeed provide an
interesting theoretical result connecting the complexity of Problems
1 and 2. In the next section however, we provide distinct algorithms for each problem when $f$ and $g$ are polymatroid functions --- this turns out to
be faster than having to resort to the iterative reductions above.}
\section{Approximation Algorithms}
We consider several algorithms for Problems 1 and 2, which can all be
characterized by the framework of
Algorithm~\ref{alg:framework}, using the surrogate functions of
the form of upper/lower bounds or approximations.\looseness-1
\subsection{Approximation Algorithms for SCSC}
\label{sec:appr-algor-scsc}
We first describe our approximation algorithms designed specifically
for SCSC, leaving to \S\ref{sec:appr-algor-scsk} the presentation of
our algorithms slated for SCSK. We first investigate a special case, the submodular set cover (SSC), and then provide two algorithms, one of them (ISSC) is very practical with a weaker theoretical guarantee, and another one (EASSC) which is slow but has the tightest guarantee.
\textbf{Submodular Set Cover (SSC): }
We start by considering a classical special case of SCSC (Problem 1)
where $f$ is already a modular function and $g$ is a submodular
function. This problem occurs naturally in a number of problems
related to active/online learning~\cite{guillory2010interactive} and summarization~\cite{linbudget,lin2011-class-submod-sum}. This
problem was first investigated by Wolsey~\cite{wolsey1982analysis},
wherein he showed that a simple greedy algorithm achieves bounded (in
fact, log-factor) approximation guarantees. We show that this greedy
algorithm can naturally be viewed in the framework of our
Algorithm~\ref{alg:framework} by choosing appropriate surrogate
functions $\hat{f_t}$ and $\hat{g_t}$. The idea is to use the modular
function $f$ as its own surrogate $\hat{f_t}$ and choose the function
$\hat{g_t}$ as a modular lower bound of $g$. Akin to the framework of
algorithms in~\cite{rkiyersemiframework2013}, the crucial factor is
the choice of the lower bound (or subgradient). Define the
\emph{greedy subgradient} \arxiv{(equivalently the \emph{greedy
permutation})} as:
\begin{align}
\label{greedyssc}
\ensuremath{\pi}(i) \in
\argmin \left\{\frac{f(j)}{g(j | S^{\ensuremath{\pi}}_{i-1})} \ \bigg\vert \ j \notin S^{\ensuremath{\pi}}_{i-1} ,
g(S^{\ensuremath{\pi}}_{i-1} \cup j) < c \right\}.
\end{align}
Once we reach an $i$ where the constraint $g(S^{\ensuremath{\pi}}_{i-1} \cup j)
< c$ can no longer be satisfied by any $j \notin S^{\ensuremath{\pi}}_{i-1}$, we
choose the remaining elements for $\ensuremath{\pi}$ arbitrarily. Let the corresponding
subgradient be referred to as $h^{\ensuremath{\pi}}$.
\arxiv{Let $N$ be the minimum $i$
such that $g(S^{\ensuremath{\pi}}_i) \geq c$ and $\theta_i = \min_{j \notin S^{\ensuremath{\pi}}_{i-1}} \frac{f(j)}{g(j |
S^{\ensuremath{\pi}}_{i-1})}$.} Then we have the following lemma,
which is an extension of \cite{wolsey1982analysis}\arxivalt{:}{, }
\arxivalt{\begin{lemma} \label{setcover} Choosing the surrogate
function $\hat{f}$ as $f$ and $\hat{g}$ as $h^{\ensuremath{\pi}}$ (with
$\ensuremath{\pi}$ defined in Eqn.~\eqref{greedyssc}) in
Algorithm~\ref{alg:framework}, at the end of the first iteration,
we are guaranteed to obtain a set $X^g$ such that
\begin{align}
\frac{f(X^g)}{f(X^*)} \leq 1 + \log_e \min\{\lambda_1, \lambda_2, \lambda_3\}
\end{align}
where $\lambda_1 = \max \{\frac{1}{1 - \kappa_g(S^{\ensuremath{\pi}}_i)} | \
i: c(S^{\ensuremath{\pi}}_i) < 1\}$, $\lambda_2 = \frac{\theta_N}{\theta_1}$
and $\lambda_3 = \frac{g(V) - g(\emptyset)}{g(V) -
g(S^{\ensuremath{\pi}}_{N-1})}$. Furthermore if $g$ is integral,
$\frac{f(X^g)}{f(X^*)} \leq H(\max_j g(j))$, where $H(d) = \sum_{i =
1}^d \frac{1}{i}$ for a positive integer $d$.
\end{lemma}
\begin{proof}
The permutation $\ensuremath{\pi}$ is chosen based on a greedy ordering
associated with the submodular set cover problem, and therefore
$h^{\ensuremath{\pi}}(S^{\ensuremath{\pi}}_N) = g(S^{\ensuremath{\pi}}_N) \geq c$ (where the
inequality follows from the definition of $N$), and thus
$S^{\ensuremath{\pi}}_N$ is a feasible solution in the surrogate problem
(where $g$ is replaced by $h^\ensuremath{\pi}$). The resulting knapsack
problem can be addressed using the greedy algorithm
\cite{kellerer2004knapsack}, but this exactly corresponds to the
greedy algorithm of submodular set cover \cite{wolsey1982analysis} and hence the guarantee
follows from \JTR{I added the theorem number, but double check that
this is really theorem 1, I think it is} Theorem 1
in~\cite{wolsey1982analysis}\RTJ{Yup, I think that's right}.
\end{proof}
}
and which is a simpler description of the result stated formally in~\cite{nipsextendedvsubcons}.\looseness-1
\begin{lemma}
\label{setcover}
The greedy algorithm for SSC~\cite{wolsey1982analysis} can be seen as an instance of
Algorithm~\ref{alg:framework} by choosing the surrogate function
$\hat{f}$ as $f$ and $\hat{g}$ as $h^{\ensuremath{\pi}}$ (with $\ensuremath{\pi}$
defined in Eqn.~\eqref{greedyssc}).
\end{lemma}
}
\arxiv{The surrogate problem in this instance is a simple knapsack problem
that can be solved nearly optimally using dynamic
programming~\cite{vazirani2004approximation}. As stated in the
proof, the greedy
algorithm for the submodular set cover
problem~\cite{wolsey1982analysis} is in fact equivalent to
using the greedy algorithm for the knapsack
problem~\cite{kellerer2004knapsack}, which is in fact suboptimal.} When
$g$ is integral, the guarantee of the greedy algorithm is $H_g
\triangleq H(\max_j g(j))$, where $H(d) = \sum_{i = 1}^d
\frac{1}{i}$~\cite{wolsey1982analysis} (henceforth we will use $H_g$ for this
quantity). This factor is tight
up to lower-order terms~\cite{feige1998threshold}. Furthermore, since
this algorithm directly solves SSC, we call it the \emph{primal greedy}. We could also
solve SSC by looking at its \emph{dual}, which is SK~\cite{sviridenko2004note}. Although SSC does
not admit any constant-factor approximation algorithms
\cite{feige1998threshold}, we can obtain a constant-factor {\em
bi-criterion} guarantee:\looseness-1
\begin{lemma}
\label{thm:dual_greedy}
Using the greedy algorithm for SK~\cite{sviridenko2004note} as the approximation oracle in
Algorithm~\ref{alg:alg2} provides a $[1 + \epsilon, 1 - e^{-1}]$
bi-criterion approximation algorithm for SSC, for any $\epsilon > 0$.
\end{lemma}
\JTR{Note that the Sviridenko algorithm of \cite{sviridenko2004note} requires
partial enumeration which is an $O(n^3)$ step rendering this approach no longer
practical. In our 2010
paper \cite{linbudget}
we show that w/o the $O(n^3)$ step still
has a $1-1/\sqrt{e} \approx 0.33$ algorithm which still works well in practice,
and that would also give a different bi-criterion approximation here
of $[1+\epsilon,1-e^{-1/2}]$.
Note that this was also shown in \url{http://las.ethz.ch/files/krause05note.pdf} ,
I added the bib here \cite{krause05note},
but
at the time in 2010, we weren't aware of Andreas's result (my fault :-).
}\RTJ{Added a note below.}
\arxiv{
\begin{proof}}We call this the \emph{dual greedy}. This result follows immediately from the guarantee of the submodular cost knapsack problem~\cite{sviridenko2004note} and Theorem~\ref{thm1}. \arxiv{\end{proof}} We remark that we can also use a simpler version of the greedy iteration at every iteration~\cite{linbudget, krause05note} and we obtain a guarantee of $(1 + \epsilon, 1/2(1 - e^{-1}))$. In practice, however, both these factors are almost $1$ and hence the simple variant of the greedy algorithm suffices.\looseness-1
\arxiv{
An interesting connection between the greedy algorithm and the induced orderings, allows us to further simplify this dual algorithm. A nice property of the greedy algorithm for the submodular knapsack problem is that it can be completely parameterized by the chain of sets (this holds for the greedy algorithm of~\cite{linbudget, krause05note} for knapsack constraints and the basic greedy algorithm of~\cite{nemhauser78} under cardinality constraints). In particular, having computed the greedy chain of sets, and given a value of $b$ or the budget, we can easily find the corresponding set in $O(\log n)$ time using binary search. Moreover, this also implies that we can do the transformation algorithms by just iterating through the chain of sets once. In particular, the linear search over the different values of $\epsilon$ is equivalent to the linear search over the different chain of sets. Moreover, we could also do the much faster binary search. Hence the complexity of the dual greedy algorithm is almost identical to the primal greedy one for the submodular set cover problem.
It is also important to put the bicriterion result into perspective. Notice that the bicriterion guarantee suggests that we find only an approximate feasible solution. In particular, the dual greedy algorithm provides a type-I bi-criterion approximation, -- that the solution obtained by running the algorithm with a cover constraint of $(1 - 1/e) c$ is competitive to the optimal solution with a cover constraint of $c$. However, we can also obtain a type-II bi-criterion approximation by running the dual greedy algorithm until it satisfies the cover constraint of $c$. In this case, we would obtain a feasible solution. The guarantee, however, would say that the resulting solution would be competitive to the optimal solution obtained with a cover constraint of $c/(1 - e^{-1})$. Furthermore, these factors are in practice close to $1$, and the primal and dual greedy algorithms would both perform very well empirically.
}
\textbf{Iterated Submodular Set Cover (ISSC): }
We next investigate an algorithm for the general SCSC problem when both $f$ and $g$
are submodular. The idea here is to iteratively solve the submodular
set cover problem which can be done by replacing $f$ by a modular
upper bound at every iteration. In particular, this can be seen as a
variant of Algorithm~\ref{alg:framework}, where we start with \arxivalt{$X^0 =
\emptyset$ and choose $\hat{f_t}(X) = m^f_{X^t, 2}(X)$ at every
iteration (alternatively, we can choose $\hat{f_t}(X) = m^f_{X^t,
1}(X)$). At the first iteration with $X^0 = \emptyset$, either
variant then corresponds to the set cover problem with the simple
modular upper bound $f(X) \leq m^f_{\emptyset}(X) = \sum_{j \in X}
f(j)$ where $m^f_{X^t}$ refers to either variant.}{$X^0 =
\emptyset$ and choose $\hat{f_t}(X) = m^f_{X^t}(X)$ at every
iteration.} The surrogate
problem at each iteration becomes\arxiv{:}
\arxivalt{
\begin{alignat*}{2}
\text{minimize } & m^f_{X^t}(X) \\
\text{subject to } & g(X) \geq c.
\end{alignat*}
}{
$\min\{ m^f_{X^t}(X) | g(X) \geq c\}$.
}Hence, each iteration is an instance of SSC and can be
solved nearly optimally using the greedy algorithm. We can continue
this algorithm for $T$ iterations or until convergence. An analysis
very similar to the ones in~\cite{rkiyeruai2012,
rkiyersemiframework2013} will reveal polynomial time convergence.
\JTR{for completeness, in supplementary material, this should ultimately be
included here in its current form.} Since each iteration is only the
greedy algorithm, this approach is also highly practical and
scalable. \arxiv{Since there are two approaches to solve the set cover
problem (the primal
approach of
Lemma~\ref{setcover}
and the dual greedy
approach of
Lemma~\ref{thm:dual_greedy}), we have two forms of ISSC,
the \emph{primal ISSC} and the \emph{dual ISSC}. The
following shows the resulting theoretical guarantees:}\looseness-1
\begin{theorem}\label{MIguar}
\arxiv{The primal} ISSC\arxiv{ algorithm} obtains an approximation factor of
$\frac{K_g H_g}{1 + (K_g - 1)(1 - \curvf{f})} \leq
\frac{n}{1 + (n - 1)(1 - \curvf{f})} H_g$ where $K_g = 1 +
\max \{|X|: g(X) < c\}$ and $H_g$ is the approximation factor
of the submodular set cover using $g$. \arxiv{Similarly the dual ISSC obtains
a bi-criterion guarantee of $\left[\frac{(1 + \epsilon) K_g}{1 + (K_g -
1)(1 - \curvf{f})}, 1 - e^{-1}\right]$.}
\end{theorem}
\JTR{again, we can get a more scalable version w/o the $O(n^3)$ step and with approx $1-e^{-1/2}$, but
in this case maybe it is no longer practical due to the need to use the
dual solver of Algorithm~\ref{alg:alg2}.}\RTJ{Yeah. I think given that we already mentioned this remark in the case of SSC, we can just keep the $1 - 1/e$ henceforth.}
\arxiv{\begin{proof}
The first part of the result follows directly from Theorem 5.4 in~\cite{rkiyersemiframework2013}. In particular, the result from~\cite{rkiyersemiframework2013} ensures a guarantee of
\begin{align}
\frac{\beta|X^*|}{1 + (|X^*| - 1)(1 - \curvf{f})}
\end{align}
for the problem of $\min\{f(X) | X \in \mathcal C\}$ where $\beta$ is
the approximation guarantee of solving a modular function over
$\mathcal C$ where $\mathcal C$ is the feasible set. In this case,
$\beta = H_g$ and $|X^*| \leq K_g$.
When using the dual greedy approach at every iteration, we
can use a similar form of the result in the bi-criterion sense. Consider only the first iteration of this algorithm (due to the monotonicity of the algorithm, we will only improve the objective value). We are then guaranteed to obtain a set $\hat{X}$ such that (denote $X^1$ as the solution after the first iteration)
\begin{align}
f(\hat{X}) \leq f(X^1) \leq m^f_{\emptyset}(X^1) \leq (1 + \epsilon)m^f_{\emptyset}(X^*) \leq \frac{K_g (1 + \epsilon)}{1 + (K_g - 1)(1 - \curvf{f})} f(X^*)
\end{align}
The inequalities above follow from the fact that the modular upper bound $m^f_{\emptyset}(X)$ satisfies~\cite{curvaturemin},
\begin{align}
m^f_{\emptyset}(X) \leq f(X) \leq \frac{|X|}{1 + (|X| - 1)(1 - \curvf{f})} f(X)
\end{align}
and the fact that $X^1$ which is the solution obtained through the dual set cover with a cover constraint $(1 - e^{-1})c$ satisfies $m^f_{\emptyset}(X^1) \leq (1 + \epsilon)m^f_{\emptyset}(X^*)$.
In the above, $X^*$ is the optimal solution to the problem $\min\{ f(X) | g(X)
\geq c\}$. We could also run the dual set cover algorithm to obtain a feasible solution (i.e., a type-II guarantee). In this case, the guarantee would compete with the optimal solution satisfying a cover constraint of $c/(1 - e^{-1})$.
\end{proof}}
From the above, it is clear that $K_g \leq n$. Notice also that $H_g$
is essentially a log-factor. We also see an interesting effect of the
curvature $\curvf{f}$ of $f$. When $f$ is modular ($\curvf{f} = 0$),
we recover the approximation guarantee of the submodular set cover
problem. Similarly, when $f$ has restricted curvature, the guarantees can be much better. \arxiv{For example, using square-root over modular function $f(X) = \sum_{i = 1}^k \sqrt{w_i(X)}$, which is common model used in applications~\cite{rkiyeruai2012, lin2012submodularity, jegelkacvpr}, the worst case guarantee is $H_g \sqrt{K_g}$. This follows directly from the results in~\cite{curvaturemin}. } Moreover, the approximation guarantee already holds after the
first iteration, so additional iterations can only further improve the
objective. \arxiv{
We remark here that a special case of ISSC, using only the first iteration (i.e., the simple modular upper bound of $f$) was considered in\arxivalt{~\cite{wan2010greedy,du2011minimum}}{~\cite{wan2010greedy}}. Our algorithm not only possibly improves upon theirs, but our approximation guarantee is also more explicit than theirs. In particular, they show a guarantee of $\nu_f H_g$, where $\nu_f = \min\{\frac{\sum_{i \in X} f(i)}{f(X)} | g(X) = g(V)\}$. Since this factor $\nu_f$ itself involves an optimization problem, it is not clear how to efficiently compute this factor. Moreover given a submodular function $f$, it is also not evident how good this factor is. While our guarantee is an upper bound of $\nu_f H_g$, it is much more explicit in it’s dependence on the parameters of the problem. It can also be computed efficiently and has an intuitive significance related to the curvature of the function. Furthermore, our bound is also tight since with, for example, $f(X) = \min\{|X|, 1\}$, our exactly matches the bound of~\cite{wan2010greedy, du2011minimum}. Lastly our algorithm also potentially improves upon theirs thanks to it’s iterative nature.}
\arxiv{For another variant of this algorithm, we can replace $g$ with
its greedy modular lower bound at every iteration. Then, rather
than solving every iteration through the greedy algorithm, we can
solve every iteration as a knapsack problem (minimizing a modular
function over a modular lower bound constraint)
\cite{kellerer2004knapsack}, using say, a dynamic programming based approach. This could potentially improve over the
greedy variant, but at a potentially higher computational
cost.}\looseness-1
\textbf{Ellipsoidal Approximation based Submodular Set Cover (EASSC): }
In this setting, we use the ellipsoidal approximation discussed in
\S\ref{background}. We can compute the $\curvf{f}$-curve-normalized{}
version of $f$ ($f^{\kappa}$, see \S\ref{background}), and then
compute its ellipsoidal approximation $\eax{f^\kappa}$. We then
define the function $\hat{f}(X) = f^{\text{ea}}(X) = \curvf{f} \sqrt{w^{f^{\kappa}}(X)} + (1 - \curvf{f}) \sum_{j \in X} f(j)$ and use this as
the surrogate function $\hat{f}$ for $f$. We choose $\hat{g}$ as $g$
itself. The surrogate problem becomes:
\begin{align}\label{EASSCeq}
\min \left\{\curvf{f} \sqrt{w^{f^{\kappa}}(X)} + (1 - \curvf{f}) \sum_{j \in X} f(j) \ \bigg\vert \ g(X) \geq c \right\}.
\end{align}
While function $\hat{f}(X)=f^{\text{ea}}(X)$ is not modular, it is a
weighted sum of a concave over modular function and a modular
function. Fortunately, we can use the result
from~\cite{nikolova2010approximation}, where they show that any
function of the form of $\sqrt{w_1(X)} + w_2(X)$ can be optimized over
any polytope $\mathcal P$ with an approximation factor of $\beta (1 +
\epsilon)$ for any $\epsilon > 0$, where $\beta$ is the approximation
factor of optimizing a modular function over $\mathcal P$. The
complexity of this algorithm is polynomial in $n$ and
$\frac{1}{\epsilon}$. \arxiv{The main idea of their paper is to reduce
the problem of minimizing $\sqrt{w_1(X)} + w_2(X)$, into $\log n$
problems of minimizing a modular function over the polytope.} We use
their algorithm to minimize $f^{\text{ea}}(X)$ over the
\emph{submodular set cover} constraint and hence we call this algorithm EASSC. \arxiv{Again we have the two variants, \emph{primal
EASSC} and \emph{dual EASSC}, which essentially use at every iteration the
primal and dual forms of set cover.}
\begin{theorem}\label{EAguar}
\arxiv{The primal}
EASSC obtains a guarantee of $O(\frac{\sqrt{n}\log n H_g}{ 1 +
(\sqrt{n}\log n - 1)(1 - \curvf{f})} )$, where $H_g$ is the
approximation guarantee of the set cover problem. \arxiv{Moreover, the dual EASSC obtains a bi-criterion approximation of
$\left[O(\frac{\sqrt{n}\log n}{ 1 + (\sqrt{n}\log n - 1)(1 -
\curvf{f})}), 1 - e^{-1}\right]$.}
\end{theorem}
\arxiv{
\begin{proof}
The idea of the proof is to use the result
from~\cite{nikolova2010approximation} where they show that any
function of the form $\lambda_1 \sqrt{m_1(X)} + \lambda_2 m_2(X)$
where $\lambda_1 \geq 0, \lambda_2 \geq 0$ and $m_1$ and $m_2$ are
positive modular functions has a FPTAS, provided a modular function
can easily be optimized over $\mathcal C$. Note that our function is
exactly of that form. Hence, $\hat{f}(X)$ can be approximately
optimized over $\mathcal C$. It now remains to show that this
translates into the approximation guarantee. From
Lemma~\ref{cor:learn}, we know that there exists a $\hat{f}$
such that $\hat{f}(X) \leq f(X) \leq \beta(n) \hat{f}(X), \forall X$
where
\begin{align}
\beta(n) = O\left(\frac{\sqrt{n} \log n}{(\sqrt{n} \log n - 1)(1 - \curvf{f}) + 1)}\right).
\end{align}
Then, if $\hat{X}$ is the $1 + \epsilon$ approximately optimal
solution for minimizing $\hat{f}$ over $\{X: g(X) \geq c\}$, we
have that:
\begin{align}
f(\hat{X})
\leq \beta(n) \hat{f}(\hat{X})
\leq H_g \beta(n) (1 + \epsilon) \hat{f}(X^*)
\leq H_g \beta(n) (1 + \epsilon) f(X^*),
\end{align}
where $X^*$ is the optimal solution. We can set $\epsilon$ to any
constant, say $1$, and we get the result.
The dual guarantee again follows in a very similar manner thanks to the guarantee for the dual SSC.
\end{proof}
}
If the function $f$ has $\curvf{f} = 1$,
\JTR{This was originally $\curvf{f} = 1$ but I changed it to $\curvf{f} = 0$,
and this occurred a few other places in the paper. Note that according
to the def of $f^{\text{ea}}$, we get this simpler form $\eaxa{f}{X}$
only when $\curvf{f} = 0$, but we should double check this, and please
lets talk about it. The thing I don't understand is that
if $\curvf{f} = 0$ then $f$ is modular and we don't need the EA approximation.
Also see the comment tagged 5XXXXX below.}\RTJ{Fixed above as well as earlier.}
we can use a much simpler
algorithm. In particular, \arxiv{since the ellipsoidal approximation is of
the form of $f^{\text{ea}}(X) = \eaxa{f}{X}$, }we can minimize $(f^{\text{ea}}(X))^2 =
w^f(X)$ at every iteration, giving a surrogate problem of the form
\arxiv{\begin{align}
\min\{ w^f(X) | g(X) \geq c\}.
\end{align}}
\notarxiv{$\min\{ w^f(X) | g(X) \geq c\}$.}
This is directly an instance of SSC, and in contrast to EASSC, we just need to solve SSC once. We call this algorithm EASSCc. \arxiv{This guarantee is tight up to $\log$ factors when $\curvf{f} = 1$.}
\begin{corollary}
\arxiv{The primal} EASSCc obtains an approximation guarantee of $O(\sqrt{n} \log n \sqrt{H_g})$. \arxiv{Similarly, the dual EASSCc obtains a bicriterion guarantee of $[O(\sqrt{n} \log n), 1 - e^{-1}]$.}
\end{corollary}
\arxiv{\begin{proof}
Let $\hat{X}$ be a set such that,
\begin{align}
w(\hat{X}) \geq H_g \min\{ w(X) | g(X) \geq c\}
\end{align}
Then denote $\alpha(n) = O(\sqrt{n}\log n)$.
\begin{align}
f(\hat{X}) \leq \alpha(n) \sqrt{w(\hat{X})} \leq \sqrt{H_g} \alpha(n) \sqrt{w(X^*)} \leq \sqrt{H_g} \alpha(n) f(X^*)
\end{align}
In the above, $X^* = \argmin\{ f(X) | g(X) \geq c\}$.
The result for the dual variant can also be similarly shown.
\end{proof}
}
\arxiv{
\begin{table}[tbh]
\small{
\begin{center}
\begin{tabular}{ |c|c|c|c| }
\hline
\multirow{2}{*}{Problem} & \multirow{2}{*}{Algorithm} & Approx. factor$^{*}$ & Hardness$^{*}$\\[1ex]
& & Bi-Criterion factor$^{\#}$ & (Bi-Criterion $\frac{\sigma}{\rho})$$^{\#}$ \\[1ex] \hline
\multirow{2}{*}{SSC} & Primal Greedy & $H(\max_j f(j))$$^{*}$ & $\log n$$^{*}$ \\[1ex]
& Dual Greedy & $(1 + \epsilon, 1 - \frac{1}{e})$$^{\#}$ & $1 - 1/e$$^{\#}$\\[1ex]
\hline
\multirow{4}{*}{SCSC} & Primal ISSC & $\frac{n H_g}{1 + (n - 1)(1 - \curvf{f})}$$^{*}$ & \multirow{4}{*}{$\Omega(\frac{\sqrt{n}}{1 + (\sqrt{n} - 1)(1 - \curvf{f})})$$^{*, \#}$} \\ [1.5ex]
& Dual ISSC & $[\frac{(1 + \epsilon)n}{1 + (n - 1)(1 - \curvf{f})}, 1 - 1/e]$$^{\#}$ & \\ [1.5ex]
& Primal EASSC & $O(\frac{\sqrt{n}\log n H_g}{1 + (\sqrt{n}\log n - 1)(1 - \curvf{f})})$$^{*}$ & \\ [1.5ex]
& Dual EASSC & $[O(\frac{\sqrt{n}\log n}{1 + (\sqrt{n}\log n - 1)(1 - \curvf{f})}, 1 - 1/e]$$^{\#}$ & \\ [1.5ex]
& Primal EASSCc & $O(\sqrt{n}\log n \sqrt{H_g})$$^{*}$ & \\ [1ex]
& Dual EASSCc & $[O(\sqrt{n}\log n), 1 - 1/e]$$^{\#}$ & \\ [1ex]
\hline
SK & Greedy & $1 - 1/e$$^{*}$ & $1 - 1/e$$^{*}$ \\ [1ex]
\hline
\multirow{3}{*}{SCSK} & Greedy & $1/n$$^{*}$ & \multirow{4}{*}{$\Omega(\frac{\sqrt{n}}{1 + (\sqrt{n} - 1)(1 - \curvf{f})})$$^{*, \#}$} \\ [1.5ex]
& ISK & $[1 - e^{-1}, \frac{K_f}{1 + (K_f - 1)(1 - \curvf{f})}]$$^{\#}$ & \\ [1.5ex]
& Primal EASK & $[1 + \epsilon, O(\frac{\sqrt{n}\log n H_g}{1 + (\sqrt{n}\log n - 1)(1 - \curvf{f})})]$$^{\#}$ & \\ [1.5ex]
& Dual EASK & $[(1 - 1/e)(1 + \epsilon), O(\frac{\sqrt{n}\log n}{1 + (\sqrt{n}\log n - 1)(1 - \curvf{f})})]$$^{\#}$ & \\ [1.5ex]
& EASKc & $[1 - 1/e, O(\sqrt{n}\log n)]$$^{\#}$ & \\ [1ex]
\hline
\end{tabular}
\end{center}
\caption{Worst case approximation factors, hardness for Problems 1
and 2 and their special cases.} }
\end{table}}
\normalsize
\subsection{Approximation Algorithms for SCSK}
\label{sec:appr-algor-scsk}
In this section, we describe our approximation algorithms for SCSK. We
note the dual nature of the algorithms in this current section to those
given in \S\ref{sec:appr-algor-scsc}. We first investigate a special case, the submodular knapsack (SK), and then provide three algorithms, two of them (Gr and ISK) being practical with slightly weaker theoretical guarantee, and another one (EASK) which is not scalable but has the tightest guarantee.\looseness-1
\textbf{Submodular Cost Knapsack (SK): }
We start with a special case of SCSK (Problem 2), where $f$ is a modular
function and $g$ is a submodular function. In this case, SCSK turns
into the SK problem for which the greedy
algorithm with partial enumeration provides a $1 - e^{-1}$
approximation~\cite{sviridenko2004note}. The greedy algorithm can be
seen as an instance of Algorithm~\ref{alg:framework} with $\hat{g}$
being the modular lower bound of $g$ and $\hat{f}$ being $f$, which is
already modular. \arxiv{In particular, we then get back the framework
of~\cite{rkiyersemiframework2013}, where the authors show that
choosing a permutation based on a greedy ordering, exactly analogous
to Eqn.~\eqref{greedyssc}, provides the bounds.} In particular,
define:
\begin{align}
\label{greedyknapsack}
\ensuremath{\pi}(i) \in
\argmax \left\{\frac{g(j | S^{\ensuremath{\pi}}_{i-1})}{f(j)} \ \bigg\vert \ {j \notin S^{\ensuremath{\pi}}_{i-1},
f(S^{\ensuremath{\pi}}_{i-1} \cup \{j\}) \leq b}\right\},
\end{align}
where the remaining elements are chosen arbitrarily. \arxiv{A slight
catch however is that for the analysis to work, \cite{sviridenko2004note} needs to
consider $n \choose 3$ instances of such orderings (partial
enumeration), chosen by fixing the first three elements in the
permutation~\cite{rkiyersemiframework2013}. We can however just choose
the simple greedy ordering in one stage, to get a slightly worse
approximation factor of $1 -
e^{-1/2}$~\cite{rkiyersemiframework2013, linbudget,krause05note}.
} \arxivalt{
\begin{lemma}\label{greedybudget}~\cite{rkiyersemiframework2013}
Choosing the surrogate function $\hat{f}$ as $f$ and $\hat{g}$ as $h^{\ensuremath{\pi}}$ in Algorithm~\ref{alg:framework} yields a set $X^g$:
\begin{equation}
\label{eq:8}
\max\{\max_{i: f(i) \leq b} g(i), g(X^g)\} \geq 1/2(1 - 1/e)g(X^*).
\end{equation}
Let $\ensuremath{\pi}^{ijk}$ be a permutation with $i,j,k$ in the first three
positions, and the remaining arrangement greedy. Running $O(n^3)$
restarts of one iteration of Algorithm~\ref{alg:framework} yields sets
$X_{ijk}$ with
\begin{equation}
\label{eq:9}
\max_{i,j,k \in V} f(X_{ijk}) \,\geq (1-1/e) f(X^*).
\end{equation}
\end{lemma}}
The following is an informal description of the result described formally in~\cite{nipsextendedvsubcons}.
\begin{lemma}
Choosing the surrogate function $\hat{f}$ as $f$ and $\hat{g}$ as $h^{\ensuremath{\pi}}$ (with $\ensuremath{\pi}$ defined in eqn~\eqref{greedyknapsack}) in Algorithm~\ref{alg:framework} with appropriate initialization obtains a guarantee of $1 - 1/e$ for SK.
\end{lemma}
}
\textbf{Greedy (Gr): }
A similar greedy algorithm can provide approximation guarantees for
the general SCSK problem, with submodular $f$ and $g$. Unlike the
knapsack case in~\eqref{greedyknapsack}, however, at iteration $i$
we choose an element $j \notin S_{i-1} :
f(S^{\ensuremath{\pi}}_{i-1} \cup \{j\}) \leq b$ which maximizes $g(j |
S_{i-1})$. In terms of Algorithm~\ref{alg:framework}, this is
analogous to choosing a permutation, $\ensuremath{\pi}$ such that:
\begin{align}
\ensuremath{\pi}(i) \in \argmax \{g(j | S^{\ensuremath{\pi}}_{i-1}) | {j \notin S^{\ensuremath{\pi}}_{i-1}, f(S^{\ensuremath{\pi}}_{i-1} \cup \{j\}) \leq b}\}.
\end{align}
\begin{theorem}\label{GrSCSK}
The greedy algorithm for SCSK obtains an approx. factor of $\frac{1}{\kappa_g} (1 - (\frac{K_f - \kappa_g}{K_f})^{k_f}) \geq \frac{1}{K_f}$, where $K_f = \max\{|X|: f(X) \leq b\}$ and $k_f = \min\{|X|: f(X) \leq b \ \& \ \forall j \in X, f(X \cup j) > b\}$.
\end{theorem}
\arxiv{
\begin{proof}
The proof of this result follows directly from~\cite{conforti1984submodular, rkiyersemiframework2013}. In particular, it holds for any down monotone constraint. It is easy to see that the constraint $\{f(X) \leq b\}$ is down-monotone when $f$ is a monotone submodular function.
\end{proof}
}
In the worst case, $k_f = 1$ and $K_f = n$, in which case the
guarantee is $1/n$. The bound above follows from a simple observation
that the constraint $\{f(X) \leq b\}$ is down-monotone for a monotone
function $f$. However, in this variant, we do not use any specific
information about $f$. In particular it holds for maximizing a
submodular function $g$ over any down monotone
constraint~\cite{conforti1984submodular}. Hence it is
conceivable that an algorithm that uses both $f$ and $g$ to choose the
next element could provide better bounds. We do not, however, currently have
the analysis for this.\looseness-1
\textbf{Iterated Submodular Cost Knapsack (ISK): } Here, we
choose $\hat{f_t}(X)$ as a modular upper bound of $f$, tight at $X^t$.
Let $\hat{g_t} = g$. Then at every iteration, we solve\arxivalt{:
\begin{align}
\max\{ g(X) | m^f_{X^t}(X) \leq b\},
\end{align}}{ $\max\{ g(X) | m^f_{X^t}(X) \leq b\}$,}
which is a submodular maximization problem subject to a knapsack
constraint (SK). As mentioned above, greedy can solve this nearly optimally. We start with $X^0 = \emptyset$, choose $\hat{f_0}(X) =
\sum_{j \in X} f(j)$ and then iteratively continue this process until
convergence (note that this is an ascent algorithm). We
have the following theoretical guarantee:
\begin{theorem}\label{ISKapprox}
Algorithm ISK obtains a set $X^t$ such that $g(X^t) \geq (1 - e^{-1})
g(\tilde{X})$, where $\tilde{X}$ is the optimal solution of $\max\left\{g(X) \ | \ f(X) \leq \frac{b (1 + (K_f - 1)(1 -
\curvf{f})}{K_f}\right\}$ and where $K_f = \max\{|X|: f(X) \leq b\}$.\looseness-1
\end{theorem}
\arxiv{\begin{proof}
We are given that $\tilde{X}$ is the optimal solution to the problem:
\begin{align}
\max\left\{g(X) \ | \ f(X) \leq \frac{b (1 + (K_f - 1)(1 - \curvf{f})}{K_f}\right\}
\end{align}
$\tilde{X}$ is also a feasible solution to the problem:
\begin{align}\label{subprobKI}
\max\left\{g(X) \ | \ \sum_{j \in X} f(j) \leq b \right\}
\end{align}
The reason for this is that:
\begin{align}
\sum_{j \in X} f(j) \leq \frac{K_f}{1 + (K_f - 1)(1 - \curvf{f})} f(X) \leq \frac{K_f}{1 + (K_f - 1)(1 - \curvf{f})}\frac{b (1 + (K_f - 1)(1 - \curvf{f}))}{K_f} \leq b \nonumber
\end{align}
\JTR{fix bugs in the above equation, both the missing parentheses and also
the divide/multiply by $b$.}\RTJ{done} Now, at the first iteration we are
guaranteed to find a set $X^1$ such that $g(X^1) \geq (1 -
1/e)g(\tilde{X})$. The further iterations will only improve the
objective since this is a ascent algorithm.
\end{proof}}
It is worth pointing out that the above bound holds even after the
first iteration of the algorithm. It is interesting to note the similarity
between this approach and \arxiv{the iterated submodular set cover}
ISSC. \notarxiv{Notice that the guarantee above is not a standard bi-criterion approximation. We show in the extended version~\cite{nipsextendedvsubcons} that with a simple transformation, we can obtain a bicriterion guarantee.}\arxiv{Notice that the bound above is a form of a bicriterion approximation factor of type-II. In particular, we obtain a feasible solution at the end of the algorithm. We can obtain a type-I bicriterion approximation bound, by running this for a larger budget constraint. In particular, we run the above algorithm with a budget constraint of $\frac{b K_f}{1 + (K_f - 1)(1 - \curvf{f})}$ instead of $b$. The following guarantee then follows.
\begin{lemma}
The ISK algorithm of type-I is guaranteed to obtain a set $X$ which has a bicriterion approximation factor of $[1 - e^{-1}, \frac{K_f}{1 + (K_f - 1)(1 - \curvf{f})}]$, where $K_f = \max\{|X|: f(X) \leq b\}$.
\end{lemma}
\begin{proof}
This result directly follows from the Lemma above, and the transformation from a type-II approximation algorithm to a type-I one.
\end{proof}
}
\textbf{Ellipsoidal Approximation based Submodular Cost Knapsack (EASK): }
Choosing the Ellipsoidal Approximation $f^{ea}$ of $f$ as a surrogate function, we obtain a simpler problem:
\begin{align}\label{EASKeq}
\max\left\{ g(X) \ \bigg\vert \ \curvf{f} \sqrt{w^{f^{\kappa}}(X)} + (1 - \curvf{f}) \sum_{j \in X} f(j) \leq b\right\}.
\end{align}
In order to solve this problem, we look at its dual
problem (i.e., Eqn.~\eqref{EASSCeq}) and use Algorithm~\ref{alg:alg1} to convert
the guarantees. \arxivalt{Recall that
Eqn.~\eqref{EASSCeq} itself admits two variants of the ellipsoidal approximation,
which we had referred to as the primal EASSC and the dual EASSC. Hence, we call the combined algorithms, primal and dual EASK, which admit the following guarantees:}
We call this procedure EASK. We then obtain guarantees very similar to Theorem~\ref{EAguar}.}
\begin{lemma}
\arxiv{The primal} EASK obtains a guarantee of $\left[1 +
\epsilon, O(\frac{\sqrt{n}\log n H_g}{ 1 + (\sqrt{n}\log n - 1)(1 -
\curvf{f})})\right]$. \arxiv{Similarly, the dual EASK obtains a
guarantee of $\left[(1 + \epsilon)(1 - 1/e), O(\frac{\sqrt{n}\log n}{ 1 +
(\sqrt{n}\log n - 1)(1 - \curvf{f})})\right]$.}
\end{lemma}
\arxiv{\begin{proof}This Lemma directly follows from
Theorem~\ref{thm1} and
Theorem~\ref{EAguar}. \end{proof}
These factors are akin to
those of Theorem~\ref{EAguar}, except for the additional factor of $1 +
\epsilon$. Unlike the greedy algorithm, ISSC and ISK, note that both EASK and EASSC are
enormously costly and complicated algorithms. Hence, it also seems at first
thought that EASK would need to run multiple versions of EASSC at each conversion round of Algorithm~\ref{alg:alg1}. Fortunately, however, we need to compute the
Ellipsoidal Approximation just once and the algorithm can reuse it for
different values of $c$. Furthermore, since construction
of the approximation is often the bottleneck, this scheme is likely to be as costly
as EASSC in practice.\looseness-1
}
In the case when the submodular function has a curvature $\curvf{f} =
1$, we can actually provide a simpler algorithm without needing to
use the conversion algorithm (Algorithm~\ref{alg:alg1}). In this case, we
can directly choose the ellipsoidal approximation of $f$ as
$\sqrt{w^f(X)}$ and solve the surrogate problem:\arxivalt{
\begin{align}
\label{EAKeqn}
\max\{g(X) : w^f(X) \leq b^2\}.
\end{align}}{ $\max\{g(X) : w^f(X) \leq b^2\}.$}
This surrogate problem is a submodular cost knapsack problem, which we can
solve using the greedy algorithm. We call this algorithm EASKc. This guarantee is tight up to $\log$ factors if $\curvf{f} = 1$.
\begin{corollary}
Algorithm EASKc obtains a bi-criterion guarantee of $[1 - e^{-1}, O(\sqrt{n}\log n)]$.
\end{corollary}
\arxiv{\begin{proof}
Let $\hat{X}$ be the approximate optimizer of
Eqn.~\eqref{EAKeqn}. First we show that $g(\hat{X}) \geq (1 -
1/e) g(X^*)$ where $X^*$ is the optimizer of problem 2. Notice
that $f(X^*) \leq b$ since it feasible in problem 2. Also,
$\sqrt{w^f(X^*)} \leq f(X^*) \leq b$ and hence $X^*$ is feasible
in Eqn.~\eqref{EAKeqn}. Hence it holds that $g(\hat{X}) \geq (1 -
1/e) g(X^*)$.
We now show that $f(\hat{X}) \leq b\sqrt{n}\log n$. Notice that,
\begin{align}
f(\hat{X}) \leq \sqrt{n} \log n \sqrt{w^f(\hat{X})} \leq b \sqrt{n} \log n
\end{align}
\end{proof}}
\subsection{Extensions beyond SCSC and SCSK}\label{extentions}
\arxivalt{SCSC is in fact more general and can be extended
to more flexible and complicated constraints which can arise naturally in many
applications~\cite{krause08robust, guillory2011simultaneous}. Notice
first that
\begin{align}
\{g(X) \geq \alpha\} \Leftrightarrow \{g^{\prime}(X) = g^{\prime}(V)\}
\end{align}
where $g^{\prime}(X) = \min\{g(X), \alpha\}$. We can also have ``and'' constraints as $g_1(X) = g_1(V)$ and $g_2(X) = g_2(V)$. These have a simple equivalence:
\begin{align}
\{g_1(X) = g_1(V) \wedge g_2(X) = g_2(V)\} \Leftrightarrow \{g(X) = g(V)\}
\end{align}
when $g(X) = g_1(X) + g_2(X)$~\cite{krause08robust}. Moreover, we can also handle $k$ `and' constraints, by defining $g(X) = \sum_{i = 1}^k g_i(X)$.
Similarly we can have ``or'' constraints, i.e., $g_1(X) = g_1(V)$ or $g_2(X) = g_2(V)$. These also
have a nice equivalence:
\begin{align}
\{g_1(X) = g_1(V) \vee g_2(X) = g_2(V)\} \Leftrightarrow \{g(X) = g(V)\}
\end{align}
by defining $g(X) = g_1(X)g_2(V) + g_2(X)g_1(V) -
g_1(X)g_2(X)$~\cite{guillory2011simultaneous}. We can also extend these recursively to multiple `or' constraints. Hence our algorithms
can directly solve all these variants of SCSC.
SCSK can also be extended to handle more complicated forms
of functions $g$. In particular, consider the function
\begin{align}
g(X)= \min\{g_1(X), g_2(X), \dots, g_k(X)\}
\end{align}
where the functions $g_1, g_2,
\dots, g_k$ are submodular. Although $g(X)$ in this case is not
submodular, this scenario occurs naturally in many applications,
particularly sensor placement~\cite{krause08robust}. The problem in~\cite{krause08robust} is in fact a special case of Problem 2, using a modular function $f$. Often, however,
the budget functions involve a cooperative cost, in which case $f$ is submodular. Using Algorithm~\ref{alg:alg1},
however, we can easily solve this by iteratively solving the dual
problem. Notice that the dual problem is in the form of Problem 1 with a non-submodular
constraint $g(X) \geq c$. It is easy to see that this is equivalent to
the constraint $g_i(X) \geq c, \forall i = 1, 2, \dots, k$, which can be solved thanks to the techniques above. \looseness-1
We can also handle multiple constraints in SCSK. In particular, consider multiple `and' constraints -- $\{f_i(X) \leq b_i, i = 1, 2, \cdots, k\}$, for monotone submodular functions $f_i$ and $g$. A first observation is that the greedy algorithm can almost directly extended to these cases, since we do not use any specific property of the constraints, while providing the approximation guarantees. Hence Theorem~\ref{GrSCSK} can directly be extended to this case. We can also provide bi-criterion approximation guarantees with `and' constraints. The approximate feasibility for a $[\rho, \sigma]$ bi-criterion approximation in this setting would be to have a set $X$ such that $f_i(X) \leq \sigma b_i$. Algorithm ISK can easily then be used in this scenario, and at every iteration we would solve a monotone submodular maximization problem subject to multiple linear (or knapsack constraints). Surprisingly this problem also has a constant factor ($1 - 1/e$) approximation guarantee~\cite{kulik2009maximizing}. Hence we can retain the same approximation guarantees as in Theorem~\ref{ISKapprox}. We can also use algorithm EASKc, and obtain a curvature-independent approximation bound for this problem. The reason for this is the Ellipsoidal Approximation is of the form $\sqrt{w^{f_i}(X)}$, for each $i$ and squaring it will lead to knapsack constraints. If we add the curvature terms (i.e., try to implement EASK in this setting), we obtain a much more complicated class of constraints, which we do not currently know how to handle.
We can also extend SCSC and SCSK to non-monotone submodular functions. In particular, recall that the submodular knapsack has constant factor approximation guarantees even when $g$ is non-monotone submodular~\cite{fiege2011submodmax}. Hence, we can obtain bi-criterion approximation guarantees for the Submodular Set Cover (SSC) problem, by solving the Submodular Knapsack (SK) problem multiple times. We can similarly do SCSC and SCSK when $f$ is monotone submodular and $g$ is non-monotone submodular, by extending ISSC, EASSC, ISK and EASK (note that in all these cases, we need to solve a submodular set cover or submodular knapsack problem with a non-monotone $g$). These algorithms, however, do not extend if $f$ is non-monotone, and we do not currently know how to implement these.
}{SCSC and SCSK can in fact be extended
to more flexible and complicated constraints which can arise naturally in many
applications~\cite{krause08robust, guillory2011simultaneous}. These include multiple covering and knapsack constraints -- i.e., $\min\{f(X) | g_i(X) \geq c_i, i = 1, 2, \cdots k\}$ and $\max\{g(X) | f_i(X) \leq b_i, i = 1, 2, \cdots k\}$, and robust optimization problems like $\max\{\min_i g_i(X) | f(X) \leq b\}$, where the functions $f, g, f_i$'s and $g_i$'s are submodular. We also consider SCSC and SCSK with non-monotone submodular functions. Due to lack of space, we defer these discussions to the extended version of this paper~\cite{nipsextendedvsubcons}.}
\subsection{Hardness}
\label{sec:hardness}
In this section, we provide the hardness for Problems 1 and 2. The
lower bounds serve to show that the approximation factors above are
almost tight.
\begin{theorem}
For any $\kappa > 0$, there exists submodular functions with curvature $\kappa$ such that no polynomial time algorithm for Problems 1 and 2
achieves a bi-criterion factor better than
$\frac{\sigma}{\rho} = \frac{n^{1/2 - \epsilon}}{1 + (n^{1/2 -
\epsilon} - 1)(1 - \kappa)}$ for any $\epsilon > 0$. \looseness-1
\end{theorem}
\arxiv{
\begin{proof}
We prove this result using the hardness construction
from~\cite{goemans2009approximating, svitkina2008submodular}. The
main idea of their proof technique is to construct two submodular
functions $f(X)$ and $f_R(X)$ that with high probability are
indistinguishable. Thus, also with high probability, no algorithm
can distinguish between the two functions and the gap in their
values provides a lower bound on the approximation. We shall see
that this lower bound in fact matches the approximation factors up
to log factors and hence this is the hardness of the problem.
Define two monotone submodular functions $f(X) = \curvf{f}\min\{|X|,
\alpha\} + (1 - \curvf{f})|X|$ and $f_R(X) = \curvf{f}\min\{\beta +
|X \cap \bar{R}|, |X|, \alpha\} + (1 - \curvf{f}) |X|$, where $R
\subseteq V$ is a random set of cardinality $\alpha$. Let $\alpha$
and $\beta$ be an integer such that $\alpha = x\sqrt{n}/5$ and
$\beta = x^2/5$ for an $x^2 = \omega(\log n)$. Both $f$ and $f_R$
have curvature $\curvf{f}$. We also assume a very simple
function $g(X) = |X|$. Given an arbitrary $\epsilon>0$, set $x^2 =
n^{2\epsilon} = \omega(\log n)$. Then the ratio between
$f_R^{\curvf{f}}(R)$ and $g^{\curvf{f}}(R)$ is
$\frac{n^{1/2-\epsilon}}{1 + (n^{1/2-\epsilon}-1)(1 -
\curvf{f})}$. A Chernoff bound analysis very similar
to~\cite{svitkina2008submodular} reveals that any algorithm that
uses a polynomial number of queries can distinguish $h^{\kappa}$ and
$f^{\kappa}_R$ with probability only $n^{-\omega(1)}$, and therefore
cannot reliably distinguish the functions with a polynomial number
of queries.
We first prove the first part of this theorem (i.e., for SCSC). In
this case, we consider the problem
\begin{align}\label{probststproof}
\min\{h(X) | |X| \geq \alpha\}
\end{align}
with $h$ chosen as $f$ and $f_R$ respectively. It is easy to see that
$R$ is the optimal set in both cases. However the ratio between the
two is
\begin{align}
\gamma(n) = \frac{n^{1/2-\epsilon}}{1 + (n^{1/2-\epsilon}-1)(1 - \curvf{f})}
\end{align}
Now suppose there exists an algorithm which is guaranteed to obtain an
approximation factor better than $\gamma(n)$. Then by running this
algorithm, we are guaranteed to find different answers for
Eqn~\eqref{probststproof} above. This must imply that this algorithm
can distinguish between $f_R$ and $g$ which is a contradiction. Hence
no algorithm for Problem 1 can obtain an approximation guarantee
$\frac{n^{1/2 - \epsilon}}{1 + (n^{1/2 - \epsilon} - 1)(1 -
\curvf{f})}$.
In order to extend the result to the bi-criterion case, we need to
show that no bi-criterion approximation algorithm can obtain a factor
better than $\frac{\sigma}{\rho} = \gamma(n)$. Assume there exists a
bi-criterion algorithm with a factor $[\sigma, \rho]$ such that:
\begin{align}
\frac{\sigma}{\rho} < \gamma(n) = \frac{n^{1/2-\epsilon}}{1 + (n^{1/2-\epsilon}-1)(1 - \curvf{f})}
\end{align}
Then, we are guaranteed to obtain a set $S$ in
equation~\eqref{probststproof} such that $h(S) \leq \sigma OPT$ and
$|S| \geq \rho \alpha$, where $\sigma \geq 1$ and $\rho \leq 1$. Now
we run this algorithm with $h = f$ and $h = f_R$ respectively. With $h
= f$, it is easy to see that the algorithm obtains a set $S_1$ such
that $f(S_1) \leq \sigma \alpha$ and $|S_1| \geq \rho
\alpha$. Similarly, with $h = f_R$, the algorithm finds a set $S_2$
such that $f(S_2) \leq \sigma (\kappa \beta + (1 - \kappa)\alpha)$ and
$|S_2| \geq \rho \alpha$. Since $f_R$ and $f$ are indistinguishable,
$S_1 = S_2 = S$.
We first assume that $|S| < \alpha$. Then $f(S) = |S| \geq \rho \alpha$. We then have that,
\begin{align}
f_R(S) &\leq \sigma (\kappa \beta + (1 - \kappa)\alpha \\
&\leq \sigma n^{2\epsilon} (1 + (n^{1/2 - \epsilon} - 1)(1 - \curvf{f}) \\
&< \rho \gamma(n) n^{2\epsilon} (1 + (n^{1/2 - \epsilon} - 1)(1 - \curvf{f}) \\
&< \rho n^{1/2 + \epsilon} \\
&< \rho \alpha
\end{align}
This implies that the algorithm can distinguish between $f_R$ and $f$
which is a contradiction. Now consider the second case when $|S| \geq
\alpha$. In this case, $f(S) = \alpha$. Again the chain of
inequalities above shows that $f_R(S) < \rho \alpha < \alpha$ since
$\rho \leq 1$. This again implies that the algorithm can distinguish
between $f_R$ and $f$ which is a contradiction. Hence no bi-criterion
approximation algorithm can obtain a factor better than $\gamma(n)$.
To show the second part (for SCSK), we can simply invoke Theorem~\ref{thm1} and argue that any $[\rho, \sigma]$ bi-criterion approximation algorithm for problem 2 with
\begin{align}
\frac{\sigma}{\rho} < \frac{\gamma(n)}{1 + \epsilon}
\end{align}
can be used in algorithm 3, to provide a $[\sigma, \rho]$ bi-criterion algorithm with $\frac{\sigma}{\rho} < \gamma(n)$. This contradicts the first part above and hence the same hardness applies to problem 2 as well.
\end{proof}
} The above result shows that EASSC and EASK meet the bounds above to
log factors. We see an interesting curvature-dependent influence
on the hardness. \arxivalt{We also see from our approximation guarantees
also that the curvature of $f$ plays a more influential role than
the curvature of $g$ on the approximation quality.}{We also see this
phenomenon in the approximation guarantees of our algorithms.} In
particular, as soon as $f$ becomes modular, the problem becomes easy,
even when $g$ is submodular. This is not surprising since the
submodular set cover problem and the submodular cost knapsack problem
both have constant factor guarantees.\looseness-1
\arxiv{\begin{figure}[t]
\centering
\hspace{-10pt}
\includegraphics[width=0.45\textwidth]{Sim2r.pdf}\hspace{-10pt}
~
\includegraphics[width=0.45\textwidth]{Sim3r.pdf} \hspace{-10pt}
\caption{The two figures show the performance of the algorithms in the text on
Problems 1 and 2 (in both cases, higher the better).
\JTR{change figure legend, I changed MI to ISSC. Also EAk should be EAK,
and KI needs to be changed as well.}
}
\label{fig:sim}
\end{figure}}
\section{Experiments}
\label{sec:experiments}
In this section, we empirically compare the performance of the various
algorithms discussed in this paper. We are motivated by the speech
data subset selection application\arxivalt{~\cite{lin2009select,lin11,jegelkanips}}{~\cite{lin2009select,lin11}} with the submodular function $f$ encouraging limited
vocabulary while $g$ tries to achieve acoustic variability. A natural
choice of the function $f$ is a function of the form
$|\Gamma(X)|$, where $\Gamma(X)$ is the neighborhood function
on a bipartite graph constructed between the utterances and the words~\cite{lin11}. For the coverage function
$g$, we use two types of coverage: one is a facility
location function \arxivalt{
\begin{align}
g_1(X) = \sum_{i \in V} \max_{j \in X} s_{ij}
\end{align}
}{$g_1(X) = \sum_{i \in V} \max_{j \in X} s_{ij}$} while the other is a saturated sum function
\arxivalt{
\begin{align}
g_2(X) = \sum_{i \in V}\min\{\sum_{j \in X} s_{ij}, \alpha \sum_{j \in V} s_{ij}\}.
\end{align}
}{$g_2(X) = \sum_{i \in V}\min\{\sum_{j \in X} s_{ij}, \alpha \sum_{j \in V} s_{ij}\}$.} Both these functions are defined in terms of a similarity matrix
$\mathbf S = \{s_{ij}\}_{i, j \in V}$, which we define on the TIMIT corpus~\cite{timit}, using the string kernel metric~\cite{rousu2006efficient} for similarity. Since some of our algorithms, like the Ellipsoidal Approximations, are computationally intensive, we restrict ourselves to $50$ utterances.
\narxiv{
\captionsetup[figure]{font=small,skip=0pt}
\begin{wrapfigure}[8]{r}{0.5\textwidth}
\vspace{-2.5ex}
\begin{minipage}{0.5\textwidth}
\centering
\includegraphics[width=0.45\textwidth]{Sim2r.pdf}\hspace{-10pt}
~
\includegraphics[width=0.45\textwidth]{Sim3r.pdf} \hspace{-10pt}
\end{minipage}
\caption{\small{Comparison of the algorithms in the text.}}
\label{fig:sim}
\end{wrapfigure}}
We compare our different algorithms on Problems 1 and 2 with $f$
being the bipartite neighborhood and $g$ being the facility location
and saturated sum respectively. \arxiv{Since the primal and dual variants of the submodular set cover problem are similar, we just
use the primal variants of ISSC and
EASSC.} Furthermore, in our
experiments, we observe that the neighborhood function $f$ has a
curvature $\curvf{f} = 1$.
Thus, it suffices to use the simpler versions of algorithm EA (i.e.,
algorithm EASSCc and EASKc). The results are shown in
Figure~\ref{fig:sim}. We observe that on the real-world instances, all
our algorithms perform almost comparably. This implies, moreover, that
the iterative variants, viz. Gr, ISSC and ISK, perform comparably to
the more complicated EA-based ones, although EASSC and EASK have
better theoretical guarantees. We also compare against a baseline of
selecting random sets (of varying cardinality), and we see that our
algorithms all perform much better. In terms of the running time,
computing the Ellipsoidal Approximation for $|\Gamma(X)|$ with $|V| =
50$ takes about $5$ hours while all the iterative variants (i.e., Gr,
ISSC and ISK) take less than a second. This difference is much more
prominent on larger instances (for example $|V| = 500$). \looseness-1
\JTR{There needs to be an additional very simple discussion here on
one-stop-shopping: i.e., you have a problem like Problem 1 or
Problem 2. Based on the theoretical and empirical results we've got,
state what we recommend for each of Problems 1 and 2.
Looks like Gprimal for problem 1 and ISSC for problem 2, but maybe
a bit more empirical results could strengthen this recommendation.}\RTJ{Yes, I will be rerunning the experiments and will modify this then.}
\section{Discussions\arxiv{ and related work}}
In this paper, we propose a unifying framework for problems 1 and 2 based on suitable surrogate functions. We provide a number of iterative algorithms which are very practical and scalable (like Gr, ISK and ISSC), and also algorithms like EASSC and EASK, which though more intensive, obtain tight approximation bounds. Finally, we empirically compare our algorithms, and show that the iterative algorithms compete empirically with the more complicated and theoretically better approximation algorithms.\arxiv{
To our
knowledge, this paper provides the first general framework of approximation
algorithms for Problems 1 and 2 for monotone submodular functions $f$ and $g$. A number of papers, however, investigate related problems and approaches. For example,~\cite{fujishige2005submodular} investigates an exact algorithm for solving problem 1, with equality instead of inequality. However, since problem 1 subsumes the problem of minimizing a monotone submodular function subject to a cardinality equality constraint, and is hence NP hard~\cite{nagano2011}. Hence this algorithm in the worst case, must be exponential. Furthermore, a similar problem was considered in~\cite{krause06near} with one specific instance of a function $f$, which is not submodular. They also use, a considerably different algorithm. Also, an algorithm equivalent to the first iteration of ISSC was proposed in~\cite{wan2010greedy, du2011minimum} and ISSC not only generalizes this, but we also provide a more explicit approximation guarantee (we provide an elaborate discussion on this in the section describing ISSC). We also point out that, a special case of SCSK was considered in~\cite{lin2012submodularity}, with $f$ being submodular, and $g$ modular (we called this the \emph{submodular span problem}). The authors there use an algorithm very similar to Algorithm~\ref{alg:alg1}, to convert this problem into an instance of minimizing a submodular function subject to a knapsack constraint, for which they use the algorithm of~\cite{svitkina2008submodular}. Unfortunately, the algorithm of~\cite{svitkina2008submodular} does not scale very well. Our algorithms for this problem, on the other hand, would continue to scale very well in practice.
Similarly, a number of approximation algorithms
have been shown for Problem 0~\cite{rkiyeruai2012, narasimhanbilmes,
kawahara2011prismatic}. The algorithms in~\cite{rkiyeruai2012,
narasimhanbilmes} are scalable and practical, but lack theoretical
guarantees. The algorithm of~\cite{kawahara2011prismatic} though exact, employs a branch and bound technique which is often inefficient in
practice (the timing analysis from~\cite{kawahara2011prismatic} also
depicts that). These facts are not surprising, since problem 0 is not only NP hard but also inapproximable. Moreover, these algorithms are not comparable to ours,
since we directly model the hard constraints and our bicriteria
results give worst-case bounds on the deviation from the constraints
and the optimal solution. This is often important, since there are hard constraints in many practical applications (in the form of power constraints, or budget constraints). Casting it as Problem 0, however, no longer
has guarantees on the deviation from the constraints.
} For future work, we would like to empirically evaluate our algorithms on many of the real world problems described above, particularly the limited vocabulary data subset selection application for speech corpora, and the machine translation application.
{\bf Acknowledgments:} Special thanks to Kai Wei and Stefanie Jegelka for discussions, to Bethany Herwaldt for going through an early draft of this manuscript and to the anonymous reviewers for useful reviews. This material is based upon work supported by the National Science
Foundation under Grant No. (IIS-1162606), a
Google and a Microsoft award, and
by the Intel Science and
Technology Center for Pervasive Computing.\looseness-1
\notarxiv{\small}
\bibliographystyle{abbrv}
|
\section{Introduction}
\subsection{Basic notation} Let
\[
\hDelta:=\frac14\bigg(\frac{\partial^2}{\partial x^2}+
\frac{\partial^2}{\partial y^2}\bigg),\qquad\diff A(z):=\diff x\diff y,
\]
denote the normalized Laplacian and the area element, respectively.
Here, $z=x+\imag y$
is the standard decomposition into real and imaginary parts. We let $\C$ denote
the complex plane. We also need the standard complex differential operators
\[
\dbar_z:=\frac{1}{2}\bigg(\frac{\partial}{\partial x}+\imag
\frac{\partial}{\partial y}\bigg),\quad \partial_z:=
\frac{1}{2}\bigg(\frac{\partial}{\partial x}-\imag
\frac{\partial}{\partial y}\bigg),
\]
so that $\hDelta$ factors as $\hDelta=\partial_z\dbar_z$. We sometimes drop
indication of the differentiation variable $z$.
\subsection{H\"ormander's solution of the $\bar\partial$-problem}
We present H\"ormander's theorem \cite{Horm1}, \cite{Horm2} in the simplest
possible case, when the domain is the entire complex plane and the weight
$\phi:\C\to\R$ is $C^{2}$-smooth with $\hDelta\phi>0$ everywhere.
\begin{thm}
{\rm(H\"ormander)} If the complex-valued function $f$ is locally area
$L^2$-integrable in the plane $\C$, then there exists a solution to the
$\dbar$-equation $\dbar u=f$ with
\[
\int_\C|u|^2\e^{-2\phi}\diff A\le\frac12
\int_\C|f|^2\frac{\e^{-2\phi}}{\hDelta\phi}\diff A.
\]
\label{thm-H1}
\end{thm}
Here, we remark that the assertion of the theorem is void unless the integral
on the right hand side is finite.
\subsection{The $\dbar$-equation with growing weights}
While Theorem \ref{thm-H1} essentially deals with decaying weights
$\e^{-2\phi}$, it is natural to ask what happens if we were to consider the
growing weights $\e^{2\phi}$ in place of $\e^{-2\phi}$.
So when could we say that there exists a solution to $\dbar u=f$ with
\begin{equation}
\int_\C|u|^2\e^{2\phi}\diff A\lesssim
\int_\C|f|^2\e^{2\phi}\diff A,
\label{eq-adjest1}
\end{equation}
where the symbol ``$\lesssim$'' is understood liberally? Here, we should have
the (Fock weight) example $\phi(z)=\frac12|z|^2$ in mind.
It is rather clear that we
cannot hope to have an estimate of the type \eqref{eq-adjest1} without an
additional condition on on the datum $f$. For instance, if
$\phi(z)=\frac12|z|^2$ and $f(z)=\e^{-|z|^2}$, it is not possible to find
such a fast-decaying function $u$ with $\dbar u=f$ (cf. Section
\ref{sec-constr} below).
It is natural to look for a class of data $f$ that would
come from functions $u$ with compact support. Let $C^\infty_c(\C)$ denote the
standard space of infinitely differentiable compactly supported test functions.
The calculation
\[
\int_\C fg\diff A=\int_\C g\dbar u\diff A=-\int_\C u\dbar g\diff A,
\qquad u\in C^\infty_c(\C),
\]
shows that the datum $f=\dbar u$ with $u\in C^\infty_c(\C)$ must satisfy,
for entire functions $g$,
\begin{equation}
\int_\C fg\diff A=0.
\label{eq-dual1}
\end{equation}
We remark here that the calculation \eqref{eq-dual1} is the basis for what is
known as Havin's lemma \cite{Hav} (see also, e.g., \cite{Hed1}).
We let $L^2(\C,\e^{-2\phi})$ and $L^2(\C,\e^{2\phi})$ be the weighted area
$L^2$-spaces with the indicated weights. The corresponding norms are
\[
\|g\|^2_{L^2(\C,\e^{-2\phi})}=\int_\C|g|^2\e^{-2\phi}\diff A,\quad
\|g\|^2_{L^2(\C,\e^{2\phi})}=\int_\C|g|^2\e^{2\phi}\diff A.
\]
The space of entire functions in $L^2(\C,\e^{-2\phi})$ is denoted by
$A^2(\C,\e^{-2\phi})$. We recall the standing assumption that $\phi$ be
$C^2$-smooth with $\Delta\phi>0$ everywhere.
\begin{thm}
Suppose $f\in L^2(\C,\e^{2\phi})$ meets the condition \eqref{eq-dual1} for all
$g\in A^2(\C,\e^{-2\phi})$. Then there exists a solution to the
$\dbar$-equation $\dbar u=f$ with
\[
\int_\C|u|^2\e^{2\phi}\hDelta\phi\diff A\le\frac12
\int_\C|f|^2\e^{2\phi}\diff A.
\]
\label{thm-H2}
\end{thm}
Superficially, this theorem looks quite different than H\"ormander's
Theorem \ref{thm-H1}. However, it is in a sense which can be made precise
dual to Theorem \ref{thm-H1}.
In our presentation, we will derive it from the same rather elementary
calculation which H\"ormander uses in e.g. \cite{Horm2}, p. 250.
As such, Theorem \ref{thm-H2} may well be known, but we have not found a
specific reference.
We might add that if the polynomials are dense in the space
$A^2(\C,\e^{-2\phi})$ of entire functions, as they are, e.g., for radial
$\phi$,
then the condition \eqref{eq-dual1} needs to be verified only for monomials
$g(z)=z^j$, $j=0,1,2,\ldots$.
\begin{rem}
Naturally, Theorem \ref{thm-H2} should generalize to the setting of several
complex variables.
\end{rem}
Under a simple condition on the weight, the solution supplied by Theorem
\ref{thm-H2} is unique.
\begin{thm}
If, in addition, $\phi$ is $C^4$-smooth and meets the curvature-type condition
\[
\frac{1}{\Delta\phi}\Delta\log\Delta\phi\ge-2 \,\,\,\,\text{on}\,\,\,\C,
\]
then the solution $u$ in Theorem \ref{thm-H2} is unique.
\label{thm-H2.1}
\end{thm}
\section{The proof of Theorem \ref{thm-H2}}
\subsection{A norm identity}
We denote by $\|\cdot\|_{L^2}$ the norm in the space $L^2(\C)$. For a function
$F$, we let $\Mop_F$ denote the operator of multiplication by $F$. The first
step is the following norm identity for $v\in C^\infty_c(\C)$:
\begin{equation}
\|\dbar v-v\dbar\phi\|^2_{L^2}-\|\partial v+v\partial\phi\|^2_{L^2}=
2\int_\C|v|^2\hDelta\phi \diff A.
\label{eq-normid1}
\end{equation}
To arrive at \eqref{eq-normid1}, we do as follows. As for the first term,
if we let $\langle\cdot,\cdot\rangle_{L^2}$ denote the standard sesquilinear
inner product of $L^2(\C)$, we see that
\[
\|\dbar v-v\dbar\phi\|^2_{L^2}=\langle \dbar v-v\dbar\phi,\dbar v-v\dbar\phi
\rangle_{L^2}=\langle (\dbar-\Mop_{\dbar\phi})v,(\dbar-\Mop_{\dbar\phi})v
\rangle_{L^2}=\langle (\dbar-\Mop_{\dbar\phi})^*
(\dbar-\Mop_{\dbar\phi})v,v\rangle_{L^2}
\]
and together with the corresponding calculation for the second term, we
find that \eqref{eq-normid1} expresses that
\begin{equation}
(\dbar-\Mop_{\dbar\phi})^*(\dbar-\Mop_{\dbar\phi})-
(\partial+\Mop_{\partial\phi})^*
(\partial+\Mop_{\partial\phi})=2\Mop_{\hDelta\phi}.
\label{eq-normid2}
\end{equation}
Here, the adjoints involved are readily expressed: $\partial^*=-\dbar$,
$\dbar^*=-\partial$, and $\Mop_F^*=\Mop_{\bar F}$. The product rule says that
$\dbar\Mop_F=\Mop_{\dbar F}+\Mop_F\dbar$ and
$\partial\Mop_F=\Mop_{\partial F}+\Mop_F\partial$. By identifying adjoints
and carrying out the necessary algebraic manipulations, \eqref{eq-normid2}
is immediate.
\subsection{Reduction to a norm inequality}
While the norm identity \eqref{eq-normid1} is interesting by itself,
we observe here that it has the consequence that
\begin{equation}
2\int_\C|v|^2\hDelta\phi\diff A\le\|\dbar v-v\dbar\phi\|^2_{L^2},\qquad
v\in C^\infty_c(\C).
\label{eq-normest3}
\end{equation}
We write $\Tope:=\dbar-\Mop_{\dbar\phi}$, and express \eqref{eq-normest3} again:
\begin{equation}
2\int_\C|v|^2\hDelta\phi\diff A\le\|\Tope v\|^2_{L^2},\qquad
v\in C^\infty_c(\C).
\label{eq-normest4}
\end{equation}
If $v_j$ is a sequence of functions such that $\Tope v_j$ converges in $L^2
(\C)$, then by \eqref{eq-normest4} the functions $v_j\sqrt{\hDelta\phi}$
converge in $L^2(\C)$ as well, and in particular, $v_j$ converges locally as
an $L^2$-function. It follows that if $h\in L^2(\C)$ is in to the
$L^2(\C)$-closure of $\Tope C^\infty_c(\C)$, then there exists a function
$v\in L^2(\C,\hDelta\phi)$ such that $\Tope v=h$ in the sense of distribution
theory, with
\begin{equation}
2\int_\C|v|^2\hDelta\phi\diff A\le\|h\|^2_{L^2}.
\label{eq-normest5}
\end{equation}
It remains to identify the $L^2(\C)$-closure of $\Tope C^\infty_c(\C)$.
To this end, we identify the orthogonal complement of $\Tope C^\infty_c(\C)$.
So, let $k\in L^2(\C)$ be such that
\begin{equation}
\langle k,\Tope v\rangle_{L^2}=0,\qquad v\in C^\infty_c(\C).
\label{eq-duality1}
\end{equation}
If we write $\Tope^*:=-\partial-\Mop_{\partial\phi}$, distribution theory
gives that \eqref{eq-duality1} is the same as
\[
\langle \Tope^* k,v\rangle_{L^2}=0,\qquad v\in C^\infty_c(\C),
\]
which in its turn expresses that $\Tope^*k=0$ holds in the sense of
distributions. Let us write $\mathrm{ker}\,\Tope^*$ for the space of all
$k\in L^2(\C)$ with $\Tope^*k=0$. We have arrived at the following result.
\begin{thm}
Suppose $h\in L^2(\C)\ominus\mathrm{ker}\,\Tope^*$. Then there exists a
function $v\in L^2(\C,\hDelta\phi)$ such that $\Tope v=h$ with
\[
2\int_\C|v|^2\hDelta\phi\diff A\le\|h\|^2_{L^2}.
\]
\label{thm-H3}
\end{thm}
Since $\Tope^*=-\Mop_{\e^{-\phi}}\partial\Mop_{\e^{\phi}}$,
$k\in\mathrm{ker}\,\Tope^*$ holds if and only if $\e^{\phi}\bar k\in
A^2(\C,\e^{-2\phi})$.
\begin{proof}[Proof of Theorem \ref{thm-H2}]
We put $h:=\e^\phi f\in L^2(\C)$. By condition \eqref{eq-dual1} for all
$g\in A^2(\C,\e^{-2\phi})$, we have -- if we write $k=\e^{-\phi}\bar g$ --
that
\[
0=\int_\C fg\diff A=\int_\C h\bar k\diff A.
\]
As the function $g$ runs over $A^2(\C,\e^{-2\phi})$, $k$ runs over all
elements
of $\mathrm{ker}\,\Tope^*$. We conclude that $h\in
L^2(\C)\ominus\mathrm{ker}\,\Tope^*$, so that Theorem \ref{thm-H3} applies,
and gives a $v\in L^2(\C,\Delta\phi)$ with $\Tope v=h$ with the norm control
\begin{equation}
2\int_\C|v|^2\hDelta\phi\diff A\le\|h\|^2_{L^2}.
\label{eq-dbarest}
\end{equation}
We put $u:=\e^{-\phi}v$. The operator $\Tope$ factors
$\Tope=\Mop_{\e^\phi}\dbar
\Mop_{\e^{-\phi}}$, which means that the equation $\Tope v=h$ is equivalent to
$\dbar u=f$. Finally, the estimate \eqref{eq-dbarest} is equivalent to the
estimate
\[
2\int_\C|u|^2\e^{2\phi}\hDelta\phi\diff A\le\int_\C|f|^2\e^{2\phi}\diff A,
\]
which concludes the proof of the theorem.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm-H2.1}]
We first observe that any two solutions of the $\bar\partial$-equation
differ by an entire function. Moreover, under the given curvature-type
condition, an entire function
$F\in L^2(\C,\e^{2\phi}\Delta\phi)$
necessarily must vanish everywhere, by the following argument. The function
$|F|^2\e^{2\phi}\Delta\phi$ is clearly nonnegative and it is also
\emph{subharmonic in} $\C$. Indeed,
if $F$ is nontrivial, we have that [in the sense of distribution theory]
$\Delta\log|F|\ge0$, and consequently
\[
\Delta\log\big[|F|^2\e^{2\phi}\Delta\phi\big]=2\Delta\log|F|+2\Delta\phi
+\Delta\log\Delta\phi\ge0,
\]
which gives that the exponentiated function $|F|^2\e^{2\phi}\Delta\phi$ is
subharmonic, as claimed. In the remaining case when $F$ vanishes identically
the claim is trivial. Next, by the estimate of each solution of the
$\bar\partial$-problem supplied by Theorem \ref{thm-H2}, it is given that
the function
$|F|^2\e^{2\phi}\Delta\phi$ is in $L^2(\C)$. If $\D(z_0,r)$ denotes the open
disk of radius $r$ about $z_0$, the sub-mean vakue property of subharmonic
functions gives that
\[
|F(z_0)|^2\e^{2\phi(z_0)}\Delta\phi(z_0)\le \frac{1}{\pi r^2}\int_{\D(z_0,r)}
|F|^2\e^{2\phi}\Delta\phi \diff A\le
\frac{1}{\pi r^2}\int_{\C}|F|^2\e^{2\phi}\Delta\phi\diff A.
\]
Letting $r\to+\infty$, we see that the left hand side vanishes. As $z_0$ is
arbitrary, it follows that $F(z)\equiv0$. This completes the proof.
\end{proof}
\begin{rem}
In Lemma 3.1 of \cite{AlRoz} appears a condition which is analogous to
\eqref{eq-dual1} in the context of distributions with compact support.
Also, in the related paper \cite{RozShir}, the Fock weight case
$\phi(z)=\frac12|z|^2$ is considered in the soft-topology setting of
solutions which are distributions.
\end{rem}
\begin{rem}
With $\phi=0$, the norm identity \eqref{eq-normid1} expresses that the
Beurling operator is an isometry on $L^2(\C)$, which implies the Grunsky
inequalities in the theory of conformal mapping (see, e.g., \cite{BarHed};
cf. \cite{BerSch}).
In the (somewhat singular) case when $\phi(z)=\theta\log|z|$,
\eqref{eq-normid1} provides the main norm identity of \cite{Hed2}, which
leads to a Prawitz-Grunsky type inequality for conformal maps. The most general
inequality of this type (with multiple ``branch points'') for conformal maps
was obtained by Shimorin \cite{Shim} (see also \cite{AbuHed}).
It would be of interest to see if the results of Shimorin may be obtained
from the general norm identity \eqref{eq-normid1}.
\end{rem}
\section{Discussion of the necessity of the orthogonality
condition on the datum}
\label{sec-constr}
\subsection{The Fock weight case}
We now narrow down the discussion to the Fock weight $\phi(z)=\frac12|z|^2$.
Since the weight is radial, the polynomials are dense in
$A^2(\C,\e^{-|z|^2})$. Also, the curvature-type condition of Theorem
\ref{thm-H2.1} is readily checked.
It follows that Theorems \ref{thm-H2} and \ref{thm-H2.1} combine to give
the following result.
\begin{thm}
Suppose $f\in L^2(\C,\e^{|z|^2})$. If the datum $f$ satisfies the moment
condition
\[
\int_\C z^jf(z)\diff A(z)=0,\qquad j=0,1,2,\ldots,
\]
then there exists a unique solution $u$ to the equation $\dbar u=f$
with
\[
\int_\C|u(z)|^2\e^{|z|^2}\diff A(z)\le \int_\C|f(z)|^2\e^{|z|^2}
\diff A(z).
\]
\label{thm-BF0}
\end{thm}
We may ask what would happen if the orthogonality condition is not satisfied.
Maybe there still exists some rapidly decaying solution $u$ anyway? The answer
is definitely no.
\begin{thm}
Suppose $f\in L^2(\C,\e^{|z|^2})$, and that $u$ solves the equation $\dbar u
=f$ while $u\in L^2(\C,\e^{\epsilon|z|^2})$ for some positive real $\epsilon$.
Then the datum $f$ has
\[
\int_\C z^jf(z)\diff A(z)=0,\qquad j=0,1,2,\ldots.
\]
\label{thm-BF1}
\end{thm}
Before we turn to the proof, we observe that if $f\in L^2(\R)$ with
\[
\int_\R |f(x)|\e^{x^2/\beta}\diff x<+\infty
\]
for some real $\beta>0$, then its Fourier transform
\[
\hat f(\xi):=\int_\R \e^{-\imag\xi x}f(x)\diff x
\]
extends to an entire function with
\[
\int_\C |\hat f(\xi)|^2\e^{-\beta|\im\xi|^2}\diff A(\xi)=
\frac{2\pi^{3/2}}{\sqrt{\beta}}\int_\R|f(x)|\e^{x^2/\beta}\diff x<+\infty.
\]
In fact, the standard Bargmann transform theory asserts
that this integrability condition characterizes the Fourier image of this
weighted $L^2$ space on $\R$ (cf. \cite{Groech1}). The two-variable extension
of the above-mentioned result maintains that if $f\in L^2(\C,\e^{|z|^2/\beta})$,
then its Fourier transform
\[
\hat f(\xi,\eta):=\int_\C \e^{-\imag(\xi x+\eta y)}f(x+\imag y)\diff x\diff y
\]
is an entire function of two variables, with
\begin{equation}
\label{eq-BT2dim}
\iint_{\C\times\C} |\hat f(\xi,\eta)|^2\e^{-\beta(|\im\xi|^2+|\im\eta|^2)}
\diff A(\xi)\diff A(\eta)=
\frac{4\pi^{3}}{\beta}\int_\C|f(z)|\e^{|z|^2/\beta}\diff A(z)<+\infty.
\end{equation}
\begin{proof}[Proof of Theorem \ref{thm-BF1}]
Since $f\in L^2(\C,\e^{|z|^2})$, the function $\hat f$ is an entire function of
two variables with the norm bound \eqref{eq-BT2dim} with $\beta=1$.
Likewise, as $u\in L^2(\C,\e^{\epsilon|z|^2})$ for some positive $\epsilon$,
$\hat u$ is an entire function of two variables. After Fourier transformation,
the relation $\dbar u=f$ reads
\[
(\imag\xi -\eta)\hat u(\xi,\eta)=2\hat f(\xi,\eta).
\]
This is only possible if $\hat f(\xi,\eta)$ vanishes when $\imag\xi-\eta=0$,
i.e., $\hat f(\xi,\imag\xi)\equiv0$. By the definition of the Fourier
transform, this means that
\[
0=\hat f(\xi,\imag\xi)=\int_\C \e^{-\imag\xi(x+\imag y)}f(x+\imag y)
\diff A(x+\imag y)=\int_\C \e^{-\imag\xi z}f(z)
\diff A(z)=\sum_{j=0}^{+\infty}\frac{(-\imag\xi)^j}{j!}\int_\C
z^jf(z)\diff A(z).
\]
By Taylor's formula, then, this implies that
\[
\int_\C z^jf(z)\diff A(z)=0,\qquad j=0,1,2,\ldots,
\]
as needed.
\end{proof}
\begin{rem}
The constant in Theorem \ref{thm-BF0} is sharp. Indeed, we may consider the
datum $f(z)=-z\e^{-|z|^2}$ for which the solution is $u(z)=\e^{-|z|^2}$.
We calculate that
\[
\int_\C|u(z)|^2\e^{|z|^2}\diff A(z)=\int_\C\e^{-|z|^2}\diff A(z)=\pi
\]
while
\[
\int_\C|f(z)|^2\e^{|z|^2}\diff A(z)=\int_\C|z|^2\e^{-|z|^2}\diff A(z)=\pi,
\]
which gives the desired sharpness.
\end{rem}
\section{Acknowledgements}
The author wishes to thank Ioannis Parissis and Serguei Shimorin
\cite{ParShi} for stimulating discussions related to the norm identity
\eqref{eq-normid1}.
The author also wishes to thank Grigori Rozenblum for an enlightening
conversion on the topic of this paper at the Euler Institute in
St-Petersburg in July, 2013.
|
\section{Introduction}
The energetic formation has a been central question for solar physics for several decades.
We know that the solar atmosphere is heated, most likely by the dissipation of magnetic energy driven by convection.
The energy input in the corona is partially dissipated by radiative losses primarily from optically thin UV/EUV emission lines of highly ionized elements.
The primary mechanism for coronal energy loss is field-aligned thermal conduction.
The energy that is input into the corona that cannot be effectively radiated is redirected through thermal conduction to the transition region and the chromosphere.
These foundational ideas were brought together by \citet[hereafter referred to as RTV]{rosner78a}.
One of basic consequences of the equations for hydrostatic balance in the corona is that the magnetic loop geometry has a determining effect on the properties of the plasma embedded along that loop.
In the constant pressure approximation imposed in RTV scaling, the loop length enters the equations in relation to the scale of thermal conduction and heat input.
In a gravitationally stratified loop \citep{serio_81}, the pressure gradient throughout the loop varies based on the gravity component parallel to the field orientation.
We extend the work of these authors by studying how variations in the geometry of flux rope loops affect the three dimensional density structure of the flux rope.
An example of a hydrostatic solutions are given in Figure 1.
\begin{figure}
\begin{center}
\includegraphics[width=.7\textwidth]{paris_hydro.eps}
\caption{Solutions for density (blue), temperature (red), and conductive flux (green) along an asymmetric flux rope field line geometry (black). ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~}
\label{fig1}
\end{center}
\end{figure}
\section{Flux Rope Model}
\cite{fangib_06} emerge a flux rope through a line-tied lower boundary into a low-beta, isothermal corona.
The model evolves through quasi-static equilibrium as the flux rope is kinematically emerged.
We extract magnetic field lines from a partially emerged equilibrium state \citep[identical dataset as in][]{gibfan_06b}.
Field lines from within the flux rope exhibit several distinct geometries.
We find only a small fraction of the flux rope volume contains field lines with magnetic dips.
Of the 1300 field lines which were extracted in uniform spacing in the r=1.02 plane, 270 were found to contain dips.
Magnetic dips are important as they are able to stably hold prominence plasma from falling into the chromosphere.
Dipped field lines form a sheath-like surface that surrounds the axis of the flux rope.
Arcade loops that surround the emerging rope have expanded to accommodate the additional flux.
\\\indent
While these loops have geometric differences discussed above, they also differ strongly in length.
Figure 2 shows how field line length varies through cross sections of the model volume.
The interior of the flux rope is composed of short field lines.
The field line geometry changes as a function of radial distance from the axis (height of 1.4 R$_s$); there is a gradual lengthening of the loops that is related to the increasing winding number.
There is a sharp boundary at the outer edge of the flux rope.
The outer flux rope field lines are 20\% longer than the neighboring external arcade field lines.
Both length and geometry affect the energy balance of the corona.
We will now discuss how these quantities enter into the equations, and how we solve for the plasma properties within these disparate structures.
\section{Hydrostatic Calculation}
The equations for hydrostatic balance within a coronal flux tube can be written as a three first-order ordinary differential equations as described in \cite{vesecky_79}.
The three equations are:
\begin{displaymath}
\frac{dn}{ds}=\frac{-m g_s-2 k F_c}{2kT}n,~\frac{dF_c}{ds}=E-n^2\Lambda(T),~\frac{dT}{ds}=F_c T^{-5/2},
\end{displaymath}
where $n$ is the density, $T$ is the temperature, $F_c$ is the conductive flux, $m$ is the mass of the proton, $g_s$ is the component of gravity parallel to $\hat{s}$, $E$ is the heat input, and $\Lambda$ is the temperature-dependent radiative loss function.
In this experiment, we use an identical $\Lambda(T)$ to \cite{vesecky_79}.
$E$ is set to 2$\times10^{-6}$ erg cm$^{-3}$ s$^{-1}$.
Unlike \cite{vesecky_79}, we integrate the hydrodynamic equations foot point to foot point.
This presents a problem as we have two foot points and three boundary conditions.
We assume that the temperature and the conductive flux are fixed at the initial foot point (3$\times10^4$K and 10$^{-3} $ erg cm$^{-2}$ s$^{-1}$, respectively) and that the temperature at far-side foot point is also 3$\times10^4$K.
We numerically integrate the hydrostatic equations using the Adams-Bashford-Moulton scheme \citep{shampine}.
The density is varied at the near-side foot point until a solution matching the far-side temperature boundary condition is met within a threshold.\\
\begin{figure}
\begin{center}
\includegraphics[width=0.78\textwidth]{topo2.eps}
\caption{Variation of field line length within and surrounding the flux rope. The left panel displays a cross-section (x=0) of the model volume viewed along the axis of the rope. The right panel displays the cross section (y=0) across the axis.}
\label{fig2}
\end{center}
\end{figure}
We have applied this method to the 1300 field lines which were extracted from the MHD model.
The density solutions from these field lines are compiled into an irregular three dimensional grid.
Figure 3 shows the density for two different lines of sight through the 3D grid: one in the flux rope (blue diamonds) and one in the arcade (red diamonds) at the same projected height (1.4 R$_S$).
There are several important elements to understand in Figure 3.
The flux rope line of sight maintains a density depletion between $\Delta x=0$ and $\Delta x=0$0.32 R$_s$.
The peak depletion is around 35\% at $\delta x=0$.
The flux rope interior and the arcade are characterized by different field line geometries, which are signified by colored arrows in Figure 3a which match with the representative field lines in Figure 3b and 3c.
The pink field line in the flux rope interior reaches a maximum height of 1.4 R$_s$ while the green field line in the arcade achieves a maximum height of 1.7 R$_s$.
In hydrostatic equilibrium with a fixed foot point temperature, the density at the foot points must increase as a function of maximum loop height; higher pressure is needed to support the additional mass in taller loops.
For the arcade line of sight, there is a decrease in density as a function of $\Delta x$.
This is caused the the arched shape of the partially emerged flux rope.
For the flux rope line of sight, there is a increase in density as a function $\Delta x$.
Our model suggests that this is related to the transition from short axial field lines to long dipped field lines (similar to light blue line in Figure 3b and 3c).
These field lines wrap around the density depleted interior field lines and are significantly higher density.
\section{Conclusions}
This experiment presents us with evidence for a simple interpretation of the contrast between the low density cavity and the high density streamer.
The cavity is composed of short field lines while the streamer is composed of taller, longer field lines.
This interpretation is completely consistent with RTV scaling as well our current models magnetic models for prominence structure.
This model makes many assumptions: constant area flux tubes, uniform heating, a highly idealized magnetic structure.
However, we believe this simple model illustrates the important role that field line geometry may play in the morphology of prominence-cavity structure.\\
\\
{\it The National Center for Atmospheric Research is funded by the National Science Foundation.
This work was aided by discussions at the International Space Science Institute in Bern.}
\begin{figure}
\begin{center}
\includegraphics[width=.785\textwidth]{los_plot2.eps}
\caption{Line of sight projection of density structure. (a) Density as a function of x-position looking through the cavity (z=1.4, y=0) is denoted by blue diamonds. A line of sight through the arcade (z=1.4, y=0.3) is denoted by red diamonds. Example field lines which intersect the lines of sight viewed along the axis (b) and across the axis (c). The color of the field lines match the colored arrows in Figure 3a showing the density of the line.}
\label{fig3}
\end{center}
\end{figure}
|
\section{Introduction}
We consider statistical analysis of functions and curves where some form of
registration or time warping is of interest. We focus on two applications involving
classification of mouse vertebrae shape outlines in evolutionary biology and the alignment of mass spectrometry
data in proteomics. Both applications require methods which can take account of arbitrary
reparameterizations of the functions or curves of interest.
In order to help choose appropriate methods and models we first describe
three different spaces of interest: the original space, the ambient space and the quotient
space. The choice of space in which to specify the statistical model is important, as
it determines what type of mean estimation and subsequent statistical analyses are carried out.
Our main contribution is to introduce a Bayesian approach to the analysis of functions and
curves, which is demonstrated to be effective in the two applications. Inference is carried out
using Markov chain Monte Carlo simulation, and prior beliefs about the amount of
time warping or registration are included as part of the model.
We wish to consider applications where the functions or curves of interest may not be in alignment. For example,
in the study of growth curves of children it makes sense to consider a time warping of the curves
so that the curves match up in a biologically meaningful way. Children reach various stages of development
such as puberty at different times, and so when comparing growth curves it is sensible to first align
the curves in time and then compare the different heights and growth rates of the children using the time-warped curves
\citep{Ramsli98}. The function registration problem
has been considered by a large number of authors, including \citet{Kneigass92,Silverman95,Ramsli98,Kneipetal00} and \citet{Srivastavaetal11a},
among many others.
Quantities such as a population mean function
and a population covariance function can then be estimated in the space of curves after alignment. In addition to the amplitude
variability of the functions post registration, it is also of interest to analyze the variability in the
registration transformations themselves, which is also known as phase variability.
When analyzing curves in two or three dimensions we have additional potential invariances,
such as translation, rotation and possibly scale invariance.
As a motivating example consider the functions in Figure \ref{fig3}, which are two mass spectrometry scans from a larger dataset.
In the left hand plot of Figure \ref{fig3} it can be seen that the scans are not well aligned, as
the large peaks are not in the same positions in the x-axis. The goal of the alignment is to register the curves with a transformation
of the x-axis so that peaks representing the same peptides can be compared between individuals.
After registration using the methodology of this paper it is clear in the right hand plot of Figure \ref{fig3} that all of the large peaks have been lined up.
In this application it is suspected that much of the alignment can be accounted for by a translation of the x-axis, and so we develop a Bayesian method
for alignment which can place strong prior information on the space of translations, if desired. The estimation of the alignments is
obtained using the posterior mean of the warping functions, and inference is carried out using Markov chain Monte
Carlo simulation. Further details are discussed in Section \ref{massspec5.2} after the methodology has been introduced, and we also consider
a problem in shape analysis where it is of interest to classify vertebrae on the basis of the outline shape.
\begin{figure}[htbp]\centering
\begin{tabular}{cc}
\includegraphics[width=6cm,height=4.125cm]{1D_pair.eps} &
\includegraphics[width=6cm,height=4.125cm]{1D_pair_median_reg.eps}
\end{tabular}
\caption{Mass spectrometry scans, before registration (left) and after registration (right).} \label{fig3}
\end{figure}
\section{The spaces of interest}
\subsection{Original, ambient and quotient spaces}
Consider data of interest in the form of functions or curves
$$f_i(t) : [0,1] \rightarrow \mathbb{R}^m, i=1,\ldots,n.$$
In functional data
analysis \citep{Ramssilv05} the function $f(t)$ is typically in $m=1$
dimension. In statistical shape analysis \citep{Klasetal03} the curve
$f(t)$ is usually in $m=2$ or $m=3$ dimensions. In practice we cannot
observe a complete continuous function but rather a finite set of
discrete points $ \{ f(t_j) \in \mathbb{R}^m : j=1,\ldots,k \}$, where the function is observed
at times $t_j, j=1,\ldots,k$.
In a general form of the registration problem let us first consider the different spaces of interest.
Each object $f$ is located in the original space
(e.g. a space of functions, a space of curves in $\mathbb{R}^m$, or a space of landmark co-ordinates).
The original space is where we represent the raw objects under study.
It is very common to standardize the objects with a preliminary transformation, such as
centering or rescaling so that the objects have unit norm, or perhaps taking a derivative with respect to time to be translation invariant.
These initial transformations
are simple in nature and carried out individually on each object, very much in the spirit of standardizing
variables to have zero mean and unit variance in univariate statistics, or taking first differences in time series.
The standardized object $f^*$
is now represented in the ambient space $S$. Given that it is straightforward to transform to the ambient space, we will assume from
now on that this initial standardization has been carried out.
Finally we
wish to investigate the equivalence class $[f] \in Q$ which is obtained by removing transformations
$\gamma \in G$ from the standardized $f^*$, where $G$ is a group of transformations and
$Q = S / G$ is a quotient space. An important observation is that in order to compute distances in
the quotient space, {optimization} over the transformation group $G$ is required.
This notion of equivalence class and quotient space is precisely that introduced by \citet{Kendall84} for
the representation of the shapes of $k$ landmarks in $\mathbb{R}^m$, where $k > m$. The $k$ landmarks are points located in
$m$ dimensions which represent the important features of the objects under study.
In this situation the original space is
the space of landmark co-ordinates $\mathbb R^{km} \setminus \{0\}$; the ambient space $S$ is the pre-shape sphere
$S^{(k-1)m-1}$ of landmark coordinates which are Helmertized (or centered) to remove location and
scaled to have unit size; and the quotient space is Kendall's shape space $\Sigma^k_m$ after quotienting out
$G = SO(m) = \{ R : R^T R = RR^T = I_m, det(R)=1 \}$, where $SO(m)$ is the special orthogonal group of $m \times m$ rotation matrices.
See \citet{Kendetal99} for a detailed description
of the geometry of this space.
In functional data analysis the registration group is a transformation of the domain of the
function, for example a translation $\gamma(t) = t+c$, affine transformation $\gamma(t) = at+c$,
or the full group of diffeomorphic transformations $\gamma : [0,1] \rightarrow [0,1]$,
such that $\gamma$ is 1-1 and onto.
The functions themselves lie in the original space, are then standardized to the
ambient space, and then finally are decomposed such that
the amplitude variability is represented in the quotient space
and the phase variability is contained in the group of transformations $G$.
In curve analysis the registration of interest is the transformation of the domain, and
in addition we may wish to register using the translation, rotation and scaling of the curve. In this case the
curves lie in the original space, standardized versions lie in the ambient space, then the shapes of the curves
are represented in the quotient space. The main spaces used in this paper are given in Table \ref{examples}.
\begin{table}[htbp]
\hskip -1.5cm
\begin{center}
\begin{small}
\begin{tabular}{|cccc|}
\hline
Original object & Ambient space & Distance & Quotient space distance\\
\hline
$X \in \mathbb{M}^{k \times m}$ & $Z = \frac{HX}{\| HX \|} \in S^{(k-1)m}$ &
$\| Z_1 - Z_2 \|$ & ${\inf}_{\Gamma \in SO(m)} \| Z_1 - Z_2 \Gamma \|$ \\
$\{ f(t) : t \in \mathbb{R} \}$ & $q = \frac{\dot f}{| \dot f |^{1/2} } \in \mathbb{L}^2$ &
$\| q_1 - q_2 \|_2$ & ${\inf}_{\gamma \in G} \| q_1 - q_2 \circ \gamma \|_2$ \\
$\{ f(t) : t \in \mathbb{R}^m \}$ & $q = \frac{\dot f}{\| \dot f \|^{1/2} } \in \mathbb{L}^2$ &
$\| q_1 - q_2 \|_2$ & ${\inf}_{\gamma \in G,\Gamma \in SO(m)} \| q_1 - (q_2 \circ \gamma) \Gamma \|_2$ \\
\hline
\end{tabular}
\end{small}
\end{center}
\caption{Three examples of original objects, ambient spaces, ambient space distances and quotients spaces.
Row 1: $k$ landmarks in $m$ dimensions, where $H$ is a Helmert sub-matrix used for removing translation
and $\Gamma$ is an $m \times m$ rotation matrix; row 2: 1-D functions, with warp $\gamma \in G$ a re-parameterization of time;
row 3: curves in $m$-D with warp $\gamma$ a re-parameterization of arc-length and $\Gamma$ is
a rotation matrix in $m$-dimensions.}\label{examples}
\end{table}
For our analysis of functions and curves, the original space and the
ambient space $S$ are standard classical spaces, such as $\mathbb{L}^2$,
$\mathbb{L}^2 \times \cdots \times \mathbb{L}^2$ or $S^{d-1}$, where statistical models
can be relatively easily formulated, and inference carried out.
In terms of statistical modelling and inference, working with objects in the
group $G$ of transformations is more challenging, but can be undertaken. The
geometry of the group is usually relatively simple and well understood.
However, the quotient space can be considerably more complicated in some situations.
For example, the similarity shape space of a finite set of landmarks in
three dimensions is very complicated, being a non-homogeneous space with
singularities \citep{Lekend93}.
So, an important question is:
in which space shall we define our statistical model, the original, ambient or
quotient space? Since the transformation from the original to the ambient space is
quite straightforward, the main issue is whether we should consider models in the
ambient space or the quotient space. Ultimately the choice of model will depend on the goals
of the study and what we are trying to make inference about.
Let us first consider two data objects $X_1$ and $X_2$,
which could both be standardized functions, curves, landmarks or any other type of object in
an ambient space $S$. How close are $X_1$ and $X_2$, ignoring arbitrary registrations
$\gamma_1, \gamma_2 \in G$? Let $[X_1]$ and $[X_2]$ denote the amplitudes (or shapes)
of $X_1, X_2$.
A natural distance between the amplitude functions is in the quotient space:
$$ d([X_1] , [X_2] ) = \inf_{\gamma \in G} d (X_1 , X_2 \circ \gamma ) , $$
where we must also have the isometric property
$$ d([X_1 \circ \gamma^*] , [X_2 \circ \gamma^*] ) = d([X_1] , [X_2] ),$$
where an arbitrary common transformation $\gamma^*$ can be applied to both
objects and the quotient distance remains unchanged. This property is
also known as a parallel orbit property, in that the orbits (transformations
of an object by $\gamma^*$) are parallel, and it is also known as ``right-invariance''.
This property is a necessity when thinking about practical statistical analyses which are
invariant to transformations.
If we apply an arbitrary transformation to our data then clearly all distances must remain invariant.
\subsection{Statistical models and inference}
Consider a distribution for a random object $X$, where it is the equivalence class
up to transformations in $\gamma \in G$ that is of interest.
We have several choices for specifying a distribution. We could
model $X$ in the ambient space with a population mean
\begin{equation}
\mu_A = {\rm arg}\inf_{\nu \in S} \int_S d( x , \nu)^2 h(x) dx , \label{muA}
\end{equation}
where $h(x)$ is the probability density function (p.d.f.) of $X$. If $d(\cdot,\cdot )$ is the $\mathbb L^2$ or
Euclidean norm then $\mu_A = E[X] = \int x h(x) dx$.
The key location parameter of
interest is then the amplitude (shape) of $\mu_A$ written as $[\mu_A]$.
Statistical models in the ambient space are quite straightforward to specify because the
ambient space is usually not complicated.
For example we specify a stochastic process/probability distribution for $X$, and then
choose some coordinates in the quotient space, which we write as $U = [X]$ together with
registration parameters $\gamma \in G$. We can specify a probability distribution for $X$ and
transform from $X$ to $U$ (where $U = X \circ \gamma^{-1} \in Q$ and
$\gamma \in G)$.
Likelihood based inference about $\mu_A$ up to transformations $\gamma$
is then carried out after {marginalization}, i.e., after integrating out
the transformations $\gamma$ from the distribution of $X$. This approach was
used by \citet{Marddryd89b,Drydmard91a,Drydmard92}
in landmark shape analysis for example.
Alternatively, we could model the equivalence class $U = [X]$ directly in the quotient space with
population \cite{Frechet48} mean
\begin{equation}
\mu_Q = \mathop{{\rm arg}\inf_{\mu \in Q}} \; \; \int_Q d( u , \mu )^2 h(u) d u , \label{muF}
\end{equation}
where $h(u)$ is the p.d.f. of $U$, and $d(\cdot,\cdot )$ is an intrinsic distance in the space.
An intrinsic distance is the length of the shortest geodesic path
between two points, where the path remains in the space at all times.
The minimized value of the expected squared distance is known as the Fr\'echet variance,
and we assume that a global minimum is obtained. If instead only a local minimum has been
found, we denote this as the Karcher mean \citep{Karcher77}.
Also, we could consider extrinsic distance between two points, where a space is embedded
in a higher dimensional Euclidean space. The extrinsic distance is taken as the Euclidean
distance between the points in the embedding space.
The population extrinsic mean
\begin{equation}
\mu_E = \mathop{{\rm arg}\inf_{\mu \in Q}} \; \; \int_Q d_E( u , \mu )^2 h(u) d u , \label{muE}
\end{equation}
where $d_E(\cdot,\cdot )$ is an extrinsic distance.
Models can be specified in the quotient space itself and we can
perform inference on $\mu_Q$ or $\mu_E$.
The method requires {optimization} over the $\gamma$ parameters in order to compute
the intrinsic distances in the shape spaces. This is the approach used in
Procrustes analysis \citep{Goodall91} in landmark shape analysis.
\begin{table}[htbp]
\begin{center}
\begin{tabular}{|ccc|}
\hline
Type of mean & Notation & Reference\\
\hline
Ambient space mean function& $\mu_A$ & Equation (\ref{muA})\\
Quotient space/Fr\'echet/Karcher mean function& $\mu_Q$ & Equation (\ref{muF})\\
Extrinsic mean function & $\mu_E$ & Equation (\ref{muE})\\
Ambient space mean vector& $\mu_A([t])$ & Section \ref{asymp}\\
Quotient space mean vector& $\mu_Q([t])$ & Section \ref{asymp}\\
\hline
\end{tabular}
\end{center}
\caption{Notation for the types of population means}\label{means}
\end{table}
A summary of the notation for the different types of population means considered in the paper is given in
Table \ref{means}.
In the next section we shall describe some methods for computing distances and carrying out
inference in quotient spaces
for functions and curves. Then, in the following section we introduce our main approach
to modelling using a Bayesian procedure in the ambient space.
\section{Quotient space}
\subsection{SRVF and quotient space}
Let $f$ be a real valued differentiable curve function in the original space, $f(t): [0,1]\rightarrow \mathbb R^m$.
From \citet{Srivetal11} the Square Root Velocity Function (SRVF) of $f$ is defined as $q :[0,1]\rightarrow \mathbb R^m$, where
$$q(t)=\frac{\dot{f}(t)}{\sqrt{\|\dot{f}(t)\|}} \; \; , $$
and $\|f\|$ denotes the standard Euclidean norm.
After taking the derivative, the $q$ function is now invariant under translation of the original function, and is thus in the ambient space.
The main interests of this paper consider situations when $m=1$ for functions and $m=2$ for planar shapes.
In the one dimensional functional case the domain $t \in [0,1]$ often represents `time' rescaled to unit length, whereas in two and higher dimensional cases
$t$ represents the proportion of arc-length along the curve.
Let $f$ be warped by a re-parameterization $\gamma\in G$, i.e., $f\circ \gamma$, where $\gamma \in G$ : $[0,1]\rightarrow [0,1]$ is a strictly increasing
differentiable warping function. The SRVF of $f\circ \gamma$ is then given as
$$q^*(t)=\sqrt{\dot{\gamma}(t)}q(\gamma(t)),$$
using the chain rule.
There are several reasons for using the $q$ representation instead of directly working with the original curve function $f$. One of the key reasons
is that we would like to consider a metric that is invariant under re-parameterization transformation $G$. The
elastic metric of \citet{Srivetal11} satisfies this desired property,
$$d_{\rm Elastic}(f_1\circ\gamma,f_2\circ\gamma)=d_{\rm Elastic}(f_1,f_2),$$
although it is quite complicated to work with directly on the functions $f_1$ and $f_2$.
However, the use of the SRVF representation
simplifies the calculation of the elastic metric to an easy-to-use
$\mathbb L^2$ metric between the SRVFs, which is attractive both theoretically and computationally.
If we define the group $G$ to be domain re-parameterization and we consider an equivalence class for $q$ functions under $G$, which is denoted as $[q]$,
then we have the equivalence class $[q]\in Q$, where $Q$ is a quotient space after removing arbitrary domain warping.
First consider the functional case in $m=1$ dimension. An elastic distance \citep{Srivetal11} defined in $Q$ is given as the following
$$d(q_1,q_2)=d([q_1],[q_2])=\inf_{\gamma\in G}\|q_1-\sqrt{\dot{\gamma}}q_2(\gamma)\|_2^2 = d_{\rm Elastic}(f_1,f_2),$$
where
$\| q \|_2 = \{ \int_0^1 q(t)^2 dt \}^{1/2} $
denotes the $\mathbb L^2$ norm of $q$. For the $m=1$ dimensional case the elastic metric is equivalent to the Fisher-Rao
metric for measuring distances between probability density functions.
If $q_1$ can be expressed as some warped version of $q_2$, i.e., they are in the same equivalence class, then $d([q_1],[q_2])=0$ in quotient space.
This elastic distance is a proper distance satisfying symmetry, non-negativity and the triangle inequality.
Note that we sometimes wish to remove scale from the function or curve, and hence we can standardize so that
\begin{equation}
\int_0^1 q(t)^2 dt = 1. \label{unitsize}
\end{equation}
In this case the ambient space would be the Hilbert sphere $S^{\infty}$.
In the $m \ge 2$ dimensional case it is common to also require invariance under rotation of the original curve. Hence we may also wish to
consider an elastic distance \citep{Joshietal07,Srivetal11} defined in $Q$ given as
$$d([q_1],[q_2])=\inf_{\gamma\in G, \Gamma \in SO(m)}\|q_1-\sqrt{\dot{\gamma}}q_2(\gamma)\Gamma\|_2 .$$
The $m=2$ dimensional elastic metric for curves was first given by \citet{Younes98}.
\subsection{Quotient space inference}
Inference can be carried out directly in the quotient space $Q$, and in this case the population
mean is most naturally the Fr\'echet/Karcher mean $\mu_Q$. Given a random sample $[q_1],\ldots,[q_n]$ we obtain the sample Fr\'echet mean by
optimizing over the warps for the 1D function case \citep{Srivastavaetal11a}:
$${\hat\mu_Q} = {\rm arg}\inf_{\mu \in Q} \sum_{i=1}^{n} \inf_{\gamma_i \in G} \| \mu-\sqrt{\dot{\gamma_i}}(q_i\circ \gamma_i)\|_2^2 . $$
In addition for the $m \ge 2$ dimensional case \citep{Srivetal11} we also need to optimize over the rotation matrices $\Gamma_i$ where
$${\hat\mu_Q} = {\rm arg}\inf_{\mu \in Q} \sum_{i=1}^{n} \inf_{\gamma_i \in G,\Gamma_i \in SO(m)} \| \mu-\sqrt{\dot{\gamma_i}}(q_i\circ \gamma_i)\Gamma_i\|_2^2.$$
This approach can be carried out using dynamic programming for pairwise matching, then
ordinary Procrustes matching for the rotation, and the sample mean is given by
$$ \hat\mu_Q = \frac{1}{n} \sum_{i=1}^n \sqrt{ {\dot{\hat\gamma}}_i } (q_i\circ \hat\gamma_i)\hat\Gamma_i . $$
Each of the parameters is then updated in an iterative algorithm
until convergence.
\section{A Bayesian ambient space model}
\subsection{The likelihood for functions}
Our main approach is to consider a model in the ambient space, and then remove the unwanted transformations by marginalization.
Since the $q$-function is a continuous function in the ambient space, naturally we consider a general
stochastic process as the modelling framework for $q$, and we first consider the $m=1$ dimensional case.
We assume a zero mean Gaussian process for the difference of two 1D $q$ functions,
i.e., $\{ q_1-q^*_2 | \gamma \} \sim GP$, where $q_1$ is untransformed and $q_2^*$ is warped by a fixed reparameterization $\gamma$, i.e.
$q^*_2(t)=\sqrt{\dot{\gamma}(t)}q_2(\gamma(t))$. The relative alignment function $\gamma$, contains the parameters of interest.
If we use $q_1([t])$ and $q^*_2([t])$ to denote $k+1$ finite points of $q_1(t)$ and $q^*_2(t)$ respectively, then the joint distribution of
these $k$ finite differences is a multivariate normal distribution based on the Gaussian process assumption, i.e,
$$\{ q_1([t])-q^*_2([t]) | \gamma \} \sim N_k( 0_k, \Sigma_{k\times k}).$$ To simplify the problem, we
assume $\Sigma_{k\times k}=\frac{1}{2\kappa} I_{k\times k}$, where $\kappa$ is a concentration parameter,
although more general covariance functions, such as the Gaussian or Mat\'ern functions, could be used.
\subsection{Prior distributions}
If we treat the re-parameterization function $\gamma \in G$: $[0,1]\rightarrow [0,1]$ as a strictly increasing cumulative distribution function (c.d.f.), then this c.d.f. can be approximated by a set of equally spaced points along its domain $[0,1]$ and linear interpolation. Let $\gamma([t])$ denote $\{\gamma([t_i]),i=0, 1, 2, \dots, M\}$, the finite collection of $M+1$ discretized points and $[t_i]=\frac{i}{M}$, then we have $\gamma([t_0])=\gamma(0)=0$ and $\gamma([t_{M}])=\gamma(1)=1$. Further, if we let $p_{i}=\gamma([t_{i+1}])-\gamma([t_{ i}])$ for $i=1,2,\dots,M$, we have $0<p_i<1$ and $\sum_{i=1}^{M}p_i=1$. If we denote $\bm{p}_M =(p_1,p_2,\dots,p_M)$ and treat $\bm{p}_M$ as a random vector, we can assign a {\rm Dirichlet} prior to $\bm{p}_M |\gamma([t])$, i.e., $\pi(\bm{p}_M)\sim {\rm {
\rm Dirichlet}}(a_1,\ldots,a_M)$. We take equal $a_i = a$ here, writing ${\rm {
\rm Dirichlet}}(a)$.
For $a=1$ the prior distribution is uniform and larger values of $a$ lead to transformations which are more concentrated on $\dot\gamma = 1$ (i.e. translations). In the limit as $M \to \infty$ the warping function is a {
\rm Dirichlet}
process. The choice of $M$ is user specific, but it should be less than the number of discrete points in the $q$ functions, i.e. $M < k$.
The prior distribution for the concentration parameter $\kappa$ is taken as a Gamma($\alpha,\beta$) distribution,
independent of $\gamma$. We use a fairly non-informative prior throughout with $\alpha = 1, \beta = 1,000$, and hence we have prior mean $E[\kappa ] = \alpha\beta = 1,000$
and prior variance $\alpha\beta^2 = 1,000,000$.
\subsection{Pairwise function comparison}
Combining the prior for $\gamma([t])$ and $\kappa$ with
the likelihood model for finite differences of two $q$ functions, the posterior distribution for $\{\gamma([t]),\kappa\}$
given $(q_1([t]),q_2([t]))$ is
$$\pi(\gamma,\kappa|q_1,q_2)\propto\kappa^{p/2}e^{-\kappa \| q_1([t])-\sqrt{\dot{\gamma}}(q_2([t])\circ \gamma)\|^2}\pi(\gamma)\pi(\kappa).$$
In the above model, $p$ represents the degrees of freedom in the model.
If there is no unit scale length constraint (\ref{unitsize}) for $q$, then $p$ would be calculated as follows: $p=km$, where $k$ is the number of finite points taken from the $q$ function, and $m$ is the original function space dimension, i.e., $m=1$ for functions and $m \ge 2$ for curves in
higher dimensions. One degree of freedom will be lost in the constrained case (\ref{unitsize}) and thus $p=km-1$.
In order to carry out inference on the warping function $\gamma$ and concentration parameter $\kappa$, we use
a Markov chain Monte Carlo (MCMC) algorithm to simulate from the joint posterior distribution. The concentration parameter
$\kappa$ is updated using a Gibbs sampler as the conditional posterior for $\kappa$ given all other parameters is still Gamma distributed.
For $\gamma([t])$ with $M+1$ points, a shift in $\gamma([t_i])$ is
proposed at each discrete point $(i = 1, . . . ,M-1)$ and accepted/rejected according to a
Metropolis-Hastings step. Note that $\gamma([t_0])=0$ and $\gamma([t_M])=1$ are both fixed and thus are not updated.
The resulting Markov chain is irreducible and aperiodic, and hence after a large number of iterations we have simulated dependent
values from the posterior distribution.
\subsection{Multiple functions}
If we are interested in multiple functions or curves, we can specify a mean process
for $q$ functions in the ambient space, i.e., $E(q^*_i)=\mu_A$, where $q^*_i=\sqrt{\dot{\gamma}_i}q_i(\gamma_i)$ is a warped version of $q_i$ through some underlying fixed $\gamma_i$. Based on the Gaussian process assumption again, we have
$$\{ q^*_i([t])-\mu_A([t])|\gamma_i([t]),\mu_A([t]) \} \sim N( 0_k, \Sigma_{k\times k})$$
for $i=1,2,\dots,n$, where $n$ is the number of $q$ functions of interest.
We take the prior distribution of $\mu_A$ to be a zero mean Gaussian process with large variance, independent of all other parameters.
The joint posterior density for $(\mu_A,\gamma_1,\dots,\gamma_n)$ is then
$$\pi(\mu_A,\gamma_1,\dots,\gamma_n|q_1,\dots,q_n)\propto
\kappa^{np/2}e^{-\kappa\sum_{i=1}^{n} \| \mu_A([t])-q_i^*([t])\|^2}\pi(\mu_A)\pi(\gamma_1,\dots,\gamma_n)\pi(\kappa)$$
To simulate from the posterior distribution we again use an MCMC algorithm, consisting of pairwise MCMC updates from each curve to the
current mean $\mu_A([t])$ and a Gibbs update for $\mu_A([t])$ itself.
In order to compute the posterior mean estimate $\hat\mu_A([t])$ it is helpful to standardize in each MCMC iteration such that
the Karcher mean of the warping functions from $\mu_A$ to each $q_i$ is the identity function, i.e. $\dot{\hat\gamma} = 1$.
\subsection{Curve Warping}
In the $m \ge 2$ dimensional case, we consider a Gaussian process for the difference of two vectorized $q$ functions in a relative orientation $\Gamma$,
i.e., $\{ {\rm vec}(q_1-q^*_2)|\gamma,\Gamma\} \sim GP$, where $q^*_2=\sqrt{\dot{\gamma}}q_2(\gamma)\Gamma$. The matrix
$\Gamma \in SO(m)$ is a rotation matrix with parameter vector $\theta$. If we assign a prior for rotation parameters (Eulerian angles) $\theta$
corresponding to rotation matrix $\Gamma$, then the joint posterior distribution of $(\gamma([t]),\theta)$, given $(q_1([t]),q_2([t]))$ is
$$\pi(\gamma,\theta|q_1,q_2)\propto\kappa^{p/2}e^{-\kappa \| q_1([t])-q^*_2([t])\|^2}\pi(\gamma)\pi(\theta)\pi(\kappa),$$
where $\gamma, \theta, \kappa$ are independent {\it a priori}.
Throughout the paper we take $\Gamma$ to have a Haar (uniform) prior on the space of rotation matrices.
For the multiple curves case, define $q^*_i(t)=\sqrt{\dot{\gamma}_i(t)}q_i(\gamma_i(t))\Gamma_i$ and $\mu_A=E(q^*_i)$ for fixed $\gamma_i$ and $\Gamma_i$, and we assume $$\mu_A([t])-q^*_i([t])\sim N(0_{km},\Sigma_{km\times km})$$ for fixed $(\gamma_i,\Gamma_i)$, $i=1,\dots,n$.
The joint posterior for $(\mu_A,\gamma_1,\dots,\gamma_n,\Gamma_1,\dots,\Gamma_n)$ is
$$\pi(\mu_A,\gamma_1,\dots,\gamma_n,\Gamma_1,\dots,\Gamma_n|q_1,\dots,q_n)\propto$$
$$\kappa^{np/2}e^{-\kappa\sum_{i=1}^{n} \| \mu_A-q_i^*\|^2} \pi(\mu_A)\pi(\gamma_1,\dots,\gamma_n)\pi(\Gamma_1,\dots,\Gamma_n)\pi(\kappa) ,$$
with warps, rotations and $\kappa$ independent {\it a priori}.
Sampling from the posterior distribution is carried out through exactly the same procedure as when $m=1$ but
with an extra Metropolis-Hastings update for rotation angles.
\section{Properties}
\subsection{Asymptotic properties}\label{asymp}
Let us write $\mathbf{\phi}$ for the vector of all the
parameters in $\{ (\gamma_i, \Gamma_i), i=1,\ldots,n. \}$, and consider $\mu_A$ to be represented by a piecewise linear function connecting a finite
number $k$ points given by $km$-vector $\mu_A([t])$.
The marginal posterior density for ambient space inference is given by
\begin{equation}
\pi_A(\mu_A([t]),\kappa | X ) = \int_{\mathbf{\phi}} \pi(\mu_A([t]),\kappa,\mathbf{\phi} | X) d \mathbf{\phi}. \label{post0}
\end{equation}
The posterior mode estimator of $(\mu_A([t]),\kappa)$ is written as $(\hat\mu_A([t]),\hat\kappa)$ and is obtained by maximizing (\ref{post0}). If the
prior distribution of $(\mu_A([t]),\kappa)$ is uniform then $(\hat\mu_A([t]),\hat\kappa)$ is the maximum likelihood estimator.
If the prior is absolutely
continuous in a neighbourhood of $\mu_A([t])$ with continuous positive density at $\mu_A([t])$ and the distribution satisfies certain regularity conditions
(including differentiable in quadratic mean with non-singular Fisher information matrix $I_{\mu_A([t])}$),
then consistency and asymptotic normality follow.
Subject to the conditions of the Bernstein-von Mises theorem \citep[p141]{vanderVaart98}, we have
$$ \sqrt{n}(\hat\mu_A([t]) - \mu_A([t])) \to N( \frac{1}{\sqrt{n}}\sum_{i=1}^n I_{\mu}^{-1} \dot \ell_{\mu_A([t])}(X_i) , I_{\mu_A([t])}^{-1})$$
in total variation norm as $n \to \infty$,
where $\dot\ell_{\mu_A([t])}(X_i)$ is the derivative of the log-likelihood corresponding to observation $i$.
If $\hat\mu_A$ is a piecewise linear function obtained from the vector $\hat\mu_A([t])$, because $\hat\mu_A([t])$ is consistent for $\mu_A([t])$
we can equivalently state that
$ \hat\mu_A \to \mu_A$
in probability as $n \to \infty$, and hence the ambient space mean is consistent. \citet{Allaetal07} and \citet{Allaetal10} give detailed discussion of
consistency in ambient space models, in particular for deformable templates in image analysis.
The sample Fr\'echet mean vector $\hat\mu_Q([t])$ is consistent for the population Fr\'echet mean vector $\mu_Q([t])$ \citep{Kendall90,Le91}
provided the distribution has support within a regular geodesic ball, and hence the corresponding piecewise linear function $\hat\mu_Q$ is consistent for $\mu_Q$.
\subsection{Comparison of the quotient and ambient space methods}
In general the population Fr\'echet mean $\mu_Q$ in the quotient space and the ambient space mean $\mu_A$ do not have the same amplitude/shape, and hence the
sample Fr\'echet mean can be inconsistent for the ambient space mean. Likewise the sample ambient space mean can be
inconsistent for the population Fr\'echet mean.
It is most natural therefore to use the appropriate estimators given the choice of
mean that is to be estimated. If we are interested in the amplitude/shape of the population ambient space mean $[\mu_A]$ then we use ambient space
inference, while if we are interested in the population Fr\'echet mean then we use the sample Fr\'echet mean. As we see below there are situations
where the sample ambient space and Fr\'echet estimators are very similar, and so our choice between them may be made on other grounds in this case, such as
ease of computation.
When the prior distributions are uniform in the parameters an identical estimator to the sample Fr\'echet mean $\hat\mu_Q$
is obtained from the posterior mode in the Bayesian model of the previous section. If the priors are not uniform then
the posterior mode is in fact a penalized quotient estimator, with the objective function
$$\hat\mu_{pen} = {\rm arg}\inf_{\mu \in Q}
\sum_{i=1}^{n} \inf_{\gamma_i \in G,\Gamma_i \in SO(m)} \{ -\log\pi(\mu,\kappa, \gamma_i,\Gamma_i | q_1,\ldots,q_n ) \} $$
for the curve case.
Note that in many practical situations the ambient space estimator and penalized quotient space estimators are quite similar.
One reason for the similarity in practice is due to a Laplace approximation, and
the marginal posterior density (for ambient space inference) is given by
\begin{equation}
\pi_A(\mu,\kappa | X ) = \int_{\mathbf{\phi}} \pi(\mu,\kappa,\mathbf{\phi} | X) d \mathbf{\phi}. \label{post1}
\end{equation}
whereas the penalized quotient space estimator is obtained by maximization of
\begin{equation}
\pi_Q( \mu,\kappa | X ) \propto \sup_{ \mathbf{\phi} } \pi(\mu,\kappa, \mathbf{\phi} | X ) . \label{post2}
\end{equation}
where $X = \{ q_1,\ldots,q_n\}$.
Often we can consider $\pi_Q(\mu,\kappa | X )$ in (\ref{post2}) to be a good approximation to the marginal density (\ref{post1}) where the integral is
approximated using Laplace's method:
\begin{eqnarray}
\int_{\mathbf{\phi}} \pi(\mu,\kappa,\mathbf{\phi} | X) d \mathbf{\phi} & = & \int_{\mathbf{\phi}}
b(\mathbf{\phi}) \exp\{ -A r(\mathbf{\phi}) \} d\mathbf{\phi} \nonumber\\
& \approx & b(\hat{\mathbf{\phi}}) \left( \frac{2\pi}{A} \right)^{p/2} | \Sigma_{\hat{\mathbf{\phi}} } |^{1/2}
\exp\{ -A r(\hat{\mathbf{\phi}} ) \} \nonumber \\
& \propto & \sup_{ \mathbf{\phi} } \; \; \; b(\mathbf{\phi}) \exp\{ -A r( \mathbf{\phi} )\} \nonumber\\
& \propto & \pi_Q( \mu,\kappa | X ) . \nonumber
\end{eqnarray}
where the gradient of $r( \mathbf{\phi} )$ is zero at $\hat{ \mathbf{\phi} }$, $\Sigma_{\hat{ \mathbf{\phi} }}$
is the inverse of the positive definite Hessian matrix of $r(\mathbf{\phi})$ at $\hat{\mathbf{\phi}}$ and $A$ is a constant.
Laplace's approximation is exact when $(\mathbf{\phi}|\mu,\kappa)$ is multivariate Gaussian, i.e.
$r(\mathbf{\phi})$ is a quadratic form in $\mathbf{\phi}$ and $b(\phi)$ is constant.
\subsection{Multimodality}\label{simtemp}
Multimodality of the posterior distribution can often be an issue with registration of functions and curves.
Simulated tempering \citep{Geyethom95} is a powerful simulation technique designed to overcome problems in moving between local modes
of the posterior. The key idea is to first jump
from the ``cold" temperature (target distribution), where it is difficult to move out of a local mode to a
``hot" temperature where movement between modes is easier
and then jump back to the ``cold" temperature. Using this procedure, the MCMC algorithm
can explore the sample space in a more efficient manner.
Let $\pi(\omega)\propto e^{-U(\omega)}$ be the unnormalized density which is the so called ``cold" distribution, where $\omega$
is the parameter vector. Often $\pi(\omega)$
has multiple local modes
when the dimension of $\omega$ is high. In order to jump out of local modes in the updating algorithm,
we need to make larger moves in the sampling space. Let $\pi_i(\omega)$ be a sequence of $T$ unnormalized
densities where $\pi_i(\omega)\propto \pi(\omega)^{\beta_i}$ for $0\leq \beta_i<1$ and $i=1,\ldots,T$. Following \citet{Liu01}
and \citet{Grametal10}
$\beta_i$ is taken as:
$$\beta_i = (1+\delta_\beta)^{1-i}$$ which is geometric spacing with $\delta_\beta>0$.
The simulated tempering algorithm is then given as follows \citep{Geyethom95}:
\begin{itemize}
\item Given $\pi_i(\omega)$ update $\omega$ using a Metropolis-Hastings step or Gibbs step.
\item Generate $j=i\pm1 $ using probabilities $q_{i,j}$, where $q_{1,2}$=$q_{T,T-1}$=1 and $q_{i,i+1}=q_{i,i-1}=\frac{1}{2}$ if $1<i<T$.
\item Accept the proposal with probability $\min(r,1)$ where
$$r = \frac{\pi_j(\omega)w_jq_{j,i}}{\pi_i(\omega)w_iq_{i,j}}.$$
\end{itemize}
Note that $w_j$ is the prior weight related to $\pi_j(\omega)$ such that each $\pi_j(\omega)$ is explored uniformly, i.e., the MCMC algorithm spends an
equal amount of time in each of the $T$ densities.
In practice, the use of simulated tempering requires much tuning, and we use a simple strategy
where the number of chains to run is $T=10$ and the spacing parameter $\delta_\beta=\frac{1}{\sqrt{N_T}}$ is used where $N_T$
is chosen such that the acceptance rate among jumping across chains is controlled to be roughly between 20\% to 40\%.
Note that the $w_i$ need to be approximated from a preliminary run in which all $w_i$ are set equal. Based on the
preliminary run, the $w_i$ is estimated to be $w_i \propto 1/n_i$ where $n_i$ is the number of samples that the MCMC algorithm takes from chain $i$.
The number of iterations in the tuning pre-run is taken as $50,000$. In case any $n_i$ are equal to 0, $\delta_\beta$ is decreased to
$\delta_\beta=\frac{1}{K\sqrt{N_T}}$ with $K=2,3,\dots$ until all $n_i > 0$.
The sampling of $\kappa$ is straightforward at each level, via a Gibbs sampler
$$\pi(\kappa|\gamma,q_1,q_2)\propto \Gamma(k_i(\frac{p}{2}+\alpha)+1-k_i,k_i(\beta+||q_1-\sqrt{\dot{\gamma}}(q_2\circ \gamma)||^2)),$$
and the sampling of $\gamma$ is via
$$\pi^{k_i}(\gamma|\kappa,q_1,q_2)\propto \{e^{-\kappa||q_1-\sqrt{\dot{\gamma}}(q_2\circ \gamma)||^2}\pi(\gamma) \}^{k_i} .$$
\section{Simulations and applications}
\subsection{Simulation Study}
We consider now a simulation study to compare estimation properties of the quotient and ambient space estimators. The quotient space
estimator $\hat{\mu}_Q$ is obtained by minimizing $\Sigma_{i=1}^{n} \| \mu-\sqrt{\dot{\gamma_i}}(q_i\circ \gamma_i)\|_2^2$ using
dynamic programming while the ambient space estimator $\hat\mu_A$ is obtained using the point-wise mean of posterior samples from MCMC iterations after convergence.
In a single Monte Carlo simulation repetition, a sample of $q$-functions in one dimension is generated through
the model $q_i([t])=\sqrt{\dot{\gamma}_i}\mu_A(\gamma_i([t]))+e_i([t])$, where $e_i\sim N(0_k,\Sigma_{k\times k}), \Sigma = \sigma^2I_{k\times k}$ and
$ \gamma_i\sim
{\rm Dirichlet}(1)$ for $i=1,\ldots,n$.
Both $\hat{\mu}_Q$ and $\hat\mu_A$ are computed and their Fisher-Rao distances to the
underlying true $\mu_A$ are calculated. Note that since the goal is to estimate $\mu_A$ in the ambient space, it is expected that $\hat\mu_A$
will be more appropriate than $\hat\mu_Q$. The MCMC algorithm for $\hat\mu_A$ is run for $50,000$ iterations with a $25,000$ iteration burn-in period.
The prior for $\gamma$ in the Bayesian model is taken as {\rm Dirichlet} with $a=1$, i.e. uniform. Given specific combinations of sample size $n \in \{5, 10, 20, 30,
50, 100, 200 \}$ and error standard deviation $\sigma \in \{ 0.1, 0.3, 0.5, 1 \}$, 100 Monte Carlo
repetitions are run and the arithmetic means of squared Fisher-Rao distances from
both estimators to $\mu_A$ are recorded.
Four examples for $\mu_A$ are considered, which are all scaled to have unit length and unit time. The functions $\mu_A$
in examples I,II,III,IV given in Figure \ref{other} are evaluated at $k$ equal to $51,51,101,51$ points respectively,
and the warping functions are parameterized using $M+1$ points, where $M$ is equal to $10,10,20,10$ respectively.
The underlying $\mu_A$ functions in example I and IV are piecewise linear, example II is a mixture of three normal
densities, and example III is the derivative of the difference of two Gamma functions (in fact it is
the derivative of the canonical haemodynamic response function often used
to model the blood oxygen level dependent signals in fMRI \citep{Glover99}).
The corresponding distances from both estimators are given in Figure \ref{FIGres}.
\begin{figure}[htbp]\centering
\begin{tabular}{cccc}
\includegraphics[width=4cm]{mu1b.eps}
\includegraphics[width=4cm]{mu2b.eps}\\
\includegraphics[width=4cm]{mu3b.eps}
\includegraphics[width=4cm]{mu4b.eps}
\end{tabular}
\caption{The true $\mu_A(t)$ functions used for simulation study. From left to right we denote the
functions as Example I,II,III and IV respectively. } \label{other}
\end{figure}
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=12cm,angle=270]{logdist2.ps}
\end{center}
\caption{The logarithm of the mean square Fisher Rao distance to the true mean $\mu_A$ versus logarithm of sample size $n$.
The full line is the ambient space estimator and the dotted
line is the quotient space estimator. The colours are red ($\sigma=0.1$), green ($\sigma=0.3$),
blue ($\sigma=0.5$) and cyan ($\sigma=1$).
}\label{FIGres}
\end{figure}
From Figure \ref{FIGres} we that see that when $\sigma$ is smaller, the average squared distance between the estimate and
true value is smaller, and as $n$ increases in general the average squared distance becomes smaller. When $\sigma$ is
small ($0.1$), the performance of both estimators is almost equivalent.
However, for larger $\sigma$ in nearly all cases there is an advantage in using the ambient space estimator. One possible
explanation could be over-warping of the quotient estimator to the noisy data due to the optimization over warpings,
compared to the integration over warpings in the ambient space estimator.
For large $\sigma \ge 0.5$ both procedures are clearly
biased for these values of $n$, but it must be borne in mind
that the signal to noise ratio is very low in these cases and so the estimation is very challenging and the discrete implementation
will have an important effect.
Overall, from these examples it does seem that there is an advantage in using the ambient space estimator as we expect,
although this is at the expense of at least twice the computational time.
\subsection{2D Mouse vertebrae}
A two-dimensional application is the study of the shape of the second thoracic (T2) vertebrae in mice
\citep{Drydmard98}. Three groups of mouse vertebrae are available: 30 Control, 23 Large and 23 Small bones. The
Large and Small group underwent genetic selection for large/small body weight, whereas the Control
group consists of unselected mice. Each bone is represented by a curve consisting of 60 points which are
determined through a semi-automatic procedure. Six landmarks are placed at points of
high absolute curvature and then nine pseudo-landmarks are equally-spaced inbetween each pair of landmarks.
The main interests here include carrying out pairwise registration, obtaining mean shapes and
credibility intervals, and carrying out classification based on the registered shapes. It is very common in many application areas
classify objects using shape information \citep{Drydmard98}, and for example in studying the fossil record there is a need to classify bones
from individuals into groups using size and/or shape as there is usually little or no other information available.
We start our analysis by performing a pairwise comparison from the ambient space model,
and we use the MCMC algorithm for pairwise matching with $50,000$ iterations. The $q$-functions are obtained by initial smoothing,
and then normalized so that $\| q \|_2 = 1$. The registration is carried out using rotation through an angle $\theta$ about the origin, and
a warping function $\gamma$.
The original and registered pair (using a posterior mean) are shown in Figure \ref{pair}
and the point-wise correspondence between the curves and a point-wise 95\% credibility interval for $\gamma(t)$ are shown in Figure \ref{pair2}.
The start point of the curve is fixed and is given by the left-most point on the curve in Figure \ref{pair2} that has a red line connecting the two bones.
The narrower regions in the credibility interval correspond well with high curvature regions in the shapes. Convergence of the MCMC scheme
was monitored by trace plots.
We also applied the multiple curve registration, as shown in Figure \ref{mult}.
\begin{figure}[htbp]\centering
\begin{tabular}{c}
\includegraphics[height=5cm,width=5cm]{fig4_1.eps}
\includegraphics[height=5cm,width=5cm]{fig4_2.eps}
\end{tabular}
\caption{Unregistered curves on left and registration through $\hat\gamma(t)_A$ on right.}\label{pair}
\end{figure}
\begin{figure}[htbp]\centering
\begin{tabular}{c}
\includegraphics[height=5.5cm,width=5.5cm]{fig5_1.eps}
\includegraphics[height=5.5cm,width=5.5cm]{fig5_2.eps}
\end{tabular}
\caption{Correspondence based on $\hat\gamma(t)_A$ and 95\% credibility interval for $\gamma(t)$. One of the bones is drawn artificially smaller in order
to better illustrate the correspondence. }\label{pair2}
\end{figure}
\begin{figure}[htbp]\centering
\begin{center}
\includegraphics[width=5.8cm,height=5.8cm]{2D_group_1.eps}
\includegraphics[width=5.8cm,height=5.8cm]{2D_group_3.eps}
\caption{The original curves from Small group, without and with registration.}\label{mult}
\end{center}
\end{figure}
In order to investigate the differences between the new Bayesian method and standard generalized Procrustes analysis on the 60
landmarks we consider
a classification study.
For classification method A, the three
group means are obtained through classical generalized Procrustes analysis \citep{Goodall91} using the {\tt shapes} package in R \citep{Dryden13}
and each test curve is assigned to the trained group which is closest in terms of Procrustes distance. The Procrustes distance is calculated
by minimizing the Euclidean sum of squares between the landmark configurations using translation, rotating and scale.
For method B, the three group means are obtained using the posterior mean from the Bayesian model and each
test curve is classified based on the elastic distance to the mean (i.e. using amplitude variability).
For method C, all training dataset curves are registered in one pooled group using generalized Procrustes analysis
and the Procrustes registered curves are used as the training data. Each test curve is aligned to the mean by
ordinary Procrustes analysis. A Support Vector Machine (SVM) \citep{Changlin01} is then trained on the
registered training curves and applied to the registered test curves. For method D, all training dataset curves are registered through the
Bayesian model and their warped, registered versions are used as the training data. Each test curve is aligned
to the mean by pairwise registration using MCMC. An SVM is then trained on the MCMC registered training curves and applied to
the registered test curves.
A total of 100 Monte Carlo
repetitions are run for each exercise, where the training data and test data are sampled from each group without replacement. In a single Monte Carlo
repetition, 16 curves from the Small group, 20 from the Control group and 16 from the Large group
(about two-thirds of the original data) are randomly selected as the training data, while the remaining
24 curves are used as the test data.
Method A gives an 80\% correct classification rate for the
test data, and method B gives 78\% correct classification. In method C, the classification rate increases rate to 87\% while method
D has the highest classification rate of 92\%. It is interesting to see that method A (Procrustes) does a little bit better than method B (Bayesian),
although they are very similar. We see an overall improvement in methods C and D compared to A and B. The main difference here is that methods
C and D are using hyperplanes to classify between distributions for each group, rather than shape distances which are isotropic in nature.
Method D demonstrates the advantage in using the Bayesian MCMC method for registration with warping, rather than just using the equally spaced
pseudo-landmarks with no warping.
\subsection{Mass spectrometry data}\label{massspec5.2}
Consider a one-dimension functional dataset of Total Ion Count (TIC)
chromatograms of five individuals with acute myeloid leukemia, each with three replicates.
The data were made available and described by \citet{Kochetal13}
at the Mathematical Bioscience Institute (MBI), Columbus, Ohio, workshop on
Statistics of Time Warpings and Phase Variations, November 2012.\footnote{{\tt http://mbi.osu.edu/2012/stwresources.html}}
After pre-processing, each of the 15 observations contains 2001 data points (intensities) from a truncated scan time of 20 to 220 minutes,
with linear interpolation at the same time points of all 15 TIC curves.
We carried out some further pre-processing including
baseline extraction (using cubic spline $\lambda = 5$) and smoothing for larger time points where excessive noise exists (with a cubic spline $\lambda = 0.4$).
Some initial analysis at the workshop was carried out by \cite{Chengetal13proteins}, and
there was a suggestion that the posterior exhibits signs of multimodality
and that the Bayesian MCMC algorithm can become stuck in a local mode. In the analysis here we use simulated tempering from Section \ref{simtemp} and compare
the results to when using the algorithm without simulated tempering.
We first consider a pairwise Bayesian registration between two TIC curves. The estimated warping function $\hat{\gamma}_A$ is taken as the posterior mean of MCMC samples after convergence and the prior for $\gamma$ is taken as ${\rm Dirichlet}(1)$, with $M=40$ in the piecewise linear approximation.
In order to deal with the multimodality of the posterior, simulated tempering is employed \citep{Geyethom95} where ten
levels of temperature are included and the initial phase for tuning to estimate the weights used in moving between levels is $50,000$ iterations.
The two curves before and after alignment are given in Figure \ref{fig3} as seen earlier, and all the peaks look well aligned.
Convergence of the MCMC algorithm is monitored by traceplots of concentration parameter $\kappa$ and
the log posterior, and there are no obvious violations of convergence (not shown).
When multiple curves are under consideration, a set of $(\hat{\gamma}_1,\dots,\hat{\gamma}_n)$ are taken from MCMC samples.
Again the convergence of the MCMC algorithm is checked by the traceplots of the concentration parameter $\kappa$ and log posterior, which seem fine for these data (not shown).
For this particular dataset 14 spiked proteins in the data have been identified in each scan by
the experimenters, which can be regarded as an answer key. If the registration results agree with the answer key, then
the positions of the 14 spiked proteins should coincide after registration.
To indicate the posterior variability of warping functions under different priors,
realizations of the warping functions are taken from the MCMC simulation and are shown in Figure \ref{warps}.
The registration results based on these samples of warping functions are also shown in Figure \ref{mcmcreg}.
The first row of integers corresponds to
the warped spike positions for individual 1, replicate 1. The second row corresponds to individual 1, replicate 2, etc.
In the ideal scenario of perfect alignment we should have all sets of numbers in 14 vertical columns.
In this analysis the MCMC algorithm was run for
50,000 iterations (after tuning the weights for simulated tempering) and we display every 1000th value after the burn-in period of 25,000
iterations, i.e. each number is shown 25 times.
From Figure \ref{mcmcreg}, we see that the main variability when $\gamma\sim {\rm Dirichlet}(1)$ lies in the position of 1, where the curve is flatter and thus contains less information, while when $\gamma\sim {\rm Dirichlet}(100)$ the variability is so small that the different numbers are only slightly different,
indicating a very tight posterior distribution. We see that the registration results under both priors look reasonably good as most positions line up in a vertical line.
However, we do notice that the stronger prior clearly helps the alignment at positions 1, 12, 13 and 14.
\begin{figure}[htbp]\centering
\begin{tabular}{c}
\includegraphics[height=6cm]{warps-lo2.ps}
\end{tabular}
\caption{Samples of groups of warping functions when $\gamma\sim {\rm Dirichlet}(1)$ and $\gamma\sim {\rm Dirichlet}(100)$. The top and bottom
rows indicate the results without and with simulated tempering respectively.} \label{warps}
\end{figure}
\begin{figure}[htbp]\centering
\begin{tabular}{c}
\includegraphics[height=6cm,width=6cm]{newfig4a-1.eps}
\includegraphics[height=6cm,width=6cm]{newfig4a-100.eps}\\
\includegraphics[height=6cm,width=6cm]{1D_multiple_a=1_key_CI.eps}
\includegraphics[height=6cm,width=6cm]{1D_multiple_a=100_key_CI.eps}
\end{tabular}
\caption{Multiple registration results based on samples of $(\gamma_1,\gamma_2,\dots,\gamma_{n})$ when $\gamma\sim {\rm Dirichlet}(1)$ (left) and $\gamma\sim {\rm Dirichlet}(100)$ (right). The top and bottom rows show the results without and with simulated tempering, respectively.} \label{mcmcreg}
\end{figure}
From Figure \ref{mcmcreg} the use of simulated tempering provides
an improvement in the MCMC
algorithm compared to not using it, where there is a danger of becoming stuck in local modes. Using simulated tempering and
$a=100$ all but two spike 2's and all but one of the spike 12's are well aligned.
Since the data mainly exhibit translational effects for registration, the strong prior ($a=100$) is particularly appropriate here.
\section{Discussion}
In this paper we state the distinctions between three spaces of interest: the original, ambient and quotient space.
We compare the ambient space estimator and quotient space estimator in simulation studies,
and explain the similarity in certain situations through a Laplace approximation. An important
component is that we incorporate prior information
about the amount of warping, which is particularly useful in the mass spectrometry application, where too much warping is not desirable. Naturally the choice of prior
is important and will of course be problem specific, however in the mass spectrometry data it was clear that translations are particularly important, and our prior
is weighted strongly towards this feature.
Note that for matching between two functions we also can use multiple alignment, which also involves estimating
the mean function, instead of the pairwise method. Although the multiple alignment method
appears to be a less efficient approach due to the need for the mean function as parameters,
it does have the property that the prior would be invariant under a common reparameterization of both curves.
Although we have focused on 1D and 2D applications the Bayesian methodology can be extended to higher dimensions, for example analysing the shape of
3D surface shapes using the square root normal fields \citep{Jermetal12}.
\bibliographystyle{apalike}
|
\section{Appendix}
In the appendix we list some steps of the calculation leading to the final result (\ref{eq:Final_W_M}). To calculate the expectation value of the Wilson loop, we need the average:
\begin{equation}
\left<g^2h^2\right>=\frac{1}{8}f^{abc}f^{\bar{a}\bar{b}c}\left<X_1^aX_1^{\bar{a}}\right>_{\rho_1}\left<X_2^bX_2^{\bar{b}}\right>_
{\rho_2}~.
\end{equation}
Using the second term in the expression for the fields $\alpha_m$ in
eq.~(\ref{eq:alpha_pert}) we have:
\begin{equation}
X_m^a=-\frac{ig^2}{2}f^{ade} \oint dx^i ~\Phi_m^d \partial^i \Phi_m^e ~,
\end{equation}
or, in momentum space:
\begin{eqnarray}
&&X_m^a = -\frac{ig^2}{2(2\pi)^3}f^{ade}R \times \nonumber \\
&&\int d^2{\bf k} d^2{\bf p}|{\bf k}| J_1(R|{\bf p}|) ~\sin(\alpha-\theta)\Phi_m^d({\bf k}) \Phi_m^e({\bf p}-{\bf k})~.~~~
\end{eqnarray}
In the above expression, $R$ is the radius of the loop, $k$ and $p$
are the momenta of the gluons shown in fig.~\ref{fig:s1}, and
$\alpha$ and $\theta$ are their corresponding azimuthal
angles. $J_1(R|{\bf p}|)$ is a Bessel function of the first
kind. Then:
\begin{eqnarray} \label{eq:4phi}
&&\left<X_m^aX_m^{\bar a}\right>_{\rho_m}=-\frac{g^4}{4(2\pi)^6}f^{ade}f^{\bar a \bar d \bar e}R^2 \times \nonumber \\
&&\int d^2{\bf k}~ d^2 {\bf p}~ d^2 {\bf \bar{k}}~ d^2 {\bf \bar{p}} ~|{\bf k}||{\bf \bar{k}}|~J_1(R|{\bf p}|) J_1(R|{\bf \bar{p}}|)~\times \nonumber \\
&&\sin(\alpha-\theta)\sin(\bar{\alpha}-\bar{\theta})\times \nonumber \\
&&\left<\Phi_m^d({\bf k}) ~\Phi_m^e({\bf p}-{\bf k})~\Phi_m^{\bar d}({\bf \bar{k}}) ~\Phi_m^{\bar e}({\bf \bar{p}}-{\bf \bar{k}})\right>_{\rho_m}~.
\end{eqnarray}
The gauge potential and the two-point function in momentum space are
\begin{eqnarray}
\Phi^a({\bf k})=-\frac{g}{k^2}\rho^a(\bf{k}) \label{eq:Phi^a}~~~~\text{and}~~~~~~~~~~~~~~ \\
\langle \rho^a({\bf k})\, \rho^b({\bf p})\rangle =\mu^2 \delta^{ab} (2 \pi)^2\delta({\bf k}+{\bf p})~.\label{eq:rhorho}
\end{eqnarray}
The four point-function in (\ref{eq:4phi}) receives a contribution
from the fourth order term in the extended Gaussian action \cite{DJP}:
\begin{eqnarray}
S_Q[\rho] = \int d^2 \bf{x} \left[
\frac{\rho^a({\bf x})\rho^a({\bf x})}{2\mu^2}
+ \frac{\rho^a({\bf x})\rho^a({\bf x})
\rho^b({\bf x})\rho^b(\bf{x})}{\kappa_4}\right]~.
\label{eq:Squartic}
\end{eqnarray}
The correction due to the $\rho^4$ operator is:
\begin{equation}
-32\pi^2\frac{\mu^8}{\kappa_4}\left(\delta^{de}\delta^{\bar d \bar e}+\delta^{d \bar d}\delta^{e \bar e}+\delta^{d \bar e}\delta^{e \bar d}\right)\delta\left(\bf{p}+\bf{\bar p}\right)~.
\end{equation}
But, the total color factor of this correction to the expectation value $\left<X_m^aX_m^{\bar a}\right>_{\rho_m}$ in (\ref{eq:4phi}) is equal to zero:
\begin{equation}
f^{ade}f^{\bar a \bar d \bar e}\left(\delta^{de}\delta^{\bar d \bar e}+\delta^{d \bar d}\delta^{e \bar e}+\delta^{d \bar e}\delta^{e \bar d}\right)=0~,
\end{equation}
and does not bring a modification to the expectation value of the Wilson loop.
With the Gaussian contractions (\ref{eq:rhorho}) the expectation value
$\left<X_m^aX_m^{\bar a}\right>_{\rho_m}$ becomes:
\begin{eqnarray} \label{delta}
&&\left<X_m^aX_m^{\bar a}\right>_{\rho_m}=-\frac{g^8\mu_m^4 }{16\pi^2}f^{ade}f^{\bar a d e} R^2 \times\nonumber \\
&&\int d^2{\bf k}~ d^2 {\bf p}~ d^2 {\bf \bar{k}}~ d^2 {\bf \bar{p}} ~\frac{J_1(R|{\bf p}|) J_1(R|{\bf \bar{p}}|)}
{|{\bf k}||{\bf \bar{k}}|({\bf p}-{\bf k})^2({\bf \bar{p}}-{\bf \bar{k}})^2}~\times \nonumber \\
&&\sin(\alpha-\theta)\sin(\bar{\alpha}-\bar{\theta})\times \\
&&\left[\delta({\bf k}+{\bf \bar{k}})\delta({\bf p}-{\bf k}+{\bf \bar{p}}-{\bf \bar{k}})-\delta({\bf k}+{\bf \bar{p}}-{\bf \bar{k}})\delta({\bf p}-{\bf k}+{\bf \bar{k}})\right]~.\nonumber
\end{eqnarray}
After performing two of the integrals using the delta functions in
(\ref{delta}), we get:
\begin{equation}
\left<X_m^aX_m^{\bar a}\right>_{\rho_m}=\frac{g^8\mu_m^4}{8}N_c\delta^{a \bar a}R^2
\int \frac{dk}{k^3}\int dp \frac{J_1^2(R|{\bf p}|)}{|{\bf p}|}~.
\end{equation}
The integral over the momentum $p$ is convergent and equal to $1/2$. The integral over $k$ is infrared divergent and we introduce a cut-off $\Lambda$ to regulate this divergence:
\begin{equation}
\int_\Lambda^\infty \frac{dk}{k^3}=\frac{1}{2\Lambda^2}~.
\end{equation}
So, finally:
\begin{equation}
\left<X_m^aX_m^{\bar a}\right>_{\rho_m}=\frac{g^8\mu_m^4}{32\Lambda^2}N_c\delta^{a \bar a}R^2~. \label{eq:<XX>}
\end{equation}
The cut-off $\Lambda$ can be thought of as due to screening of the
gauge potential. Introducing screened propagators in (\ref{eq:Phi^a}):
\begin{equation}
\Phi^a({\bf k})=-\frac{g}{k^2+m^2}\rho^a(\bf{k})~,
\end{equation}
reproduces the result (\ref{eq:<XX>}) with $\Lambda^2$ replaced by
$m^2$. A self-consistent resummation of screening effects is beyond
the purpose of the present analysis.
In terms of the saturation scale (\ref{eq:Qs}) the final result is:
\begin{equation}
\left<g^2h^2\right>=\frac{\pi^2N_c^7}{32\left(N_c^2-1\right)^3}\frac{Q_{s1}^4Q_{s2}^4}{\Lambda^4}A^2~,
\end{equation}
and
\begin{equation}
W_M(R) \simeq 1-\frac{\pi^2 N_c^6}{64(N_c^2-1)^3}\frac{Q_{s1}^4Q_{s2}^4}{\Lambda^4}A^2~.
\end{equation}
|
\section{INTRODUCTION}
Understanding the formation and evolution of our Galaxy has always been
one of the key quests in modern astrophysics for decades \cite[e.g.][]{freeman02}.
Since RRLs are easily identifiable and
they can provide a powerful means to probe the chemical compositions
and dynamical properties of the old stellar populations,
RRLs in the Galactic globular cluster (GC) systems or in the field
are of particular importance to address the question of the early history
of our Galaxy \cite[see, for example,][]{smith95}.
Also, being a primary distance indicator,
the distance to RRLs can be accurately measured and
RRLs can help to yield important insights into the structure of our Galaxy.
Recent studies by \citet{drake13} or \citet{pietrukowicz} are excellent
examples of making use of RRLs to understand the substructures
in the Galactic halo expected from the theory of the hierarchical structure formation,
and to delineate the bar structure in the central part of our Galaxy.
With the advent of large aperture telescopes, the utility of RRL will
increase further in the future in order to learn more about
the early buildup of our Galaxy or nearby galaxies.
For example, the Large Synoptic Survey Telescope (LSST) will be capable of
detecting RRLs in the local group of galaxies out to $\approx$ 1 Mpc scale
in several years of time span \citep{lsst}, providing us wonderful opportunities
to investigate the substructures of not only the outer halo in our Galaxy
but also the nearby dwarf galaxies.
Furthermore, the future 30- to 40-m class telescopes
will have the capability of detecting RRLs in nearby groups of galaxies
and the utility of RRLs not only as a primary distance indicator
but also as an old stellar population tracer will become more important.
In doing so, we need to calibrate the absolute dimensions of
physical parameters of RRLs in our Galaxy using well understood
samples of RRLs and then we can apply derived transformation relations of RRLs,
such as a period-luminosity-metallicity relation,
to establish a cosmic distance scale.
In recent years, the Fourier decomposition analysis of RRLs has become very popular,
since this method can provide an efficient means
to investigate fundamental stellar parameters,
such as metallicity, absolute magnitude, intrinsic color, temperature, etc
\citep[see for example,][]{simon81, kovacs96, jurcsik98, kk98, kovacs99,
kovacs01, kk07}.
The practical application of Fourier parameters to characterize
the RRLs in our Galaxy or in the Large Magellanic Cloud
appears to be successful, but some worrisome signs
also started to emerge lately.
In their study of RRLs in the globular cluster M3,
\citet{cacciari05} noticed that some fundamental physical parameters,
such as the intrinsic color or the absolute visual magnitude, from Fourier or
pulsational parameters do not agree with the observed ones.
Naturally, both the intrinsic color and the absolute visual magnitude
of RRLs can affect the derived distance scale of RRLs.
The uncertainty in the intrinsic color of RRLs can affect
the foreground interstellar reddening value estimates and,
subsequently, the distance scale of RRLs under consideration will be affected too,
although defining a proper color of RRLs, especially that of RRab type variables
with a large pulsational amplitude,
is not a trivial task \citep[e.g.][]{carney92}.
Achieving an accuracy of a 0.1 mag level for the absolute magnitude of RRLs
is non-trivial and has been a difficult task for decades \citep{smith95}.
Putting aside the discontinuity in the metallicity-luminosity relation
of RRLs, $M_V$ = $a$ + $b$[Fe/H], between the Oosterhoff I and II GCs
seen in our Galaxy \citep{lc99b},
the zero-point $a$ in the above relation is poorly determined
and is a major source of uncertainties in the distance measurement based on RRLs.
Note that an uncertainty of 0.1 mag in the absolute magnitude results
in a $\approx$ 5\% error in the distance measurement.
The recent results from the absolute trigonometric parallaxes of the five
nearby RRLs using Fine Guidance Sensor (FGS)
on Hubble Space Telescope (HST) prefer a brighter absolute
visual magnitude level of RRL than those previously known,
with $M_V$ = 0.45 $\pm$ 0.05 at [Fe/H] = $-$1.50 dex\footnote{
It should be noted that the field RRLs also appear to show
the Oosterhoff dichotomy, which is thought to be originated from
the hierarchical formation of our Galaxy \citep{lc99b}.
If we consider metallicity as a first approximation for the Oosterhoff classes
of the field RRLs at the boundary of [Fe/H] $\approx$ $-$1.6 dex,
XZ Cyg, UV Oct and RR Lyr correspond to the Oosterhoff group I
while RZ Cep and SU Dra correspond to the Oosterhoff group II.}
\citep{benedict11}, which is 0.07 mag brighter than those from
others, i.e.\ $M_V$ = 0.52 mag at [Fe/H] = $-$1.5 dex
\citep[e.g. see][and references therein]{walker12,cacciari13}.
Hence, if the recent measurement by \cite{benedict11} are true,
the previous calibrations from the Fourier parameters or
the theoretical horizontal-branch (HB) models
based on the fainter absolute visual magnitude of RRL are in error.
It would be very desirable to re-calibrate the zero-point in old transformation
relations and to set up new distance scale of RRLs.
It is expected that the zero-point of the absolute visual magnitude and
the accuracy of the measurement will be dramatically improved with Gaia
during the next decade \citep[e.g.][]{cacciari09}.
This is the first of a series of papers addressing the distance scale from RRLs.
In this paper, we present a high precision $BV$ CCD photometry of RRLs in NGC~6723.
Our main goal is to study the RRL population in NGC~6723
and scrutinize the distance scale from RRLs of the cluster
by employing various methods.
In section 2, we address the interesting aspects of the metal-rich
inner halo GC NGC~6723.
In section 3, data acquisition and data reductions are discussed,
including our photometric calibration and the color-magnitude diagram (CMD).
Section 4 describes the methods of period searching and the methods of
searching for new variables.
In section 5, we discuss the fundamental physical properties of RRLs
from Fourier and pulsational parameters.
In particular, we discuss the distance scale from RRLs in detail
and we derive the new zero-point correction terms for the absolute visual
magnitude of RRLs.
In section 6, we show a relation between the degree of photometric
contamination of RRLs
and the apparent crowdedness of the GC systems.
In the appendix, we discuss the evolutionary effect and the photometric
contamination, such as blending, on the three RRLs in NGC~6723, V14, V15 and V16.
We also show that two $\delta$ Sct variable stars
and five W UMa type eclipsing binaries are off-cluster field stars
based on the distance from the center of NGC~6723 and the true distance
from the Sun.
\section{THE METAL-RICH INNER HALO GC NGC~6723}
NGC~6723 is an old metal-rich GC
located $\approx$ 2.6 kpc from the Galactic center \citep{harris96}.
It has long been thought that NGC~6723 is a genuine Galactic bulge
GC owing to its rather high metallicity and
its proximity to the Galactic center.
For example, \citet{vdB93} claimed that NGC~6723 is
on a circular orbit about the Galactic center and
it is a true resident of the central region of the Galaxy.
However, by employing the absolute proper motion study of the cluster,
\citet{dinescue03} showed that NGC~6723 has a highly inclined polar orbit.
They suggested that the kinematics of NGC~6723 is more likely that of a halo object
and is not likely a member of the rotationally supported system,
although the bulge/disk/halo decomposition is not a trivial task
in the central region of our Galaxy.
Nevertheless, several aspects still make NGC~6723 a very interesting
metal-rich inner halo GC in investigating the formation of the Galactic inner halo:
(i) It is metal-rich ([Fe/H]\ $\approx -$1.2) and its metallicity is
similar to (or slightly more metal-poor than) the mean metallicity of RRLs
in the Baade's Window \citep{walker91,kunder08,pietrukowicz}.
(ii) It is very old \citep{fulltonphd,ACS}.
(iii) It is rich in the RRL population \citep{carney92}.
(iv) It suffers only small amount of foreground interstellar reddening,
$E(B-V)$\ = 0.05 mag \citep{harris96} and, therefore, is very suitable
to investigate photometric parameters of RRLs.
(v) Since NGC~6723 has been confined in the Galactic bulge (or the inner halo)
region and never left the bulge (or the inner halo)
due to lower total kinetic energy \citep{dinescue03},
NGC~6723 can play a role as the GC counterpart
for the numerous field inner halo RRLs in Baade's Window.
Therefore, one may infer the formation epoch of
the inner halo RRLs from that of NGC~6723, whose RRLs share common physical properties
as the inner halo RRLs.
In spite of its importance, only a few
studies have been done for the cluster.
\citet{menzies74} studied the variable star population in NGC~6723 and he
presented $BV$ light curves for 29 RRLs.
He suggested that NGC~6723 has a heavily populated blue HB.
\citet{ldz94} also found that NGC~6723 has a rather blue HB morphology,
$(B-R)/(B+V+R)$\footnotemark[1]\footnotetext[1]{$B$, $V$, and $R$
represent the numbers of blue HB stars, RRLs,
and red HB stars, respectively.}
= $-$0.08 $\pm$ 0.08, leading them to conclude that NGC~6723
is a very old GC.
\citet{rosenberg99} carried out a homogenous $VI$ photometry study
of 34 GCs. They measured the magnitude difference
between the turn-off (TO) and HB, $\Delta V_{TO}^{HB}$, and
reached the same conclusion that NGC~6723 is one the oldest GCs
in our Galaxy.
Various studies by others \citep{fulltonphd,fullton96,alcaino99,ACS}
using the model isochrone fitting method, for example,
also showed that NGC~6723 has a very old age.
The previous metallicity measurements of NGC~6723 suggest that
the metallicity of the cluster ranges from [Fe/H]\ $\approx$ $-$0.7
to $\approx$ $-$1.4.
Smith (1981 and references therein) measured the metallicity of
NGC~6723 using the $\Delta S$ method of five RRLs
and he obtained [Fe/H]\ $\approx$ $-$0.7 on Butler's abundance scale.
\citet{zw84} derived [Fe/H]\ = $-$1.09 from
the CMD morphology and the Q39 integrated light index.
Later, \citet{smith86} obtained [Fe/H]\ = $-$1.14
from the DDO photometry system and
\citet{geisler86} obtained [Fe/H]\ = $-$1.35 from
the Washington photometry system.
Later, \citet{fullton96} performed a high resolution
spectroscopic study of three red-giant branch (RGB) stars
and they obtained [Fe/H]\ = $-$1.26 $\pm$ 0.09 for NGC~6723.
They also found that NGC~6723 appears to have
enhanced $\alpha$-elemental abundances, consistent with its old age.
\citet{rutledge97} measured the metallicity of the cluster using the Ca~II triplet lines
of the RGB stars in near infrared passband and they obtained
[Fe/H]\ = $-$1.12 $\pm$ 0.07 on the Zinn \& West's abundance scale
and $-$0.96 $\pm$ 0.04 on the Carretta \& Gratton's abundance scale
\citep{zw84,cg97}.
\section{OBSERVATIONS AND DATA REDUCTION}
Our observations of the cluster were made with seven runs (29 nights in total)
spanning nearly ten years, from August 2002 to July 2012,
using the CTIO 0.9m and the CTIO 1.0m telescopes.
The journal of observations is presented in Table~\ref{tab:obs}.
Also shown in the table are the number of observations per filter and per night.
The CTIO 0.9m telescope was equipped with the Tektronix 2048 No. 3 CCD,
providing a plate scale of 0.40 arcsec pixel$^{-1}$
and a field of view of 13.5 $\times$ 13.5 arcmin.
We used the CTIO 0.9m telescope for four runs in 2002, 2004 and 2012.
The CTIO 1.0m telescope was equipped with an STA 4k$\times$4k CCD camera,
providing a plate scale of 0.289 arcsec pixel$^{-1}$ and a field of view of
20 $\times$ 20 arcmin.
We used the CTIO 1.0m telescope for three runs in 2008, 2009 and 2010.
The typical exposure times for the cluster were 300 s for Johnson $B$-band and
200 s for $V$-band filters for most runs.
In total, we collected 383 frames in $B$ and 401 frames in $V$.
The raw data were processed using the standard
IRAF\footnotemark[2] \footnotetext[2]{IRAF (Image Reduction
and Analysis Facility) is distributed by the National Optical
Astronomy Observatories, which are operated by the Association of
Universities for Research in Astronomy, Inc., under contract with the
National Science Foundation.} packages.
The raw image frames were trimmed and bias-corrected.
For 2002 and 2004 seasons, dome and sky flat images were applied to remove
pixel to pixel variations and illumination corrections using
sky flats were applied to eliminate the large scale structure.
For other seasons, only sky flat images were applied.
Since both telescopes are equipped with an iris type shutter,
the illumination across the CCD chips is not uniform and
the shutter shading correction is non-trivial especially for short exposures.
Therefore, we applied the shutter shading correction for science frames with
less than 10 s exposure time.
During the 2002 season, we obtained the series of dome flats with the
exposure times of 1 s and 60 s using the CTIO 0.9m telescope.
We normalized these two sets of dome flats and then
we calculated the residual in the 60 s dome flats in the time unit
and calculated shutter delay time across the CCD chip.
The exposure time with the CTIO 0.9m telescope is $\approx$ 70 ms longer
in the center than the upper-left or lower-right corners of the CCD chip
and the shutter delay time is not uniform across the chip.
For the CTIO 1.0m telescope, we also obtained shutter delay time correction
images in each of three individual seasons.
The shutter delay time for the CTIO 1.0m telescope is slightly shorter
than that of the CTIO 0.9m telescope.
The exposure time with the CTIO 1.0m telescope is $\approx$ 50 ms longer
at the center of the CCD chip.
We also found that the shutter delay time for the CTIO 1.0m telescope
does not appear to have any temporal variations.
During August 2002, May 2004 and July 2012 seasons, we observed 26, 35 and
36 photometric standard stars from \citet{landolt92,landolt07,landolt09}.
We selected stars with a wide range of colors and magnitudes.
All standard frames were analyzed using the PHOTOMETRY task in DAOPHOTII,
DAOGROW and CCDSTD \citep{pbs90,pbs93,pbs95}.
In order to derive the transformation relations
from the instrumental system to the standard system, we adopt the
following equations:
\begin{eqnarray}
v &=& V + \alpha_v (B-V) + \beta_v X + \gamma_v, \\
b &=& B + \alpha_b (B-V) + \beta_b X + \gamma_b, \nonumber
\end{eqnarray}
where $X$ is the airmass and $V$, $B$ and ($B-V$) are
the magnitudes and color on the standard system.
In Figure~\ref{fig:landolt}, we show the residuals of $V$ and ($B-V$)
in the sense \citet{landolt92,landolt07,landolt09} minus our measurements and
we show our transformation coefficients and residuals in $V$ magnitude
and $(B-V)$ color in Table~\ref{tab:coeff}.
Point-spread function (PSF) photometry for all NGC~6723 science frames
was performed using DAOPHOTII-ALLSTAR, ALLFRAME, COLLECT-CCDAVE-NEWTRIAL
packages \citep{pbs87,pbs93,pbs94,pbs95,turner95}.
To achieve good PSFs on individual science frames,
we used about 100 to 250 isolated bright stars, depending on the seeing conditions
and exposure times, and at least three iterations were required to remove
neighboring stars from our PSF stars.
For the PSF calculation, we adopted a quadratically variable PSF for most cases.
The total number of stars measured from our ALLFRAME run
was more than 31,000 and our final CMD is shown in Figure~\ref{fig:cmd}.
Finally, astrometric solutions for individual stars in our field
have been derived using the positional data extracted from
the Naval Observatory Merged Astrometric Dataset \citep[NOMAD,][]{nomad}.
We achieved rms scatter in the residuals of 0.05 arcsec using the IRAF IMCOORS package.
Then the astrometric fit was applied to calculate the equatorial coordinates
for all stars measured in our science frames.
In Figures \ref{fig:menzies}, \ref{fig:alvarado} and \ref{fig:pbs},
we show comparisons of our results with the photoelectric photometry data of
\citet{menzies74} and \citet{alvarado94} and the CCD photometry data
of \citet{pbs_stds}.
As can be seen in the figures, our photometry is in excellent agreement with
previous measurements by others.
\section{RR LYRAE VARIABLES}
\subsection{Periods and light curves}
\citet{menzies74} and \citet{clement01} list 32 variables in NGC~6723:
29 RRLs, two red variables and one $H_\alpha$ variable.
We used the finding charts provided by \citet{menzies74} to identify
those variables. Then we used his periods
to estimate the photometric phases for our list of new standard
$V$ and $B$ magnitudes and Heliocentric Julian Day (HJD),
and drew light curves for each variable star.
For all variables, the light curves showed phase shifts
in their maxima or minima indicating that period adjustments were needed.
Therefore, we determined new ephemerides and periods
for the 29 previously known RRLs.
To determine the periods, we used three methods:
the string-length method \citep{lafler65},
the phase dispersion minimization method \citep{pdm},
and the information entropy minimization method \citep{cincotta95}.
In all cases, the periods were determined from both $B$ and $V$ data.
In Figure~\ref{fig:rrlc}, we show new light curves for previously known
RRLs \citep{menzies74,clement01}.
In Table~\ref{tab:rrlyrae}, we show positions, new periods,
pulsational amplitudes in $B$ and $V$ bandpasses and mean magnitudes
for individual variable stars.
In the table, $\langle V \rangle_\mathrm{mag}$ and
$\langle B \rangle_\mathrm{mag}$ denote magnitude-weighted integrated magnitudes
and $\langle V \rangle_\mathrm{int}$ and $\langle B \rangle_\mathrm{int}$
denote intensity-weighted integrated magnitudes of variables.
For the RRLs, we also provide static mean magnitudes and
colors as defined by \citet{marconi03};
\begin{eqnarray}
\langle V \rangle_\mathrm{s} &=& -0.345 \langle V \rangle_\mathrm{mag} + 1.345 \langle V \rangle_\mathrm{int}, \\
\langle B-V \rangle_\mathrm{s} &=& 0.488 \langle B-V \rangle_\mathrm{mag} + 0.479 \langle B-V \rangle_\mathrm{int}.
\end{eqnarray}
In Figure~\ref{fig:cmdHB}, we show the CMD of the HB region of NGC~6723.
For RRLs, we adopted magnitude-weighted integrated colors,
$\langle B-V\rangle_\mathrm{mag}$, and the intensity-weighted integrated magnitudes, $\langle V\rangle_\mathrm{int}$.
It should be noted that the results derived from Fourier parameters presented in
Section \ref{s:fc} are for $\langle B-V\rangle_\mathrm{mag}$\ and $\langle V\rangle_\mathrm{int}$.
Since the energy generated by nuclear reaction inside
the variable star is unaffected by the pulsation of the variable star,
the $\langle V\rangle_\mathrm{int}$\ should be the same as would be observed
were the star not pulsating \citep{smith95}.
It is known that the $\langle B-V\rangle_\mathrm{mag}$\ gives the closest approximation to the
color-temperature relation of the ``equivalent static star'' \citep{sandage90}.
However, a caution should be advised on the average color of RRLs,
in particular that of RRab type variable stars with large pulsational amplitude.
As noted by \citet{carney92}, either $\langle B-V\rangle_\mathrm{mag}$\ or $\langle B-V\rangle_\mathrm{int}$\
(and $\langle B-V\rangle_\mathrm{s}$) do not represent the correct temperature
measure for the RRLs due to excess emission
in the $B$ passband during the rise from minimum to maximum radius.
In Figure~\ref{fig:compBV}, we show comparisons of integrated colors
of RRLs in NGC~6723. In the Figure, we show differences in colors,
$\langle B-V\rangle_\mathrm{mag}$\ $-$ $\langle B-V\rangle_\mathrm{int}$\ and $\langle B-V\rangle_\mathrm{mag}$\ $-$ $\langle B-V\rangle_\mathrm{s}$,
as functions of $\langle B-V\rangle_\mathrm{mag}$\ and the blue amplitude, $A_B$.
As can be seen in the Figure, the differences in color between different color
systems are very small, $\leq$ 0.01 mag, for RRc type variable stars with
symmetric sinusoidal light curves.
However, the difference in color can be as large as 0.07 mag
for RRab type variable stars with asymmetric light curves with large $A_B$ values.
Figure~\ref{fig:compBV} (b) and (d) suggest that the difference in the color of RRLs
is closely correlated with the blue amplitude $A_B$, in the sense that
as the blue amplitude increases the discrepancy between
different color systems increases.
\subsection{New variables}\label{s:nv}
Recent CCD photometry surveys of GCs reveal many new variables,
especially RRLs with low amplitudes or in the crowded central part
of GCs and we attempt to search for new variables
in NGC~6723 following the method similar to \citet{lc99a}.
We collected PSF magnitudes for about 460 variable candidates
returned from NEWTRIAL based on the variability index
and calculated magnitude variations as a function of the HJD.
We then calculated light curves of each variable candidate
with the best estimated period.
We employed the information entropy minimization method
with an 8$\times$8 grid size and the string-length method for this purpose.
Visual examinations of plots of magnitude variations against the HJD and
the phase-folded light curves for all variable candidate stars
revealed 22 new variables;
14 RRLs, one SX Phe type and two $\delta$ Scuti type variable stars and
five W UMa type eclipsing binaries.
We also ``rediscovered" all 29 previously known RRLs.
Positions and photometric parameters for new variables
are presented in Table~\ref{tab:rrlyrae} and light curves
for the new variables are shown in Figure~\ref{fig:newrrlc}.
As expected, most of the newly found variables lie near
the central part of the cluster.
Identification charts for the new variables are presented
in Figures~\ref{fig:chart1} and \ref{fig:chart2} for the outer
(with a field of view of 15 $\times$ 15 arcmin) and inner part
(4 $\times$ 4 arcmin) of the cluster, respectively.
We also marked the previously known RRLs in the figures.
Note that all new RRLs lie within the tidal radius of NGC~6723.
In an attempt to confirm the membership of the new RRLs,
we examined the CMD near the HB region (see Figure~\ref{fig:cmdHB})
and the period-amplitude diagram (see Figure~\ref{fig:PvsA_B}),
since there is no previous proper motion study or
radial velocity measurements are available for the new variables.
As shown in Figure~\ref{fig:cmdHB}, the RRc type variable NV1
appears to be too bright to be a regular RRc type variable star of NGC~6723.
The average intensity-weighted integrated magnitude $\langle V\rangle_\mathrm{int}$\
of the remaining seven RRc type variables including NV7 and NV11
is 15.480 $\pm$ 0.019 mag and NV1 is 0.748 mag brighter than the mean magnitude.
However, it should be noted that
NV1 is located within the core radius of the cluster\footnote{The radial distance
of NV1 is 36 arcsec from the center of the cluster and the core radius
of the cluster is 56.4 arcsec \citep{harris96}.}
and its photometric metallicity, [Fe/H] = $-$1.28 dex, from the Fourier analysis
presented in section~\ref{s:feh} is consistent with the mean metallicity of the cluster,
[Fe/H] $\approx$ $-$1.2 dex.\footnote{As discussed above,
the Galactic bulge RRLs are slightly more metal-rich than NGC~6723 is.
Therefore, metallicity alone may not be a proper indicator to distinguish
the cluster member star.}
It is suspected that NV1 may be a member of a binary system.
We attempted to find out the secondary periodicity for NV1 using pre-whitened
data but we failed to get any reliable secondary period of the variable.
Since it lies about 36 arcsec from the center of the cluster,
it may suffer from blending effect with nearby bright stars.
We visually examined the point-spread function of NV1 in individual
science frames and found no sign of blending.
We also found no evidence of blending for NV1 from the images taken with
the HST Advanced Camera for Surveys (ACS) by \citet{sarajedini07} which were retrieved
via the Mikulski Archive for Space Telescopes (MAST)\footnote{STScI is operated
by the Association of Universities for Research in Astronomy, Inc.,
under NASA contract NAS5--26555. Support for MAST for non-HST data is provided
by the NASA Office of Space Science via grant NNX09AF08G and by other grants
and contracts.},
leading us to believe that NV1 is most likely a non-member of the cluster
and is most likely a foreground RRc type variable.
However, follow-up monitoring of NV1 in the future would be desirable to reveal
the potential binary nature of NV1.
The RRab type variable NV14 is slightly fainter
and redder than the rest of RRab type variables of the cluster.
However, the radial distance of NV14 from the center of the cluster
is 264 arcsec and NV14 lies well within the tidal radius of the cluster,
$\approx$ 630 arcsec.
NV14 is in good agreement with other RRab type variable stars in NGC~6723
on the period-amplitude diagram in Figure~\ref{fig:PvsA_B},
and on the Wesenheit function-period diagram in Figure~\ref{fig:W}.
Also other physical parameters of NV14, such as metallicity,
derived from the Fourier analysis do not disagree with those from
other variable stars.
This leads us to believe that NV14 is a cluster member with
an excess differential foreground reddening effect by $\gtrsim$ 0.03 mag.
In appendix, we present the membership status of other types of new variable
stars in NGC~6723. The SX Phe type variable star NV19 is most likely a cluster
member star while the two $\delta$ Sct type variable stars, NV16 and NV18, and
the five W UMa type eclipsing binaries NV15, NV17, NV20, NV21 and NV22
do not appear to be cluster members, based on the radial distance from
the center of the cluster and the true distance modulus.
With the discovery of new variables, the new mean periods of RRab and RRc variables are
$\langle$P$_{ab}$$\rangle$ = 0.541 $\pm$ 0.066 day (35 stars) and
$\langle$P$_{c}$$\rangle$ = 0.292 $\pm$ 0.030 day (7 stars without NV1), respectively.
The errors are those of the mean.
The new RRL distribution is n(c)/n(ab+c) = 0.167.
Our mean periods are in good agreement with those of \citet{menzies74},
who obtained 0.540 day and 0.291 day for $\langle$P$_{ab}$$\rangle$
and $\langle$P$_{c}$$\rangle$, respectively.
However, our new n(c)/n(ab+c) value is slightly smaller than
that of Menzies, n(c)/n(ab+c) = 0.172.
\section{THE FOURIER ANALYSIS}\label{s:fc}
During the last two decades, Fourier decomposition methods have been
successfully applied to characterize the RRLs and other types of variables
\citep[see for example,][]{simon81, kovacs96, jurcsik98, kovacs99,
kovacs01,cacciari05}.
We performed the Fourier analysis and derived Fourier coefficients and
relevant physical parameters of the variables from the new photometry of the cluster.
For our analysis, we used the FORTRAN programs kindly provided by Dr.\ Kov\'acs.
\subsection{Fourier Coefficients}
Fourier coefficients of RRLs in NGC~6723 were derived
for the $V$ band photometry by assuming that the variations
in the observed $V$ magnitude can be reproduced by a Fourier sine series;
\begin{equation}
V = A_0 + \sum^{N}_{i=1}A_i\sin(i\omega t + \phi_i),
\end{equation}
where $\omega$ = 2$\pi/P$ is the angular frequency, $t$ is the epoch
of observations, and the $A_i$ and $\phi_i$ are the Fourier coefficients.
We reconstructed artificial light curves using the derived
Fourier coefficients for each variable.
Each artificial light curve was visually inspected by comparing it to the observed data.
We show the Fourier coefficients for RRLs
in Table~\ref{tab:rrabfc}, where
$R_{i1} = A_i/A_1$, and $\phi_{i1} = \phi_i - i\phi_1$
($0 \leq \phi_{i1} < 2\pi$).
\subsection{The Metallicity}\label{s:feh}
\citet{jk96} derived an empirical relation to estimate metallicity of
RRab type variables from the Fourier parameters,
\begin{equation}
\mathrm{[Fe/H]_\mathrm{JK96}} = -5.038 - 5.394P + 1.345\phi_{31}.
\end{equation}
To apply this relation, \citet{jk96} and \citet{kk98}
introduced the deviation parameter $D_F$
$ = |F_{obs} - F_{calc}|/\sigma_F$ to measure relative accuracy of
the prediction, where $F_{obs}$ is the observed parameter, $F_{calc}$
is the calculated value from the other observed parameters, $\sigma_F$
is the respective standard deviation in their Table 6 \citep[see also][]{kk98}.
The compatibility condition parameter $D_m$, which is defined to be the maximum
of the deviation parameters $D_F$, represents a compatibility test
on the regularity of the shape of the light curves.
\citet{jk96} suggested that physical parameters, such as metallicity,
of RRab type from the Fourier decomposition can be securely applicable
if $|D_m|$ $<$ 3.
In the last column of Table \ref{tab:rrabfc}, we show deviation parameters $D_m$
for all RRab variables in NGC~6723 based on \citet{jk96}.
It should be noted that V13, V18, NV2, NV3, NV5 and NV13
appear to be Blazhko variables in their light curves
as shown in Figures~\ref{fig:rrlc} and \ref{fig:newrrlc},
but their $|D_m|$ values are less than 3,
suggesting that the compatibility condition may not be reliable.
\citet{cacciari05} also found that the compatibility condition by
\citet{jk96}, {\em i.e.} $|D_m|$ $<$ 3, may not be suitable to differentiate
Blazhko variables. They noted that they detected as much as $\approx$ 40\% of
Blazhko variables with $|D_m|$ $<$ 3 and $\approx$ 60\% with $|D_m|$ $<$ 5.
Using the original relation by \citet{jk96}, we obtained a mean metallicity of
[Fe/H]$_\mathrm{JK96}$ = $-$0.93 $\pm$ 0.22 dex from all 18 RRab type variable stars
and $-$0.94 $\pm$ 0.12 dex from the 10 regular RRab variable stars\footnote{
We do not use Blazhko variables V13, V18, NV2, NV3, NV5, NV13 and NV14 in our calculations.
In addition to these, we do not include V14, which appear to suffer from
blending effect with nearby faint companions. See Appendix~\ref{ap:s:notes}.} for NGC~6723.
To calculate more widely used metallicity scales for GC systems by \citet{zw84},
we used the transformation relation given by \citet{jurcsik95},
\begin{equation}
\mathrm{[Fe/H]_\mathrm{JK96}} = (1.431 \pm 0.006)\mathrm{[Fe/H]}_\mathrm{ZW} +
(0.880 \pm 0.010), \label{eq:zw}
\end{equation}
and we obtained [Fe/H]$_\mathrm{ZW}$ = $-$1.27 $\pm$ 0.09 dex from 10 RRab
type variable stars in NGC~6723.
We show the metallicity of each variable in column (2) of Table~\ref{tab:rrabpam}.
We also used an empirical relation given by \citet{morgan07}
to obtain Zinn \& West metallicity scale for RRc type variables,
\begin{eqnarray}
\mathrm{[Fe/H]_\mathrm{ZW}} &=& 52.466P^2 - 30.075P + 0.131\phi^{(c)2}_{31} + \nonumber\\
& & 0.982\phi^{(c)}_{31} - 4.198\phi^{(c)}_{31}P + 2.424,
\end{eqnarray}
where $\phi^{(c)}_{31}$ is the phase for cosine series.
This cosine phase was transformed
from our sine series phase using the following relation,
$\phi^{(c)}_{31} = mod(\phi_{31} + \pi, 2\pi)$, where
$mod$ is the modulus operator, and therefore,
$0 \leq \phi^{(c)}_{31} < 2\pi$.
We obtained [Fe/H]$_\mathrm{ZW}$ = $-$1.18 $\pm$ 0.15 dex for all 7 RRc variable
stars and $-$1.18 $\pm$ 0.20 dex for 4 regular RRc variable,
consistent with that from RRab type variables of the cluster.
We show our results for RRc type variable stars in Table \ref{tab:rrcpam}.
The unweighted mean of the photometric metallicity of NGC~6723
in the Zinn \& West metallicity scale becomes
$\langle$[Fe/H]$_\mathrm{ZW}\rangle$ = $-$1.23 $\pm$ 0.11 dex.
We conclude that our metallicity from the Fourier analysis of RRLs
of NGC~6723 is in excellent agreement with
that of \citet{fullton96}, who obtained [Fe/H]\ = $-$1.26 $\pm$ 0.09
using high resolution spectroscopic abundance study
of the three red-giant branch stars in the cluster.
\subsection{Intrinsic Color $(B-V)_0$ of RRab type Variable Stars
and Interstellar Reddening $E(B-V)$}
The interstellar reddening values of GCs are of great interest
since, for example, they affect the derived photometric temperature
or metallicity from the observed color, the true distance modulus, etc.
The previous estimates of interstellar reddening values of NGC~6723
range from $\approx$ 0.0 \citep{menzies74} to 0.11 \citep{alcaino99}.
We explore the interstellar reddening by comparing the observed colors and
intrinsic colors of RRLs in NGC~6723.
We calculated the intrinsic color of RRab type variables using the
empirical relations given by \citet{jurcsik98},
\begin{equation}
(B-V)_0 = 0.308 + 0.163P - 0.187A_1, \label{eq:bv0:1}
\end{equation}
and by \citet{kovacs01},
\begin{equation}
(B-V)_0 = 0.460 + 0.189\log P - 0.313A_1 + 0.293A_3. \label{eq:bv0:2}
\end{equation}
Note that the above relations are for $\langle B-V\rangle_\mathrm{mag}$\ and these relations are
based on the same program.
Not surprisingly, the intrinsic colors $(B-V)_0$\ of individual variable stars
from both relations [Equations (\ref{eq:bv0:1}) and (\ref{eq:bv0:2})]
are in excellent agreement to within 0.002 $\pm$ 0.004 and
we show average $\langle (B-V)_0\rangle$ of
RRab type variable stars in column (3) of Table~\ref{tab:rrabpam}.
In column (4) of the table, we also show our interstellar reddening $E(B-V)$\ estimates
for each variable by comparing the intrinsic color with the observed
magnitude-weighted integrated color of RRab type variables.
For the mean interstellar reddening value of NGC~6723,
we obtained $E(B-V)$\ = 0.087 $\pm$ 0.014 mag from all 18 RRab type variable stars
and 0.086 $\pm$ 0.007 mag from 10 regular RRab type variable stars .
We also make use of an empirical period-amplitude-color-metallicity
relation given by \citet{piersimoni},
\begin{equation}
(B-V)_0 = 0.507 - 0.052A_B + 0.223\log P + 0.036\mathrm{[Fe/H]},\label{eq:bv0:3}
\end{equation}
assuming [Fe/H]\ = $-$1.23 dex for NGC~6723 and
we obtained $E(B-V)$\ = 0.089 $\pm$ 0.013 for 35 RRab stars
and $E(B-V)$\ = 0.086 $\pm$ 0.010 for 23 regular RRab type variable stars
in NGC~6723.
The interstellar reddening values from Equations (\ref{eq:bv0:1}),
(\ref{eq:bv0:2}) and (\ref{eq:bv0:3}) are in good agreement with
the mean value of $E(B-V)$\ = 0.086 $\pm$ 0.017.
Note that our mean interstellar reddening estimate
appears to be slightly larger than the previous
interstellar reddening value\footnote{\citet{schlegel98} have provided
tools to estimate total reddening along essentially any line of sight,
which should reflect the total reddening to the cluster.
Their result is $E(B-V)$\ = 0.16 mag for the cluster and it is about three
times larger than that of \citet{harris96}.
\citet{dutra00} noted that interstellar reddening estimates
by \citet{schlegel98} for clusters near the Galactic plane
appear to be larger than those based on stellar contents,
possibly due to background dust.
\citet{arce98} have cautioned their readers that \citet{schlegel98}
may over-estimate reddening by a factor of 1.3 to 1.5
in regions with $E(B-V)$\ $> 0.2$ or so. Our results appear to
confirm these cautionary remarks about the use of the \citet{schlegel98}
reddening maps in regions of high reddening
\citep[see also][]{lcs01,lc06}.} by \citet{harris96}, $E(B-V)$\ = 0.05 but
smaller than that of \citet{alcaino99}, who obtained $E(B-V)$\ = 0.11 $\pm$ 0.01
based on $UBVRI$ photometry of the cluster.
It should be noted that the calculated intrinsic colors of RRab type
variable stars from Equations (\ref{eq:bv0:1}) -- (\ref{eq:bv0:3})
have a significant gradient against the observed colors of RRab variables
and weak gradients against periods and blue amplitudes $A_B$, suggesting that
Equations (\ref{eq:bv0:1}) -- (\ref{eq:bv0:3}) are slightly wrong.
In Figure~\ref{fig:ebv} (a), we show the calculated intrinsic
color of individual RRab type variable stars using
Equations (\ref{eq:bv0:1}) -- (\ref{eq:bv0:3}) as a function of
the observed $\langle B-V\rangle_\mathrm{mag}$\ and we obtained the following relation,
\begin{equation}
(B-V)_0 = (0.730 \pm 0.036) \langle B-V\rangle_\mathrm{mag} + (0.026 \pm 0.015).\label{eq:bv0_bvmag}
\end{equation}
The slope of this relation is not unity and, as a consequence,
$E(B-V)$\ [ = $\langle B-V\rangle_\mathrm{mag} - (B-V)_0$]
has a substantial gradient against the observed color
as shown in Figure~\ref{fig:ebv} (b).
Our result strongly suggests that the intrinsic color estimation for
RRab type stars from the Fourier analysis or the pulsational properties
fails to work. We propose that the previous empirical relations
to derive the intrinsic color of RRab type variables
using their pulsational parameters, such as Equations
(\ref{eq:bv0:1}) -- (\ref{eq:bv0:3}), should be re-examined in the future.
In their study of RRLs in M3,
\citet{cacciari05} also found that there existed
significant discrepancies between the intrinsic colors derived from
Fourier parameters and the observed ones.
They found that the distribution of the intrinsic colors $(B-V)_0$\ derived from
Fourier parameters using the transformation relation by \citet{kovacs99}
is somewhat compressed and slightly blue-shifted with
respect to the observe $\langle B-V\rangle_\mathrm{mag}$\ distribution.
The narrowness of the range of their calculated intrinsic color
can be naturally understood if our derived relation between
the calculated $(B-V)_0$\ and the observed $\langle B-V\rangle_\mathrm{mag}$\ [Equation (\ref{eq:bv0_bvmag})]
is still valid for M3 RRLs.
As stated above, Equations (\ref{eq:bv0:1}) and (\ref{eq:bv0:2})
are for $\langle B-V\rangle_\mathrm{mag}$\ but we also investigate the interstellar reddening value
from $\langle B-V\rangle_\mathrm{s}$, i.e.\ $E(B-V)$\ = $\langle B-V\rangle_\mathrm{s}$\ $-$ $(B-V)_0$.
We obtained the mean interstellar reddening value of $E(B-V)$\ = 0.064 $\pm$ 0.021
from 18 RRab type variables and
$E(B-V)$\ = 0.061 $\pm$ 0.014
from 10 regular RRab type variables.
We show our results in the column (5) of Table~\ref{tab:rrabpam}.
Note that this value is in good agreement with that of \cite{harris96}.
However, as can be seen in the relation between the observed $\langle B-V\rangle_\mathrm{mag}$\ and the calculated $(B-V)_0$\
in Equation (\ref{eq:bv0_bvmag}), the slope of the relation between
the observed $\langle B-V\rangle_\mathrm{s}$\ and the calculated $(B-V)_0$\ deviates significantly from unity
[see Figure~\ref{fig:ebv} (c) and (d)].
We obtained the following relation;
\begin{equation}
(B-V)_0 = (0.567 \pm 0.033) \langle B-V\rangle_\mathrm{s} + (0.110 \pm 0.013).\label{eq:bv0_bvs}
\end{equation}
Recently \citet{ACS} presented interstellar reddening values of about 60 GCs
in the HST ACS photometric system, $E(6-8) [= E(F606W-F814W)]$,
using isochrone fitting and they presented $E(6-8)$ values for individual
GCs in their Table 2.
We compared $E(6-8)$ by \citet{ACS} to $E(B-V)$\ by \citet{harris96}
for 61 GCs and we obtained the following transformation relation,
\begin{equation}
E(B-V) = (0.959 \pm 0.027) E(6 - 8) - (0.005 \pm 0.006).
\end{equation}
For NGC~6723, \citet{ACS} obtained $E(6-8)$ = 0.073 and their value
is equivalent to $E(B-V)$\ = 0.065 $\pm$ 0.027.
This result is in excellent agreement with our interstellar reddening estimate from
the static color of RRab variables of the cluster, $E(B-V)$\ = 0.061 $\pm$ 0.014,
and is in agreement with that from $\langle B-V\rangle_\mathrm{mag}$, $E(B-V)$\ = 0.087 $\pm$ 0.014,
within measurement errors.
For the foreground interstellar reddening value of
NGC~6723, we adopt the unweighted mean of those
from the static color $\langle B-V\rangle_\mathrm{s}$\ of RRLs and from the isochrone fitting by \cite{ACS},
$E(B-V)$\ = 0.063 $\pm$ 0.015.
\subsection{The visual magnitude of RR Lyrae variables}
The average intensity-weighted integrated magnitude of the 7 RRc and
30 RRab type variables (excluding V14, V15, V16, NV1, NV2
and NV14)\footnote{In Appendix~\ref{ap:s:notes}, we discuss that
V16 is a more evolved RRab type variable while V15 suffers from
blending with nearby faint objects.
V14 appears to suffer both effects. } in NGC~6723 is
$\langle V\mathrm{(RR)}\rangle_\mathrm{int}$ = 15.459 $\pm$ 0.055 mag.
Our new $\langle V\mathrm{(RR)}\rangle_\mathrm{int}$ is in good agreement with
$\langle V\mathrm{(HB)}\rangle$ magnitudes
by \citet{menzies74}, \citet{fulltonphd} and \citet{rosenberg99},
$\langle V\mathrm{(HB)}\rangle$ = 15.48, 15.47 $\pm$ 0.02,
and 15.45 $\pm$ 0.05, respectively.
It should be noted that our $\langle V\mathrm{(RR)}\rangle_\mathrm{int}$ magnitude
is not for the zero-age HB (ZAHB) but includes RRLs evolved away from
the ZAHB as shown in Figure~\ref{fig:cmdHB}.
Therefore, the maximum $\langle V\rangle_\mathrm{int}$\ of regular RRLs of NGC~6723,
which is $\langle V\rangle_\mathrm{int}$\ = 15.514 mag for V7, should be close to the ZAHB.
\subsection{The absolute visual magnitude of RR Lyrae variables}
\subsubsection{Old calibrations based on the fainter $M_V(\mathrm{RR})$}
\citet{jurcsik98} provided an empirical relation to calculate
the absolute visual magnitude of RRab type variables employing
Fourier parameters,
\begin{equation}
M_V\mathrm{(RRab)} = 1.221 - 1.396P - 0.477A_1 + 0.103\phi_{31}.\label{eq:mv:1}
\end{equation}
Using this relation, we obtained an average absolute visual magnitude
$\langle M_V\mathrm{(RRab)}\rangle$ = 0.854 $\pm$ 0.029 mag
from 18 RRab type variables and
$\langle M_V\mathrm{(RRab)}\rangle$ = 0.852 $\pm$ 0.020 mag
from 10 regular RRab type variables in NGC~6723.
We show the absolute visual magnitude for each star
in column (6) of Table~\ref{tab:rrabpam}.
To estimate the absolute visual magnitude of RRc type variables,
we used the relation given by \citet{kovacs98},
\begin{equation}
M_V\mathrm{(RRc)} = 1.261 - 0.961P - 0.044\phi_{21} - 4.447A_4. \label{eq:mv:2}
\end{equation}
We obtained
$\langle M_V\mathrm{(RRc)}\rangle$ = 0.809 $\pm$ 0.053 mag
for the seven RRc type variables and
$\langle M_V\mathrm{(RRc)}\rangle$ = 0.820 $\pm$ 0.021 mag
for the four regular RRc type variables in NGC~6723.
The absolute magnitudes of each variable star are listed in Table~\ref{tab:rrcpam}.
The average absolute visual magnitude from the four regular RRc type variables is
0.032 mag brighter than that from 10 regular RRab type variables in NGC~6723
but both absolute magnitude scales are in good agreement within the errors.
The combined absolute visual magnitude of NGC~6723
becomes $\langle M_V(\mathrm{RR})\rangle$\
= 0.843 $\pm$ 0.025 mag from 10 regular RRab and four regular RRc variable stars.
As a consistency check, we compare our $\langle M_V(\mathrm{RR})\rangle$\ for NGC~6723
derived from Fourier parameters with the recent estimates of $M_V(\mathrm{RR})$.
As nicely summarized by \cite{cacciari13}, recent theoretical or empirical
$M_V(\mathrm{RR})$\ versus [Fe/H]\ relations prefer $M_V(\mathrm{RR})$\ = 0.52 mag at [Fe/H]\ = $-$1.50 dex.
As a consequence, $M_V(\mathrm{RR})$\ at [Fe/H]\ = $-$1.23 dex (i.e.\ the metallicity of NGC~6723)
should be 0.578 mag or 0.588 mag depending
on the slope of the $M_V(\mathrm{RR})$\ versus [Fe/H]\ relation, 0.214 or 0.250, respectively.
Our $\langle M_V(\mathrm{RR})\rangle$\ for NGC~6723 from Equation (\ref{eq:mv:1}) or (\ref{eq:mv:2})
is about 0.26 mag fainter than the current standard value.
\citet{cacciari05} also found that the derived absolute visual magnitude of
RRLs from Fourier parameters is somewhat faint and
they used the empirical relation by \citet{kovacs98}
[Equation (\ref{eq:mv:2})] to derive
RRc type variable stars in M3, with a brighter zero point
by 0.2 mag [i.e.\ from 1.261 to 1.061 in their Equation (16)].
These results suggest that the $M_V(\mathrm{RR})$\ transformation relations
using Fourier parameters based on the $M_V(\mathrm{RR})$\ from
Baade-Wesselink method by \citet{jurcsik98}
and \citet{kovacs98} may not be reliable and a revision
of the transformation relations should be performed in the future.
\subsubsection{Is RR Lyrae an evolved star?}
Recently, \citet{benedict11} measured absolute trigonometric
parallaxes and proper motions of five nearby RRLs using the HST FGS.
They presented the absolute visual magnitude of RRLs as a function of metallicity,
\begin{equation}\label{eq:hstmv}
M_V\mathrm{(RR)} = (0.214 \pm 0.047)(\mathrm{[Fe/H]} + 1.5) + (0.45 \pm 0.05).
\end{equation}
The metallicity-luminosity relation by \citet{benedict11} gives
$M_V\mathrm{(RR)}$ = 0.45 mag at [Fe/H]\ = $-$1.50 dex, which is about 0.07 mag
brighter than those from others
\citep[see for examples][]{federici12,cacciari13,kollmeier13}.
Although \citet{benedict11} claimed that their relation gives the absolute
visual distance modulus of LMC of
$(m-M)_\mathrm{0}$ = 18.55 $\pm$ 0.05 mag,
the zero-point of \citet{benedict11}, 0.45 $\pm$ 0.05 mag,
and the dereddened mean visual magnitude of RRLs in LMC
by \cite{clementini03}, 19.064 $\pm$ 0.064 mag at [Fe/H]\ = $-$1.50 dex,
can be translated into $(m-M)_\mathrm{0}$ = 18.61 $\pm$ 0.08 mag for LMC,
significantly larger than those from other methods \citep{walker12,cacciari13}.
As kindly noted by the referee, the evolutionary status of RRLs studied by
\cite{benedict11} does not appear to be well established.
\cite{fernley98} suspected that SU Dra is likely an evolved RRL.
More recently \cite{catelan08} claimed that RR Lyr is indeed an evolved RRL.
They performed a careful study of the star using Str\"omgren photometry and
they claimed that RR Lyr is overluminous by 0.06 mag for its metallicity.
In their Equation (4a), \cite{catelan08} provided the following
metallicity-luminosity relation;
\begin{eqnarray}\label{eq:catelan08}
M_V\mathrm{(RR)}_\mathrm{ZW} &=& (0.23 \pm 0.04)\mathrm{[Fe/H]} +
(0.984 \pm 0.127). \nonumber
\end{eqnarray}
Using this relation, one can find $M_V(\mathrm{RR})$\ = 0.64 mag at [Fe/H]\ = $-$1.50 dex,
which is 0.12 mag fainter than the current standard value, 0.52 mag.
It should be noted that the result by \cite{catelan08} is based
on the theoretical HB models by \cite{catelan04}, whose RRL magnitude levels
are 0.08 mag fainter than those from other works
\citep[for example, see Table 1 of][]{cacciari13}.
At [Fe/H]\ = $-$1.50 dex in the Zinn-West metallicity scale,
the theoretical models by \cite{catelan04} predicts $M_V(\mathrm{RR})$\ = 0.60 mag
and their models are 0.08 mag fainter than the current standard value,
$M_V(\mathrm{RR})$\ = 0.52 mag.
It is suspected that if one can make the zero-point of
the theoretical models of \cite{catelan04} 0.08 mag brighter,
then RR Lyr could have been thought as a normal RRL near the ZAHB.
Among 5 RRLs studied by \cite{benedict11}, RR Lyr is of particular importance
at least for two reasons.
First, RR Lyr is the nearest RRL in their sample.
Their absolute trigonometric parallax of RR Lyr using the HST is
$\pi_\mathrm{abs}$ = 3.77 $\pm$ 0.13 mas,
equivalent to $M_V$ = 0.54 $\pm$ 0.07 mag.
Second, the duration of their HST observations of RR Lyr is over 13 years
and they claimed that their parallax measurement for RR Lyr
has been significantly improved compared to their previous study of the star
\citep{benedict02}, $\pi_\mathrm{abs}$ = 3.82 $\pm$ 0.2 mas
(equivalent to $M_V$ = 0.61 $\pm$ 0.10 mag).
If we take their new absolute magnitude ($M_V$ = 0.54 $\pm$ 0.07 mag),
the metallicity of RR Lyr ([Fe/H]\ = 1.41 $\pm$ 0.13 dex)
and the slope in the $M_V(\mathrm{RR})$-[Fe/H]\ relation
by \cite{clementini03} or \cite{benedict11} (0.214 $\pm$ 0.047),
the absoulte visual magnitude of RRLs becomes
$M_V(\mathrm{RR})$\ = 0.52 $\pm$ 0.13 mag\footnote{The error in
$M_V(\mathrm{RR})$\ is dominated by the uncertainty in the metallicity measurement
of RR Lyr, $\pm$0.13 dex.} at [Fe/H]\ = $-$1.50 dex.
This zero-point is in excellent agreement with the current standard value
adopted by others.
Therefore, Equation (\ref{eq:hstmv}) can be re-written as
\begin{equation}\label{eq:rev_hstmv}
M_V\mathrm{(RR)} = (0.214 \pm 0.047)(\mathrm{[Fe/H]} + 1.5) + (0.52 \pm 0.13).
\end{equation}
Our revised metallicity-luminosity relation for RRLs gives
$(m-M)_0$ = 18.54 $\pm$ 0.13 mag for LMC,
if we use $\langle V_0(\mathrm{RRL})\rangle$ = 19.064 $\pm$ 0.064
at [Fe/H]\ = $-$1.50 for RRLs in LMC by \cite{clementini03}.
\subsubsection{$M_V(\mathrm{RR})$\ from Fourier parameters}
Using the same method by \citet{kinman02},
we re-derive the $M_V\mathrm{(RRab)}$ of NGC~6723.
\citet{kinman02} used RR Lyrae and
the empirical transformation relation by \citet{kovacs01},
\begin{equation}
M_V\mathrm{(RR)} = -1.876\log P - 1.158 A_1 + 0.821 A_3 + K,\label{eq:mv:3}
\end{equation}
where $K$ is a zero-point constant which should be determined.
Using $P$ = 0.566837 day, $A_1$ = 0.31539 mag, $A3$ = 0.09768 mag
and $M_V(\mathrm{RR~Lyr})$ = 0.61 $\pm$ 0.10 of \cite{benedict02},
\cite{kinman02} obtained $K$ = 0.43.
As discussed above, the absolute magnitude of RR Lyr from \cite{benedict02}
is 0.07 mag fainter than that of \cite{benedict11}.
Using the same $P$, $A_1$, $A_3$ values,
but $M_V(\mathrm{RR~Lyr})$ = 0.54 $\pm$ 0.07
of \cite{benedict11} for RR Lyr, we obtained $K$ = 0.36.
As a consistency check, we show $M_V(\mathrm{RR})$\ of the globular cluster M3 using
Equation (\ref{eq:mv:3}) with our revised $K$ value and the apparent
visual distance modulus in Figure~\ref{fig:M3RRab}.
In panels (a) and (d), we show observed $\langle V\rangle_\mathrm{int}$\ of RRab type variables in M3
by \cite{cacciari05} and \cite{jurcsik12}, respectively.
As can be seen in the figure, both results are in good agreement.
Panels (b) and (e) show the calculated $M_V(\mathrm{RR})$\ using Equation (\ref{eq:mv:3})
with our revised zero-point, $K$ = 0.36.
Note that \cite{cacciari05} obtained $M_V$ = 0.57 $\pm$ 0.02 mag for
M3 RRab stars and 0.57 $\pm$ 0.04 mag for M3 RRc stars in their analyses,
and their results are in good agreement with our results,
0.553 $\pm$ 0.015 mag and 0.555 $\pm$ 0.016 mag from
RRL data by \cite{cacciari05} and \cite{jurcsik12}, respectively.
In panels (c) and (f), we show apparent visual distance modulus for M3
by comparing the observed $\langle V\rangle_\mathrm{int}$\ and the calculated $M_V$ of RRab stars.
We obtain apparent visual distance modulus values of 15.098 $\pm$ 0.028 mag
and 15.113 $\pm$ 0.024 mag from \cite{cacciari05} and \cite{jurcsik12}, respectively.
Our apparent visual distance modulus values for M3 are in excellent
agreement with that of \cite{federici12}, who presented $(m-M)$ = 15.11
for the cluster, indicating that our revised $K$ value works correctly.
Also shown in the figure with red dashed lines are $M_V(\mathrm{RR})$\ and $(m-M)$ for
M3 RRab type variables using the metallicity-luminosity zero-point
by \cite{catelan08}.
If we use the zero-point by \cite{catelan08}, $M_V(\mathrm{RR})$\ is fainter and
$(m-M)$ is smaller by about 0.05 mag than our results for M3.
\subsubsection{$M_V(\mathrm{RR})$\ for NGC~6723}
We show $M_V$ of individual RRab variable stars in NGC~6723
using Equation (\ref{eq:mv:3}) with $K$ = 0.36
in column (10) of Table~\ref{tab:rrabpam}.
The mean absolute visual magnitude in this case is
$\langle M_V\mathrm{(RRab)}\rangle$ = 0.579 $\pm$ 0.041 mag
from all 18 RRab type variables and
$\langle M_V\mathrm{(RRab)}\rangle$ = 0.581 $\pm$ 0.038 mag
from 10 regular RRab type variables in NGC~6723.
This value is in excellent agreement with our that expected from
our revised metallicity-luminosity relation,
$\langle M_V\mathrm{(RRab)}\rangle$ = 0.578 $\pm$ 0.133 mag.
The difference in the absolute magnitude between that from the original
relation by \citet{jurcsik98} and that from the revised relation of \citet{kovacs01}
with our new zero-point constant becomes 0.270 $\pm$ 0.020 mag.
Note that if we use the zero-point constant by \citet{kinman02}, $K$ = 0.43,
then $\langle M_V\mathrm{(RRab)}\rangle$ = 0.651 $\pm$ 0.038 mag
for 10 regular RRab type variables in NGC~6723.
Our results also support the idea that the zero-point constant
in Equation (\ref{eq:mv:2}) should be decreased
by about 0.2 mag as suggested by \citet{cacciari05}.
If we use the reduced zero-point constant by 0.2 mag
for RRc type variable stars in NGC~6723,
$\langle M_V\mathrm{(RRc)}\rangle$ = 0.620 $\pm$ 0.021 mag for
four regular RRc variable stars is in marginal agreement with
that from RRab type stars within the measurement errors.
\subsection{The distance modulus}\label{s:dm}
\subsubsection{Fourier parameters}
We derive the distance modulus for NGC~6723 by comparing the observed $\langle V\rangle_\mathrm{int}$\
with the absolute visual magnitude of RRab variables $M_V\mathrm{(RRab)}$.
Using Equation (\ref{eq:mv:1})
and our interstellar reddening value, $E(B-V)$\ = 0.063 $\pm$ 0.015 mag,
we obtained $\langle (m-M)_0\rangle$ = 14.405 $\pm$ 0.083 mag from 18 RRab stars
and $\langle (m-M)_0\rangle$ = 14.418 $\pm$ 0.032 mag from 10 regular RRab stars
in NGC~6723.
We show our true distance modulus $(m-M)_0$
for individual variables in columns (7) of Table~\ref{tab:rrabpam}.
As discussed previously, Equation (\ref{eq:mv:1}) is based
on the fainter RRab absolute magnitude scale, hence
our $\langle (m-M)_0\rangle$ value is under-estimated.
Our derived $\langle (m-M)_0\rangle$ value should be corrected
by 0.270 $\pm$ 0.020 mag.
Therefore, if we apply this correction term,
our true distance modulus of NGC~6723 using 10 regular RRab stars
becomes $\langle (m-M)_0\rangle$ = 14.681 $\pm$ 0.038 mag,
corresponding to a distance from the Sun of 8.63 $\pm$ 0.15 kpc.
We also make use of Equation (\ref{eq:mv:3}) with $K$ = 0.36.
We obtained $\langle (m-M)_0\rangle$ = 14.476 $\pm$ 0.074 mag
from 18 RRab stars and 14.684 $\pm$ 0.036 mag from 10 regular RRab stars
using $E(B-V)$\ = 0.063 $\pm$ 0.015 mag and
the distance of the cluster from the Sun is 8.65 $\pm$ 0.14 kpc.
We show our results in column (11) of Table~\ref{tab:rrabpam}.
Our true distance modulus values from RRab type variable stars in NGC~6723
using the the relation by
\cite{jurcsik98} with our correction term and
by \cite{kovacs01} with our revised zero-point
are in excellent agreement
with that of \citet{harris96}, $(m-M)_0$ = 14.69 mag.
We calculated the apparent and the true distance moduli
from RRc variable stars and
we obtained the average apparent distance modulus
$\langle (m-M)\rangle$ = 14.671 $\pm$ 0.096 mag
from all seven RRc variable stars and
14.643 $\pm$ 0.076 mag from four regular RRc variable stars.
Using $E(B-V)$\ = 0.063 mag, we obtained
$\langle (m-M)_0\rangle$ = 14.476 $\pm$ 0.096 mag
from all seven RRc variable stars and
14.448 $\pm$ 0.076 mag from four regular RRc variable stars,
which are equivalent to distances from the Sun of 7.86 $\pm$ 0.35 kpc
and 7.76 $\pm$ 0.27 kpc, respectively.
If we apply the zero-point correction term for the absolute magnitude
of RRc type variables from Fourier parameters suggested by \citet{cacciari05},
0.20 $\pm$ 0.02 mag, the true distance modulus from four RRc variable stars
becomes $(m-M)_0$\ = 14.648 $\pm$ 0.079 mag, corresponding to 8.50 $\pm$ 0.31 kpc,
for four regular RRc variable stars.
If we apply the same zero-point correction term as for RRab type variable stars
in Equation (\ref{eq:mv:3}), 0.270 $\pm$ 0.020,
the true distance modulus for NGC~6723 from 4 regular RRc variable stars becomes
$\langle (m-M)_0\rangle$ = 14.718 $\pm$ 0.079 mag,
corresponding to the distance from the Sun of 8.78 $\pm$ 0.32 kpc.
Applying the correction term of 0.270 mag for RRc type variable stars
appears to make the distance modulus over-estimated
and the correction term by \cite{cacciari05}, 0.20 mag, appears to
be more adequate for RRc type variables.
\subsubsection{The Wesenheit function from Fourier parameters: $W(BV)_0$}
\citet{kovacs98} introduced an alternative empirical method to estimate
true distance modulus of RRab type variables using the Wesenheit function;
\begin{equation}
(m-M)_0 = W(BV) - W_0(BV),
\end{equation}
where $W(BV)$ is the apparent Wesenheit function in $BV$ photometry defined to be
\begin{equation}
W(BV) = V - R_V(B-V),
\end{equation}
and we adopt $R_V$ = 3.10.
$W_0(BV)$ is the true Wesenheit function and
\citet{kovacs98} derived the following transformation relation,
\begin{equation}
W_0(BV) = 0.676 - 1.943 P + 0.315 A_1 + 0.068 \phi_{41}. \label{eq:W0}
\end{equation}
To apply this transformation relation, we used $\langle V\rangle_\mathrm{int}$\ and $\langle B-V\rangle_\mathrm{mag}$.
We show the calculated $W_0(BV)$ for individual RRab variable stars
in column (8) and the true distance modulus based on this method, $(m-M)_0$\,
in column (9) of Table~\ref{tab:rrabpam}.
We obtained $\langle (m-M)_0\rangle$ = 14.302 $\pm$ 0.037 mag,
which is equivalent to the distance from the Sun of 7.25 $\pm$ 0.12 kpc.
Our result is significantly smaller than that of \citet{harris96}.
However, the original calibration of Equation (\ref{eq:W0})
by \citet{kovacs98} was also based on the fainter magnitude of RRab variable stars
and the zero-point correction by 0.270 $\pm$ 0.020 mag is required
as we discussed above.
If we apply the zero-point correction term in Equation (\ref{eq:W0}),
the true distance modulus of NGC~6723
using the Wesenheit function by \citet{kovacs98} becomes
$\langle (m-M)_0\rangle$ = 14.572 $\pm$ 0.042 mag,
resulted in the distance from the Sun of 8.21 $\pm$ 0.16 kpc.
As can be seen in Table~\ref{tab:dm}, the distance modulus of NGC~6723 from the
Wesenheit function by \cite{kovacs98} is significantly smaller than those
from other methods. The improvement of Equation~(\ref{eq:W0}) in the future
would be very desirable.
\subsubsection{The Wesenheit function from theoretical models:
$W(BV)_{th}$ }\label{ss:W_th}
\citet{cassisi04} studied the slope and the zero-point of
the theoretical Wesenheit function incorporating evolutionary
and pulsational properties for various metallicity and HB type.
They provided the theoretical Wesenheit function $W(BV)_{th}$,
\begin{equation}
W(BV)_{th} = W_{BV}^{-0.3} + b_{BV}^{W}(\log P_\mathrm{f} + 0.3),
\end{equation}
where $P_f$ is the fundamental period and
$W_{BV}^{-0.3}$ is the $W(BV)$ value at $\log P_f = -0.3$.
By interpolating the data given in Table 3 of \citet{cassisi04},
we obtained $W_{BV}^{-0.3}$ = $-$0.329 and $b_{BV}^{W}$ = $-$2.341
for [Fe/H]\ = $-$1.23 dex and the HB type of $-$0.08 for NGC~6723.
Then the true distance modulus, $(m-M)_0$, of NGC~6723 can be written as
\begin{eqnarray}
(m-M)_0 &=& W(BV) - W(BV)_{th} \nonumber \\
&=& W(BV) + 2.341\log P_\mathrm{f} + 1.031,
\end{eqnarray}
at a given $\log P_\mathrm{f}$.
In Figure~\ref{fig:W}, we show a plot of $W(BV)$ of individual RRab variable
stars as a function of their $\log P_\mathrm{f}$.
We calculated the least square fit of the theoretical Wesenheit
function to the observed data with a fixed slope
on the $\log P_\mathrm{f}$ versus $W(BV)$ plane.
We obtained the true distance modulus of
$(m-M)_0$\ = 14.531 $\pm$ 0.061 for all 35 RRab variable stars
and 14.518 $\pm$ 0.043 for 14 regular RRab variable stars\footnote{We do not
use NV14 in our calculation because it appears to suffer from larger
differential reddening effect than other RRLs in NGC~6723.
However, as can be seen in Figure~\ref{fig:W}, the location of NV14 agrees well
with other RRLs in NGC~6723.
Since the Wesenheit function is reddening free, the location of NV14
in Figure~\ref{fig:W} strongly suggests that NV14 is a member variable star.}.
As shown in Figure~\ref{fig:W}, the regular RRab variables V14 and V15
significantly deviate from the theoretical Wesenheit function shown with
the dashed line in the plot. Note that V14 and V15 have somewhat brighter
$\langle V\rangle_\mathrm{int}$\ than other regular RRab type variable stars.
If we exclude V14, V15, V16 and NV2 (see Appendix~\ref{ap:s:notes}),
our true distance modulus of NGC~6723
based on the theoretical Wesenheit function by \citet{cassisi04} becomes
$(m-M)_0$\ = 14.530 $\pm$ 0.017 from 10 RRab variable stars
and the distance from the Sun is 8.05 $\pm$ 0.06 kpc.
The theoretical Wesenheit function of RRL can provide a powerful means
to determine the distance of GCs because it is reddening-free and, furthermore,
it incorporates evolutionary and pulsational properties for various metallicities
and HB types of stellar populations.
However, our distance modulus of NGC~6723
from the theoretical Wesenheit function by \citet{cassisi04} appears
to be slightly smaller than those from other methods (see Table~\ref{tab:dm}).
In Figure~\ref{fig:cassisi}, we show a plot of $M_V(\mathrm{RR})$\
from the absolute trigonometric parallaxes by \citet{benedict11}
as a function of metallicity along with our revised metallicity-luminosity relation
as presented in Equation (\ref{eq:rev_hstmv}).
Also shown are the theoretical mean absolute visual magnitudes of
RRLs with different HB types at fixed metallicity
\citep{cassisi04}.
As can be seen in the figure, our revised linear fit to the measured $M_V(\mathrm{RR~Lyr})$
by \citet{benedict11} does not match with the absolute visual magnitudes of RRLs
from theoretical model predictions by \citet{cassisi04}.
We performed a least-square fit to the theoretical models
using a fixed slope adopted by \citet{benedict11}, 0.214,
and we obtained the offset value in absolute visual magnitude of
$\delta M_V$ = 0.080 $\pm$ 0.130 mag, in the sense that
the theoretical model predictions by \citet{cassisi04} are fainter.
\cite{cassisi04} also noted that their theoretical prediction for
$M_V(\mathrm{RR})$\ is in good agreement with $M_V(\mathrm{RR~Lyr})$ by \cite{benedict02},
whose $M_V(\mathrm{RR~Lyr})$ is 0.08 mag fainter than that of \cite{benedict11}.
If we apply the zero-point correction term for the theoretical model
predictions by \citet{cassisi04}, the true distance modulus of NGC~6723
from the theoretical Wesenheit function becomes
$(m-M)_0$\ = 14.610 $\pm$ 0.131 mag and the distance from the Sun is 8.36 $\pm$ 0.50 kpc.
It would be very desirable to perform detailed calculation of
the Wesenheit function from the theoretical models with
an updated absolute visual magnitude scale of RRLs in the future.
\subsubsection{$M_V(RR)$ versus $V(RR)$}
Finally, we also derive the apparent and true distance moduli of NGC~6723
by using the average magnitude of the RRLs.
In Figure \ref{fig:cmdHB}, we show the CMD of the HB region
using the magnitude-weighted integrate colors and
the intensity-weighted integrated magnitudes for RRLs.
As shown in the Figure, the average magnitude of RRLs
(excluding V14, V15, V16, NV1, NV2 and NV14; see Appendix~\ref{ap:s:notes}) is
$\langle V\mathrm{(RR)}\rangle_\mathrm{int}$ = 15.459 $\pm$ 0.055 mag.
If we use the absolute visual magnitude of RRLs
from our revised metallicity-luminosity relation
presented in Equation (\ref{eq:rev_hstmv}),
$M_V\mathrm{(RR)}$ = 0.578 $\pm$ 0.133 mag at [Fe/H]\ = $-$1.23 $\pm$ 0.11 dex,
and the apparent distance modulus of NGC~6723 is
$(m-M)$ = 14.881 $\pm$ 0.143 mag.
Then the true distance modulus of NGC~6723 becomes
$(m-M)_0$\ = 14.686 $\pm$ 0.143 mag if we use $E(B-V)$\ = 0.063 mag,
corresponding to the distance from the Sun of 8.65 $\pm$ 0.57 kpc.
\medskip
In Table~\ref{tab:dm}, we summarize our derived true distance
modulus and the distance from the Sun for NGC~6723 using various methods.
Our $(m-M)_0$\ ranges from 14.57 to 14.69 mag with the unweighted mean value
of 14.64 $\pm$ 0.05 mag and
the distance of NGC~6723 from the Sun ranges from 8.21 to 8.65 kpc
with the unweighted mean value of 8.47 $\pm$ 0.19 kpc
(the errors are those of the mean).
\section{$V(RR)$ versus RADIAL DISTANCE}\label{s:vhb}
Recently, \citet{majaess12} claimed that photometry of RRLs
in GCs depends on the radial distance from the center
and the photometric contamination in the crowded central region
of GCs may affect the inferred parameters, such as distance,
absolute magnitude, etc.
With the exception of M3, the difference in the distance modulus
derived from all RRLs and using only those in the outskirt
of the cluster can be as large as 0.25 mag,
in the sense that the RRLs in the outskirt of the GCs
are fainter than those in the crowded central regions.
They found an average difference for the five GCs in their Table~1 of
about 0.1 mag.
We explore the variation of magnitude against the radial distance from
the center for all the RRLs found in NGC~6723.
We calculated the angular distance $r$ of each RRL
from the center of the cluster using the Spherical law of cosine,
\begin{equation}
r = \arccos[\sin\delta_0\sin\delta + \cos\delta_0\cos\delta\cos(\alpha_0 - \alpha)],
\end{equation}
where $\alpha_0$ and $\delta_0$ are the right ascension and the declination
of the center of NGC~6723 and $\alpha$ and $\delta$ are those of individual RRLs.
For our calculations, we adopted the coordinates for the cluster center
measured by \citet{glodsbury};
$\alpha_0$ = 18:59:33.15 and $\delta_0$ = $-$36:37:56.1.
In Figure~\ref{fig:magVSrad}, we show plots of $BV$ magnitudes of
all RRLs in NGC~6723 as functions of the radial distance.
Also shown are the average magnitudes and the dispersions in the magnitude
($\pm\sigma$ levels) in the central region ($r$ $\leq$ $r_c$, the core radius),
in the intermediate region ($r_c$ $\leq$ $r$ $\leq$ $r_h$, the half-light radius)
and in the outskirt region ($r$ $\geq$ $r_h$) of the cluster.
The $BV$ magnitudes of RRLs
from three different regions agree well within measurement
errors and the radial gradient does not appear to exist, suggesting that
the photometric contamination in the central part of NGC~6723 is not severe.
An alternative approach to examine a potential radial gradient
of RRL magnitudes can be found in Figure~\ref{fig:Wrad}, where we show
plots of $W(BV)$ versus $\log P_f$ in three different radial regions.
We calculate the distance modulus in each radial zone
using the theoretical Wesenheit function with a fixed slope.
Also shown in the figure is the theoretical Wesenheit function
with $(m-M)_0$\ = 14.531 $\pm$ 0.061 mag using all 35 RRab type variable stars
in NGC~6723 as a reference (see section \ref{ss:W_th}).
Again, the distance moduli measured from the theoretical Wesenheit function
from three different regions agree well within measurement errors
and the radial gradient cannot be seen.
To understand the radial gradient of RRL magnitude,
in Figure~\ref{fig:blend} we show
a plot of $\log \rho_c d^2$ versus $\bar{\mu} - \mu_r$
for 6 GCs from Majaess et al.\ (2012) and NGC~6723 from our current study,
where $\rho_c$ and $d$ are the central luminosity density
in units of solar luminosities per cubic parsec and
the distance from the Sun in kpc, respectively, from \citet{harris96}.
We define $\rho_c d^2$ to be the apparent crowdedness of the central region
assuming that the mass-segregation is negligible in GCs under consideration.
Also $\bar{\mu} - \mu_r$ is defined to be the distance spread between
the average computed using all 35 RRLs and only those near
the periphery (Majaess et al.\ 2012) and it is a measure
of the photometric contamination of the central region of a given GC system.
Since $\rho_c d^2$ is a measure of the apparent crowdedness
of the central region of GC systems,
it is natural to expect that as $\rho_c d^2$ increases,
the degree of photometric contamination, such as resulted from blending, increases.
Therefore one expects to have
\begin{equation}\label{eq:blend}
|\bar{\mu} - \mu_r| \propto \log \rho_c d^2
\end{equation}
in GC systems as shown in Figure~\ref{fig:blend}.
Since the LSST, for example, can detect RRLs in the local group of galaxies
out to $\approx$ 1 Mpc scale in several years of time span \citep{lsst}
and the future 30-m to 40-m class telescopes are expected to detect RRLs
in nearby groups of galaxies, the utility of RRLs
as a primary distance indicator will become more important.
A more thorough study of the relation between $|\bar{\mu} - \mu_r|$ and
$\log \rho_c d^2$ with an extended sample of GCs
or high density regions would be desirable in the future, because if real,
this relation can play a significant role to mitigate the discrepancy in
establishing cosmic distance scale using RRLs.
\section{SUMMARY}
We have presented $BV$ CCD photometry for 54 variables in the metal-rich
inner halo globular cluster NGC~6723, including 22 newly discovered variables.
We found that 13 RRLs and one SX Phe variable star are cluster member stars
and two $\delta$ Sct variable stars and five W UMa type eclipsing binary stars
are most likely off-cluster field stars.
New light curves, periods and photometric parameters for all those variables
are presented. With the discovery of new RRLs, the mean periods of the RRab and
RRc variables in the cluster are $\langle$P$_{ab}$$\rangle$ = 0.541 $\pm$ 0.066 day
from 35 RRab type variable stars and $\langle$P$_{c}$$\rangle$ = 0.292 $\pm$ 0.030 day
from seven RRc type variable stars, respectively.
The number ratio of the RRc type variable stars to the total number of
the RRLs is n(c)/n(ab+c) = 0.167.
Our new mean periods and the RRL distribution of NGC~6723
are in good agreement with those of a typical Oosterhoff group I globular cluster.
We also obtained $\langle V\rangle_\mathrm{int}$\ = 15.459 $\pm$ 0.055 mag from 30 RRab type
and seven RRc type variable stars and it is in good agreement with
previous measurements by others.
We carried out the Fourier decomposition analysis for the RRLs in NGC~6723.
Using empirical transformation relations by \citet{jk96} and \citet{morgan07},
we obtained [Fe/H]$_\mathrm{ZW}$ = $-$1.27 $\pm$ 0.09 dex from RRab
and $-$1.18 $\pm$ 0.15 dex from RRc variable stars.
The unweighted mean of the photometric metallicity of NGC~6723
in the Zinn \& West metallicity scale becomes
$\langle$[Fe/H]$_\mathrm{ZW}\rangle$ = $-$1.23 $\pm$ 0.11 dex,
consistent with the previous metallicity measurement of the cluster
from the high resolution spectroscopic study of red-giant branch stars
by \citet{fullton96}.
We investigated the interstellar reddening value by comparing
the intrinsic color using empirical transformation relations from
the Fourier and the pulsational parameters and the observed color
of individual RRab type variable stars in NGC~6723.
We also provided a new calibration of the interstellar reddening values
on the HST ACS photometric system using the ACS GC survey by \citet{ACS}.
When compared to the interstellar reddening values from
the HST ACS main-sequence photometry of NGC~6723,
the $E(B-V)$\ values from the Fourier and the pulsational parameters
of RRab type variable stars appear to slightly over-estimate
the foreground interstellar reddening value.
Furthermore, we found that the slope in the relation between
the intrinsic color derived from the Fourier and the pulsational parameters
and the observed color of RRab type variable stars is not unity.
As a consequence, the interstellar reddening values
from the Fourier and the pulsational parameters
have substantial gradient against the observed color
and compressed the derived $(B-V)_0$\ distribution of RRab type variable stars.
The later effect may explain the compressed $(B-V)_0$\ distribution of
M3 RRLs noted by \citet{cacciari05} when they used
the empirical transformation relation by \citet{kovacs99}.
Our results also suggest that derived intrinsic colors of individual variable stars
from the Fourier and the pulsational parameters may not be reliable, confirming
the similar results by \citet{cacciari05}.
For the interstellar reddening value of NGC~6723, we adopted the unweighted mean
of those from the static color $\langle B-V\rangle_\mathrm{s}$\ of RRLs of our current study
and from the isochrone fitting by \cite{ACS}, finding $E(B-V)$\ = 0.063 $\pm$ 0.015.
We discussed that RR Lyr may not be an evolved star, in contrast to
the recent Str\"omgren photometry study by \cite{catelan08} who
claimed that RR Lyr is about 0.06 mag brighter than the ZAHB.
The reason for having brighter luminosity for RR Lyr by \cite{catelan08}
is that their analysis was relied on the theoretical HB models by
\cite{catelan04}, whose HB absolute magnitude levels are 0.08 mag
fainter than the current standard value.
Using the recent HST absolute trigonometric parallax measurements of RR Lyr
by \cite{benedict11}, we obtained the following metallicity-luminosity
relation,
\begin{eqnarray}
M_V\mathrm{(RR)} &=& (0.214 \pm 0.047)(\mathrm{[Fe/H]} + 1.5) + (0.52 \pm 0.13). \nonumber
\end{eqnarray}
Our revised relation is in excellent agreement with the current
standard absolute visual magnitude of RRLs,
$M_V(\mathrm{RR})$\ = 0.52 mag at [Fe/H]\ = $-$1.50 dex.
Also our revised metallicity-luminosity relation for RRLs gives
$(m-M)_0$ = 18.54 $\pm$ 0.13 mag for LMC,
if we use the $V_0(\mathrm{RRL})$ value in LMC by \cite{clementini03}.
Our results for $M_V(\mathrm{RR})$\ of NGC~6723 confirmed that original transformation relations
of the absolute visual magnitude of RRLs by \citet{jurcsik98} and \citet{kovacs98}
are 0.270 $\pm$ 0.020 mag fainter than the current standard $M_V\mathrm{(RR)}$ value.
We calibrated the new zero-point constant of the empirical transformation
to derive the absolute visual magnitude of RRab type variable star by \citet{kovacs01}.
Using our own zero-point constant, we obtained
$\langle M_V\mathrm{(RR)}\rangle$ = 0.581 $\pm$ 0.038 mag for NGC~6723.
We then derived the distance modulus of NGC~6723 using
(i) $M_V\mathrm{(RR)}$ from the Fourier parameters,
(ii) Wesenheit functions of the clusters,
and (iii) the $M_V\mathrm{(RR)}$-[Fe/H]\ relation.
Our true distance modulus of NGC~6723 ranges from 14.57 to 14.69 mag
with the unweighted mean value of 14.65 $\pm$ 0.05 mag and
the distance of NGC~6723 from the Sun ranges from 8.21 to 8.65 kpc
with the unweighted mean value of 8.47 $\pm$ 0.17 kpc.
Finally, we examined the RRL magnitude dependencies on the radial distance
from the center of the cluster, as claimed by \citet{majaess12}.
We found no evidence of the radial gradient in the RRL magnitude in NGC~6723.
We showed that it is natural to expect to have more severe photometric
contamination with more severe degree of the apparent crowdedness.
We suggested that there appears to exist a relation between the degree
of photometric contamination and the apparent crowdedness of the central
region of globular cluster systems,
\begin{eqnarray}
|\bar{\mu} - \mu_r| &\propto& \log \rho_c d^2. \nonumber
\end{eqnarray}
If real, this relation can play a significant role to mitigate the discrepancy
in establishing cosmic distance scale using RRLs
in resolved stellar populations.
\acknowledgments
J.-W.L. acknowledges financial support from the Basic Science
Research Program (grant No. 2010-0024954) and the Center
for Galaxy Evolution Research through the National Research
Foundation of Korea.
J.-W.L. also thanks Dr.\ Kov\'acs for providing Fourier decomposition
FORTRAN programs.
We thank the anonymous referee for a thorough and beneficial review.
Some of the data presented in this paper were based on observations
made with telescopes operated by the SMARTS consortium.
|
\section{\@startsection {section}{1}{\z@}{-3.5ex plus -1ex minus
\newcommand{{\textstyle *}}{{\textstyle *}}
\newcommand{{\rm F}}[2]{F_{#1}^{#2}}
\newcommand{{\rm FQ}}[2]{{\rm PF}_{{#1\rm\mbox{-}T}}^{#2}}
\newcommand{{\rm Q}}[2]{\P_{{#1\rm\mbox{-}T}}^{#2}}
\newcommand{\FQ_\|}[2]{{\rm PF}_{{#1\rm\mbox{-}tt}}^{#2}}
\newcommand{\Q_\|}[2]{\P_{{#1\rm\mbox{-}tt}}^{#2}}
\newcommand{\P/\poly}{\P/{\rm poly}}
\newcommand{\P^{\NP[\log]}}{\P^{{\rm NP}[\log]}}
\newcommand{\pleq}[1]{\leq_{#1}^{p}}
\def\abs#1{\left|#1\right|}
\def\exp#1{{\rm exp}\left(#1\right)}
\newcommand{\transfig}[1]{\begin{center}\vspace{-6pt}\input{#1.tex}\vspace{-12pt}\end{center}}
\newif\ifshortconferences
\shortconferencesfalse
\newif\ifmediumconferences
\mediumconferencesfalse
\def\ending#1{{\count1=#1\relax
\count2=\count1
\divide\count2 by 100
\multiply\count2 by 100
\advance\count1 by -\count2
\ifnum\count1=11
th%
\else \ifnum\count1=12
th%
\else \ifnum\count1=13
th%
\else
\count2=\count1
\divide\count1 by 10
\multiply\count1 by 10
\advance\count2 by -\count1
\ifnum\count2=1
st%
\else \ifnum\count2=2
nd%
\else \ifnum\count2=3
rd%
\else th%
\fi\fi\fi\fi\fi\fi
}}
\def\conf{STOC}{\conf{STOC}}
\def\ifshortconferences ACM STOC\else\ifmediumconferences Ann. ACM Symp. Theor. Comput.\else Annual ACM Symposium on Theory of Computing\fi\fi{\ifshortconferences ACM STOC\else\ifmediumconferences Ann. ACM Symp. Theor. Comput.\else Annual ACM Symposium on Theory of Computing\fi\fi}
\def68{68}
\def\conf{FOCS}{\conf{FOCS}}
\def\ifshortconferences IEEE FOCS\else\ifmediumconferences Ann. Symp. Found. Comput. Sci.\else IEEE Symposium on Foundations of Computer Science\fi\fi{\ifshortconferences IEEE FOCS\else\ifmediumconferences Ann. Symp. Found. Comput. Sci.\else IEEE Symposium on Foundations of Computer Science\fi\fi}
\def59{59}
\def\conf{FSTTCS}{\conf{FSTTCS}}
\def\ifshortconferences FST\&TCS\else\ifmediumconferences Ann. Conf. Found. Softw. Theor. and Theor. Comp. Sci.\else Conference on Foundations of Software Theory and Theoretical Computer Science\fi\fi{\ifshortconferences FST\&TCS\else\ifmediumconferences Ann. Conf. Found. Softw. Theor. and Theor. Comp. Sci.\else Conference on Foundations of Software Theory and Theoretical Computer Science\fi\fi}
\def80{80}
\def\conf{Complexity}{\conf{Complexity}}
\def\ifshortconferences Conf. Computational Complexity\else\ifmediumconferences Ann. Conf. Computational Complexity\else Annual Conference on Computational Complexity\fi\fi{\ifshortconferences Conf. Computational Complexity\else\ifmediumconferences Ann. Conf. Computational Complexity\else Annual Conference on Computational Complexity\fi\fi}
\def85{85}
\defStructure in Complexity Theory{Structure in Complexity Theory}
\def\conf{Structures}{\conf{Structures}}
\def\ifshortconferences Ann. Conf. Structure in Complexity Theory\else\ifmediumconferences Ann. Conf. Structure in Complexity Theory\else Annual Conference on Structure in Complexity Theory\fi\fi{\ifshortconferences Ann. Conf. Structure in Complexity Theory\else\ifmediumconferences Ann. Conf. Structure in Complexity Theory\else Annual Conference on Structure in Complexity Theory\fi\fi}
\def85{85}
\defStructure in Complexity Theory{Structure in Complexity Theory}
\def\conf{SODA}{\conf{SODA}}
\def\ifshortconferences ACM-SIAM Symp. on Discrete Algorithms\else Annual ACM-SIAM Symposium on Discrete Algorithms\fi{\ifshortconferences ACM-SIAM Symp. on Discrete Algorithms\else Annual ACM-SIAM Symposium on Discrete Algorithms\fi}
\def89{89}
\def\conf{STACS}{\conf{STACS}}
\def\ifshortconferences STACS\else\ifmediumconferences\else Annual Symposium on Theoretical Aspects of Computer Science\fi\fi{\ifshortconferences STACS\else\ifmediumconferences\else Annual Symposium on Theoretical Aspects of Computer Science\fi\fi}
\def83{83}
\def\conf{SPAA}{\conf{SPAA}}
\def\ifshortconferences ACM SPAA\else\ifmediumconferences Ann. ACM Symp. Par. Alg. Arch.\else Annual ACM Symposium on Parallel Algorithms and Architectures\fi\fi{\ifshortconferences ACM SPAA\else\ifmediumconferences Ann. ACM Symp. Par. Alg. Arch.\else Annual ACM Symposium on Parallel Algorithms and Architectures\fi\fi}
\def88{88}
\def\ifshortconferences Proc.\else\ifmediumconferences Proc.\else Proceedings\fi\fi{\ifshortconferences Proc.\else\ifmediumconferences Proc.\else Proceedings\fi\fi}
\def\ifshortconferences Proc.\else\ifmediumconferences Proc.\else Proceedings of the\fi\fi{\ifshortconferences Proc.\else\ifmediumconferences Proc.\else Proceedings of the\fi\fi}
\newcounter{confnum}
\def\conf#1#2{%
\setcounter{confnum}{#2}%
\addtocounter{confnum}{-\csname #1zero\endcsname}%
\ifnum\value{confnum}=1%
\expandafter\ifx\csname #1One\endcsname\relax%
\ifshortconferences Proc.\else\ifmediumconferences Proc.\else Proceedings of the\fi\fi\ \arabic{confnum}\ending{\value{confnum}}\ \csname #1name\endcsname%
\else \csname #1One\endcsname\fi%
\else%
\ifshortconferences Proc.\else\ifmediumconferences Proc.\else Proceedings of the\fi\fi\
\arabic{confnum}\ending{\value{confnum}}\ \csname #1name\endcsname\fi}
\def\vrule width0.7ex height0.9em depth0ex{\vrule width0.7ex height0.9em depth0ex}
\newif\ifqed\qedtrue
\def\global\qedfalse{\global\qedfalse}
\def\qed{\ifqed{\penalty1000\unskip\nobreak\hfil\penalty50
\hskip2em\hbox{}\nobreak\hfil\vrule width0.7ex height0.9em depth0ex
\parfillskip=0pt \finalhyphendemerits=0\par\medskip}\fi\global\qedtrue}
\def\QEDcomment#1{\ifqed{\penalty1000\unskip\nobreak\hfil\penalty50
\hskip2em\hbox{}\nobreak\hfil\vrule width0.7ex height0.9em depth0ex\ #1
\parfillskip=0pt \finalhyphendemerits=0\par\medskip}\fi\global\qedtrue}
\makeatletter
\def\eqnqed{\global\qedfalse
\def\@tempa{equation}
\ifx\@tempa\@currenvir\def\@eqnnum{\vrule width0.7ex height0.9em depth0ex}%
\addtocounter{equation}{-1}\else%
\def\@@eqncr{\let\@tempa\relax
\ifcase\@eqcnt \def\@tempa{& & &}\or \def\@tempa{& &}%
\else \def\@tempa{&}\fi
\@tempa {\def\@eqnnum{{\vrule width0.7ex height0.9em depth0ex}}\@eqnnum
\global\@eqnswtrue\global\@eqcnt\z@\cr}\fi}
\def\eqnlabel#1#2{\if@filesw {\let\thepage\relax%
\def\protect{\noexpand\noexpand\noexpand}%
\edef\@tempa{\write\@auxout{\string
\newlabel{#2}{{{#1}}{\thepage}}}}%
\expandafter}\@tempa%
\if@nobreak \ifvmode\nobreak\fi\fi\fi%
\def\@tempa{equation}
\ifx\@tempa\@currenvir\def\theequation{{#1}
\addtocounter{equation}{-1}\else%
\def\@@eqncr{\let\@tempa\relax
\ifcase\@eqcnt \def\@tempa{& & &}\or \def\@tempa{& &}%
\else \def\@tempa{&}\fi
\@tempa {\def\@eqnnum{{#1}}\@eqnnum
\global\@eqnswtrue\global\@eqcnt\z@\cr}\fi}
\makeatother
\newcommand{\listqed}{\qed\global\qedfalse}
\def\littleqed{\ifqed{\penalty1000\unskip\nobreak\hfil\penalty50
\hskip2em\hbox{}\nobreak\hfil\vrule width0.6ex height0.6em depth0ex
\parfillskip=0pt \finalhyphendemerits=0\par\medskip}\fi\global\qedtrue}
\def\vrule width0.6ex height0.6em depth0ex{\vrule width0.6ex height0.6em depth0ex}
\def\qed{\qed}
\makeatother
\def\aftereqnskip{\vskip-3pt}
\newcommand{\SPASAM}[1]{{\rm MOL{\rm\mbox{-}} A}(#1)}
\newcommand{\SPASM}[1]{{\rm MOL'}(#1)}
\newcommand{\SPAM}[1]{{\rm MOL}(#1)}
\newcommand{\NPbits}[1]{{\rm NPbits}(#1)}
\newcommand{\NPinit}[1]{{\rm NPinit}(#1)}
\newcommand{\NPpaths}[1]{{\rm NPpaths}(#1)}
\newcommand{\bigvee}{\bigvee}
\newcommand{\set}[1]{\left\{{#1}\right\}}
\section{Introduction}
The polynomial identity testing problem (PIT) is to test whether a polynomial computed by an
arithmetic circuit is identical to zero. PIT problem plays a
significant role in the field of computational complexity. It is
known that every polynomial computed by a polynomial size circuit
can be determined if it is identical to zero by a polynomial time
randomized algorithm~\cite{Schwartz80,Zippel79}.
The results of Impagliazzo and
Widgerson~\cite{ImpagliazzoWigderson97} suggested that every
randomized polynomial time algorithm can be derandomized into a
deterministic polynomial time algorithm. They proved that $\P={\rm BPP}$
if E contains any problem that requires $2^{\Omega(n)}$ size boolean
circuits. It has been a long standing open problem in complexity
theory to separate NEXP from BPP.
Kabanets, Impagliazzo and Wigderson showed that derandomizing
Promise-BPP implies ${\rm NEXP}\not\subseteq
{\rm P/poly}$~\cite{ImpagliazzoKabanetsWigderson}. Building upon the work
of Kabanets, Impagliazzo and
Wigderson~\cite{ImpagliazzoKabanetsWigderson}, Kabanets and
Impagliazzo~\cite{KabanetsImpagliazzo04} proved that to derandomize
the polynomial identity testing problem in the integer ring, one
must prove that ${\rm NEXP}$ has no polynomial size boolean circuits or
permanent has no polynomial size arithmetic circuits.
The proof of Kabanets and Impagliazzo's theorem was
simplified by Aaronson and Melkebeek~\cite{AaronsonMelkebeek10}.
Many papers have been
published toward the derandomization of the polynomial identity
problems~(see for examples,
\cite{AgrawalBiswas03,KlivansSpielman01,KabanetsImpagliazzo04,LewinVadhan98,ChenKao00,ShpilkaVolkovich08,LiptonVishnoi03,Zippel79,Schwartz80,RazShpilk05,KayalSaxena07,KarninShpilka07,DvirShpilk07a,Saxena08}).
We have not found any existing result that shows derandomization of
PIT over a finite field implies ${\rm NEXP}\not\subseteq {\rm P/poly}$. It is
essential to identify the connection between derandomizing PIT over
finite fields and complexity classes separation. In this paper, we
study the implication of polynomial identity problem over finite
fields to separations in computational complexity theory. Our
results are derived without using permanent problem, and give a
direct implication for complexity classes separation via
derandomization of PIT.
We show that for any finite field $F$, if ${\rm PIT}$ over $F$ is in
${\rm NSUBEXP}$, then ${\rm NEXP}\not\subseteq {\rm P/poly}$. We also show that if
${\rm PIT}$ over a finite field $F$ is in ${\rm NP}$, then ${\rm NTIME}(n^{\log
n})\not\subseteq {\rm PH}\cap {\rm P/poly}$. It implies that if there exists a
polynomial time deterministic algorithm for polynomial identity
testing problem over any finite field $F$, then ${\rm NTIME}(n^{\log
n})\not\subseteq {\rm BPP}$.
Our proof is different from Kabanets and
Impagliazzo's~\cite{KabanetsImpagliazzo04} work that is for PIT over
$Z$.
Their methods involve
permanent that is $\#\P$-hard~\cite{Valiant},
and Toda's theorem ${\rm PH}\subseteq \P^{\#\P}$~\cite{TodaPPPH}. Our
method works for the PIT over any finite fields, but it does not
imply that derandomizing PIT over $Z$ separates NEXP from ${\rm P/poly}$.
Therefore, Kabanets and Impagliazzo's~\cite{KabanetsImpagliazzo04}
work and this paper are complementary to each other to support the
importance of derandomizing the polynomial identity testing problem,
and its implication to nonuniform lower bounds.
\section{Notations}
A {\it boolean circuit} is a circuit with AND ($\bigwedge$), OR
($\bigvee$), and NEGATION ($\bar{x}$) gates with fan-in at most
two, and no feedback. An {\it arithmetic circuit} is a circuit with
$+,-$ and $*$ gates over a finite field. The {\it size} of a circuit
$C(.)$ is the number of gates and is denoted by $|C(.)|$.
The {\it polynomial identity testing problem} is to test if a
polynomial computed by an arithmetic circuit is identical to zero.
We use ${\rm PIT}_q$ to represent the polynomial identity testing problem
over the field $F(q)$ of size $q$. Let ${\rm PIT}_Z$ be the polynomial
identity testing problem over the integers $Z$, which is an integral
ring.
The basic knowledge of algebra can be found in standard algebra
textbooks such as~\cite{Hungerford74}. Every finite field $F(q)$ of
size $q$ has $q=p^k$ for some prime number $p$ and integer $k$. For
an element $a$ in a finite field $F(q)$, its {\it order} is the
least integer $r\ge 1$ with $a^r=1$.
The permanent maps square matrices to values. Let
$A=(a_{i,j})_{n\times n}$ be an $n\times n$ matrix over integers.
Define permanent to be the function
${\rm perm}(A)=\sum_{\sigma}\prod_{i=1}^n a_{i,\sigma(i)}$, where
$\sigma$ is over all permutations of $1,2,\cdots, n$.
Define $N=\{0,1,2,\cdots\}$ to be the set of nonnegative integers.
Let $t(n):N\rightarrow N$ be a nondecreasing function. Define
${\rm NTIME}(t(n))$ to be the class of languages accepted by nondeterministic
Turing machines in time $O(t(n))$.
For a function $f(n): N\rightarrow N$, it is {\it time
constructible} if given an integer $n$, $f(n)$ can be computed in
$O(f(n))$ steps by a deterministic Turing machine.
Theorem~\ref{Zak-theorem} is a separation of nondeterministic
complexity classes due to Zak\cite{Zak83}.
\begin{theorem}[\cite{Zak83}]\label{Zak-theorem}
If $t_1(.)$ and $t_2(.)$ are time-constructible nondecreasing
functions from $N$ to $N$, and $t_1(n+1)=o(t_2(n))$, then
${\rm NTIME}(t_2(n))$ is strictly contained in ${\rm NTIME}(t_1(n))$.
\end{theorem}
Assume that $M(.)$ is an oracle Turing machine. A decision
computation
$M^A(x)$ returns either $0$ or $1$ when the input is $x$ and
oracle is $A$. For a class $C$ of languages, we use ${\rm NP}^C={\rm NP}_{\rm
T}(C)$ to represent the class of languages that can be reducible to
the languages in $C$ via polynomial time nondeterministic Turing
reductions. Define ${\rm NEXP}=\cup_{c=1}^{\infty} {\rm NTIME}(2^{n^c})$ and
${\rm NP}=\cup_{c=1}^{\infty} {\rm NTIME}(n^c)$.
Let ${\rm PH}$ be the class of polynomial time
hierarchy~\cite{Stockmeyer} ${\rm PH}=\cup_{i=1}^{\infty} \sum_i^{\rm
P}$, where $\sum_1^{\rm P}={\rm NP}$, and $\sum_{i+1}^{\rm
P}={\rm NP}^{\sum_i^{\rm P}}$ for all $i\ge 1$. Define the subexponential
time nondeterministic class to be
${\rm NSUBEXP}=\cap_{\epsilon>0}{\rm NTIME}(2^{n^{\epsilon}})$. Define ${\rm P/poly}$
to be the class of languages that have nonuniform polynomial size
circuits. BPP, which stands for bounded-error probabilistic
polynomial time, is the class of decision problems solvable by a
probabilistic Turing machine in polynomial time, with an error
probability of at most 1/3 for all instances.
An {\it instance of {\rm 3SAT}} is a 3CNF that is a conjunction of
clauses of at most three literals. For example, $(x_1\bigvee
\overline{x_2}\bigvee x_3)\bigwedge (\overline{x_1}\bigvee
\overline{x_2}\bigvee x_3) \bigwedge (x_1\bigvee
\overline{x_4}\bigvee x_5)$. A formula is said to be satisfiable if
it can be made true by assigning appropriate logical values (i.e.
TRUE (1), FALSE(0)) to its variables. The {\rm 3SAT}~is, given a
3CNF, to check whether it is satisfiable. It is well known that
{\rm 3SAT}~ is NP-complete problem~\cite{Cook-NP-complete}. The
number of variables of a {\rm 3SAT}~instance and its length is
polynomially related.
\section{Our Results}
Theorem~\ref{main-thm} is the main theorem of this paper. It will be
proved in Section~\ref{proof-section}. Theorem~\ref{main-thm} is
stated in a format so that we have a self-contained proof. Some
corollaries that involve some existing results are stronger than the
main theorem.
\begin{theorem}\label{main-thm}
Let $t(n)$ and $t'(n)$ be time constructible nondecreasing
superpolynomial functions from $N$ to $N$ with $t'(n+1)=o(t(n))$.
Let $h(n)$ be an nondecreasing function from $N$ to $N$ such that
for every fixed $c>0$, $h(n^{c})+n^c\le t'(n)$ for all large $n$.
Let $F(q)$ be a finite field of size $q$. If ${\rm PIT}_q\in {\rm NTIME}(h(n))$,
then ${\rm NTIME}(t(n))\not\subseteq {\rm NP}^{{\rm NP}}\cap {\rm P/poly}$.
\end{theorem}
Using the main theorem, we have some corollaries. Their proofs need
to combine Theorem~\ref{main-thm} with some existing well known
theorems in the computational complexity theory.
\begin{corollary}\label{coro-first-corollary}
Let $t(n)$ and $t'(n)$ be time constructible nondecreasing
superpolynomial functions from $N$ to $N$ with $t'(n+1)=o(t(n))$.
Let $h(n)$ be a nondecreasing function from $N$ to $N$ such that for
every fixed $c>0$, $h(n^{c})+n^c\le t'(n)$ for all large $n$. Let
$F(q)$ be a finite field of size $q$. If ${\rm PIT}_q\in {\rm NTIME}(h(n))$, then
${\rm NTIME}(t(n))\not\subseteq {\rm PH}\cap {\rm P/poly}$.
\end{corollary}
\begin{proof}
Assume ${\rm NTIME}(t(n))\subseteq {\rm PH}\cap {\rm P/poly}$. By Karp and Lipton's
thereom~\cite{KarpLipton}, we have ${\rm PH}=\sum_2^P={\rm NP}^{{\rm NP}}$. It
follows from Corollary~\ref{coro-first-corollary}.
\end{proof}
\begin{corollary}
If ${\rm PIT}_q\in {\rm NSUBEXP}$ for a finite field $F(q)$, then
${\rm NEXP}\not\subseteq {\rm P/poly}$.
\end{corollary}
\begin{proof}
Assume that ${\rm PIT}_q\in {\rm NSUBEXP}$ and ${\rm NEXP}\subseteq {\rm P/poly}$. By
Impagliazzo, Kabanets, and Wigderson's
theorem~\cite{ImpagliazzoKabanetsWigderson}, ${\rm NEXP}={\rm PH}$. We have a
contradiction by Theorem~\ref{main-thm}.
\end{proof}
\begin{corollary}\label{NEXP-BPP-corollary}
If ${\rm PIT}_q\in {\rm NSUBEXP}$ for a finite field $F(q)$, then ${\rm NEXP}\not=
{\rm BPP}$.
\end{corollary}
\begin{proof}
Assume that ${\rm PIT}_q\in {\rm NSUBEXP}$ and ${\rm NEXP}={\rm BPP}$. It is well known
that Adleman~\cite{Adleman} proved
${\rm BPP}\subseteq {\rm NP}^{{\rm NP}}$. We have a contradiction by Theorem~\ref{main-thm}.
\end{proof}
\begin{corollary}
If ${\rm PIT}_q\in {\rm NP}$ for a finite field $F(q)$, then ${\rm NTIME}(n^{\log
n})\not\subseteq {\rm BPP}$.
\end{corollary}
\begin{proof}
It is similar to the proof of Corollary~\ref{NEXP-BPP-corollary}.
\end{proof}
\section{Overview of Our Method}
In this section, we give a brief review of our method. The main
theorem will be proved by contradiction.
A special version of our main
theorem is formulated as ${\rm PIT}_2\in {\rm NP}\Rightarrow
{\rm NEXP}\not\subseteq {\rm NP}^{{\rm NP}}\cap {\rm P/poly}$.
Assume ${\rm PIT}_2\in {\rm NP}$ and ${\rm NEXP}\subseteq {\rm NP}^{{\rm NP}}\cap {\rm P/poly}$.
Let $K$ be a complete language of the class ${\rm NEXP}$ and let
{\rm 3SAT}~be computed by a polynomial size circuit $C(.)$. Our main
technical contribution is a method that transforms a boolean circuit
$C(.)$ into an arithmetic circuit $A^*_C(.)$ over a finite field
$F(q)$ such that $C(.)$ decides {\rm 3SAT}~if and only if $A^*_C(.)$
is identical to zero.
As the {\rm 3SAT}~problem is not arithmetically defined as permanent.
If each instance of {\rm 3SAT}~is encoded as a binary string that
will be easy to decode, then there are some binary strings that do
not encode valid instances of {\rm 3SAT}. In other words, a mapping
from instances of {\rm 3SAT}~to binary strings may not be both
one-one and onto.
We construct a
special polynomial size arithmetic function $G(Y)$ that is zero if $Y$ is not an instance of {\rm 3SAT},
and nonzero otherwise. For an instance $f(x_1,x_2,\cdots,x_n)$ of {\rm 3SAT},
it is satisfiable if and only if at least one of $f(0,x_2,\cdots,x_n)$ and
$f(1,x_2,\cdots,x_n)$ is satisfiable. This recursive relation is
also converted into an arithmetic circuit $A^*_C(.)$ as PIT problem to verify whether $C(.)$ decides {\rm 3SAT}. The arithmetic circuit $A^*_C(.)$ is expressed as $H(f)G(f)$. The arithmetic circuit $H(f)$ is used to verify
the recursive relationship of
circuit $C(.)$ for deciding {\rm 3SAT}. As the input of $A^*_C(.)$ has many cases that do not encode any instance of {\rm 3SAT}, the function $G(.)$ has a value zero to pass the identity testing among those cases.
We will show that $K$ can be computed by $M^{{\rm 3SAT}}(.)$ for a
polynomial time oracle Turing machine $M(.)$. A polynomial time
nondeterministic computation will be derived to compute $K$. A
circuit $C(.)$ will be guessed and is checked via converting to
${\rm PIT}_2$, which is verified again in a nondeterministic polynomial
time. Thus, we have $K\in {\rm NP}$. This contradicts the well know
nondeterministic computational complexity hierarchy, which is stated
in Theorem~\ref{Zak-theorem} and implies ${\rm NEXP}\not= {\rm NP}$.
This approach lets us obtain lower bound for NEXP under the
existence of the derandmoization of PIT over an arbitrary finite
field without using permanent that is $\#\P$ hard. This paper has
almost self-contained proof for the main theorem. A reader is able
to understand our main theorem just by knowing
Theorem~\ref{Zak-theorem} and that {\rm 3SAT}~is
NP-complete~\cite{Cook-NP-complete}, which can be found in a
standard textbook of theory of computation.
\section{Proof of Main Theorem}\label{proof-section}
In this section, we derive our main theorem. Some lemmas are
provided to convert boolean circuits into arithmetic circuits. It is
divided into several subsections to prove the main theorem.
\subsection{From Boolean Circuits to Arithmetic Circuits}
In this section, we show how to transform boolean circuits into an
arithmetic circuits.
For a finite field $F(q)$, we have the following property that is
often called ``Fermat Little Theorem" for the case that $q$ is a
prime number. Its proof can be found in a standard algebra textbook.
For completeness, its proof is included here.
\begin{lemma}\label{classical-lemma}
Let $F(q)$ be a finite field. For any $a\in F(q)-\{0\}$,
$a^{q-1}=1$.
\end{lemma}
\begin{proof} Assume that $F$ is a finite field.
Let $[a]=\{a, a^2, a^3,\cdots, \}$ be the set of elements in $F(q)$
generated by $a$. $([a],.)$ forms a subgroup of $F(q)^*=F(q)-\{0\}$,
where ``." is the multiplication operation over field $F(q)$.
Therefore, the order $r$ of $a$ is the size of $[a]$. Therefore, $r$
is a divisor of $q-1$. So, $a^r=a^{q-1}=1$.
\end{proof}
We give Lemma~\ref{encoding-lemma} to convert an instance
of~{\rm 3SAT}~into a binary string. We give
Definition~\ref{normalized-def} to normalize the input of an
instance of~{\rm 3SAT}.
\begin{definition}\label{normalized-def}
Assume that an instance $C_1\bigwedge C_2\bigwedge\cdots \bigwedge C_m$ of {\rm 3SAT}~ of $n$ variables and satisfies the conditions
below:
\begin{enumerate}[1.]
\item
each $C_i$ has at most three literals,
\item
no variable appears in two literals of the same clause,
\item
all
clauses have a different set of literals, and
\item
its $n$ variables are $x_1,x_2,\cdots,x_n$ that are indexed from $1$
to $n$
\end{enumerate}
We have the following definitions:
\begin{itemize}
\item
Define $E_l(x_i)=(i, 1)$ and
$E_l(\overline{x_i})=(i, 0)$.
\item
Define $E_c((y_i\bigvee y_j\bigvee
y_k)))=(E_l(y_i), E_l(y_j), E_l(y_k))$ for each
clause
$(y_i\bigvee y_j\bigvee y_k)$.
\item
Define the normalized representation of $C_1\bigwedge
C_2\bigwedge\cdots \bigwedge C_m$ of {\rm 3SAT}~to be $(E_c(C_1),
E_c(C_2),\cdots, E_c(C_m))$.
\item
The logical value TRUE ($1$) is treated as special instance of 3SAT,
and we define its normalized representation to be $(1)$.
\item
The logical value FALSE ($0$) is treated as special instance of
3SAT, and we define its normalized representation to be $(0)$.
\end{itemize}
\end{definition}
\begin{lemma}\label{encoding-lemma}
Assume an instance $f$ of {\rm 3SAT}~ is a normalized representation
as Definition~\ref{normalized-def}.
\begin{enumerate}
\item\label{case1}
There is a polynomial time encoding method $E(.)$ such that given an
instance $f$ of at most $n$ variables of {\rm 3SAT}, $E(n,f)$ is a
0,1-string of length $8n^4$.
\item\label{case2}
There is a polynomial time decoding method $D(.)$ such that given a
0,1-string $s=E(n,f)$ for some instance $f$ with at most $n$
variables of {\rm 3SAT}, $D(s)=f$.
\item\label{case3}
There is a polynomial time algorithm $H(.)$ such that $H(1^n)$
generates a polynomial size boolean circuit $V_n(.)$ such that given
a
0,1-string $s$ of length $8n^4$, $V_n(s)= 1$ if $s=E(n,f)$ for some instance of at most $n$ variables of {\rm 3SAT}, and 0 otherwise.
\end{enumerate}
\end{lemma}
\begin{proof} We prove the three statements below:
Statement~\ref{case1}: Given a normalized representation an instance
of {\rm 3SAT}, just replace each symbol with ASCI table to transfer
it into a binary string. Append $10^k$ for some $k$ so that the
total length is exactly equal to $8n^4$. Each 3CNF instance has at
most $24{n\choose 3}<4n^3$ different clauses. $8n^4$ binary bits are
enough to encode any 3CNF instance of at most $n$ variables.
Statement~\ref{case2}: It is straight forward to decode the binary
string into an instance of {\rm 3SAT}~ by using the ASCI table.
Statement~\ref{case3}: With a polynomial time, we can check if a
binary string is a binary string to encode an valid instance of a
{\rm 3SAT}. It can be converted into a polynomial size boolean
circuit.
\end{proof}
\begin{definition}
Let $C(x_1,x_2,\cdots, x_n):\{0,1\}^n\rightarrow \{0,1\}$ be a
boolean circuit, and $A(y_1,y_2,\cdots, y_n): F(q)^n\rightarrow
F(q)$ be an arithmetic circuit over a finite field $F(q)$. We say
$C(.)$ and $A(.)$ are {\it equivalent } if for any $a_1,a_2,\cdots,
a_n\in \{0,1\}$, $C(a_1,a_2,\cdots, a_n)=0 \Leftrightarrow
A(a_1,a_2,\cdots, a_n)=0$ in the field $F(q)$; and
$C(a_1,a_2,\cdots, a_n)=1 \Leftrightarrow A(a_1,a_2,\cdots, a_n)=1$
in the field $F(q)$.
\end{definition}
The following Lemma~\ref{bool-arithmetic-lemma} shows how a boolean
circuit is converted into an equivalent arithmetic circuit with a
similar size.
\begin{lemma}\label{bool-arithmetic-lemma}
For any boolean circuit $C(.)$, then there is an equivalent
arithmetic circuit $A_C(.)$ over a field $F(q)$ such that
$|A_C(.)|=O(|C(.)|)$. Furthermore, $A_C(.)$ can be constructed from
$C(.)$ in a polynomial time of $|C(.)|$.
\end{lemma}
\begin{proof} We just show how to simulate the three AND, OR, and
NOT gates in a boolean circuit with arithmetic operations. The
arithmetic circuit is constructed by simulating the boolean circuit
$C(.)$ gate by gate. For an AND operation $a\bigwedge b$, it can be
converted into product $a\cdot b$ over $F(q)$. For an OR operation
$a \bigvee b$, it can be converted into $1-(1-a)(1-b)$. For an NOT
operation $\neg{a}$, it is converted into $1-a$. Since each gate in
$C(.)$ is transformed into
$O(1)$ gates in $A_C(.)$, we have $|A_C(.)|=O(|C(.)|)$. It is easy to
see that the total time to construct $A_C(.)$ is a polynomial time
of $|C(.)|$.
\end{proof}
\begin{definition}\label{nor-encoding-def}
Let $f$ be a normalized representation of an instance of {\rm 3SAT}.
Define $E(f)$ to be the {\it normalized binary encoding} of $f$,
where $E(.)$ is as defined in Lemma~\ref{encoding-lemma}. Define
$one_n=E(n, (1))$ and $zero_n=E(n,(0))$ for the normalized binary
representation of true and false respectively, where $E(.)$ is given
in Lemma~\ref{encoding-lemma}.
\end{definition}
\begin{lemma}\label{exclude-lemma}
Let $F(q)$ be a fixed finite field. Then there is a polynomial time
algorithm that given an unary integer $1^n$, it generates an
arithmetic circuit $G_n(x_1,x_2,\cdots,x_m)$ with $m=8n^4$ such that
\begin{enumerate}
\item
$G_n(x_1,x_2,\cdots,x_m)=0$ if at least one of $x_1,x_2,\cdots, x_m$
is not in $\{0,1\}$;
\item
$G_n(x_1,x_2,\cdots,x_m)=0$ if $x_1x_2\cdots x_m$ is not a
normalized binary encoding of an instance of {\rm 3SAT}~ with at most
$n$ variables; and
\item
$G_n(x_1,x_2,\cdots,x_m)\not=0$ if $x_1x_2\cdots x_m$ is a
normalized binary encoding of an instance of {\rm 3SAT}~ with at most
$n$ variables.
\end{enumerate}
\end{lemma}
\begin{proof} By Lemma~\ref{encoding-lemma}, we let $V_n(f)$ be a boolean circuit such that $V_n(f)\not=0$ if and only if
$f$ is a normalized binary encoding of an instance of~{\rm 3SAT}. Let
$A_V(.)$ be the arithmetic circuit defined by
Lemma~\ref{bool-arithmetic-lemma}.
Define $R(x)=1-(x(x-1))^{q-1}$.
It is easy to see that $R(x)\not=0$ if and only if $x\in \{0,1\}$ by
Lemma~\ref{classical-lemma}.
Finally, we define $G_n(x_1,x_2,\cdots,x_m)=R(x_1)R(x_2)\cdots
R(x_m)A_{V_n}(x_1,x_2,\cdots, x_m)$. It is easy to see that $G_n(.)$
satisfies expected properties.
\end{proof}
\begin{lemma}\label{reducing-lemma}
Assume that each input instance of {\rm 3SAT}~ is a normalized binary
encoding (see Definition~\ref{nor-encoding-def}. Then there is a
polynomial time algorithm such that given $1^n$, it generates
$n^{O(1)}$ size arithmetic circuits $S_{n,0}(.)$, and $S_{n,1}(.)$
such that the following are satisfied:
\begin{enumerate}
\item
$S_{n,0}(f)$ generates a normalized binary encoding for $g(0,
x_2,\cdots, x_k)$ if $f$ is a normalized binary encoding of a
{\rm 3SAT}~ instance $g(x_1, x_2,\cdots, x_k)$ with $0\le k\le n$;
\item
$S_{n,1}(f)$ generates a
normalized binary encoding for $g(1, x_2,\cdots, x_k)$ if $f$ is a
normalized binary encoding of a {\rm 3SAT}~ instance $g(x_1,
x_2,\cdots, x_k)$ with $0\le k\le n$; and
\item
$S_{n,i}(f)=f$ for $i\in \{0,1\}$ and $f\in\{zero_k,one_k\}$ with
$0\le k\le n$.
\end{enumerate}
\end{lemma}
\begin{proof}
It is easy to see that there is a polynomial time algorithm to
generate the formulas $g(0, x_2,\cdots, x_k)$, $g(1, x_2,\cdots,
x_k)$ with $0\le k\le n$.
Thus, we can get a boolean circuits to generate
them. By Lemma~\ref{bool-arithmetic-lemma}, we can get the
equivalent arithmetic circuits to generate them respectively.
\end{proof}
\subsection{From Arithmetic Circuits to PIT}
In this section, we show how to convert the arithmetic expressions
developed in the last section and a circuit for {\rm 3SAT}~into a PIT
problem.
The following Lemma~\ref{trans-arithm-lemma} shows how to use the
PIT problem over a finite field to check if a boolean circuit
decides {\rm 3SAT}. It transform a boolean circuit into an arithmetic
circuit in a polynomial number of steps.
\begin{lemma}\label{trans-arithm-lemma}Let $F(q)$ be a fixed finite
field. Then there is a polynomail time algorithm such that given a
circuit $C_n(.)$, it generates another arithmetic circuit
$A^*_{C_n}(.)$ over a finite field $F(q)$ such that $C_n(.)$
decides instances for {\rm 3SAT}~with at most $n$ variables if and
only if $A^*_{C_n}(.)$ is identical to zero.
\end{lemma}
\begin{proof} We assume that all instances of {\rm 3SAT}~ with at most $n$
variables have normalized binary encoding of length $8n^4$ as input
for $C_n(.)$. Let $S_{n,0}(.)$ and $S_{n,1}(.)$ be defined as in
Lemma~\ref{reducing-lemma}. Let $A_{C_n}(.)$ be the arithmetic
circuit that is equivalent to $C_n(f)$ by
Lemma~\ref{bool-arithmetic-lemma}. Let $one_n$ be the normalized
binary encoding of logical constant TRUE $(1)$, and let $zero_n$ be
the normalized binary encoding of logical constant FALSE $(0)$ (see
Definition~\ref{nor-encoding-def}). Let $y_0, y_1, y_2$
be new variables that do not appear in $A_{C_n(.)}$. We have the
arithmetic circuit
\begin{eqnarray*}
H(f, y_0, y_1, y_
)=&&
y_0(A_{C_n}(one_n)-1)+y_1A_{C_n}(zero_n)+\\
&&y_2(A_{C_n}(f)-(1-(1-A_{C_n}(S_{n,0}(f)))(1-A_{C_n}(S_{n,1}(f)))).
\end{eqnarray*}
Let $G_n(.)$ be the arithmetic circuit defined by
Lemma~\ref{exclude-lemma}. Define $A^*_{C_n}(f, y_0, y_1, y_2)=H(f,
y_0, y_1, y_2)G_n(f)$.
Assume that circuit $C_n(.)$ decides {\rm 3SAT}~for all instance of
at most $n$ variables, and takes the normalized binary encoding of
length $8n^4$ as input. For each normalized binary encoding $f$ of
an instance of {\rm 3SAT}, we have $H(f, y_0, y_1, y_2)=0$. This is
because recursive relation for each decider of {\rm 3SAT}. If $f$ is
not a normalized binary encoding of a valid instance of {\rm 3SAT},
we have $G_n(f)=0$. Therefore, $A^*_{C_n}$ is identical to zero.
Assume that $A^*_{C_n}$ is identical to zero. We need to verify that
${C_n}(.)$ is a circuit for {\rm 3SAT}. For each valid instance $f$
with at most $n$ variables of {\rm 3SAT}, we have $G_n(f)\not=0$ by
Lemma~\ref{exclude-lemma}. Thus, $H(f, y_0, y_1, y_2)=0$. It
confirms the $C_n(.)$ satisfies the recursive relation for a
{\rm 3SAT}~decider. For each instance $g(x_1,\cdots, x_n)$ with $n$
variables for {\rm 3SAT}, let $a_1,\cdots, a_k\in \{0,1\}$ be an
assignment for its first $k$ variables with $0\le k\le n$. We can
still find a normalized binary encoding $f_k$ for $g(a_1,\cdots,a_k,
x_{k+1},\cdots, x_n)$ and $f_k$ has length $8n^4$. Since
$G_n(f_k)\not=0$ and $H(f_k, y_0, y_1, y_2)=0$, $C_n(.)$ satisfies
the recursive relation for a {\rm 3SAT}~decider at the cases
$g(a_1,\cdots,a_k, x_{k+1},\cdots, x_n)$ for all $0\le k\le n$..
We have that ${C_n}(.)$ is a circuit for {\rm 3SAT}~if and only if
$A^*_{C_n}(f, y_0, y_1, y_2)$ is zero since it verifies if the
circuit ${C_n}(.)$ satisfies the recursion for~{\rm 3SAT}~instance
satisfiability.
\end{proof}
\subsection{From Derandomization to Separations}
In this section, we show that derandomizing PIT over a finite field
implies separation of computational complexity classes. The proof of
main theorem is given here.
\begin{proof}[Theorem~\ref{main-thm}]
Assume ${\rm NTIME}(t(n))\subseteq {\rm NP}^{{\rm NP}}\cap {\rm P/poly}$.
Let $K$ be an arbitrary language of ${\rm NTIME}(t(n))$ under polynomial
time many-one reduction. Let $M^{{\rm 3SAT}}(.)$ be a polynomial time
nondeterministic Turing machine to accept $K$, and runs in a
polynomial time $p(n)$ that is nondecreasing function from $N$ to
$N$. Let $N(.)$ be an $O(h(n))$ time nondeterministic Turing machine
that decides ${\rm PIT}_q$. We have the following nondeterministic
algorithm for $K$.
\vskip 10pt
{\bf Nondeterministic Algorithm for $K$}
Input $x$ of length $n$,
\begin{enumerate}[1.]
\item
\qquad Guess a circuit $C_{p(n)}(.)$ for deciding the instances of
variables at most $p(n)$ for {\rm 3SAT};
\item
\qquad Generate an arithmetic circuit $A^*_{C_{p(n)}}(.)$ ${\rm PIT}_q$
problem to verify $C_{p(n)}(.)$ by Lemma~\ref{trans-arithm-lemma};
\item\label{verify-step}
\qquad Run $N(A^*_{C_{p(n)}}(.))$ nondeterministically to decide if
$A^*_{C_{p(n)}}(.)$ is identical to zero;
\item
\qquad Nondeterministically select a path $P^*$ in
$M^{{\rm 3SAT}}(x)$;
\item
\qquad If step~\ref{verify-step} is successful, use
$A^*_{C_{p(n)}}(.)$ to answer all the queries to~{\rm 3SAT}~in path
$P^*$;
\item
\qquad Output yes, if $P^*$ accepts;
\end{enumerate}
{\bf End of Algorithm}
\vskip 10pt
Since $M(.)$ runs in polynomial time $p(n)$, the instance of queries
made by $M(.)$ has at most $p(n)$ variables. Since ${\rm NTIME}(t(n))\in
{\rm P/poly}$ implies {\rm 3SAT}$\in {\rm P/poly}$, there is a polynomial size
boolean circuit $C_{p(n)}(.)$ to decide {\rm 3SAT}~ for all instances
with at most $p(n)$ variables. Let $q_1(n)$
be a polynomial with $|C_{p(n)}(.)|\le q_1(n)$. We have an
arithmetic circuits $A^*_{C_{p(n)}}(.)$ that is equivalent with
$C_{p(n)}(.)$ by Lemma~\ref{bool-arithmetic-lemma}. We also have $|A^*_{C_{p(n)}}(.)|\le
q_2(n)$ for some polynomial $q_2(n)$. Step~\ref{verify-step} in the
algorithm takes $O(h(q_2(n)))$ nondeterministic steps. For some constant $c>0$, the entire
computation is in $O(h(n^c)+n^c)=O(t'(n))$ nondeterministic steps.
The nondeterministic algorithm above shows that $K$ is in
${\rm NTIME}(t'(n))$. Since $K$ is an arbitrary language in ${\rm NTIME}(t(n))$, we
have ${\rm NTIME}(t(n))\subseteq {\rm NTIME}(t'(n))$. This contradicts the well
known hierarchy theorem (see Theorem~\ref{Zak-theorem}) for
nondeterministic computation classes.
\end{proof}
\section{Generalization to Bounded Depth Circuits}
In this section, we consider the problem for the ${\rm PIT}$ with bounded
depth arithmetic circuits, and its connection to the
super-polynomial lower bounds of bounded depth boolean circuits. It
is an open problem to prove ${\rm NEXP}\not\subseteq {\rm NC1/poly}$, where
${\rm NC1/poly}$ is the class of languages that have polynomial size $O(\log
n)$-bounded depth boolean circuits.
\begin{definition}
Let $d(n)$ be a function from $N$ to $N$. Let $F(q)$ be a finite
field of size $q$. Define ${\rm PIT}_q(d(n))$ to be the polynomial
identity testing problem that decides if a polynomial computed by an
arithmetic circuit of depth at most $d(n)$ is identical to zero over
field $F(q)$.
\end{definition}
\begin{definition}
Let $d(n)$ be a function from $N$ to $N$. A {\it $d(n)$-bounded
depth} boolean circuits is the class of boolean circuits that
consists of AND, OR, and NOT gates with unbounded fan-in for AND and
OR gates. Define {\it Depth$(d(n))$-PC} to be the class of languages
that have polynomial size $d(n)$-bounded depth boolean circuits.
\end{definition}
\begin{theorem}\label{main2-thm}
Let $t(n)$ and $t'(n)$ be time constructible nondecreasing
superpolynomial functions from $N$ to $N$ with $t'(n+1)=o(t(n))$.
Let $h(n)$ be a nondecreasing function from $N$ to $N$ such that for
every fixed $c>0$, $h(n^{c})+n^c\le t'(n)$ for all large $n$. Let
$F(q)$ be a finite field of size $q$. Let $d(n)$ be a function from
$N$ to $N$ with $d(n)\ge \log n$. If ${\rm PIT}_q(O(d(n))\in {\rm NTIME}(h(n))$,
then ${\rm NTIME}(t(n))\not\subseteq {\rm NP}^{{\rm NP}}\cap $ Depth$(d(n))$-PC.
\end{theorem}
\begin{proof}[Sketch]
The proof is similar to that of Theorem~\ref{main-thm}. We need to
have a similar lemma like Lemma~\ref{bool-arithmetic-lemma} to show
that a bounded depth $k$ boolean circuit has an equivalent bounded
depth $O(k)$ arithmetic circuit. With $d(n)\ge \log n$, we also have
a lemma similar to Lemma~\ref{trans-arithm-lemma} such that a
bounded $d(n)$ depth boolean
circuit for~{\rm 3SAT}~can be converted into a
${\rm PIT}_q(O(d(n)))$ problem to verify it. This is because $S_{n,0}(.)$
and $S_{n,1}(.)$ have polynomial size $O(\log n)$-depth boolean
circuits.
\end{proof}
\begin{corollary}
Let $t(n)$ and $t'(n)$ be time constructible nondecreasing
superpolynomial functions from $N$ to $N$ with $t'(n+1)=o(t(n))$.
Let $h(n)$ be an nondecreasing function from $N$ to $N$ such that
for every fixed $c>0$, $h(n^{c})+n^c\le t'(n)$ for all large $n$.
Let $d(n)$ be a function from $N$ to $N$ with $d(n)\ge \log n$. Let
$F(q)$ be a finite field of size $q$. If ${\rm PIT}_q(O(d(n)))\in
{\rm NTIME}(h(n))$, then ${\rm NTIME}(t(n))\not\subseteq {\rm PH}\cap $
Depth$(O(d(n)))$-PC.
\end{corollary}
\begin{proof}
Assume ${\rm NTIME}(t(n))\subseteq {\rm PH}\cap$Depth$(O(d(n)))$-PC. By Karp and
Lipton's thereom~\cite{KarpLipton}, we have
${\rm PH}=\sum_2^P={\rm NP}^{{\rm NP}}$. It follows from Theorem~\ref{main2-thm}.
\end{proof}
\begin{corollary}
If ${\rm PIT}_q(O(\log n))\in {\rm NSUBEXP}$ for a finite field $F(q)$, then
${\rm NEXP}\not\subseteq {\rm NC1/poly}$.
\end{corollary}
\begin{proof}
Assume that ${\rm PIT}_q\in {\rm NSUBEXP}$ and ${\rm NEXP}\subseteq
{\rm NC1/poly}=$Depth$(O(\log n))$-PC. By Impagliazzo, Kabanets, and
Wigderson's theorem~\cite{ImpagliazzoKabanetsWigderson},
${\rm NEXP}={\rm PH}$. We have a contradiction by Theorem~\ref{main2-thm}.
\end{proof}
\section{Conclusion
}
The result developed in this shows that derandomizing the PIT in any
finite field implies NEXP does not have nonuniform polynomial size
circuits. It gives right motivation to study the derandomization of
PIT in finite fields that the computational complexity community has
spent much efforts. We hope that the results in this paper brings a
tool to achieve the separation of NEXP from BPP via derandomizing
${\rm PIT}_p$ for a prime number $p$ such as $2$. Since there exists an
oracle to collapse NEXP to BPP by Heller~\cite{Heller}, separating
${\rm NEXP}$ from ${\rm BPP}$ requires a new way to go through the barrier of
relativization.
Another interesting open problem is if derandomizing ${\rm PIT}$ over $Z$
implies ${\rm NEXP}\not\subseteq{\rm P/poly}$ (In other words, ${\rm PIT}_Z\in
{\rm NSUBEXP}\Rightarrow {\rm NEXP}\not\subseteq{\rm P/poly}$?). Our technology
fails on integers $Z$. We cannot obtain a similar result as
Lemma~\ref{exclude-lemma} over the ring $Z$ of integers.
\vskip 10pt
{\bf Acknowledegments:} The author is grateful to Bohan Fan, Cynthia
Fu, and Feng Li for their proofreading and suggestions for an
earlier version of this paper. This research is supported in part by
NSF Early Career Award CCF-0845376.
|
\section{Optimal Experimental Design}
\label{intro}
Consider the following linear model
\begin{eqnarray} \label{eq:mod_rand} y = x^T(t) \theta + \epsilon(t),\end{eqnarray}
where components of $x^T(t) = (x_1(t), x_1(t), \dots, x_n(t))$ are $n$ linearly independent continuous functions on some compact space and $\theta\in {\rm I\kern-3.2pt R}^n$ is a vector of unknown parameters to be estimated. Let the error terms $\epsilon(t)$ follow a multivariate normal distribution with mean 0 and the error in each observation be independent from the others. Without loss of generality, suppress the dependency of the vector $x(t)$ on the actual experimental conditions $t$ and work with a model function such as \begin{eqnarray}\label{eq:lin_mod} y=x^T\theta +\epsilon,\end{eqnarray} in which the vector $x$ will be referred to as the regression or design vector. Let $\mathcal{X}=\{x_1,\dots, x_m\}\subset {\rm I\kern-3.2pt R}^n$ be the set of regression vectors, assume henceforth that the $x_i$'s span ${\rm I\kern-3.2pt R}^n$, and $X$ denote a matrix of order $n\times m$ whose columns consist of these vectors. (Frequently, the regression points are chosen from
some fixed compact set, here suppose
that some large fixed subset $\cal{X}$ has been preselected.)
\begin{definition} An \emph{experimental design of size $N$} is given by a finite number of regression points $x_1,\dots,x_m$ in ${\rm I\kern-3.2pt R}^n$ and nonnegative integers (representing the number of repetitions at each respective point) $n_1,\dots,n_m$ such that $\sum_{i=1}^m{n_i}=N$.\end{definition}
In this setting, the \emph{dispersion matrix} related to the optimal (unbiased) estimator for the parameter vector is \begin{eqnarray}\label{eq:disp} D=\sigma^2\left(\sum_{i=1}^m{n_ix_ix_i^T}\right)^{-1}=\frac{\sigma^2}{N}\left(\sum_{i=1}^m{\frac{n_i}{N}x_ix_i^T}\right)^{-1}.\end{eqnarray} \emph{Optimal experimental design} focuses on finding integers $n_i$ so that the dispersion matrix, which is a measure of the variance (or the error) of the estimator, is minimized in some sense. The dispersion matrix is positive definite (i.e., $D\succ 0$) and usually the minimization is with respect to the Loewner ordering over the cone of positive semidefinite matrices
($A\succeq B \iff A-B \in {\it S}\R^{n \times n}_+$). Since this is an antitonic ordering, minimizing the dispersion matrix is equivalent to maximizing the \emph{information matrix} $$M=\frac{N}{\sigma^2}\sum_{i=1}^m{\frac{n_i}{N}x_ix_i^T}.$$
When the total number of experiments $N$ is finite, experimental design problems become integer programming problems which are quite hard to attack especially for large $m$. Hence the case where $N$ tends to infinity is studied instead. In this case we maximize $M(u):=\sum_{i=1}^m{u_ix_ix_i^T}$, where $u_i \geq0$, for $i = 1,\dots, m$, and $\sum_{i=1}^m{u_i} = 1$. Note that an experimental design with an infinite sample size $N$ defines a probability distribution which assigns all its weight to a finite number of points. The points with positive weight are the \emph{support points} of the experimental design. One can refer to Chapter 12 in \cite{Puk93} or \cite{TMM09} for a valuable discussion on how to come up with an exact experimental design for a finite sample size once the optimal design for an infinite sample size is found.
\begin{definition} An \emph{information function} is a function $\phi$ from the cone of positive semidefinite matrices to the real line, $\phi: {\it S}\R^{n \times n}_+ \rightarrow {\rm I\kern-3.2pt R},$ which is positively homogeneous, superadditive, nonnegative, nonconstant, and upper semicontinuous. \end{definition}
It is easy to see that information functions are concave. They order the information matrices according to their informative value and preserve the Loewner ordering. The most common information functions are matrix means.
\begin{definition} Let $\lambda(C)$ denote the eigenvalues of a matrix $C$. If $C$ is a positive definite matrix, i.e., $C\succ0$, the matrix mean $\phi_p$ is a function defined as
$$\phi_p(C)=\left\{
\begin{array}{ccc}
\lambda_{\max}(C) & {\rm for } & p=\infty;\\
\left(\frac{1}{n}{\rm Trace} C^p\right)^{1/p} & {\rm for } &p\neq0,\pm\infty;\\
(\det C)^{1/n} & {\rm for } &p=0;\\
\lambda_{\min}(C) & {\rm for } &p=-\infty.
\end{array}\right.
$$ If $C$ is a singular positive semidefinite matrix, then
$$\phi_p(C)=\left\{
\begin{array}{ccc}
\lambda_{\max}(C) & {\rm for } & p=\infty;\\
\left(\frac{1}{n}{\rm Trace} C^p\right)^{1/p} & {\rm for } &p\neq0,\infty;\\
0 & {\rm for } &p\leq0.\\
\end{array}\right.
$$
\end{definition}
Matrix means satisfy the necessary properties of information functions when $p\leq1$. Using these functions, the \emph{general optimal experimental design problem} is defined as follows:
$$
\begin{array}{rrrcl}
& \max_u & g_p(u)& :=& \ln \phi_p(M(u))\\
(\mathcal{D}_p) & & e^T u & = & 1, \\
& & u & \geq & 0,
\end{array}$$ where $e$ is a vector of ones in ${\rm I\kern-3.2pt R}^m$.
Each value of the parameter $p$ gives rise to a different criterion with different applications. We will study one of the special cases (when $p=-1$) in great detail in Section \ref{sec:Aopt} forward.
\section{Ellipsoidal Inclusion Problems}
Assume that we have a set of points $\mathcal{X}=\{x_1,\dots,x_m\}\subset {\rm I\kern-3.2pt R}^n$, which spans ${\rm I\kern-3.2pt R}^n$ and is symmetric with respect to the origin. We are interested in approximating (especially enclosing) the convex hull of these points with an ellipsoid. Note that the idea is to approximate the complex structure of the convex hull with a simple geometric object. Boxes, balls, ellipsoids, and cylinders are used in the literature. Ellipsoids are preferred in many applications since they are smooth and flexible, and testing membership in or optimizing a linear function over an ellipsoid is a straightforward task.
The set $$\mathcal{E}(\bar{x},H):=\{x\in{\rm I\kern-3.2pt R}^n : (x-\bar{x})^T H (x-\bar{x})\leq n\}$$ for $\bar{x}\in {\rm I\kern-3.2pt R}^n$ and $H\succ0$ is an ellipsoid in ${\rm I\kern-3.2pt R}^n$. It is centered at $\bar{x}$ and its shape is defined by $H$. It can be viewed as a unit ball under an affine map where each point $\tilde{x}$ in the unit ball is mapped to a point $x=\bar{x}+\sqrt{n}L\tilde{x}$ in the ellipsoid, where $L$ satisfies $LL^T=H^{-1}$. Geometric properties of the ellipsoid such as its volume, length of its semi-axes, etc., are determined by the shape matrix $H$. For example, its volume is $\frac{n^{n/2}}{\sqrt{\det{H}}}$ times that of the unit ball.
The convex hull of a set of finitely many points can be enclosed by an infinite number of ellipsoids. Obviously we are only interested in ellipsoids which are centered at the origin (since $\mathcal{X}$ is symmetric around the origin) and resemble the convex hull in some sense. Although the enclosing ellipsoid which has the minimum volume is a natural choice from both theoretical and practical points of view, as discussed in detail in \cite{TYil07} and \cite{AST08}, defining the problem using a more general criterion is quite insightful since other criteria can be needed in certain applications.
For $q\leq 1$, consider the following problem:
$$\begin{array}{rrrcl}
& \min_H & f_q(H) &:=& - \ln \phi_q(H) \\
(\mathcal{P}_q) & & x_i^T H x_i & \leq & n, \, i =
1,\dots,m,\\
& & H & \succ & 0.
\end{array} $$ For each value of $q$, this problem finds an ellipsoid which encloses all points in $\mathcal{X}$, is centered at the origin, and has a shape matrix with the largest matrix mean $\phi_q$. Each value of the parameter $q$ leads to a different problem with a different geometric interpretation. For example, when $q=0$, the objective function becomes (a multiple of) $\ln \det(H^{-1})$ and hence $(\mathcal{P}_q)$ is equivalent to the Minimum-Volume Enclosing Ellipsoid problem discussed in the previously mentioned references. Similarly, for the extreme case of $q=-\infty$, we have $\ln (\lambda_{\min}(H))^{-1}$ as the objective function and hence the problem becomes that of finding the Minimum Enclosing Ball of $\mathcal{X}$. (See F\cite{Yil08} and \cite{AY08} for efficient algorithms for this problem.) When $q=1/2$, $(\mathcal{P}_q)$ maximizes the trace of $H^{1/2}$ and leads to a less familiar geometric problem in which we would like to maximize the sum of the inverses of the semi-axes of the enclosing ellipsoid. This problem has important applications in statistics and solving this problem is the main topic of this paper. We will refer to the general problem $(\mathcal{P}_q)$ as the \emph{ellipsoidal inclusion} problem.
\section{Duality}
We now show that the two problems introduced above are closely related.
\begin{lemma}\label{lemm:weak}{\rm [Weak Duality]} Let $p$ and $q$ be a pair of conjugate numbers in $\left(-\infty,1\right)$, i.e., they satisfy $pq=p+q$. Then we have $f_q(H)\geq g_p(u)$ for any $H$ and $u$
feasible in $(\mathcal{P}_q)$ and $(\mathcal{D}_p)$, respectively.\end{lemma}
\bf Proof: \rm }%\par \noindent We have
\begin{eqnarray*} f_q(H)-g_p(u) & = & - \ln\phi_q(H) - \ln\phi_p(M(u)) \\
& = & - \ln \left(\phi_q(H)\phi_p(M(u))\right)\\
&\geq &-\ln\left(\frac{1}{n}H\bullet M(u)\right)\\
&\geq& -\ln 1=0,
\end{eqnarray*} where $\bullet$ denotes the trace product of two symmetric matrices, i.e., $A\bullet B ={\rm Trace} (AB)$. The first inequality is an application of the H\"{o}lder's inequality (on the eigenvalues of the matrices at hand) and a detailed proof can be found in \cite{Puk93}. The second inequality follows from the feasibility of the solutions $H$ and $u$. Indeed, $\frac{1}{n}H\bullet(M(u)) = \frac{1}{n}\sum_{i=1}^m \left(u_i H\bullet(x_ix_i^T)\right) \leq \frac{1}{n}\sum_{i=1}^m{\left( u_i(x_i^THx_i)\right)}\leq \frac{n}{n}=1.$\ifmmode\else\unskip\quad\fi\hbox{\rlap{$\sqcap$}$\sqcup$}
\begin{theorem} \label{theo:opt_cond} {\rm [Strong Duality]} Let $p$ and $q$ be a pair of conjugate numbers in $\left(-\infty,1\right)$. There exist optimal solutions for problems $(\mathcal{P}_q)$ and $(\mathcal{D}_p)$. Furthermore, the following conditions, together with primal and dual feasibility, are necessary and sufficient for optimality in both
$(\mathcal{P}_q)$ and $(\mathcal{D}_p)$: \begin{itemize}
\item[a.] $H=\frac{n}{{\rm Trace}{(M(u))^{p}}}(M(u))^{p-1}$ and
\item[b.] $x_i^THx_i=n$ if $u_i>0$.
\end{itemize} \end{theorem}
\bf Proof: \rm }%\par \noindent Let $H$ be a feasible solution for problem $(\mathcal{P}_q)$. Summing up
the linear constraints, we must have $\sum_{i=1}^m{x_i^THx_i}=H\bullet
XX^T\leq nm$. Since $XX^T\succ0$ and $nm>0$, $\{H \succeq 0 : H\bullet
XX^T\leq nm\}$ is a compact set. Hence the feasible region for
problem $(\mathcal{P}_q)$ is also a compact set (since it is the intersection of
a compact set with a finite set of halfspaces). Moreover, $H = \epsilon I$
is feasible for $(\mathcal{P}_q)$ for sufficiently small positive $\epsilon$, and we can add the constraint
that $f_p(H) \leq f_p(\epsilon I)$ without loss of generality.
The objective function is (finite and) continuous on this
modified compact feasible region, so an optimal solution
exists for problem $(\mathcal{P}_q)$. Existence of an optimal solution for
$(\mathcal{P}_q)$ implies the existence of an optimal solution for
$(\mathcal{D}_p)$ as will be discussed later.
Sufficiency follows from the previous lemma, since the conditions imply equality in the weak duality inequality. In order to prove necessity, let $\tilde{H}$ be an optimal solution for $(\mathcal{P}_q)$. The KKT conditions must hold for this solution, i.e., there exist nonnegative multipliers $\tilde{u}\in{\rm I\kern-3.2pt R}^m$ such that the following equalities hold:
\begin{eqnarray} \label{eq:hu}-\frac{n}{{\rm Trace}{\tilde{H}}^{q}}{\tilde{H}}^{q-1}+M(\tilde{u})&=&0,\\\label{eq:comsl}
\tilde{u}_i(n-x_i^T\tilde{H}x_i)&=&0, \quad i=1,\dots,m.
\end{eqnarray}
These equalities imply that $\sum_{i=1}^m{\tilde{u}_i}=1$, since
\begin{eqnarray*} \sum_{i=1}^m{\tilde{u}_i} & = & \frac{\sum_{i=1}^m \tilde{u}_ix_i^T\tilde{H}x_i}{n}\\
& = & {\rm Trace} \left(\frac{\tilde{H} M(\tilde{u})}{n}\right)\\
& = & {\rm Trace} \left(\frac{\tilde{H} \left(\frac{n}{{\rm Trace}{\tilde{H}}^{q}}{\tilde{H}}^{q-1}\right)}{n}\right)\\
& = & \frac{n{\rm Trace} H^q}{n{\rm Trace} H^q}=1,\end{eqnarray*}
and hence $\tilde{u}$ is a feasible solution for $(\mathcal{D}_p)$. Strong duality holds for the solution pair $\tilde{H}$ and $\tilde{u}$, so strong duality holds for any pair of optimal solutions $H$ and $u$. Conditions (a) and (b) are direct consequences of Equations (\ref{eq:hu}) and (\ref{eq:comsl}), and hence they are necessary. \ifmmode\else\unskip\quad\fi\hbox{\rlap{$\sqcap$}$\sqcup$}
Let $\beta_i(u):=x_i^T(M(u))^{p-1}x_i$. The following identity will be used extensively. \begin{eqnarray}\label{eq:base}
u^T\beta(u)&=&\sum_{i=1}^m{u_i\beta_i(u)}\nonumber\\
&=& \sum_{i=1}^m{{\rm Trace}\left(u_ix_i^T(M(u))^{p-1}x_i\right)}\nonumber\\
&=& {\rm Trace}\left((M(u))^{p-1}\sum_{i=1}^m{u_ix_ix_i^T}\right)\nonumber\\
&=& {\rm Trace}{(M(u))^{p}}.
\end{eqnarray} Using (\ref{eq:base}), we can write the necessary and sufficient conditions for $u$ to be optimal in $(\mathcal{D}_q)$ (the optimal
$H$ for $(P_q)$ follows from (a)) as
\begin{itemize}
\item[(i)] $\beta_i(u)\leq {u}^T\beta(u)$ for all $i$, and
\item[(ii)] $\beta_i(u)={u}^T\beta(u)$ if $u_i>0$,
\end{itemize} which motivates the following definitions.
\begin{definition}\label{def:sol} Given a positive $\epsilon$, we call a dual feasible point $u$ an \emph{$\epsilon$-primal feasible solution} if $\beta_i(u)\leq u^T\beta(u)(1+\epsilon)$ for all $i$, and say that it satisfies the \emph{$\epsilon$-approximate optimality conditions} or it is an \emph{$\epsilon$-approximate optimal solution} if moreover $\beta_i(u)\geq u^T\beta(u)(1-\epsilon)$ whenever $u_i>0$.
\end{definition}
The following lemma justifies the notation and proves that an $\epsilon$-primal feasible solution for $(\mathcal{D}_p)$ is close to being optimal in a well-defined way.
\begin{lemma} \label{lem:dual_gap} Let $p$ and $q$ be a pair of conjugate numbers in $\left(-\infty,1\right)$. Given a dual feasible solution $u$ which is $\epsilon$-primal feasible, $H=\frac{n}{(1+\epsilon){\rm Trace}{(M(u))^{p}}}(M(u))^{p-1}$ is feasible in $(\mathcal{P}_q)$ and we have
$0\leq g_p^*-g_p(u)\leq \ln
(1+\epsilon)$ where $g_p^*$ is the optimal objective function value
of $(\mathcal{D}_p)$. \end{lemma}
\bf Proof: \rm }%\par \noindent The $\epsilon$-primal feasibility implies that
$H=\frac{n}{(1+\epsilon){\rm Trace}{(M(u))^{p}}}(M(u))^{p-1}$ is feasible for the primal
problem $(\mathcal{P}_q)$. Let us first assume that $p,q\neq 0$. Then by weak duality, we have
\begin{eqnarray*} 0 & \leq & f_q(H)-g_p^* \\
&=&-\frac{1}{q}\ln\left(\frac{1}{n}{\rm Trace}\left(\frac{n(M(u))^{p-1}}{(1+\epsilon){\rm Trace}{(M(u))^{p}}}\right)^q\right)-g_p^*\\
&=&\ln(1+\epsilon)-\frac{1}{q}\ln\left(\frac {n^{q-1}{\rm Trace}(M(u))^{(p-1)q}}{({\rm Trace}{(M(u))^{p}})^q}\right)-g_p^*\\
&=&\ln(1+\epsilon)+\ln\left(n^{\frac{1-q}{q}}\left({\rm Trace}(M(u))^{p}\right)^{\frac{q-1}{q}}\right)-g_p^*\\
&\leq&\ln(1+\epsilon)+\frac{1}{p}\ln\left(\frac{1}{n}{\rm Trace}(M(u))^{p}\right)-g_p^*\\
g_p^*-g_p(u)&\leq&\ln(1+\epsilon).
\end{eqnarray*} The case where $p=q=0$ is similar and the proof can be found in \cite{AST08}.\ifmmode\else\unskip\quad\fi\hbox{\rlap{$\sqcap$}$\sqcup$}
\begin{lemma} \label{lem:init_gapKHA} $u^0=\frac{1}{m}(1,1,\dots,1)$ is an ($m$-$1$)-primal feasible solution.\end{lemma}
\bf Proof: \rm }%\par \noindent We have \begin{eqnarray*} \sum_{i=1}^m\frac{1}{m}\beta_i(u^0)&=& (u^0)^T\beta(u^0), {\rm or}\\
\sum_{i=1}^m\beta_i(u^{0}) & = & m(u^0)^T\beta(u^0), {\rm so \ that}\\
\max_{1\leq i \leq m}\beta_i(u^0) & \leq & (1+(m-1))(u^0)^T\beta(u^0),\end{eqnarray*} and the result follows from the definition of an ($m$-1)-primal feasible solution.\ifmmode\else\unskip\quad\fi\hbox{\rlap{$\sqcap$}$\sqcup$}
So far, we have developed the duality relation between problems $(\mathcal{P}_q)$ and $(\mathcal{D}_p)$ and characterized the optimal solutions of these problems. We also have an initial solution for $(\mathcal{D}_p)$ which is somewhat close to optimality and we can assess the quality of the solutions at hand. (Note that we will refer to this initialization method as ``Khachiyan's Initialization" since it was used by Khachiyan in \cite{Kha96} for $p=q=0$.) In other words, we know how to start and end an algorithm for $(\mathcal{D}_p)$ and now we need to figure out how to move from a given solution to a better one. The selection of the iterate and the analysis of the algorithm changes with respect to the specific parameter, namely $p$, of the optimal experimental design. In the following section, we will develop a Frank-Wolfe type first-order algorithm for the case when $p=-1$ (and hence $q=1/2$). This problem is referred to as the A-optimal experimental design in statistics.
\begin{remark} \label{rem:lit} We would like to note that most of the results in this section are not entirely new to the statistic community. What is new, and hopefully useful, is the treatment of the subject using a standard mathematical programming approach that builds the necessary machinery in devising algorithms and analysing their convergence properties.
Specifically, (i) Theorem \ref{theo:opt_cond} in this section (and Theorem \ref{theo:opt_condTR} below, which is a special case of Theorem \ref{theo:opt_cond}) can be obtained by following Theorems 7.12, 7.19, and 7.20 in \cite{Puk06}; and (ii) Lemma \ref{lem:dual_gap} is similar to Proposition IV.28 in \cite{Paz86}. Instead of borrowing these results directly from literature, we have provided a consistent and comprehensive treatment of the subject here. We strongly believe this is a simpler and -in some sense- more intuitive approach for building algorithms. Understanding the relationship between primal and dual problems, and the derivation of the optimality conditions based on this relation is necessary to follow the rest of the paper. One exception is possibly Lemma \ref{lem:init_gapKHA}, which was only obtained for the $p=q=0$ case in \cite{Kha96}. The generalized result provided here is novel according to our knowledge.
\end{remark}
\begin{remark} \label{rem:geo} The duality relationship between problems $(\mathcal{P}_q)$ and $(\mathcal{D}_p)$ presented in this chapter, provides a geometric and non-trivial insight to the design problem: Finding the best experimental design is equivalent to covering the induced design space with a 'minimum volume' ellipsoid (where the measure of the volume is dictated by the criterion used for the design problem). This interpretation is also well-known to the statistics community for the case $p=q=0$. (See: \cite{S72},\cite{ST73}, \cite{H93}, and more recently in \cite{AST08}). In \cite{BDZ06}, the authors provide a similar discussion about the geometric interpretation of $ (\mathcal{D}_p)$-optimal design problems for all values of $p$ for models with two parameters. Our discussion is more general since it is independent of the number of parameters in the model. Understanding the geometric interpretation plays a significant role in internalizing several pieces of the machinery developed in this paper, especially in construction of approximate solutions, quantification of the duality gap associated with them and choosing pivots for the algorithm. A similar geometric interpretation exists for the ${\rm D}_k$-optimal experimental design problem: A generalization of the D-optimal experimental design problem where we are only interested in estimating the first $k$ out of $n$ parameters in a general linear model. In this case, finding the best experimental design is equivalent to covering the induced design space with a minimum-area ellipsoidal cylinder with special properties about its base and axis as discussed in \cite{AT13}. It is easy to see that although this paper discusses only the D-criterion, the geometric interpretation carries to other criteria in a straightforward way.
\end{remark}
\begin{remark} Finally, before continuing our discussion towards algorithms for the A-optimal experimental design problem ($p=-1$ and $q=1/2$) below, we would like to mention that `in principle' algorithms for problems with other values of $p$ (and respective $q$) can be designed and analysed following the steps outlined here. Nevertheless, the step sizes and convergence analysis need to be customized for each criterion, and can be challenging in some cases. One can refer to \cite{AST08} for a detailed analysis of similar algorithms for the D-optimal experimental design problem. \end{remark}
\section{Existing Algorithms} \label{sec:LIT}
Many Frank-Wolfe type algorithms have been devised to solve experimental design problems, especially for the D-optimal experimental design problem. Some of these were developed by statisticians: \cite{Fed72} and \cite{W72} provided algorithms that maximize a linearization of the objective function over the unit simplex at each iteration. These algorithms only allow iterations that increase the weight of one of the coordinates of the solution. These were improved significantly by \cite{Atw73} where decreasing the weight of the chosen coordinate was also considered, paralleling the addition of \emph{Wolfe's Away Steps} to Frank's algorithm (see \cite{FW56} for the original Frank-Wolfe algorithm). Recently, these algorithms were analysed rigorously by the optimization community, motivated by the ellipsoidal inclusion problem rather than the design problem. The algorithms in \cite{Kha96} and \cite{KumYil05} were equivalent to that of \cite{Fed72}. In addition, \cite{KumYil05} proposed an initialization scheme that produces optimal solutions with significantly smaller number of nonzero weights than previous algorithms. This was accompanied by introducing the concept of \emph{core sets}, and the authors were able to provide upper bounds on the number of nonzero weights in the optimal design. Later, \cite{TYil07} extended the analysis to include Wolfe's away steps, hence providing rigorous complexity results for an algorithm equivalent to that of \cite{Atw73}. During this period, \cite{HarPro07} proved a simple condition that can be used to identify and eliminate points that do not lie on the boundary of the optimal ellipsoid, i.e., points that are guaranteed to have zero weight in the optimal design. (Recently, this result has been extended for all values of $p$ in \cite{HarPro13}.) Incorporating this condition to any Frank-Wolfe type algorithm is very easy and improves the computational time significantly (see Chapter 2 in \cite{Ahi09}). In addition, \cite{AST08} proved that the Frank-Wolfe type algorithms with an exact line search have favorable local convergence properties and therefore can be used to obtain very accurate solutions. In the following section, we will devise and analyse an algorithm which is a Frank-Wolfe type algorithm with Wolfe's aways steps. It can be viewed as applying Atwood's approach to the A-optimal experimental design problem. The global and local convergence properties that will be established below are in line with those developed recently by the optimization community for the D-optimal experimental design problem.
In contrast to Frank-Wolfe type algorithms, multiplicative algorithms update all weights simultaneously. Several versions were developed for various criteria: C-optimality in \cite{Fell74}, D-optimality in \cite{Tit76}, and A-optimality in \cite{Tor83}. Recently, faster algorithms were developed in \cite{Yu11} for D-optimality and in \cite{YBT13} for the general experimental design problem, i.e., problem ($\mathcal{D}_p$) discussed in this paper. A relatively recent survey on multiplicative algorithms together with a new multiplicative approach can also be found in \cite{TMM09}.
Another interesting and modern approach to the experimental design problem is using semidefinite programming reformulations as discussed in \cite{VB99}. This approach fails to solve large problems due to the lack of efficient solvers as demonstrated in Section \ref{sec:TRcompSDP}.
\section{The A-Optimal Experimental Design Problem}\label{sec:Aopt}
Let $\mathcal{X}=\{x_1,\dots, x_m\}\subset {\rm I\kern-3.2pt R}^n$ be a set of regression vectors and $X$ denote a matrix of order $n\times m$ whose columns consist of these vectors. Finding a design which minimizes the mean dispersion of the parameters in (\ref{eq:mod_rand}) amounts to solving $$
\begin{array}{rrrcl}
& \max_u & \hat{g}(u) &:=& -{\rm Trace} (M(u))^{-1} \\
(\hat{\mathcal{D}}) & & e^T u & = & 1, \\
& & u & \geq & 0,
\end{array}
$$
where $e$ is a vector
of ones in ${\rm I\kern-3.2pt R}^m$ as in the previous sections. Problem $(\hat{\mathcal{D}})$ is referred to as the A-optimal experimental design problem in statistics. In \cite{Fed72}, Fedorov proved that a Frank-Wolfe type algorithm converges to an optimal design and discussed the conditions under which D-optimal and A-optimal designs coincide. In this paper, we will introduce a pair of problems dual to each other and closely related to $(\hat{\mathcal{D}})$. Using the interplay between these problems, we will develop various Frank-Wolfe type algorithms and prove that an $\epsilon$-approximate solution (defined as in Section \ref{intro}) can be obtained in $\mathcal{O}(n\ln n+\epsilon^{-1})$ or $\mathcal{O}(\ln m+\epsilon^{-1})$ iterations. Each step of the algorithm can be performed in $\mathcal{O}(nm)$ arithmetic operations. In Section \ref{sec:TRloc}, we will prove that some of these algorithms possess a local linear convergence property. These algorithms are also preferable in practice as illustrated by the computational results in Section \ref{sec:TRcomp}.
Consider the following two problems:
$$\begin{array}{cccl}
\min & f(H) := - 2\ln {\rm Trace} H^{1/2} \\
(\mathcal{P}) & x_i^T H x_i \leq 1, \, i =
1,\dots,m,
\end{array}$$ and
$$
\begin{array}{cccl}
\max_{u} & g(u) := -\ln {\rm Trace} (M(u))^{-1} \\
(\mathcal{D}) & e^T u = 1, \\
& u \geq 0.
\end{array}
$$ $(\mathcal{P})$ is a special case of $(\mathcal{P}_q)$ in Section \ref{intro} in which $q=1/2$. From a geometric point of view, it is the problem of finding an ellipsoid which encloses all data points in $\mathcal{X}$ and has the largest sum of inverses of its semi-axes. Also $(\mathcal{D})$ is a special case of $(\mathcal{D}_p)$ in Section \ref{intro} where $p=-1$. This problem is equivalent to the statistical problem $(\hat{\mathcal{D}})$ introduced above. We will use both $(\mathcal{D})$ and $(\hat{\mathcal{D}})$ in order to develop and analyze first-order algorithms for solving all of the three problems mentioned above simultaneously. We will first establish weak duality:
\begin{lemma}\label{lemm:weakTR}{\rm [Weak Duality]} We have $f(H)\geq g(u)$ for any $H$ and $u$ feasible in $(\mathcal{P})$ and $(\mathcal{D})$, respectively.\end{lemma}
\bf Proof: \rm }%\par \noindent Follows from Lemma \ref{lemm:weak} since $p=-1$ and $q=1/2$ are conjugate numbers in $\left(-\infty,1\right]$. Note that we have omitted an additive constant in the objective functions of $(\mathcal{P})$ and $(\mathcal{D})$ in this section unlike Section \ref{intro}.\ifmmode\else\unskip\quad\fi\hbox{\rlap{$\sqcap$}$\sqcup$}
We next show that having two feasible solutions $H$ and $u$ such that $f(H)=g(u)$ is not just sufficient but also necessary for optimality.
\begin{theorem} \label{theo:opt_condTR} {\rm [Strong Duality]} There exist optimal solutions $H^*$ and $u^*$ for problems $(\mathcal{P})$ and $(\mathcal{D})$, respectively. Furthermore, the following conditions, together with primal and dual feasibility, are necessary and sufficient for optimality in both
$(\mathcal{P})$ and $(\mathcal{D})$: \begin{itemize}
\item[a.] $H^*=\frac{(M(u^*))^{-2}}{{\rm Trace}{(M(u^*))^{-1}}}$,
\item[b.] $x_i^TH^*x_i=1$ if $u_i^*>0$.
\end{itemize} \end{theorem}
\bf Proof: \rm }%\par \noindent As in the previous lemma, the proof follows from Theorem \ref{theo:opt_cond} for $p=-1$ and $q=1/2$.\ifmmode\else\unskip\quad\fi\hbox{\rlap{$\sqcap$}$\sqcup$}
After some simplification, the necessary and sufficient conditions for $u^*$ to be optimal in $(\mathcal{D})$ can be written as
\begin{itemize}
\item[(i)] $\alpha_i(u^*)\leq {u^*}^T\alpha(u^*)$ for all $i$, and
\item[(ii)] $\alpha_i(u^*)={u^*}^T\alpha(u^*)$ if $u^*_i>0$,
\end{itemize} where $
\alpha(u) := \nabla \hat{g}(u) = (x_i^T (M(u))^{-2} x_i)_{i=1}^m$. We say that a feasible solution $u$ for $(\mathcal{D})$ is \emph{$\epsilon$-primal feasible} if
$\alpha_i(u)\leq u^T\alpha(u)(1+\epsilon)$ for all $i$, and say that it satisfies the \emph{$\epsilon$-approximate optimality conditions} or it is an \emph{$\epsilon$-approximate optimal solution} if moreover $\alpha_i(u)\geq u^T\alpha(u)(1-\epsilon)$ for all $i$ such that $u_i>0$. (Note that these definitions can be deduced from those in Section \ref{intro} for $p=-1$ and $q=1/2$.)
\begin{lemma} \label{lem:dual_gapTR} Let $u$ be an $\epsilon$-primal feasible solution. Then we have \begin{itemize}\item[i.] $0\leq g^*-g(u)\leq \ln
(1+\epsilon)$
\item[ii.] $1\leq \frac{\hat{g}(u)}{\hat{g}^*}\leq 1 +\epsilon,$
\end{itemize} where $g^*$ and $\hat{g}^*$ are the optimal objective function values
of $(\mathcal{D})$ and $(\hat{\mathcal{D}})$, respectively.\end{lemma}
\bf Proof: \rm }%\par \noindent Since $u$ is an $\epsilon$-primal feasible solution, $\frac{(M(u))^{-2}}{(1+\epsilon){\rm Trace}{(M(u))^{-1}}}$ is feasible with respect to $(\mathcal{P})$. Let $H^*$ and $u^*$ be optimal solutions of $(\mathcal{P})$
and $(\mathcal{D})$, respectively. Then we have \begin{eqnarray} -2\ln{\rm Trace}\left(
\frac{(M(u))^{-2}}{(1+\epsilon){\rm Trace}{(M(u))^{-1}}}\right)^{1/2}+2\ln{\rm Trace} H^{*1/2}&\geq&0, {\rm or}\nonumber\\
\ln(1+\epsilon)-\ln{\rm Trace} (M(u))^{-1}-g(u^*)&\geq&0, {\rm from \ which} \nonumber\\
0\leq g^*-g(u)\leq \ln(1+\epsilon),&& \end{eqnarray} which proves (i). Property (ii) follows from $g=-\ln(-\hat{g})$.\ifmmode\else\unskip\quad\fi\hbox{\rlap{$\sqcap$}$\sqcup$}
\section{Algorithms and Analysis} \label{sec:TR}
In the rest of this paper, we will develop various iterative (Frank-Wolfe type) algorithms for solving $(\mathcal{D})$ and $(\hat{\mathcal{D}})$. We will assume that the following assumption holds, for every feasible solution $u$ produced by these algorithms.
\begin{assumption}\label{ass:upb} The dual feasible variable $u$ satisfies
$\omega_j(u):=x_j^T(M(u))^{-1}x_j\leq \omega$ for all $j \in \{1,\dots,m\}$ and for some $\omega>1$.\end{assumption}
The objective function $\hat{g}$ of $(\hat{\mathcal{D}})$ is a concave function with gradient $\alpha(u)$
and that, with
\begin{equation}\label{eq:updateuTR}
u_+ := (1-\tau) u + \tau e_j,
\end{equation}
rank-one update formulae give \begin{eqnarray} \label{eq:imping}\hat{g}(u_+)&=& -{\rm Trace}(M(u_+))^{-1}\nonumber\\
&=&-{\rm Trace}\left((1+\lambda)\left((M(u))^{-1}-\frac{\lambda (M(u))^{-1}x_jx_j^T(M(u))^{-1}}{1+\lambda \omega_j(u)}\right)\right)\nonumber\\
&=&-(1+\lambda)\left({\rm Trace}(M(u))^{-1}-\frac{\lambda{\rm Trace}\left((M(u))^{-1}x_jx_j^T(M(u))^{-1}\right)}{1+\lambda \omega_j(u)}\right)\nonumber\\
&=&(1+\lambda)\hat{g}(u)+\frac{\lambda(1+\lambda)}{1+\lambda \omega_j(u)}\alpha_j(u),\end{eqnarray} where $\lambda=\frac{\tau}{1-\tau}$. The partial derivative of the objective function is equal to \begin{eqnarray}\frac{\partial \hat{g}(u_+)}{\partial \lambda}&=&\hat{g}(u)+\frac{\lambda^2\omega_j(u)+2\lambda+1}{(1+\lambda \omega_j(u))^2}\alpha_j(u).\end{eqnarray} Let $\hat{g}$, $\omega_j$, and $\alpha_j$ be shorthand for $\hat{g}(u)$, $\omega_j(u)$, and $\alpha_j(u)$, respectively. The numerator of the partial derivative is equal to the left-hand side of the following equation (the denominator is positive):
\begin{eqnarray} \label{eq:TRroot}(\omega_j^2\hat{g}+\omega_j \alpha_j)\lambda^2+\lambda (2\omega_j \hat{g}+2\alpha_j)+\hat{g}+\alpha_j=0.\end{eqnarray} We can find the best step size $\tau^*$ (or $\lambda^*$) by investigating the roots of the quadratic equation (\ref{eq:TRroot}) and the boundary condition ($\lambda^*\geq-u_j$) arising from the nonnegativity of the dual feasible solutions as follows:
\begin{itemize}
\item if we have $\omega_j \hat{g}+\alpha_j = 0$, then the partial derivative is negative for all values and hence $\lambda^*=-u_j$;
\item if $(1 - \omega_j)(\alpha_j + \omega_j \hat g) < 0$ (which is equivalent to $\omega_j < 1$ since $\alpha_j + \omega_j \hat g\leq0$ for any feasible solution), the discriminant of the quadratic (\ref{eq:TRroot}) is negative. Furthermore, $\hat{g} + \alpha_j < 0$ (since $0 \leq \omega_j < 1$, $\alpha_j + \omega_j \hat g < 0$, and $\hat{g} < 0$), and hence the quadratic (\ref{eq:TRroot}) has no real roots and everywhere negative. Therefore, $\lambda^*=-u_j$;
\item otherwise $\lambda^*$ is equal to one of the roots of the quadratic (\ref{eq:TRroot}), which are \begin{eqnarray}\lambda^*_{1,2}&=&\frac{-\omega_j \hat{g}-\alpha_j\pm\sqrt{(\omega_j \hat{g}+\alpha_j)^2-(\omega_j^2\hat{g}+\omega_j \alpha_j)(\hat{g}+\alpha_j)}}{(\omega_j^2\hat{g}+\omega_j \alpha_j)}\nonumber\\
&=& \frac{-\omega_j \hat{g}-\alpha_j\pm\sqrt{\alpha_j(1-\omega_j)(\alpha_j+\omega_j \hat{g})}}{(\omega_j^2\hat{g}+\omega_j \alpha_j)}\nonumber,\end{eqnarray} or $-u_j$ whichever is feasible and gives the greatest improvement in the objective function.
\end{itemize}
Once we find the step size, we can calculate $\omega(u_+)$ and $\alpha(u_+)$ from \begin{eqnarray} \label{update_xi}\omega_i(u_+)&=&x_i^T(M(u_+))^{-1}x_i\nonumber\\
&=&x_i^T\left((1+\lambda)\left((M(u))^{-1}-\frac{\lambda (M(u))^{-1}x_jx_j^T(M(u))^{-1}}{1+\lambda \omega_j(u)}\right)\right)x_i\nonumber\\
&=&(1+\lambda)\omega_i(u)-\frac{(1+\lambda)\lambda}{1+\lambda \omega_j(u)}\omega_{ij}(u)^2\nonumber\\
&=& (1+\lambda)(\omega_i(u)-\eta\omega_{ij}(u)^2),
\end{eqnarray}
and
\begin{eqnarray}
\alpha_i(u_+)&=&x_i^T(M(u_+))^{-2}x_i\nonumber\\
&=& x_i^T((1+\lambda)\left((M(u))^{-1}-\frac{\lambda (M(u))^{-1}x_jx_j^T(M(u))^{-1}}{1+\lambda \omega_j(u)}\right)...\nonumber\\
&& (1+\lambda)\left((M(u))^{-1}-\frac{\lambda (M(u))^{-1}x_jx_j^T(M(u))^{-1}}{1+\lambda \omega_j(u)}\right))x_i\nonumber\\
&=& (1+\lambda)^2 x_i^T((M(u))^{-2}-\frac{2\lambda}{1+\lambda \omega_j(u)}(M(u))^{-2}x_jx_j^T(M(u))^{-1}...\nonumber\\
&& +\frac{\lambda^2}{(1+\lambda \omega_j(u))^2}(M(u))^{-1}x_jx_j^T(M(u))^{-2}x_jx_j^T(M(u))^{-1})x_i\nonumber\\
&=& (1+\lambda)^2\alpha_i(u)-2\frac{(1+\lambda)^2\lambda}{1+\lambda \omega_j(u)}\omega_{ij}(u)\alpha_{ij}(u)+\frac{(1+\lambda)^2\lambda^2}{(1+\lambda \omega_j(u))^2}\omega_{ij}(u)^2\alpha_{j}(u)\nonumber\\
&=& (1+\lambda^2)(\alpha_i(u)-2\eta\omega_{ij}(u)\alpha_{ij}(u)+\eta^2\omega_{ij}(u)^2\alpha_j(u)),
\end{eqnarray}
where $\eta:=\frac{\lambda}{1+\lambda \omega_j(u)}$,
$\omega_{ij}(u):=x_i^T(M(u))^{-1}x_j$, and
$\alpha_{ij}(u):=x_i^T(M(u))^{-2}x_j$. Note that all updates can be performed cheaply
(in $\mathcal{O}(nm)$ operations).
Now we describe two Frank-Wolfe type algorithms. The first algorithm (Algorithm \ref{TR-KH}) uses positive step sizes and seeks an $\epsilon$-primal feasible solution; whereas the second one (Algorithm \ref{TR-TY}) may also have negative step sizes and stops when an $\epsilon$-approximate optimal solution is found. This algorithm is an extension of the first one with Wolfe's away steps. We will show that although these algorithms have similar global complexity results, away steps are necessary in order to achieve high accuracy, a phenomenon that is also observed for the D-Optimal Experimental Design Problem in \cite{AST08}.
\begin{center}
\framebox{ \vbox{\hsize=1in
\begin{algorithm}\label{TR-KH}\begin{tabbing}
\= \hspace*{.25in} \= \hspace*{.25in} \= \hspace*{.25in} \=
\hspace*{.25in} \=\kill
\\\>\textbf{Input:} $X \in {\rm I\kern-3.2pt R}^{n \times m}$,
$\epsilon > 0$.
\\\> \textbf{Step 0.} Let $u=(1/m)e$. Compute $\omega(u)$ and $\alpha(u)$.\\
\>\textbf{Step 1.} Find $j := \arg\max_t \{\alpha_t(u) - u^T\alpha(u)\}$. \\
\>\>\textbf{If} $\frac{\alpha_j(u)}{u^T\alpha(u)}-1 \leq \epsilon$, \\
\>\>\>\textbf{STOP:} $u$ is an $\epsilon$-primal feasible solution.\\
\>\textbf{Step 2.} Replace $u$ as in (\ref{eq:updateuTR}), where
$\tau
> 0$
is chosen to maximize $\hat{g}$.\\
\>\textbf{Step 3.} Update $\omega(u)$ and $\alpha(u)$. \textbf{Go to} Step 1.
\end{tabbing}\end{algorithm}}
}
\end{center}
\begin{center}
\framebox{ \vbox{\hsize=1in
\begin{algorithm}\label{TR-TY}\begin{tabbing}
\= \hspace*{.25in} \= \hspace*{.25in} \= \hspace*{.25in} \=
\hspace*{.25in} \=\kill
\\\>\textbf{Input:} $X \in {\rm I\kern-3.2pt R}^{n \times m}$,
$\epsilon > 0$.
\\\> \textbf{Step 0.} Let $u=(1/m)e$. Compute $\omega(u)$ and $\alpha(u)$.\\
\>\textbf{Step 1.} Find $j := \arg\max_t \{\alpha_t(u) - u^T\alpha(u)\}$ and $i := \arg\min_t\{\alpha_t(u) - u^T\alpha(u):u_t>0\}$. \\
\>\>\textbf{If} $\frac{\alpha_j(u)}{u^T\alpha(u)}-1 \leq \epsilon$ and $1-\frac{\alpha_i(u)}{u^T\alpha(u)} \leq \epsilon$, \\
\>\>\>\textbf{STOP:} $u$ is an $\epsilon$-approximate optimal solution.\\
\>\>\textbf{Else,} \\
\>\>\>\textbf{if} $\alpha_j(u) - u^T\alpha(u) > u^T\alpha(u) - \alpha_i(u) $, \textbf{go to} Step 2;\\
\>\>\>\textbf{else,} \textbf{go to} Step 3. \\
\>\textbf{Step 2.} Replace $u$ as in (\ref{eq:updateuTR}), where
$\tau
> 0$
is chosen to maximize $g$. \textbf{Go to} Step 4.\\
\>\textbf{Step 3.} Replace $u$ by $u_+:= (1 - \tau)u + \tau e_i$,
where now $\tau$ is
chosen from \\ \>\>negative values to maximize $\hat{g}$ subject to $u_+$
remaining feasible. \\
\>\textbf{Step 4.} Update $\omega(u)$ and $\alpha(u)$. \textbf{Go to} Step 1.
\end{tabbing}\end{algorithm}}
}
\end{center}
If we look closely at these algorithms, we can identify three different types of iterations. Let $u^l$ be the dual feasible solution at hand at iteration number $l$, $e_{j_l}$ be the vertex that we use in our update and $\tau_l$ be the step size associated with this update. We refer to iteration $l$ as
\begin{itemize}\item[-]an \emph{add/increase step} if $\tau_l>0$,
\item[-]a \emph{decrease step} if $u^{l}_{j_l}>0$ and $\frac{-u^{l}_{j_l}}{1-u^{l}_{j_l}}<\tau_l<0$, and
\item[-]a \emph{drop step} if $u^{l}_{j_l}>0$ and $\tau_l=\frac{-u^{l}_{j_l}}{1-u^{l}_{j_l}}$.\end{itemize}
We only have add/increase steps in Algorithm \ref{TR-KH}, whereas all types of steps can be performed in Algorithm \ref{TR-TY}. Note that after a drop step we have $u^{l+1}_{j_l}=0$. In such a step, we may not be able to improve the objective function as much as we desire. Fortunately, the number of drop steps is bounded above by the number of add steps plus a constant (the
number of positive components of the initial solution), and hence studying only the
first two types of steps will be enough to obtain convergence results.
\begin{lemma} \label{lem:init_KHA_TR} $u^0=(1/m)e=\frac{1}{m}(1,1,\dots,1)$ is an $(m-1)$-primal feasible solution.\end{lemma}
\bf Proof: \rm }%\par \noindent Follows from Lemma \ref{lem:dual_gap} in Section \ref{intro}. \ifmmode\else\unskip\quad\fi\hbox{\rlap{$\sqcap$}$\sqcup$}
We now analyze the first algorithm closely:
\begin{lemma}\label{lemm:CAlg1} As long as $u^l$ satisfy Assumption \ref{ass:upb} for all $l=1,2,\dots$, Algorithm \ref{TR-KH} finds an $\epsilon$-primal feasible solution in at most \begin{eqnarray}
\label{{Alg1iter}}\mathcal{L}(\epsilon)=\mathcal{O}(\ln m+\epsilon^{-1})\end{eqnarray} steps. The constants hidden in the `big oh' are linearly dependent on the constant $\omega$ in Assumption \ref{ass:upb}.\end{lemma}
\bf Proof: \rm }%\par \noindent Given a dual solution $u^l$ (the iterate at iteration $l$), we define $ \epsilon_l = \max \{\frac{\alpha_j(u^l)}{(u^l)^T\alpha(u^l)}-1 ,1-\frac{\alpha_i(u^l)}{(u^l)^T\alpha(u^l)} \}$, where $j := \arg\max_t \{\alpha_t(u^l) - (u^l)^T\alpha(u^l)\}$ and $i := \arg\min_t\{\alpha_t(u^l) - (u^l)^T\alpha(u^l):u^l_t>0\}$. (Note that the algorithm stops at iteration $k$ if $\epsilon_k \leq \epsilon$.)
We will first prove that \begin{eqnarray} \label{L1}\mathcal{L}(1)=\min\{l|\epsilon_l\leq1\}=\mathcal{O}(\ln m).\end{eqnarray} Let $j_l$ be the index of the pivot point at iteration $l$, $\tau_l$ be the step size, and $\lambda_l=\frac{\tau_l}{1-\tau_l}$. (Remember that all values of $\hat{g}$ are negative by definition.) At each iteration $l$ with $\epsilon_l\geq1$, from (\ref{eq:imping}), we have
\begin{eqnarray} \hat{g}(u^{l+1})-\hat{g}(u^l)& = & \lambda_l\hat{g}(u^l)+\frac{\lambda_l(1+\lambda_l)}{1+\lambda_l\omega_{j_l}(u^l)}\alpha_{s_l}\nonumber\\
& \geq & \frac{1}{2\omega_{j_l}(u^l)}\hat{g}(u^l)-\frac{\frac{1}{2\omega_{j_l}(u^l)}}{1+\frac{1}{2\omega_{j_l}(u^l)}\omega_{j_l}(u^l)}2\hat{g}(u^l)\nonumber\\
& \geq & \frac{1}{2\omega_{j_l}(u^l)}\hat{g}(u^l)\left(1-\frac{2}{1+\frac{1}{2}}\right)\nonumber\\
& \geq & -\frac{\hat{g}(u^l)}{6\omega}.
\end{eqnarray} The first inequality follows since the improvement obtained from choosing the best step length is at least as good as the improvement obtained by using any step length; in particular, it can be bounded by plugging in $\lambda_l=\frac{1}{2\omega_{j_l}(u^l)}$.
Hence we have \begin{eqnarray}\label{ggu} \hat{g}(u^{l+1})\geq (1-\frac{1}{6\omega})\hat{g}(u^l). \end{eqnarray} Using Lemmas \ref{lem:dual_gapTR} and \ref{lem:init_KHA_TR}, \begin{eqnarray} \label{gg0}\hat{g}(u^0)\geq m\hat{g}^*.\end{eqnarray} Combining inequalities (\ref{ggu}) and (\ref{gg0}), we obtain \begin{eqnarray} \hat{g}^*\geq \hat{g}(u^l)\geq & (1-\frac{1}{6\omega})^l\hat{g}(u^0)\geq (1-\frac{1}{6\omega})^lm\hat{g}^*\geq e^{-\frac{l}{6\omega}}m\hat{g}^*.\end{eqnarray} Hence we must have $\mathcal{L}(1)\leq 6\omega \ln(m)=\mathcal{O}(\ln m).$
Now assume that $\epsilon_l\leq 1$ and define $h(\epsilon_l):=\min\{h|\epsilon_{l+h}\leq\epsilon_l/2\}$. As long as $\epsilon_{l+h}\geq\epsilon/2$, from (\ref{eq:imping}) we also have
\begin{eqnarray} \label{impsmall}\hat{g}(u^{l+h+1})-\hat{g}(u^{l+h}) &\geq &\hat{g}(u^{l+h})\frac{\epsilon_l}{4\omega_{l+h}(u^l)}\left(1-\frac{1+\epsilon_l/2}{1+\frac{\epsilon_l}{4\omega_{l+h}(u^l)}\omega_{l+h}(u^l)}\right)\nonumber\\ &\geq & -\frac{\epsilon_l^2}{32\omega}\hat{g}^*.\end{eqnarray} Again, the first inequality is obtained by setting $\lambda_l=\frac{\epsilon_l}{4\omega_{l+h}(u^l)}$. On the other hand, Lemma \ref{lem:dual_gapTR} gives\begin{eqnarray} \label{gapsmall}\frac{\hat{g}(u^l)}{\hat{g}^*}\leq {1+\epsilon_l}.\end{eqnarray} Combining equations (\ref{impsmall}) and (\ref{gapsmall}), we get $h(\epsilon_l)\leq \frac{32\omega}{\epsilon_l}$. Therefore \begin{eqnarray} \label{Leps} \mathcal{H}(\epsilon)&=&h(\epsilon_l)+h(\epsilon_l/2)+h(\epsilon_l/4)+\dots+h(\epsilon_l/2^{\left\lceil \ln{\epsilon_l}/\epsilon\right\rceil-1})\nonumber\\
&\leq & 32\omega\left(\frac{1}{\epsilon_l}+\frac{2}{\epsilon_l}+\frac{4}{\epsilon_l}+\dots+\frac{2^{\left\lceil \ln{\epsilon_l}/\epsilon\right\rceil-1}}{\epsilon_l}\right)\leq\frac{64\omega}{\epsilon}=\mathcal{O}(\epsilon^{-1}),\end{eqnarray} iterations are required to obtain an $\epsilon$-primal feasible solution starting with a solution $\epsilon_l\leq 1$. Combining (\ref{Leps}) and (\ref{L1}) completes the proof. \ifmmode\else\unskip\quad\fi\hbox{\rlap{$\sqcap$}$\sqcup$}
Once we take care of the drop steps, the analysis of the algorithm with away steps is no more complicated.
\begin{lemma}\label{lemm:CAlg2} As long as $u^l$ satisfy Assumption \ref{ass:upb} for all $l=1,2,\dots$, Algorithm \ref{TR-TY} finds an $\epsilon$-approximate optimal solution in at most \begin{eqnarray}\label{Alg2iter}\mathcal{L}(\epsilon)=\mathcal{O}(m+\epsilon^{-1})\end{eqnarray} steps. The constants hidden in the `big oh' are linearly dependent on the constant $\omega$ in Assumption \ref{ass:upb}.\end{lemma}
\bf Proof: \rm }%\par \noindent We can only have add/increase steps when $\epsilon_l\geq 1$; hence Algorithms \ref{TR-KH} and \ref{TR-TY} take the same steps until the first solution $u_{\hat{l}}$ with $\epsilon_{\hat{l}}\leq 1$ is encountered. So that \begin{eqnarray} \label{L2}\mathcal{L}(1)=\min\{l|\epsilon_l\leq1\}=\mathcal{O}(\ln m) \end{eqnarray} holds for Algorithm \ref{TR-TY} as well.
Now assume that $\epsilon_l\leq 1$ and define $h(\epsilon_l):=\min\{h|\epsilon_{l+h}\leq\epsilon_l/2\}$ as before. Let us look at the improvement in the objective function at the $(l+h)^{th}$ iteration. There are three cases:
\begin{enumerate}
\item If this is an add/increase step, then
\begin{eqnarray}
\hat{g}(u^{l+h+1})-\hat{g}(u^{l+h})&\geq & -\frac{\epsilon_l^2}{32\omega}\hat{g}^*
\end{eqnarray}
from (\ref{impsmall});
\item if it is a decrease step, we have
\begin{eqnarray}
\hat{g}(u^{l+h+1})-\hat{g}(u^{l+h})&\geq &\hat{g}(u^{l+h})\frac{-\epsilon_l}{4\omega_{l+h}}\left(1-\frac{1-\epsilon_l/2}{1-\frac{\epsilon_l}{4\omega_{l+h}}\omega_{l+h}}\right)\nonumber\\
&\geq & -\frac{\epsilon_l^2}{16\omega}\hat{g}^*;
\end{eqnarray}
\item otherwise (it is a drop step), we can only conclude that
\begin{eqnarray}
\hat{g}(u^{l+h+1})-\hat{g}(u^{l+h})&\geq 0.
\end{eqnarray}
\end{enumerate}
Hence we have \begin{eqnarray} \label{impsmall2} \hat{g}(u^{l+h+1})-\hat{g}(u^{l+h})&\geq & -\frac{\epsilon_l^2}{32\omega}g^*,\end{eqnarray} whenever we have an add/increase or decrease step.
On the other hand, using Lemma \ref{lem:dual_gapTR} we have\begin{eqnarray} \label{gapsmall2}\frac{\hat{g}(u^l)}{\hat{g}^*}\leq {1+\epsilon_l}.\end{eqnarray}
Combining equations (\ref{impsmall2}) and (\ref{gapsmall2}), we need to perform at most
$$h(\epsilon_l)\leq \frac{32\omega}{\epsilon_l}$$ add/increase and decrease steps to obtain an $\epsilon_l/2$-approximate optimal solution starting with an $\epsilon_l$-approximate optimal solution. Applying this argument repeatedly, we conclude that we need at most \begin{eqnarray} \mathcal{H}(\epsilon)&=&h(\epsilon_l)+h(\epsilon_l/2)+h(\epsilon_l/4)+\dots+h(\epsilon_l/2^{\left\lceil \ln{\epsilon_l}/\epsilon\right\rceil-1})\nonumber\\
&\leq & 32\omega\left(\frac{1}{\epsilon_l}+\frac{2}{\epsilon_l}+\frac{4}{\epsilon_l}+\dots+\frac{2^{\left\lceil \ln{\epsilon_l}/\epsilon\right\rceil-1}}{\epsilon_l}\right)\nonumber\\&\leq&\frac{64\omega}{\epsilon}=\mathcal{O}(\epsilon^{-1} ),\end{eqnarray} add/increase and decrease iterations to obtain an $\epsilon$-approximate optimal solution starting with an $\epsilon_l$-approximate optimal solution where $\epsilon_l\in(0,1]$. Since the number of drop steps is bounded above by the number of add steps plus $m$ (the number of positive components of the initial solution $u^0$), (\ref{Alg2iter}) is immediate.\ifmmode\else\unskip\quad\fi\hbox{\rlap{$\sqcap$}$\sqcup$}
The following lemma shows that (for the same set of data points) an approximate solution to the D-optimal design problem is also close to the optimal solution of the A-optimal design problem in some sense.
\begin{lemma} \label{lem:init_gapKY} Let $u^D$ be a $\delta$-primal feasible solution for the D-optimal design (as defined as in Definition \ref{def:sol} with $p=q=0$), then $u^D$ is an $(n+n\delta-1)$-primal feasible solution for $(\mathcal{D})$.\end{lemma}
\bf Proof: \rm }%\par \noindent For all $1\leq j \leq m$, we have \begin{eqnarray*}
x_j^T(M(u^D)X^T)^{-2}x_j&=&{\rm Trace}((M(u^D)X^T)^{-1}(M(u^D)X^T)^{-1/2}x_jx_j^T(M(u^D)X^T)^{-1/2})\\
&\leq&{\rm Trace}((M(u^D)X^T)^{-1}){\rm Trace}((M(u^D)X^T)^{-1/2}x_jx_j^T(M(u^D)X^T)^{-1/2})\\
&\leq&{\rm Trace}((M(u^D)X^T)^{-1}){\rm Trace}(x_j^T(M(u^D)X^T)^{-1}x_j)\\
&\leq&{\rm Trace}((M(u^D)X^T)^{-1})(n+n\delta),\end{eqnarray*}
where $U^D={\rm Diag}(u^D)$. This proves that $u^D$ is an $(n+n\delta-1)$-primal feasible solution for $(\mathcal{D})$.\ifmmode\else\unskip\quad\fi\hbox{\rlap{$\sqcap$}$\sqcup$}
Let us call the algorithm which finds a 1-approximate optimal solution for the D-optimal design problem using WA-TY method described in \cite{TYil07} and proceeds with Steps 1, 2, and 3 of Algorithm \ref{TR-KH} as Algorithm \ref{TR-KH}-MV; and that proceeds with Steps 1, 2, 3, and 4 of Algorithm \ref{TR-TY} as Algorithm \ref{TR-TY}-MV. When $m\gg n$, these algorithms perform significantly better than the original ones as the following lemma suggests. In addition, we are able to obtain core-set results for free.
\begin{lemma} \label{lemm:CAlgD} As long as $u^l$ satisfy Assumption \ref{ass:upb} for all $l=1,2,\dots$,
\begin{itemize}
\item[a.] Algorithm \ref{TR-KH}-MV finds an $\epsilon$-primal feasible solution in at most \begin{eqnarray}\label{Alg1Diter}\mathcal{L}(\epsilon)=\mathcal{O}(n\ln n+\epsilon^{-1})\end{eqnarray} steps;
\item[b.] Algorithm \ref{TR-TY}-MV finds an $\epsilon$-approximate optimal solution in at most \begin{eqnarray}\label{Alg2Diter}\mathcal{L}(\epsilon)=\mathcal{O}(\min\{m, n\ln n\}+\epsilon^{-1})\end{eqnarray} steps;
\item[c.] furthermore, Algorithm \ref{TR-KH}-MV identifies a set $\mathcal{A}\subset\mathcal{X}$ such that
$$|\mathcal{A}|\leq\mathcal{O}(n\ln n+\epsilon^{-1})$$
and an $\epsilon$-primal feasible solution $u$ for the A-optimal design problem defined over data set $\mathcal{A}$ is also an $\epsilon$-primal feasible solution for the A-optimal design problem defined over data set $\mathcal{X}$; and
\item[d.] Algorithm \ref{TR-TY}-MV identifies a set $\mathcal{A}\subset\mathcal{X}$ such that
$$|\mathcal{A}|\leq\mathcal{O}(n\ln n+\epsilon^{-1})$$
and an $\epsilon$-approximate optimal solution $u$ for the A-optimal design problem defined over data set $\mathcal{A}$ is also an $\epsilon$-approximate optimal solution for the A-optimal design problem defined over data set $\mathcal{X}$.
\end{itemize}\end{lemma}
\bf Proof: \rm }%\par \noindent It is proved in \cite{TYil07} that a 1-approximate optimal solution for the D-optimal design problem can be obtained in $\mathcal{O}(n\ln n)$ iterations. Let $u^0$ be such a solution. Lemmas \ref{lem:dual_gapTR} and \ref{lem:init_gapKY} give \begin{eqnarray} \label{ggTY}\hat{g}(u^0)\geq 2n\hat{g}^*.\end{eqnarray} Replacing (\ref{gg0}) with (\ref{ggTY}) in the proof of Lemma \ref{lemm:CAlg1}, gives $\mathcal{L}(1)=\mathcal{O}(\ln n)$ for Algorithm \ref{TR-KH}-MV. Since the rest of the proof is unchanged, Algorithm \ref{TR-KH}-MV finds an $\epsilon$-primal feasible solution in $\mathcal{L}(\epsilon)=\mathcal{O}(n\ln n+\ln n +\epsilon^{-1})=\mathcal{O}(n\ln n+\epsilon^{-1})$ iterations, which proves (a).
Similarly, (b) follows from Lemma \ref{lemm:CAlg2} with replacing $\mathcal{L}(1)=\mathcal{O}(\ln n)$ and noticing that the number of positive components in $u^0$ is bounded above by $\mathcal{O}(\min\{m, n\ln n\})$ as proved in \cite{KumYil05}.
Let $\hat{u}$ be the output of Algorithm \ref{TR-KH}-MV. Letting $\mathcal{A}=\{x_i: \hat{u}_i>0\}$ proves (c) since the number of positive components of $\hat{u}$ is bounded above by the number of positive components in the initial solution (which is $2n$ as discussed in \cite{KumYil05}) plus the number of add steps (which is less than the total number of iterations proved in part (a)). Similar arguments can be used to prove part (d).\ifmmode\else\unskip\quad\fi\hbox{\rlap{$\sqcap$}$\sqcup$}
\begin{remark} The complexity results we have presented in this section depend on the constant $\omega$ in Assumption \ref{ass:upb}. It is easy to see that $\omega$ is proportional to the inverse of the infimum of the set of eigenvalues of the matrices $M(u^l)$, $l = 1,2, \dots$ generated by the algorithm. Alternatively, $\omega$ is the supremum of the ellipsoidal distances of the data points with respect to the ellipsoids centered at the origin that have shape matrices $M(u^l)$. Therefore, $\omega$ depends on the geometry of the design points and the steps taken by the algorithm. When the design points are very thinly spread around a proper subspace of ${\rm I\kern-3.2pt R}^n$, the ellipsoids generated by the algorithm will have elongated axes in some directions and extremely short axes in others, potentially leading to large ellipsoidal distances for some data points. The sequence of positive definite matrices, $M(u^l)$, $l = 1,2, \dots$, generated by the algorithm converge to a single limit point, say $M(u^*)$, which is positive definite and has smallest eigenvalue, say $\lambda^*$. Therefore, there exists an integer $N(\frac{\lambda^*}{2})$, such that the eigenvalues of $M(u^l)$, for $l\geq N\left(\frac{\lambda^*}{2}\right)$ are lower bounded by $\lambda^*/2$. That guarantees that $\omega$ is finite.
\end{remark}
\section{Local Convergence Properties}\label{sec:TRloc}
In this section, we will show that Algorithms \ref{TR-TY} and \ref{TR-TY}-MV are locally linearly convergent, i.e., the number of iterations grows with $\mathcal{O}(\ln\epsilon^{-1})$ not $\mathcal{O}(\epsilon^{-1})$ asymptotically under certain assumptions. The typical behavior of the algorithms is demonstrated in Figure \ref{fig:figlinTR}. Unfortunately, this bound depends on the data of the problem as well as the dimensions and the constant $\omega$ defined as in Lemma \ref{lemm:CAlg2}, and so does not provide global complexity bounds better than those above.
\begin{figure}
\begin{tabular}{ll}
\centering
\includegraphics[width=3in]{gap_knTR.jpg}
&
\centering
\includegraphics[width=3in]{loglinTR.jpg}
\end{tabular}
\caption{Behavior of Algorithm \ref{TR-TY} for $(m,n)=(10000,100)$.}
\label{fig:figlinTR} \end{figure}
Let us look at the following perturbation of the primal problem ($\mathcal{P}$):
$$\begin{array}{rrrcl}
& \min & f(H) &:=& - 2\ln {\rm Trace} H^{1/2} \\
(\mathcal{P}(\kappa)) & & x_i^T H x_i & \leq & 1+\kappa_i, \, i =
1,\dots,m.
\end{array} $$ Given $u$ satisfying the $\epsilon$-approximate optimality conditions, let $H(u):=\frac{(M(u))^{-2}}{{\rm Trace}(M(u))^{-1}}$ and define $\kappa:=\kappa(u,\epsilon)$ as
\[
\kappa_i(u,\epsilon) := \left\{ \begin{array}{ll} \epsilon & \mbox{if } u_i = 0, \\
x_i^T H(u) x_i - 1 & \mbox{else.} \end{array} \right.
\] Note that, each component of perturbation vector $\kappa$ is absolutely bounded by $\epsilon$ and $u^T\kappa=\frac{\sum_{j:u_j>0}u_jx_j^T(M(u))^{-2}x_j}{{\rm Trace}(M(u))^{-1}}-1=1-1=0$. $H(u)$ is optimal w.r.t. $\mathcal{P}(\kappa(u,\epsilon))$, since it is feasible and $u$ provides the corresponding Lagrangian multipliers. Let $\phi(\kappa)$ be the value function, the optimal value of ($\mathcal{P}(\kappa)$). If $u^*$ is a vector of multipliers corresponding to the optimal solution of ($\mathcal{P}$), then $u^*$ is a subgradient of $\phi$ at 0. For any $\epsilon$-approximate optimal solution $u$ and $\kappa:=\kappa(u,\epsilon)$, we have \begin{eqnarray} g(u)=f(H(u))&=&\phi(\kappa)\geq\phi(0)+u^{*^T}\kappa\nonumber\\
&=&g^*+(u^*-u)^T\kappa\geq g^*-\|u-u^*\|\|\kappa\|.\end{eqnarray}
Since $f(H)$ is strongly convex near any $H\succ0$ and the constraints are linear, Robinson's second order condition holds at $(H,\hat{u})$ for any $\mathcal{P}(\kappa)$, where $H$ is the optimal solution and $\hat{u}$ is any Lagrangian multiplier. Moreover, the linear constraints are regular at any feasible point and they are polyhedral, therefore Robinson's Corollary 4.3 (\cite{Rob82}) applies, which shows that
$$\|u-u^*\|\leq L\|\kappa\|\leq L\sqrt{m}\epsilon,$$ where $L$ is a data-dependent constant and
whenever $\|\kappa\|$ is sufficiently small. Hence we conclude \begin{eqnarray}\label{gap}g^*-g(u)\leq M\epsilon^2\end{eqnarray}
for some $M$ depending on the data of the problem ($\mathcal{P}$). Using inequality (\ref{gap}), we can find a constant $\hat{c}$ such that \begin{eqnarray} \label{gap2} \frac{\hat{g}(u^l)}{\hat{g}^*}\leq e^{M\epsilon_l^2}\leq 1+\hat{c}\epsilon_l^2,\end{eqnarray} for any $\epsilon_l$-approximate solution $u^l$, as long as $\epsilon_l$ is small enough. Using (\ref{gap2}) instead of (\ref{gapsmall2}) in the last part of the proof of Lemma \ref{lemm:CAlg2} we obtain the following lemma:
\begin{lemma} \label{local22} Under the assumption of Lemma \ref{lemm:CAlg2}, there exists a data-dependent constant $Q$ such that Algorithms \ref{TR-TY} and \ref{TR-TY}-MV discussed above converges to an $\epsilon$-approximate optimal solution in $\mathcal{O}(Q+\ln(1/\epsilon))$ steps. \end{lemma}
\section{Computational Study} \label{sec:TRcomp}
In this section we present some computational results for Algorithms \ref{TR-KH} and \ref{TR-TY}, using different initialization strategies: the Khachiyan initialization (KH) strategy, where the initial feasible solution $u$ is the center of the unit simplex, i.e., $u_i=1/m$ for all $i=1,\dots,m$; the Kumar-Y{\i}ld{\i}r{\i}m initialization (KY) strategy introduced in \cite{KumYil05}; and a new strategy (MV) where the initial solution is set to be a 1-approximate optimal solution obtained by the WA-TY method of \cite{TYil07}. All experiments were carried out on a 3.40 GHz Pentium
IV processor with 1.0 GB RAM using MATLAB version R2006b. We assume a general linear model in this section and next. Note that we do not generate our regression points as fixed grid of support points on a compact interval as many other papers do. Instead we generate a large set of random regression points following \cite{SunFre02}. According to our past experience from \cite{AST08}, instances generated by this method are quite challenging. For all algorithms we study below, we report the total computational time inclusive of the time spent on the initialization schemes.
In Table \ref{comp_1TR}, we compare the computation time of the algorithms described above with three initializations on small- to medium-sized data sets. The data sets are generated as in \cite{SunFre02}. The results presented are the geometric means of the solution times for 10 random problems to obtain an $\epsilon$-primal feasible (for Algorithm \ref{TR-KH}) or an $\epsilon$-approximate optimal solution (for Algorithm \ref{TR-TY}) where $\epsilon=10^{-3}$. It is clear from the results that Algorithm \ref{TR-TY} preforms significantly better than Algorithm \ref{TR-KH} showing that away steps are necessary for developing efficient algorithms. For these instances, it is hard to make conclusions on the performances of the initialization strategies.
\begin{table}[ht]
\begin{center}
\caption{Geometric mean of solution times of Algorithms \ref{TR-KH} and \ref{TR-TY} for small-medium sized problems with different initializations}\label{comp_1TR} \vspace{.1in}
\begin{tabular}{|c|r||r|r|r||r|r|r|}\hline
& & \multicolumn{6}{|c|}{Geometric Mean of Time (Seconds)} \\\cline{3-8}
& & \multicolumn{3}{|c||}{Algorithm \ref{TR-KH}}&\multicolumn{3}{|c|}{Algorithm \ref{TR-TY}} \\
\cline{3-8} n & m & Kha &KY & MV & Kha & KY &MV \\ \hline
10 & 50 & 9.1 & 8.5 & 8.5 & 1.6 & 0.7 & 0.8 \\
10 & 100 & 10.5 & 10.3 & 10.1 & 1.2 & 1.3 & 1.9 \\
10 & 200 & 10.8 & 9.9 & 10.6 & 0.6 & 1.4 & 1.1 \\
10 & 400 & 11.9 & 11.2 & 12.5 & 0.4 & 0.8 & 1.0 \\
10 & 600 & 13.3 & 13.0 & 12.7 & 0.6 & 1.1 & 0.8 \\
10 & 800 & 13.9 & 13.4 & 13.4 & 1.0 & 1.5 & 1.2 \\
20 & 200 & 37.9 & 36.4 & 35.3 & 1.2 & 0.8 & 0.6 \\
20 & 300 & 39.6 & 40.0 & 39.2 & 1.4 & 1.1 & 1.0 \\
20 & 400 & 38.3 & 38.5 & 39.7 & 0.7 & 1.7 & 1.6 \\
20 & 600 & 49.2 & 49.2 & 45.7 & 0.9 & 2.0 & 2.9 \\
20 & 800 & 52.6 & 54.5 & 52.3 & 1.2 & 2.5 & 3.4 \\
20 & 1000 & 57.1 & 54.4 & 53.1 & 1.7 & 3.4 & 3.4 \\
20 & 1200 & 58.7 & 56.4 & 56.6 & 1.8 & 5.3 & 5.0 \\
30 & 450 & 108.6 & 100.1 & 93.9 & 2.0 & 2.9 & 2.8 \\
30 & 900 & 130.0 & 119.6 & 127.5 & 1.5 & 4.7 & 4.5 \\
30 & 1350 & 142.3 & 121.3 & 120.9 & 2.3 & 6.5 & 5.8 \\
30 & 1800 & 154.2 & 131.3 & 128.9 & 3.5 & 7.6 & 7.7 \\
\hline
\end{tabular}\end{center}
\end{table}
\begin{table}[ht]
\begin{center}
\caption{Geometric mean of solution times of Algorithms \ref{TR-KH} and \ref{TR-TY} for large problems with different initializations}\label{comp_2TR} \vspace{.1in}
\begin{tabular}{|c|r||r|r|r||r|r|r|}\hline
& & \multicolumn{6}{|c|}{Geometric Mean of Time (Seconds)} \\\cline{3-8}
& & \multicolumn{3}{|c||}{Algorithm \ref{TR-KH}}&\multicolumn{3}{|c|}{Algorithm \ref{TR-TY}} \\
\cline{3-8} n & m & Kha &KY & MV & Kha & KY &MV \\ \hline
5 & 10000 & 17.267 & 12.208 & 11.641 & 35.236 & 3.5327 & 3.5428 \\
5 & 20000 & 26.57 & 20.417 & 20.905 & 55.491 & 7.8292 & 7.4747 \\
5 & 30000 & 35.941 & 29.808 & 30.374 & 43.136 & 7.9607 & 9.8677 \\
5 & 50000 & 58.433 & 54.698 & 52.828 & 98.456 & 28.159 & 28.715 \\
10 & 10000 & 43.677 & 32.431 & 32.173 & 38.017 & 5.7187 & 5.5486 \\
10 & 20000 & 76.886 & 67.377 & 66.554 & 138.93 & 10.604 & 10.154 \\
10 & 30000 & 103.56 & 87.166 & 90.091 & 126.69 & 17.158 & 15.499 \\
20 & 10000 & 141.76 & 113.23 & 117.45 & 48.849 & 18.482 & 19.234 \\
20 & 20000 & 211.44 & 186.48 & 183.35 & 196.31 & 40.659 & 39.256 \\
20 & 30000 & 287.15 & 253.81 & 252.65 & 385.37 & 53.223 & 45.749 \\
20 & 50000 & 426.9 & 395.6 & 402.68 & 543.22 & 99.232 & 91.305 \\
30 & 10000 & 295.09 & 247.77 & 243.47 & 59.061 & 27.439 & 31.508 \\
30 & 20000 & 451.68 & 395.66 & 402.26 & 220.01 & 74.113 & 61.231 \\
30 & 30000 & 606.04 & 536.8 & 528.98 & 500.77 & 89.2 & 96.194 \\
50 & 50000 & 2308.2 & 2154.5 & 2142.8 & 1992.3 & 370.77 & 327.79 \\
\hline
\end{tabular}\end{center}
\end{table}
\begin{table}[ht]
\begin{center}
\caption{Geometric mean of solution times of Algorithm \ref{TR-TY}-MV with different (update) selection strategies for small instances}\label{comp_3TR} \vspace{.1in}
\begin{tabular}{|c|r||r|r||r|r|}\hline
& & \multicolumn{2}{|c||}{Time (Seconds)}&\multicolumn{2}{|c|}{Iterations} \\
\cline{3-6} n & m & ALL & Orig. & ALL & Orig. \\ \hline
20 & 200 & 0.54 & 0.85 & 510.7 & 1697.9\\
20 & 300 & 0.67 & 1.16 & 638.5 & 2252\\
20 & 400 & 0.91 & 1.72 & 772.08 & 3122\\
20 & 600 & 1.45 & 2.02 & 904.9 & 3254\\
20 & 800 & 2.01 & 2.57 & 1028.9 & 3918.6\\
20 & 1000 & 2.67 & 3.41 & 1189.9 & 4836.6\\
20 & 1200 & 3.00 & 5.35 & 1195.3 & 6397\\
30 & 450 & 1.26 & 2.90 & 963.3 & 4467.3\\
30 & 900 & 2.82 & 4.71 & 1314.6 & 5723.7\\
30 & 1350 & 4.68 & 6.59 & 1660.4 & 6976.3\\
30 & 1800 & 6.33 & 7.67 & 1782.9 & 7706.8\\
20 & 1000 & 2.54 & 3.49 & 1168.4 & 4694.9\\
\hline
\end{tabular}\end{center}
\end{table}
\begin{table}[ht]
\begin{center}
\caption{Geometric mean of solution times of Algorithm \ref{TR-TY}-MV with different (update) selection strategies for large instances}\label{comp_4TR} \vspace{.1in}
\begin{tabular}{|c|r||r|r||r|r|}\hline
& & \multicolumn{2}{|c||}{Time (Seconds)}&\multicolumn{2}{|c|}{Iterations} \\
\cline{3-6} n & m & ALL & Orig. & ALL & Orig. \\ \hline
10 & 10000 & 13.33 & 5.71 & 875.8 & 2656.5\\
20 & 10000 & 26.08 & 18.48 & 1634.5 & 6072.5\\
20 & 20000 & 59.32 & 40.61 & 1879.8 & 6852.7\\
20 & 30000 & 102.14 & 62.41 & 2220.3 & 7854.7\\
30 & 10000 & 42.95 & 27.43 & 2547.9 & 7100.8\\
30 & 20000 & 101.86 & 74.11 & 3085.6 & 10515\\
30 & 30000 & 140.42 & 89.2 & 2876.5 & 8899.2\\
50 & 50000 & 428.3 & 370.7 & 5106.4 & 15979\\
\hline
\end{tabular}\end{center}
\end{table}
Table \ref{comp_2TR} presents the performance of the algorithms on larger data sets. Again, the results are the geometric means of the solution times of 10 random problems generated as in \cite{SunFre02} for each parameter set. The results indicate that for these instances where $m\gg n$, the MV initialization is outperforming the Khachiyan initialization as Lemma \ref{lemm:CAlgD} suggests. Since the KY initialization is somehow close to the MV initialization, its performance it similar to the MV initialization. One should not be surprised by the fact that Algorithm \ref{TR-TY} with the Khachiyan initialization is very slow on these instances, since the initial solution has many entries with positive weights and the algorithm needs to take many drop steps before converging to the optimal solution. Fortunately, other two initializations are able to find accurate solutions in short time. We have tried even larger data sets to explore the limits of the algorithms. We were able to find $10^{-4}$-approximate optimal solutions to instances where $n=500$ and $m=10000$ (generated as before) with Algorithm \ref{TR-TY} using KY initialization under 30 minutes.
The number of iterations required can be significantly decreased if we make the best possible update (not just one of the two arguments used in Step 1) at each iteration. This can be done by calculating the improvement related to each index and choosing the best. We have coded a version of Algorithm \ref{TR-TY}-MV and experimented on some of the data sets above. The (mean) solution times and number of iterations are compared in Tables \ref{comp_3TR} and \ref{comp_4TR}. The unmodified version of the algorithm is represented in the columns labeled with `'Orig.`' while the version with optimal decisions is labeled with `'ALL'`. It is obvious that as the number of points in the data set increase calculating the possible improvement for each index becomes expensive; hence considering only two promising vertices is a wise choice. Obviously some hybrid versions, which choose the best of a small set of carefully selected indices, can perform better for certain instances; so can other versions with active set strategies.
\section{Semidefinite Programming Reformulation and Comparison} \label{sec:TRcompSDP}
Any reader with some familiarity with nonlinear optimization would know that semidefinite programming has gained significant attention in last two decades. As discussed in \cite{WSV00}, many interesting problems in science and engineering can be reformulated as SDPs and solved via one of the freely available SDP solvers such as SDPT3 or SEDUMI. The D-optimal and A-optimal design problems are no exception. Section 4 of \cite{VB99} provides the reformulations of both of these problems. Following their discussion, problem $(\mathcal{D})$ is equivalent to:
$$\begin{array}{cccl}
\min & \sum_{i=1}^n{t_i} \\
(\mathcal{SDP}) & \quad \left(\begin{array}{cc} M(u) & e_i \\ e_i^T & t_i \end{array} \right) \succeq 0 , \, i =
1,\dots,n,\\
& e^Tu = 1,\\
& u \geq 0,
\end{array}$$
where $e_i$ is the $i^{th}$ unit vector in ${\rm I\kern-3.2pt R}^n$, and the variables are $u \in {\rm I\kern-3.2pt R}^m$ and $t \in {\rm I\kern-3.2pt R}^n$.
Although many problems can be cast as semidefinite programs, not many semidefinite formulations can be solved efficiently yet due to high memory requirements and slow convergence rate of the state-of-the-art methods. We compare one of our algorithms (Algorithm \ref{TR-TY} with KY initialization) versus the SDPT3 algorithm using the CVX platform on MATLAB, which is a classic platform to solve SDPs. The results presented in Table \ref{comp_SDP} are mean solution times for 5 random problems to obtain an $\epsilon$-approximate optimal solution with Algorithm \ref{TR-TY} in the third column and with the SDP solver on the forth. For fair comparison, we run both algorithms until a very accurate solution is obtained (i.e., $\epsilon=10^{-7}$), especially since being able find accurate solutions is one of the strong points of the SDP approach. In this section, we test only 5 instances of each problem since the SDP solver takes very long amount of time and the conclusion is obvious even with small number of instances considered. The instances are generated as before following \cite{SunFre02}. The solutions obtained from the two methods were identical (to be precise: the norm of their distance was smaller than $10^{-7}$ as expected). It is clear that our first-order technique dominates the SDP method, sometimes it is more than 300 times faster. Furthermore, it is impossible to solve large instances of the SDP formulation due to memory restrictions and time limitations. For example, we can not solve problems with $n=30$ and $m=600$ with the SDP solver.
\begin{table}[ht]
\begin{center}
\caption{Mean solution times of SDPT3 and Algorithm 2-MV for small-medium sized problems}\label{comp_SDP} \vspace{.1in}
\begin{tabular}{|c|c|c|c||c|}\hline
n & m & Algorithm \ref{TR-TY}& SDP & speed-up \\ \hline
10 & 50 & 0.33 & 0.58 & 3.20 \\
10 & 100 & 0.24 & 0.45 & 3.63 \\
10 & 200 & 0.70 & 1.51 & 1.90 \\
10 & 400 & 0.76 & 2.80 & 5.32 \\
10 & 600 & 1.56 & 7.83 & 3.01 \\
10 & 800 & 1.37 & 9.90 & 12.18 \\
10 & 1000 & 0.88 & 11.43 & 7.85 \\
20 & 50 & 0.08 & 2.66 & 42.40 \\
20 & 100 & 0.33 & 4.83 & 8.20 \\
20 & 200 & 0.44 & 11.40 & 19.64 \\
20 & 400 & 0.92 & 33.46 & 55.77 \\
20 & 600 & 1.15 & 66.73 & 54.49 \\
20 & 800 & 2.02 & 120.82 & 66.19 \\
20 & 1000 & 1.84 & 197.45 & 115.66 \\
30 & 50 & 0.05 & 17.07 & 334.33 \\
30 & 100 & 0.19 & 22.36 & 145.58 \\
30 & 200 & 0.89 & 58.38 & 38.88 \\
30 & 400 & 1.07 & 140.58 & 187.42 \\
30 & 600 & 2.13 & 337.95 & 139.66 \\
\hline
\end{tabular}\end{center}
\end{table}
\section{Conclusions}
In this paper, we will develop a Frank-Wolfe type algorithm for the A-optimal experimental design problem. Our approach is similar to the Frank-Wolfe type algorithms developed for the D-optimal experimental design problem. Nevertheless, we are the first to discuss global and local convergence of the algorithms rigorously for the A-optimal experimental design problems.
\section{Acknowledgements}
The author would like to express her gratitude to Prof. Mike Todd for bringing the experimental design problem to her attention during her PhD candidacy at Cornell University and his constant support and mentorship. She is also thankful to the two anonymous reviewers for their valuable comments and suggestions.
|
\section{Introduction}
\label{sec:entropy-errors}
SPH solutions to the compressible Euler equations are characterized by
an anomalous ``blip'' or kink in the pressure at the contact
discontinuity. The density profile is accurate which means the
internal energy shows a corresponding heating/cooling to mirror the
pressure jump. The error, once introduced, neither grows nor
attenuates without dissipation and is simply advected with the
particles at the local material velocity. Monaghan and
Gingold~\cite{monaghan-gingold-shock} were the first to observe this
behaviour when they applied SPH to simulate shock-tube problems.
Their observations lead them to ascribe the phenomenon to general
``starting'' errors when discontinuous initial profiles are
used. Presumably, SPH struggles with the discontinuous thermal
energy. Resolving this behaviour has been the focus of numerous
researchers over the last thirty years as this is manifestly a grave
drawback of the method. Despite this error, SPH has been found to be
useful within the astrophysics community, it's application often
preceded by ``code-comparisons'' with existing Eulerian
techniques. One of the early comparisons was undertaken by Davies et
al.~\cite{davies93}, who compared SPH simulations of stellar
collisions with the Piecewise Parabolic Method (PPM). They suggest
that the advantages of each approach are mutually exclusive, although
the two approaches were qualitatively similar. Caution is advised in
extending this observation for other calculations in which different
hydrodynamic effects determine the solution. About the same time,
Steinmetz and M\"uller~\cite{steinmetz-muller1993} had also suggested
that SPH and finite difference methods should be looked upon as
complimentary methods to solve hydrodynamic problems. In their seminal
work, Agertz et al.~\cite{sphvgrids} performed a comprehensive
comparison of astrophysical codes (using GADGET~\cite{gadget-code} for
SPH) for the simulation of interacting multi-phase fluids. The
un-physical pressure jump at a density gradient, as produced by SPH in
it's standard formulation was found to render the method incapable of
resolving hydrodynamic instabilities like the Kelvin-Helmholtz (KHI)
or Rayleigh-Taylor instabilities (RTI). A similar comparison was
carried out by Tasker et al.~\cite{taskeretal08} for test problems
with analytical solutions, therefore enabling a more quantitative
comparison. While SPH was found to be generally comparable in it's
accuracy with the Eulerian schemes, a major difference was the
pressure jump at the contact discontinuity, which is absent for
grid-based codes. For the hydrodynamics of multi-phase fluids (more
generally at a density gradient), this spurious pressure jump behaves
like an artificial surface tension force, inhibiting the development
of density driven instabilities like KHI. In another study, Okamoto et
al.~\cite{okamoto03} observed that the erroneous pressure jump can
also result in spurious momentum transfer across shearing flows,
significantly affecting numerical results. These code comparisons
rekindled the need to resolve the spurious pressure at the contact
discontinuity, with the Kelvin-Helmholtz instability (KHI) often used
as a canonical ``mixing'' problem exposing the method's
vulnerability. \newline Among the many tricks for
SPH~\cite{herant-tricks}, arguably the oldest one is a judicious use
of artificial dissipation. Thermal conduction is as old as artificial
viscosity itself with Monaghan~\cite{monaghan-review} and
Brookshaw~\cite{brookshaw94} being early advocates for it's use in
treating ``wall-heating'' errors. It has been used for example, by
Sigalotti et al.~\cite{sigalotti2006-shock,sigalotti2008-shock} and
Rosswog and Price~\cite{rosswog-price07} for strong shock problems in
hydrodynamics and magnetohydrodynamics respectively. Addressing the
mixing problem originally highlighted by Agertz et
al.~\cite{sphvgrids}, Price~\cite{price2008-khi} demonstrated that a
judicious use of thermal conduction enables a suitable description of
density driven instabilities like KHI. The conduction terms are
formulated using the signal-based artificial
viscosity~\cite{monaghan1997-riemann} and are constructed to result in
a diffusion of energy across the contact discontinuity. The use of and
need for similar thermal conduction terms was also suggested by
Wadsley et al.~\cite{Wadsley08}, Garc\'ia-Senz et
al.~\cite{Garcia-Senz01012009} and and Valcke et al.~\cite{valcke10}
for mixing problems in astrophysics. Some authors also suggest that
apart from the use of the thermal conduction, the magnitude of the
pressure jump can be curtailed by relaxing the initial conditions and
by using a modified kernel with an increased sampling. Thermal
conduction is necessary for the long term simulation and to avoid
``oily''~\cite{valcke10} or ``gloopy''~\cite{read10} features in the
solution. Following Price, Valdarnini~\cite{valdarnini12}, Kawata et
al.~\cite{kawata13} and Rosswog~\cite{rosswog09-review} have also
advocated the use of artificial thermal conduction. By using an error
and stability analysis, Read et al.~\cite{read10} showed that the
inability of SPH to adequately resolve mixing was due in part to a
``Local Mixing Instability'' (LMI), whereby, particles are inhibited
to mix on the kernel scale due to entropy conservation, which in turn
results in a pressure discontinuity. The LMI is therefore another term
for the pressure ``blip'' in the context of hydrodynamic mixing. The
LMI was cured by using a modified density estimate, similar to that
employed by Ritchie and Thomas~\cite{RT01}, to ensure a single valued
pressure throughout the flow. The modified density approaches
(\cite{RT01,MarriWhite03,read10}) are designed for a more accurate
density estimation for multi-phase fluids (mixing problems) in
pressure equilibrium. Consequently, they perform poorly for flows with
strong shocks. Indeed, in a recent article, Read and
Hayfield~\cite{read12} discuss a new high-order dissipation switch for
adaptive viscosity in which they forgo the modified density approach
in favour of an artificial heating term as proposed by
Price~\cite{price2008-khi}.\newline
Moving away from adding thermal conduction in a somewhat ad-hoc
manner, Price~\cite{price2012} argues that the assumption of a
differentiable density is the cause of the spurious pressure jump. The
density estimate plays a central role in the variational formulation
of SPH and is used to define an implicit particle volume through the
ratio of particle mass to particle density. Saitoh and
Makino~\cite{saitoh12} took cue from this idea to develop a
density-independent SPH (DISPH) by replacing the mass density by an
equivalent pressure density and it's arbitrary
function. Hopkins~\cite{hopkins12} also considered the idea of
replacing the particle volume, traditionally defined by the mass
density, by an arbitrary smoothed function. A family of equivalent
Lagrangian schemes are derived by different choices of the
function. In particular, the pressure-entropy formulation was shown to
be superior at resolving mixing in the hydrodynamic context. However,
there appear to be problems for shocked flows (due to the
non-isentropic nature of the flow), similar to the modified density
approach occurs for this formulation as well~\cite{sphgal}.\newline
The SPH formulations discussed hitherto were of the variational kind
with the use of explicit dissipation
terms. Inutsuka~\cite{inutsuka-riemann} developed an
artificial-viscosity free scheme that requires the solution of a
Riemann problem between interacting particle pairs. The Riemann solver
introduces the necessary and sufficient dissipation required to
stabilize the scheme. Although the pressure blip is present for these
Godunov SPH (GSPH) schemes, it is less pronounced. The result is a
more suitable description of fluid instabilities like KHI. Indeed, Cha
et al.~\cite{cha-inutsuka-khi} found GSPH to be superior to the
standard SPH for the development of KHI. They argue in favour of the
linear consistency in the GSPH momentum equation and a more accurate
Lagrangian function used in GSPH. Murante et al.~\cite{murante-gsph}
observed similar advantages of GSPH for the simulation of
hydrodynamical instabilities vis-a-vis standard SPH. Another
artificial-viscosity free SPH scheme was proposed by
Lanzafame~\cite{lanzafame09} by considering shock flows as
non-equilibrium events. The equation of state is reformulated
according to a Riemann problem to introduce the necessary
dissipation. Incidentally, these GodunovSPH schemes typically include
an implicit thermal conduction term which is known
(\cite{price2008-khi}) to ameliorate the pressure jump. Indeed, the
authors have shown~\cite{purigsph13} that GSPH with a class of
approximate Riemann solvers is essentially equivalent to the standard
SPH with artificial viscosity and thermal conduction terms.\newline
It is worth noting that despite the numerous efforts to address this
pressure jump, the error is at best ameliorated, with it's adverse
effects kept to within acceptable limits. Care must be exercised with
the use of thermal conduction so as to avoid excessive smearing and a
resulting loss of accuracy. This is achieved through viscosity
limiters~\cite{cullen-walter,read12,taylor12,price2012} and solution
dependent conductivity coefficients
\cite{sigalotti2008-shock,price2012}. Without a unifying theory
however, the different approaches seem serendipitous and somewhat
ad-hoc. Success of a particular method notwithstanding, a discernible
pattern among all proposed solutions is the introduction of a certain
dissipation into the equations of motion to handle the pressure
jump. The dissipation is introduced either in the form of an explicit
(\cite{springel2010,price2008-khi,price2012}) or implicit (GSPH)
thermal conduction
(\cite{inutsuka-riemann,cha-inutsuka-khi,murante-gsph}), or by more
subtle means via the state equation \cite{lanzafame09}, density
estimate~\cite{read10} or particle
volume~\cite{saitoh12,hopkins12}.\newline That dissipation can be used
to progressively smear the pressure blip once it is created should be
fairly obvious. A more fundamental question that can be asked perhaps,
pertains to the origin of this error. Towards this goal, we search for
similar behaviour in finite difference/volume schemes. These
grid-based schemes have received a great deal of attention and success
within the CFD community and it is therefore helpful to study them. In
particular, if we can relate the SPH errors to those generated with a
suitable finite volume scheme, we can gain new insights and a more
satisfying explanation as to why the aforementioned approaches
work.\newline
As it turns out, a pressure blip, exactly analogous to SPH, is
produced when using the Lagrange plus Eulerian remap version of the
PPM code, PPMLR~\cite{colella-woodward-ppm,vh1,cmhog}. Remap schemes
involve a Lagrangian advection step in which the cells move, followed
by a conservative remap onto the original Eulerian grid. We find
(agreeing with the argument of Davies et al.~\cite{davies93}) it
highly unlikely that two fundamentally different approaches result in
the same erroneous features. Interestingly, the Eulerian version of
the PPM (PPMDE) and indeed, other Eulerian schemes
\cite{toro-book,leveque-book} do not exhibit this anomaly. Lagrangian
finite difference codes have traditionally fallen out in favour of
their Eulerian counterparts and we are led to conjecture therefore
that the difference between the two versions of the PPM scheme can
provide an answer to origin of the pressure jump in SPH.\newline
This work is outlined as follows. In Sec.~\ref{sec:evidence}, we use a
one-dimensional shock tube problem to provide numerical evidence to
the claim that PPMLR exhibits, qualitatively the same errors as the
SPH pressure jump at the contact discontinuity. In
Sec.~\ref{sec:blip-discussion}, we use the spatiotemporal behaviour of
the SPH error to draw an analogy with ``wall-heating'' errors for
traditional finite difference schemes and argue that the pressure jump
is a result of a spurious entropy generation in the initial transient
phase of shock formation. In Sec.~\ref{sec:explanation}, through a
comparison of PPMLR and PPMDE schemes, a lack of diffusion in the
material wave (contact) is highlighted as the source of the error and
we demonstrate how it may be eliminated for PPMLR by using a more
diffusive Lagrangian advection step. Finally, in
Sec.~\ref{sec:sph_application}, we use these ideas to propose a minor
modification to the GSPH scheme, where the requisite dissipation is
added through a suitable choice of an approximate Riemann solver
(similar to the method proposed by Shen et al.~\cite{ShenYan10}). The
scheme is applied to standard test problems in one and two dimensions
to demonstrate it's effectiveness in mitigating the pressure jump for
flows with strong shocks. We summarize this work in
Sec.~\ref{sec:summary}, with conclusions drawn from this work and
suggestions for possible extensions. All numerical results presented
in this manuscript are generated using the code (SPH2D)\footnote{SPH2D
is available at https://bitbucket.org/kunalp/sph2d}, which is freely
available for validation and use.
\section{Numerical evidence of the pressure blip}
\label{sec:evidence}
In this section, we provide numerical evidence for the existence of
the pressure discontinuity when using SPH and the Lagrange plus remap
version of the Piecewise Parabolic Method (PPMLR
\cite{colella-woodward-ppm,vh1}) finite difference code. The error is
particularly severe for SPH and generally arises across a density
gradient (contact discontinuity). We consider two one-dimensional
shock tube problems that work well to highlight the anomalous
behaviour. The initial conditions are defined with the the left (l)
and right (r) states displayed in Table.~\ref{tab:cases}
\begin{table}[h]
\centering
\begin{tabular}{|l | c | c | c| c| c| r|} \hline
Test & $\rho_l$ & $p_l$ & $u_l$ & $\rho_r$ & $p_r$ &$u_r$\\ \hline
Sod's shock-tube & 1& 1& 0& 0.125& 0.1& 0 \\ \hline
Blast-wave & 1 & 1000 & 0 & 1 & 0.01 & 0 \\ \hline
\end{tabular}
\caption{Initial data for the test problems to highlight the pressure
discontinuity}
\label{tab:cases}
\end{table}
The first test is the famous Sod's shock tube~\cite{sod1978-survey}
problem. This test represents the bare minimum a numerical scheme for
the compressible Euler equations should hope to resolve. The initial
conditions result in a right moving contact discontinuity sandwiched
between a left moving rarefaction and a right moving shock wave. The
solution is self similar with four constant states separated by the
three waves. The second test is a more stringent version, referred to
as the blast-wave problem, in which the magnitude of the initial
pressure jump ($p_l/p_r$) is $10^5$. The wave structure is similar
with a strong, right facing shock wave ($M = 198$) moving into the low
pressure gas. \newline For the SPH results, we use our SPH2D
\cite{Puri2013-I} code that implements the variational formulation
described by Price \cite{price2012}. Thermal conduction is turned off
for both test problems to highlight the errors at the contact
discontinuity. For the PPMLR method, we have used John Blondin's VH-1
\cite{vh1} and Jim Stone's CMHOG \cite{cmhog} codes independently to
verify our results. Both codes implement the PPMLR algorithm proposed
by Colella and Woodward \cite{colella-woodward-ppm}.
\subsection{Test $1$: Sod's shock-tube problem}
\label{sec:sod-shock-evidence}
\begin{figure}[!h]
\begin{center}
\includegraphics[width=\columnwidth]{sodshock-sph-ppmlr.pdf}
\caption{Numerical pressure for Sod's shock-tube problem using SPH
(dots), PPMLR (solid blue line) at $t = 0.15s$, compared with
the exact solution (solid black line). The pressure ``blip'' is
clearly visible in the SPH solution at the contact $x\approx
0.17$. A close-up of the solution around the contact is shown in
the inset.}
\label{fig:sodshock-pressure-evidence}
\end{center}
\end{figure}
The numerical pressure profiles for the SPH and PPMLR schemes is shown
in Fig.~\ref{fig:sodshock-pressure-evidence}. The SPH simulation was
performed using a total of $450$ particles. Initially, $400$ particles
were placed to the left of the initial discontinuity ($x = 0$) with
spacing $\Delta x_l = 0.00125$. The remaining $50$ particles were
placed to the right of $x = 0$, with a spacing of $\Delta x_r =
0.01$. The particle mass was set equal to the inter-particle spacing
$\Delta x_l$ so that $m/\Delta x$ reproduces the desired
density. For the PPMLR results, we used a total of $500$ grid cells
(zones). The pressure discontinuity is clearly visible for the SPH
results. A close up of the solution in the vicinity of the contact is
shown in the inset. For this relatively simple problem, PPMLR does not
exhibit the anomalous pressure jump at the contact. The situation is
different for the blast-wave problem discussed next.
\subsection{Test $2$: Blast-wave problem}
\label{sec:blastwave-evidence}
\begin{figure}[!h]
\begin{center}
\includegraphics[width=\columnwidth]{blastwave-sph-ppmlr.pdf}
\caption{Numerical pressure for the blast-wave problem using SPH
(dots), PPMLR (solid blue line) at $t = 0.01s$ compared with the
exact solution (solid black line). The pressure blip is visible
for both SPH and PPMLR (see inset) schemes.}
\label{fig:blastwave-pressure-evidence}
\end{center}
\end{figure}
Numerical pressure profiles for the SPH and PPMLR schemes are shown in
Fig.~\ref{fig:blastwave-pressure-evidence}. We notice that the blip
which was absent in the PPMLR solution for the Sod's shock-tube
problem is now present. Another striking feature is the huge jump in
the pressure for the SPH solution. The results were generated using a
total of $500$ particles for SPH and $500$ grid cells for PPMLR. At
the outset, it may seem that the SPH results are no-where in
comparison to PPMLR but this is not the case. This is because the
PPMLR scheme has an inherent diffusion for the thermal energy which
works to dissipate the error with time. Recall that we explicitly
switched off the thermal conduction for the SPH scheme. With a small
amount of thermal conduction, the results of the two schemes are
similar as
\begin{figure}[!h]
\begin{center}
\includegraphics[width=\columnwidth]{blastwave-sphcond-ppmlr.pdf}
\caption{Numerical pressure for Blast-wave problem using SPH with
thermal conduction (dots), PPMLR (solid blue line) at $t =
0.01s$ compared with the exact solution (black line). The
pressure blip produced by each of the schemes is now similar.}
\label{fig:blastwave-pressure-evidence2}
\end{center}
\end{figure}
can be seen in Fig.~\ref{fig:blastwave-pressure-evidence2}, where the
magnitude of the SPH pressure jump is dramatically reduced and is only
slightly larger than that of the PPMLR scheme. Having established the
presence of the numerical error for both schemes, we now proceed to a
discussion on the nature of the error and provide an explanation for
it's occurrence.
\section{A discussion on the error}
\label{sec:blip-discussion}
In the previous section, we provided numerical evidence to support our
original claim that the ubiquitous pressure jump at the contact
discontinuity for SPH solutions, occurs for a class of finite
difference schemes. In particular, the Lagrange plus remap version of
the Piecewise Parabolic Method (PMPLR) results in an error remarkably
similar to SPH. We note that although code comparisons for SPH and PPM
in different contexts have previously been conducted
(\cite{davies93,oshea-etal05,sphvgrids,taskeretal08}), this error has
rather strangely gone unnoticed or has not been reported. It is
possible that the comparisons were carried out with the fully Eulerian
version of PPM, as two step (Lagrange plus remap) codes have
traditionally fallen out in favour of their fully Eulerian
counterparts. Indeed, Woodward and Colella
\cite{woodward-colella1984}, in their comparison of numerical schemes
for flows with strong shocks showed that the cell-centred, direct
Eulerian version of the PPM scheme (PPMDE) was the most accurate. In a
more recent study, Pember and Anderson \cite{pember-anderson00} argue
otherwise, stating that the two approaches yield generally equivalent
results. Nevertheless, direct Eulerian schemes are undoubtedly more
prevalent and higher order versions of these schemes do not exhibit
the SPH-like pressure jump at the contact, as can be verified by any
of the schemes presented in the monographs by Toro \cite{toro-book}
and LeVeque \cite{leveque-book}. We believe the differences between
the remap and direct Eulerian finite difference schemes could provide
insight into the nature of the pressure jump in SPH. We would like to
remind the reader that although we know of a suitable ``fix'' in the
form of thermal conduction, we are looking for a consistent
explanation for it's origin and a justification for the myriad
approaches outlined in the introduction.
\subsection{The nature of the error}
\label{sec:error-nature}
A natural question to ask of a numerical scheme is convergence to the
physically correct solution. The pressure jump at the contact, being
clearly erroneous, raises valid questions as to the behaviour of the
error with the spatial resolution. It is instructive therefore, to
catalogue known features of the pressure jump in the SPH context. We
continue with the strong shock problem of
Sec.~\ref{sec:blastwave-evidence} as the canonical example exposing
this behaviour for SPH. Thermal conduction is switched off to avoid
cosmetic smoothing of the results. Concerning the question of
numerical convergence, we first examine the error as we increase the
number of particles.
\begin{figure}[!h]
\begin{center}
\includegraphics[width=\columnwidth]{blastwave-sph-convergence.pdf}
\caption{Numerical pressure for Blastwave problem using SPH with
different resolutions. The solution profiles are expectedly more
crisp for higher resolutions. The spread of the error reduces
with increasing resolution but peak magnitude of the error
remains constant.}
\label{fig:blastwave-pressure-convergence}
\end{center}
\end{figure}
Fig.~\ref{fig:blastwave-pressure-convergence} shows the results for
the blastwave problem when we have used $100$, $200$ and $400$
particles respectively. The shock transition region is sharper with
higher resolution as expected. We notice that the spread of the error
\emph{reduces} with increasing resolution but the peak point-wise
error remains \emph{constant}. Convergence in $L_{\infty}$ is
therefore not possible in SPH. Convergence in $L_{1}$ with a
convergence rate or approximately $2$ has been observed previously
\cite{springel2010}. The temporal behaviour of the error can be
studied through Fig.~\ref{fig:blastwave-pressure-thist}, which shows
snapshots of the pressure profile at the times $t=0.0025$, $t =
0.005$, $t = 0.0075$ and $t=0.01$. These results highlight another
feature of the error.
\begin{figure}[!h]
\begin{center}
\includegraphics[width=\columnwidth]{blastwave-sph-time-evolution.pdf}
\caption{Numerical pressure for Blastwave problem at the time
instants $t=0.0025$, $t = 0.005$, $t = 0.0075$ and $t=0.01$
seconds, when simulated with SPH. Once the pressure ``blip'' is
generated, it is advected without dissipation. Thermal
conduction was explicitly turned off to highlight this
behaviour.}
\label{fig:blastwave-pressure-thist}
\end{center}
\end{figure}
Namely, once the error is created, the pressure jump simply advects
with the flow and neither grows nor attenuates without explicit
thermal conduction. Artificial viscosity has no discernible effect on
the solution. This is not surprising when we consider that the
artificial viscosity is added to generate the correct entropy jump
across the shock and therefore, has no effect at the contact
discontinuity. To summarize, the pressure jump at the contact
discontinuity, in the context of SPH has the following properties:
\begin{itemize}
\item The error, once created is simply advected with the material
velocity.
\item With the increase of spatial resolution, the spread of the
error decreases whilst maintaining a constant peak magnitude.
\item Artificial viscosity has no role to play in suppressing or
mitigating it's effect.
\end{itemize}
\subsection{The relation to wall-heating errors}
\label{sec:wall-heating}
A multi-valued pressure in the presence of an accurate density profile
results, in a corresponding jump in the thermal energy, via the
equation of state. The pressure jump in SPH can therefore be thought
of as a spurious heating/cooling of the fluid. For the numerical
solution of the compressible Euler equations, ``wall-heating'' is the
canonical term given to errors that result in a spurious rise in
thermal energy (heating). The problem has generated considerable
interest and has received the attention of several
researchers~\cite{RiderWallHeating}. The term was coined by
Noh.~\cite{noh1978-wallshock} when he considered the problem of shock
reflection in planar, cylindrical and spherical geometries. The reason
for our interest in these errors is the similarity it bears with the
SPH errors at the contact discontinuity. For instance, under mesh
refinement, Noh observed that the overheating decreases in spatial
range while maintaining it's peak point-wise value. This leads Noh to
conjecture that the error is built into the exact solution of the
modified viscous equations. He further argues that any numerical
method based on ``shock-smearing'' (artificial viscosity) will
demonstrate this error. By considering the asymptotic solution of the
governing equations with artificial viscosity,
Menikoff~\cite{Menikoff94} argues that entropy errors, resulting in
spurious heating/cooling are generated not only for shock reflection,
but also for shock interaction or when a shock passes through a change
in mesh spacing. The errors are generated over a short transient phase
as a result of the numerical width of the shock. After the initial
transient phase, the entropy errors are \emph{frozen} and simply
advected with the fluid and the only recourse is to add thermal
conduction which leads to a diffusion of the energy. A similar
analysis in Lagrangian coordinates was carried out by Shen et
al.~\cite{ShenYan10}. Once again, entropy errors were observed to
occur over a short, initial transient phase of shock
interaction/reflection.\newline These observations for wall heating
errors in the context of traditional finite volume methods bear a
striking similarity to the SPH pressure/entropy errors. Thus, we
conjecture that the SPH errors are a form of spurious heating that
occurs in an initial transient phase. Once the entropy errors are
generated, diffusion of entropy (thermal conduction) is the only
recourse to mitigate it's effect.
\section{An Explanation for the Error}
\label{sec:explanation}
We argued that the pressure jump at the contact discontinuity for SPH
has symptoms of wall-heating errors previously observed for
traditional finite difference/volume codes. Essentially, an entropy
error is generated during the initial transient phase of shock
formation, the magnitude of which is independent of the spatial
resolution. After the initial transient, the error is convected along
the particle trajectories without dissipation. Monaghan and
Gingold.~\cite{monaghan-gingold-shock} had originally ascribed this
anomalous behaviour for SPH, to generic ``starting'' errors. More than
twenty years later, Tasker et al.~\cite{taskeretal08} had also
suggested that the discontinuous initial conditions give rise to an
entropy error. Since the errors are generated at start-up, and
subsequently passively advected with the particles, thermal conduction
is required to mitigate it's effect. This was the also the conclusion
drawn by Noh.~\cite{noh1978-wallshock} when he proposed an artificial
heat flux for finite difference schemes. While this reasoning serves
to justify the artificial conductivity approach in treating the error,
it sheds no light onto the origins of the error. Towards this aim, we
adopt a different perspective by studying grid-based schemes.\newline
Recall that in Sec.~\ref{sec:evidence}, the errors for the Lagrange
plus Eulerian remap version of the PPM scheme (PPMLR) was shown to be
qualitatively similar to those of SPH. Agreeing with the reasoning of
Davies et al.~\cite{davies93}, we find it highly unlikely that two
fundamentally different schemes (SPH and PPMLR) would result in the
same erroneous features. This leads us to believe that both schemes
solve a similar modified equation, the solution of which exhibits the
heating and corresponding pressure jump at the contact. A scaling
argument similar to Noh.~\cite{noh1978-wallshock} shows that the
magnitude of the error is independent of the spatial resolution. This
justifies to some extent the claim that the error is built into the
exact solution of the discrete SPH equations. It is therefore
reasonable to assume that a solution to the problem within the finite
volume context using PPMLR might provide clues for a similar
resolution in SPH. This can be done by comparing the Lagrangian plus
remap PPM scheme, PPMLR, with it's direct Eulerian counterpart, PPMDE,
for which the error is absent.
\subsection{PPMLR and PPMDE}
\label{sec:ppm}
The Piecewise Parabolic Method (PPM) \cite{colella-woodward-ppm} is a
high order, Godunov finite difference method that has, as it's
building block, a third order advection scheme. Along with the
ENO/WENO type schemes \cite{JiangShu96}, PPM is considered to be a
highly accurate method for the compressible Euler equations
\cite{woodward-colella1984}. The scheme, as originally proposed by
Colella and Woodward~\cite{colella-woodward-ppm}, can be formulated to
follow either Lagrangian or Eulerian hydrodynamics. Although the two
equation sets are mathematically the same, their numerical solutions
exhibit differences which we are interested in.\newline In what
follows, we focus on the development of the one-dimensional PPM scheme
since multi-dimensional extensions are constructed with the
dimensional splitting approach. Thus, essential details of the method
and the error producing mechanism in particular are contained within
the one-dimensional scheme. In a such a scheme, the conserved
variables are the specific volume $1/\rho$, velocity $u$, and the
specific energy $\hat{e} = e + \frac{1}{2} u^2$ for the Lagrangian
formulation, and density $\rho$, momentum $\rho u$, and total energy
$\rho \hat{e}$, for the Eulerian formulation. Both versions of the PPM
scheme (PPMLR, PPMDE) advance the solution (vector of conserved
variables $\boldsymbol{U}$) over physical zones or cells. The generic
$i^{\text{th}}$ cell has it's center at $x_i$, left and right faces at
$x_{i-\frac{1}{2}}$ and $x_{i+\frac{1}{2}}$ respectively and $\Delta
x_i = x_{i+\frac{1}{2}} - x_{i-\frac{1}{2}}$ denotes the cell volume.
\subsubsection{PPMLR}
\label{sec:ppmlr}
The conservative equations for Lagrangian hydrodynamics are given as
\begin{subequations}
\label{eq:lagrangian-hydrodynamics}
\begin{align}
\tau_t - u_{\xi} &= 0 \\
u_t + p_{\xi} &=0 \\
\hat{e}_t + (pu)_{\xi} &=0
\end{align}
\end{subequations}
where $\tau = 1/\rho$ is the specific volume, $\hat{e} =
\frac{1}{2}u^2 + e$ is the total energy per unit mass and the time
derivative is to be understood as a derivative moving with the fluid
(material derivative) $\frac{d}{dt}(*) = \frac{\partial}{\partial
t}(*) + u \nabla(*)$. $\xi$ is the \emph{mass} coordinate, which is
related to the spatial coordinate $x$, through the transformation
$d\xi = \rho dx$. The system is hyperbolic with eigenvalues $\lambda_1
= -C$, $\lambda_2 = 0$ and $\lambda_3 = C$, where $C = (\gamma p
\rho)^{\frac{1}{2}}$ is the Lagrangian sound speed. The convective
terms are absent in this formulation which is reflected as a wave of
speed $0$.\newline In PPMLR, the procedure to advance the solution is
carried out in two steps. In the first step, piecewise parabolic
interpolations of the pressure, velocity and density are used to
compute the effective \emph{left} and \emph{right} states for a
Riemann problem between two adjacent cells. Since the zone edge is
moving with the fluid velocity, the input state is determined purely
by the acoustic modes $\lambda_1$ and $\lambda_3$. With the input
(\emph{left}, \emph{right}) states constructed from the parabolic
reconstructions, a Riemann problem is solved to calculate fluxes
through the cell boundaries. The vector of conserved variables are
updated using the conservative differencing equations:
\begin{subequations}
\label{eq:ppmlr-advection}
\begin{align}
x_{j+\frac{1}{2}}^{n + 1} &= x_{j+\frac{1}{2}}^{n} + \Delta t
\overline{u}_{j+\frac{1}{2}} \\ \tau_{j+\frac{1}{2}}^{n
+ 1} &= \frac{1}{\Delta m_j}\left(x_{j+\frac{1}{2}}^{n + 1} -
x_{j-\frac{1}{2}}^{n + 1}\right) \\ u_{j+\frac{1}{2}}^{n
+ 1} &= u_{j+\frac{1}{2}}^{n} + \frac{\Delta t}{\Delta
m_j}\left(\overline{p}_{j-\frac{1}{2}} -
\overline{p}_{j+\frac{1}{2}}\right)
\\ \hat{e}_{j+\frac{1}{2}}^{n + 1} &= u_{j+\frac{1}{2}}^{n} +
\frac{\Delta t}{\Delta
m_j}\left(\overline{u}_{j-\frac{1}{2}}\overline{p}_{j-\frac{1}{2}}
- \overline{u}_{j+\frac{1}{2}}\overline{p}_{j+\frac{1}{2}}\right)
\end{align}
\end{subequations}
In these equations, $\overline{p}$ and $\overline{u}$ denote the
intermediate pressure and velocity that results from the solution to
Riemann problem at a zone edge. Convection is introduced through the
cell motion. After the advection step, the solution is conservatively
re-mapped onto the original Eulerian grid. A necessary condition for
high-order Godunov schemes, non-linearity is introduced by using
limiters and monotonicity constraints in the piecewise parabolic data
reconstruction.
\subsubsection{PPMDE}
\label{sec:ppmde}
The one-dimensional equations for Eulerian, inviscid hydrodynamics are
given as
\begin{subequations}
\label{eq:eulerian-hydrodynamics}
\begin{align}
\rho_t + (\rho u)_x &= 0 \\
(\rho u)_t + (\rho u^2 + p)_x &= 0\\
(\rho \hat{e})_t + (\rho \hat{e} u + pu)_x &=0
\end{align}
\end{subequations}
where the symbols have the same meaning and $(*)_t$ is the partial
derivative with respect to time. The system of equations is again
hyperbolic with eigenvalues $\lambda_1 = u-c$, $\lambda_2 = u$ and
$\lambda_3 = u + c$, where $c = \rho C$, is the Eulerian sound
speed. The procedure to update the solution in the one-step, direct
Eulerian formulation is essentially the same as PPMLR. Piecewise
parabolic interpolations of the dependent variables are used to
compute effective \emph{left} and \emph{right} states for Riemann
problems between adjacent cells. Zone fluxes, computed from the
Riemann solution are used in a conservative differencing scheme.
\begin{subequations}
\label{eq:ppmde-step}
\begin{align}
\boldsymbol{U}_{j+\frac{1}{2}}^{n+1} &=
\boldsymbol{U}_{j+\frac{1}{2}}^{n} + \frac{\Delta t}{\Delta x_j}
\left( \boldsymbol{F}(\overline{\boldsymbol{U}}_{j-\frac{1}{2}}) -
\boldsymbol{F}(\overline{\boldsymbol{U}}_{j+\frac{1}{2}}) \right)
\\ \overline{\boldsymbol{U}}_{j+\frac{1}{2}} &\approx
\frac{1}{\Delta
t}\int_{t^n}^{t^{n+1}}\boldsymbol{U}(x_{j+\frac{1}{2}}, t) dt
\end{align}
\end{subequations}
The difference is in the construction of the input states for the
Riemann problem. This is now more complicated than the Lagrangian case
as there may now be as many as $3$ and as few as $0$ waves impinging
on a zone edge from a given side. A consequence of this is that in
general, for the input state at a given zone edge, contributions from
each wave family must be accounted for \cite{colella-woodward-ppm}.
\subsection{Diffusion in the material wave}
\label{sec:ppmde-diffusion}
Both PPMLR and PPMDE employ piecewise parabolic interpolations of the
cell centered density, pressure and velocity to construct the input
left and right states as integral averages over the characteristic
domain of dependence. The domain of dependence for a given wave family
is defined by tracing back the path of the wave if it impinges on the
zone edge from a given side. For the Lagrangian formulation, we have
two waves corresponding to the acoustic modes, $\lambda_1$ and
$\lambda_3$. The material wave is absent as the cells are assumed to
move with the local fluid velocity ($\lambda_2 = 0$). For the Eulerian
formulation, waves from each of the three families can impinge on an
edge from a given side. The input state in this case is constructed
such that the amount of wave associated with each family of
characteristics transported across a zone edge is correct up to terms
of second order \cite{colella-woodward-ppm}. Thus, the additional
material wave must be accounted for in the Eulerian formulation.
Diffusion across this wave is the main difference between the two
versions of PPM. Indeed, Eulerian Godunov schemes are known to be more
diffusive at the contact when compared to Lagrangian
formulations. This was highlighted by Woodward and Colella
\cite{woodward-colella1984} and later by Pember and
Anderson~\cite{pember-anderson00}, when they compared remap and direct
Eulerian finite volume schemes.
\begin{figure}[!h]
\begin{center}
\includegraphics[width=\columnwidth]{blastwave-ppmlr-pcmlr.pdf}
\caption{Results for the blast-wave problem using a remap code
with piecewise parabolic (PPMLR) and piecewise constant (PCMLR)
interpolations. The entropy error (pressure jump) is eliminated
with when using the diffusive PCMLR method.}
\label{fig:bwave-ppmlr-pcmlr}
\end{center}
\end{figure}
This suggests a \emph{lack} of dissipation is actually the cause of
the SPH-like entropy error for PPMLR as observed in
Sec.~\ref{sec:evidence}. It was suggested by Jim Stone (private
communication, August 2013) that the low dissipation in PPM causes
these ``start-up'' errors. The discontinuous initial conditions gives
rise to additional waves on a discrete level which is captured by the
low dissipative schemes like PPMLR. One is then tempted to verify the
hypothesis by constructing a more diffusive version of the PPMLR
scheme. The easiest way to do this is to use a piecewise constant
reconstruction instead of the parabolic reconstruction used in PPMLR.
Results for the blast-wave problem using such a scheme (PCMLR) is shown
in Fig.~\ref{fig:bwave-ppmlr-pcmlr}. The solution is expectedly less
crisp than PPMLR but remarkably, the pressure jump at the contact is
eliminated. We note that it is not possible for the remap phase of
PPMLR to introduce the error. This is because remapping can be viewed
as a projection and is inherently a diffusive process. As a result, no
new extrema can occur in this step. The error is therefore generated
in the Lagrangian advection phase.\newline What does dissipation in
the material wave look like? To answer this question, we consider the
eigenstructure of the equations in the Lagrangian formulation
(Eq.~\ref{eq:lagrangian-hydrodynamics}). The right and left
eigenvectors for this hyperbolic system is given as (\cite{Rider94})
\begin{equation}
\label{eq:lagrangian-eigenstructure}
R =
\begin{pmatrix}
1 & 1 &1 \\ C &0 &-C \\ uC -p &\frac{p}{\gamma -1} &-uC -p
\end{pmatrix},
\,\,\,R^{-1} =
\begin{pmatrix}
\frac{1}{2\gamma}& \frac{1}{2C} + \frac{u(\gamma-1)}{2\gamma p} & \frac{1-\gamma}{2\gamma p} \\
\frac{\gamma-1}{\gamma}& \frac{u(\gamma-1)}{\gamma p} & \frac{\gamma-1}{\gamma p} \\
\frac{1}{2\gamma}& -\frac{1}{2C} + \frac{u(\gamma-1)}{2\gamma p} & \frac{1-\gamma}{2\gamma p}
\end{pmatrix}
\end{equation}
where $C = (\gamma p \rho)^{\frac{1}{2}}$ is the Lagrangian sound
speed. A conservative finite volume scheme with a general diffusive
flux contribution can be defined as
\begin{equation}
\label{eq:lagrangian-diffusive-flux}
\left(\boldsymbol{F}_{j+\frac{1}{2}}\right)_{\text{diss}} = -
\frac{1}{2}\sum_{k=1}^{3}\hat{\boldsymbol{r}}_k
|\hat{\lambda}_k|\left(\hat{\boldsymbol{r}}_k^{-1} \cdot \Delta
\hat{\boldsymbol{U}}_{j+\frac{1}{2}} \right),
\end{equation}
where the caret denotes a suitably averaged value at the zone
interface $x_{j+\frac{1}{2}}$. This is the numerical flux for
\emph{linearized} schemes such as Roe's scheme~\cite{Roe81} and is
applicable to SPH~\cite{monaghan1997-riemann,purigsph13}. The
diffusive flux in Eq.~\ref{eq:lagrangian-diffusive-flux} computes
\emph{jumps} across each wave family. The magnitude of the jump (wave
strength), $\alpha = \hat{\boldsymbol{r}}_k^{-1} \cdot \Delta
\hat{\boldsymbol{U}}$, is weighted by the wave speed. The final
contribution to the conserved variables is determined by the right
eigenvector for that wave family. For the Lagrangian scheme, this
contribution vanishes for the material wave ($k = 2$), since
$\lambda_2 = 0$. This is the contribution we are interested in, the
algebraic form of which is given by
\begin{equation}
\label{eq:lagrangian-material-wave-dissipation}
\left(f_{j+\frac{1}{2}}^2\right)_{\text{diss}} =
-|\lambda_2|\frac{1}{2}\,\begin{bmatrix} \frac{1}{2C}
+\frac{u(\gamma-1)}{2\gamma p}& \frac{u(1-\gamma)}{\gamma p} &
-\frac{1}{2C} +\frac{u(\gamma-1)}{2\gamma p}\end{bmatrix}
\cdot \begin{bmatrix} \Delta \hat{\tau} \\ \Delta\hat{u} \\ \Delta
\hat{e} \end{bmatrix} \begin{bmatrix} 1 \\ 0 \\ \frac{p}{\gamma
-1} \end{bmatrix}
\end{equation}
Due to the structure of the right eigenvector ($\boldsymbol{r}_2$ in
Eq.~\ref{eq:lagrangian-eigenstructure}), the dissipation acts on the
first and third components of $\boldsymbol{U}$. For the Lagrangian
formulation, these are the specific volume $\tau = 1/\rho$, and total
energy per uni mass $\hat{e} = \frac{1}{2}u^2 + e$ respectively. Thus,
dissipation in the material wave would result in an additional density
and energy diffusion \emph{simultaneously} that is absent in a purely
Lagrangian formulation.
\section{Application to SPH}
\label{sec:sph_application}
The lack of dissipation in the material wave, coupled with the low
diffusion of PPMLR is responsible for the entropy errors, and hence
the SPH-like pressure jump at the contact. The introduction of
dissipation helped eliminate the error for the finite volume remap
code in Sec.~\ref{sec:explanation}. It is then reasonable to assume
that an improvement of results can be expected for SPH if this
dissipation is somehow introduced. Indeed, this has been the adopted
practice within the SPH community, with dissipation often introduced
directly through thermal conduction
\cite{sigalotti2006-shock,price2008-khi,Wadsley08,Garcia-Senz01012009,rosswog09-review,valcke10,read12},
or via surrogate means such as using a smoother estimate to define
particle volume~\cite{saitoh12,hopkins12}, relaxing initial conditions
\cite{valcke10,read10} and a modification to the equation of state
\cite{lanzafame09}. The problem with the dissipation introduced in
these schemes is that they appear serendipitous and their reasoning
belies the simplicity of the SPH formulation. The requirement that
dissipation should act across the material wave provides a consistent
explanation as to why the aforementioned approaches work. From
Eq.~\ref{eq:lagrangian-material-wave-dissipation}, we know that a
combination of dissipation in the density and energy variables is
required to suppress the entropy errors. Dissipation for velocity
(artificial viscosity) has no role to play. This was verified
numerically in Sec.~\ref{sec:blip-discussion}. Armed with this
knowledge, we can attempt to introduce the requisite dissipation in a
consistent manner for SPH, thereby validating our hypothesis.
\subsection{Adding diffusion to SPH}
\label{sec:sph_diffusion}
We consider the GSPH formulation \cite{inutsuka-riemann} with an
approximate Riemann solver. We have shown (\cite{purigsph13}) that
with a suitable choice of an approximate Riemann solver, this
formulation is equivalent to a variational SPH scheme with artificial
dissipation and thermal conduction. The advantage of this formulation
is the explicit control of the dissipation through numerical fluxes
akin to finite difference/volume schemes. The discrete SPH equations
in this formulation, for the density, velocity and thermal energy are
given as
\begin{subequations}
\label{eq:gsph-equations-simple}
\begin{align}
\label{eq:gsph-summation-density}
\rho_a &= \sum_{b\in\mathcal{N}(a)} m_b W_{ab}(h_a)\\
\label{eq:gsph-simple-mom}
\ddot{\boldsymbol{x}}_a &= -\sum_{b \in \mathcal{N}(a)}m_b
p^*_{ab}\left(\frac{1}{\rho_a^2}\nabla W_{ab}(h_a) +
\frac{1}{\rho_b^2}\nabla W_{ab}(h_b)\right) \\
\label{eq:gsph-simple-enr}
\dot{e}_a &= -\sum_{b \in \mathcal{N}(a)}m_b
p^*_{ab}[\boldsymbol{v}_{ab}^{*} - \dot{\boldsymbol{x}}_a] \left(
\frac{1}{\rho_a^2}\nabla W_{ab}(h_a) +
\frac{1}{\rho_b^2}\nabla W_{ab}(h_b)\right)
\end{align}
\end{subequations}
where, the starred quantities ($p^*, \, \boldsymbol{v}^*$) are the
intermediate state arising from the solution of a Riemann problem
between two interacting particles. The solution to the Riemann problem
introduces the minimum and necessary dissipation required to stabilize
the scheme.\newline We follow the approach proposed by Shen et
al.~\cite{ShenYan10} by constructing a hybrid scheme in which a
regular Riemann solver is used in the momentum equation equation, and
a \emph{diffusive} Riemann solver is used for the energy
equation. Recall that the flux vector for the conservative equations
in the Lagrangian formulation are $\left(-u, p, pu\right)$. Thus,
using a more \emph{dissipative} intermediate velocity is akin to
introducing dissipation in the density and energy equations. In
\cite{purigsph13}, we evaluated $5$ different approximate Riemann
solvers in Lagrangian coordinates for use with GSPH. From an analysis
of an accuracy test for the Euler equations, we find that the Harten,
Lax, van Leer and Einfeldt (HLLE)~\cite{Rider94} solver is a suitable
choice for the \emph{diffusive} approximate Riemann solver. For the
regular Riemann solver, we can use any one from the exact
\cite{vanleer-muscl5}, Ducowicz \cite{Ducowicz85} or Roe \cite{Roe81}
approximate Riemann solvers.
\begin{figure}[!h]
\begin{center}
\includegraphics[width=\columnwidth]{gsph-hybrid-blastwave-pressure.pdf}
\caption{Numerical pressure profiles for the blastwave problem
using standard GSPH (blue) and the hybrid GSPH (green) using the
HLLE approximate Riemann solver, compared with the exact
solution (ref). The HLLE solver is successful in suppressing the
entropy errors as can be seen in the inset plot.}
\label{fig:gsph-hybrid-blastwave-pressure}
\end{center}
\end{figure}
We construct such a scheme, where the two approximate Riemann solvers
used are the van Leer exact and the HLLE approximate Riemann
solver. In particular, the diffusive contribution is constructed as
\begin{subequations}
\label{eq:gsph-hybrid-vstar}
\begin{align}
u^* &= \frac{1}{t_f}(tu^*_{\text{regular}} +
(t_f - t)u^*_{\text{diff}}), \\
p^* &= \frac{1}{t_f}(tp^*_{\text{regular}} +
(t_f - t)p^*_{\text{diff}}),
\end{align}
\end{subequations}
where, $t_f$ is the final time in the simulation. This corresponds to
a linear blending of the two estimates with a more diffusive estimate
used in the initial stages of the computation. Since the errors are
expected to be generated at start-up, the blending avoids excessive
dissipation that may ruin the
solution. Fig.~\ref{fig:gsph-hybrid-blastwave-pressure} shows the
numerical pressure profiles for the standard (blue) and hybrid GSPH
(green), compared with the exact solution (red), when using $1000$
equal mass particles. As expected, the dissipation helps to suppress
the entropy error, with only a slight kink visible in the inset plot.
\begin{figure}[!h]
\begin{center}
\includegraphics[width=\columnwidth]{gsph-hybrid-blastwave-all-variables.pdf}
\caption{Numerical solution (dots) for the hybrid GSPH compared
with the exact solution (red). The hybrid scheme shows good
agreement with the exact solution with a negligible pressure
jump. The dissipation has the desired effect of acting on the
contact discontinuity.}
\label{fig:gsph-hybrid-blastwave-all-variables}
\end{center}
\end{figure}
The suppression of the pressure jump at the contact discontinuity does
not have an adverse effect on the profiles of the other physical
variables as can be seen in
Fig.~\ref{fig:gsph-hybrid-blastwave-all-variables}, which shows the
numerical solution for the hybrid GSPH scheme (dots), compared with
the exact solution (red line). The dissipation has the desired effect
of acting on the contact discontinuity as can be observed by the
slightly smeared density and thermal energy profiles. near $x \approx
0.15$.
We would like to point out that this method produces improved results
than with using a traditional scheme with larger thermal conduction
parameters. Fig.~\ref{fig:mpm-conduction-parameter} shows the result
of simply increasing the thermal conduction parameter $\alpha_u$, for
the variational scheme of Monaghan, Price and
Morris~\cite{monaghan1997-riemann,morris-monaghan1997-switch,price2012}.
\begin{figure}[!h]
\begin{center}
\includegraphics[width=\columnwidth]{mpm-conduction-parameter.pdf}
\caption{Effect of increasing the thermal conduction parameter
$\alpha_u$ for the MPM scheme. Larger values of the parameter
(right) mitigate the jump in the pressure (upper panel) but also
generates a corresponding dip in the velocity profile (lower
panel).}
\label{fig:mpm-conduction-parameter}
\end{center}
\end{figure}
Larger values of the thermal conduction parameter works to supress the
presure discontinuity as expected. However, results in an unwanted dip
in the velocity profile at the contact discontinuity. This behaviour
has been recently reported by Sirotkin and Yoh~\cite{sirotkin_yoh2013}
in their SPH scheme with approximate Riemann solvers. In comparison,
the hybrid GSPH scheme (cf
Fig.~\ref{fig:gsph-hybrid-blastwave-all-variables}) does not produce
this behaviour.
\subsection{Consequences of adding dissipation}
\label{sec:consequences}
The linear blending of the two estimates for $\boldsymbol{v}^*$
through Eq.~\ref{eq:gsph-hybrid-vstar} was shown to work well to
suppress the pressure jump for the one-dimensional blast-wave
problem. Since the errors are expected to be generated at start-up,
it's use can be detrimental for long time simulations.
\begin{figure}[!h]
\begin{center}
\includegraphics[width=\columnwidth]{wcblast-linear-blending.pdf}
\caption{Density profiles at $t = 0.038$ for the Woodward and
Colella blast-wave problem using standard GSPH (blue) and the
hybrid modification (green) using
Eq.~\ref{eq:gsph-hybrid-vstar}. The contact discontinuities near
$x = 0.1$ and $x = 0.3$ are heavily smeared for the hybrid
scheme and the solution within $x \in [0.15, 0.3]$ has lost some
detail.}
\label{fig:wcblast-linear-blending}
\end{center}
\end{figure}
For example, Fig.~\ref{fig:wcblast-linear-blending} shows the density
profiles for the Woodward and Colella blast-wave problem
\cite{woodward-colella1984,Puri2013-I} using standard GSPH (blue) and
the hybrid modification (green) at the time $t = 0.038$. The extra
dissipation in the hybrid scheme has resulted in an excessive smearing
of the contact discontinuities near $x = 0.1$ and $x = 0.3$, and a
loss of detail within the region $x \in [0.15, 0.3]$. Note that adding
dissipation to the material wave (contact discontinuity) is exactly
what we set out to do, although Eq.~\ref{eq:gsph-hybrid-vstar} results
in an over diffusive scheme. This can be corrected by defining the intermediate states as
\begin{subequations}
\label{eq:gsph-hybrid-vstar-exp}
\begin{align}
u^* &= u^*_{\text{regular}} +
e^{-\alpha\frac{t}{t_f}}\left(u^*_{\text{diff}} -
u^*_{\text{regular}}\right) \\
p^* &= p^*_{\text{regular}} +
e^{-\alpha\frac{t}{t_f}}\left(p^*_{\text{diff}} -
p^*_{\text{regular}}\right)
\end{align}
\end{subequations}
This corresponds to an exponential decay with time, for the diffusive
component. The \emph{parameter} $\alpha$, controls the rate of decay
(growth) of the two velocity estimates, with higher values resulting
in a more rapid decay (growth).
\begin{figure}[!h]
\begin{center}
\includegraphics[width=\columnwidth]{wcblast-exp-blending.pdf}
\caption{Density profiles at $t = 0.038$ for the Woodward and
Colella blast-wave problem using standard GSPH (blue) and the
hybrid modification (green) using
Eq.~\ref{eq:gsph-hybrid-vstar-exp} and $\alpha = 10$. The
solution is expectedly more crisp with a good agreement in the
region $x \in [0.15, 0.3]$}
\label{fig:wcblast-exp-blending}
\end{center}
\end{figure}
Fig.~\ref{fig:wcblast-exp-blending} shows the density profiles at $t =
0.038$ for the same problem when we have this blending function with
$\alpha = 10$. The density profile for the hybrid GSPH (green) scheme
is expectedly more crisp with an improved agreement with the second
order GSPH (blue) scheme in the region $x \in [-0.15, 0.3]$. \newline
Fig.~\ref{fig:wcblast-exp-blending} shows the density profiles at $t =
0.038$ for the same problem when we have this blending function with
$\alpha = 10$. The density profile for the hybrid GSPH (green) scheme
is expectedly more crisp with an improved agreement with the second
order GSPH (blue) scheme in the region $x \in [-0.15, 0.3]$. \newline
We concede that a \emph{tuning} parameter to control dissipation
perhaps goes against the ethos of a GSPH scheme that is inherently
parameter free. While one can argue that the choice of the Riemann
solver itself in the GSPH scheme can be thought of as a
\emph{parameter}, the issue we want to highlight is the subtle role of
dissipation, that is needed for stability but criticized when used in
excess. An ideal scheme should use just the right amount of
dissipation for \emph{all} problems, the requisite amount, in turn,
ideally determined by the scheme itself (adaptive schemes). High-order
Godunov methods (MUSCL, PPM, ENO/WENO) are generally accepted to fit
this ideal. However, Quirk \cite{quirk94} famously pointed out several
instances where these schemes fail or produce erroneous
results. Moreover, he suggests that most of these errors can be
overcome by a judicious use of artificial dissipation. The trick is to
avoid a proliferation of \emph{tuning} parameters to determine the
requisite dissipation. We are faced with a similar conundrum while
constructing our hybrid GSPH scheme. Dissipation must be somehow
introduced into the purely inviscid equations and in this work we have
argued in favour of a specific form, acting across the energy and
density variables to suppress entropy errors. Given the transient
nature of the error, we are forced to introduce a \emph{parameter}
that limits the extra dissipation to when it is needed.
\subsection{Extension to higher dimensions}
\label{sec:2d}
We used the one-dimensional blast-wave problem as the canonical test
highlighting the SPH entropy errors. The manner in which we chose to
introduce the dissipation is however, not limited to the
one-dimensional case. This can bee seen in
Fig.~\ref{fig:gsph-hybrid-blastwave2d-pressure}, which shows the
numerical (dots) pressure for the hybrid scheme (green) compared with
the reference second order GSPH scheme (blue) with the van Leer
Riemann solver, for a two-dimensional blastwave problem.
\begin{figure}[!h]
\begin{center}
\includegraphics[width=\columnwidth]{gsph-hybrid-blastwave2d-pressure.pdf}
\caption{Numerical pressure for the hybrid GSPH (green) scheme
with $\alpha = 3.0$ and with the choice of HLLE as the diffusive
Riemann solver, compared with the standard second order GSPH
scheme using the van Leer exact Riemann solver (blue). The
results are exactly analogous to the one-dimensional case
(Fig.~\ref{fig:gsph-hybrid-blastwave-pressure}), with the hybrid
scheme working well to limit the pressure jump.}
\label{fig:gsph-hybrid-blastwave2d-pressure}
\end{center}
\end{figure}
For the hybrid scheme, we use the exponential smoothing given by
Eqs.~\ref{eq:gsph-hybrid-vstar-exp} with $\alpha = 3.0$, and choose
the HLLE solver as the diffusive Riemann solver. The suppression of
the pressure blip has no discernible adverse effect on the other
variables as can be seen in
Fig.~\ref{fig:gsph-hybrid-blastwave2d-all-variables}, which shows the
density (left), velocity (center) and thermal energy (right) for the
hybrid GSPH scheme (lower panel), compared with a second order GSPH
scheme using the van Leer exact Riemann solver (upper panel).
\begin{figure}[!h]
\begin{center}
\includegraphics[width=\columnwidth]{gsph-hybrid-blastwave2d-all-variables.pdf}
\caption{Numerical solution for the hybrid GSPH scheme (lower
panel) with $\alpha = 3$ and with the choice of HLLE as the
diffusive Riemann solver, compared with the standard second
order GSPH scheme (upper panel). The hybrid scheme works to
suppress the pressure blip
(Fig.~\ref{fig:gsph-hybrid-blastwave2d-pressure}) and has no
discernible adverse effect on the other physical variables. In
particular, the spike in the thermal energy behind the contact
is eliminated. Additionally, the contact discontinuity is
slightly smeared for the density.}
\label{fig:gsph-hybrid-blastwave2d-all-variables}
\end{center}
\end{figure}
The hybrid scheme eliminates the spike in the thermal energy behind
the contact. Additionally, analogous to the one-dimensional case, the
contact discontinuity is slightly smeared as can be seen in the
density plot around $x \approx 0.6$. The results were generated using
a total of $20000$ equal mass particles initially distributed in a
hexagonal close paced arrangement.
\section{Summary and further work}
\label{sec:summary}
In this work, we attempted to provide an explanation for the origin of
the ubiquitous pressure jump in SPH simulations of the compressible
Euler equations. The anomalous behaviour has been observed since the
dawn of SPH~\cite{monaghan-gingold-shock} and has received attention
ever since. It has been highlighted as a drawback of the method when
compared with traditional Eulerian
schemes~\cite{sphvgrids,taskeretal08}, and has led some researchers to
develop particle tessellation techniques as an alternative to
SPH~\cite{springel-moving-mesh10,hess-tessellation10}. Within the SPH
community, the use of dissipation, introduced via various means has
been the general recourse to mitigate the effects of the
error.\newline
Through an analogy with ``wall heating'' errors for finite
difference/volume codes, we argued that the pressure jump is a result
of entropy errors generated over an initial transient phase and
thereafter, passively advected with the particles. We highlighted that
a qualitatively similar error is present for the Lagrange plus remap
version of the Piecewise Parabolic Method (PPM) finite volume code
PPMLR. Through a comparison of PPMLR with it's direct Eulerian
counterpart, PPMDE, a lack of diffusion across the material wave was
identified as the origin for the error. By examining the
eigenstructure of the Lagrangian equations of motion, we showed that
the requisite diffusion needs to act on the density and energy
equations simultaneously. This explanation also justifies the myriad
techniques employed by different researchers to ``cure'' the
problem. Using our hypothesis, we construct a hybrid GSPH scheme that
introduces the requisite dissipation by using a more diffusive flux
for the energy equation. We verified our hypothesis by using the
blast-wave problem as the canonical test highlighting the pressure
anomaly in SPH. The results using the new scheme are shown to be
better than simply increasing the magnitude of thermal conduction for
SPH schemes that rely on explicit dissipation. A \emph{tuning}
parameter is introduced to limit the dissipation to the initial stages
of the computation.\newline
We expect the added dissipation to be disadvantageous for certain
problems. An example is the Sj\"ogreen's strong rarefaction
test~(1-2-3 problem \cite{toro-book}), for which, numerical
dissipation should be kept to a minimum. Indeed, one of the advantages
of the SPH artificial viscosity is the ability to switch it off
entirely when not required. Additionally, we believe that the hybrid
GSPH scheme can certainly be improved upon by using an alternative
hybridization to Eqs.~\ref{eq:gsph-hybrid-vstar-exp}. These equations
were constructed for the specific example to validate our hypothesis
concerning the origin of the spurious pressure jump in SPH. The
construction of an adaptive hybridization and validation for a general
suite of multi-dimensional problems is left as an area for future
investigation.
\section{}
\input{introduction}
\bibliographystyle{model6-num-names}
\newpage
|
\section{Introduction}
\label{sec-mainres-1}
The cubic Gross-Pitaevskii (GP) hierarchy is a system of infinitely many coupled linear PDE's
describing a Bose gas of infinitely many
particles, interacting via two-body delta interactions (repulsive in the defocusing case, and
attractive in the focusing case). It emerges in the derivation of the nonlinear Schr\"odinger equation (NLS) from
a bosonic $N$-particle Schr\"odinger system in the limit as $N\rightarrow\infty$,
where the pair interaction potentials
tend to a delta distribution.
In this paper, we prove the existence of scattering states for solutions to the defocusing cubic
GP hierarchy in ${\mathbb R}^3$. Moreover, we show that an exponential energy growth condition commonly
used in the well-posedness theory
of the GP hierarchy is, in a specific sense, necessary.
Our approach uses the {\em quantum de Finetti theorem} as presented in the work
of Lewin, Nam and Rougerie \cite{lnr} (see Section \ref{ssec-deFinetti-1}).
We previously applied it in \cite{chpa} to give a new, short proof of the unconditional uniqueness of solutions
to the cubic GP hierarchy in the energy space. The quantum de Finetti theorem allows
us to lift a variety of results that hold for the corresponding NLS to the GP hierarchy.
In particular, we use this approach in the work at hand
to establish the existence of scattering states for the cubic defocusing
GP hierarchy in ${\mathbb R}^3$.
Another main goal of this paper is to illuminate an important
exponential energy growth condition that is
invoked in all works on the well-posedness
of the GP hierarchy equations in the literature. We show that if this condition is
removed, the focusing GP hierarchy equations
become ill-posed. Again, the de Finetti theorem allows us to access this
previously elusive problem by relating it to the blowup in $H^1$ of solutions to the corresponding
focusing cubic NLS.
The first derivation of the nonlinear Hartree (NLH) equation from an interacting Bose gas
was given by Hepp in \cite{he}, via second quantization and coherent states.
Lanford, in his fundamental analysis of the $N\rightarrow\infty$ limit of $N$-particle
systems in classical mechanics, made central use of the BBGKY hierarchy \cite{Lan-1,Lan-2}.
The latter was subsequently employed by Spohn for a different derivation of the NLH,
in \cite{sp}. Fr\"ohlich, Tsai and Yau revisited this topic more recently in \cite{frtsya}.
Subsequently, Erd\"os, Schlein and Yau gave the derivation of
the NLS and NLH for a wide range of situations in their landmark works \cite{esy1,esy2,esy3,esy4}.
In their approach, proving the {\em uniqueness of solutions} to the GP hierarchy in a space of marginal density matrices
$L^\infty_{t\in[0,T)}{\mathfrak H}^1$ (defined in \eqref{frakH-def-1} below) is a crucial ingredient.
Their approach involves sophisticated singular integral estimates organized with
Feynman graph expansions, and introduces an important combinatorial
method that controls the large number of such graphs.
Subsequently, by combining a reformulation of the combinatorial method of \cite{esy1,esy2,esy3,esy4}
with methods from the theory of dispersive PDE's,
Klainerman and Machedon \cite{KM} gave a shorter proof of uniqueness of solutions
in a different solution space,
but under the assumption of an a priori condition on the solutions.
Their approach was used by various authors for the derivation of the NLS
from interacting Bose gases
\cite{chpa,CPBBGKY,xch3,CheHol-2013,kiscst,CT1,zxie}.
The analysis of the Cauchy problem for the GP hierarchy was initiated in \cite{chpa2} and continued e.g. in \cite{GreSohSta-2012,CT1}.
In \cite{CHPS-1}, we gave a new proof of unconditional uniqueness for solutions to the cubic GP hierarchy
in ${\mathbb R}^3$. Our result is equivalent to the uniqueness result in \cite{esy1,esy2,esy3,esy4}; the proof
combines the Erd\"os-Schlein-Yau combinatorial method \cite{esy1,esy2,esy3,esy4}
in boardgame formulation \cite{KM}, with an application of the {\em quantum de Finetti theorem} \cite{lnr},
see Section \ref{ssec-deFinetti-1}.
There exists a variety of different
approaches to the derivation of the NLS and NLH
from many-body quantum dynamics, due to the contributions of many authors; we refer to
\cite{esy1,esy2,esy3,esy4,ey,kiscst,rosc} and the references therein,
and also \cite{adgote,anasig,frgrsc,frknpi,frknsc,grma,grmama,he,pick,pick2}.
These dispersive nonlinear PDE's give a mean field description of the dynamics of Bose-Einstein condensates,
whose formation was first experimentally verified in 1995,
\cite{anenmawico,dameandrdukuke}. For the mathematical study of Bose-Einstein condensation,
we refer to \cite{ailisesoyn,lise,lisesoyn,liseyn, liseyn2} and the references therein.
\subsection{Definition of the GP hierarchy}
The cubic defocusing GP hierarchy on ${\mathbb R}^3$ for an infinite sequence of
bosonic
marginal density matrices $\Gamma=(\gamma^{(k)})_{k\in{\mathbb N}}$ is defined as the initial value problem
\begin{eqnarray} \label{eq-def-GP}
i\partial_t \gamma^{(k)} =\sum_{j=1}^k [-\Delta_{x_j},\gamma^{(k)}]
\, + \, \lambda B_{k+1} \gamma^{(k+1)}
\nonumber\\
\gamma^{(k)}(0)=\gamma_0^{(k)} \;,
\;\;\; k\in{\mathbb N} \;,
\end{eqnarray}
where $\lambda\in\{1,-1\}$, and
where $\gamma^{(k)}(t;{\underline{x}}_k;{\underline{x}}_k')$ is fully symmetric under permutations separately of the
components of ${\underline{x}}_k:=(x_1,\dots,x_k)$, and of the components of ${\underline{x}}_k':=(x_1',\dots,x_k')$.
We call \eqref{eq-def-GP} {\em defocusing} if $\lambda=1$, and {\em focusing} if $\lambda=-1$.
The interaction term for the $k$-particle marginal is defined by
\begin{eqnarray} \label{eq-def-b}
B_{k+1}\gamma^{(k+1)}
\, = \, B^+_{k+1}\gamma^{(k+1)}
- B^-_{k+1}\gamma^{(k+1)} \, ,
\end{eqnarray}
where
\begin{eqnarray}\label{eq-Bplus-GP-def-1}
B^+_{k+1}\gamma^{(k+1)}
= \sum_{j=1}^k B^+_{j;k+1 }\gamma^{(k+1)},
\end{eqnarray}
and
\begin{eqnarray}
B^-_{k+1}\gamma^{(k+1)}
= \sum_{j=1}^k B^-_{j;k+1 }\gamma^{(k+1)},
\end{eqnarray}
with
\begin{align}\label{eq-Bplusmin-def-1-1}
& \left(B^+_{j;k+1}\gamma^{(k+1)}\right)(t,x_1,\dots,x_k;x_1',\dots,x_k') \nonumber\\
& \quad \quad = \int dx_{k+1} dx_{k+1}' \nonumber\\
& \quad\quad\quad\quad
\delta(x_j-x_{k+1})\delta(x_j-x_{k+1}' )
\gamma^{(k+1)}(t,x_1,\dots,x_{k+1};x_1',\dots,x_{k+1}') \nonumber\\
& \quad \quad =
\gamma^{(k+1)}(t,x_1,\dots,x_j,\dots,x_k,x_j;x_1',\dots,x_k',x_j),
\end{align}
and
\begin{align}\label{eq-Bplusmin-def-1-2}
& \left(B^-_{j;k+1}\gamma^{(k+1)}\right)(t,x_1,\dots,x_k;x_1',\dots,x_k') \nonumber\\
& \quad \quad = \int dx_{k+1} dx_{k+1}' \nonumber\\
& \quad\quad\quad\quad
\delta(x'_j-x_{k+1})\delta(x'_j-x_{k+1}' )
\gamma^{(k+1)}(t,x_1,\dots,x_{k+1};x_1',\dots,x_{k+1}')\nonumber\\
& \quad \quad =
\gamma^{(k+1)}(t,x_1,\dots,x_k,x_j';x_1',\dots,x_j',\dots,x_k',x_j') \,.
\end{align}
We say that $B^+_{j;k+1}$ {\em contracts} the triple of variables $x_j,x_{k+1},x_{k+1}'$, and that
$B^-_{j;k+1}$ contracts the triple of variables $x_j',x_{k+1},x_{k+1}'$.
In \cite{esy1,esy2,esy3,esy4} and \cite{CHPS-1}, the
well-posedness of \eqref{eq-def-GP} is studied in the space of solutions
\begin{align}\label{frakH-def-1}
{\mathfrak H}^1:=\Big\{ \,(\gamma^{(k)})_{k\in{\mathbb N}} \, \Big| \,
{\rm Tr}(| S^{(k,1)} [\gamma^{(k)}] |) < R^{2k} \; \mbox{for some constant }R<\infty \, \Big\}
\end{align}
where $S^{(k,\alpha)}:=\prod_{j=1}^k(1-\Delta_{x_j})^{\alpha/2}(1-\Delta_{x_j'})^{\alpha/2}$ for $\alpha>0$.
We write
\begin{eqnarray}
U^{(k)}(t) := \prod_{\ell=1}^k e^{it(\Delta_{x_\ell}-\Delta_{x_\ell'})}
\end{eqnarray}
for the free $k$-particle propagator.
A {\em mild solution} to \eqref{eq-def-GP} in the space $L^\infty_{t\in[0,T]}{\mathfrak H}^1$
is a sequence of marginal density matrices $\Gamma=(\gamma^{(k)}(t))_{k\in{\mathbb N}}$
solving the integral equation
\begin{eqnarray}
\gamma^{(k)}(t) = U^{(k)}(t)\gamma^{(k)}(0) + i \int_0^t U^{(k)}(t-s) B_{k+1}\gamma^{(k+1)}(s) ds \,
\;\;\;,\;\;\; k\in{\mathbb N}\,,
\end{eqnarray}
satisfying
\begin{eqnarray}
\sup_{t\in[0,T]}{\rm Tr}(|S^{(k,1)}[\gamma^{(k)}(t)]|) < R^{2k}
\end{eqnarray}
for a finite constant $R$ independent of $k$.
\subsection{The cubic NLS}
In the special case of factorized initial data,
\begin{eqnarray}\label{eq-gammak-factorized-1}
\gamma^{(k)}_0({\underline{x}}_k;{\underline{x}}_k') \, = \, \prod_{j=1}^k \phi_0(x_j) \, \overline{\phi_0(x_j')} \,,
\end{eqnarray}
the condition that $(\gamma_0^{(k)})\in{\mathfrak H}^1$ implies
\begin{eqnarray}
{\rm Tr}(|S^{(k,1)}[\gamma_0^{(k)}] |) = \|\phi_0\|_{H^1}^{2k} < R^{2k}
\;\;\;,\;\;\;k\in{\mathbb N}\,,
\end{eqnarray}
and is equivalent to the condition $\|\phi_0\|_{H^1} < R$.
A particular solution to \eqref{eq-def-GP} with initial data \eqref{eq-gammak-factorized-1} is given by
$\Gamma=(\gamma^{(k)}(t))_{k\in{\mathbb N}}$ where for all $k\in{\mathbb N}$,
\begin{eqnarray}
\gamma^{(k)}(t;{\underline{x}}_k;{\underline{x}}_k') \, = \,
\prod_{j=1}^{k} \phi(t,x_j) \, \overline{ {\phi}(t,x'_j) }
\end{eqnarray}
is factorized.
In particular, the 1-particle wave function $\phi$ satisfies the
cubic NLS
\begin{eqnarray}\label{eq-NLS-def-1}
i\partial_t\phi(t)=-\Delta\phi(t)+ \lambda|\phi(t)|^2\phi(t) \;\;\;,\;\;\;\phi(0)=\phi_0\in H^1\,,
\end{eqnarray}
which is defocusing if $\lambda=1$ and focusing if $\lambda=-1$.
Solutions to \eqref{eq-NLS-def-1} conserve the $L^2$-mass
\begin{eqnarray}
M[\phi(t)] =\|\phi(t)\|_{L^2_x}^2 = M[\phi_0] \,,
\end{eqnarray}
the momentum
\begin{eqnarray}
P[\phi(t)] = i \int \overline{\phi(t,x)}\nabla\phi(t,x) dx \,,
\end{eqnarray}
angular momentum
\begin{eqnarray}
L[\phi(t)] = i\int \overline{\phi(t,x)}\,x\wedge\nabla\phi(t,x) dx \,,
\end{eqnarray}
and the energy
\begin{eqnarray}
E[\phi(t)] = \frac12\| \nabla_x \phi(t) \|_{L^2_x}^2 + \frac\lambda4\|\phi(t)\|_{L^4_x}^4 = E[\phi_0] \,.
\end{eqnarray}
The cubic NLS in ${\mathbb R}^3$ \eqref{eq-NLS-def-1} is $L^2$-supercritical and $H^1$-subcritical,
and is globally well-posed in $H^1$ if $\lambda=1$,
and locally well-posed if $\lambda=-1$, \cite{TaoBook}.
\subsubsection{The defocusing NLS}
In the defocusing case $\lambda=1$, \eqref{eq-NLS-def-1} is globally well-posed and
displays the existence of scattering states
and asymptotic completeness:
\begin{theorem}\label{thm-NLSscattering-1}
Let $S_t:\phi_0\mapsto \phi(t)$ denote the flow map associated to \eqref{eq-NLS-def-1},
for $t\in{\mathbb R}$ and $\lambda=1$.
Then, there exist continuous bijections (wave operators)
$W_+,W_-:H^1({\mathbb R}^3)\rightarrow H^1({\mathbb R}^3)$, such that
the strong limit
\begin{eqnarray}
\lim_{t\rightarrow\pm\infty}e^{-it\Delta}S_t(\phi_0) = \phi_\pm
\;\;\;\; , \;\;\;\; \phi_0=W_\pm(\phi_\pm)
\end{eqnarray}
holds for all $\phi_0\in H^1({\mathbb R}^3)$.
\end{theorem}
We refer to Section 3.6 in \cite{TaoBook} for a detailed discussion and a proof.
\subsection{The quantum de Finetti theorem}
\label{ssec-deFinetti-1}
As shown in our recent work \cite{CHPS-1},
solutions to the GP hierarchy and solutions to the NLS are closely interconnected via the
{\em quantum de Finetti theorem}, which
is a quantum analogue of the Hewitt-Savage theorem in probability theory, \cite{HewittSavage}.
We quote it in the formulation presented by Lewin, Nam and Rougerie in \cite{lnr} who coined the
notions of the strong and weak quantum de Finetti theorems (here collected into a single theorem).
\begin{theorem}\label{thm-strongDeFinetti-1}
Let ${\mathcal H}$ be a separable Hilbert space and let
${\mathcal H}^k = \bigotimes_{sym}^k{\mathcal H}$ denote the corresponding bosonic $k$-particle space.
Let $\Gamma$ denote a collection of
bosonic density matrices on ${\mathcal H}$, i.e.,
$$
\Gamma = (\gamma^{(1)},\gamma^{(2)},\dots)
$$
with $\gamma^{(k)}$ a non-negative trace class operator on ${\mathcal H}^k$.
Then, the following hold:
\begin{itemize}
\item
(Strong Quantum de Finetti theorem, \cite{HudsonMoody,Stormer-69,lnr})
Assume that $\Gamma$ is admissible, i.e., $\gamma^{(k)}={\rm Tr}_{k+1} \gamma^{(k+1)}$,
where ${\rm Tr}_{k+1}$ denotes the partial trace over the $(k+1)$-th factor, $\forall k\in{\mathbb N}$.
Then, there exists a unique Borel probability measure $\mu$,
supported on the unit sphere in ${\mathcal H}$, and invariant under multiplication of
$\phi \in {\mathcal H}$ by complex numbers of modulus one, such that
\begin{equation}\label{gkdf}
\gamma^{(k)} = \int d\mu(\phi) ( | \phi \rangle \langle \phi | )^{\otimes k}
\;\;\;,\;\;\;\forall k\in{\mathbb N}\,.
\end{equation}
\item
(Weak Quantum de Finetti theorem, \cite{lnr,AmmariNier-2008,AmmariNier-2011})
Assume that $\gamma_N^{(N)}$ is an arbitrary sequence of mixed states on ${\mathcal H}^N$, $N\in{\mathbb N}$,
satisfying $\gamma_N^{(N)}\geq 0$ and ${\rm Tr}_{{\mathcal H}^N}(\gamma_N^{(N)})=1$, and assume
that its $k$-particle marginals have weak-* limits
\begin{eqnarray}
\gamma^{(k)}_{N}:={\rm Tr}_{k+1,\cdots,N}(\gamma^{(N)}_N)
\; \rightharpoonup^* \; \gamma^{(k)} \;\;\;\; (N\rightarrow\infty)\,,
\end{eqnarray}
in the trace class on ${\mathcal H}^k$ for all $k\geq1$ (here, ${\rm Tr}_{k+1,\cdots,N}(\gamma^{(N)}_N)$
denotes the partial trace in the $(k+1)$-st up to $N$-th component).
Then, there exists a unique Borel probability measure $\mu$ on the unit ball in
${\mathcal H}$, and invariant under multiplication of
$\phi \in {\mathcal H}$ by complex numbers of modulus one,
such that
\eqref{gkdf} holds
for all $k\geq0$.
\end{itemize}
\end{theorem}
We note that the limiting hierarchies of marginal density matrices obtained via weak-* limits from
the BBGKY hierarchy of bosonic $N$-body
Schr\"odinger systems as in \cite{esy1,esy2,esy3,esy4} do not necessarily satisfy admissibility.
For the problems considered in this paper, the Hilbert space is given by ${\mathcal H}=L^2({\mathbb R}^3)$.
In \cite{CHPS-1}, we have used Theorem \ref{thm-strongDeFinetti-1} to present a new, shorter proof of the
{\em unconditional} uniqueness of solutions to the GP hierarchy
in $L^\infty_{t\in[0,T)}{\mathfrak H}^1$; we thereby also obtain a direct correspondence between solutions
to the GP hierarchy and solutions to the NLS which will be crucial for our proof of the main results in this paper.
The unconditional uniqueness part itself is equivalent to the uniqueness result proven in
\cite{esy1,esy2,esy3,esy4}.
Our main result in \cite{CHPS-1} states the following.
\begin{theorem}\label{thm-uniqueness-1} (Chen-Hainzl-Pavlovi\'c-Seiringer, \cite{CHPS-1})
Let $(\gamma^{(k)}(t))_{k\in{\mathbb N}}$ be a mild solution in $L^\infty_{t\in[0,T)}{\mathfrak H}^1$
to the (de)focusing cubic GP hierarchy in ${\mathbb R}^3$ with initial data $(\gamma^{(k)}(0))_{k\in{\mathbb N}}\in{\mathfrak H}^1$,
which is either admissible, or obtained
at each $t$ from a weak-* limit as described in Theorem \ref{thm-strongDeFinetti-1}.
Then, $(\gamma^{(k)})_{k\in{\mathbb N}}$ is the unique solution for the given initial data.
Moreover, assume that the initial data $(\gamma^{(k)}(0))_{k\in{\mathbb N}} \in{\mathfrak H}^1$ satisfy
\begin{eqnarray}
\gamma^{(k)}(0) = \int d\mu(\phi)(|\phi\rangle\langle\phi|)^{\otimes k}
\;\;\;,\;\;\;\forall k\in{\mathbb N}\,,
\end{eqnarray}
where $\mu$ is a Borel probability measure
supported either on the unit sphere or on the unit ball in $L^2({\mathbb R}^3)$,
and invariant under multiplication of
$\phi \in {\mathcal H}$ by complex numbers of modulus one.
Then,
\begin{eqnarray}\label{eq-GPdeF-NLS-sol-1}
\gamma^{(k)}(t) = \int d\mu(\phi)(|S_t(\phi)\rangle\langle S_t(\phi)|)^{\otimes k}
\;\;\;,\;\;\;\forall k\in{\mathbb N}\,,
\end{eqnarray}
where $S_t:\phi\mapsto \phi(t)$ is the flow map of the cubic (de)focusing NLS, for $t\in[0,T)$. That is, $\phi(t)$
satisfies \eqref{eq-NLS-def-1} with initial data $\phi$.
Accordingly,
\begin{eqnarray}\label{eq-GPdeF-pullb-sol-1}
\gamma^{(k)}(t) = \int d\mu_t(\phi)(|\phi\rangle\langle\phi|)^{\otimes k}
\;\;\;,\;\;\;\forall k\in{\mathbb N}\,,
\end{eqnarray}
where $d\mu_t(\phi):=d\mu(S_{-t}(\phi))$ is the push-forward measure under the NLS flow.
\end{theorem}
\section{Statement of main resuilts}
In this paper, we prove the existence of scattering
states for the defocusing cubic GP hierarchy in ${\mathbb R}^3$.
Moreover, we investigate the necessity of the energy growth condition in the definition of the
solution spaces ${\mathfrak H}^1$, see \eqref{frakH-def-1}.
\subsection{Scattering for the cubic GP hierarchy in ${\mathbb R}^3$}
We prove the existence of scattering states using the quantum de Finetti theorems,
Theorem \ref{thm-NLSscattering-1},
and Theorem \ref{thm-uniqueness-1}, which was proved in our earlier paper \cite{CHPS-1}.
The initial data for the GP hierarchy $\Gamma_0=(\gamma_0^{(k)})_{k\in{\mathbb N}}$ have the form
\begin{eqnarray}\label{eq-gamma0-def-1}
\gamma_0^{(k)} = \int d\mu(\phi)\big( |\phi\rangle\langle\phi|\big)^{\otimes k} \,.
\end{eqnarray}
We consider the defocusing cubic NLS with $\lambda=1$, and assume that
\begin{eqnarray}\label{eq-muEphi-cond-1}
\int d\mu(\phi) (E[\phi])^{2k} \leq R^k
\end{eqnarray}
holds for some finite constant $R>0$, and all $k\in{\mathbb N}$, where
\begin{eqnarray}
E[\phi] = \frac12\int |\nabla\phi|^2dx +\frac14 \int |\phi|^4 dx \,,
\end{eqnarray}
is the energy functional for the cubic defocusing NLS in ${\mathbb R}^3$.
The condition \eqref{eq-muEphi-cond-1} is equivalent to $\mu$ having support in a
ball in $H^1$; see Lemma \ref{lm-Chebyshev-1}, below.
We note that while the de Finetti theorems provide the existence and uniqueness of a
measure $\mu$, $\mu$ is in general not explicitly known.
Therefore, it is important to express the
condition \eqref{eq-muEphi-cond-1}, directly at the level of density matrices.
This is addressed in Section \ref{ssec-higheren-1} below, where we review {\em higher order
energy functionals} for GP hierarchies that were first introduced in \cite{CPHE}.
The first main result of this paper establishes the existence of scattering states for the
cubic defocusing GP hierarchy on ${\mathbb R}^3$, and provides the construction
of the corresponding asymptotic measures for the de Finetti representation \eqref{eq-GPdeF-pullb-sol-1}.
This has been a longstanding open problem
despite much activity in the field.
With our approach via the de Finetti theorem, it follows from the scattering theory for the NLS.
\begin{theorem}\label{thm-main-1}
Let $\Gamma_0=(\gamma_0^{(k)})_{k\in{\mathbb N}}$ be as in \eqref{eq-gamma0-def-1},
and $\lambda=1$ (the defocusing case).
We assume that $\mu$ satisfies
\eqref{eq-muEphi-cond-1}.
Let $\gamma^{(k)}(t)=\int d\mu(\phi)(|S_t\phi\rangle\langle S_t\phi|)^{\otimes k}$, for $k\in{\mathbb N}$,
denote the unique solution to \eqref{eq-def-GP}
satisfying $\gamma^{(k)}(0)=\gamma_0^{(k)}$, for $k\in{\mathbb N}$.
Then, there exist unique asymptotic measures $\mu_+$, $\mu_-$ such that
\begin{eqnarray}
\gamma_\pm^{(k)}:=\int d\mu_\pm(\phi) (|\phi\rangle\langle\phi|)^k
\end{eqnarray}
are scattering states $\gamma_+^{(k)}$, $\gamma_-^{(k)}$
on $L^2({\mathbb R}^{3k})$ satisfying
\begin{eqnarray}
\lim_{t\rightarrow\pm\infty}
{\rm Tr}\Big( \, \Big| \, S^{(k,1)} \Big[ \, U^{(k)}(-t)\gamma^{(k)}(t)
- \gamma_\pm^{(k)} \, \Big] \, \Big| \, \Big)
=0
\end{eqnarray}
for all $k\in{\mathbb N}$.
In particular,
\begin{eqnarray}
d\mu_\pm(\phi) = d\mu(W_\pm(\phi)) \,
\end{eqnarray}
where the continuous bijections $W_+$, $W_-:H^1\rightarrow H^1$ are the wave operators
from Theorem \ref{thm-NLSscattering-1}.
\end{theorem}
More generally, our method allows to transfer knowledge about the non-linear
Schr\"odinger equation (as given in Theorem \ref{thm-NLSscattering-1}) to
results about the GP hierarchy. For instance, if the existence of scattering states for the
focusing NLS can be shown for a suitable set of initial data (see for instance \cite{DuyHolRou}), one can also infer a
corresponding result for the GP hierarchy for initial states with de Finetti measure
$\mu$ supported on that set.
\subsubsection{Higher order energy functionals}
\label{ssec-higheren-1}
The condition on $\mu$ given in \eqref{eq-muEphi-cond-1}
can be formulated directly at the level of marginal density matrices.
This is of importance because the initial data for the GP hierarchy is
usually provided at the level of density matrices $\gamma_0^{(k)}$,
without explicit determination of the measure $\mu$.
To this end, we recall the higher order energy functionals that were introduced in \cite{CPHE}.
In the case of the cubic GP hierarchy, they are defined by
\begin{align}\label{eq-HE-def-1}
\langle K^{(m)}\rangle_{\Gamma(t)}:={\rm Tr}(K^{(m)}\gamma^{(2m )}(t))
\end{align}
for $m\in\mathbb{N}$, where
\begin{eqnarray}
K_\ell&:=&\frac{1}{2}(1-\Delta_{x_\ell})\text{Tr}_{\ell+1}+\frac{1}{4}B^+_{\ell;\ell+1}
\;\;\;,\;\;\;\ell\in{\mathbb N}\,,
\nonumber\\
K^{(m)}&:=&K_1K_{3}\cdots K_{2m- 1}.
\end{eqnarray}
In \cite{CPHE}, it is shown that these higher order energy functionals are conserved.
We note that
\begin{eqnarray}\label{eq-mu-condfoc-2-0}
{\rm Tr}(K_1K_3\cdots K_{2k-1} \gamma_0^{(2k)}) =
\int d\mu(\phi) (E[\phi])^k
\end{eqnarray}
corresponding to \eqref{eq-muEphi-cond-1}; see Section 4 of \cite{CPHE}.
\subsection{Energy growth condition}
Results on the well-posedness of the Cauchy problem for the GP hierarchy are usually
obtained in solution spaces of marginal density matrices where an exponential
growth condition either of the form
\begin{eqnarray}\label{eq-H1growth-1}
{\rm Tr}|S^{(k,1)}[\gamma^{(k)}]| < R^{2k} \;\;\;\forall k\in{\mathbb N}
\end{eqnarray}
holds in the trace norm, or of the form
\begin{eqnarray}\label{eq-H1growth-2}
\|S^{(k,1)}[\gamma^{(k)}] \|_{\rm HS} < R^{2k} \;\;\;\forall k\in{\mathbb N}
\end{eqnarray}
in the Hilbert-Schmidt norm.
In the works \cite{chpa,chpa2,CPHE,CPBBGKY,chpatz1,chpatz2}
and \cite{xch3,CheHol-2013,kiscst,CT1,zxie}, well-posedness is studied
in solution spaces incorporating the condition \eqref{eq-H1growth-2}.
In \cite{esy1,esy2} and the paper at hand, only the case \eqref{eq-H1growth-1} is considered;
a condition of this form is an important technical ingredient
for these uniqueness proofs.
We would like to address the crucial question whether the
energy growth condition \eqref{eq-H1growth-1} in the definition of the space ${\mathfrak H}^1$
is necessary for a well-posedness theory.
We introduce the quantity
\begin{eqnarray}\label{eq-logbound-1}
{\mathcal R}_{H^1}(\mu):= \exp\Big[ \, \limsup_{k\rightarrow\infty}\frac1{2k}\log\Big(\
\int d\mu(\phi) \|\phi\|_{H^1}^{2k} \, \Big) \, \Big] \,,
\end{eqnarray}
which corresponds to the radius of the smallest ball in $H^1$ that contains the support of $\mu$.
We observe that \eqref{eq-H1growth-1}, expressed
via the de Finetti theorem as
\begin{eqnarray}
\int d\mu(\phi)\|\phi\|_{H^1}^{2k}<R^{2k} \;\;,\;\;\;\forall k\in{\mathbb N}\,,
\end{eqnarray}
is equivalent to
the condition that $\mu$ satisfies
\begin{eqnarray}
{\mathcal R}_{H^1}(\mu) \, < \, R \,,
\end{eqnarray}
for $R<\infty$.
Hence, \eqref{eq-H1growth-1} simply means that $\mu$ has bounded support in $H^1$.
Here, we prove that if a faster than exponential growth rate is
admitted, so that ${\mathcal R}_{H^1}(\mu)=\infty$,
the {\em focusing} cubic GP hierarchy is {\em ill-posed}, in the sense that there exist
initial data at $t=0$ for which the solution blows up instantaneously;
that is, the norm
${\rm Tr}(|S^{(k,1)}\gamma^{(k)}(t)|)$ diverges for any positive $t>0$.
This result is a consequence of the following well-known
result about the blowup in $H^1$ of solutions of the cubic NLS in the focusing case $\lambda=-1$.
Eq. \eqref{eq-NLS-def-1} is locally well-posed; given any initial data $\phi_0\in H^1$,
there exists $\tau=\tau(\phi_0)>0$
and a unique solution $\phi(t)\in H^1$ for $t\in[0,\tau)$.
However, the solution might only exist for a finite time.
Let
\begin{eqnarray}
V[\phi](t) := \|x\phi(t)\|_{L^2}^2
\end{eqnarray}
denote the quadratic moment in $x$ with respect to $\phi(t)$.
Then, blowup in finite time occurs whenever $E[\phi_0]<0$ and $V[\phi_0]<\infty$.
This is proven by use of the {\em virial identities}
(Vlasov-Petrishchev-Talanov \cite{vlpetal}, Zakharov \cite{zakh}, Glassey \cite{glassey})
\begin{eqnarray}\label{eq-virial-1}
\partial_t V[\phi](t) = 2 \Im \int x\cdot \overline{\phi(x)}\nabla\phi(x) dx
\end{eqnarray}
and
\begin{eqnarray}\label{eq-virial-2}
\partial_t^2 V[\phi](t) = 16 E[\phi_0] - 2 \|\phi(t)\|_{L^4}^4 \,.
\end{eqnarray}
In fact, if $E[\phi_0]<0$, the r.h.s. of \eqref{eq-virial-2}
is strictly negative, and therefore, $\|x\phi(t)\|_{L^2}$ tends to
zero in finite time.
However, by the Heisenberg uncertainty principle,
\begin{eqnarray}
\|\phi_0\|_{L^2}^2
\leq C \|x\phi(t)\|_{L^2}\|\phi(t)\|_{H^1} \,,
\end{eqnarray}
see for instance \cite{TaoBook}.
Hence, a bound of the form $\|x\phi(t)\|_{L^2}< b(t)$ with $b(t) \searrow0$ as $t\nearrow T= T(\phi_0)$
implies that the solution blows up in $H^1$, that is,
$\|\phi(t)\|_{H^1}\nearrow\infty$ as $t\nearrow \tau=\tau(\phi_0)$ for some $\tau\leq T$.
We refer to $\tau(\phi_0)$ as the blowup time corresponding to the initial data $\phi_0\in H^1$.
One can easily derive an upper bound on the blowup time as follows. From
the virial identity \eqref{eq-virial-1}, it follows that
\begin{eqnarray}
|\partial_t \|x\phi(t)\|_{L^2}^2 | \leq 2 \|x\phi(t)\|_{L^2} \|\phi(t)\|_{\dot H^1} \,,
\end{eqnarray}
and from \eqref{eq-virial-2} that
\begin{eqnarray}
\partial_t^2 \|x\phi(t)\|_{L^2}^2 < 16 E[\phi(t)] = 16 E[\phi] \,,
\end{eqnarray}
where $\phi(t)$ solves the focusing cubic NLS with initial data $\phi$.
From second order Taylor expansion in $t$, we thus find that
\begin{eqnarray}\label{eq-blowup-quadeq-1}
\|x\phi(t)\|_{L^2}^2 &\leq& \|x\phi\|_{L^2}^2 +
2 t \|x\phi\|_{L^2} \|\phi\|_{\dot H^1} +8t^2 E[\phi] \,.
\end{eqnarray}
While the left hand side is non-negative, the right hand side becomes negative in
finite time if $E[\phi]<0$, which implies that the solution blows up in $H^1$.
If $E[\phi]<0$, it follows that the quadratic equation on the right hand side has
precisely one positive and one negative root.
The positive root $T(\phi)>0$ is an upper bound
on the blowup time $\tau(\phi)$.
Combining this with the de Finetti representation \eqref{eq-GPdeF-NLS-sol-1}
for solutions to the GP hierarchy, we obtain the following main result.
\begin{theorem}\label{thm-main-2}
Consider the set of probability measures $\mu$ on the unit ball in $L^2({\mathbb R}^3)$.
Then, the following dichotomy holds for the focusing cubic GP hierarchy \eqref{eq-def-GP}
(where we have $\lambda=-1$):
\begin{itemize}
\item
For the subset of probability measures satisfying
\begin{eqnarray}\label{eq-logbound-0}
{\mathcal R}_{H^1}(\mu) < \infty\,,
\end{eqnarray}
the following holds.
Given $\mu_0\in\{\mu\,|\,{\mathcal R}_{H^1}(\mu) <\infty\}$, there exists a unique solution to
the focusing cubic GP hierarchy in $L^\infty_{[0,T)}{\mathfrak H}^1$,
for some $T=T(\mu_0)>0$, with the initial data
\begin{eqnarray}\label{eq-gamma0-deF-mainthm2}
\Big(\,\gamma^{(k)}_0 = \int d\mu_0(\phi) (|\phi\rangle\langle\phi|)^{\otimes k}
\, \Big)_{k\in{\mathbb N}}
\end{eqnarray}
in ${\mathfrak H}^{1}$.
\item
For the subset of probability measures satisfying
\begin{eqnarray}\label{eq-logbound-1-0}
{\mathcal R}_{H^1}(\mu) = \infty\,,
\end{eqnarray}
the following holds.
For any $\delta>0$, there exist probability measures $\mu_0\in \{\mu\,|\,{\mathcal R}_{H^1}(\mu) =\infty\}$
with the following properties:
\begin{itemize}
\item
The right hand side of \eqref{eq-logbound-1} diverges at a rate at most $\exp(c k^\delta)$
as $k\rightarrow\infty$,
\begin{eqnarray}\label{eq-logbound-2}
\exp\Big[ \, \frac1{2k}\log\Big(\
\int d\mu_0(\phi) \|\phi\|_{H^1}^{2k} \, \Big) \, \Big] \, < \, C e^{c k^\delta} \,.
\end{eqnarray}
\item
The initial data defined by $\mu_0$ as in
\eqref{eq-gamma0-deF-mainthm2}
satisfies ${\rm Tr}(|S^{(k,1)}\gamma_0^{(k)}|)<\infty$ for all $k\in{\mathbb N}$,
but the associated solution to the cubic focusing GP hierarchy
displays instantaneous blowup (see below for the precise definition).
\end{itemize}
\end{itemize}
\end{theorem}
\subsection{Remarks}
We make the following remarks concerning the case \eqref{eq-logbound-1-0}:
\begin{itemize}
\item
The precise meaning of instantaneous blowup that we are considering is as follows.
Let $A_{R}:=\{\phi\in L^2 | \,\|\phi\|_{L^2}=1 \,,\, \|\phi\|_{H^1}\leq R \}$ for $R>0$, and
denote by ${\bf 1}_{A_R}$ the corresponding characteristic function. Then,
for every $R>0$, there exists $T=T(R)>0$ such that the sequence of regularized density matrices
\begin{eqnarray}
\Big( \, \gamma^{(k)}_R(t) := \int d\mu_0(\phi) {\bf 1}_{A_R}(\phi)
(|S_t(\phi)\rangle\langle S_t(\phi)|)^{\otimes k} \Big)_{k\in{\mathbb N}}
\end{eqnarray}
is a solution to the focusing cubic GP hierarchy in $L^\infty_{t\in[0,T(R))}{\mathfrak H}^{1}$.
However, in the limit $R\rightarrow\infty$,
\begin{eqnarray}\label{eq-limlog-infty-1}
\lim_{R\rightarrow\infty}{\rm Tr}(|S^{(k,1)}[\gamma_R^{(k)}(t)] |) = \infty \;\;\;\forall t>0 \,,
\end{eqnarray}
for any $k\in{\mathbb N}$.
It is in this sense that we say that ${\rm Tr}(|S^{(k,1)}[\gamma^{(k)}(t)] |)$ blows up instantaneously for $t>0$.
\\
\item
We note that for local well-posedness to hold, it is necessary that
$\mu_0$-almost surely, the blowup time,
$\tau(\phi)>\epsilon>0$, is bounded away from zero.
In our analysis of the case \eqref{eq-logbound-1-0},
we will construct measures $\mu_0$ for which $\tau(\phi)$ can be arbitrarily small
on the support of $\mu_0$. This is only possible when $\|\phi\|_{H^1}$ can be arbitrarily large
on the support of $\mu_0$.
\end{itemize}
\section{Proof of Theorem \ref{thm-main-1}}
In this section, we apply the quantum de Finetti theorem to prove
the existence of scattering states for solutions
to the defocusing cubic GP hierarchy in 3 dimensions.
To begin with, we observe that the condition \eqref{eq-muEphi-cond-1}
implies that $E[\phi]\leq R$ holds $\mu$-almost surely.
\begin{lemma}\label{lm-Chebyshev-1}
Assume that
\begin{eqnarray}\label{eq-EnCond-1}
\int d\mu(\phi) (E[\phi])^{2k} \leq R^{2k}
\end{eqnarray}
holds for some finite constant $R>0$, and all $k\in{\mathbb N}$. Then,
\begin{eqnarray}
\mu\Big( \Big\{ \, \phi\in L^2({\mathbb R}^3) \, \Big| \, \|\phi\|_{L^2}=1 \;,\; E[\phi] > R \,\Big\} \Big) \, = \, 0 \,.
\end{eqnarray}
\end{lemma}
\begin{proof}
From Chebyshev's inequality, we have that
\begin{eqnarray}
\lefteqn{
\mu\Big( \, \Big\{ \, \phi\in L^2({\mathbb R}^3) \, \Big| \, \|\phi\|_{L^2}=1 \;,\; E[\phi] > \lambda \,\Big\} \, \Big)
}
\nonumber\\
&\leq& \frac1{\lambda^{2k}}\int d\mu(\phi) (E[\phi])^{2k} \leq \frac{R^{2k}}{\lambda^{2k}} \,,
\end{eqnarray}
and for $\lambda> R$, the right hand side tends to zero when $k\rightarrow\infty$.
\end{proof}
Recalling that $\lambda=1$,
the representation \eqref{eq-gamma0-def-1} immediately yields
\begin{eqnarray}\label{eq-defoc-H1bd-1}
{\rm Tr}(|S^{(k,1)}[\gamma_0^{(k)}] |) &=& \int d\mu(\phi)\|\phi\|_{H^1}^{2k}
\nonumber\\
&\leq&
\int d\mu(\phi) (1+2E[\phi])^{k} \leq (1+2 R)^{k} \;\;\; \forall k\in {\mathbb N} \,.
\end{eqnarray}
This implies that $\mu$-almost surely, $\|\phi\|_{H^1}^2\leq 1+2E[\phi] \leq 1+2R$, by the
same argument as in Lemma \ref{lm-Chebyshev-1}.
Thus, Theorem \ref{thm-NLSscattering-1} implies that $\mu$-almost surely, there exists a unique
solution to the defocusing cubic NLS \eqref{eq-NLS-def-1} with initial data $\phi(0)=\phi$ which exhibits scattering and
asymptotic completeness. For notational convenience
further below, we denote $g_\pm(\phi):=\phi_\pm$, such that
\begin{eqnarray}
\lim_{t\rightarrow\pm\infty} \|e^{-it\Delta}S_t(\phi)-g_\pm(\phi)\|_{H^1} = 0 \,.
\end{eqnarray}
Then, $g_\pm(\phi)=W_\pm^{-1}(\phi)$.
Using the de Finetti representation of
the $k$-particle marginal
\begin{eqnarray}
\gamma^{(k)} = \int d\mu(\phi)\big( |\phi\rangle\langle \phi |\big)^{\otimes k} \,,
\end{eqnarray}
we let
\begin{eqnarray}\label{eq-gpm-def-1}
\gamma_\pm^{(k)}
&:=& \int d\mu(\phi)\big( |g_\pm(\phi)\rangle\langle g_\pm(\phi)|\big)^{\otimes k}
\nonumber\\
&=&\int d\mu_\pm(\phi)\big( |\phi \rangle\langle \phi |\big)^{\otimes k} \,,
\end{eqnarray}
where $d\mu_\pm(\phi)=d\mu(W_\pm(\phi))$.
It follows from energy conservation and positivity of the
potential energy term $\lambda\|\phi\|_{L^4}^4$ that $\mu$-almost surely,
\begin{eqnarray}\label{eq-phitg-bd-1}
\|S_t(\phi)\|_{H^1}^2 &\leq& 1+2E[\phi] \; \leq \; 1+2R
\nonumber\\
\|g_+(\phi)\|_{H^1}^2 &\leq& 1+2 E[\phi] \; \leq \; 1+2R \,.
\end{eqnarray}
For $\phi\in H^1$ satisfying $E[\phi]<R$, we have
\begin{eqnarray}\label{eq-scattdiff-bd-1}
\|e^{-it\Delta}S_t(\phi)-g_\pm(\phi)\|_{H^1}&\leq& \|e^{-it\Delta}S_t(\phi)\|_{H^1}
+\|g_\pm(\phi)\|_{H^1}
\nonumber\\
&\leq& 2(1+2E[\phi])^{1/2} < 2(1+2R)^{1/2}
\end{eqnarray}
uniformly in $\phi$, and uniformly in $t\in{\mathbb R}$.
Thus, we obtain that
\begin{eqnarray}\label{eq-domconv-1}
\lefteqn{
\lim_{t\rightarrow\pm\infty}
\int d\mu(\phi) \|e^{-it\Delta}S_t(\phi)
-g_\pm(\phi)\|_{H^1}
}
\nonumber\\
&=&
\int d\mu(\phi) \lim_{t\rightarrow\pm\infty}\|e^{-it\Delta}S_t(\phi)
-g_\pm(\phi)\|_{H^1}
\nonumber\\
&=&
0 \, ,
\end{eqnarray}
from the dominated convergence theorem.
We may now prove the existence of scattering states at the level of the GP hierarchy.
Using Theorem \ref{thm-main-1} and \eqref{eq-gpm-def-1}, we obtain that
\begin{eqnarray}
\lefteqn{
{\rm Tr}\Big(\Big|S^{(k,1)}\Big[U^{(k)}(-t)\gamma^{(k)}(t) - \gamma_+^{(k)}) \Big]\Big|\Big)
}
\nonumber\\
&=&\int d\mu(\phi) {\rm Tr}\Big(\Big|S^{(k,1)}
\Big[ \,
\big( \, |U(-t)S_t(\phi)\rangle\langle U(-t)S_t(\phi)| \, \big)^{\otimes k}
\nonumber\\
&&\hspace{5cm}
- \big( |g_+(\phi)\rangle\langle g_+(\phi)|\big)^{\otimes k} \Big) \Big]\Big|\Big)
\label{eq-auxdiff-1-1}
\end{eqnarray}
Using the identity
\begin{eqnarray}
A_0^{\otimes k}-A_1^{\otimes k}
&=&
\sum_{j=0}^{k-1}
A_{1}^{\otimes j}\otimes (A_0-A_1) \otimes A_{0}^{\otimes k-1-j}
\end{eqnarray}
with $A_0:=|U(-t)S_t(\phi)\rangle\langle U(-t)S_t(\phi)|$ and $A_1:=|g_+(\phi)\rangle\langle g_+(\phi)|$,
and
\begin{eqnarray}
{\rm Tr}(|S^{(1,1)}[A_0-A_1]|) \leq \|e^{-it\Delta}S_t(\phi) - g_+(\phi)\|_{H^1} \big( \, \|S_t(\phi)\|_{H^1} +
\|g_+(\phi)\|_{H^1} \, \big)
\end{eqnarray}
we find
\begin{eqnarray}
\eqref{eq-auxdiff-1-1}
&\leq&\sum_{j=0}^{k-1}
\int d\mu(\phi){\rm Tr}(|S^{(1,1)}[A_0-A_1]|)({\rm Tr}(|S^{(1,1)}[A_1]|))^j{\rm Tr}(|S^{(1,1)}[A_0]|)^{k-j-1}
\nonumber\\
&\leq&
\int d\mu(\phi)\|e^{-it\Delta}S_t(\phi) - g_+(\phi)\|_{H^1} ( \, \|S_t(\phi)\|_{H^1} +
\|g_+(\phi)\|_{H^1} \, )^{2k-1}
\nonumber\\
&\leq&\sum_{j=0}^{k-1}
\Big(\int d\mu(\phi)\|e^{-it\Delta}S_t(\phi) - g_+(\phi)\|_{H^1}^{2k}\Big)^{\frac1{2k}}
\nonumber\\
&&
\hspace{1cm}
\Big(\int d\mu(\phi)( \, \|S_t(\phi)\|_{H^1}+\|g_+(\phi) \|_{H^1} \, )^{2k}\Big)^{\frac{2k-1}{2k}} \,.
\label{eq-scattdiff-1}
\end{eqnarray}
It follows from \eqref{eq-defoc-H1bd-1} that $\mu$-almost
surely, $E[\phi(t)]=E[\phi]<R$.
Together with \eqref{eq-phitg-bd-1}, this implies
\begin{eqnarray}
\eqref{eq-scattdiff-1} \leq
2^k
\Big(\int d\mu(\phi)\|e^{-it\Delta}S_t(\phi) - g_+(\phi)\|_{H^1}^{2k}\Big)^{\frac 1{2k}}
(1+2R)^{\frac{2k-1}{2k}} \,.
\end{eqnarray}
The right hand side converges to zero as $t\rightarrow\infty$, as a consequence of
\eqref{eq-scattdiff-bd-1} and
\eqref{eq-domconv-1}.
This concludes the proof of Theorem \ref{thm-main-1}.
\qed
\section{Proof of Theorem \ref{thm-main-2}}
\subsection{The case ${\mathcal R}_{H^1}(\mu)<\infty$}
Given ${\mathcal R}_{H^1}(\mu)<R$ for some $R<\infty$,
it follows from Lemma \ref{lm-Chebyshev-1} that $\mu$-almost surely,
$\|\phi\|_{H^1}<R$.
The focusing cubic NLS, with flow map $\phi\mapsto S_t(\phi)$,
is locally well-posed in $H^1({\mathbb R}^3)$. In particular, there exist constants $T>0$ and $M<\infty$
such that
$\|S_t(\phi)\|_{H^1}<M$ for $t\in[0,T]$ where $T=T(\|\phi\|_{H^1})$
is monotonically decreasing, and where $M=M(\|\phi\|_{H^1})<\infty$ is monotonically
increasing in $\|\phi\|_{H^1}$ (more details are given in Section \ref{ssec-TMmonot-1} below).
Thus, by monotonicity of $T$ and $M$ with respect to $ \|\phi\|_{H^1}$,
it follows that $\mu$-almost surely,
$\|S_t(\phi)\|_{H^1}<M(R)<\infty$ for $t\in[0,T(R)]$.
Therefore, $\gamma^{(k)}(t)$ as given in \eqref{eq-GPdeF-NLS-sol-1}, with $k\in{\mathbb N}$,
satisfy
\begin{eqnarray}
\sup_{t\in[0,T(R)]}{\rm Tr}(|S^{(k,1)}[\gamma^{(k)}]|)<(M(R))^{2k} \;,\;\;\forall k\in{\mathbb N}\,,
\end{eqnarray}
and hence,
$(\gamma^{(k)})_{k\in{\mathbb N}}\in L^\infty_{t\in[0,T(R)]}{\mathfrak H}^1$. This proves the existence of a solution,
and its uniqueness follows from Theorem \ref{thm-uniqueness-1}.
\subsubsection{Monotonicity of the constants $T$ and $M$ with respect to $\|\phi\|_{H^1}$}
\label{ssec-TMmonot-1}
We remark that one can take $T(\|\phi\|_{H^1})\sim \|\phi\|_{H^1}^{-\beta}$ for some $\beta>0$ and
$M(\|\phi\|_{H^1})\sim \|\phi\|_{H^1}$.
For example, this can be easily obtained from applying the estimate
(3.42) in \cite{chpa2} to factorized solutions to the GP hierarchy
$\Gamma(t)=((|S_t(\phi)\rangle\langle S_t(\phi)|)^{\otimes k})_{k\in{\mathbb N}}$ with initial data of the focusing
cubic NLS satisfying
$\|\phi\|_{H^1}<R$, and for parameters $\xi_1=\frac{1}{2R}$ and $\xi_2=\frac{1}{4R}$.
In this case, we note that $\|\Gamma(t)\|_{{\mathcal H}_{\xi_2}^1}=\sum_{k\geq1}\xi_2^k\|S_t(\phi)\|_{H^1}^{2k}$,
etc,
in the notation of \cite{chpa2}.
\subsection{The case ${\mathcal R}_{H^1}(\mu)=\infty$}
We will explicitly construct a family of probability measures on the unit sphere
in $L^2({\mathbb R}^3)$ satisfying
\begin{eqnarray}\label{eq-limlog-infty-1}
{\mathcal R}_{H^1}(\mu)=\infty \,
\end{eqnarray}
with a prescribed maximum rate of divergence, together with
\begin{eqnarray}\label{eq-mumom-cond-1}
\int d\mu(\phi)\|x\phi\|_{L^2}^2<\infty \,,
\end{eqnarray}
and ${\rm Tr}(|S^{(k,1)}\gamma_0^{(k)}|)<\infty$ for all $k\in{\mathbb N}$,
such that instantaneous blowup occurs for the corresponding initial data.
In fact, we will be more specific, and construct measures $\mu$ such that the sequence
$(\gamma^{(k)} = \int d\mu(\phi)\big( |\phi\rangle\langle \phi |\big)^{\otimes k})_{k\in{\mathbb N}}$ belongs to the set
\begin{align}\label{eq-frakHeps-def-1}
{\mathfrak H}^{\alpha,r}:=\Big\{ \,(\gamma^{(k)})_{k\in{\mathbb N}} \, \Big| \,
{\rm Tr}(| S^{(k,\alpha)} [\gamma^{(k)}] |) < e^{ck^{r}} \; \mbox{for some constant }c<\infty
\,\Big\} \,
\end{align}
for $r\geq1$, where evidently,
${\mathfrak H}^{\alpha}={\mathfrak H}^{\alpha,1}$.
Instead of an exponential growth of order
$\int d\mu(\phi)\|\phi\|_{H^1}^{2k}\leq R^k = O( e^{ck})$, our aim is to
admit a growth of order $O(e^{ck^r})$ for some arbitrary $r>1$.
We note that any probability measure $\mu$ on $L^2({\mathbb R}^3)$ having the property that
\begin{eqnarray}\label{eq-mugamma-Hrcond-1}
\Big( \, \int d\mu(\phi) (|\phi\rangle\langle\phi|)^{\otimes k}
\, \Big)_{k\in{\mathbb N}} \,
\in {\mathfrak H}^{1,r}\setminus{\mathfrak H}^{1} \,
\end{eqnarray}
satisfies \eqref{eq-limlog-infty-1}. The parameter $r>1$ determines the rate of divergence
of \eqref{eq-logbound-1}.
To construct a measure $\mu$ satisfying \eqref{eq-mumom-cond-1}
and \eqref{eq-mugamma-Hrcond-1}, we may, for simplicity, pick $\mu$ to be supported on the unit sphere
\begin{eqnarray}
{\mathcal S}:=\{\psi\in L^2({\mathbb R}^3)| \, \|\psi\|_{L^2}=1\}\,.
\end{eqnarray}
We consider the dyadic decomposition of ${\mathcal S}=\cup_{j\in{\mathbb N}_0}{\mathcal N}_j$ based on the sets
\begin{eqnarray}
{\mathcal N}_j &:=& \Big\{\phi\in L^2({\mathbb R}^3)\,\Big|\, \|\phi\|_{L^2}=1 \,,\,
2^{j-1}<\|\phi\|_{\dot H^1}\leq 2^{j} \Big\}
\nonumber\\
{\mathcal N}_0 &:=& \Big\{\phi\in L^2({\mathbb R}^3)\,\Big|\, \|\phi\|_{L^2}=1 \,,\,
\|\phi\|_{\dot H^1}\leq 1\, \Big\} \,,
\end{eqnarray}
where
$$\|f\|_{\dot H^1}=(\int d\xi |\xi|^{2}|\widehat f(\xi)|^2)^{1/2}\,.$$
We define
$$d\mu_j(\phi):=d\mu(\phi){\bf 1}_{{\mathcal N}_j}(\phi)\,.$$
Our goal is to introduce subsets ${\mathcal M}_j\subset{\mathcal N}_j$, for $j\in{\mathbb N}_0$, such that for initial data $\phi^{(j)}\in{\mathcal M}_j$,
the blowup time $\tau(\phi^{(j)})$ for the cubic focusing NLS
tends to zero as $j\rightarrow\infty$.
For $\phi\in{\mathcal N}_j$, one observes that if $E[\phi]=\frac12\|\nabla\phi\|_{L^2}^2-\frac14\|\phi\|_{L^4}^4<0$,
then
\begin{eqnarray}
\|\phi\|_{L^4}\geq 2^{-\frac14} 2^{\frac j2} \,.
\end{eqnarray}
On the other hand, from the Gagliardo-Nirenberg inequality,
\begin{eqnarray}
\|\phi\|_{L^4}\leq C \|\phi\|_{\dot H^1}^{3/4}\|\phi\|_{L^2}^{1/4} \leq C 2^{\frac{3j}{4}} \,.
\end{eqnarray}
These are the only restrictions on the size of $\|\phi\|_{L^4}$ on $ {\mathcal N}_j $.
Moreover, from the uncertainty principle
\begin{eqnarray}
\|\phi\|_{L^2}^2\leq C \|x\phi\|_{L^2} \|\phi\|_{\dot H^1} \,,
\end{eqnarray}
it follows that for $\phi\in{\mathcal N}_j$,
\begin{eqnarray}
\|x\phi\|_{L^2} > C 2^{-j} \,.
\end{eqnarray}
Thus, we define subsets of ${\mathcal N}_j$ given by
\begin{align}
{\mathcal M}_j &:= \Big\{\phi\in L^2({\mathbb R}^3)\,\Big|\, \|\phi\|_{L^2}=1 \,, \|x\phi\|_{L^2} < b \,, \,
2^{j-1}<\|\phi\|_{\dot H^1}\leq 2^{j} \,,
\|\phi\|_{L^4} > C 2^{\frac{5j}{8}}\Big\}
\nonumber\\
{\mathcal M}_0 &:= \Big\{\phi\in L^2({\mathbb R}^3)\,\Big|\, \|\phi\|_{L^2}=1\,, \|x\phi\|_{L^2} < b \,, \,
\|\phi\|_{\dot H^1}\leq 1
\Big\} \,,
\end{align}
where $b>0$ is a fixed constant.
These sets are non-empty; an example of a function $f_j\in{\mathcal M}_j$ is given by
\begin{eqnarray}
f_j(x) = 2^{3j/2} g(2^{j}x) \,,
\end{eqnarray}
where $g(x)=e^{-x^2}$ is the standard Gaussian.
We define measures $\mu_j$ on $L^2({\mathbb R}^3)$ satisfying
\begin{eqnarray}\label{eq-mujMj-def-1}
\mu_j({\mathcal M}_j) = \kappa_r (j^{j^{1/\delta} })^{-j}
\end{eqnarray}
for $r>1$ and $\delta:=r-1$, where the constant $\kappa_r$ ensures that $\mu:=\sum \mu_j$ is a probability measure
on ${\mathcal S}$. For instance, we can think of $\mu_j$ as the uniform measure concentrated
on $\{e^{i\theta}f_j\}_{\theta\in[0,2\pi)}$, which is invariant under multiplication by a phase.
Then, we let
\begin{eqnarray}
\gamma^{(k)} := \int d\mu(\phi) (|\phi\rangle\langle\phi|)^{\otimes k}
\;\;\; \forall k\in{\mathbb N} \,,
\end{eqnarray}
and obtain that
\begin{eqnarray}
{\rm Tr}(|S^{(k,1)}\gamma^{(k)}|)
&=&{\rm Tr}\Big| \, S^{(k,1)}\int d\mu(\phi) (|\phi\rangle\langle\phi|)^{\otimes k} \, \Big|
\nonumber\\
&=&\sum_j\int d\mu_j(\phi) \|\phi\|_{H^1}^{2k}
\nonumber\\
&\leq&C \sum_j (j^{j^{1/\delta} })^{-j} 2^{2jk}
\nonumber\\
&\leq&C e^{c k^r} \,,
\label{eq-superexp-bd-1}
\end{eqnarray}
see Lemma \ref{lm-exponbd-1} below. Thus, $\gamma^{(k)}\in{\mathfrak H}^{1,r}$ for $r>1$.
On the other hand, $(\gamma^{(k)})_{k\in{\mathbb N}}\not\in{\mathfrak H}^{1,1}$.
This is because if $(\gamma^{(k)})_{k\in{\mathbb N}}\in{\mathfrak H}^{1,1}$,
it follows from Chebyshev's inequality (similar to Lemma \ref{lm-Chebyshev-1}) that
\begin{eqnarray}
\mu\Big(\Big\{\phi\in L^2({\mathbb R}^3)\Big| \|\phi\|_{H^1}>R\Big\}\Big) = 0 \,,
\end{eqnarray}
for some $R<\infty$. But this implies that
there are some constants $0<c<C<\infty$ independent of $R$, and $J>0$
such that $c\log R<J< C \log R$ for all $R>1$ sufficiently large,
and $\mu({\mathcal M}_j)=0$ for all $j>J$.
But then, $\mu_j({\mathcal M}_j)=0$ for all $j>J$, which contradicts \eqref{eq-mujMj-def-1}.
For $\phi\in {\mathcal M}_j$, we have that
\begin{eqnarray}
E[\phi] &=& \frac12 \|\nabla\phi\|_{L^2}^2 - \frac14 \|\phi\|_{L^4}^4
\nonumber\\
&<&\frac14(2^{2j}-2^{\frac52 j})
\nonumber\\
&<&- C 2^{\frac 52j}
\end{eqnarray}
for a constant $C>0$ independent of $j$. Therefore, by the blowup criterion
of Vlasov-Petrishchev-Talanov \cite{vlpetal}, Zakharov \cite{zakh}, and Glassey \cite{glassey},
the solution $\phi(t)$ with initial data $\phi(0)=\phi$ blows up in finite time in $H^1$.
Next, we derive an upper bound $T_j$ on the blowup time for solutions of the
focusing cubic NLS with initial data $\phi\in{\mathcal M}_j$.
From \eqref{eq-blowup-quadeq-1}, we obtain the quadratic inequality
\begin{eqnarray}\label{eq-quadrineq-1}
0&=& \|x\phi(0)\|_{L^2}^2 +
2 t \|x\phi\|_{L^2} \|\phi\|_{\dot H^1} +8t^2 E[\phi]
\nonumber\\
&\leq& b^2 + 2 t b - 8 t^2 C 2^{\frac52 j} \,.
\end{eqnarray}
The positive zero $T_j>0$ of the quadratic polynomial in $t$ on the lower line
provides an upper bound on the blowup time of the solution $\phi(t)$.
From \eqref{eq-quadrineq-1}, we get
\begin{eqnarray}
T_j < C 2^{-\frac{5j}2}
\end{eqnarray}
for a positive constant $C$ independent of $j$.
Hence, for any $\epsilon>0$, there exists $J=J(\epsilon)>c |\log \epsilon|>0$ such that
\begin{eqnarray}
{\rm Tr}\Big(\Big|S^{(k,1)}\Big[
\sum_{j=0}^J \int d\mu_j(\phi) (|S_t(\phi) \rangle
\langle S_t(\phi)|)^{\otimes k}\Big]
\Big|\Big)
\end{eqnarray}
blows up in a time interval $[0,2^{-cJ})\subset[0,\epsilon)$.
Letting $\epsilon\rightarrow0$ so that $J\rightarrow\infty$, we obtain that
${\rm Tr}(|S^{(k,1)}[\gamma^{(k)}(t)] |)$ blows up instantaneously.
This completes the proof of Theorem \ref{thm-main-2}.
\qed
Finally, we prove the last step in \eqref{eq-superexp-bd-1}.
\begin{lemma}\label{lm-exponbd-1}
Assume that $r>1$, and let $\delta:=r-1$. Then, for $k\in{\mathbb N}$ sufficently large (depending
only on $\delta$),
\begin{eqnarray}
\sum_j (j^{j^{1/\delta} })^{-j} 2^{2jk}
\leq e^{c k^r}
\end{eqnarray}
for a finite constant $c>0$.
\end{lemma}
\begin{proof}
Clearly,
\begin{eqnarray}
\sum_j (j^{j^{1/\delta} })^{-j} 2^{2jk} =
\sum_j \Big( \frac{2^{2k} }{j^{j^{1/\delta} }}\Big)^j \,.
\end{eqnarray}
Let $J=J(k) = k^\delta$. Then,
$$\frac{2^{2k} }{j^{j^{1/\delta} }} < \frac{2^{2k}}{k^{\delta k}}< \frac12$$
for all $j>J$,
if $k$ is large enough (depending only on $\delta$).
Therefore,
\begin{eqnarray}
\sum_{j>J} \Big( \frac{2^{2k} }{j^{j^{1/\delta} }}\Big)^j
< \sum_{j>J} \Big( \frac12\Big)^j < 1 \,,
\end{eqnarray}
for $k$ sufficiently large.
On the other hand,
\begin{eqnarray}
\sum_{0\leq j\leq J} \Big( \frac{2^{2k} }{j^{j^{1/\delta}}}\Big)^j
\leq \sum_{0\leq j\leq J} 2^{2kj}
\leq J 2^{2kJ} = k^\delta 2^{2k^{1+\delta}} \leq e^{c k^r} \,,
\end{eqnarray}
for a suitable constant $c>0$, as claimed.
\end{proof}
\subsection*{Acknowledgements}
We are grateful to the anonymous referees for their very useful comments.
The work of T.C. was supported by NSF grants DMS-1009448
and DMS-1151414 (CAREER).
The work of N.P. was supported by NSF grant DMS-1101192.
The work of R.S. was supported by NSERC.
|
\section{Introduction}
Single electron or hole spins confined in semiconductor quantum dots
(QDs) have been suggested\cite{SchliemannKhaetskiiLoos2003,
HansonSpinQdotsRMP2007} as prime candidates for the realization of
solid-state qubits. Single-shot readout of the
electron spin has been demonstrated in gate controlled\cite{Elzerman04} QDs, and a very
high degree coherent control of spins has been achieved in
self-assembled
QDs.\cite{Bonadeo1998,GreilichBayer2007,FokinaBayer2010,SpatzekGreilichBayer2011}
While the strong confinement of the electronic wave function in QDs
reduces the interaction with the substrate and, therefore,
suppresses electronic decoherence mechanisms, it simultaneously
enhances the hyperfine interaction between the confined electronic
spin and the nuclear spin bath formed by the underlying lattice.
Although spin-lattice relaxation processes might contribute to the
spin dephasing, it is
believed\cite{SchliemannKhaetskiiLoos2003,HansonSpinQdotsRMP2007,Merkulov2002,KhaetskiiLoss2003,CoishLoss2004}
that the hyperfine interaction dominates the spin relaxation in such
systems.
The dynamics of a single-electron spin coupled to a nuclear spin
bath of non-interacting
spins\cite{HansonSpinQdotsRMP2007,FischerLoss2008,Testelin2009} is
described by the Gaudin's central spin model\cite{Gaudin1976}
(CSM). Even though the CSM is exactly solvable\cite{Gaudin1976} using a Bethe ansatz
(BA), up to now there exist no thermodynamic Bethe
ansatz equations for this model as for the Kondo
model.\cite{AndreiFuruyaLowenstein83} The explicit solution of the BA
equations are restricted to a finite size system of $N<25$ bath
spins,\cite{BortzStolze2007,Bortz2010} while larger spin bath sizes
require stochastical
techniques\cite{FaribautSchuricht2013a,FaribautSchuricht2013b} to
extract the spin dynamics of the central spin and are still limited to
a small number of nuclear spins of $N<50$.
Recent spin-noise measurements performed on
quantum dots\cite{CrookerBayerSpinNoise2010,Dahbashi2012,LiBayer2012,*ZapasskiiGreilichBayer2013}
charged with single electrons or holes
aimed to directly reveal the intrinsic dynamics of central spins
interacting with a nuclear spin bath. The spin-noise
spectrum measured in $z$-direction shifts upon increasing the transversal magnetic field $B_x$ to
higher frequencies and traces the Larmor frequency $\omega_L$, while the
low-frequency range crosses over from a nearly Lorentz shape to a
$1/f$
noise.\cite{CrookerBayerSpinNoise2010,LiBayer2012,*ZapasskiiGreilichBayer2013}
In this paper, we investigate the spin-noise spectra for the
anisotropic CSM\cite{Gaudin1976,FischerLoss2008,Testelin2009} using
the Chebyshev expansion technique (CET). The CET has been developed
30 years ago\cite{TalEzer-Kosloff-84,*Kosloff-94,Fehske-RMP2006} and
offers an accurate way to calculate the time evolution of a single
initial state $|\psi_0 \rangle$ under the influence of a general
time-independent Hamiltonian ${\cal H}$ operating on a finite
dimensional Hilbert space. This approach has been
proposed\cite{Dobrovitski2003} as an efficient scheme for numerical
simulations of the spin-bath decoherence and applied to the isotropic
CSM\cite{ZangHarmon2006} as well as two coupled spins 1/2 in contact
with a spin bath.\cite{Dobrovitski2003,Yuan2008}
The original application\cite{Dobrovitski2003,ZangHarmon2006,Yuan2008}
of the CET was restricted to the propagation of a single wave
function. We have extended the approach (i) to thermodynamic
ensembles and also (ii) have averaged over many randomly generated
hyperfine coupling constant configurations. While (i) significantly
increases the accuracy of the CET for incoherent spin baths
in the high temperature limit relevant to the
experiments,\cite{CrookerBayerSpinNoise2010,Dahbashi2012,LiBayer2012,*ZapasskiiGreilichBayer2013}
(ii) turns out to be crucial for obtaining a smooth noise
spectrum. In any finite-size system, the exact spectral functions are
given by a finite number of $\delta$-functions. Since the
eigenenergies depend on the configuration of hyperfine couplings,
averaging over many configurations mimics a much larger system and
smoothens the superposition of $\delta$-functions to a continuous
function when introducing a very small, but finite broadening
similar to the $z$-averaging\cite{YoshidaWithakerOliveira1990}
used in the time-dependent numerical renormalization group
approach\cite{BullaCostiPruschke2008,AndersSchiller2005,*AndersSchiller2006}
to the non-equilibrium dynamics.
For a rigorous
solution\cite{Gaudin1976,BortzStolze2007,Bortz2010,FaribautSchuricht2013a,FaribautSchuricht2013b}
of the central spin dynamics the surrounding spin bath must be taken
into account exactly. Spin baths differ fundamentally from bosonic
baths\cite{Leggett1987} due to their degeneracies and their finite
dimensional Hilbert space. Over the last decade, very intuitive
pictures for the qualitative understanding of the decoherence
induced by a spin bath have emerged. The separation of time scales
\cite{Merkulov2002} -- a fast electronic precession around an
effective nuclear magnetic field, and slow nuclear spin precessions
around the fluctuating electronic spin -- has motivated various
quasi-static
approximations\cite{Merkulov2002,KhaetskiiLoss2003,ErlingssonNazarov2004,Cucchietti2005,Merkulov2010}
(QSA) and semiclassical
approximations\cite{Al-Hassanieh2006,ChenBalents2007,Sinitsyn2012} which describe
very well the short-time dynamics but predict a non-decaying
fraction of the central spin polarization. Early on, it became
clear\cite{Merkulov2002} that non-Markovian
corrections\cite{KhaetskiiLoss2003,CoishLoss2004} caused by slowly
fluctuating nuclear bath configurations generate corrections to this
non-decaying fraction as well as the long-time decay of spin
polarization. The functional form is non-universal and depends on the
details of the distribution function of the hyperfine coupling
constants.\cite{KhaetskiiLoss2003,CoishLoss2004} A crossover from
$\propto 1/\log^\alpha(t)$ in the absence of an external magnetic
field to a $\propto 1/t \log^2(t)$ in a finite field has been
predicted\cite{KhaetskiiLoss2003,CoishLoss2004} where the exponent
$\alpha$ is non-universal and depends on the distribution function.
Semiclassical
approaches\cite{Merkulov2002,KhaetskiiLoss2003,ErlingssonNazarov2004,Cucchietti2005}
have been employed to access the long-time dynamics of the
spin-dynamics using either a spin-coherent-state P
representation\cite{Al-Hassanieh2006,Sinitsyn2012} or a path integral
formulation,\cite{ChenBalents2007} both based on coherent spin
states. Spin fluctuations in quantum dot ensembles have been addressed
by a semi-classical Langevin term in the Bloch
equation.\cite{Glazov2012,StanekRaasUhrig2013}
All quasi-statical and semiclassical approximations or truncation
schemes in quantum-master equations allow to access the thermodynamic
limit by neglecting higher order correlations in the spin bath. While
such an approximation is very useful for tracing the spin-decay of
an initially polarized central spin coupled to an incoherent spin
bath, these approximations become questionable in coherent control
experiments\cite{GreilichBayer2007,FokinaBayer2010,SpatzekGreilichBayer2011}
where a pulsed pump laser induces frequency focusing of electron spin
coherence\cite{GreilichBayer2007} by a non-equilibrium nuclear spin
polarization.
Numerical methods, however, accurately include the
entanglement between the central spin and the spin bath but are
limited to finite spin bath size. Recently, the time-dependent density matrix
renormalization group
(TD-DMRG)\cite{White92,*SchollwoeckDMRG2005,*Schollwoeck2011} has been
adapted to the central spin
model\cite{FriedrichPhD2006,StanekRaasUhrig2013} and has been able to
push this limit to up to $N\approx 1000$ bath
spins\cite{StanekRaasUhrig2013} in calculation for the short time
dynamics.
While most of the theoretical literature has focused on the dynamics
of the central spin in a single QD, only recently a semi-classical
approach\cite{Glazov2012} has
been applied to the calculation of spin-noise in an ensemble of
QDs charged with single electrons or holes
as investigated in experiments.\cite{CrookerBayerSpinNoise2010,Dahbashi2012,LiBayer2012,ZapasskiiGreilichBayer2013}
We have extended the CET approach to ensembles of QDs and present spin-noise
spectra for this case as well.
We find an evolution of a Gaussian-type spin-noise spectrum of
a single QD to a more Lorentzian
shape spectrum at finite transversal magnetic field in a QD ensemble
similar to the observed experimental
data on electron spins \cite{CrookerBayerSpinNoise2010}
or hole spins.\cite{LiBayer2012}
Most of those approaches predict a non-decaying fraction of the central spin polarization
and a very slow, non-exponential decay of the spin in the long-time limit
which has not been observed in the experiments.
Therefore, it has been suggested that taking into account additional nuclear quadrupole
couplings\cite{Sinitsyn2012} can lead to an exponential decay when these quadrupole
coupling constants exceed the hyperfine coupling strength.
In this paper, however, we do not consider such an additional quadrupole
term. We restrict ourselves to the minimal anisotropic CSM
and focus on an exact finite size calculation
including all correlations between the electronic spin and the nuclear spins.
\subsection{Preliminaries}
We have calculated the spin-correlation function
$S(t)=[\expect{S^z(t) S^z} +\expect{S^z S^z(t) }]/2$ and its Fourier transformation, the spin-noise spectrum
$S(\omega)$, for the anisotropic CSM. Its isotropic limit is relevant
to QDs charged with a single electron, while the maximally anisotropic case, the
exactly solvable Ising limit, can be applied to dot-confined heavy-hole
spins.\cite{LeeEtAl2005,HansonSpinQdotsRMP2007,Testelin2009} The
generic anisotropic regime interpolates between these two extreme
cases and accounts for dot-confined hole spin of
arbitrary mixtures of light-hole and heavy-hole contributions.
We have studied $S(\omega)$ as a function of the
external transversal magnetic field $B_x$: while an external
longitudinal field $B_z$ suppresses the spin decay of a spin
initially polarized in longitudinal direction, a transverse field
$B_x$ induces a Larmor precession with the Larmor frequency
$\omega_L\propto |\vec{B}|$ of the electronic spin.
The following qualitative picture has emerged for the spin-noise
spectrum $S(\omega)$. In addition to a $\delta$-peak at zero frequency
whose spectral weight is given by the non-decaying fraction of the
spin polarization, we find a Gaussian type noise spectrum plus
corrections in the isotropic CSM. The width and center of the
spectrum are given by the intrinsic energy scale of the fluctuating
nuclear hyperfine field $1/T^*$. A finite transversal magnetic field
destroys the $\delta$-peak and the center of the spectrum is shifted
to larger frequencies which is given by the Larmor frequency in the
limit $\omega_L T^*\gg 1$.
In the Ising limit, the quasi-static approximation becomes exact. In
zero-field, $S_z$ cannot decay at all, and at any finite magnetic
field a finite non-decaying fraction of the spin polarization
remains after averaging over all randomly precessing configurations
in the long-time limit. In the thermodynamic limit, the finite
frequency part of the noise spectrum shows a threshold behavior, where
the threshold frequency $\omega_{\rm th}$ is given by $\omega_L$. Above the
threshold, we find $S(\omega) \propto (\omega-\omega_{\rm th})^{-\alpha}$ for
$\omega>0$ where the fits to our numerics are consistent with the
predictions\cite{KhaetskiiLoss2003,CoishLoss2004} of $\alpha=1/2$. Far
away from the threshold, the noise spectrum contains non-universal
parts and is cutoff sharply at the largest eigenenergy difference
$\omega_{\rm max}<\sqrt{\omega^2_L + A^2_s/4}$, where $A_s$ is determined by
the details of the electronic wave function of the confined
hole in the QD.
In the anisotropic CSM, the spin-noise spectrum depends strongly on the
anisotropy parameter $\lambda$ and the external magnetic field. We
find a crossover from a more threshold-like noise spectrum for
transversal fields that are small compared to $\lambda$ to a more
Gaussian type shape but with a renormalized width $1/T^*_\lambda$
which depends on the asymmetry parameter.
\subsection{Plan of the Paper}
As outlined above, the main objective of the paper is the discussion
of the electronic spin noise in the anisotropic CSM. Since the two
extreme limits, the isotropic CSM and the Ising limit, show two distinct
spectral properties, we divide the part on the results in three sections.
But first we begin with an introduction of the model
in Sec.\ \ref{sec:model} and
the discussion of a realistic distribution of hyperfine coupling
constants in Sec.\ \ref{sec distribution}. That distribution depends
not only on the envelope of the electronic wave-function but also on
the finite volume $V\propto r_0^3$ that encloses the QDs since with
increasing volume the number of nuclear spins which couple
exponentially weak is increasing. We briefly review the CET in Sec.\
\ref{Sec:Chebyshev} before we state the expansion of $S(\omega)$ in terms
of Chebychev polynomials in Sec.\ \ref{sec:CET-spin-noise}.
Sec.\ \ref{sec:results:isotropic-CSM} is devoted to the results for
the isotropic CSM while Sec.\ \ref{sec:results-ising-limit} focuses
on the Ising limit. In Sec.\ \ref{sec:results-anisotropic-CSM} we
present our data for the fully anisotropic case and investigate the
crossover from small to large transversal magnetic fields.
In order to establish the accuracy of the CET approach, we compare
the CET results with exact diagonalization (ED) for small bath sizes
in Sec.\ \ref{sec:benchmark}; we also augmented our data with the
prediction of QSA for the short time dynamics that has been reviewed
in Sec.\ \ref{sec mf}. Sec.\ \ref{sec:influence-distribution-function}
is devoted to an investigation of the influence of the distribution function
on the real-time dynamics while we extract the cutoff dependence of
the non-decaying fraction of spin polarization in Sec.\ \ref{sec:influence-of-r0}.
In Sec.\ \ref{sec:QD-ensemble-average} we present results for ensemble averaged
spin-noise spectra for parameters which closely resemble the recent
experiments.\cite{CrookerBayerSpinNoise2010,Dahbashi2012,LiBayer2012,*ZapasskiiGreilichBayer2013}
We explicitly demonstrate that a distribution of characteristic time scales of the quantum dots
modifies the spectral properties from a more Gaussian like shape to an ensemble averaged spectrum
which can be fitted with a Lorentzian. We will discuss the $g$-factor induced and
hyperfine interaction induced broadening of the single QD spectra. We summarize our findings in
Sec.\ \ref{sec:discussion-outlook} and give a brief outlook.
\section{Theory}
\label{sec:theory}
\subsection{Modelling of the quantum dots}
\label{sec:model}
For the spin decoherence in semiconductor QDs various interactions
play a role. As main contributions three sources have been
identified for relativistic electrons confined in a semiconductor QD:
the (i) Fermi contact hyperfine interaction, (ii) the dipole-dipole
interaction and (iii) the coupling of the orbital angular momentum
to the nuclear spin.\cite{FischerLoss2008} The Fermi contact
hyperfine interaction provides the largest energy scale of the three
contributions.\cite{FischerLoss2008}
Since the atomic contribution\cite{LeeEtAl2005,FischerLoss2008} to
the electron wave function stems mainly from 4s-orbitals in Ga and
As, the Fermi-contact hyperfine-interaction dominates. The wave
functions for light and heavy holes, however, are dominated by
4p-orbitals which vanish at the nuclei. Therefore, the sources (ii)
and (iii) govern the coupling for light and heavy holes to the
nuclear spins.
Fischer et al.\cite{FischerLoss2008} have shown that all cases can be
casted into an anisotropic CSM\cite{Testelin2009} given
by the Hamiltonian
\begin{eqnarray} H &= &\omega_L \vec{S} \vec{n}_B + \sum_{k=1}^{N} A_k
\left( S^z I_k^z + \frac{1}{\lambda} \left( S^x I_k^x + S^y I_k^y
\right) \right)
\label{eq hamiltonian}
\end{eqnarray}
where $ \vec{S}$ denotes the electron spin operator, $
\vec{I}_k$ the nuclear spin of the $k$-th nucleus, $N$ the number of
nuclear spins, and $\vec{n}_B =\vec{B}/|\vec{B}|$ is
the unit vector of the external magnetic field direction.
We include the electron or hole $g$-factor as well as the external
magnetic field strength $B= |\vec{B}|$ into the Larmor frequency
$\omega_L=g \mu_B B$. The anisotropy parameter $\lambda$ distinguishes
the three different cases: $\lambda = 1$ for electrons,
$\lambda= 1/2$ for light holes and $\lambda \rightarrow \infty$ for
heavy holes. For mixed heavy and light hole states $1<\lambda <\infty$
holds. We will review a realistic distribution of the $A_k$ and an
estimate of the orders of magnitude in the next section below. Since
$\lambda$ introduces an anisotropy axis, we use the term longitudinal
for external magnetic fields in $z$-direction and call $B_x$ a
transverse field.
For $\lambda = 1$, we recover the standard isotropic central-spin
model\cite{Gaudin1976} which conserves the total spin $\vec{J} =
\vec{S}+\sum_k \vec{I}_k$ of the coupled system in absence of an
external field, and the spin component of the total spin in the
direction of the applied field. For $\lambda \neq 1$ only the
component $J^z = S^z+\sum_k I^z_k$ of the total spin commutes with the
Hamiltonian for the absence of a transversal external
field. Throughout the paper we will use the convention $\hbar=1$,
$k_B=1$ unless otherwise stated.
Recently, the effect of additional nuclear quadrupole couplings
in the Hamiltonian on the central spin dynamics
have been investigated.\cite{Sinitsyn2012}
Such nuclear quadrupole terms significantly
change the bath characteristics. They lift the very large degeneracies of
the nuclear spin bath and lead to an exponential
spin decay\cite{Sinitsyn2012} once the nuclear quadrupole coupling strength exceeds
the hyperfine interaction. In this paper, however, we do not include these additional
nuclear quadrupole couplings and present an exact finite size calculation that avoids any
truncations or factorization of correlations which typically changes
the type of long-time dynamics.
{\em Definition of a time scale:} In addition to the Larmor frequency
$\omega_L$, the fluctuations of the transversal and longitudinal component
of an unpolarized nuclear spin bath in the absence of an external field
defines the time scale $T^*_\lambda$
\begin{eqnarray} [T_\lambda^*]^{-2} &=& \frac1{\lambda^2}
\sum_{k=1}^{N} A_k^2
\end{eqnarray} and $T^*=T^*_{\lambda=1}$ respectively. These scales
govern the short-time spin decay of the electronic spin polarized
along the $z$-axis. We use the transversal time scale to define the
dimensionless hyperfine couplings $a_k = A_k T_\lambda^*$ which enters
the dimensionless Hamiltonian $\tilde H = H T^*_\lambda$
\begin{eqnarray}
\label{eq hamiltonian-dim-less}
\tilde H &=& \lambda b \vec{S} \vec{n}_B
+
\sum_{k=1}^{N} a_k \left( S^z I_k^z + \frac{1}{\lambda} \left( S^x
I_k^x + S^y I_k^y \right) \right)
\,.
\end{eqnarray} The longitudinal scale $T^*$ has been used to define
the dimensionless magnetic field $b= \omega_L T^*$.
\subsection{Distributions of the coupling constants $A_k$}
\label{sec distribution}
In numerical simulations of the CSM either a
model\cite{BortzStolze2007,Bortz2010,StanekRaasUhrig2013,FaribautSchuricht2013a,FaribautSchuricht2013b}
distribution function $P(A)$ for the hyperfine coupling constants
$A_k$, or a more realistic\cite{SchliemannKhaetskiiLoos2003}
distribution based on the envelope function (\ref{eqn:psi-R}) have
been used.
In materials the coupling constants $A_k$ are given
by\cite{AbragamNMR1961,HansonSpinQdotsRMP2007,
LeeEtAl2005,FischerLoss2008}
\begin{align}
A_k &= \frac{16\mu_B \mu_N \gamma_k}{I_k} \left\lvert \Psi
(\vec{R}_k) \right\rvert^2 \eta_k
& = A_s v_0 \eta_k \left\lvert \Psi (\vec{R}_k) \right\rvert^2,
\label{eq akdef}
\end{align}
where $\mu_B$ is the Bohr magneton, and $\mu_N$ the nuclear magneton,
and $\gamma_k$ the gyro-magnetic factor of the $k$-th nuclei.
$I_k$ denotes the spin of
the $k$-th nucleus. $v_0$ is the average volume
occupied by a single nucleus within the crystal, and
$A_s = 16\mu_B \mu_k/v_0 I_k$.
The electron (or hole) wave function $\psi
(\vec{r}) = \Psi (\vec{r}) u(\vec{r})$ is divided into a slowly
varying envelope function $\Psi (\vec{r})$, that appears in (\ref{eq
akdef}), and a fast varying dimensionless Bloch factor $u (\vec{r})$
describing the wave function in the individual unit cells at the
nuclei $k$ and determining $\eta_k$. The factor $\eta_k$
encodes the symmetries of the Bloch factor and differs for electrons
and holes:\cite{Testelin2009}
\begin{align}
\eta^e_k &= \frac\pi3 \left\lvert u(\vec{R}_k)
\right\rvert^2\\ \eta^h_k &= \frac85 v_0 \left\langle
\frac{1}{\left\lvert \vec{r} - \vec{R}_k \right\rvert^3} \right\rangle
\label{eq etaholes} \, .
\end{align}
(For details on the definition of
Eq.\ (\ref{eq etaholes}) see appendix A of Ref.\
\onlinecite{Testelin2009}.)
In case of a simple Bloch wave of a free electron,
$|u(\vec{R}_k)|^2=1$, and different factors $\eta^e_k$ and $\eta^h_k$
are discussed in the
literature\cite{SchliemannKhaetskiiLoos2003,Testelin2009} to account
for the different values of the electronic wave function at Ga and As
nuclear sites in the unit cell.
In this paper, however, we neglect these differences and restrict our
investigation to a generic spin bath with $j=1/2$. We set $\eta^{e(h)}_k=1$
and absorb the value into the definition of $A_s$.
Since $\Psi (\vec{R}_k)$ varies slowly over the volume of a single
nucleus, $|\Psi (\vec{R}_k) |^2$ is taken as constant over the volume
$v_0$, and the normalization integral can be approximated by a
discrete sum over all nuclei
\begin{align}
1 &= \int \text{d}^3 R \left\lvert \Psi (\vec{R}_k)
\right\rvert^2 \approx \sum_k v_0 \left\lvert \Psi (\vec{R}_k)
\right\rvert^2
\end{align} from which we conclude:
\begin{eqnarray}
A_s &=& \sum_k A_k
\end{eqnarray}
and is constant independent of the details of the wave function
as a consequence of the wave function normalization.
$A_s$ is
typically\cite{LeeEtAl2005,HansonSpinQdotsRMP2007,Testelin2009} of the
order $O(10^{-5}\,\text{eV})$ for electrons and predicted\cite{Testelin2009} about a factor
$10-1000$ times smaller for holes yielding a much larger decoherence
time $T^*$ for hole-spins.
A typical distribution of $A_k$ for an (InGa)As self-assembled quantum
dot with base diameter of $15\,\text{nm}$ is depicted in Fig.\ 2 of
Ref.~\onlinecite{LeeEtAl2005}. This suggests a normalized envelope
function of
\begin{eqnarray}
\label{eqn:psi-R} \Psi(\vec{r})&=& (\sqrt{\pi} L_0)^{-3/2}
e^{-\frac{r^2}{2L_0^2}}
\end{eqnarray} varying on a length scale of $L_0\approx 5\,\text{nm}$.
Using\cite{LeeEtAl2005} the lattice constant of GaAs of $\approx
0.5\,\text{nm}$, $L_0=5\,\text{nm}$ and the ${\rm max}\{ A_k\}\approx 8\,\text{neV}$,
we obtain a realistic estimate\cite{LeeEtAl2005,HansonSpinQdotsRMP2007,Testelin2009}
for $A_s\approx 20\,\mu \text{eV}$.
\begin{figure}[t]
\begin{center}
\includegraphics[width=80mm]{fig-1}
\caption{Probability density $P(A)$ vs $A_s/A_{max}$
for three different radii $R/L_0=2,3,4$. The dashed lines
of the same color depict the histogram of the distribution generated by $10^{8}$ random picks
for $A_k$.}
\label{fig:probability-pa}
\end{center}
\end{figure}
Apparently, the probability to find a nucleus with coupling $A(r)$
increases quadratically with radius $r$: the distribution function
$P(A)$ diverges for $A\to 0$ and requires a finite cutoff radius
$R>L_0$ for which the envelope wave function has almost vanished and
$\Psi(\vec{r})$ is approximately normalized in that sphere of radius
$R$. Although the physics must be independent of this artificial
cutoff $R$, only those nuclei with a significant coupling constant
contribute to the dynamics of the central spin. Using the probability
density for point of fixed radius $r$, $P(r)= 3r^2/R^3$ and
$P(A)dA=P(r)dr$, we derive the probability distribution
\begin{eqnarray}
\label{eqn:p-a} P(A,r_0) &=& \frac{3}{2} \frac{1}{r_0^3 A}
\sqrt{\log(A_{max}/A)}
\end{eqnarray}
where $A_{max}=A(0) =A_s v_0/(\sqrt{\pi} L_0)^3$ and
the ratio $r_0$ is defined as $r_0=R/L_0$. This distribution is shown
in Fig. \ref{fig:probability-pa} for three different ratios $r_0 =
R/L_0=2,3,4$. We also added the histogram from $10^{8}$ randomly
generated $A_k$ values as dashed line in the same color. They are
nearly indistinguishable from the analytic function stated in
(\ref{eqn:p-a}). Note that this distribution function is almost identical to the one
used by Coish and Loss.\cite{CoishLoss2004}
Using this probability distribution $P(A)$ it is straight forward to
calculate the average $\bar A = \int dA P(A) A \approx A_s/N(R)$ where
$N(R)= n_0 4\pi R^3/3$ denotes the number of nuclei in the sphere of
radius $R$. The square average $\bar A^2$ is approximately given by
$\bar A^2 \approx \sqrt {2} A_s^2 [3\sqrt{\pi} N(R)N(L_0)]^{-1}$.
Both approximations become exact for $R\to\infty$.
Since the choice of the radius $R$ should be arbitrary, as long as
$R>L_0$ and $|\Psi(R)|^2\approx 0$, the physical properties of the
central spin model must not depend on $R$. This is clearly the case
for $A_s=\sum_k A_k = N(R) \bar A$ since we have $N(R)$ different
coupling constants. For the time scale $T^*$, the inverse rms of the Overhauser
field, which governs the short
time dynamics, we obtain
\begin{eqnarray}
[T^*]^{-2} &=& \sum_{k=1}^{N} A_k^2 = N(R) \bar A^2 =
A_s^2 \frac{1}{3\sqrt{\pi}N(L_0)}
\label{eqn-11}
\end{eqnarray}
which is also independent of the radius $R$. It is only
dependent on $A_s$ and the number of nuclei which are located within
the sphere defined by the length scale $L_0$ of the electronic
envelope function. Using the parameters from above an estimate for
$T^* \approx 30$ ns and a characteristic frequency
$f=1/T^*\approx 32$ Mhz for electrons and a factor
$10^{-1}-10^{-3}$ times smaller value for hole spins.
Although the cutoff $r_0$ controls the width of the distribution function
$P(A,r_0)$, the physics must remain invariant in the thermodynamic limit,
when sending $N\to\infty$ first and then $r_0\to\infty$.
In a finite size calculation, however, each random configuration
of hyperfine couplings $\{ A_k \}$ generated by $P(A,r_0)$
yields a slightly different dynamics.
To bridge between the typically $N(L_0)\approx 10^5$ nuclear spins in
real QDs and the numerical CET simulations of a spin bath with $N=20$,
each configuration $\{ A_k\}$ is normalized to a fixed $1/T^* = \sqrt{\sum_k A_k^2}$
which is the energy unit used in all calculations.
Therefore, each configuration is
characterized by exactly the same short time dynamics, and by
averaging over typically $n=50-100$ different configurations we mimic
a much larger spin-bath.
A word is in order about varying the cutoff $r_0$. For very large $r_0$, the ratio $a$
between the largest and the smallest hyperfine coupling is exponentially large and
the probability of generating exponentially low coupling constants is large. In this case, we will end
up with one or two large couplings $a_k=A_K T^*\to 1$, while all other are exponentially small for
a fixed $N$. The resulting unphysical dynamics will be discussed in Sec.\ \ref{sec:benchmark} below.
In order to avoid this effect one could demand that the largest dimensionless coupling
$a_{max} = {\rm max}\{a_k\}$ must be a constant when varying $r_0$. This requires
a simultaneous change of the bath size $N$ when varying $r_0$.
Choosing $a_{max}=0.5$ ensures a reasonable distribution of dimensionless coupling constants $\{a_k\}$
since the sum of all other couplings squares
must be $3/4$. For $r_0=1.5$ we can fulfill this condition with
$N=18$ while for $r_0=1$ only $N=8$ would be sufficient. For $r_0=2$, however, we would
need $N=42$ nuclear spins which is beyond the reach of the CET.
For $r_0=2$ and $N=18$ we find $a_{max}=0.64$: this is only slightly larger than $0.5$
and implies that all other hyperfine constants still contribute $\approx 60\%$ to
$T^*$. In our simulations, we typically us $r_0=1.5$ and $N=18$.
\subsection{Definition of the spin-noise function}
Experimentally the spin-noise is measured via fluctuations of the
Faraday rotation angle using a linearly polarized probe laser in
$z$-direction of the sample. The auto-correlation function of the
Faraday rotation angle is equivalent to the symmetrized fluctuation
function
\begin{eqnarray}
\label{eqn:def-spin-noise} S (t) &=& \frac{1}{2}\left[ \expect{ S^z(t)
S^z} +\expect{ S^z S^z(t)} \right] - \expect{S^z}^2
\end{eqnarray} where $\expect{S^z}$ denotes the average
spin-polarization which vanishes in the absence of an external
magnetic field. The probe laser only weakly perturbs the system, and
all expectation values are calculated using the equilibrium density
operator. Since $S(t)$ is symmetric in time, the spin-noise spectrum
\begin{eqnarray}
S (\omega)&=& \int_{-\infty}^\infty S (t) e^{-i\omega t} dt =
\int_{-\infty}^\infty S (t)\cos(\omega t) dt \, . \label{s_alpha}
\end{eqnarray}
From these definitions, we obtain the obvious sum-rule
\begin{eqnarray}
\int_{-\infty}^\infty \frac{d\omega}{2\pi} \, S (\omega) &=&
S(0) = \expect{(S^z)^2} -\expect{S^z}^2
\label{eq sum-rule}
\end{eqnarray}
for the spin-noise spectrum. In the absence of an
external magnetic field, its value is fixed to $1/4$ for a QD
filled with a single electron or hole spin. This sum-rule is useful
to test the accuracy of any numerical spin-noise calculation.
\subsection{Connection between the spin-noise function and the
real-time dynamics}
\label{sec:spin-noise-real-time-dynamics}
The spin-noise measurements are performed in thermal
equilibrium\cite{CrookerBayerSpinNoise2010} at $T \approx 5\,\text{K}
\rightarrow k_B T \approx 4\cdot 10^{-4}\,\text{eV}$. Since the
intrinsic energy scale $A_s$ of the system is of the order
$O(10^{-5}\,\text{eV})$ for electrons and even one order of magnitude
smaller for holes, $\beta A_s\ll 1$ holds, and we can consider the
coupled system consisting of the nuclear spin bath and the central
spin in the limit of high temperature and being characterized by the
initial density operator
\begin{align} \rho_0 &= \frac1D \underline{\underline{1}} \, ,
\end{align} where $D$ is the dimension of the Hilbert space and
$\underline{\underline{1}}$ is the identity matrix. Using the
commutator $\left[ \rho_0, S^z \right] = 0$, we conclude:
$\expect{S^z(t) S^z} = \expect{S^z S^z(t)}$.
If we prepare an initially fully polarized electron (hole) spin along
the $z$-direction coupled to an incoherent nuclear spin bath, the
density operator for such a system is given by
\begin{align}
\rho_p &= \rho_0 (\underline{\underline{1}} + 2 S^z)
\,\, .
\end{align}
The time evolution $\expect{S_z (t)}$ for this initial
condition
\begin{eqnarray}
\expect{S^z (t)}_{\rho_p} &= & \Tr{\rho_pS^z (t)} =
\left( \expect{S^z} + 2 \expect{S^z (t) S^z} \right) \nonumber \\ &=& 2
\expect{S^z (t) S^z} = 2 S(t)
\label{eq:sz-s-vs-t}
\end{eqnarray}
is equivalent to twice the correlation function $S (t)$
where the expectation values are calculated with respect to $\rho_0$.
Therefore, $S(t)$ can also be interpreted as the dynamics of an
initially fully polarized spin coupled to a bath at high
temperature. In this limit, we still can neglect the spin polarization
in Eq.\ (\ref{eqn:def-spin-noise}) in a magnetic field, since the
large field limit discussed below implies a large magnetic field in
comparison to the hyperfine energy scale but still small compared to
the temperature.
\subsection{Quasi-static approximation for $\vec{B} =0$}
\label{sec mf}
Merkulov\cite{Merkulov2002} et al.\ proposed a quasi-static approximation (QSA) to
calculate the short-time dynamics of an initially polarized electron
spin which later has been extended to dot-confined hole spins by Testelin et al.\cite{Testelin2009}
It is based on a separation of energy scales
and, therefore, time scales. While a single nuclear spin just is
exposed to the field generated by the single central spin whose
magnitude is proportional to $A_k\ll A_s$, the electron spin precesses
in a constant effective magnetic field $\vec{B}_{\text{eff}}$ provided
by a frozen nuclear spin bath configuration $\ket{\psi_{\rm bath}}$
\begin{eqnarray}
\vec{B}_{\text{eff}} &=& \frac{1}{\mu_B g_e} \sum_k
A_k \bra{\psi_{\rm bath}}\vec{I}_k\ket{\psi_{\rm bath}} =
B_{\text{eff}}\vec{n}
\label{glg beff}
\end{eqnarray}
for the time scale defined by the effective Larmor
frequency $\omega_{\text{eff}} = g_e \mu_B |\vec{B}_{\text{eff}}|$
which is of the order of $O(A_s)$.
In this momentarily frozen field in the direction
$\vec{n} = \vec{B}_{\text{eff}} /|\vec{B}_{\text{eff}} |$,
the Bloch equations for the electronic spin dynamics have the simple
solution
\begin{eqnarray}
\label{eq eom}
\expect{\vec{S}(t)} &=& \left( \vec{S}_0 \cdot \vec{n} \right)
\vec{n} + \left( \vec{S}_0 - \left( \vec{S}_0 \cdot \vec{n} \right)
\vec{n} \right) \text{cos}(\omega_{\text{eff}} t)
\nonumber
\\
&& + \left[
\vec{n} \times \left( \vec{S}_0 - \left( \vec{S}_0 \cdot \vec{n}
\right) \vec{n} \right) \right] \text{sin}(\omega_{\text{eff}} t),
\end{eqnarray}
with initial polarization of the electron spin
$\vec{S}_0$.
The effective magnetic field is generated by a large number of small
contributions from randomly oriented nuclear spins. Therefore, the
direction is isotropically distributed over a unit-sphere, and, in the
limit of large $N$, the magnitude of the effective field is described
by the Gaussian probability distribution
\begin{align}
W (\vec{B}_{\text{eff}}) &= \frac{1}{\pi^{3 / 2}
\Delta_B^3} \text{exp}\left( -
\frac{\vec{B}_{\text{eff}}^2}{\Delta_B^2} \right), \label{glg
feldverteilung} \\ \Delta_B^2 &= \frac{1}{2(\mu_B g_e)^2} \sum_k A_k^2
= \frac{1}{2(\mu_B g_e T^{*})^2}
\end{align}
whose width is defined by the fluctuation time scale
$T^*$.
Averaging the central spin dynamics (\ref{eq eom}) over the
distribution function $W (\vec{B}_{\text{eff}}) $
\begin{eqnarray}
\left\langle \vec{S} (t) \right\rangle &= & \int d\Omega \int_0^{\infty}
B^2 \text{d}B \, W(\vec{B}_{\text{eff}})\vec{S}(t)
\end{eqnarray}
the QSA result\cite{Merkulov2002} for $\vec{S}(t) =
S^z(t) \vec{e}_z$
\begin{eqnarray}
\left\langle S^z (t) \right \rangle &=& \frac{S^z_0}{3}
\left[ 1 + 2 \left(1 - \left(\frac{t}{2T^{*}} \right)^2 \right)
e^{-\frac{1}{2}\left(\frac{t}{2T^{*}} \right)^2} \right] \nonumber \\ &=& 2M(t)
\label{glg merkulov}
\end{eqnarray}
has been obtained. It is straight forward to calculate
its Fourier transformation $M(\omega)$
\begin{eqnarray}
\label{eq:m-w} M(\omega) &=& \int_{-\infty}^\infty \text{d}t\,
\text{e}^{-i\omega t} M(t) \\ &=& \frac{S_0}{3} \left[ 2\pi
\delta(\omega) + \omega^2 (\sqrt{8}T^*)^3 \sqrt{\pi}
\text{e}^{-2(\omega T^*)^2} \right] \, , \nonumber
\end{eqnarray}
where $S_0 = S^z_0 / 2$, because $S_0$ refers to the correlation
function $S(t)$ instead of $\left\langle S^z (0) \right\rangle$.
Since the QSA generates decoherence only by angular averaging, it
lacks long-time decay and contains a large non-decaying contribution
of $1/3$ of the initial spin-polarization $S_0$. This large
non-decaying contribution defines the weight of the spin-noise
$\delta$-function at $\omega=0$.
\subsection{The Chebyshev expansion technique}
\label{Sec:Chebyshev}
\subsubsection{Expansion of the time evolution operator}
Since all our results have been obtained using the
CET,\cite{TalEzer-Kosloff-84,Kosloff-94,Fehske-RMP2006} we briefly
review the CET, in order to introduce the notation used below.
The CET\cite{TalEzer-Kosloff-84,Kosloff-94,Fehske-RMP2006} has been
developed 30 years ago and offers an accurate way to calculate the
time evolution of a single initial state $|\psi_0 \rangle$ under the
influence of a general time-independent and finite-dimensional
Hamiltonian ${\cal H}$:
\begin{eqnarray} |\psi(t) \rangle &=& e^{-i{\cal H}t}|\psi_0\rangle .
\label{stationarySolution}
\end{eqnarray} The main idea of the method is to construct a stable
numerical approximation for the time-evolution operator $e^{-i{\cal
H}t}$ that is independent of the initial state $|\psi_0\rangle$ and
whose error can be reduced to machine precision for any given time
$t$. Its limitation lies in the need to explicitly store certain
states in the course of the calculation, which limits the size of the
Hilbert space that can be handled.
There are different ways to expand the time-evolution operator. The
most direct one is the conventional expansion of the exponent in
powers of ${\cal H}$ using the definition of any operator function.
One would like, however, to use an expansion that converges uniformly,
independent of the initial state $|\psi_0\rangle$. The Chebyshev
polynomials turned out to be such a suitable
choice\cite{TalEzer-Kosloff-84}. They are defined by the recursion
relation
\begin{equation} T_{n+1}(z) = 2zT_n(z)-T_{n-1}(z) ,
\end{equation} subject to the initial conditions $T_0(z) = 1$ and
$T_1(z) = z$. Those polynomials can be used to expand any function
$f(z)$ on the interval $-1 \leq z \leq 1$. Explicitly, $f(z)$ is
expressed as an infinite series
\begin{eqnarray}
f(z) &= &\sum_{n=0}^{\infty} b_n T_n(z) ,
\label{f-z-expantion}
\end{eqnarray}
where the expansion coefficients $b_n$ are given by
\begin{eqnarray}
b_n &= &\frac{2 - \delta_{n, 0}}{\pi} \int_{-1}^{1}
dx\, \frac{f(x)T_n(x)}{\sqrt{1-x^2}} \, .
\label{b_n_integral}
\end{eqnarray}
Using the integral
representation\cite{AbramowitzStegun} of the Bessel function
\begin{eqnarray}
J_n(z) &=& \frac{i^{-n}}{\pi} \int_0^\pi d\vartheta
e^{iz\cos\vartheta} \cos(n\vartheta)
\end{eqnarray}
and $T_n(\cos \vartheta)= \cos(n\vartheta)$, we
immediately arrive for $z\in[-1:1]$ at
\begin{eqnarray}
e^{-iz \tau } &=& \sum_{n = 0}^{\infty} b_n T_n(z)
\end{eqnarray}
with the expansion coefficients $b_{n} =
(2-\delta_{0,n}) i^n J_n (\tau)$.
If the spectrum of the Hamiltonian is bound to $E_{\rm min}\le E\le
E_{\rm max}$, the time-evolution operator $e^{-i {\cal H} t}$ can be
expanded in the same fashion after mapping the Hamiltonian to the
dimensionless $H' = (H-\alpha)/\Delta E$ where we have defined the
center of the energy spectrum $\alpha = (E_{\rm max}+E_{\rm
min})/2$ and its half-width $\Delta E = (E_{\rm max}-E_{\rm
min})/2$. Identifying $\tau=\Delta E t$ we arrive at
\begin{equation}
e^{-i H t} = \sum_{n = 0}^{\infty} b_n(t) T_n(H') \,
\label{Chebyshev-exp-e^-iH}
\end{equation} with
\begin{eqnarray}
\label{eqn:b_n} b_n(t)&=& (2-\delta_{0,n} ) i^n e^{-i\alpha t} J_n (
\Delta E t).
\end{eqnarray}
Finally, applying Eq.~(\ref{Chebyshev-exp-e^-iH}) to the initial state
$|\psi_0\rangle$ one obtains
\begin{equation}
|\psi(t) \rangle = \sum_{n = 0}^{\infty} b_n(t)
|\phi_n \rangle ,
\label{CET}
\end{equation} where the infinite set of states $|\phi_n \rangle =
T_n({\cal H}') |\psi_0\rangle$ obey the recursion
relation\cite{Fehske-RMP2006}
\begin{equation}
|\phi_{n+1} \rangle = 2{\cal H}' |\phi_n \rangle -
|\phi_{n-1} \rangle ,
\label{chebyshev-recursion-relation}
\end{equation}
subject to the initial condition $|\phi_0 \rangle =
|\psi_0 \rangle$ and $|\phi_1 \rangle = {\cal H}' |\psi_0 \rangle$.
Several comments are in order. First, all time dependence is confined
in Eq.~(\ref{CET}) to the expansion coefficients $b_n(t)$, which are
independent of the initial state $|\psi_0\rangle$. Second, the
Chebyshev recursion relation of
Eq.~(\ref{chebyshev-recursion-relation}) reveals the iterative nature
of the calculations. Starting from the initial state $|\psi_0\rangle$,
one constructs all subsequent states $|\phi_n\rangle$ using repeated
applications of the ``transformed'' Hamiltonian ${\cal H}'$. Third,
since $J_n(x) \sim (e x/2 n)^n$ for large order $n$, the Chebyshev
expansion converges quickly as $n$ exceeds $\Delta E t$. This allows
to terminate the series (\ref{CET}) after a finite number of elements
$N_C$ guaranteeing an exact result up to a well defined order.
Finally, the Chebyshev expansion has the virtue that numerical errors
are practically independent of $t$, allowing access to very long
times. The main limitation of the approach, as commented above, stems
from the size of the Hilbert space, since each of the states
$|\phi_n\rangle$ must be constructed explicitly.
For the application of the CET to the central spin model, an
estimation of the upper and lower bound of the Hamiltonian is required
entering the center of the energy spectrum $\alpha$ and its half-width
$\Delta E$. Applying the power iteration method, the series
\begin{eqnarray} \left\lvert \varphi_n \right\rangle = \frac{{\cal H}^{n} \left\lvert \varphi_0
\right\rangle}{\sqrt{\left\langle \varphi_0 \right\rvert {\cal H}^{n} {\cal H}^{n} \left\lvert \varphi_0
\right\rangle}}
\end{eqnarray} converges to the eigenvector associated with the
eigenvalue $\bar E$ the largest absolute value ${\rm max}\{
|E_{\text{min}}|, |E_{\text{max}}| \}$. For $\lambda=1$, and $A_k>0$,
one can show that the eigenvalue obtained by the power iteration
determines $E_{\text{min}}$ while $E_{\text{max}} = \omega_L/2 +
A_s /4$. For the Ising regime, $\lambda\to\infty$, the largest
eigenvalue $E_+$ and the smallest eigenvalue $E_-$ are exactly known
\begin{eqnarray} E_\pm &=& \pm\sqrt{\left( \frac{\omega_L}2 \right)^2 +
\left( \frac{A_s}{4} \right)^2}
\end{eqnarray} and for any finite $\lambda$, we interpolate between
these to limits. Alternatively, one can set $\alpha=0$ and only use
the eigenenergy $\bar E$ to define $\Delta E = 2\bar E$. In either
case, $\alpha$ and $\Delta E$ entering the Chebyshev expansion are
easily obtained.
\subsubsection{Evaluating traces}
\label{sec:traces}
The original application\cite{Dobrovitski2003} of the
CET\cite{TalEzer-Kosloff-84,*Kosloff-94} focused on the dynamics of a
single wave-function. We have extended the approach to thermodynamic
ensembles to incorporate the incoherent spin-bath at high-temperature
relevant to the experiments.
The expectation value of an arbitrary time-dependent observable $O$ is
given by
\begin{eqnarray}
\left \langle O(t) \right \rangle &= &\text{Tr}\left[
\rho_0 O (t) \right] = \sum_{i = 1}^{D} \left \langle i \right \rvert
\rho_0 \text{e}^{iHt} O \text{e}^{-iHt} \left \lvert i \right \rangle
\nonumber \\ &=& \sum_{i = 1}^{D} \bra{i'(t) }O \ket{i(t)}
\label{glg spurbildung}
\end{eqnarray}
where $\left \lvert i \right \rangle$ denotes a state
of the complete basis set. $D= 2^{N+1}$ grows exponentially with the
number of bath spins, and the direct evaluation of the trace cannot be
computed in moderate time for large $N$. In addition, the CET
provides only the time evolution of a single state $\ket{i(t)}=
\text{e}^{-iHt} \ket{i}$ and $\ket{i'(t)} = \text{e}^{-iHt}\rho_0
\ket{i}$.
Therefore, we employ a stochastical method discussed by Weisse et
al..\cite{Fehske-RMP2006} It is based on the generation of $N_s$
random states $\ket{r}$ of the form
\begin{eqnarray}
\left \lvert r \right \rangle &= &\sum_{i = 1}^{D}
\xi_{ri} \left \lvert i \right \rangle
\label{eq ranstate}
\end{eqnarray}
with the real coefficients $\xi_{si}$ fulfilling the
relations
\begin{eqnarray}
\left \langle \left \langle \xi_{ri} \right \rangle
\right \rangle &=& 0, \\
\left \langle \left \langle \xi_{ri}
\xi_{r'j} \right \rangle \right \rangle &= &\delta_{r,r'} \delta_{i,j}
\end{eqnarray}
where $\left \langle \left \langle \cdots \right
\rangle \right \rangle$ refers to the statistical average of these
random numbers. Note that $\ket{r}$ is not a normalized state for
$\xi_{si}$ fulfilling those relations. However, the trace of an
operator $\hat A$ can be evaluated\cite{Fehske-RMP2006} by
\begin{eqnarray}
\left \langle \left \langle \frac{1}{N_s} \sum_{r =1}^{N_s}
\bra{r} \hat A \ket{r} \right \rangle \right \rangle &=&
\frac{1}{N_s} \sum_{r = 1}^{N_s} \sum_{i,j = 1}^{D} \left \langle
\left \langle \xi_{ri} \xi_{rj} \right \rangle \right \rangle
\bra{i}\hat A\ket{j} \nonumber\\
&=& \sum_{i= 1}^{D} \bra{i}\hat
A\ket{i}
\end{eqnarray}
by statistical average of the random numbers.
Using the self-averaging properties of $\xi_{ri}$ drawn from a
Gaussian distribution, the trace is approximated\cite{Fehske-RMP2006}
by
\begin{eqnarray}
\frac{1}{N_s} \sum_{r = 1}^{N_s} \bra{r} \hat A
\ket{r} &=& \sum_{i= 1}^{D} \bra{i}\hat A\ket{i} +
O\left( \frac{1}{\sqrt{N_s D}} \right) \, .
\end{eqnarray}
The error is well controlled and scales with $(N_s
D)^{-1/2}$: only a few states $N_s$ are needed for an exponentially
large Hilbert space. In our simulations we typically use $N_s=5$
different randomly generated states for the evaluation of the
traces.
For very small Hilbert-spaces $N<10$, we have the reverse situation:
the number $N_s$ of random states required for a small error might
exceed the dimension of the Hilbert-space $D$. In such cases, the
trace has been evaluated exactly.
\subsection{Spin-noise spectra obtained from Chebyshev polynomial
expansion}
\label{sec:CET-spin-noise}
Since the time-dependent coefficients of the CET are known and are
stated in Eq.\ (\ref{b_n_integral}), we can analytically perform the
Fourier transformation of $S(\omega)$ in Eq.\ (\ref{s_alpha}) and derive
an explicit expression for the spin noise in terms of the momentum
$\mu_{n,m}$ and a convolution of two Chebyshev polynomials
\begin{eqnarray}
\label{eqn-CET-spin-noise-full}
S(\omega)
&=& \frac{2\pi}{\Delta E} \sum_{n,m = 0}^\infty \mu_{n,m}
\int_{-1}^{1-\frac{\omega}{\Delta E}} \text{d}\tilde\omega \,
\nonumber \\
&& \times \frac{T_{n}(\tilde\omega) T_m (\tilde\omega +
\frac{\omega}{\Delta E})} {\sqrt{(1 - \tilde\omega^2) (1 -
(\tilde\omega + \frac{\omega}{\Delta E})^2)}}
\label{eq calcspec}
\end{eqnarray}
for $\omega \geq 0$. While the convolution of two
Chebyshev polynomials only depends on the half-width $\Delta E$ of the
spectrum of $H$ and is independent of the dynamics, the momentum
$\mu_{n,m}$ gather all Hamiltonian dependent information about the
dynamics and are defined as
\begin{eqnarray}
\label{eqn:momenta-nm}
\mu_{n,m} &= &\frac{2 - \delta_{n,0}}{\pi}
\frac{2 - \delta_{m,0}}{\pi} \nonumber \\
&&\times \text{Tr}\left\{
\rho_0 T_n (H') S^z T_m (H') S^z \right\} \, .
\end{eqnarray}
and evaluated with the method presented in Sec.\
\ref{Sec:Chebyshev}.
The additional prefactors $g_n$
\begin{eqnarray}
g_n &=& \frac{(N_C - n + 2)\text{cos} \frac{\pi
n}{N_C + 2} +
\text{sin} \frac{\pi n}{N_C + 2}\text{cot}\frac{\pi}{N_C
+ 2}}{N_C + 2},
\end{eqnarray}
referring to as Jackson kernel,\cite{Fehske-RMP2006}
considerably reduce the truncation error\cite{Fehske-RMP2006} when
evaluating the truncated series
\begin{eqnarray}
S(\omega) &=& \frac{2\pi}{\Delta E}
\sum_{n,m = 0}^{N_C} g_n g_m \mu_{n,m}
I_{n,m}(\frac{\omega}{\Delta E})
\label{eqn-CET-spin-noise-NC}
\end{eqnarray}
instead of the true infinite series given by Eq.\
(\ref{eqn-CET-spin-noise-full}).
Since the function $I_{n,m}(x)$ defined as
\begin{eqnarray}
I_{n,m}(x) &=& \int_{-1}^{1- x} \text{d}\tilde\omega \,
\frac{T_{n}(\tilde\omega) T_m (\tilde\omega +x)} {\sqrt{(1 - \tilde\omega^2) (1 - (\tilde\omega +
x)^2)}}
\end{eqnarray}
is independent of the Hamiltonian,
it can be calculated and stored independently,
and later used in the summation (\ref{eqn-CET-spin-noise-NC}) of the momenta.
From the orthogonality relation of the Chebyshev polynomials we can
immediately conclude that only the momentum $\mu_{0,0}$ contributes to
the spin noise sum-rule (\ref{eq sum-rule}):
\begin{eqnarray}
S(t = 0) &= &\frac{1}{2\pi} \int_{-\infty}^\infty
\text{d}\omega\, S(\omega) \\
&=& \sum_{n,m = 0}^{N_C} \text{Tr}\left[
\rho_0 T_n (H') S^z T_m (H') S^z \right] \delta_{n,0} \delta_{m,0}
\nonumber \\
&= &\text{Tr}\left[ \rho_0 (S^z)^2 \right] = \frac{1}{4}.
\end{eqnarray}
Thus all spectral functions calculated from the CET
exactly fulfill the sum-rule independent of the number $N_C$ of included
Chebyshev polynomials.
\section{Dephasing of an electron spin: the isotropic central spin
model}
\label{sec:results:isotropic-CSM}
We begin with the discussion of the spin dynamics for a single
electron confined in a single quantum dot by investigating the isotropic
CSM with $\lambda = 1$. For all simulations of the real-time
dynamics, we used the CET for the evolution of the states in a system
of $N$ bath spins for a fixed configuration $\{ A_k\}$ drawn from the
probability distribution $P(A,r_0)$ stated in Eq.\
(\ref{eqn:p-a}). Since the number of bath sites is limited to $N\approx
20$, we average over typically $50$ configurations $\{ A_k\}$. By this
averaging we minimize finite size oscillations and essentially mimic
an effectively larger bath.
\subsection{Benchmarks}
\label{sec:benchmark}
In order to establish the virtue and the limitations of the CET
approach in combination with a statistical evaluation of the traces,
we have investigated the influence of (i) the number of bath spins
$N$, (ii) the order of the largest polynomial $N_C$, (iii) the number
of random states $N_s$.
\begin{figure} [tb] \centering
\includegraphics[width=75mm]{fig_2_a_b}
\includegraphics[width=75mm]{fig_2_c}
\caption{(color online) Benchmark for the CET calculating the time
evolution of the spin correlation function $S(t)$ for uniform coupling
constants $A_k = \frac{A_s}{N}$. Two different orders $N_C$ of the
expansion are compared to ED for $N = 2$ (a) and $N = 6$ (b) bath
spins. After the time scale indicated by the arrows the CET-error exceeds
$O(10^{-3})$. In (a)-(b) the occurring traces have been calculated
exactly and no error from the statistical evaluation of traces
enters. (c) treats this error for $N=10$, by comparing results for a
varying number $N_s$ of random states to a result taking the full
trace into account with $N_C = 66$. }
\label{fig benchmark}
\end{figure}
To benchmark the CET in small test systems accessible to exact
diagonalisation (ED), we restrict ourselves to uniform coupling
constants $A_k = A_s/N$ at first, defining $T_1^* = T^* =
\sqrt{N}/A_s$. We compare results obtained by ED with CET
calculations for $N=2,6$ and three different CET orders $N_C$ in Figs.\
\ref{fig benchmark} (a)-(b). For such small systems, we evaluate the
traces for the momenta $\mu_{n,m}$ in (\ref{eqn:momenta-nm}) exactly,
since the number $N_s$ of randomly generated states needed for an
accurate statistical evaluation of the traces exceeds the dimension of
the Hilbert space. Therefore, the only error of the CET data at longer
times in Figs.\ \ref{fig benchmark} (a)-(b) arises from the finite
$N_C$ while for the short-time dynamics up to $t<t_{\rm max}(N_C)$ an
essentially exact result is obtained.
The convergence of the CET for any given time $t$ is ensured by the
analytic properties of the Bessel functions of large order given by $J_n
(\Delta E t) \sim \left(e\Delta E t/2 n \right)^n$. The vertical
arrows in Figs.\ \ref{fig benchmark} (a)-(b) indicate where this estimate
exceeds the value $10^{-3}$ for $N_C = 10$ and $20$. Apparently the
CET reproduces the exact ED results accurately up to this time. Thus,
the order of the CET for all further calculations is determined by the
smallest $N_C$ fulfilling the condition $\left( e\Delta E t_{\rm max}
/2 N_C \right)^{N_C} \leq 10^{-3}$, where $t_{\rm max}$ is the largest
time of interest. For $N=6$ and $t_{\text{max}}(N_C)/T^*=50$ this estimate yields
$N_C = 56$. Consequently, the CET renders the ED result exactly up to
$t/T^*=50$. To illustrate the deviations at larger time for an
insufficiently large $N_C$, $S(t)$ is plotted for the additional two
values $N_C=10,20<56$ in Fig.\ \ref{fig benchmark} (b).
Fig.\ \ref{fig benchmark}(c) illustrates the effect of the error
arising from the statistical evaluation of traces for $N=10$ bath
spins. The CET-result where the traces have been exactly calculated
(blue line) is serving as reference. Since the error of the
statistical evaluation is of the order $O ( (N_s D)^{-1/2})$, each
ascending value for $N_s$ shown in Fig.\ \ref{fig benchmark} (c)
reduces the remaining error by a factor $2$ independent of the time $t$.
This decrease of the statistical error is clearly visible. For
$N=10$, where the Hilbert space has the dimension of $D=2^{11}=2048$,
our results with $N_s=16$ random states already converged enough to be
optically almost undistinguishable from the exact calculations.
By choosing either a large number of random states $N_s$ or a large
number of nuclei $N$, we are able to obtain an accurate representation
of the exact evaluation of the traces.
The physics of the CSM with uniform coupling constants $A_k = A/N$ is
well understood.\cite{KhaetskiiLoss2003} For a system with only two
bath spins we observe a coherent oscillation as depicted in Fig.\
\ref{fig benchmark}(a) since the central spin effectively only
interacts with the triplet state formed by the two bath spins while
the singlet is decoupled. The oscillation frequency is given by the
full width $2\Delta E = E_{\text{max}} - E_{\text{min}}$ of the
Hamiltonian's spectrum. For larger systems the dynamics is still
coherent and of the form as exemplarily shown for $N=6$ and $N=10$ in
Figs.\ \ref{fig benchmark} (b)-(c). The short time dynamics for $N >
3$ is governed by $T^*$ and the recurrence time $T_{\rm rec}$, where
$S(T_{\rm rec}+t)=S(t)$, increases linearly with the bath size since
the differences of the eigenenergies are commensurable. We find
$T_{\rm rec}\approx 30T^*$ for $N=6$ and $T_{\rm rec}\approx 40T^*$
for $N=10$.
\subsection{Results in the absence of an external magnetic field}
\subsubsection{Influence of the distribution function on the real-time
dynamics}
\label{sec:influence-distribution-function}
\begin{figure} [tb] \centering
\includegraphics[width=80mm]{fig_3_a.eps}
\includegraphics[width=80mm]{fig_3_b.eps}
\caption{(color online) The spin correlation function $S(t)$ for
randomly generated coupling constants $A_k$ calculated via CET. Each
shown curve has been averaged over $n=50$ different random
realizations of couplings. (a) Short-time evolution of $S(t)$ for
increasing $r_0$ with $N = 18$. (b) The analytical result $M(t)$ in
comparison to calculated data for increasing bath size $N$ based on a
fixed ratio $r_0 =R/L_0 = 1.5$. The inset shows the area marked
by the box. }
\label{fig b0szt}
\end{figure}
Now we discuss the influence of randomly generated coupling constants
$A_k$ onto the time evolution of the spin-correlation function $S(t)$.
In numerically accurate simulations of the real-time dynamics using a
statistical evaluation of the Bethe-ansatz
equations,\cite{FaribautSchuricht2013a,FaribautSchuricht2013b} the
maximum number of bath spins $N\approx 45 $ still remains several
orders of magnitude smaller than the $10^{5}$ nuclear spins present in
experimental samples. The dynamics of small systems is influenced by
the range of coupling constants defining the ratio
$a=A_{\text{max}}/A_{\text{min}}$ between the largest and the smallest
coupling constant. Increasing $a$ at constant $N$ increases the number
of nuclear spins which are only very weakly coupled to the electronic
spin. In addition the deviation of the average square, $\expect{A}^2$
and the fluctuation $\expect{A^2}$ increase. The short and intermediate
dynamics is dominated by a decreasing number of bath spins for a fixed $N$,
while thevery weakly coupled spins are only contributing significantly at
extremely long times.
In a statistical evaluation of the exact Bethe ansatz
equations\cite{FaribautSchuricht2013a,FaribautSchuricht2013b} the
fixed set of $A_k= A/N \exp[-(k-1)/(N-1)]$ has been chosen, leading to
the ratio $a={\rm e}$. A recent TD-DMRG\cite{StanekRaasUhrig2013}
study has pushed the limit to up to $N=100-1000$ nuclear bath spins.
In this study, the configurations of $\{A_k\}$ have been drawn from
$P(A)=const$ on the interval $A_0[1/2,1]$, corresponding to a ratio
$a=2$. For the first distribution, the fluctuation
$u=\expect{A}^2/\expect{A^2}\approx (1+1/e)/(1-1/e)/2\approx 1.082$,
while the distribution $P(A) =const.$ yields $u=28/27$. In both cases
$u\approx 1$ holds which does not differ significantly from
$A_k=const$. Therefore, the non-decaying fraction of the spin
polarization remains close to the QSA result.
In the distribution function $P(A,r_0)$, defined in Eq.\
(\ref{eqn:p-a}), the cutoff ratio $r_0$ directly translates into the
ratio $a = \exp(r_0^2)$. For large $r_0$ the probability $P(A,r_0)$ is
high for adding more and more nuclei to the system whose interaction
with the central spin is negligible, e.~g.\ for $r_0 = 4$ the ratio
between the largest and smallest coupling constant $A_k$ has already
reached $A_{\text{max}}/A_{\text{min}} \approx 10^{7}$. In order to
obtain results faithfully representing a larger system, a set of
several $A_k$ must be taken into account for each order of magnitude
which is impossible for a system size of only $N= 20$ bath spins.
Fig.\ \ref{fig b0szt} (a) illustrates the influence of the cutoff
$r_0$ onto the real-time dynamics. The results are calculated for $N
= 18$ bath spins and averaged over $n = 50$ different random
configurations $\{ A_k\}$ to reduce the influence of fluctuations and
effectively take more nuclei into account. We added the QSA result
$M(t)$ stated in Eq.\ (\ref{glg merkulov}) as a guide for the
short-time dynamics obtained from a random nuclear field approximation
in the thermodynamic limit $N\to\infty$. All CET curves perfectly coincide
with $M(t)$ for very short time scales $t/T^*<3$.
For $r_0=1$, the ratio $a= e$, and
only significant coupling constants of the same order of magnitude are
taken into account. The polarization saturates approximately at
the value $S(0)/3$ as predicted by the QSA. However, slight deviations
between the CET curve and $M(t)$ are observed for times $t/T^*>3$.
Nevertheless the CET short-time dynamics and results obtained by other
approaches\cite{FaribautSchuricht2013a,FaribautSchuricht2013b,StanekRaasUhrig2013}
agree remarkably well with each other and with the QSA.
This indicates that the generic dynamics can already be obtained by a rather
small numbers of bath spins.
For increasing cutoffs $r_0>1$, the short-time dynamics of $S (t)$
evolves from passing through a single minimum as described by the QSA
curve to a damped oscillation as depicted in Fig.\ \ref{fig b0szt}(a).
This behavior is easily understood by the distribution of coupling
constants in any of the random configurations $\{ A_k\}$. For a large
cutoff $r_0$ and a fixed number of bath spins $N$, the number of nuclei
which couple to the central spin with a coupling constant
$A_k/A_s=O(1)$ becomes very small in $\{ A_k\}$ due to the increasing
probability to find a small coupling. Essentially we see a similar
coherent motion as in Fig.\ \ref{fig benchmark}(a) involving only one
or two bath spins with the strongest coupling constants, while the
slow dephasing is induced by the remaining very weakly coupled nuclear
spins. Therefore, we conclude that such choices of the cutoff $r_0$ do
not render the dynamics for $N\to\infty$ when working with a fixed and
small $N$.
Fig.\ \ref{fig b0szt}(b) focuses on the bath-size dependency of the
short-time evolution of $S(t)$ for $N=14,18,22$ and fixed $r_0 =
1.5$. This cutoff would correspond to $N(R) \approx 10^5$
in a real system, implying a ratio $a \approx 8$.
The figure compares exact simulations for three
different bath sizes to the QSA result $M(t)$ and demonstrate the
fast convergent with the bath size for $r_0=1.5$.
The initial decay of
$S(t)$ is well described by $M(t)$ and for increasing $N$ the exact
finite size curves approach the QSA solution for short-time
scales. But after the initial decay the central spin's polarisation
drops below the value predicted by the QSA approach. This deviation
arises from the contribution of the small coupling constants, whose
interaction with the central spin is too weak to have major influence
on the short time behavior of $S(t)$, but on large time scales the
small couplings become dominant. Since the QSA result is based on the
assumption of a static bath, it is not surprising that it is only able
to describe the short-time and intermediate-time evolution of the
central spin that is dominated by the strongly coupling nuclei.
\subsubsection{The influence of $r_0$ onto the long-time limit}
\label{sec:influence-of-r0}
The deviation of the non-decaying fraction of the spin-polarization
from the QSA value of $S(0)/3$ observed in Fig.\ \ref{fig b0szt}(b)
justifies a more detailed analysis.
The influence of $r_0$ onto the long-time limit is depicted in Fig.\
\ref{fig-4}(a). With increasing $r_0$ we observe two effects: (i) the
non-decaying part of the polarization $S_\infty=\lim_{t\to\infty}
S(t)$ is decreasing, (ii) the relaxation time from the
pre-equilibrated intermediate state reached after a short transient
time of the order of $O(10T^*)$ into the steady-state is increasing.
Since the weakly-coupled nuclei can only contribute on large-time
scales, the second observation is intuitively clear due to the
increasing number of weakly coupling nuclei with increasing $r_0$ and
fixed $N$.
The first observation can be also understood within a simple argument.
In the QSA approach, no spin polarization transfer between the central
spin and the spin bath can occur since the nuclear magnetic field has
been treated statically. The spin decay is purely driven through
dephasing by averaging over the random and isotropic effective
magnetic field distribution yielding a finite steady-state value
$S_\infty=S(0)/3$.
Recent Bethe-ansatz
calculations\cite{FaribautSchuricht2013a,FaribautSchuricht2013b} up to
$N=44$ nuclei confirm that the non-decaying fraction of the
spin-polarization depends on the distribution of the coupling
constants. In any finite size representation of the model with a
small number of coupling constants $\{ A_k\}$ a finite non-decaying
fraction of the spin-polarization is found. This fraction, however,
decreases when additionally weak coupling nuclear spins have been
added. Faribault et al.\ gave an analytical
argument\cite{FaribautSchuricht2013b} why there must be a finite
non-decaying fraction of the spin-polarization in any finite-size
system, where the distribution of coupling constants $\{ A_k\}$ is
limited to the same order of magnitude. This agrees perfectly with
our findings for a finite size system.
\begin{figure} [tb] \centering
\hspace*{3mm}\includegraphics[width=75mm]{fig_4_a.eps}
\includegraphics[width=75mm]{fig_4_b.eps}
\caption{(color online) The behavior of $S(t)$ on long-time scales
for different values of $r_0$ and $N=18$ (a). Each shown curve has
been averaged over $n=50$ different random configurations $\{ A_k\}$
of couplings. The colored arrows indicate the long-time values for
$S_\infty$ predicted\cite{Merkulov2002} from the fluctuation ratio
$u=\expect{A^2}/\expect{A}^2$. (b) The non-decaying fraction
$S_\infty$, obtained by averaging the data over the calculated data
points in the interval $\frac t{T^*} \in [450:500]$, vs. the cutoff
$r_0$ for two different system sizes $N=14,20$. The straight lines are
linear fits. The analytical predictions using the fluctuation ratio $u$
are added for comparison. }
\label{fig-4}
\end{figure}
In contrast to exact evaluations of small systems, approximate
treatments\cite{Merkulov2002,
KhaetskiiGlazman2002,Zhang2004,CoishLoss2004} of the model allows to
access the thermodynamic limit. Such treatments require the
neglecting of higher order correlation effects and predict a finite
non-decaying fraction. Taking into account non-Markovian contributions
in second order of the transverse coupling,\cite{CoishLoss2004} leads
to a non-exponential correction to the mean-field solution stated in
Eq.\ (\ref{glg merkulov}) of the form $1/\log t$ in the absence of a
magnetic field in the decay to a finite steady-state limit. Even
though we observe a non-trivial transient behavior with a very slow
decay between $100<t/T^*<1000$
the data is not sufficient to extrapolate a $[\log t]^{-\alpha}$ correction or a
power-law decay to the non-decaying fraction from the data presented
in Fig.\ \ref{fig-4}(a).
However, we have been able to extract a power-law behavior from the
low-frequency properties of spin-noise spectra which will be discussed
in Sec.\ \ref{sec:spin-noise-spectra}.
For bridging to the experiments we can ask the question what is the
asymptotic non-decaying fraction of spin-polarization in the
thermodynamic limit $r_0\to \infty$ and $N\to\infty$.
In order to shed some light on this question within the framework
of the CET approach, we have investigated the scaling properties of the
non-decaying fraction of the polarization with respect to $r_0$ for
two bath sizes $N=14,20$ in the interval $1\le r_0\le 2$, in which
$S(t)$ has reached a steady-state limit as shown in Fig.\
\ref{fig-4}(a). The short time dynamics is always governed by the
time-scale $T^*$ and agrees very well with the QSA result.
The long-time limit is plotted as function of the cutoff $r_0$ for two
different bath sizes $N$ in Fig.\ \ref{fig-4}(b). The steady-state
polarisation decreases linearly with $r_0$ for $r_0 \leq 2$. Since
the CET dynamics does not properly represent the large $N$ limit for
$r_0$ exceeding $r_0>2$, as illustrated in panel (a) and discussed
above, no data is shown for such cutoffs.
The linear scaling of $S_\infty$ as function of $r_0$ suggests
that the non-decaying part of the central-spin polarization should
vanish when the influence of very large numbers of small coupling
constants is taken into account.
Extrapolating our linear fit to the data for $N=20$ indicates that
this is the case for $r_0 \approx 3.7$, corresponding to a ratio
$a\approx 10^6$. However, with increasing of the number of bath spins
$N$, the predicted cutoff $r_0$ for that $S_\infty$ should vanish
decreases as exemplified by the fit to two different spin bath
dimensions in Fig.\ \ref{fig-4}(b).
Our scaling analysis indicates that the spin correlations will
completely decay at infinitely long times in the thermodynamic limit,
e.g.\ $N\to \infty$ and then $r_0\to \infty$. This finding is fully consistent with an
extension of the QSA which takes into account the long-time
fluctuations of the nuclear magnetic field. Averaging Eq.\ (\ref{eq
eom}) over times larger than the electron spin-precession time but
much smaller than the nuclear spin precession time, the spin
precession contribution vanishes and
only the term $ (\vec{S}_0 \vec{n}) \vec{n} $ survices.
After inclusion of the explicit
time-dependence of the nuclear field $\vec{B}_{\rm eff} = B_{\rm eff}
\vec{n}(t)$ and spin, the ensemble average is given by Eq.\ (12) in
Ref.\ [\onlinecite{Merkulov2002}]
\begin{eqnarray}
\expect{\vec{S}(t)} &=& \expect{\vec{n}(t)
[\vec{n}(t)\vec{S}(t)] }
\label{eqn:merkulov-eqn-12}
\end{eqnarray}
Since $[\vec{B}_{\rm eff}(t)\vec{S}(t)]$ accounts for the total energy
of the central-spin model, it is a conserved quantity and time
independent: $[\vec{B}_{\rm eff}(t)\vec{S}(t)] = [\vec{B}_{\rm
eff}(0)\vec{S}(0)]$. Furthermore the nuclear-spin-spin correlation
function is isostropic leading to
\begin{eqnarray}
\expect{\vec{S}(t)} &=& \gamma(t) \frac{\vec{S}(0)}{3}
\end{eqnarray}
where $\gamma(t)= \expect{\vec{n}(t)\vec{n}(0)}$ is
defined as correlation function of the nuclear spin orientation.
For the long-time limit, $\gamma(t)$ approaches a stationary
value\cite{Merkulov2002} and is only dependent on the ratio
$u=\expect{A^2}/\expect{A}^2$. We added our estimates for $S_\infty$
using $\gamma(u)$ derived in Ref.\ \onlinecite{Merkulov2002} as
horizontal arrows in Fig.\ \ref{fig-4}(a) as well as crosses labeled
``analytic`` into Fig.\ \ref{fig-4}(b).
Although our finite size scaling qualitatively agrees with a
decreasing $\gamma(u)$ for increasing $r_0$,
the functional form of $\gamma(u)$ differs
from our linear scaling. This might be related to
the change of the largest hyperfine coupling when varying $r_0$
for fixed $N$. Keeping both $T^*$ and the largest hyperfine coupling
fixed requires the increase of $N$ when increasing $r_0$. This
would accelerate the decrease of $S_\infty$ when increasing $r_0$
in closer agreement with $\gamma(u)$.
Another argument of why the non-decaying fraction $S_\infty$ must
vanish in the long-time limit for $N\to \infty$ in QDs with smooth
electronic confinement potential was given by Chen et
al.\cite{ChenBalents2007} based on the distribution function of the
$A_k$. Although the total angular momentum in the isotropic CSM is
conserved it will be equally redistributed onto all nuclear spins at
large times. Since angular momentum transfer from the central spin to
the nuclear spin $k$ occurs on a time scale $t>1/A_k$, only those
spins within a given radius $R(t) = L_0 [\ln(A_0 t)]^{1/2}$ can
contribute to the spin decay using the Gaussian envelope function
(\ref{eqn:psi-R}) and $A_0=A(R_k=0)$. Only the strongly coupling
spins in the sphere with radius $R_s=R(T^*)$ contribute to the
short-time dynamics up to the time $t$. Therefore, the central spin
should decay as
\begin{eqnarray}
\expect{S^z(t)} &\propto & \frac{N(R_s)}{N(R(t))}
\propto [\ln(A_0 t)]^{-3/2}
\end{eqnarray} in the long time limit $t\gg T^*$ in three dimensions
and its expectation value vanishes for $t\to \infty$. Hence,
the non-decaying fraction of the central spin-polarization must vanish in
the thermodynamic limit. In any finite size
calculation,\cite{BortzStolze2007,Bortz2010,StanekRaasUhrig2013,FaribautSchuricht2013a,FaribautSchuricht2013b}
however, there exist a smallest coupling constant $A_{\rm min}$ which
limits the time scale beyond which the exact finite-size calculation
will deviate from the thermodynamic limit.
\subsubsection{Summary}
Before we move on to the discussion of the spin noise spectra, we
briefly summarize the results of this section. We have demonstrated
the accuracy of the CET by a comparison of the real-time dynamics with
small systems exactly solvable using ED. While in principle arbitraryly
long times could be reached with the CET approach, it is limited by
the largest polynomial order $N_C$ which has been included in the
calculation. We have established the quality of the statistical
evaluation of the momentum $\mu_{n,m}$ entering the extension of the CET
approach to ensemble averages. The short-time dynamics is governed by
the time scale $T^*$ and agrees qualitatively well with the QSA
result. However, we observe small deviations which can be traced to
(i) the distribution of the coupling constants $A_k$, to (ii) the
number of bath spins and to (iii) the ratio between the largest and
the smallest $A_k$. The larger this ratio is for fixed number of bath
spins $N$, the smaller the number of bath spins which couple with an
$A_k=O(1/T^*)$, the less bath spins contribute effectively to the
short time dynamics. The finite value of $S_\infty$ in a finite-size
system depends on the distribution function $P(A)$.
\subsection{Spin noise spectra}
\label{sec:spin-noise-spectra}
In recent
experiments,\cite{CrookerBayerSpinNoise2010,Dahbashi2012,LiBayer2012,*ZapasskiiGreilichBayer2013}
the spin noise spectra have been measured in QD ensembles.
Assuming independent QDs, it is sufficient to average the
generic spin-noise spectrum $S(\omega)$ over the distribution of time
scales $T^*$ and $g$-factors to make a connection to the experimental
data. Therefore, we focus on calculating the spin-noise spectrum $S(\omega)$
for a single QD
first and postpone the discussion to spin-noise spectra
for QD ensembles to Sec.\ \ref{sec:QD-ensemble-average}.
\begin{figure} [tb] \centering
\includegraphics[width=75mm]{fig_5_a}
\includegraphics[width=75mm]{fig_5_b}
\caption{(color online) The spin-noise spectrum $S(\omega)$ in the
absence of an external field for $r_0 = 1.5$. Panel (a) illustrates
the convergence properties of the CET for varying $N_C$ and the effect
of averaging results for several ($n$) configurations of the coupling
constants $A_k$ based on $N=18$ bath spins and $r_0=1.5$.
Panel (b) focuses on the finite bath size effects where we have
averaged over $n = 50$ different configurations $\{A_k\}$. $M
(\omega)$ denotes the Fourier transformation of the semi-classical
result where the $\delta$-peak in $M(\omega)$ at $\omega = 0$ has
been approximated by a Lorentzian. }
\label{fig b0spec}
\end{figure}
A realistic modelling of the QD requires the treatment of the order
$O(10^5)$ nuclei. Even on a fine frequency scale, the noise spectrum
will be continuous while $S(\omega)$ obtained from an exact simulation
for $\sim 20$ bath spins with a single distribution $\{ A_k\} $
clearly reflects the spectrum's discrete character.
In order to recover a continuous spectrum from the finite size CET
calculation, we use two different ingredients: (i) averaging over
random distributions $\{ A_k\} $ and (ii) choosing a rather low order
$N_C$ in the CET calculations. The averaging over several random
distributions $\{ A_k\} $ mimics a larger number of nuclei than
contained in a single configuration. By artificially reducing $N_C$,
we can effectively add a broadening to the individual $\delta$-peaks of
the finite size spectrum which would only be precisely recovered in the
limit $N_C\to \infty$.
Increasing $N_C$ systematically increases the frequency resolution
which we have employed to reveal a power-law in the low-frequency
spin-noise spectrum.
Note that no spectral weight is lost by this
procedure since the sum-rule (\ref{eq sum-rule}) is exactly fulfilled
for arbitrary $N_C$.
In order to illustrate the effect of those two ingredients, we show a
direct comparison of $S(\omega)$ for a single configuration ($n=1$) and
data averaged over $n=100$ random configurations $\{ A_k\}$ with $N_C
= 400$, $r_0 = 1.5$ and $N=18$ in Fig.\ \ref{fig b0spec}(a).
Additionally, the corresponding spectrum for a lower order $N_C = 100$
of the CET and a reduced number of configurations $n=50$ has been
added. While $S(\omega)$ obtained from a single configuration displays a
clear signature of a superposition of discrete peaks a
quasi-continuous spectrum is generated by the configuration averaging.
The calculations with a lower Chebychev order $N_C = 100$ reproduce
the calculations for $N_C = 400$ excellently, except at small
frequencies being consistent with linear scaling of the largest
accessible time scale with $N_C$. Even though it would be sufficient
to use rather small $N_C$ for an accurate description of the
short-time dynamics, the numerical effort increases substantially to
access the low frequency behavior of the spin-noise spectrum.
Based on the largest accessible time discussed in Sec.\
\ref{sec:benchmark}, the smallest accessible frequency of the CET is
given by $\omega_{\text{min}} \geq 10^{3/N_C}\pi \text{e} \Delta E/N_C$. Note
that there are two limiting factors to the accessibility of small
frequencies in our simulations: the finite $N_C$ and the finite cutoff
$r_0$ which set the boundary to the lowest $A_k$ and therefore, the
lowest non-zero excitation energy of the system. Choosing a larger $N_C$
than required by the means of this lowest excitation energy does not add
additional information to the finite frequency spectrum but only
sharpens the $\delta(\omega)$-peak.
The spectral functions $S(\omega)$ depicted in Fig.\ \ref{fig b0spec}(b)
for increasing number of bath spins and $r_0 = 1.5$ correspond to the
time resolved data shown in Fig.\ \ref{fig b0szt}(b). We note the fast
convergence of $S(\omega)$ as function of $N$. We also added the QSA spin
noise $M(\omega)$ stated in Eq.\ (\ref{eq:m-w}). The $\delta$-peak in
$M(\omega)$ at $\omega = 0$ has been approximated by a Lorentzian,
and its spectral weight is given by the non-decaying fraction of
the spin polarization.
The
high-frequency part of $S(\omega)$ agrees remarkably well with the QSA
spin-noise spectrum $M(\omega)$ rendering the excellent agreement in the
short-time dynamics between both approaches. As expected, the broad
high-frequency peak is centered at $1/T^*$ and its width given by
$1/2T^*$. However, we notice significant deviations between $M(\omega)$
and $S(\omega)$ for smaller frequencies. Those differences also reflect
the different transient behavior at times $t\gg T^*$ depending on the
configurations of $\{ A_k \}$.
\begin{figure}[t]
\begin{center}
\includegraphics[width=85mm]{fig_6}
\caption{(color online)
$S(\omega)$ vs $\omega T^*$ on a log-log scale for different values of $r_0=1.5,1.75,2$.
The dashed line is a fit to the low-frequency behavior
above the resolution-broadened $\delta(\omega)$-peak. The smallest accessible frequency $\omega_{\rm min}$
and the expectation value of the smallest contributing coupling constant $\overline{A}_{\text{min}}$
are indicated by arrows.
Parameters: $N_C=1000,n=100,N=18,b=0$.
}
\label{fig:s-w-isotrop-log}
\end{center}
\end{figure}
At low frequencies a small
shoulder around $\omega T^*\approx 0.1$ is observed in $S(\omega)$ (green
curve) in Fig.\ \ref{fig b0spec}(a) when using $N_C=400$ and $n=100$
configuration averages. This indicates the existence of an additional
low-frequency feature in $S(\omega)$, that is not covered by the QSA
result, located above the resolution-broadened zero-frequency $\delta$-peak
and below the Gaussian type high-energy peak around $\omega T^*\approx 1$.
In order to reveal the shape and nature of the low frequency part of
the spin-noise spectrum in greater details, we have pushed the CET
order to $N_C= 1000$ to significantly increase the frequency
resolution to $\omega_{\rm min}T^*= 8.5\cdot 10^{-3}$. As depicted in
Fig.\ \ref{fig:s-w-isotrop-log}, now the resolution-broadened
$\delta(\omega)$-peak, whose full-width half maximum can be estimated by
$2\omega_{\rm min}$, is well separated from the remaining low-frequency
part of $S(\omega)$: the shoulder has evolved into a threshold-type
behavior with a crossover around $\omega T^*\approx 0.03$ defined by the
smallest excitation energy of the finite size system. A power law
$\propto \omega^{-3/4}$ can be fitted in this region over approximately
one decade and is indicated as dashed line.
At larger frequencies, the previously discussed Gaussian-like peak
remains visible and is centered around $\omega T^*\approx 1$.
The $r_0$-dependency of the crossover scale separating the low-energy
feature in $S(\omega)$ from the resolution-broadened $\delta(\omega)$-peak is
clearly visible in Fig.\ \ref{fig:s-w-isotrop-log}. Increasing $r_0$
extends the frequency range which can be fitted by $\propto \omega^{-3/4}$
to lower frequencies. The cutoff parameter $r_0$ determines the
smallest hyperfine coupling $A_k$ contained in the configurations $\{
A_k\}$ and, therefore, the smallest excitation energy in the system.
We have indicated the value $\bar A_{\rm min}(r_0)= \expect{{\rm min}[
A_k ]} $ averaged over all configurations $\{ A_k\}$ by an additional
vertical arrow in Fig.\ \ref{fig:s-w-isotrop-log}. Apparently,
$\bar A_{\rm min}(r_0)$ determines the low-frequency crossover scale
to the power-law behavior.
We have demonstrated in Fig.\ \ref{fig-4} that the non-decaying
fraction of the central spin polarization and, hence the spectral
weight of the $\delta(\omega)$ peak decreases with increasing $N$ and
$r_0$ and eventually vanishes in the thermodynamic
limit.\cite{ChenBalents2007} Hence, spectral weight of the
$\delta(\omega)$-peak is transferred to the low-frequency part of $S(\omega)$
dominating the long-time properties of the spin correlation function
$S(t)$. We conjecture that the observed finite low-energy crossover
scale is approaching zero-frequency in the limit $r_0\to\infty$ and
$N\to\infty$.
Restricted by the finite size of the spin bath and the frequency
resolution, our numerical data is not accurate enough to predict the
precise analytic form of $S(\omega)$ in the thermodynamic limit. The
guide-to-the-eye fit to our data, however, would suggest $S(\omega)\propto
\omega^{-3/4}$ implying a $1/t^{1/4}$ decay at long-time scales. This
finding is consistent with the prediction $S(t) \propto
\ln^{-3/2}(t)$, because in the intermediate time regime we can access
via the CET both findings are almost indistinguishable in the time
domain. To predict a deviation a fit over more than one decade is
necessary.
\subsection{Magnetic field dependence of the spin noise}
\subsubsection{Transversal magnetic field}
\begin{figure} [tbp]
\centering
\includegraphics[width=85mm]{fig_7.eps}
\caption{(color online) All shown simulations have been averaged
over $n = 50$ configurations. The spectral functions are all based on
$N_C = 100$. (a) The correlation function $S(t)$ for two different
values of $b$. To distinguish the results for different $b$ an offset
of $0.5$ has been added to the upper curve. The function $E(t) =
\frac14 \text{exp}\left(-\frac12\left(\frac{t}{2T^*}\right)^2\right)$
approximates the envelope of the signals. (b) $S(t)$ for small
external field strengths. (c) The renormalized spin-noise function
$S(\omega)$ for $b = 2,4,...,10$. (d) The same data as shown in (c)
shifted by $\omega^* = \sqrt{b^2 + \frac12}$, pointing out that the
width of the occuring maxima is given by $T^*$ and that $\omega^*$ is
the exact frequency of the occuring spin precession. Parameters:
$N=20,r_0=1.5$.}
\label{fig bxstrong}
\end{figure}
Adding a transversal magnetic field $B_x$ to the system has two
important effects on the time evolution of the correlation function
$S(t)$. First, it breaks the conservation of the total polarization
along the $z$-axis. Second, the short-time dynamics is now governed by
a shifted time scale $1/T^{'*}(b) = \sqrt{(T^{*})^{-2} +
2\omega^2_L}$ which evolves continuously from $(T^{*})^{-1}$ with
the external magnetic field. This can be either analytically derived
from the von-Neumann equation or can be extracted from the numerical
data for large $b$ as depicted Fig.\ \ref{fig bxstrong}(d).
Applying a large external magnetic field $b=\omega_L T^* \gg 1$ causes
a damped oscillation in $S(t)$ whose dimensionless frequency is given
by $\omega^* = \sqrt{b^2 + 1/2}$. Since the hyperfine interaction
remains the origin of dephasing, the characteristic time scale of the
decay is governed by the intrinsic time scale $T^*$. Fig.\ \ref{fig
bxstrong}(a) demonstrates this behavior for two different magnetic
field strengths fulfilling $\omega_L T^* > 1$. Additionally the
function
\begin{eqnarray}
E(t) &=& \frac{1}{4}
\exp\left[-\frac{1}{2}\left(\frac{t}{2T^*}\right)^2 \right]
\end{eqnarray}
has been added as an approximation to the envelope of
the real-time dynamics.
Fig.\ \ref{fig bxstrong}(c) shows the spin-noise spectrum $S(\omega)$
for five different magnetic field strengths $b=2,4,6,8,10$. The
resulting spectrum contains a small contribution near $\omega = 0$
that vanishes quickly for growing magnetic field strength.
The main contribution of the spectrum, however,
consists of a central peak around
$\omega^*=\sqrt{b^2+1/2}$ with a field independent width proportional to
$1/T^*$ and is well separated from the low-frequency part
of the spectrum. The universality of $S(\omega)$ is revealed by shifting the
dimensionless frequency $\Delta\omega = \omega T^* -\omega^*$. All curves collapse
onto this universal curve independent of $b$ as depicted in Fig.\
\ref{fig bxstrong}(d). This also proves the claim that the envelope
function of the spin-decay is governed by $T^*$ independent of
$b$. Furthermore, we note that $S(\omega)$ is independent of the number of
bath spins for a fixed $r_0$.
Fig.\ \ref{fig bxstrong}(b) focuses on the time evolution $S(t)$ for a
weak external field. The short-time evolution of the central spin
remains unaffected by the application of a very small magnetic field,
$b=\omega_L T^*\ll 1$. By symmetry breaking, $S_\infty=0$, so that
the correlation function $S(t)$ approaches zero in the long time
limit. Thus the $\delta$-peak occurring at $\omega=0$ in the spin-noise
function $S(\omega)$ must already vanish for an infinitesimal small
transversal field. Due to the finite-time resolution of the CET, we
only can show the transient behavior, while no change can be resolved
in $S(\omega)$ for $\omega\to 0$ (not shown) compared to Fig.\ \ref{fig
b0spec}.
The long-time transients are not governed by the nuclear-field
fluctuation time $T^*$ but by $\omega_L$. This is clearly visible in
Fig.\ \ref{fig bxstrong}(b) where we plotted $S(t)$ for three
different weak magnetic field values as function of the dimensionless
time $\omega_L t$. We find universality on the intermediate time scale
depicting a very slow decay for $t\to\infty$. This findings agree with
the Bethe-ansatz data of Faribault et
al.\cite{FaribautSchuricht2013a,FaribautSchuricht2013b}
We can ask how does the spectrum evolve from $b=0$ to finite $b>0$.
Since the eigenvalue spectrum evolves adiabatically from $b=0$, we
must observe two effects. (i) the $\delta(\omega)$-peak vanishes and its
spectral weight is shifted to finite frequencies, and (ii) all finite
frequency excitation energies will shift with the magnetic field.
Therefore, the center of the Gaussian shaped peak with a width of
$T^*$ gradually evolves to $\sqrt{(T^{*})^{-2} + 2\omega^2_L}$ as
function of the magnetic field as depicted in Fig.\ \ref{fig bxstrong}(d).
\subsubsection{Longitudinal magnetic field $B_z$}
\begin{figure}[tb]
\begin{center}
\includegraphics[width=85mm]{fig_8_a}
\includegraphics[width=85mm]{fig_8_b}
\caption{(color online) The spin correlation function $S(t)$ for
increasing applied longitudinal fields (a), $N = 18$ and an average
over $n=10$ different configurations. (b) Corresponding calculated
spectral functions $S(\omega)$. The inset depicts the spectral weight
of the $\delta(\omega)$-peak as function of $b$ corresponding to the
non-decaying fraction of $S(t)$. Parameter: $N_C = 100$, $n = 50$ and
$r_0=1.5$.
}
\label{fig-7-Sw-for-finite-Bz}
\end{center}
\end{figure}
Applying a magnetic field $B_z$ in longitudinal direction induces a
finite energy difference between the different eigenstates of the
total spin component $J_z$, and the spin-flip processes of the central
spin are suppressed by this additional energy
barrier. The initial spin-decay is almost independent of $b$ for
ultra-short time scales until $t\approx 1/b$ where the further decay
of $S(t)$ starts to be oppressed. Consequently, the non-decaying
fraction of the spin polarization is increasing with $b$: the
spin-decay is completely suppressed for $b\to\infty$ as can be seen in
Fig.\ \ref{fig-7-Sw-for-finite-Bz}(a).
The corresponding spin-noise spectrum is displayed in Fig.\
\ref{fig-7-Sw-for-finite-Bz}(b). The increasing non-decaying fraction
$S_\infty$ in $S(t)$ corresponds to an increasing spectral weight of
the $\delta$-peak located at $\omega=0$. Since the total spectral weight
is conserved, there is a spectral weight transfer from the broader
peak centered around $\omega^* T^*\approx 1$ to the $\delta(\omega)$-peak.
The broad finite-frequency peak is shifted to higher frequencies and
is again centered at $\omega^*\approx \sqrt{b^2+1/2}$ at large magnetic
field. The major difference to the application of a transversal
magnetic field is the loss of spectral weight at finite frequencies in
favor of the zero-frequency peak.
As discussed above, the CET is restricted to a finite accessible
$t_{\rm max}\propto N_C$ which defines the lowest frequency resolution
$\Delta \omega$. Therefore, the $\delta(\omega)$ cannot be accurately
resolved. We define a low frequency cutoff $\omega_{\rm min}$ and
calculate the total spectral weight $S_+$
\begin{eqnarray}
S_+ &=& \int_{\omega_{min}}^\infty \frac{d\omega}{2\pi} \, S(\omega)
\end{eqnarray}
of the finite frequency part of the spin-noise
spectrum. Since the CET spin-noise exactly fulfills the spectral
sum-rule, we obtain the spectral weight $S_0=2\pi S_\infty$ of the
$\delta(\omega)$-peak from the difference $S_0 = 1/4-S_+$. The resulting
$S_\infty$ as function of $b$ is plotted as inset in Fig.\
\ref{fig-7-Sw-for-finite-Bz}(b) and reveals the increase of the weight
with increasing longitudinal magnetic field.
\section{Spin dynamics of the anisotropic CSM}
\subsection{Spin dynamics in the Ising limit}
\label{sec:results-ising-limit}
Up until now, the results were restricted to the isotropic case,
$\lambda=1$. In order to set the stage for the anisotropic model with
finite $\lambda<\infty$, we focus on the opposite limit $\lambda
\rightarrow \infty$, the Ising limit, in this section. Then, the
Hamiltonian (\ref{eq hamiltonian}) of the CSM reduces to an Ising
interaction between the central spin and the nuclear spins
\begin{align} H &= \omega_L \vec{S} \vec{n}_B + \sum_k A_k S^z I_k^z
\end{align} and can be solved in a closed analytical form in some
limiting cases.\cite{Koppens2007,FischerLoss2008,Testelin2009} This
limit describes a pure heavy-hole spin.\cite{Koppens2007,FischerLoss2008,Testelin2009}
Obviously, all nuclear spin operators $I_k^z$ commute with $H$ and are
conserved. Therefore, the spin-bath is static and fully determined by
the eigenvalue configuration $\{ m_k \}$ of all $I_k^z$. In the
absence of an external magnetic field, the eigenenergies are given by
$E^*(\sigma, \{ m_k \})$
\begin{eqnarray}
E^*(\sigma, \{ m_k \}) &=&
\sigma \sum_k A_k m_k = \sigma E^* \{ m_k \})
\end{eqnarray}
where $\sigma$ is the eigenvalue of $S^z$.
\begin{figure} [tb]
\centering
\includegraphics[width=85mm]{fig_9}
\caption{(color online) The spin noise function $S(t)$ for the Ising
limit $\lambda \rightarrow \infty$ of the anisotropic central spin
model plotted vs the dimensionless time $\tau = t/4bT^*$ for three
values of the external magnetic field $b=2,4,8$. An offset of $0.5$
has been added to distinguish the three individual curves. The
envelope of $E_z (\tau)$ given by Eq.\ (\ref{eq eble}) has been added
as a black line to each simulation. All curves have been calculated
for $r_0 = 1.5$, $n = 20$ and $N = 18$. }
\label{fig an0}
\end{figure}
In a finite external magnetic field, the Hamilton matrix decomposes in
$2\times 2$ subblocks for each fixed nuclear configuration $\{ m_k
\}$. For an external magnetic field in $x$-direction, we obtain the
two eigenenergies $E_\pm( \{ m_k \})) = \pm\sqrt{\omega_L^2 +[E^* \{ m_k
\})]^2}/2$. The electronic spin precesses around the resulting
effective magnetic field $\vec{B}_{\rm eff}(\{ m_k \}) = ( \omega_L, 0,
E^* (\{ m_k \}) )^T$ which depends on the bath
configuration.\cite{Koppens2007}
Averaging over the Larmor oscillations, the projection of the spin
component $\vec{n} [\vec{n}\vec{S}(0)]$ onto the magnetic field
direction $\vec{n}$ survives as already discussed above in the context
of Eq.\ (\ref{eqn:merkulov-eqn-12}). For a spin initially polarized
in $z$-direction, the $z$-component of the non precessing contribution
$(\vec{e}_z \vec{n})^2 |\vec{S}(0)|$ has to be averaged over all
nuclear configurations $\{ m_k \}$
\begin{eqnarray}
\label{eqn:non-decaying-fraction-ising}
\expect{ (\vec{e}_z
\vec{n})^2}_{ \{ m_k \} } &=&
\expect{\frac{[E^* (\{ m_k \})
]^2}{\omega_L^2 +[E^* \{ m_k \})]^2}}_{ \{ m_k \} }
\end{eqnarray}
where $\vec{e}_z $ is the unit vector in $z$-direction.
For large magnetic fields, this average yields $\expect{ (\vec{e}_z
\vec{n})^2} = 1/4b^2$ while in zero magnetic field $\expect{
(\vec{e}_z \vec{n})^2}=1$.
When applying a transversal external magnetic field $B_x$, the fast
dynamics of a spin initially polarized in $z$-direction is determined
by the Larmor frequency $\omega_L\propto b$. The spin-decay, however, is
governed by a slowly varying envelope function. Fig.\ \ref{fig an0}
depicts the real-time dynamics of $S(t)$ for three values of a large
external magnetic field for the Ising limit of the Hamiltonian. An
offset of $0.5$ has been added to distinguish the different curves.
Testelin and collaborators\cite{Testelin2009} have extended the
semi-classical approach of Merkulov et.\ al.\cite{Merkulov2002} to the
Ising limit of the CSM and derived the analytic decay function
$E_z(t)$
\begin{eqnarray}
E_z (\tau) &=& \frac14
\left[\frac{\text{cos}(4b^2\tau + \frac12 \text{arctan}(\tau))}{(1 +
\tau^2)^{1/4}} \right . \nonumber \\ && \left . + \frac1{4b^2} \left( 1 -
\frac{\text{cos}(4b^2\tau + \frac32 \text{arctan}(\tau))}{(1 +
\tau^2)^{3/4}} \right)\right],
\label{eq eble}
\end{eqnarray}
for the spin-component in $z$-direction in the limit of
large magnetic fields $b\gg 1$ where the new dimensionless time scale
$\tau = t/4bT^*$ has been introduced. It consists of two oscillatory
terms governed by the Larmor frequency $4b^2\tau= \omega_L t$ with an
additional phase shift term and two decaying envelope functions $\propto
(1+\tau^2)^{1/4}$ and $\propto (1+\tau^2)^{3/4}$. The functional form
of this non-exponential decay has been derived\cite{Koppens2007} by
averaging the coherent spin-precession of the central spin for a given
nuclear configuration $\{ m_k \}$ over all nuclear configurations
using a Gaussian distribution of the nuclear field. The value of the
non-decaying fraction $1/4b^2$ agrees exactly with the prediction of
Eq.\ (\ref{eqn:non-decaying-fraction-ising}). The increase of the
dephasing time $T_{\rm deph} = 4b T^*$ with increasing field strength
can be understood by the suppression of the effective field fluctuation.
We have added the envelope of the function $E_z(\tau)$ to Fig.\ \ref{fig an0}
for the three different magnetic field strengths. The semi-classical
approach excellently reproduces the exact simulations for large
magnetic fields $b \gg 1$. Our numerical results confirm the slow
decay of the longitudinal spin component $\propto 1/\sqrt{t}$ for
large times in this limit.
\begin{figure} [tb]
\centering
\includegraphics[width=80mm]{fig_10}
\caption{(color online) The spin noise spectra $S(\omega)$ for
$(\omega-\omega_L)\geq 0$ in the Ising limit calculated by exact Fourier
transformation of (\ref{eqn:exact-ising}) for different external
magnetic fields $b$. The inset shows the non-decaying fraction
$S_\infty$ which defines the spectral weight of the $\delta$-peak at
$\omega=0$. Parameters: $N=18$, $r_0=1.5$. }
\label{fig-10-ising-s-w-exact}
\end{figure}
While the power-law decay for large magnetic fields is well
established\cite{Koppens2007,FischerLoss2008,Testelin2009} it is not
obvious whether the analytic form of the long-time envelope prevails
in the crossover regime to small fields. In order to reveal the
envelope function for the long time decay, one can subtract the
non-decaying fraction from $S(t)$ and multiply the remaining
oscillatory part with the dominating long-time decay. We find that the
amplitude of the oscillatory function
$(S(t)-S_\infty)(1+\tau^2)^{1/4}$ reaches a time-independent long time
limit -- not shown here.
To substantiate these findings, we expand the initial
spin polarized state $\ket{\uparrow, \{ m_k \}} $
\begin{eqnarray}
\ket{\uparrow, \{ m_k \}} &=& c_+( \{ m_k \}) \ket{+,
\{ m_k \}} + c_-( \{ m_k \}) \ket{-, \{ m_k \}} \nonumber
\end{eqnarray}
using the exact
eigenstates $\ket{\pm, \{ m_k \}}$ for a given configuration $ \{ m_k \}$ and exactly calculate the spin-noise function $S(t)=
\expect{S^z(t)}_{\rho_p}/2$ by averaging over all excitation energies
$\Delta E( \{ m_k \})= \sqrt{\omega_L^2 +[E^* \{ m_k \})]^2}$:
\begin{eqnarray}
\label{eqn:exact-ising} S(t) &= & S_\infty + \sum_{ \{ m_k \}} c_+^2
c_- ^2\cos(\Delta E( \{ m_k \}) t) \\ S_\infty &=& \sum_{ \{ m_k \}}
\frac{1}{4}\left(c_+^2-c_-^2\right)^2
\,\, .
\end{eqnarray}
While the time-independent part $S_\infty$ is
equivalent to (\ref{eqn:non-decaying-fraction-ising}) and defines the
spectral weight of $S(\omega)$ at zero frequency, the Fourier
transformation of the second part can be calculated analytically and
yields the exact finite frequency contribution to the noise spectra.
For the frequency distribution of $\sqrt{\omega_L^2 +[E^* \{ m_k \})]^2}$
we conclude that (i) there exists a finite threshold frequency $\omega_{\rm
th}$ below which $S(0<\omega<\omega_{\rm th})=0$ and (ii)
$\omega_L=\lim_{N\to\infty} \omega_{\rm th}$. Consequently, the spectral gap
in $S(\omega)$ below $\omega_{\rm th}$ will prevail in the thermodynamic
limit. Since the largest frequency in (\ref{eqn:exact-ising}) is
limited by the two fully polarized configurations $\{ m_k =\uparrow\},
\{ m_k =\downarrow\}$, the spin noise spectrum $S(\omega)$ also must vanish
for $|\omega|> \sqrt{\omega_L^2 +[E^*_{\rm max}]^2}$.
Fig.\ \ref{fig-10-ising-s-w-exact} depicts the exact $S(\omega)$ as
function of $\omega-\omega_L$ for different magnetic fields. $S(\omega)$ can be
fitted by a power law $(\omega-\omega_L)^{-\alpha}$ very close to the Larmor
frequency. We have extracted a universal exponent $\alpha\approx 1/2$
independent of $b$. For small $b$, however, the non-decaying fraction
of the spin dominates and the finite frequency spectral function is
very small. Therefore, the fitting accuracy decreases for $b\to 0$.
Approximating $S(\omega)$ by $C (\omega-\omega_L)^{-\alpha}$ close to the threshold
frequency up to some finite cutoff frequency $\omega_{\rm max}$ and a
normalization constant $C$, we can analytically extrapolate the
long-time approach to $S_\infty$ from
\begin{eqnarray}
\Delta S(t) &= &S(t)-S_\infty =
2\Re e
\left[\int_{\omega_L}^{\omega_{\rm max}} \frac{d\omega}{2\pi}
\frac{C}{(\omega-\omega_L)^\alpha} e^{i\omega t}\right] \nonumber \\ &=& \frac{C}{\pi
t^{1-\alpha}} \Re e \left[ e^{i\omega_L t} \int_0^{(\omega_{\rm max} -\omega_L)t} d
u \frac{e^{iu}}{u^\alpha} \right]
\end{eqnarray}
which can be approximated for $t\to \infty$ and
$\alpha=1/2$ to
\begin{eqnarray} \Delta S(t) &\approx & \frac{a C}{\pi}
\frac{\cos(\omega_L t+\pi/4) }{\sqrt{t}}
\end{eqnarray} where $a\approx 1.25331$. This result recovers the
long-term limit of $E_z(t)$ stated in Eq.\ (\ref{eq eble}) independent
from the $b\gg 1$ limit.
\subsection{Spin dynamics in the generic anisotropic CSM}
\label{sec:results-anisotropic-CSM}
\begin{figure} [t]
\begin{center}
\includegraphics[width=84mm]{fig_11_a}
\hspace*{5mm}\includegraphics[width=78mm]{fig_11_b}
\end{center}
\caption{(color online) (a) $S(t)$ vs $t/T^*_\lambda$ for different
values of the anisotropy factor $\lambda=1,2,4,8$. (b) Spin-noise
spectra vs $\omega T^*$ for the same parameters as in panel (a) and $N_C = 100$.
A $\delta$-peak at zero frequency resolved by the CET with the same number
of moments has been added as reference.
Parameters: $N=18, n = 50, r_0=1.5$. }
\label{fig-11-b=0-anisotropic-CSM}
\end{figure}
\begin{figure} [bt]
\centering
\includegraphics[width=75mm]{fig_12.eps}
\caption{(color online) Comparison between the analytical prediction
(blue) of Eq (26c) of Ref.\ [\onlinecite{Testelin2009}] using a QSA
and the CET approach for $b/\lambda>1$. All curves have been calculated
for $b = 10$ and $\lambda=2$ or $\lambda=10$. We have added an offset
of $0.5$ to the $\lambda=2$-curves for better comparison. Parameters:
$N=18,r_0=1.5$. }
\label{fig-CET-vs-QSA-strong-field}
\end{figure}
\begin{figure} [bt]
\begin{center}
\includegraphics[width=70mm]{fig_13_a}
\includegraphics[width=70mm]{fig_13_b}
\includegraphics[width=70mm]{fig_13_c}
\end{center}
\caption{(color online) $S (t)$ in a transversal magnetic field: (a)
for a fixed value $\lambda=10$ and $b=2,6,10$, and (b) for a fixed
value $b=4$ and four different values of $\lambda=2,20,50,100$. For
all cases we have added the envelope function
$0.25\exp[-(t/T^*_\lambda)^2/8]$ to reveal the difference to the
short-time dynamics of the isotropic CSM. (c) Rescaled spin
correlation function
$S(t)(1+\tau^2)^{1/4}$ vs
$t/T^*_\lambda$ for the same parameters as in panel (b) with $\tau =
t/4bT^*_1$.
Parameters: $N=18, n=20, r_0=1.5$. }
\label{fig-12-anisotrop-CMS-St-finite-bt}
\end{figure}
So far we have discussed the two extreme limits of the Hamiltonian
(\ref{eq hamiltonian}): the isotropic CSM and the Ising limit. Now we
investigate the influence of an arbitrary anisotropy factor $\lambda$
onto the spin-noise spectra as function of a transversal external
magnetic field.
In the absence of an external magnetic field, the decay of the spin
correlation function $S(t)$ is only caused by the transversal
terms in the Hamiltonian whose relative strength is controlled by the factor
$1/\lambda$. Presenting the data as function of the dimensionless
time $t/T^*_\lambda = t/\lambda T^*$ clearly reveals the
$\lambda$-dependence of the intrinsic time scale: the location of the
minimum in the short-time dynamics remains almost independent of $\lambda$
as depicted in Fig.~\ref{fig-11-b=0-anisotropic-CSM}(a). With increasing
$\lambda$, the non-decaying fraction of $S(t)$ monotonically increases,
and the curves basically interpolate between the isotropic CSM limit
and Ising limit where $S(t)=1/4=const.$ does not decay at all.
The corresponding spin-noise spectrum $S(\omega)$ is shown as a function of
the dimensionless frequency $\omega T^*_1$ in
Fig.~\ref{fig-11-b=0-anisotropic-CSM}(b). We have added the result for
the isotropic CSM as reference for comparison. Since the CET approach
has a finite resolution of a $\delta(\omega)$-peak for $\omega\to 0$, the data
for $\lambda=\infty$ serves as reference and indicates the change of the
spectral function at finite frequencies as function of $\lambda$. The
height of the Gaussian shaped peak close to $\omega \approx 1/T^*_\lambda$
of the isotropic CSM decreases with increasing $\lambda$, and spectral
weight is shifted in parts into the $\delta$-peak at $\omega=0$ and to
smaller frequencies $\omega>0$. For $\lambda>3$, the peak close to $\omega
\approx 1/T^*_\lambda$ has disappeared and the spectra are
monotonically increasing for decreasing $\omega$.
Now we proceed to the spin response in a transversal magnetic field.
A word is in order to clarify the definition of a strong external
field in the anisotropic CSM. The Hamiltonian
(\ref{eq hamiltonian}) contains two terms driving (i) the decoherence
and (ii) the dephasing of the central spin in $z$-direction: (i) the
transversal spin-flip term is governed by the characteristic timescale
$T^*_\lambda$ and (ii) the external transversal field is governed by
$T^*_\lambda / \lambda b$. If $b\cdot\lambda \gg 1$ holds, the
external magnetic field dominates over the spin-flip term, and,
therefore, is called a strong external field. The central spin
performs several Larmor precessions before the decoherence due to the nuclei
sets in to the spin bath. In the opposite limit, $b\cdot\lambda \ll 1$, the
central spin cannot complete a single Larmor precession on the time
scale $T^*_\lambda$ on which $S(t)$ decays due to the hyperfine
interaction.
We begin to investigate the strong field regime defined by $b\cdot\lambda > 1$. We
compare the analytical predictions\cite{Testelin2009} by the QSA with our
CET results for short-time and intermediate-time scales as depicted in
Fig.~\ref{fig-CET-vs-QSA-strong-field}. While for $\lambda=2,b=10$ good
agreement is found between both methods significant deviations are
observed for an increased asymmetry $\lambda=10,b=10$ already at short
times. While the numerical exact results show a monotonic decay of the
envelope function, the QSA predicts an oscillation of the spin
polarization amplitude. It turns out that this result can be generalized
to the statement, that Eq (26c) of Ref.\ [\onlinecite{Testelin2009}] using
the QSA only describes $S(t)$ in a transversal magnetic field adequately if
the parameter $b / \lambda$ is large
compared to $1$. This indicates the importance of the ratio $b/\lambda$
for separating different types of the dynamics in the anisotropic central spin
model.
To gain more information on the influence of the ratio $b / \lambda$,
CET results for $S(t)$ are shown at either a
fixed $\lambda=10$ and different magnetic field strengths $b=2,6,10$ in
panel (a), or at fixed magnetic field strength $b=4$ and several values
of $\lambda=2,10,20,50$ in panel (b) of
Fig.~\ref{fig-12-anisotrop-CMS-St-finite-bt}.
Augmenting the numerical data with the envelope funtion
$0.25\exp[-(t/T^*_\lambda)^2/8]$ demonstrates that the short-time
response is governed by the characteristic time scale $T^*_\lambda$
and the decay is well captured by a Gaussian envelope function for
$b/\lambda\ge 1$ as predicted by the QSA.\cite{Testelin2009}
Deviations from such a Gaussian decay increase with decreasing ratio
$b/\lambda<1$ and are very pronounced for $b/\lambda=1/25$. For these
parameters, the initial decay occurs much faster but on long-time
scales significantly slower than described by a Gaussian envelope
function. The dependency of the crossover on the ratio $b/\lambda$
can be understood from the fact that the hyperfine-field asymmetry
increases with increasing $\lambda$, and a larger transversal external
magnetic field $b$ is required to induce a dynamics similar to the
isotropic CSM. Only at very strong transversal magnetic fields $B_x$,
the relevance of anisotropy of the hyperfine interaction is reduced
and the Gaussian type of decay is expected for $b/\lambda > 1$.
For strong external fields, $b\cdot\lambda > 1$, the ratio $b/\lambda$
solely dictates the form of the envelope of $S(t)$:
for several different combinations of $b$ and $\lambda$, the envelope functions
for $S(t)-S_\infty$ are identical - being not explicitly shown here.
For further illustration of the crossover from a Gaussian to the Ising
type decay, we plotted the rescaled spin correlation function
$S(t)[1+\tau^2]^{1/4}$ vs $t/T^*_\lambda$ in
Fig.\ \ref{fig-12-anisotrop-CMS-St-finite-bt}(c), where $\tau = t/4b T^*$.
This reveals an Ising type power law decay for long-time scales that is
recovered for $\lambda\to\infty$.
For $b=4$ and $b/\lambda=2$, we clearly observe a Gaussian dominated
decay. For decreasing $b/\lambda$, an increasing intermediate-time
regime develops where the envelope function follows a slower power-law
type decay. Note that $S(t)[1+\tau^2]^{1/4}$ yields a non-decaying
oscillation in the Ising limit, which can not be shown in figure
\ref{fig-12-anisotrop-CMS-St-finite-bt}(c) due to the underlying
time-scale.
\begin{figure}[t] \centering
\includegraphics[width=75mm]{fig_14}
\caption{(color online) Spin noise spectra $S(\omega)$ for $b = 4$ and
$\lambda = 2,10,20,50$, plotted vs $(\omega-\omega_L)T^*_\lambda$. Since the
characteristic time-scale $T^*_\lambda$ increases with increasing
$\lambda$, we adjusted the number of Chebychev momenta to
$N_C(\lambda=2)=100$, $N_C(\lambda=10)=300$, $N_C(\lambda=20)=600$,
and $N_C(\lambda=50)=1500$ to provide an adequate frequency
resolution. The chosen parameters are equivalent to those in
FIG. \ref{fig-12-anisotrop-CMS-St-finite-bt}. }
\label{fig-spin-noise-rescaled-St}
\end{figure}
We have been able to identify three different regimes for $b\cdot\lambda>1$.
For $b> \lambda$, (i) the decay is of Gaussian type\cite{Testelin2009} similar
to the isotropic CSM. As long as $\lambda$ is less than one order of
magnitude larger than $b$, (ii) the decay of $S(t)$ deviates from
the Gaussian envelope, but the long-time decay is still governed by
the tail of the envelope function. In the last regime, (iii) where
$\lambda \gg b$ holds, the behavior of the spin-noise function
approaches the Ising regime discussed in the previous section.
The corresponding spin-noise spectrum for this crossover regime
$b/\lambda<1$ is depicted as function of $(\omega-\omega_L)T_\lambda^*$ in
Fig.\ \ref{fig-spin-noise-rescaled-St}. In leading order, the peak
position of $S(\omega)$ is clearly given by the Larmor frequency
$\omega_L$, while the peak width is governed by $T^*_\lambda$. Note that
the smallest and the largest value of $\lambda$ differ by a factor of
$25$. Therefore, the increase of $T^*_\lambda$ requires a
significant increase of the Chebychev order $N_C$ for a reliable
resolution of the spectra. $S(\omega)$ evolves from a Gaussian shape for
$b/\lambda=2$, to a precursor of a threshold behavior for
$b/\lambda=2/25$. While the high frequency tail can be fitted with a
Lorentzian, the low frequency spectrum is rapidly suppressed below the
Larmor frequency for $0<\omega<\omega_L$ with increasing anisotropy. In the
limit $\lambda\to \infty$, the Ising limit of the spectrum, as shown
in Fig.~\ref{fig-10-ising-s-w-exact}, must be recovered. Therefore,
the increasing low frequency gap with increasing asymmetry prevails in
the thermodynamic limit.
Let us now turn to small magnetic fields characterized by the
condition $b\cdot\lambda \ll 1$. In this case the central spin
experiences decoherence induced by the spin bath before a whole Larmor
precession can occur. The short-time dynamics in this regime is
governed by $T^*_\lambda$ and is of the same form as shown for
different values of $\lambda$ in Fig.\
\ref{fig-11-b=0-anisotropic-CSM}(a). After the initial spin decay,
$S(t)$ will approach zero on a large timescale that is dictated by
$\lambda/\omega_L$. This long-time behavior of $S(t)$ is analogous to
the isotropic CSM as shown in Fig.\ \ref{fig bxstrong}(b). Combining
the information provided by both figures fully describes the long-time
and short-time behavior of $S(t)$ in the weak field limit of the
anisotropic CSM, and we do not show further explicit results for this
regime.
So far we did not discuss the correlation function $S(t)$ for a light hole spin.
Such QDs are characterized by an anisotropy factor
$\lambda = 1/2< 1$, so that the spin-flip terms dominate over the
Ising contribution. Since $b/\lambda > 1/\lambda^2>1$ is bound in the
strong field regime, no Ising-type behavior can be observed and $S(t)$
always shows a Gaussian type decay for $b\cdot\lambda > 1$. However, all other findings
discussed above remain valid for the case $\lambda < 1$: an analogous
behavior for the weak field limit $b\cdot\lambda \ll 1$ is found.
Overall, the spin noise is very similar to the isotropic CSM but with a faster
characteristic time scale $T^*\to T^*_\lambda<T^*$.
\section{Spin noise spectra in quantum dot ensembles}
\label{sec:QD-ensemble-average}
\begin{figure}[t]
\begin{center}
\includegraphics[width=75mm]{fig_15_a}
\includegraphics[width=75mm]{fig_15_b}
\caption{(color online)
(a)
Ensemble averaged spin-noise spectra vs $\omega$ for five different values of $B$
for a dot-confined electron spin, i.~e.\ $\lambda=1$.
Defining the average timescale $\overline{T}^* =\expect{T^*}=1$ ns as reference,
and using $|g_e|=0.54$ as in Ref.~\onlinecite{CrookerBayerSpinNoise2010},
$S(\omega)$ has been obtained by averaging over $n=100$ different spectra
generated from a constant distribution of $T^* \in [0.2\,{\rm ns},1.8\,{\rm ns}]$
of characteristic single QD time scales.
For better visibility, an offset proportional to $b=\omega_L \overline{T}^*$ has been added to $S(\omega)$.
Parameters: $N_C=15,N=18$. The black lines indicate Lorentzian fits to the
individual curves.
(b) Evolution of the ensemble noise spectrum with increasing $N_C$ at fixed $B=300$ G.
}
\label{fig:s-w-ensemble-average-electron}
\end{center}
\end{figure}
The spin noise spectrum has been
measured\cite{CrookerBayerSpinNoise2010,Dahbashi2012,LiBayer2012,*ZapasskiiGreilichBayer2013}
on QD ensembles charged with a single electron or hole.
To bridge between the single QD calculations and the recent experiments,
we have performed an ensemble average of $n$ single independent QD spectra
with different configurations $\{A_k\}$. We have used the average timescale
$\overline{T^*} =\expect{T^*}$ as reference timescale with
equally distributed individual $T^*\in [0.2\overline{T^*}, 1.8 \overline{T^*}]$ for each individual QD.
\begin{figure}[tb]
\begin{center}
\includegraphics[width=80mm,height=50mm]{fig_16_a}
\includegraphics[width=80mm,height=50mm]{fig_16_b}
\caption{(color online)
(a) Ensemble averaged spin-noise spectra in the asymmetric CSM for $\lambda=5$ (red)
and $\lambda=50$ (blue) for different $b$.
$S(\omega)$ has been obtain by averaging over $n=2000$ different single QD
spectra generated from a constant distribution of $T^*_i\in [0.2 T^*, 1.8
T^*]$, and the individual $g$ factors are taken from a Gaussian distribution
with $ \Delta g / g=0.2$.
(b) Influence of the $g$-factor distribution function on the shape of the
spin-noise spectrum: comparing the $\lambda=5$ data of (a) with an
ensemble averaged $S(\omega)$ where the individual $g$ factors $g_i$
are taken from a constant distribution of $g_i/g \in [0.8,1.2]$ (green curves).
For better visibility, an offset proportional to $b$ has been added to $S(\omega)$.
Parameters: $N = 18, N_C=50, r_0 = 1.5$.
}
\label{fig:s-w-ensemble-average-hole}
\end{center}
\end{figure}
For QD ensembles with dot-confined electrons
the distribution of $g$-factors is very sharp,\cite{GreilichBayer2007} and the spread
of the $g$ factors can be neglected. The resulting ensemble averaged spin-noise spectra
are depicted in Fig.\ \ref{fig:s-w-ensemble-average-electron}(a). An offset proportional
to the magnetic field $b$ has been added to $S(\omega)$ for better
visibility. The black lines indicate a Lorentzian fit to each individual curve.
For the absolute units in Fig.\
\ref{fig:s-w-ensemble-average-electron}, we have set the average electron
dephasing time $\overline{T^*} = 1\,\text{ns}$ and have used the
experimentally determined\cite{CrookerBayerSpinNoise2010} electron
$g$-factor $|g_e|=0.54$. Our results demonstrate that phenomenological
Lorentz fits to the experimental spin noise spectra are justified for
strong external magnetic fields $b=\omega_L \overline{T^*} >1$. For small
$b$ significant deviations from a simple Lorentzian fit are observed
in the ensemble averaged spectra as being reported in the experimental
data.\cite{CrookerBayerSpinNoise2010} These deviations are related to
the low-frequency features of the spin-noise ignored in such a simple
Lorentzian fit.
Our results are in good agreement with the experimental data of Fig.\
4(a) in Ref.\ \onlinecite{CrookerBayerSpinNoise2010}. For larger
Chebyshev orders the ensemble averaging still provides peaks around
the Larmor-frequency in $S(\omega)$ that can be approximated by a
Lorentzian, but a clear gap between zero frequency and $\omega_L$
occurs as discussed before.
Note that the spectra presented in Fig.\
\ref{fig:s-w-ensemble-average-electron}(a) have been calculated for a
relatively small Chebyshev order $N_C = 15$
corresponding to a frequency resolution of $\omega> 20$ MHz
for a magnetic field of $B=300$ G up to $\omega> 60$ MHz for $B=800$ G.
This would mimic a limited resolution in the experiments.
The effect of an increase of the frequency resolution (increasing $N_C$)
on the spectral properties
of the ensemble averaged spin noise is depicted in Fig.\
\ref{fig:s-w-ensemble-average-electron}(b) for a fixed magnetic field $B=300$ G.
While the broadened
peak around $200$ MHz converges very quickly by a small amount of narrowing,
the additional low-frequency feature can now be resolved. This would
apply to all spectra shown in \ref{fig:s-w-ensemble-average-electron}(a): the additional
frequency resolution does not change the qualitative properties of the spin noise in
a transverse magnetic field but only reveals additional spectral information at low frequency.
Now we turn to hole spin ensembles modeled by the anisotropic CSM.
In this case the distribution
of the $g$-factors in the QDs is crucial when making connection
to measurements\cite{LiBayer2012,*ZapasskiiGreilichBayer2013}
of spin noise in those QD ensembles. To account for
this spread, we have used a Gaussian probability distribution for the
$g$-factors centered at $g$ with standard deviation $\Delta
g / g = 0.2$ to include a random value of $g$ for each
individual QD in our simulations. In addition,
the average dephasing time $\overline{T^*} =\expect{T^*}$ is set as reference time
(and reciprocal energy) scale, and
the individual time scales $T^*$ have been generated
from a constant distribution $T^*\in [0.2 T^*, 1.8
T^*]$ for each of the $n=2000$ individual QD calculations entering the ensemble
average. The resulting averaged $S(\omega)$ are depicted in Fig.\ \ref{fig:s-w-ensemble-average-hole}(a)
for two different values of $\lambda=5,50$.
It is striking that the width of the ensemble averaged $S(\omega)$ hardly depends
of the anisotropy factor $\lambda$, although
the peak width is given by $1/T^*_\lambda=1/\lambda T^*$ in a single QD.
In a QD ensemble of hole spins, however, the peak width of the noise spectra is determined by the
distribution of $g$-factors, and, consequently, the spectral width of $S(\omega)$
increases with increasing
external magnetic field.
To illustrate the influence of the $g$-factor distribution function
onto the shape of ensemble averaged $S(\omega)$,
we have combined
the $\lambda=5$ spectra of Fig.\ \ref{fig:s-w-ensemble-average-hole}(a)
with another set of ensemble averaged $S(\omega)$ calculated by
using a constant
distribution function $b_i/b \in [0.8,1.2]$
but otherwise identical parameters (green curves)
in Fig.\ \ref{fig:s-w-ensemble-average-hole}(b). Clearly, the peak shape
of $S(\omega)$ obtained from a box-shaped $g$-factor distribution function
resembles strongly this distribution function while the spectra
of Fig.\ \ref{fig:s-w-ensemble-average-hole}(a)
(red curves) exhibit a more Gaussian-type shape.
As a consequence, the shape of ensemble averaged spin-noise spectra
is mainly governed by the two distribution functions for $g$ and $T^*$ and reveal
less information about the dephasing mechanism in a single quantum dot.
Therefore, it would be desirable to be able to measure the spin noise directly on a single
quantum dot in order to obtain additional information on the microscopic dephasing
mechanism of an electron or hole spin in a QD.
\section{Discussion and outlook}
\label{sec:discussion-outlook}
We have presented a comprehensive study of the spin-noise in the minimal
model for dephasing in QDs, the anisotropic CSM. We have investigated
the dependency of $S(t)$ and its Fourier transform $S(\omega)$ on the
external magnetic field and the anisotropy parameter $\lambda$ using a
CET approach. This approach provides a time-evolution of an arbitrary
initial state whose error can be reduced to machine precision for any
give time $t$. We have extended the CET to ensemble averaging of
traces needed to account for a fully incoherent spin-bath.
In the absence of an external magnetic field, $S(t)$ exhibits an
universal time evolution for the isotropic CSM on a very short time
scale governed by the bath-spin fluctuation time $T^*$. At
intermediate times $t/T^*<100$, the dynamics depend on the
distribution function of the hyperfine coupling constants. While a
semi-classical approximation predicts a finite non-decaying fraction
of the spin polarization of $S^{\text{QSA}}_\infty=S(0)/3$ by averaging over
a distribution of static nuclear field configurations, our CET results
are typically below $S^{\text{QSA}}_\infty$. Fluctuations of the nuclear
spin bath occur on a much longer time scale and can lead to
significant deviations from the quasi-static approximation when the
hyperfine coupling constants cover a large parameter range. Typically
we find a very slow decay at intermediate and long-time scales driven
by the distribution of the small hyperfine interaction constants
$A_k$. Finite size scaling of the non-decaying fraction of the spin
polarization of $S_\infty$ as function of the cutoff $r_0$ indicates
vanishing $S_\infty$ in the thermodynamic limit in agreement with the
intuitive but not rigorous argument of Chen et
al.\cite{ChenBalents2007}
For the isotropic CSM, we find that the spin-noise spectrum consists
of three parts: (i) a zero-frequency $\delta$-peak whose spectral
weight is given by the non-decaying fraction of the central spin, (ii)
a low-frequency threshold behavior which follows approximately a
power-law and is cut off at the smallest non-zero excitation energy of
the finite-size system determined by $\bar A_{\rm min}(r_0)$ and (iii)
a Gaussian-type high frequency peak around $\omega T^*\approx 1$ which
determines the short time dynamics on time scales of $T^*$. We find a
remarkable agreement with a $\omega^{-3/4}$ fit to the low-frequency part
of our high resolution CET spin-noise spectrum which suggests an
asymptotic $1/t^{1/4}$ spin-decay at long times which is not
contradicting the approximative solutions in the literature: on
intermediate time scales $1/t^{1/4}$ and $1/\log(t)^{3/2}$ are nearly
indistinguishable and only deviate significantly in their asymptotic behavior.
We conjecture that in the thermodynamic limit the zero-frequency
$\delta$-peak will disappear, and the low-frequency spectra will show
a pronounced threshold behavior which will govern the long-time
spin-decay. For a detailed analysis of its analytical form, other
approaches have to be employed.
The Ising limit of the CSM is exactly solvable by treating the
$2\times2$ Hamiltonian subspaces for each frozen nuclear spin
configuration leading to a strict threshold behavior of the noise
spectra. The spectrum consists of a $\delta(0)$ peak whose spectral
weight decreases with increasing $b$ and close to the threshold a
$\omega^{-1/2}$ fit can be applied to the data, yielding a $t^{-1/2}$ long
time decay independently of the applied magnetic field. In the Ising
limit pure dephasing occurs and we find that an increase of the
external field yields an increasing coherence time of the central spin
due to the decreasing spread of the eigenenergies.
In the anisotropic CSM without an external magnetic field the
timescale $T^*_\lambda$ determines the short-time evolution of the
central spin. We have demonstrated that in a transversal magnetic
field the value of $b \cdot\lambda$ separates hyperfine interaction
driven spin dynamics from magnetic field dominated spin dynamics.
For $b \cdot\lambda > 1$, the ratio $b / \lambda$ describes the
crossover from Ising-like dynamics ($b / \lambda \ll 1$) to a more
Gaussian type decay ($b / \lambda > 1$.) For a single QD in a strong
external field we always find a clear energy gap in the spin-noise
between $\omega = 0$ and the broadened peak close to the Larmor
frequency whose width is given by $1/T^*_\lambda$.
Furthermore, we have performed an ensemble average of the
spin-noise spectra to compare our
calculations to recent spin-noise measurements on QD ensembles. By
using realistic characteristic time scales $T^*$, the experimentally
determined distribution of $g$-factors, we find qualitative
agreement for QD ensembles of both
hole spins and electron spins.
Our calculations reveal the shape dependence of
ensemble averaged noise spectra on the distribution functions for
the $g$-factors and the characteristic time scales $T^*$:
the averaged noise spectrum contains less information
on the dephasing mechanism than the data for an individual quantum dot.
We acknowledge fruitful discussions with Manfred Bayer, Scott Crooker,
Alex Greilich, Nikolai Sinitsyn,
Joachim Stolze, G\"otz Uhrig, and Dmitri Yakovlev.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.